1 /* Output routines for GCC for ARM.
2 Copyright (C) 1991-2018 Free Software Foundation, Inc.
3 Contributed by Pieter `Tiggr' Schoenmakers (rcpieter@win.tue.nl)
4 and Martin Simmons (@harleqn.co.uk).
5 More major hacks by Richard Earnshaw (rearnsha@arm.com).
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it
10 under the terms of the GNU General Public License as published
11 by the Free Software Foundation; either version 3, or (at your
12 option) any later version.
14 GCC is distributed in the hope that it will be useful, but WITHOUT
15 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
16 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public
17 License for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 #define IN_TARGET_CODE 1
26 #define INCLUDE_STRING
28 #include "coretypes.h"
37 #include "stringpool.h"
44 #include "diagnostic-core.h"
46 #include "fold-const.h"
47 #include "stor-layout.h"
51 #include "insn-attr.h"
57 #include "sched-int.h"
58 #include "common/common-target.h"
59 #include "langhooks.h"
65 #include "target-globals.h"
67 #include "tm-constrs.h"
69 #include "optabs-libfuncs.h"
74 /* This file should be included last. */
75 #include "target-def.h"
77 /* Forward definitions of types. */
78 typedef struct minipool_node Mnode
;
79 typedef struct minipool_fixup Mfix
;
81 /* The last .arch and .fpu assembly strings that we printed. */
82 static std::string arm_last_printed_arch_string
;
83 static std::string arm_last_printed_fpu_string
;
85 void (*arm_lang_output_object_attributes_hook
)(void);
92 /* Forward function declarations. */
93 static bool arm_const_not_ok_for_debug_p (rtx
);
94 static int arm_needs_doubleword_align (machine_mode
, const_tree
);
95 static int arm_compute_static_chain_stack_bytes (void);
96 static arm_stack_offsets
*arm_get_frame_offsets (void);
97 static void arm_compute_frame_layout (void);
98 static void arm_add_gc_roots (void);
99 static int arm_gen_constant (enum rtx_code
, machine_mode
, rtx
,
100 unsigned HOST_WIDE_INT
, rtx
, rtx
, int, int);
101 static unsigned bit_count (unsigned long);
102 static unsigned bitmap_popcount (const sbitmap
);
103 static int arm_address_register_rtx_p (rtx
, int);
104 static int arm_legitimate_index_p (machine_mode
, rtx
, RTX_CODE
, int);
105 static bool is_called_in_ARM_mode (tree
);
106 static int thumb2_legitimate_index_p (machine_mode
, rtx
, int);
107 static int thumb1_base_register_rtx_p (rtx
, machine_mode
, int);
108 static rtx
arm_legitimize_address (rtx
, rtx
, machine_mode
);
109 static reg_class_t
arm_preferred_reload_class (rtx
, reg_class_t
);
110 static rtx
thumb_legitimize_address (rtx
, rtx
, machine_mode
);
111 inline static int thumb1_index_register_rtx_p (rtx
, int);
112 static int thumb_far_jump_used_p (void);
113 static bool thumb_force_lr_save (void);
114 static unsigned arm_size_return_regs (void);
115 static bool arm_assemble_integer (rtx
, unsigned int, int);
116 static void arm_print_operand (FILE *, rtx
, int);
117 static void arm_print_operand_address (FILE *, machine_mode
, rtx
);
118 static bool arm_print_operand_punct_valid_p (unsigned char code
);
119 static const char *fp_const_from_val (REAL_VALUE_TYPE
*);
120 static arm_cc
get_arm_condition_code (rtx
);
121 static bool arm_fixed_condition_code_regs (unsigned int *, unsigned int *);
122 static const char *output_multi_immediate (rtx
*, const char *, const char *,
124 static const char *shift_op (rtx
, HOST_WIDE_INT
*);
125 static struct machine_function
*arm_init_machine_status (void);
126 static void thumb_exit (FILE *, int);
127 static HOST_WIDE_INT
get_jump_table_size (rtx_jump_table_data
*);
128 static Mnode
*move_minipool_fix_forward_ref (Mnode
*, Mnode
*, HOST_WIDE_INT
);
129 static Mnode
*add_minipool_forward_ref (Mfix
*);
130 static Mnode
*move_minipool_fix_backward_ref (Mnode
*, Mnode
*, HOST_WIDE_INT
);
131 static Mnode
*add_minipool_backward_ref (Mfix
*);
132 static void assign_minipool_offsets (Mfix
*);
133 static void arm_print_value (FILE *, rtx
);
134 static void dump_minipool (rtx_insn
*);
135 static int arm_barrier_cost (rtx_insn
*);
136 static Mfix
*create_fix_barrier (Mfix
*, HOST_WIDE_INT
);
137 static void push_minipool_barrier (rtx_insn
*, HOST_WIDE_INT
);
138 static void push_minipool_fix (rtx_insn
*, HOST_WIDE_INT
, rtx
*,
140 static void arm_reorg (void);
141 static void note_invalid_constants (rtx_insn
*, HOST_WIDE_INT
, int);
142 static unsigned long arm_compute_save_reg0_reg12_mask (void);
143 static unsigned long arm_compute_save_core_reg_mask (void);
144 static unsigned long arm_isr_value (tree
);
145 static unsigned long arm_compute_func_type (void);
146 static tree
arm_handle_fndecl_attribute (tree
*, tree
, tree
, int, bool *);
147 static tree
arm_handle_pcs_attribute (tree
*, tree
, tree
, int, bool *);
148 static tree
arm_handle_isr_attribute (tree
*, tree
, tree
, int, bool *);
149 #if TARGET_DLLIMPORT_DECL_ATTRIBUTES
150 static tree
arm_handle_notshared_attribute (tree
*, tree
, tree
, int, bool *);
152 static tree
arm_handle_cmse_nonsecure_entry (tree
*, tree
, tree
, int, bool *);
153 static tree
arm_handle_cmse_nonsecure_call (tree
*, tree
, tree
, int, bool *);
154 static void arm_output_function_epilogue (FILE *);
155 static void arm_output_function_prologue (FILE *);
156 static int arm_comp_type_attributes (const_tree
, const_tree
);
157 static void arm_set_default_type_attributes (tree
);
158 static int arm_adjust_cost (rtx_insn
*, int, rtx_insn
*, int, unsigned int);
159 static int arm_sched_reorder (FILE *, int, rtx_insn
**, int *, int);
160 static int optimal_immediate_sequence (enum rtx_code code
,
161 unsigned HOST_WIDE_INT val
,
162 struct four_ints
*return_sequence
);
163 static int optimal_immediate_sequence_1 (enum rtx_code code
,
164 unsigned HOST_WIDE_INT val
,
165 struct four_ints
*return_sequence
,
167 static int arm_get_strip_length (int);
168 static bool arm_function_ok_for_sibcall (tree
, tree
);
169 static machine_mode
arm_promote_function_mode (const_tree
,
172 static bool arm_return_in_memory (const_tree
, const_tree
);
173 static rtx
arm_function_value (const_tree
, const_tree
, bool);
174 static rtx
arm_libcall_value_1 (machine_mode
);
175 static rtx
arm_libcall_value (machine_mode
, const_rtx
);
176 static bool arm_function_value_regno_p (const unsigned int);
177 static void arm_internal_label (FILE *, const char *, unsigned long);
178 static void arm_output_mi_thunk (FILE *, tree
, HOST_WIDE_INT
, HOST_WIDE_INT
,
180 static bool arm_have_conditional_execution (void);
181 static bool arm_cannot_force_const_mem (machine_mode
, rtx
);
182 static bool arm_legitimate_constant_p (machine_mode
, rtx
);
183 static bool arm_rtx_costs (rtx
, machine_mode
, int, int, int *, bool);
184 static int arm_address_cost (rtx
, machine_mode
, addr_space_t
, bool);
185 static int arm_register_move_cost (machine_mode
, reg_class_t
, reg_class_t
);
186 static int arm_memory_move_cost (machine_mode
, reg_class_t
, bool);
187 static void emit_constant_insn (rtx cond
, rtx pattern
);
188 static rtx_insn
*emit_set_insn (rtx
, rtx
);
189 static rtx
emit_multi_reg_push (unsigned long, unsigned long);
190 static int arm_arg_partial_bytes (cumulative_args_t
, machine_mode
,
192 static rtx
arm_function_arg (cumulative_args_t
, machine_mode
,
194 static void arm_function_arg_advance (cumulative_args_t
, machine_mode
,
196 static pad_direction
arm_function_arg_padding (machine_mode
, const_tree
);
197 static unsigned int arm_function_arg_boundary (machine_mode
, const_tree
);
198 static rtx
aapcs_allocate_return_reg (machine_mode
, const_tree
,
200 static rtx
aapcs_libcall_value (machine_mode
);
201 static int aapcs_select_return_coproc (const_tree
, const_tree
);
203 #ifdef OBJECT_FORMAT_ELF
204 static void arm_elf_asm_constructor (rtx
, int) ATTRIBUTE_UNUSED
;
205 static void arm_elf_asm_destructor (rtx
, int) ATTRIBUTE_UNUSED
;
208 static void arm_encode_section_info (tree
, rtx
, int);
211 static void arm_file_end (void);
212 static void arm_file_start (void);
213 static void arm_insert_attributes (tree
, tree
*);
215 static void arm_setup_incoming_varargs (cumulative_args_t
, machine_mode
,
217 static bool arm_pass_by_reference (cumulative_args_t
,
218 machine_mode
, const_tree
, bool);
219 static bool arm_promote_prototypes (const_tree
);
220 static bool arm_default_short_enums (void);
221 static bool arm_align_anon_bitfield (void);
222 static bool arm_return_in_msb (const_tree
);
223 static bool arm_must_pass_in_stack (machine_mode
, const_tree
);
224 static bool arm_return_in_memory (const_tree
, const_tree
);
226 static void arm_unwind_emit (FILE *, rtx_insn
*);
227 static bool arm_output_ttype (rtx
);
228 static void arm_asm_emit_except_personality (rtx
);
230 static void arm_asm_init_sections (void);
231 static rtx
arm_dwarf_register_span (rtx
);
233 static tree
arm_cxx_guard_type (void);
234 static bool arm_cxx_guard_mask_bit (void);
235 static tree
arm_get_cookie_size (tree
);
236 static bool arm_cookie_has_size (void);
237 static bool arm_cxx_cdtor_returns_this (void);
238 static bool arm_cxx_key_method_may_be_inline (void);
239 static void arm_cxx_determine_class_data_visibility (tree
);
240 static bool arm_cxx_class_data_always_comdat (void);
241 static bool arm_cxx_use_aeabi_atexit (void);
242 static void arm_init_libfuncs (void);
243 static tree
arm_build_builtin_va_list (void);
244 static void arm_expand_builtin_va_start (tree
, rtx
);
245 static tree
arm_gimplify_va_arg_expr (tree
, tree
, gimple_seq
*, gimple_seq
*);
246 static void arm_option_override (void);
247 static void arm_option_save (struct cl_target_option
*, struct gcc_options
*);
248 static void arm_option_restore (struct gcc_options
*,
249 struct cl_target_option
*);
250 static void arm_override_options_after_change (void);
251 static void arm_option_print (FILE *, int, struct cl_target_option
*);
252 static void arm_set_current_function (tree
);
253 static bool arm_can_inline_p (tree
, tree
);
254 static void arm_relayout_function (tree
);
255 static bool arm_valid_target_attribute_p (tree
, tree
, tree
, int);
256 static unsigned HOST_WIDE_INT
arm_shift_truncation_mask (machine_mode
);
257 static bool arm_sched_can_speculate_insn (rtx_insn
*);
258 static bool arm_macro_fusion_p (void);
259 static bool arm_cannot_copy_insn_p (rtx_insn
*);
260 static int arm_issue_rate (void);
261 static int arm_first_cycle_multipass_dfa_lookahead (void);
262 static int arm_first_cycle_multipass_dfa_lookahead_guard (rtx_insn
*, int);
263 static void arm_output_dwarf_dtprel (FILE *, int, rtx
) ATTRIBUTE_UNUSED
;
264 static bool arm_output_addr_const_extra (FILE *, rtx
);
265 static bool arm_allocate_stack_slots_for_args (void);
266 static bool arm_warn_func_return (tree
);
267 static tree
arm_promoted_type (const_tree t
);
268 static bool arm_scalar_mode_supported_p (scalar_mode
);
269 static bool arm_frame_pointer_required (void);
270 static bool arm_can_eliminate (const int, const int);
271 static void arm_asm_trampoline_template (FILE *);
272 static void arm_trampoline_init (rtx
, tree
, rtx
);
273 static rtx
arm_trampoline_adjust_address (rtx
);
274 static rtx_insn
*arm_pic_static_addr (rtx orig
, rtx reg
);
275 static bool cortex_a9_sched_adjust_cost (rtx_insn
*, int, rtx_insn
*, int *);
276 static bool xscale_sched_adjust_cost (rtx_insn
*, int, rtx_insn
*, int *);
277 static bool fa726te_sched_adjust_cost (rtx_insn
*, int, rtx_insn
*, int *);
278 static bool arm_array_mode_supported_p (machine_mode
,
279 unsigned HOST_WIDE_INT
);
280 static machine_mode
arm_preferred_simd_mode (scalar_mode
);
281 static bool arm_class_likely_spilled_p (reg_class_t
);
282 static HOST_WIDE_INT
arm_vector_alignment (const_tree type
);
283 static bool arm_vector_alignment_reachable (const_tree type
, bool is_packed
);
284 static bool arm_builtin_support_vector_misalignment (machine_mode mode
,
288 static void arm_conditional_register_usage (void);
289 static enum flt_eval_method
arm_excess_precision (enum excess_precision_type
);
290 static reg_class_t
arm_preferred_rename_class (reg_class_t rclass
);
291 static void arm_autovectorize_vector_sizes (vector_sizes
*);
292 static int arm_default_branch_cost (bool, bool);
293 static int arm_cortex_a5_branch_cost (bool, bool);
294 static int arm_cortex_m_branch_cost (bool, bool);
295 static int arm_cortex_m7_branch_cost (bool, bool);
297 static bool arm_vectorize_vec_perm_const (machine_mode
, rtx
, rtx
, rtx
,
298 const vec_perm_indices
&);
300 static bool aarch_macro_fusion_pair_p (rtx_insn
*, rtx_insn
*);
302 static int arm_builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost
,
304 int misalign ATTRIBUTE_UNUSED
);
305 static unsigned arm_add_stmt_cost (void *data
, int count
,
306 enum vect_cost_for_stmt kind
,
307 struct _stmt_vec_info
*stmt_info
,
309 enum vect_cost_model_location where
);
311 static void arm_canonicalize_comparison (int *code
, rtx
*op0
, rtx
*op1
,
312 bool op0_preserve_value
);
313 static unsigned HOST_WIDE_INT
arm_asan_shadow_offset (void);
315 static void arm_sched_fusion_priority (rtx_insn
*, int, int *, int*);
316 static bool arm_can_output_mi_thunk (const_tree
, HOST_WIDE_INT
, HOST_WIDE_INT
,
318 static section
*arm_function_section (tree
, enum node_frequency
, bool, bool);
319 static bool arm_asm_elf_flags_numeric (unsigned int flags
, unsigned int *num
);
320 static unsigned int arm_elf_section_type_flags (tree decl
, const char *name
,
322 static void arm_expand_divmod_libfunc (rtx
, machine_mode
, rtx
, rtx
, rtx
*, rtx
*);
323 static opt_scalar_float_mode
arm_floatn_mode (int, bool);
324 static unsigned int arm_hard_regno_nregs (unsigned int, machine_mode
);
325 static bool arm_hard_regno_mode_ok (unsigned int, machine_mode
);
326 static bool arm_modes_tieable_p (machine_mode
, machine_mode
);
327 static HOST_WIDE_INT
arm_constant_alignment (const_tree
, HOST_WIDE_INT
);
329 /* Table of machine attributes. */
330 static const struct attribute_spec arm_attribute_table
[] =
332 /* { name, min_len, max_len, decl_req, type_req, fn_type_req,
333 affects_type_identity, handler, exclude } */
334 /* Function calls made to this symbol must be done indirectly, because
335 it may lie outside of the 26 bit addressing range of a normal function
337 { "long_call", 0, 0, false, true, true, false, NULL
, NULL
},
338 /* Whereas these functions are always known to reside within the 26 bit
340 { "short_call", 0, 0, false, true, true, false, NULL
, NULL
},
341 /* Specify the procedure call conventions for a function. */
342 { "pcs", 1, 1, false, true, true, false, arm_handle_pcs_attribute
,
344 /* Interrupt Service Routines have special prologue and epilogue requirements. */
345 { "isr", 0, 1, false, false, false, false, arm_handle_isr_attribute
,
347 { "interrupt", 0, 1, false, false, false, false, arm_handle_isr_attribute
,
349 { "naked", 0, 0, true, false, false, false,
350 arm_handle_fndecl_attribute
, NULL
},
352 /* ARM/PE has three new attributes:
354 dllexport - for exporting a function/variable that will live in a dll
355 dllimport - for importing a function/variable from a dll
357 Microsoft allows multiple declspecs in one __declspec, separating
358 them with spaces. We do NOT support this. Instead, use __declspec
361 { "dllimport", 0, 0, true, false, false, false, NULL
, NULL
},
362 { "dllexport", 0, 0, true, false, false, false, NULL
, NULL
},
363 { "interfacearm", 0, 0, true, false, false, false,
364 arm_handle_fndecl_attribute
, NULL
},
365 #elif TARGET_DLLIMPORT_DECL_ATTRIBUTES
366 { "dllimport", 0, 0, false, false, false, false, handle_dll_attribute
,
368 { "dllexport", 0, 0, false, false, false, false, handle_dll_attribute
,
370 { "notshared", 0, 0, false, true, false, false,
371 arm_handle_notshared_attribute
, NULL
},
373 /* ARMv8-M Security Extensions support. */
374 { "cmse_nonsecure_entry", 0, 0, true, false, false, false,
375 arm_handle_cmse_nonsecure_entry
, NULL
},
376 { "cmse_nonsecure_call", 0, 0, true, false, false, true,
377 arm_handle_cmse_nonsecure_call
, NULL
},
378 { NULL
, 0, 0, false, false, false, false, NULL
, NULL
}
381 /* Initialize the GCC target structure. */
382 #if TARGET_DLLIMPORT_DECL_ATTRIBUTES
383 #undef TARGET_MERGE_DECL_ATTRIBUTES
384 #define TARGET_MERGE_DECL_ATTRIBUTES merge_dllimport_decl_attributes
387 #undef TARGET_LEGITIMIZE_ADDRESS
388 #define TARGET_LEGITIMIZE_ADDRESS arm_legitimize_address
390 #undef TARGET_ATTRIBUTE_TABLE
391 #define TARGET_ATTRIBUTE_TABLE arm_attribute_table
393 #undef TARGET_INSERT_ATTRIBUTES
394 #define TARGET_INSERT_ATTRIBUTES arm_insert_attributes
396 #undef TARGET_ASM_FILE_START
397 #define TARGET_ASM_FILE_START arm_file_start
398 #undef TARGET_ASM_FILE_END
399 #define TARGET_ASM_FILE_END arm_file_end
401 #undef TARGET_ASM_ALIGNED_SI_OP
402 #define TARGET_ASM_ALIGNED_SI_OP NULL
403 #undef TARGET_ASM_INTEGER
404 #define TARGET_ASM_INTEGER arm_assemble_integer
406 #undef TARGET_PRINT_OPERAND
407 #define TARGET_PRINT_OPERAND arm_print_operand
408 #undef TARGET_PRINT_OPERAND_ADDRESS
409 #define TARGET_PRINT_OPERAND_ADDRESS arm_print_operand_address
410 #undef TARGET_PRINT_OPERAND_PUNCT_VALID_P
411 #define TARGET_PRINT_OPERAND_PUNCT_VALID_P arm_print_operand_punct_valid_p
413 #undef TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA
414 #define TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA arm_output_addr_const_extra
416 #undef TARGET_ASM_FUNCTION_PROLOGUE
417 #define TARGET_ASM_FUNCTION_PROLOGUE arm_output_function_prologue
419 #undef TARGET_ASM_FUNCTION_EPILOGUE
420 #define TARGET_ASM_FUNCTION_EPILOGUE arm_output_function_epilogue
422 #undef TARGET_CAN_INLINE_P
423 #define TARGET_CAN_INLINE_P arm_can_inline_p
425 #undef TARGET_RELAYOUT_FUNCTION
426 #define TARGET_RELAYOUT_FUNCTION arm_relayout_function
428 #undef TARGET_OPTION_OVERRIDE
429 #define TARGET_OPTION_OVERRIDE arm_option_override
431 #undef TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE
432 #define TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE arm_override_options_after_change
434 #undef TARGET_OPTION_SAVE
435 #define TARGET_OPTION_SAVE arm_option_save
437 #undef TARGET_OPTION_RESTORE
438 #define TARGET_OPTION_RESTORE arm_option_restore
440 #undef TARGET_OPTION_PRINT
441 #define TARGET_OPTION_PRINT arm_option_print
443 #undef TARGET_COMP_TYPE_ATTRIBUTES
444 #define TARGET_COMP_TYPE_ATTRIBUTES arm_comp_type_attributes
446 #undef TARGET_SCHED_CAN_SPECULATE_INSN
447 #define TARGET_SCHED_CAN_SPECULATE_INSN arm_sched_can_speculate_insn
449 #undef TARGET_SCHED_MACRO_FUSION_P
450 #define TARGET_SCHED_MACRO_FUSION_P arm_macro_fusion_p
452 #undef TARGET_SCHED_MACRO_FUSION_PAIR_P
453 #define TARGET_SCHED_MACRO_FUSION_PAIR_P aarch_macro_fusion_pair_p
455 #undef TARGET_SET_DEFAULT_TYPE_ATTRIBUTES
456 #define TARGET_SET_DEFAULT_TYPE_ATTRIBUTES arm_set_default_type_attributes
458 #undef TARGET_SCHED_ADJUST_COST
459 #define TARGET_SCHED_ADJUST_COST arm_adjust_cost
461 #undef TARGET_SET_CURRENT_FUNCTION
462 #define TARGET_SET_CURRENT_FUNCTION arm_set_current_function
464 #undef TARGET_OPTION_VALID_ATTRIBUTE_P
465 #define TARGET_OPTION_VALID_ATTRIBUTE_P arm_valid_target_attribute_p
467 #undef TARGET_SCHED_REORDER
468 #define TARGET_SCHED_REORDER arm_sched_reorder
470 #undef TARGET_REGISTER_MOVE_COST
471 #define TARGET_REGISTER_MOVE_COST arm_register_move_cost
473 #undef TARGET_MEMORY_MOVE_COST
474 #define TARGET_MEMORY_MOVE_COST arm_memory_move_cost
476 #undef TARGET_ENCODE_SECTION_INFO
478 #define TARGET_ENCODE_SECTION_INFO arm_pe_encode_section_info
480 #define TARGET_ENCODE_SECTION_INFO arm_encode_section_info
483 #undef TARGET_STRIP_NAME_ENCODING
484 #define TARGET_STRIP_NAME_ENCODING arm_strip_name_encoding
486 #undef TARGET_ASM_INTERNAL_LABEL
487 #define TARGET_ASM_INTERNAL_LABEL arm_internal_label
489 #undef TARGET_FLOATN_MODE
490 #define TARGET_FLOATN_MODE arm_floatn_mode
492 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
493 #define TARGET_FUNCTION_OK_FOR_SIBCALL arm_function_ok_for_sibcall
495 #undef TARGET_FUNCTION_VALUE
496 #define TARGET_FUNCTION_VALUE arm_function_value
498 #undef TARGET_LIBCALL_VALUE
499 #define TARGET_LIBCALL_VALUE arm_libcall_value
501 #undef TARGET_FUNCTION_VALUE_REGNO_P
502 #define TARGET_FUNCTION_VALUE_REGNO_P arm_function_value_regno_p
504 #undef TARGET_ASM_OUTPUT_MI_THUNK
505 #define TARGET_ASM_OUTPUT_MI_THUNK arm_output_mi_thunk
506 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
507 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK arm_can_output_mi_thunk
509 #undef TARGET_RTX_COSTS
510 #define TARGET_RTX_COSTS arm_rtx_costs
511 #undef TARGET_ADDRESS_COST
512 #define TARGET_ADDRESS_COST arm_address_cost
514 #undef TARGET_SHIFT_TRUNCATION_MASK
515 #define TARGET_SHIFT_TRUNCATION_MASK arm_shift_truncation_mask
516 #undef TARGET_VECTOR_MODE_SUPPORTED_P
517 #define TARGET_VECTOR_MODE_SUPPORTED_P arm_vector_mode_supported_p
518 #undef TARGET_ARRAY_MODE_SUPPORTED_P
519 #define TARGET_ARRAY_MODE_SUPPORTED_P arm_array_mode_supported_p
520 #undef TARGET_VECTORIZE_PREFERRED_SIMD_MODE
521 #define TARGET_VECTORIZE_PREFERRED_SIMD_MODE arm_preferred_simd_mode
522 #undef TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES
523 #define TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES \
524 arm_autovectorize_vector_sizes
526 #undef TARGET_MACHINE_DEPENDENT_REORG
527 #define TARGET_MACHINE_DEPENDENT_REORG arm_reorg
529 #undef TARGET_INIT_BUILTINS
530 #define TARGET_INIT_BUILTINS arm_init_builtins
531 #undef TARGET_EXPAND_BUILTIN
532 #define TARGET_EXPAND_BUILTIN arm_expand_builtin
533 #undef TARGET_BUILTIN_DECL
534 #define TARGET_BUILTIN_DECL arm_builtin_decl
536 #undef TARGET_INIT_LIBFUNCS
537 #define TARGET_INIT_LIBFUNCS arm_init_libfuncs
539 #undef TARGET_PROMOTE_FUNCTION_MODE
540 #define TARGET_PROMOTE_FUNCTION_MODE arm_promote_function_mode
541 #undef TARGET_PROMOTE_PROTOTYPES
542 #define TARGET_PROMOTE_PROTOTYPES arm_promote_prototypes
543 #undef TARGET_PASS_BY_REFERENCE
544 #define TARGET_PASS_BY_REFERENCE arm_pass_by_reference
545 #undef TARGET_ARG_PARTIAL_BYTES
546 #define TARGET_ARG_PARTIAL_BYTES arm_arg_partial_bytes
547 #undef TARGET_FUNCTION_ARG
548 #define TARGET_FUNCTION_ARG arm_function_arg
549 #undef TARGET_FUNCTION_ARG_ADVANCE
550 #define TARGET_FUNCTION_ARG_ADVANCE arm_function_arg_advance
551 #undef TARGET_FUNCTION_ARG_PADDING
552 #define TARGET_FUNCTION_ARG_PADDING arm_function_arg_padding
553 #undef TARGET_FUNCTION_ARG_BOUNDARY
554 #define TARGET_FUNCTION_ARG_BOUNDARY arm_function_arg_boundary
556 #undef TARGET_SETUP_INCOMING_VARARGS
557 #define TARGET_SETUP_INCOMING_VARARGS arm_setup_incoming_varargs
559 #undef TARGET_ALLOCATE_STACK_SLOTS_FOR_ARGS
560 #define TARGET_ALLOCATE_STACK_SLOTS_FOR_ARGS arm_allocate_stack_slots_for_args
562 #undef TARGET_ASM_TRAMPOLINE_TEMPLATE
563 #define TARGET_ASM_TRAMPOLINE_TEMPLATE arm_asm_trampoline_template
564 #undef TARGET_TRAMPOLINE_INIT
565 #define TARGET_TRAMPOLINE_INIT arm_trampoline_init
566 #undef TARGET_TRAMPOLINE_ADJUST_ADDRESS
567 #define TARGET_TRAMPOLINE_ADJUST_ADDRESS arm_trampoline_adjust_address
569 #undef TARGET_WARN_FUNC_RETURN
570 #define TARGET_WARN_FUNC_RETURN arm_warn_func_return
572 #undef TARGET_DEFAULT_SHORT_ENUMS
573 #define TARGET_DEFAULT_SHORT_ENUMS arm_default_short_enums
575 #undef TARGET_ALIGN_ANON_BITFIELD
576 #define TARGET_ALIGN_ANON_BITFIELD arm_align_anon_bitfield
578 #undef TARGET_NARROW_VOLATILE_BITFIELD
579 #define TARGET_NARROW_VOLATILE_BITFIELD hook_bool_void_false
581 #undef TARGET_CXX_GUARD_TYPE
582 #define TARGET_CXX_GUARD_TYPE arm_cxx_guard_type
584 #undef TARGET_CXX_GUARD_MASK_BIT
585 #define TARGET_CXX_GUARD_MASK_BIT arm_cxx_guard_mask_bit
587 #undef TARGET_CXX_GET_COOKIE_SIZE
588 #define TARGET_CXX_GET_COOKIE_SIZE arm_get_cookie_size
590 #undef TARGET_CXX_COOKIE_HAS_SIZE
591 #define TARGET_CXX_COOKIE_HAS_SIZE arm_cookie_has_size
593 #undef TARGET_CXX_CDTOR_RETURNS_THIS
594 #define TARGET_CXX_CDTOR_RETURNS_THIS arm_cxx_cdtor_returns_this
596 #undef TARGET_CXX_KEY_METHOD_MAY_BE_INLINE
597 #define TARGET_CXX_KEY_METHOD_MAY_BE_INLINE arm_cxx_key_method_may_be_inline
599 #undef TARGET_CXX_USE_AEABI_ATEXIT
600 #define TARGET_CXX_USE_AEABI_ATEXIT arm_cxx_use_aeabi_atexit
602 #undef TARGET_CXX_DETERMINE_CLASS_DATA_VISIBILITY
603 #define TARGET_CXX_DETERMINE_CLASS_DATA_VISIBILITY \
604 arm_cxx_determine_class_data_visibility
606 #undef TARGET_CXX_CLASS_DATA_ALWAYS_COMDAT
607 #define TARGET_CXX_CLASS_DATA_ALWAYS_COMDAT arm_cxx_class_data_always_comdat
609 #undef TARGET_RETURN_IN_MSB
610 #define TARGET_RETURN_IN_MSB arm_return_in_msb
612 #undef TARGET_RETURN_IN_MEMORY
613 #define TARGET_RETURN_IN_MEMORY arm_return_in_memory
615 #undef TARGET_MUST_PASS_IN_STACK
616 #define TARGET_MUST_PASS_IN_STACK arm_must_pass_in_stack
619 #undef TARGET_ASM_UNWIND_EMIT
620 #define TARGET_ASM_UNWIND_EMIT arm_unwind_emit
622 /* EABI unwinding tables use a different format for the typeinfo tables. */
623 #undef TARGET_ASM_TTYPE
624 #define TARGET_ASM_TTYPE arm_output_ttype
626 #undef TARGET_ARM_EABI_UNWINDER
627 #define TARGET_ARM_EABI_UNWINDER true
629 #undef TARGET_ASM_EMIT_EXCEPT_PERSONALITY
630 #define TARGET_ASM_EMIT_EXCEPT_PERSONALITY arm_asm_emit_except_personality
632 #endif /* ARM_UNWIND_INFO */
634 #undef TARGET_ASM_INIT_SECTIONS
635 #define TARGET_ASM_INIT_SECTIONS arm_asm_init_sections
637 #undef TARGET_DWARF_REGISTER_SPAN
638 #define TARGET_DWARF_REGISTER_SPAN arm_dwarf_register_span
640 #undef TARGET_CANNOT_COPY_INSN_P
641 #define TARGET_CANNOT_COPY_INSN_P arm_cannot_copy_insn_p
644 #undef TARGET_HAVE_TLS
645 #define TARGET_HAVE_TLS true
648 #undef TARGET_HAVE_CONDITIONAL_EXECUTION
649 #define TARGET_HAVE_CONDITIONAL_EXECUTION arm_have_conditional_execution
651 #undef TARGET_LEGITIMATE_CONSTANT_P
652 #define TARGET_LEGITIMATE_CONSTANT_P arm_legitimate_constant_p
654 #undef TARGET_CANNOT_FORCE_CONST_MEM
655 #define TARGET_CANNOT_FORCE_CONST_MEM arm_cannot_force_const_mem
657 #undef TARGET_MAX_ANCHOR_OFFSET
658 #define TARGET_MAX_ANCHOR_OFFSET 4095
660 /* The minimum is set such that the total size of the block
661 for a particular anchor is -4088 + 1 + 4095 bytes, which is
662 divisible by eight, ensuring natural spacing of anchors. */
663 #undef TARGET_MIN_ANCHOR_OFFSET
664 #define TARGET_MIN_ANCHOR_OFFSET -4088
666 #undef TARGET_SCHED_ISSUE_RATE
667 #define TARGET_SCHED_ISSUE_RATE arm_issue_rate
669 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
670 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD \
671 arm_first_cycle_multipass_dfa_lookahead
673 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD
674 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD \
675 arm_first_cycle_multipass_dfa_lookahead_guard
677 #undef TARGET_MANGLE_TYPE
678 #define TARGET_MANGLE_TYPE arm_mangle_type
680 #undef TARGET_ATOMIC_ASSIGN_EXPAND_FENV
681 #define TARGET_ATOMIC_ASSIGN_EXPAND_FENV arm_atomic_assign_expand_fenv
683 #undef TARGET_BUILD_BUILTIN_VA_LIST
684 #define TARGET_BUILD_BUILTIN_VA_LIST arm_build_builtin_va_list
685 #undef TARGET_EXPAND_BUILTIN_VA_START
686 #define TARGET_EXPAND_BUILTIN_VA_START arm_expand_builtin_va_start
687 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
688 #define TARGET_GIMPLIFY_VA_ARG_EXPR arm_gimplify_va_arg_expr
691 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
692 #define TARGET_ASM_OUTPUT_DWARF_DTPREL arm_output_dwarf_dtprel
695 #undef TARGET_LEGITIMATE_ADDRESS_P
696 #define TARGET_LEGITIMATE_ADDRESS_P arm_legitimate_address_p
698 #undef TARGET_PREFERRED_RELOAD_CLASS
699 #define TARGET_PREFERRED_RELOAD_CLASS arm_preferred_reload_class
701 #undef TARGET_PROMOTED_TYPE
702 #define TARGET_PROMOTED_TYPE arm_promoted_type
704 #undef TARGET_SCALAR_MODE_SUPPORTED_P
705 #define TARGET_SCALAR_MODE_SUPPORTED_P arm_scalar_mode_supported_p
707 #undef TARGET_COMPUTE_FRAME_LAYOUT
708 #define TARGET_COMPUTE_FRAME_LAYOUT arm_compute_frame_layout
710 #undef TARGET_FRAME_POINTER_REQUIRED
711 #define TARGET_FRAME_POINTER_REQUIRED arm_frame_pointer_required
713 #undef TARGET_CAN_ELIMINATE
714 #define TARGET_CAN_ELIMINATE arm_can_eliminate
716 #undef TARGET_CONDITIONAL_REGISTER_USAGE
717 #define TARGET_CONDITIONAL_REGISTER_USAGE arm_conditional_register_usage
719 #undef TARGET_CLASS_LIKELY_SPILLED_P
720 #define TARGET_CLASS_LIKELY_SPILLED_P arm_class_likely_spilled_p
722 #undef TARGET_VECTORIZE_BUILTINS
723 #define TARGET_VECTORIZE_BUILTINS
725 #undef TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION
726 #define TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION \
727 arm_builtin_vectorized_function
729 #undef TARGET_VECTOR_ALIGNMENT
730 #define TARGET_VECTOR_ALIGNMENT arm_vector_alignment
732 #undef TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE
733 #define TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE \
734 arm_vector_alignment_reachable
736 #undef TARGET_VECTORIZE_SUPPORT_VECTOR_MISALIGNMENT
737 #define TARGET_VECTORIZE_SUPPORT_VECTOR_MISALIGNMENT \
738 arm_builtin_support_vector_misalignment
740 #undef TARGET_PREFERRED_RENAME_CLASS
741 #define TARGET_PREFERRED_RENAME_CLASS \
742 arm_preferred_rename_class
744 #undef TARGET_VECTORIZE_VEC_PERM_CONST
745 #define TARGET_VECTORIZE_VEC_PERM_CONST arm_vectorize_vec_perm_const
747 #undef TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST
748 #define TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST \
749 arm_builtin_vectorization_cost
750 #undef TARGET_VECTORIZE_ADD_STMT_COST
751 #define TARGET_VECTORIZE_ADD_STMT_COST arm_add_stmt_cost
753 #undef TARGET_CANONICALIZE_COMPARISON
754 #define TARGET_CANONICALIZE_COMPARISON \
755 arm_canonicalize_comparison
757 #undef TARGET_ASAN_SHADOW_OFFSET
758 #define TARGET_ASAN_SHADOW_OFFSET arm_asan_shadow_offset
760 #undef MAX_INSN_PER_IT_BLOCK
761 #define MAX_INSN_PER_IT_BLOCK (arm_restrict_it ? 1 : 4)
763 #undef TARGET_CAN_USE_DOLOOP_P
764 #define TARGET_CAN_USE_DOLOOP_P can_use_doloop_if_innermost
766 #undef TARGET_CONST_NOT_OK_FOR_DEBUG_P
767 #define TARGET_CONST_NOT_OK_FOR_DEBUG_P arm_const_not_ok_for_debug_p
769 #undef TARGET_CALL_FUSAGE_CONTAINS_NON_CALLEE_CLOBBERS
770 #define TARGET_CALL_FUSAGE_CONTAINS_NON_CALLEE_CLOBBERS true
772 #undef TARGET_SCHED_FUSION_PRIORITY
773 #define TARGET_SCHED_FUSION_PRIORITY arm_sched_fusion_priority
775 #undef TARGET_ASM_FUNCTION_SECTION
776 #define TARGET_ASM_FUNCTION_SECTION arm_function_section
778 #undef TARGET_ASM_ELF_FLAGS_NUMERIC
779 #define TARGET_ASM_ELF_FLAGS_NUMERIC arm_asm_elf_flags_numeric
781 #undef TARGET_SECTION_TYPE_FLAGS
782 #define TARGET_SECTION_TYPE_FLAGS arm_elf_section_type_flags
784 #undef TARGET_EXPAND_DIVMOD_LIBFUNC
785 #define TARGET_EXPAND_DIVMOD_LIBFUNC arm_expand_divmod_libfunc
787 #undef TARGET_C_EXCESS_PRECISION
788 #define TARGET_C_EXCESS_PRECISION arm_excess_precision
790 /* Although the architecture reserves bits 0 and 1, only the former is
791 used for ARM/Thumb ISA selection in v7 and earlier versions. */
792 #undef TARGET_CUSTOM_FUNCTION_DESCRIPTORS
793 #define TARGET_CUSTOM_FUNCTION_DESCRIPTORS 2
795 #undef TARGET_FIXED_CONDITION_CODE_REGS
796 #define TARGET_FIXED_CONDITION_CODE_REGS arm_fixed_condition_code_regs
798 #undef TARGET_HARD_REGNO_NREGS
799 #define TARGET_HARD_REGNO_NREGS arm_hard_regno_nregs
800 #undef TARGET_HARD_REGNO_MODE_OK
801 #define TARGET_HARD_REGNO_MODE_OK arm_hard_regno_mode_ok
803 #undef TARGET_MODES_TIEABLE_P
804 #define TARGET_MODES_TIEABLE_P arm_modes_tieable_p
806 #undef TARGET_CAN_CHANGE_MODE_CLASS
807 #define TARGET_CAN_CHANGE_MODE_CLASS arm_can_change_mode_class
809 #undef TARGET_CONSTANT_ALIGNMENT
810 #define TARGET_CONSTANT_ALIGNMENT arm_constant_alignment
812 /* Obstack for minipool constant handling. */
813 static struct obstack minipool_obstack
;
814 static char * minipool_startobj
;
816 /* The maximum number of insns skipped which
817 will be conditionalised if possible. */
818 static int max_insns_skipped
= 5;
820 extern FILE * asm_out_file
;
822 /* True if we are currently building a constant table. */
823 int making_const_table
;
825 /* The processor for which instructions should be scheduled. */
826 enum processor_type arm_tune
= TARGET_CPU_arm_none
;
828 /* The current tuning set. */
829 const struct tune_params
*current_tune
;
831 /* Which floating point hardware to schedule for. */
834 /* Used for Thumb call_via trampolines. */
835 rtx thumb_call_via_label
[14];
836 static int thumb_call_reg_needed
;
838 /* The bits in this mask specify which instruction scheduling options should
840 unsigned int tune_flags
= 0;
842 /* The highest ARM architecture version supported by the
844 enum base_architecture arm_base_arch
= BASE_ARCH_0
;
846 /* Active target architecture and tuning. */
848 struct arm_build_target arm_active_target
;
850 /* The following are used in the arm.md file as equivalents to bits
851 in the above two flag variables. */
853 /* Nonzero if this chip supports the ARM Architecture 4 extensions. */
856 /* Nonzero if this chip supports the ARM Architecture 4t extensions. */
859 /* Nonzero if this chip supports the ARM Architecture 5T extensions. */
862 /* Nonzero if this chip supports the ARM Architecture 5TE extensions. */
865 /* Nonzero if this chip supports the ARM Architecture 6 extensions. */
868 /* Nonzero if this chip supports the ARM 6K extensions. */
871 /* Nonzero if this chip supports the ARM 6KZ extensions. */
874 /* Nonzero if instructions present in ARMv6-M can be used. */
877 /* Nonzero if this chip supports the ARM 7 extensions. */
880 /* Nonzero if this chip supports the Large Physical Address Extension. */
881 int arm_arch_lpae
= 0;
883 /* Nonzero if instructions not present in the 'M' profile can be used. */
884 int arm_arch_notm
= 0;
886 /* Nonzero if instructions present in ARMv7E-M can be used. */
889 /* Nonzero if instructions present in ARMv8 can be used. */
892 /* Nonzero if this chip supports the ARMv8.1 extensions. */
895 /* Nonzero if this chip supports the ARM Architecture 8.2 extensions. */
898 /* Nonzero if this chip supports the FP16 instructions extension of ARM
900 int arm_fp16_inst
= 0;
902 /* Nonzero if this chip can benefit from load scheduling. */
903 int arm_ld_sched
= 0;
905 /* Nonzero if this chip is a StrongARM. */
906 int arm_tune_strongarm
= 0;
908 /* Nonzero if this chip supports Intel Wireless MMX technology. */
909 int arm_arch_iwmmxt
= 0;
911 /* Nonzero if this chip supports Intel Wireless MMX2 technology. */
912 int arm_arch_iwmmxt2
= 0;
914 /* Nonzero if this chip is an XScale. */
915 int arm_arch_xscale
= 0;
917 /* Nonzero if tuning for XScale */
918 int arm_tune_xscale
= 0;
920 /* Nonzero if we want to tune for stores that access the write-buffer.
921 This typically means an ARM6 or ARM7 with MMU or MPU. */
922 int arm_tune_wbuf
= 0;
924 /* Nonzero if tuning for Cortex-A9. */
925 int arm_tune_cortex_a9
= 0;
927 /* Nonzero if we should define __THUMB_INTERWORK__ in the
929 XXX This is a bit of a hack, it's intended to help work around
930 problems in GLD which doesn't understand that armv5t code is
931 interworking clean. */
932 int arm_cpp_interwork
= 0;
934 /* Nonzero if chip supports Thumb 1. */
937 /* Nonzero if chip supports Thumb 2. */
940 /* Nonzero if chip supports integer division instruction. */
941 int arm_arch_arm_hwdiv
;
942 int arm_arch_thumb_hwdiv
;
944 /* Nonzero if chip disallows volatile memory access in IT block. */
945 int arm_arch_no_volatile_ce
;
947 /* Nonzero if we should use Neon to handle 64-bits operations rather
948 than core registers. */
949 int prefer_neon_for_64bits
= 0;
951 /* Nonzero if we shouldn't use literal pools. */
952 bool arm_disable_literal_pool
= false;
954 /* The register number to be used for the PIC offset register. */
955 unsigned arm_pic_register
= INVALID_REGNUM
;
957 enum arm_pcs arm_pcs_default
;
959 /* For an explanation of these variables, see final_prescan_insn below. */
961 /* arm_current_cc is also used for Thumb-2 cond_exec blocks. */
962 enum arm_cond_code arm_current_cc
;
965 int arm_target_label
;
966 /* The number of conditionally executed insns, including the current insn. */
967 int arm_condexec_count
= 0;
968 /* A bitmask specifying the patterns for the IT block.
969 Zero means do not output an IT block before this insn. */
970 int arm_condexec_mask
= 0;
971 /* The number of bits used in arm_condexec_mask. */
972 int arm_condexec_masklen
= 0;
974 /* Nonzero if chip supports the ARMv8 CRC instructions. */
975 int arm_arch_crc
= 0;
977 /* Nonzero if chip supports the AdvSIMD Dot Product instructions. */
978 int arm_arch_dotprod
= 0;
980 /* Nonzero if chip supports the ARMv8-M security extensions. */
981 int arm_arch_cmse
= 0;
983 /* Nonzero if the core has a very small, high-latency, multiply unit. */
984 int arm_m_profile_small_mul
= 0;
986 /* The condition codes of the ARM, and the inverse function. */
987 static const char * const arm_condition_codes
[] =
989 "eq", "ne", "cs", "cc", "mi", "pl", "vs", "vc",
990 "hi", "ls", "ge", "lt", "gt", "le", "al", "nv"
993 /* The register numbers in sequence, for passing to arm_gen_load_multiple. */
994 int arm_regs_in_sequence
[] =
996 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
999 #define ARM_LSL_NAME "lsl"
1000 #define streq(string1, string2) (strcmp (string1, string2) == 0)
1002 #define THUMB2_WORK_REGS (0xff & ~( (1 << THUMB_HARD_FRAME_POINTER_REGNUM) \
1003 | (1 << SP_REGNUM) | (1 << PC_REGNUM) \
1004 | (1 << PIC_OFFSET_TABLE_REGNUM)))
1006 /* Initialization code. */
1010 enum processor_type scheduler
;
1011 unsigned int tune_flags
;
1012 const struct tune_params
*tune
;
1015 #define ARM_PREFETCH_NOT_BENEFICIAL { 0, -1, -1 }
1016 #define ARM_PREFETCH_BENEFICIAL(num_slots,l1_size,l1_line_size) \
1023 /* arm generic vectorizer costs. */
1025 struct cpu_vec_costs arm_default_vec_cost
= {
1026 1, /* scalar_stmt_cost. */
1027 1, /* scalar load_cost. */
1028 1, /* scalar_store_cost. */
1029 1, /* vec_stmt_cost. */
1030 1, /* vec_to_scalar_cost. */
1031 1, /* scalar_to_vec_cost. */
1032 1, /* vec_align_load_cost. */
1033 1, /* vec_unalign_load_cost. */
1034 1, /* vec_unalign_store_cost. */
1035 1, /* vec_store_cost. */
1036 3, /* cond_taken_branch_cost. */
1037 1, /* cond_not_taken_branch_cost. */
1040 /* Cost tables for AArch32 + AArch64 cores should go in aarch-cost-tables.h */
1041 #include "aarch-cost-tables.h"
1045 const struct cpu_cost_table cortexa9_extra_costs
=
1052 COSTS_N_INSNS (1), /* shift_reg. */
1053 COSTS_N_INSNS (1), /* arith_shift. */
1054 COSTS_N_INSNS (2), /* arith_shift_reg. */
1056 COSTS_N_INSNS (1), /* log_shift_reg. */
1057 COSTS_N_INSNS (1), /* extend. */
1058 COSTS_N_INSNS (2), /* extend_arith. */
1059 COSTS_N_INSNS (1), /* bfi. */
1060 COSTS_N_INSNS (1), /* bfx. */
1064 true /* non_exec_costs_exec. */
1069 COSTS_N_INSNS (3), /* simple. */
1070 COSTS_N_INSNS (3), /* flag_setting. */
1071 COSTS_N_INSNS (2), /* extend. */
1072 COSTS_N_INSNS (3), /* add. */
1073 COSTS_N_INSNS (2), /* extend_add. */
1074 COSTS_N_INSNS (30) /* idiv. No HW div on Cortex A9. */
1078 0, /* simple (N/A). */
1079 0, /* flag_setting (N/A). */
1080 COSTS_N_INSNS (4), /* extend. */
1082 COSTS_N_INSNS (4), /* extend_add. */
1088 COSTS_N_INSNS (2), /* load. */
1089 COSTS_N_INSNS (2), /* load_sign_extend. */
1090 COSTS_N_INSNS (2), /* ldrd. */
1091 COSTS_N_INSNS (2), /* ldm_1st. */
1092 1, /* ldm_regs_per_insn_1st. */
1093 2, /* ldm_regs_per_insn_subsequent. */
1094 COSTS_N_INSNS (5), /* loadf. */
1095 COSTS_N_INSNS (5), /* loadd. */
1096 COSTS_N_INSNS (1), /* load_unaligned. */
1097 COSTS_N_INSNS (2), /* store. */
1098 COSTS_N_INSNS (2), /* strd. */
1099 COSTS_N_INSNS (2), /* stm_1st. */
1100 1, /* stm_regs_per_insn_1st. */
1101 2, /* stm_regs_per_insn_subsequent. */
1102 COSTS_N_INSNS (1), /* storef. */
1103 COSTS_N_INSNS (1), /* stored. */
1104 COSTS_N_INSNS (1), /* store_unaligned. */
1105 COSTS_N_INSNS (1), /* loadv. */
1106 COSTS_N_INSNS (1) /* storev. */
1111 COSTS_N_INSNS (14), /* div. */
1112 COSTS_N_INSNS (4), /* mult. */
1113 COSTS_N_INSNS (7), /* mult_addsub. */
1114 COSTS_N_INSNS (30), /* fma. */
1115 COSTS_N_INSNS (3), /* addsub. */
1116 COSTS_N_INSNS (1), /* fpconst. */
1117 COSTS_N_INSNS (1), /* neg. */
1118 COSTS_N_INSNS (3), /* compare. */
1119 COSTS_N_INSNS (3), /* widen. */
1120 COSTS_N_INSNS (3), /* narrow. */
1121 COSTS_N_INSNS (3), /* toint. */
1122 COSTS_N_INSNS (3), /* fromint. */
1123 COSTS_N_INSNS (3) /* roundint. */
1127 COSTS_N_INSNS (24), /* div. */
1128 COSTS_N_INSNS (5), /* mult. */
1129 COSTS_N_INSNS (8), /* mult_addsub. */
1130 COSTS_N_INSNS (30), /* fma. */
1131 COSTS_N_INSNS (3), /* addsub. */
1132 COSTS_N_INSNS (1), /* fpconst. */
1133 COSTS_N_INSNS (1), /* neg. */
1134 COSTS_N_INSNS (3), /* compare. */
1135 COSTS_N_INSNS (3), /* widen. */
1136 COSTS_N_INSNS (3), /* narrow. */
1137 COSTS_N_INSNS (3), /* toint. */
1138 COSTS_N_INSNS (3), /* fromint. */
1139 COSTS_N_INSNS (3) /* roundint. */
1144 COSTS_N_INSNS (1) /* alu. */
1148 const struct cpu_cost_table cortexa8_extra_costs
=
1154 COSTS_N_INSNS (1), /* shift. */
1156 COSTS_N_INSNS (1), /* arith_shift. */
1157 0, /* arith_shift_reg. */
1158 COSTS_N_INSNS (1), /* log_shift. */
1159 0, /* log_shift_reg. */
1161 0, /* extend_arith. */
1167 true /* non_exec_costs_exec. */
1172 COSTS_N_INSNS (1), /* simple. */
1173 COSTS_N_INSNS (1), /* flag_setting. */
1174 COSTS_N_INSNS (1), /* extend. */
1175 COSTS_N_INSNS (1), /* add. */
1176 COSTS_N_INSNS (1), /* extend_add. */
1177 COSTS_N_INSNS (30) /* idiv. No HW div on Cortex A8. */
1181 0, /* simple (N/A). */
1182 0, /* flag_setting (N/A). */
1183 COSTS_N_INSNS (2), /* extend. */
1185 COSTS_N_INSNS (2), /* extend_add. */
1191 COSTS_N_INSNS (1), /* load. */
1192 COSTS_N_INSNS (1), /* load_sign_extend. */
1193 COSTS_N_INSNS (1), /* ldrd. */
1194 COSTS_N_INSNS (1), /* ldm_1st. */
1195 1, /* ldm_regs_per_insn_1st. */
1196 2, /* ldm_regs_per_insn_subsequent. */
1197 COSTS_N_INSNS (1), /* loadf. */
1198 COSTS_N_INSNS (1), /* loadd. */
1199 COSTS_N_INSNS (1), /* load_unaligned. */
1200 COSTS_N_INSNS (1), /* store. */
1201 COSTS_N_INSNS (1), /* strd. */
1202 COSTS_N_INSNS (1), /* stm_1st. */
1203 1, /* stm_regs_per_insn_1st. */
1204 2, /* stm_regs_per_insn_subsequent. */
1205 COSTS_N_INSNS (1), /* storef. */
1206 COSTS_N_INSNS (1), /* stored. */
1207 COSTS_N_INSNS (1), /* store_unaligned. */
1208 COSTS_N_INSNS (1), /* loadv. */
1209 COSTS_N_INSNS (1) /* storev. */
1214 COSTS_N_INSNS (36), /* div. */
1215 COSTS_N_INSNS (11), /* mult. */
1216 COSTS_N_INSNS (20), /* mult_addsub. */
1217 COSTS_N_INSNS (30), /* fma. */
1218 COSTS_N_INSNS (9), /* addsub. */
1219 COSTS_N_INSNS (3), /* fpconst. */
1220 COSTS_N_INSNS (3), /* neg. */
1221 COSTS_N_INSNS (6), /* compare. */
1222 COSTS_N_INSNS (4), /* widen. */
1223 COSTS_N_INSNS (4), /* narrow. */
1224 COSTS_N_INSNS (8), /* toint. */
1225 COSTS_N_INSNS (8), /* fromint. */
1226 COSTS_N_INSNS (8) /* roundint. */
1230 COSTS_N_INSNS (64), /* div. */
1231 COSTS_N_INSNS (16), /* mult. */
1232 COSTS_N_INSNS (25), /* mult_addsub. */
1233 COSTS_N_INSNS (30), /* fma. */
1234 COSTS_N_INSNS (9), /* addsub. */
1235 COSTS_N_INSNS (3), /* fpconst. */
1236 COSTS_N_INSNS (3), /* neg. */
1237 COSTS_N_INSNS (6), /* compare. */
1238 COSTS_N_INSNS (6), /* widen. */
1239 COSTS_N_INSNS (6), /* narrow. */
1240 COSTS_N_INSNS (8), /* toint. */
1241 COSTS_N_INSNS (8), /* fromint. */
1242 COSTS_N_INSNS (8) /* roundint. */
1247 COSTS_N_INSNS (1) /* alu. */
1251 const struct cpu_cost_table cortexa5_extra_costs
=
1257 COSTS_N_INSNS (1), /* shift. */
1258 COSTS_N_INSNS (1), /* shift_reg. */
1259 COSTS_N_INSNS (1), /* arith_shift. */
1260 COSTS_N_INSNS (1), /* arith_shift_reg. */
1261 COSTS_N_INSNS (1), /* log_shift. */
1262 COSTS_N_INSNS (1), /* log_shift_reg. */
1263 COSTS_N_INSNS (1), /* extend. */
1264 COSTS_N_INSNS (1), /* extend_arith. */
1265 COSTS_N_INSNS (1), /* bfi. */
1266 COSTS_N_INSNS (1), /* bfx. */
1267 COSTS_N_INSNS (1), /* clz. */
1268 COSTS_N_INSNS (1), /* rev. */
1270 true /* non_exec_costs_exec. */
1277 COSTS_N_INSNS (1), /* flag_setting. */
1278 COSTS_N_INSNS (1), /* extend. */
1279 COSTS_N_INSNS (1), /* add. */
1280 COSTS_N_INSNS (1), /* extend_add. */
1281 COSTS_N_INSNS (7) /* idiv. */
1285 0, /* simple (N/A). */
1286 0, /* flag_setting (N/A). */
1287 COSTS_N_INSNS (1), /* extend. */
1289 COSTS_N_INSNS (2), /* extend_add. */
1295 COSTS_N_INSNS (1), /* load. */
1296 COSTS_N_INSNS (1), /* load_sign_extend. */
1297 COSTS_N_INSNS (6), /* ldrd. */
1298 COSTS_N_INSNS (1), /* ldm_1st. */
1299 1, /* ldm_regs_per_insn_1st. */
1300 2, /* ldm_regs_per_insn_subsequent. */
1301 COSTS_N_INSNS (2), /* loadf. */
1302 COSTS_N_INSNS (4), /* loadd. */
1303 COSTS_N_INSNS (1), /* load_unaligned. */
1304 COSTS_N_INSNS (1), /* store. */
1305 COSTS_N_INSNS (3), /* strd. */
1306 COSTS_N_INSNS (1), /* stm_1st. */
1307 1, /* stm_regs_per_insn_1st. */
1308 2, /* stm_regs_per_insn_subsequent. */
1309 COSTS_N_INSNS (2), /* storef. */
1310 COSTS_N_INSNS (2), /* stored. */
1311 COSTS_N_INSNS (1), /* store_unaligned. */
1312 COSTS_N_INSNS (1), /* loadv. */
1313 COSTS_N_INSNS (1) /* storev. */
1318 COSTS_N_INSNS (15), /* div. */
1319 COSTS_N_INSNS (3), /* mult. */
1320 COSTS_N_INSNS (7), /* mult_addsub. */
1321 COSTS_N_INSNS (7), /* fma. */
1322 COSTS_N_INSNS (3), /* addsub. */
1323 COSTS_N_INSNS (3), /* fpconst. */
1324 COSTS_N_INSNS (3), /* neg. */
1325 COSTS_N_INSNS (3), /* compare. */
1326 COSTS_N_INSNS (3), /* widen. */
1327 COSTS_N_INSNS (3), /* narrow. */
1328 COSTS_N_INSNS (3), /* toint. */
1329 COSTS_N_INSNS (3), /* fromint. */
1330 COSTS_N_INSNS (3) /* roundint. */
1334 COSTS_N_INSNS (30), /* div. */
1335 COSTS_N_INSNS (6), /* mult. */
1336 COSTS_N_INSNS (10), /* mult_addsub. */
1337 COSTS_N_INSNS (7), /* fma. */
1338 COSTS_N_INSNS (3), /* addsub. */
1339 COSTS_N_INSNS (3), /* fpconst. */
1340 COSTS_N_INSNS (3), /* neg. */
1341 COSTS_N_INSNS (3), /* compare. */
1342 COSTS_N_INSNS (3), /* widen. */
1343 COSTS_N_INSNS (3), /* narrow. */
1344 COSTS_N_INSNS (3), /* toint. */
1345 COSTS_N_INSNS (3), /* fromint. */
1346 COSTS_N_INSNS (3) /* roundint. */
1351 COSTS_N_INSNS (1) /* alu. */
1356 const struct cpu_cost_table cortexa7_extra_costs
=
1362 COSTS_N_INSNS (1), /* shift. */
1363 COSTS_N_INSNS (1), /* shift_reg. */
1364 COSTS_N_INSNS (1), /* arith_shift. */
1365 COSTS_N_INSNS (1), /* arith_shift_reg. */
1366 COSTS_N_INSNS (1), /* log_shift. */
1367 COSTS_N_INSNS (1), /* log_shift_reg. */
1368 COSTS_N_INSNS (1), /* extend. */
1369 COSTS_N_INSNS (1), /* extend_arith. */
1370 COSTS_N_INSNS (1), /* bfi. */
1371 COSTS_N_INSNS (1), /* bfx. */
1372 COSTS_N_INSNS (1), /* clz. */
1373 COSTS_N_INSNS (1), /* rev. */
1375 true /* non_exec_costs_exec. */
1382 COSTS_N_INSNS (1), /* flag_setting. */
1383 COSTS_N_INSNS (1), /* extend. */
1384 COSTS_N_INSNS (1), /* add. */
1385 COSTS_N_INSNS (1), /* extend_add. */
1386 COSTS_N_INSNS (7) /* idiv. */
1390 0, /* simple (N/A). */
1391 0, /* flag_setting (N/A). */
1392 COSTS_N_INSNS (1), /* extend. */
1394 COSTS_N_INSNS (2), /* extend_add. */
1400 COSTS_N_INSNS (1), /* load. */
1401 COSTS_N_INSNS (1), /* load_sign_extend. */
1402 COSTS_N_INSNS (3), /* ldrd. */
1403 COSTS_N_INSNS (1), /* ldm_1st. */
1404 1, /* ldm_regs_per_insn_1st. */
1405 2, /* ldm_regs_per_insn_subsequent. */
1406 COSTS_N_INSNS (2), /* loadf. */
1407 COSTS_N_INSNS (2), /* loadd. */
1408 COSTS_N_INSNS (1), /* load_unaligned. */
1409 COSTS_N_INSNS (1), /* store. */
1410 COSTS_N_INSNS (3), /* strd. */
1411 COSTS_N_INSNS (1), /* stm_1st. */
1412 1, /* stm_regs_per_insn_1st. */
1413 2, /* stm_regs_per_insn_subsequent. */
1414 COSTS_N_INSNS (2), /* storef. */
1415 COSTS_N_INSNS (2), /* stored. */
1416 COSTS_N_INSNS (1), /* store_unaligned. */
1417 COSTS_N_INSNS (1), /* loadv. */
1418 COSTS_N_INSNS (1) /* storev. */
1423 COSTS_N_INSNS (15), /* div. */
1424 COSTS_N_INSNS (3), /* mult. */
1425 COSTS_N_INSNS (7), /* mult_addsub. */
1426 COSTS_N_INSNS (7), /* fma. */
1427 COSTS_N_INSNS (3), /* addsub. */
1428 COSTS_N_INSNS (3), /* fpconst. */
1429 COSTS_N_INSNS (3), /* neg. */
1430 COSTS_N_INSNS (3), /* compare. */
1431 COSTS_N_INSNS (3), /* widen. */
1432 COSTS_N_INSNS (3), /* narrow. */
1433 COSTS_N_INSNS (3), /* toint. */
1434 COSTS_N_INSNS (3), /* fromint. */
1435 COSTS_N_INSNS (3) /* roundint. */
1439 COSTS_N_INSNS (30), /* div. */
1440 COSTS_N_INSNS (6), /* mult. */
1441 COSTS_N_INSNS (10), /* mult_addsub. */
1442 COSTS_N_INSNS (7), /* fma. */
1443 COSTS_N_INSNS (3), /* addsub. */
1444 COSTS_N_INSNS (3), /* fpconst. */
1445 COSTS_N_INSNS (3), /* neg. */
1446 COSTS_N_INSNS (3), /* compare. */
1447 COSTS_N_INSNS (3), /* widen. */
1448 COSTS_N_INSNS (3), /* narrow. */
1449 COSTS_N_INSNS (3), /* toint. */
1450 COSTS_N_INSNS (3), /* fromint. */
1451 COSTS_N_INSNS (3) /* roundint. */
1456 COSTS_N_INSNS (1) /* alu. */
1460 const struct cpu_cost_table cortexa12_extra_costs
=
1467 COSTS_N_INSNS (1), /* shift_reg. */
1468 COSTS_N_INSNS (1), /* arith_shift. */
1469 COSTS_N_INSNS (1), /* arith_shift_reg. */
1470 COSTS_N_INSNS (1), /* log_shift. */
1471 COSTS_N_INSNS (1), /* log_shift_reg. */
1473 COSTS_N_INSNS (1), /* extend_arith. */
1475 COSTS_N_INSNS (1), /* bfx. */
1476 COSTS_N_INSNS (1), /* clz. */
1477 COSTS_N_INSNS (1), /* rev. */
1479 true /* non_exec_costs_exec. */
1484 COSTS_N_INSNS (2), /* simple. */
1485 COSTS_N_INSNS (3), /* flag_setting. */
1486 COSTS_N_INSNS (2), /* extend. */
1487 COSTS_N_INSNS (3), /* add. */
1488 COSTS_N_INSNS (2), /* extend_add. */
1489 COSTS_N_INSNS (18) /* idiv. */
1493 0, /* simple (N/A). */
1494 0, /* flag_setting (N/A). */
1495 COSTS_N_INSNS (3), /* extend. */
1497 COSTS_N_INSNS (3), /* extend_add. */
1503 COSTS_N_INSNS (3), /* load. */
1504 COSTS_N_INSNS (3), /* load_sign_extend. */
1505 COSTS_N_INSNS (3), /* ldrd. */
1506 COSTS_N_INSNS (3), /* ldm_1st. */
1507 1, /* ldm_regs_per_insn_1st. */
1508 2, /* ldm_regs_per_insn_subsequent. */
1509 COSTS_N_INSNS (3), /* loadf. */
1510 COSTS_N_INSNS (3), /* loadd. */
1511 0, /* load_unaligned. */
1515 1, /* stm_regs_per_insn_1st. */
1516 2, /* stm_regs_per_insn_subsequent. */
1517 COSTS_N_INSNS (2), /* storef. */
1518 COSTS_N_INSNS (2), /* stored. */
1519 0, /* store_unaligned. */
1520 COSTS_N_INSNS (1), /* loadv. */
1521 COSTS_N_INSNS (1) /* storev. */
1526 COSTS_N_INSNS (17), /* div. */
1527 COSTS_N_INSNS (4), /* mult. */
1528 COSTS_N_INSNS (8), /* mult_addsub. */
1529 COSTS_N_INSNS (8), /* fma. */
1530 COSTS_N_INSNS (4), /* addsub. */
1531 COSTS_N_INSNS (2), /* fpconst. */
1532 COSTS_N_INSNS (2), /* neg. */
1533 COSTS_N_INSNS (2), /* compare. */
1534 COSTS_N_INSNS (4), /* widen. */
1535 COSTS_N_INSNS (4), /* narrow. */
1536 COSTS_N_INSNS (4), /* toint. */
1537 COSTS_N_INSNS (4), /* fromint. */
1538 COSTS_N_INSNS (4) /* roundint. */
1542 COSTS_N_INSNS (31), /* div. */
1543 COSTS_N_INSNS (4), /* mult. */
1544 COSTS_N_INSNS (8), /* mult_addsub. */
1545 COSTS_N_INSNS (8), /* fma. */
1546 COSTS_N_INSNS (4), /* addsub. */
1547 COSTS_N_INSNS (2), /* fpconst. */
1548 COSTS_N_INSNS (2), /* neg. */
1549 COSTS_N_INSNS (2), /* compare. */
1550 COSTS_N_INSNS (4), /* widen. */
1551 COSTS_N_INSNS (4), /* narrow. */
1552 COSTS_N_INSNS (4), /* toint. */
1553 COSTS_N_INSNS (4), /* fromint. */
1554 COSTS_N_INSNS (4) /* roundint. */
1559 COSTS_N_INSNS (1) /* alu. */
1563 const struct cpu_cost_table cortexa15_extra_costs
=
1571 COSTS_N_INSNS (1), /* arith_shift. */
1572 COSTS_N_INSNS (1), /* arith_shift_reg. */
1573 COSTS_N_INSNS (1), /* log_shift. */
1574 COSTS_N_INSNS (1), /* log_shift_reg. */
1576 COSTS_N_INSNS (1), /* extend_arith. */
1577 COSTS_N_INSNS (1), /* bfi. */
1582 true /* non_exec_costs_exec. */
1587 COSTS_N_INSNS (2), /* simple. */
1588 COSTS_N_INSNS (3), /* flag_setting. */
1589 COSTS_N_INSNS (2), /* extend. */
1590 COSTS_N_INSNS (2), /* add. */
1591 COSTS_N_INSNS (2), /* extend_add. */
1592 COSTS_N_INSNS (18) /* idiv. */
1596 0, /* simple (N/A). */
1597 0, /* flag_setting (N/A). */
1598 COSTS_N_INSNS (3), /* extend. */
1600 COSTS_N_INSNS (3), /* extend_add. */
1606 COSTS_N_INSNS (3), /* load. */
1607 COSTS_N_INSNS (3), /* load_sign_extend. */
1608 COSTS_N_INSNS (3), /* ldrd. */
1609 COSTS_N_INSNS (4), /* ldm_1st. */
1610 1, /* ldm_regs_per_insn_1st. */
1611 2, /* ldm_regs_per_insn_subsequent. */
1612 COSTS_N_INSNS (4), /* loadf. */
1613 COSTS_N_INSNS (4), /* loadd. */
1614 0, /* load_unaligned. */
1617 COSTS_N_INSNS (1), /* stm_1st. */
1618 1, /* stm_regs_per_insn_1st. */
1619 2, /* stm_regs_per_insn_subsequent. */
1622 0, /* store_unaligned. */
1623 COSTS_N_INSNS (1), /* loadv. */
1624 COSTS_N_INSNS (1) /* storev. */
1629 COSTS_N_INSNS (17), /* div. */
1630 COSTS_N_INSNS (4), /* mult. */
1631 COSTS_N_INSNS (8), /* mult_addsub. */
1632 COSTS_N_INSNS (8), /* fma. */
1633 COSTS_N_INSNS (4), /* addsub. */
1634 COSTS_N_INSNS (2), /* fpconst. */
1635 COSTS_N_INSNS (2), /* neg. */
1636 COSTS_N_INSNS (5), /* compare. */
1637 COSTS_N_INSNS (4), /* widen. */
1638 COSTS_N_INSNS (4), /* narrow. */
1639 COSTS_N_INSNS (4), /* toint. */
1640 COSTS_N_INSNS (4), /* fromint. */
1641 COSTS_N_INSNS (4) /* roundint. */
1645 COSTS_N_INSNS (31), /* div. */
1646 COSTS_N_INSNS (4), /* mult. */
1647 COSTS_N_INSNS (8), /* mult_addsub. */
1648 COSTS_N_INSNS (8), /* fma. */
1649 COSTS_N_INSNS (4), /* addsub. */
1650 COSTS_N_INSNS (2), /* fpconst. */
1651 COSTS_N_INSNS (2), /* neg. */
1652 COSTS_N_INSNS (2), /* compare. */
1653 COSTS_N_INSNS (4), /* widen. */
1654 COSTS_N_INSNS (4), /* narrow. */
1655 COSTS_N_INSNS (4), /* toint. */
1656 COSTS_N_INSNS (4), /* fromint. */
1657 COSTS_N_INSNS (4) /* roundint. */
1662 COSTS_N_INSNS (1) /* alu. */
1666 const struct cpu_cost_table v7m_extra_costs
=
1674 0, /* arith_shift. */
1675 COSTS_N_INSNS (1), /* arith_shift_reg. */
1677 COSTS_N_INSNS (1), /* log_shift_reg. */
1679 COSTS_N_INSNS (1), /* extend_arith. */
1684 COSTS_N_INSNS (1), /* non_exec. */
1685 false /* non_exec_costs_exec. */
1690 COSTS_N_INSNS (1), /* simple. */
1691 COSTS_N_INSNS (1), /* flag_setting. */
1692 COSTS_N_INSNS (2), /* extend. */
1693 COSTS_N_INSNS (1), /* add. */
1694 COSTS_N_INSNS (3), /* extend_add. */
1695 COSTS_N_INSNS (8) /* idiv. */
1699 0, /* simple (N/A). */
1700 0, /* flag_setting (N/A). */
1701 COSTS_N_INSNS (2), /* extend. */
1703 COSTS_N_INSNS (3), /* extend_add. */
1709 COSTS_N_INSNS (2), /* load. */
1710 0, /* load_sign_extend. */
1711 COSTS_N_INSNS (3), /* ldrd. */
1712 COSTS_N_INSNS (2), /* ldm_1st. */
1713 1, /* ldm_regs_per_insn_1st. */
1714 1, /* ldm_regs_per_insn_subsequent. */
1715 COSTS_N_INSNS (2), /* loadf. */
1716 COSTS_N_INSNS (3), /* loadd. */
1717 COSTS_N_INSNS (1), /* load_unaligned. */
1718 COSTS_N_INSNS (2), /* store. */
1719 COSTS_N_INSNS (3), /* strd. */
1720 COSTS_N_INSNS (2), /* stm_1st. */
1721 1, /* stm_regs_per_insn_1st. */
1722 1, /* stm_regs_per_insn_subsequent. */
1723 COSTS_N_INSNS (2), /* storef. */
1724 COSTS_N_INSNS (3), /* stored. */
1725 COSTS_N_INSNS (1), /* store_unaligned. */
1726 COSTS_N_INSNS (1), /* loadv. */
1727 COSTS_N_INSNS (1) /* storev. */
1732 COSTS_N_INSNS (7), /* div. */
1733 COSTS_N_INSNS (2), /* mult. */
1734 COSTS_N_INSNS (5), /* mult_addsub. */
1735 COSTS_N_INSNS (3), /* fma. */
1736 COSTS_N_INSNS (1), /* addsub. */
1748 COSTS_N_INSNS (15), /* div. */
1749 COSTS_N_INSNS (5), /* mult. */
1750 COSTS_N_INSNS (7), /* mult_addsub. */
1751 COSTS_N_INSNS (7), /* fma. */
1752 COSTS_N_INSNS (3), /* addsub. */
1765 COSTS_N_INSNS (1) /* alu. */
1769 const struct addr_mode_cost_table generic_addr_mode_costs
=
1773 COSTS_N_INSNS (0), /* AMO_DEFAULT. */
1774 COSTS_N_INSNS (0), /* AMO_NO_WB. */
1775 COSTS_N_INSNS (0) /* AMO_WB. */
1779 COSTS_N_INSNS (0), /* AMO_DEFAULT. */
1780 COSTS_N_INSNS (0), /* AMO_NO_WB. */
1781 COSTS_N_INSNS (0) /* AMO_WB. */
1785 COSTS_N_INSNS (0), /* AMO_DEFAULT. */
1786 COSTS_N_INSNS (0), /* AMO_NO_WB. */
1787 COSTS_N_INSNS (0) /* AMO_WB. */
1791 const struct tune_params arm_slowmul_tune
=
1793 &generic_extra_costs
, /* Insn extra costs. */
1794 &generic_addr_mode_costs
, /* Addressing mode costs. */
1795 NULL
, /* Sched adj cost. */
1796 arm_default_branch_cost
,
1797 &arm_default_vec_cost
,
1798 3, /* Constant limit. */
1799 5, /* Max cond insns. */
1800 8, /* Memset max inline. */
1801 1, /* Issue rate. */
1802 ARM_PREFETCH_NOT_BENEFICIAL
,
1803 tune_params::PREF_CONST_POOL_TRUE
,
1804 tune_params::PREF_LDRD_FALSE
,
1805 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
1806 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
1807 tune_params::DISPARAGE_FLAGS_NEITHER
,
1808 tune_params::PREF_NEON_64_FALSE
,
1809 tune_params::PREF_NEON_STRINGOPS_FALSE
,
1810 tune_params::FUSE_NOTHING
,
1811 tune_params::SCHED_AUTOPREF_OFF
1814 const struct tune_params arm_fastmul_tune
=
1816 &generic_extra_costs
, /* Insn extra costs. */
1817 &generic_addr_mode_costs
, /* Addressing mode costs. */
1818 NULL
, /* Sched adj cost. */
1819 arm_default_branch_cost
,
1820 &arm_default_vec_cost
,
1821 1, /* Constant limit. */
1822 5, /* Max cond insns. */
1823 8, /* Memset max inline. */
1824 1, /* Issue rate. */
1825 ARM_PREFETCH_NOT_BENEFICIAL
,
1826 tune_params::PREF_CONST_POOL_TRUE
,
1827 tune_params::PREF_LDRD_FALSE
,
1828 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
1829 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
1830 tune_params::DISPARAGE_FLAGS_NEITHER
,
1831 tune_params::PREF_NEON_64_FALSE
,
1832 tune_params::PREF_NEON_STRINGOPS_FALSE
,
1833 tune_params::FUSE_NOTHING
,
1834 tune_params::SCHED_AUTOPREF_OFF
1837 /* StrongARM has early execution of branches, so a sequence that is worth
1838 skipping is shorter. Set max_insns_skipped to a lower value. */
1840 const struct tune_params arm_strongarm_tune
=
1842 &generic_extra_costs
, /* Insn extra costs. */
1843 &generic_addr_mode_costs
, /* Addressing mode costs. */
1844 NULL
, /* Sched adj cost. */
1845 arm_default_branch_cost
,
1846 &arm_default_vec_cost
,
1847 1, /* Constant limit. */
1848 3, /* Max cond insns. */
1849 8, /* Memset max inline. */
1850 1, /* Issue rate. */
1851 ARM_PREFETCH_NOT_BENEFICIAL
,
1852 tune_params::PREF_CONST_POOL_TRUE
,
1853 tune_params::PREF_LDRD_FALSE
,
1854 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
1855 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
1856 tune_params::DISPARAGE_FLAGS_NEITHER
,
1857 tune_params::PREF_NEON_64_FALSE
,
1858 tune_params::PREF_NEON_STRINGOPS_FALSE
,
1859 tune_params::FUSE_NOTHING
,
1860 tune_params::SCHED_AUTOPREF_OFF
1863 const struct tune_params arm_xscale_tune
=
1865 &generic_extra_costs
, /* Insn extra costs. */
1866 &generic_addr_mode_costs
, /* Addressing mode costs. */
1867 xscale_sched_adjust_cost
,
1868 arm_default_branch_cost
,
1869 &arm_default_vec_cost
,
1870 2, /* Constant limit. */
1871 3, /* Max cond insns. */
1872 8, /* Memset max inline. */
1873 1, /* Issue rate. */
1874 ARM_PREFETCH_NOT_BENEFICIAL
,
1875 tune_params::PREF_CONST_POOL_TRUE
,
1876 tune_params::PREF_LDRD_FALSE
,
1877 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
1878 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
1879 tune_params::DISPARAGE_FLAGS_NEITHER
,
1880 tune_params::PREF_NEON_64_FALSE
,
1881 tune_params::PREF_NEON_STRINGOPS_FALSE
,
1882 tune_params::FUSE_NOTHING
,
1883 tune_params::SCHED_AUTOPREF_OFF
1886 const struct tune_params arm_9e_tune
=
1888 &generic_extra_costs
, /* Insn extra costs. */
1889 &generic_addr_mode_costs
, /* Addressing mode costs. */
1890 NULL
, /* Sched adj cost. */
1891 arm_default_branch_cost
,
1892 &arm_default_vec_cost
,
1893 1, /* Constant limit. */
1894 5, /* Max cond insns. */
1895 8, /* Memset max inline. */
1896 1, /* Issue rate. */
1897 ARM_PREFETCH_NOT_BENEFICIAL
,
1898 tune_params::PREF_CONST_POOL_TRUE
,
1899 tune_params::PREF_LDRD_FALSE
,
1900 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
1901 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
1902 tune_params::DISPARAGE_FLAGS_NEITHER
,
1903 tune_params::PREF_NEON_64_FALSE
,
1904 tune_params::PREF_NEON_STRINGOPS_FALSE
,
1905 tune_params::FUSE_NOTHING
,
1906 tune_params::SCHED_AUTOPREF_OFF
1909 const struct tune_params arm_marvell_pj4_tune
=
1911 &generic_extra_costs
, /* Insn extra costs. */
1912 &generic_addr_mode_costs
, /* Addressing mode costs. */
1913 NULL
, /* Sched adj cost. */
1914 arm_default_branch_cost
,
1915 &arm_default_vec_cost
,
1916 1, /* Constant limit. */
1917 5, /* Max cond insns. */
1918 8, /* Memset max inline. */
1919 2, /* Issue rate. */
1920 ARM_PREFETCH_NOT_BENEFICIAL
,
1921 tune_params::PREF_CONST_POOL_TRUE
,
1922 tune_params::PREF_LDRD_FALSE
,
1923 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
1924 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
1925 tune_params::DISPARAGE_FLAGS_NEITHER
,
1926 tune_params::PREF_NEON_64_FALSE
,
1927 tune_params::PREF_NEON_STRINGOPS_FALSE
,
1928 tune_params::FUSE_NOTHING
,
1929 tune_params::SCHED_AUTOPREF_OFF
1932 const struct tune_params arm_v6t2_tune
=
1934 &generic_extra_costs
, /* Insn extra costs. */
1935 &generic_addr_mode_costs
, /* Addressing mode costs. */
1936 NULL
, /* Sched adj cost. */
1937 arm_default_branch_cost
,
1938 &arm_default_vec_cost
,
1939 1, /* Constant limit. */
1940 5, /* Max cond insns. */
1941 8, /* Memset max inline. */
1942 1, /* Issue rate. */
1943 ARM_PREFETCH_NOT_BENEFICIAL
,
1944 tune_params::PREF_CONST_POOL_FALSE
,
1945 tune_params::PREF_LDRD_FALSE
,
1946 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
1947 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
1948 tune_params::DISPARAGE_FLAGS_NEITHER
,
1949 tune_params::PREF_NEON_64_FALSE
,
1950 tune_params::PREF_NEON_STRINGOPS_FALSE
,
1951 tune_params::FUSE_NOTHING
,
1952 tune_params::SCHED_AUTOPREF_OFF
1956 /* Generic Cortex tuning. Use more specific tunings if appropriate. */
1957 const struct tune_params arm_cortex_tune
=
1959 &generic_extra_costs
,
1960 &generic_addr_mode_costs
, /* Addressing mode costs. */
1961 NULL
, /* Sched adj cost. */
1962 arm_default_branch_cost
,
1963 &arm_default_vec_cost
,
1964 1, /* Constant limit. */
1965 5, /* Max cond insns. */
1966 8, /* Memset max inline. */
1967 2, /* Issue rate. */
1968 ARM_PREFETCH_NOT_BENEFICIAL
,
1969 tune_params::PREF_CONST_POOL_FALSE
,
1970 tune_params::PREF_LDRD_FALSE
,
1971 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
1972 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
1973 tune_params::DISPARAGE_FLAGS_NEITHER
,
1974 tune_params::PREF_NEON_64_FALSE
,
1975 tune_params::PREF_NEON_STRINGOPS_FALSE
,
1976 tune_params::FUSE_NOTHING
,
1977 tune_params::SCHED_AUTOPREF_OFF
1980 const struct tune_params arm_cortex_a8_tune
=
1982 &cortexa8_extra_costs
,
1983 &generic_addr_mode_costs
, /* Addressing mode costs. */
1984 NULL
, /* Sched adj cost. */
1985 arm_default_branch_cost
,
1986 &arm_default_vec_cost
,
1987 1, /* Constant limit. */
1988 5, /* Max cond insns. */
1989 8, /* Memset max inline. */
1990 2, /* Issue rate. */
1991 ARM_PREFETCH_NOT_BENEFICIAL
,
1992 tune_params::PREF_CONST_POOL_FALSE
,
1993 tune_params::PREF_LDRD_FALSE
,
1994 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
1995 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
1996 tune_params::DISPARAGE_FLAGS_NEITHER
,
1997 tune_params::PREF_NEON_64_FALSE
,
1998 tune_params::PREF_NEON_STRINGOPS_TRUE
,
1999 tune_params::FUSE_NOTHING
,
2000 tune_params::SCHED_AUTOPREF_OFF
2003 const struct tune_params arm_cortex_a7_tune
=
2005 &cortexa7_extra_costs
,
2006 &generic_addr_mode_costs
, /* Addressing mode costs. */
2007 NULL
, /* Sched adj cost. */
2008 arm_default_branch_cost
,
2009 &arm_default_vec_cost
,
2010 1, /* Constant limit. */
2011 5, /* Max cond insns. */
2012 8, /* Memset max inline. */
2013 2, /* Issue rate. */
2014 ARM_PREFETCH_NOT_BENEFICIAL
,
2015 tune_params::PREF_CONST_POOL_FALSE
,
2016 tune_params::PREF_LDRD_FALSE
,
2017 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
2018 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
2019 tune_params::DISPARAGE_FLAGS_NEITHER
,
2020 tune_params::PREF_NEON_64_FALSE
,
2021 tune_params::PREF_NEON_STRINGOPS_TRUE
,
2022 tune_params::FUSE_NOTHING
,
2023 tune_params::SCHED_AUTOPREF_OFF
2026 const struct tune_params arm_cortex_a15_tune
=
2028 &cortexa15_extra_costs
,
2029 &generic_addr_mode_costs
, /* Addressing mode costs. */
2030 NULL
, /* Sched adj cost. */
2031 arm_default_branch_cost
,
2032 &arm_default_vec_cost
,
2033 1, /* Constant limit. */
2034 2, /* Max cond insns. */
2035 8, /* Memset max inline. */
2036 3, /* Issue rate. */
2037 ARM_PREFETCH_NOT_BENEFICIAL
,
2038 tune_params::PREF_CONST_POOL_FALSE
,
2039 tune_params::PREF_LDRD_TRUE
,
2040 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
2041 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
2042 tune_params::DISPARAGE_FLAGS_ALL
,
2043 tune_params::PREF_NEON_64_FALSE
,
2044 tune_params::PREF_NEON_STRINGOPS_TRUE
,
2045 tune_params::FUSE_NOTHING
,
2046 tune_params::SCHED_AUTOPREF_FULL
2049 const struct tune_params arm_cortex_a35_tune
=
2051 &cortexa53_extra_costs
,
2052 &generic_addr_mode_costs
, /* Addressing mode costs. */
2053 NULL
, /* Sched adj cost. */
2054 arm_default_branch_cost
,
2055 &arm_default_vec_cost
,
2056 1, /* Constant limit. */
2057 5, /* Max cond insns. */
2058 8, /* Memset max inline. */
2059 1, /* Issue rate. */
2060 ARM_PREFETCH_NOT_BENEFICIAL
,
2061 tune_params::PREF_CONST_POOL_FALSE
,
2062 tune_params::PREF_LDRD_FALSE
,
2063 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
2064 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
2065 tune_params::DISPARAGE_FLAGS_NEITHER
,
2066 tune_params::PREF_NEON_64_FALSE
,
2067 tune_params::PREF_NEON_STRINGOPS_TRUE
,
2068 FUSE_OPS (tune_params::FUSE_MOVW_MOVT
),
2069 tune_params::SCHED_AUTOPREF_OFF
2072 const struct tune_params arm_cortex_a53_tune
=
2074 &cortexa53_extra_costs
,
2075 &generic_addr_mode_costs
, /* Addressing mode costs. */
2076 NULL
, /* Sched adj cost. */
2077 arm_default_branch_cost
,
2078 &arm_default_vec_cost
,
2079 1, /* Constant limit. */
2080 5, /* Max cond insns. */
2081 8, /* Memset max inline. */
2082 2, /* Issue rate. */
2083 ARM_PREFETCH_NOT_BENEFICIAL
,
2084 tune_params::PREF_CONST_POOL_FALSE
,
2085 tune_params::PREF_LDRD_FALSE
,
2086 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
2087 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
2088 tune_params::DISPARAGE_FLAGS_NEITHER
,
2089 tune_params::PREF_NEON_64_FALSE
,
2090 tune_params::PREF_NEON_STRINGOPS_TRUE
,
2091 FUSE_OPS (tune_params::FUSE_MOVW_MOVT
| tune_params::FUSE_AES_AESMC
),
2092 tune_params::SCHED_AUTOPREF_OFF
2095 const struct tune_params arm_cortex_a57_tune
=
2097 &cortexa57_extra_costs
,
2098 &generic_addr_mode_costs
, /* addressing mode costs */
2099 NULL
, /* Sched adj cost. */
2100 arm_default_branch_cost
,
2101 &arm_default_vec_cost
,
2102 1, /* Constant limit. */
2103 2, /* Max cond insns. */
2104 8, /* Memset max inline. */
2105 3, /* Issue rate. */
2106 ARM_PREFETCH_NOT_BENEFICIAL
,
2107 tune_params::PREF_CONST_POOL_FALSE
,
2108 tune_params::PREF_LDRD_TRUE
,
2109 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
2110 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
2111 tune_params::DISPARAGE_FLAGS_ALL
,
2112 tune_params::PREF_NEON_64_FALSE
,
2113 tune_params::PREF_NEON_STRINGOPS_TRUE
,
2114 FUSE_OPS (tune_params::FUSE_MOVW_MOVT
| tune_params::FUSE_AES_AESMC
),
2115 tune_params::SCHED_AUTOPREF_FULL
2118 const struct tune_params arm_exynosm1_tune
=
2120 &exynosm1_extra_costs
,
2121 &generic_addr_mode_costs
, /* Addressing mode costs. */
2122 NULL
, /* Sched adj cost. */
2123 arm_default_branch_cost
,
2124 &arm_default_vec_cost
,
2125 1, /* Constant limit. */
2126 2, /* Max cond insns. */
2127 8, /* Memset max inline. */
2128 3, /* Issue rate. */
2129 ARM_PREFETCH_NOT_BENEFICIAL
,
2130 tune_params::PREF_CONST_POOL_FALSE
,
2131 tune_params::PREF_LDRD_TRUE
,
2132 tune_params::LOG_OP_NON_SHORT_CIRCUIT_FALSE
, /* Thumb. */
2133 tune_params::LOG_OP_NON_SHORT_CIRCUIT_FALSE
, /* ARM. */
2134 tune_params::DISPARAGE_FLAGS_ALL
,
2135 tune_params::PREF_NEON_64_FALSE
,
2136 tune_params::PREF_NEON_STRINGOPS_TRUE
,
2137 tune_params::FUSE_NOTHING
,
2138 tune_params::SCHED_AUTOPREF_OFF
2141 const struct tune_params arm_xgene1_tune
=
2143 &xgene1_extra_costs
,
2144 &generic_addr_mode_costs
, /* Addressing mode costs. */
2145 NULL
, /* Sched adj cost. */
2146 arm_default_branch_cost
,
2147 &arm_default_vec_cost
,
2148 1, /* Constant limit. */
2149 2, /* Max cond insns. */
2150 32, /* Memset max inline. */
2151 4, /* Issue rate. */
2152 ARM_PREFETCH_NOT_BENEFICIAL
,
2153 tune_params::PREF_CONST_POOL_FALSE
,
2154 tune_params::PREF_LDRD_TRUE
,
2155 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
2156 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
2157 tune_params::DISPARAGE_FLAGS_ALL
,
2158 tune_params::PREF_NEON_64_FALSE
,
2159 tune_params::PREF_NEON_STRINGOPS_FALSE
,
2160 tune_params::FUSE_NOTHING
,
2161 tune_params::SCHED_AUTOPREF_OFF
2164 /* Branches can be dual-issued on Cortex-A5, so conditional execution is
2165 less appealing. Set max_insns_skipped to a low value. */
2167 const struct tune_params arm_cortex_a5_tune
=
2169 &cortexa5_extra_costs
,
2170 &generic_addr_mode_costs
, /* Addressing mode costs. */
2171 NULL
, /* Sched adj cost. */
2172 arm_cortex_a5_branch_cost
,
2173 &arm_default_vec_cost
,
2174 1, /* Constant limit. */
2175 1, /* Max cond insns. */
2176 8, /* Memset max inline. */
2177 2, /* Issue rate. */
2178 ARM_PREFETCH_NOT_BENEFICIAL
,
2179 tune_params::PREF_CONST_POOL_FALSE
,
2180 tune_params::PREF_LDRD_FALSE
,
2181 tune_params::LOG_OP_NON_SHORT_CIRCUIT_FALSE
, /* Thumb. */
2182 tune_params::LOG_OP_NON_SHORT_CIRCUIT_FALSE
, /* ARM. */
2183 tune_params::DISPARAGE_FLAGS_NEITHER
,
2184 tune_params::PREF_NEON_64_FALSE
,
2185 tune_params::PREF_NEON_STRINGOPS_TRUE
,
2186 tune_params::FUSE_NOTHING
,
2187 tune_params::SCHED_AUTOPREF_OFF
2190 const struct tune_params arm_cortex_a9_tune
=
2192 &cortexa9_extra_costs
,
2193 &generic_addr_mode_costs
, /* Addressing mode costs. */
2194 cortex_a9_sched_adjust_cost
,
2195 arm_default_branch_cost
,
2196 &arm_default_vec_cost
,
2197 1, /* Constant limit. */
2198 5, /* Max cond insns. */
2199 8, /* Memset max inline. */
2200 2, /* Issue rate. */
2201 ARM_PREFETCH_BENEFICIAL(4,32,32),
2202 tune_params::PREF_CONST_POOL_FALSE
,
2203 tune_params::PREF_LDRD_FALSE
,
2204 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
2205 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
2206 tune_params::DISPARAGE_FLAGS_NEITHER
,
2207 tune_params::PREF_NEON_64_FALSE
,
2208 tune_params::PREF_NEON_STRINGOPS_FALSE
,
2209 tune_params::FUSE_NOTHING
,
2210 tune_params::SCHED_AUTOPREF_OFF
2213 const struct tune_params arm_cortex_a12_tune
=
2215 &cortexa12_extra_costs
,
2216 &generic_addr_mode_costs
, /* Addressing mode costs. */
2217 NULL
, /* Sched adj cost. */
2218 arm_default_branch_cost
,
2219 &arm_default_vec_cost
, /* Vectorizer costs. */
2220 1, /* Constant limit. */
2221 2, /* Max cond insns. */
2222 8, /* Memset max inline. */
2223 2, /* Issue rate. */
2224 ARM_PREFETCH_NOT_BENEFICIAL
,
2225 tune_params::PREF_CONST_POOL_FALSE
,
2226 tune_params::PREF_LDRD_TRUE
,
2227 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
2228 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
2229 tune_params::DISPARAGE_FLAGS_ALL
,
2230 tune_params::PREF_NEON_64_FALSE
,
2231 tune_params::PREF_NEON_STRINGOPS_TRUE
,
2232 FUSE_OPS (tune_params::FUSE_MOVW_MOVT
),
2233 tune_params::SCHED_AUTOPREF_OFF
2236 const struct tune_params arm_cortex_a73_tune
=
2238 &cortexa57_extra_costs
,
2239 &generic_addr_mode_costs
, /* Addressing mode costs. */
2240 NULL
, /* Sched adj cost. */
2241 arm_default_branch_cost
,
2242 &arm_default_vec_cost
, /* Vectorizer costs. */
2243 1, /* Constant limit. */
2244 2, /* Max cond insns. */
2245 8, /* Memset max inline. */
2246 2, /* Issue rate. */
2247 ARM_PREFETCH_NOT_BENEFICIAL
,
2248 tune_params::PREF_CONST_POOL_FALSE
,
2249 tune_params::PREF_LDRD_TRUE
,
2250 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
2251 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
2252 tune_params::DISPARAGE_FLAGS_ALL
,
2253 tune_params::PREF_NEON_64_FALSE
,
2254 tune_params::PREF_NEON_STRINGOPS_TRUE
,
2255 FUSE_OPS (tune_params::FUSE_AES_AESMC
| tune_params::FUSE_MOVW_MOVT
),
2256 tune_params::SCHED_AUTOPREF_FULL
2259 /* armv7m tuning. On Cortex-M4 cores for example, MOVW/MOVT take a single
2260 cycle to execute each. An LDR from the constant pool also takes two cycles
2261 to execute, but mildly increases pipelining opportunity (consecutive
2262 loads/stores can be pipelined together, saving one cycle), and may also
2263 improve icache utilisation. Hence we prefer the constant pool for such
2266 const struct tune_params arm_v7m_tune
=
2269 &generic_addr_mode_costs
, /* Addressing mode costs. */
2270 NULL
, /* Sched adj cost. */
2271 arm_cortex_m_branch_cost
,
2272 &arm_default_vec_cost
,
2273 1, /* Constant limit. */
2274 2, /* Max cond insns. */
2275 8, /* Memset max inline. */
2276 1, /* Issue rate. */
2277 ARM_PREFETCH_NOT_BENEFICIAL
,
2278 tune_params::PREF_CONST_POOL_TRUE
,
2279 tune_params::PREF_LDRD_FALSE
,
2280 tune_params::LOG_OP_NON_SHORT_CIRCUIT_FALSE
, /* Thumb. */
2281 tune_params::LOG_OP_NON_SHORT_CIRCUIT_FALSE
, /* ARM. */
2282 tune_params::DISPARAGE_FLAGS_NEITHER
,
2283 tune_params::PREF_NEON_64_FALSE
,
2284 tune_params::PREF_NEON_STRINGOPS_FALSE
,
2285 tune_params::FUSE_NOTHING
,
2286 tune_params::SCHED_AUTOPREF_OFF
2289 /* Cortex-M7 tuning. */
2291 const struct tune_params arm_cortex_m7_tune
=
2294 &generic_addr_mode_costs
, /* Addressing mode costs. */
2295 NULL
, /* Sched adj cost. */
2296 arm_cortex_m7_branch_cost
,
2297 &arm_default_vec_cost
,
2298 0, /* Constant limit. */
2299 1, /* Max cond insns. */
2300 8, /* Memset max inline. */
2301 2, /* Issue rate. */
2302 ARM_PREFETCH_NOT_BENEFICIAL
,
2303 tune_params::PREF_CONST_POOL_TRUE
,
2304 tune_params::PREF_LDRD_FALSE
,
2305 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
2306 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
2307 tune_params::DISPARAGE_FLAGS_NEITHER
,
2308 tune_params::PREF_NEON_64_FALSE
,
2309 tune_params::PREF_NEON_STRINGOPS_FALSE
,
2310 tune_params::FUSE_NOTHING
,
2311 tune_params::SCHED_AUTOPREF_OFF
2314 /* The arm_v6m_tune is duplicated from arm_cortex_tune, rather than
2315 arm_v6t2_tune. It is used for cortex-m0, cortex-m1, cortex-m0plus and
2317 const struct tune_params arm_v6m_tune
=
2319 &generic_extra_costs
, /* Insn extra costs. */
2320 &generic_addr_mode_costs
, /* Addressing mode costs. */
2321 NULL
, /* Sched adj cost. */
2322 arm_default_branch_cost
,
2323 &arm_default_vec_cost
, /* Vectorizer costs. */
2324 1, /* Constant limit. */
2325 5, /* Max cond insns. */
2326 8, /* Memset max inline. */
2327 1, /* Issue rate. */
2328 ARM_PREFETCH_NOT_BENEFICIAL
,
2329 tune_params::PREF_CONST_POOL_FALSE
,
2330 tune_params::PREF_LDRD_FALSE
,
2331 tune_params::LOG_OP_NON_SHORT_CIRCUIT_FALSE
, /* Thumb. */
2332 tune_params::LOG_OP_NON_SHORT_CIRCUIT_FALSE
, /* ARM. */
2333 tune_params::DISPARAGE_FLAGS_NEITHER
,
2334 tune_params::PREF_NEON_64_FALSE
,
2335 tune_params::PREF_NEON_STRINGOPS_FALSE
,
2336 tune_params::FUSE_NOTHING
,
2337 tune_params::SCHED_AUTOPREF_OFF
2340 const struct tune_params arm_fa726te_tune
=
2342 &generic_extra_costs
, /* Insn extra costs. */
2343 &generic_addr_mode_costs
, /* Addressing mode costs. */
2344 fa726te_sched_adjust_cost
,
2345 arm_default_branch_cost
,
2346 &arm_default_vec_cost
,
2347 1, /* Constant limit. */
2348 5, /* Max cond insns. */
2349 8, /* Memset max inline. */
2350 2, /* Issue rate. */
2351 ARM_PREFETCH_NOT_BENEFICIAL
,
2352 tune_params::PREF_CONST_POOL_TRUE
,
2353 tune_params::PREF_LDRD_FALSE
,
2354 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* Thumb. */
2355 tune_params::LOG_OP_NON_SHORT_CIRCUIT_TRUE
, /* ARM. */
2356 tune_params::DISPARAGE_FLAGS_NEITHER
,
2357 tune_params::PREF_NEON_64_FALSE
,
2358 tune_params::PREF_NEON_STRINGOPS_FALSE
,
2359 tune_params::FUSE_NOTHING
,
2360 tune_params::SCHED_AUTOPREF_OFF
2363 /* Auto-generated CPU, FPU and architecture tables. */
2364 #include "arm-cpu-data.h"
2366 /* The name of the preprocessor macro to define for this architecture. PROFILE
2367 is replaced by the architecture name (eg. 8A) in arm_option_override () and
2368 is thus chosen to be big enough to hold the longest architecture name. */
2370 char arm_arch_name
[] = "__ARM_ARCH_PROFILE__";
2372 /* Supported TLS relocations. */
2380 TLS_DESCSEQ
/* GNU scheme */
2383 /* The maximum number of insns to be used when loading a constant. */
2385 arm_constant_limit (bool size_p
)
2387 return size_p
? 1 : current_tune
->constant_limit
;
2390 /* Emit an insn that's a simple single-set. Both the operands must be known
2392 inline static rtx_insn
*
2393 emit_set_insn (rtx x
, rtx y
)
2395 return emit_insn (gen_rtx_SET (x
, y
));
2398 /* Return the number of bits set in VALUE. */
2400 bit_count (unsigned long value
)
2402 unsigned long count
= 0;
2407 value
&= value
- 1; /* Clear the least-significant set bit. */
2413 /* Return the number of bits set in BMAP. */
2415 bitmap_popcount (const sbitmap bmap
)
2417 unsigned int count
= 0;
2419 sbitmap_iterator sbi
;
2421 EXECUTE_IF_SET_IN_BITMAP (bmap
, 0, n
, sbi
)
2430 } arm_fixed_mode_set
;
2432 /* A small helper for setting fixed-point library libfuncs. */
2435 arm_set_fixed_optab_libfunc (optab optable
, machine_mode mode
,
2436 const char *funcname
, const char *modename
,
2441 if (num_suffix
== 0)
2442 sprintf (buffer
, "__gnu_%s%s", funcname
, modename
);
2444 sprintf (buffer
, "__gnu_%s%s%d", funcname
, modename
, num_suffix
);
2446 set_optab_libfunc (optable
, mode
, buffer
);
2450 arm_set_fixed_conv_libfunc (convert_optab optable
, machine_mode to
,
2451 machine_mode from
, const char *funcname
,
2452 const char *toname
, const char *fromname
)
2455 const char *maybe_suffix_2
= "";
2457 /* Follow the logic for selecting a "2" suffix in fixed-bit.h. */
2458 if (ALL_FIXED_POINT_MODE_P (from
) && ALL_FIXED_POINT_MODE_P (to
)
2459 && UNSIGNED_FIXED_POINT_MODE_P (from
) == UNSIGNED_FIXED_POINT_MODE_P (to
)
2460 && ALL_FRACT_MODE_P (from
) == ALL_FRACT_MODE_P (to
))
2461 maybe_suffix_2
= "2";
2463 sprintf (buffer
, "__gnu_%s%s%s%s", funcname
, fromname
, toname
,
2466 set_conv_libfunc (optable
, to
, from
, buffer
);
2469 static GTY(()) rtx speculation_barrier_libfunc
;
2471 /* Set up library functions unique to ARM. */
2473 arm_init_libfuncs (void)
2475 /* For Linux, we have access to kernel support for atomic operations. */
2476 if (arm_abi
== ARM_ABI_AAPCS_LINUX
)
2477 init_sync_libfuncs (MAX_SYNC_LIBFUNC_SIZE
);
2479 /* There are no special library functions unless we are using the
2484 /* The functions below are described in Section 4 of the "Run-Time
2485 ABI for the ARM architecture", Version 1.0. */
2487 /* Double-precision floating-point arithmetic. Table 2. */
2488 set_optab_libfunc (add_optab
, DFmode
, "__aeabi_dadd");
2489 set_optab_libfunc (sdiv_optab
, DFmode
, "__aeabi_ddiv");
2490 set_optab_libfunc (smul_optab
, DFmode
, "__aeabi_dmul");
2491 set_optab_libfunc (neg_optab
, DFmode
, "__aeabi_dneg");
2492 set_optab_libfunc (sub_optab
, DFmode
, "__aeabi_dsub");
2494 /* Double-precision comparisons. Table 3. */
2495 set_optab_libfunc (eq_optab
, DFmode
, "__aeabi_dcmpeq");
2496 set_optab_libfunc (ne_optab
, DFmode
, NULL
);
2497 set_optab_libfunc (lt_optab
, DFmode
, "__aeabi_dcmplt");
2498 set_optab_libfunc (le_optab
, DFmode
, "__aeabi_dcmple");
2499 set_optab_libfunc (ge_optab
, DFmode
, "__aeabi_dcmpge");
2500 set_optab_libfunc (gt_optab
, DFmode
, "__aeabi_dcmpgt");
2501 set_optab_libfunc (unord_optab
, DFmode
, "__aeabi_dcmpun");
2503 /* Single-precision floating-point arithmetic. Table 4. */
2504 set_optab_libfunc (add_optab
, SFmode
, "__aeabi_fadd");
2505 set_optab_libfunc (sdiv_optab
, SFmode
, "__aeabi_fdiv");
2506 set_optab_libfunc (smul_optab
, SFmode
, "__aeabi_fmul");
2507 set_optab_libfunc (neg_optab
, SFmode
, "__aeabi_fneg");
2508 set_optab_libfunc (sub_optab
, SFmode
, "__aeabi_fsub");
2510 /* Single-precision comparisons. Table 5. */
2511 set_optab_libfunc (eq_optab
, SFmode
, "__aeabi_fcmpeq");
2512 set_optab_libfunc (ne_optab
, SFmode
, NULL
);
2513 set_optab_libfunc (lt_optab
, SFmode
, "__aeabi_fcmplt");
2514 set_optab_libfunc (le_optab
, SFmode
, "__aeabi_fcmple");
2515 set_optab_libfunc (ge_optab
, SFmode
, "__aeabi_fcmpge");
2516 set_optab_libfunc (gt_optab
, SFmode
, "__aeabi_fcmpgt");
2517 set_optab_libfunc (unord_optab
, SFmode
, "__aeabi_fcmpun");
2519 /* Floating-point to integer conversions. Table 6. */
2520 set_conv_libfunc (sfix_optab
, SImode
, DFmode
, "__aeabi_d2iz");
2521 set_conv_libfunc (ufix_optab
, SImode
, DFmode
, "__aeabi_d2uiz");
2522 set_conv_libfunc (sfix_optab
, DImode
, DFmode
, "__aeabi_d2lz");
2523 set_conv_libfunc (ufix_optab
, DImode
, DFmode
, "__aeabi_d2ulz");
2524 set_conv_libfunc (sfix_optab
, SImode
, SFmode
, "__aeabi_f2iz");
2525 set_conv_libfunc (ufix_optab
, SImode
, SFmode
, "__aeabi_f2uiz");
2526 set_conv_libfunc (sfix_optab
, DImode
, SFmode
, "__aeabi_f2lz");
2527 set_conv_libfunc (ufix_optab
, DImode
, SFmode
, "__aeabi_f2ulz");
2529 /* Conversions between floating types. Table 7. */
2530 set_conv_libfunc (trunc_optab
, SFmode
, DFmode
, "__aeabi_d2f");
2531 set_conv_libfunc (sext_optab
, DFmode
, SFmode
, "__aeabi_f2d");
2533 /* Integer to floating-point conversions. Table 8. */
2534 set_conv_libfunc (sfloat_optab
, DFmode
, SImode
, "__aeabi_i2d");
2535 set_conv_libfunc (ufloat_optab
, DFmode
, SImode
, "__aeabi_ui2d");
2536 set_conv_libfunc (sfloat_optab
, DFmode
, DImode
, "__aeabi_l2d");
2537 set_conv_libfunc (ufloat_optab
, DFmode
, DImode
, "__aeabi_ul2d");
2538 set_conv_libfunc (sfloat_optab
, SFmode
, SImode
, "__aeabi_i2f");
2539 set_conv_libfunc (ufloat_optab
, SFmode
, SImode
, "__aeabi_ui2f");
2540 set_conv_libfunc (sfloat_optab
, SFmode
, DImode
, "__aeabi_l2f");
2541 set_conv_libfunc (ufloat_optab
, SFmode
, DImode
, "__aeabi_ul2f");
2543 /* Long long. Table 9. */
2544 set_optab_libfunc (smul_optab
, DImode
, "__aeabi_lmul");
2545 set_optab_libfunc (sdivmod_optab
, DImode
, "__aeabi_ldivmod");
2546 set_optab_libfunc (udivmod_optab
, DImode
, "__aeabi_uldivmod");
2547 set_optab_libfunc (ashl_optab
, DImode
, "__aeabi_llsl");
2548 set_optab_libfunc (lshr_optab
, DImode
, "__aeabi_llsr");
2549 set_optab_libfunc (ashr_optab
, DImode
, "__aeabi_lasr");
2550 set_optab_libfunc (cmp_optab
, DImode
, "__aeabi_lcmp");
2551 set_optab_libfunc (ucmp_optab
, DImode
, "__aeabi_ulcmp");
2553 /* Integer (32/32->32) division. \S 4.3.1. */
2554 set_optab_libfunc (sdivmod_optab
, SImode
, "__aeabi_idivmod");
2555 set_optab_libfunc (udivmod_optab
, SImode
, "__aeabi_uidivmod");
2557 /* The divmod functions are designed so that they can be used for
2558 plain division, even though they return both the quotient and the
2559 remainder. The quotient is returned in the usual location (i.e.,
2560 r0 for SImode, {r0, r1} for DImode), just as would be expected
2561 for an ordinary division routine. Because the AAPCS calling
2562 conventions specify that all of { r0, r1, r2, r3 } are
2563 callee-saved registers, there is no need to tell the compiler
2564 explicitly that those registers are clobbered by these
2566 set_optab_libfunc (sdiv_optab
, DImode
, "__aeabi_ldivmod");
2567 set_optab_libfunc (udiv_optab
, DImode
, "__aeabi_uldivmod");
2569 /* For SImode division the ABI provides div-without-mod routines,
2570 which are faster. */
2571 set_optab_libfunc (sdiv_optab
, SImode
, "__aeabi_idiv");
2572 set_optab_libfunc (udiv_optab
, SImode
, "__aeabi_uidiv");
2574 /* We don't have mod libcalls. Fortunately gcc knows how to use the
2575 divmod libcalls instead. */
2576 set_optab_libfunc (smod_optab
, DImode
, NULL
);
2577 set_optab_libfunc (umod_optab
, DImode
, NULL
);
2578 set_optab_libfunc (smod_optab
, SImode
, NULL
);
2579 set_optab_libfunc (umod_optab
, SImode
, NULL
);
2581 /* Half-precision float operations. The compiler handles all operations
2582 with NULL libfuncs by converting the SFmode. */
2583 switch (arm_fp16_format
)
2585 case ARM_FP16_FORMAT_IEEE
:
2586 case ARM_FP16_FORMAT_ALTERNATIVE
:
2589 set_conv_libfunc (trunc_optab
, HFmode
, SFmode
,
2590 (arm_fp16_format
== ARM_FP16_FORMAT_IEEE
2592 : "__gnu_f2h_alternative"));
2593 set_conv_libfunc (sext_optab
, SFmode
, HFmode
,
2594 (arm_fp16_format
== ARM_FP16_FORMAT_IEEE
2596 : "__gnu_h2f_alternative"));
2598 set_conv_libfunc (trunc_optab
, HFmode
, DFmode
,
2599 (arm_fp16_format
== ARM_FP16_FORMAT_IEEE
2601 : "__gnu_d2h_alternative"));
2604 set_optab_libfunc (add_optab
, HFmode
, NULL
);
2605 set_optab_libfunc (sdiv_optab
, HFmode
, NULL
);
2606 set_optab_libfunc (smul_optab
, HFmode
, NULL
);
2607 set_optab_libfunc (neg_optab
, HFmode
, NULL
);
2608 set_optab_libfunc (sub_optab
, HFmode
, NULL
);
2611 set_optab_libfunc (eq_optab
, HFmode
, NULL
);
2612 set_optab_libfunc (ne_optab
, HFmode
, NULL
);
2613 set_optab_libfunc (lt_optab
, HFmode
, NULL
);
2614 set_optab_libfunc (le_optab
, HFmode
, NULL
);
2615 set_optab_libfunc (ge_optab
, HFmode
, NULL
);
2616 set_optab_libfunc (gt_optab
, HFmode
, NULL
);
2617 set_optab_libfunc (unord_optab
, HFmode
, NULL
);
2624 /* Use names prefixed with __gnu_ for fixed-point helper functions. */
2626 const arm_fixed_mode_set fixed_arith_modes
[] =
2629 { E_UQQmode
, "uqq" },
2631 { E_UHQmode
, "uhq" },
2633 { E_USQmode
, "usq" },
2635 { E_UDQmode
, "udq" },
2637 { E_UTQmode
, "utq" },
2639 { E_UHAmode
, "uha" },
2641 { E_USAmode
, "usa" },
2643 { E_UDAmode
, "uda" },
2645 { E_UTAmode
, "uta" }
2647 const arm_fixed_mode_set fixed_conv_modes
[] =
2650 { E_UQQmode
, "uqq" },
2652 { E_UHQmode
, "uhq" },
2654 { E_USQmode
, "usq" },
2656 { E_UDQmode
, "udq" },
2658 { E_UTQmode
, "utq" },
2660 { E_UHAmode
, "uha" },
2662 { E_USAmode
, "usa" },
2664 { E_UDAmode
, "uda" },
2666 { E_UTAmode
, "uta" },
2677 for (i
= 0; i
< ARRAY_SIZE (fixed_arith_modes
); i
++)
2679 arm_set_fixed_optab_libfunc (add_optab
, fixed_arith_modes
[i
].mode
,
2680 "add", fixed_arith_modes
[i
].name
, 3);
2681 arm_set_fixed_optab_libfunc (ssadd_optab
, fixed_arith_modes
[i
].mode
,
2682 "ssadd", fixed_arith_modes
[i
].name
, 3);
2683 arm_set_fixed_optab_libfunc (usadd_optab
, fixed_arith_modes
[i
].mode
,
2684 "usadd", fixed_arith_modes
[i
].name
, 3);
2685 arm_set_fixed_optab_libfunc (sub_optab
, fixed_arith_modes
[i
].mode
,
2686 "sub", fixed_arith_modes
[i
].name
, 3);
2687 arm_set_fixed_optab_libfunc (sssub_optab
, fixed_arith_modes
[i
].mode
,
2688 "sssub", fixed_arith_modes
[i
].name
, 3);
2689 arm_set_fixed_optab_libfunc (ussub_optab
, fixed_arith_modes
[i
].mode
,
2690 "ussub", fixed_arith_modes
[i
].name
, 3);
2691 arm_set_fixed_optab_libfunc (smul_optab
, fixed_arith_modes
[i
].mode
,
2692 "mul", fixed_arith_modes
[i
].name
, 3);
2693 arm_set_fixed_optab_libfunc (ssmul_optab
, fixed_arith_modes
[i
].mode
,
2694 "ssmul", fixed_arith_modes
[i
].name
, 3);
2695 arm_set_fixed_optab_libfunc (usmul_optab
, fixed_arith_modes
[i
].mode
,
2696 "usmul", fixed_arith_modes
[i
].name
, 3);
2697 arm_set_fixed_optab_libfunc (sdiv_optab
, fixed_arith_modes
[i
].mode
,
2698 "div", fixed_arith_modes
[i
].name
, 3);
2699 arm_set_fixed_optab_libfunc (udiv_optab
, fixed_arith_modes
[i
].mode
,
2700 "udiv", fixed_arith_modes
[i
].name
, 3);
2701 arm_set_fixed_optab_libfunc (ssdiv_optab
, fixed_arith_modes
[i
].mode
,
2702 "ssdiv", fixed_arith_modes
[i
].name
, 3);
2703 arm_set_fixed_optab_libfunc (usdiv_optab
, fixed_arith_modes
[i
].mode
,
2704 "usdiv", fixed_arith_modes
[i
].name
, 3);
2705 arm_set_fixed_optab_libfunc (neg_optab
, fixed_arith_modes
[i
].mode
,
2706 "neg", fixed_arith_modes
[i
].name
, 2);
2707 arm_set_fixed_optab_libfunc (ssneg_optab
, fixed_arith_modes
[i
].mode
,
2708 "ssneg", fixed_arith_modes
[i
].name
, 2);
2709 arm_set_fixed_optab_libfunc (usneg_optab
, fixed_arith_modes
[i
].mode
,
2710 "usneg", fixed_arith_modes
[i
].name
, 2);
2711 arm_set_fixed_optab_libfunc (ashl_optab
, fixed_arith_modes
[i
].mode
,
2712 "ashl", fixed_arith_modes
[i
].name
, 3);
2713 arm_set_fixed_optab_libfunc (ashr_optab
, fixed_arith_modes
[i
].mode
,
2714 "ashr", fixed_arith_modes
[i
].name
, 3);
2715 arm_set_fixed_optab_libfunc (lshr_optab
, fixed_arith_modes
[i
].mode
,
2716 "lshr", fixed_arith_modes
[i
].name
, 3);
2717 arm_set_fixed_optab_libfunc (ssashl_optab
, fixed_arith_modes
[i
].mode
,
2718 "ssashl", fixed_arith_modes
[i
].name
, 3);
2719 arm_set_fixed_optab_libfunc (usashl_optab
, fixed_arith_modes
[i
].mode
,
2720 "usashl", fixed_arith_modes
[i
].name
, 3);
2721 arm_set_fixed_optab_libfunc (cmp_optab
, fixed_arith_modes
[i
].mode
,
2722 "cmp", fixed_arith_modes
[i
].name
, 2);
2725 for (i
= 0; i
< ARRAY_SIZE (fixed_conv_modes
); i
++)
2726 for (j
= 0; j
< ARRAY_SIZE (fixed_conv_modes
); j
++)
2729 || (!ALL_FIXED_POINT_MODE_P (fixed_conv_modes
[i
].mode
)
2730 && !ALL_FIXED_POINT_MODE_P (fixed_conv_modes
[j
].mode
)))
2733 arm_set_fixed_conv_libfunc (fract_optab
, fixed_conv_modes
[i
].mode
,
2734 fixed_conv_modes
[j
].mode
, "fract",
2735 fixed_conv_modes
[i
].name
,
2736 fixed_conv_modes
[j
].name
);
2737 arm_set_fixed_conv_libfunc (satfract_optab
,
2738 fixed_conv_modes
[i
].mode
,
2739 fixed_conv_modes
[j
].mode
, "satfract",
2740 fixed_conv_modes
[i
].name
,
2741 fixed_conv_modes
[j
].name
);
2742 arm_set_fixed_conv_libfunc (fractuns_optab
,
2743 fixed_conv_modes
[i
].mode
,
2744 fixed_conv_modes
[j
].mode
, "fractuns",
2745 fixed_conv_modes
[i
].name
,
2746 fixed_conv_modes
[j
].name
);
2747 arm_set_fixed_conv_libfunc (satfractuns_optab
,
2748 fixed_conv_modes
[i
].mode
,
2749 fixed_conv_modes
[j
].mode
, "satfractuns",
2750 fixed_conv_modes
[i
].name
,
2751 fixed_conv_modes
[j
].name
);
2755 if (TARGET_AAPCS_BASED
)
2756 synchronize_libfunc
= init_one_libfunc ("__sync_synchronize");
2758 speculation_barrier_libfunc
= init_one_libfunc ("__speculation_barrier");
2761 /* On AAPCS systems, this is the "struct __va_list". */
2762 static GTY(()) tree va_list_type
;
2764 /* Return the type to use as __builtin_va_list. */
2766 arm_build_builtin_va_list (void)
2771 if (!TARGET_AAPCS_BASED
)
2772 return std_build_builtin_va_list ();
2774 /* AAPCS \S 7.1.4 requires that va_list be a typedef for a type
2782 The C Library ABI further reinforces this definition in \S
2785 We must follow this definition exactly. The structure tag
2786 name is visible in C++ mangled names, and thus forms a part
2787 of the ABI. The field name may be used by people who
2788 #include <stdarg.h>. */
2789 /* Create the type. */
2790 va_list_type
= lang_hooks
.types
.make_type (RECORD_TYPE
);
2791 /* Give it the required name. */
2792 va_list_name
= build_decl (BUILTINS_LOCATION
,
2794 get_identifier ("__va_list"),
2796 DECL_ARTIFICIAL (va_list_name
) = 1;
2797 TYPE_NAME (va_list_type
) = va_list_name
;
2798 TYPE_STUB_DECL (va_list_type
) = va_list_name
;
2799 /* Create the __ap field. */
2800 ap_field
= build_decl (BUILTINS_LOCATION
,
2802 get_identifier ("__ap"),
2804 DECL_ARTIFICIAL (ap_field
) = 1;
2805 DECL_FIELD_CONTEXT (ap_field
) = va_list_type
;
2806 TYPE_FIELDS (va_list_type
) = ap_field
;
2807 /* Compute its layout. */
2808 layout_type (va_list_type
);
2810 return va_list_type
;
2813 /* Return an expression of type "void *" pointing to the next
2814 available argument in a variable-argument list. VALIST is the
2815 user-level va_list object, of type __builtin_va_list. */
2817 arm_extract_valist_ptr (tree valist
)
2819 if (TREE_TYPE (valist
) == error_mark_node
)
2820 return error_mark_node
;
2822 /* On an AAPCS target, the pointer is stored within "struct
2824 if (TARGET_AAPCS_BASED
)
2826 tree ap_field
= TYPE_FIELDS (TREE_TYPE (valist
));
2827 valist
= build3 (COMPONENT_REF
, TREE_TYPE (ap_field
),
2828 valist
, ap_field
, NULL_TREE
);
2834 /* Implement TARGET_EXPAND_BUILTIN_VA_START. */
2836 arm_expand_builtin_va_start (tree valist
, rtx nextarg
)
2838 valist
= arm_extract_valist_ptr (valist
);
2839 std_expand_builtin_va_start (valist
, nextarg
);
2842 /* Implement TARGET_GIMPLIFY_VA_ARG_EXPR. */
2844 arm_gimplify_va_arg_expr (tree valist
, tree type
, gimple_seq
*pre_p
,
2847 valist
= arm_extract_valist_ptr (valist
);
2848 return std_gimplify_va_arg_expr (valist
, type
, pre_p
, post_p
);
2851 /* Check any incompatible options that the user has specified. */
2853 arm_option_check_internal (struct gcc_options
*opts
)
2855 int flags
= opts
->x_target_flags
;
2857 /* iWMMXt and NEON are incompatible. */
2859 && bitmap_bit_p (arm_active_target
.isa
, isa_bit_neon
))
2860 error ("iWMMXt and NEON are incompatible");
2862 /* Make sure that the processor choice does not conflict with any of the
2863 other command line choices. */
2864 if (TARGET_ARM_P (flags
)
2865 && !bitmap_bit_p (arm_active_target
.isa
, isa_bit_notm
))
2866 error ("target CPU does not support ARM mode");
2868 /* TARGET_BACKTRACE cannot be used here as crtl->is_leaf is not set yet. */
2869 if ((TARGET_TPCS_FRAME
|| TARGET_TPCS_LEAF_FRAME
) && TARGET_ARM_P (flags
))
2870 warning (0, "enabling backtrace support is only meaningful when compiling for the Thumb");
2872 if (TARGET_ARM_P (flags
) && TARGET_CALLEE_INTERWORKING
)
2873 warning (0, "enabling callee interworking support is only meaningful when compiling for the Thumb");
2875 /* If this target is normally configured to use APCS frames, warn if they
2876 are turned off and debugging is turned on. */
2877 if (TARGET_ARM_P (flags
)
2878 && write_symbols
!= NO_DEBUG
2879 && !TARGET_APCS_FRAME
2880 && (TARGET_DEFAULT
& MASK_APCS_FRAME
))
2881 warning (0, "-g with -mno-apcs-frame may not give sensible debugging");
2883 /* iWMMXt unsupported under Thumb mode. */
2884 if (TARGET_THUMB_P (flags
) && TARGET_IWMMXT
)
2885 error ("iWMMXt unsupported under Thumb mode");
2887 if (TARGET_HARD_TP
&& TARGET_THUMB1_P (flags
))
2888 error ("can not use -mtp=cp15 with 16-bit Thumb");
2890 if (TARGET_THUMB_P (flags
) && TARGET_VXWORKS_RTP
&& flag_pic
)
2892 error ("RTP PIC is incompatible with Thumb");
2896 if (target_pure_code
|| target_slow_flash_data
)
2898 const char *flag
= (target_pure_code
? "-mpure-code" :
2899 "-mslow-flash-data");
2901 /* We only support -mpure-code and -mslow-flash-data on M-profile targets
2903 if (!TARGET_HAVE_MOVT
|| arm_arch_notm
|| flag_pic
|| TARGET_NEON
)
2904 error ("%s only supports non-pic code on M-profile targets with the "
2905 "MOVT instruction", flag
);
2907 /* Cannot load addresses: -mslow-flash-data forbids literal pool and
2908 -mword-relocations forbids relocation of MOVT/MOVW. */
2909 if (target_word_relocations
)
2910 error ("%s incompatible with -mword-relocations", flag
);
2914 /* Recompute the global settings depending on target attribute options. */
2917 arm_option_params_internal (void)
2919 /* If we are not using the default (ARM mode) section anchor offset
2920 ranges, then set the correct ranges now. */
2923 /* Thumb-1 LDR instructions cannot have negative offsets.
2924 Permissible positive offset ranges are 5-bit (for byte loads),
2925 6-bit (for halfword loads), or 7-bit (for word loads).
2926 Empirical results suggest a 7-bit anchor range gives the best
2927 overall code size. */
2928 targetm
.min_anchor_offset
= 0;
2929 targetm
.max_anchor_offset
= 127;
2931 else if (TARGET_THUMB2
)
2933 /* The minimum is set such that the total size of the block
2934 for a particular anchor is 248 + 1 + 4095 bytes, which is
2935 divisible by eight, ensuring natural spacing of anchors. */
2936 targetm
.min_anchor_offset
= -248;
2937 targetm
.max_anchor_offset
= 4095;
2941 targetm
.min_anchor_offset
= TARGET_MIN_ANCHOR_OFFSET
;
2942 targetm
.max_anchor_offset
= TARGET_MAX_ANCHOR_OFFSET
;
2945 /* Increase the number of conditional instructions with -Os. */
2946 max_insns_skipped
= optimize_size
? 4 : current_tune
->max_insns_skipped
;
2948 /* For THUMB2, we limit the conditional sequence to one IT block. */
2950 max_insns_skipped
= MIN (max_insns_skipped
, MAX_INSN_PER_IT_BLOCK
);
2953 /* True if -mflip-thumb should next add an attribute for the default
2954 mode, false if it should next add an attribute for the opposite mode. */
2955 static GTY(()) bool thumb_flipper
;
2957 /* Options after initial target override. */
2958 static GTY(()) tree init_optimize
;
2961 arm_override_options_after_change_1 (struct gcc_options
*opts
)
2963 /* -falign-functions without argument: supply one. */
2964 if (opts
->x_flag_align_functions
&& !opts
->x_str_align_functions
)
2965 opts
->x_str_align_functions
= TARGET_THUMB_P (opts
->x_target_flags
)
2966 && opts
->x_optimize_size
? "2" : "4";
2969 /* Implement targetm.override_options_after_change. */
2972 arm_override_options_after_change (void)
2974 arm_configure_build_target (&arm_active_target
,
2975 TREE_TARGET_OPTION (target_option_default_node
),
2976 &global_options_set
, false);
2978 arm_override_options_after_change_1 (&global_options
);
2981 /* Implement TARGET_OPTION_SAVE. */
2983 arm_option_save (struct cl_target_option
*ptr
, struct gcc_options
*opts
)
2985 ptr
->x_arm_arch_string
= opts
->x_arm_arch_string
;
2986 ptr
->x_arm_cpu_string
= opts
->x_arm_cpu_string
;
2987 ptr
->x_arm_tune_string
= opts
->x_arm_tune_string
;
2990 /* Implement TARGET_OPTION_RESTORE. */
2992 arm_option_restore (struct gcc_options
*opts
, struct cl_target_option
*ptr
)
2994 opts
->x_arm_arch_string
= ptr
->x_arm_arch_string
;
2995 opts
->x_arm_cpu_string
= ptr
->x_arm_cpu_string
;
2996 opts
->x_arm_tune_string
= ptr
->x_arm_tune_string
;
2997 arm_configure_build_target (&arm_active_target
, ptr
, &global_options_set
,
3001 /* Reset options between modes that the user has specified. */
3003 arm_option_override_internal (struct gcc_options
*opts
,
3004 struct gcc_options
*opts_set
)
3006 arm_override_options_after_change_1 (opts
);
3008 if (TARGET_INTERWORK
&& !bitmap_bit_p (arm_active_target
.isa
, isa_bit_thumb
))
3010 /* The default is to enable interworking, so this warning message would
3011 be confusing to users who have just compiled with
3012 eg, -march=armv4. */
3013 /* warning (0, "ignoring -minterwork because target CPU does not support THUMB"); */
3014 opts
->x_target_flags
&= ~MASK_INTERWORK
;
3017 if (TARGET_THUMB_P (opts
->x_target_flags
)
3018 && !bitmap_bit_p (arm_active_target
.isa
, isa_bit_thumb
))
3020 warning (0, "target CPU does not support THUMB instructions");
3021 opts
->x_target_flags
&= ~MASK_THUMB
;
3024 if (TARGET_APCS_FRAME
&& TARGET_THUMB_P (opts
->x_target_flags
))
3026 /* warning (0, "ignoring -mapcs-frame because -mthumb was used"); */
3027 opts
->x_target_flags
&= ~MASK_APCS_FRAME
;
3030 /* Callee super interworking implies thumb interworking. Adding
3031 this to the flags here simplifies the logic elsewhere. */
3032 if (TARGET_THUMB_P (opts
->x_target_flags
) && TARGET_CALLEE_INTERWORKING
)
3033 opts
->x_target_flags
|= MASK_INTERWORK
;
3035 /* need to remember initial values so combinaisons of options like
3036 -mflip-thumb -mthumb -fno-schedule-insns work for any attribute. */
3037 cl_optimization
*to
= TREE_OPTIMIZATION (init_optimize
);
3039 if (! opts_set
->x_arm_restrict_it
)
3040 opts
->x_arm_restrict_it
= arm_arch8
;
3042 /* ARM execution state and M profile don't have [restrict] IT. */
3043 if (!TARGET_THUMB2_P (opts
->x_target_flags
) || !arm_arch_notm
)
3044 opts
->x_arm_restrict_it
= 0;
3046 /* Enable -munaligned-access by default for
3047 - all ARMv6 architecture-based processors when compiling for a 32-bit ISA
3048 i.e. Thumb2 and ARM state only.
3049 - ARMv7-A, ARMv7-R, and ARMv7-M architecture-based processors.
3050 - ARMv8 architecture-base processors.
3052 Disable -munaligned-access by default for
3053 - all pre-ARMv6 architecture-based processors
3054 - ARMv6-M architecture-based processors
3055 - ARMv8-M Baseline processors. */
3057 if (! opts_set
->x_unaligned_access
)
3059 opts
->x_unaligned_access
= (TARGET_32BIT_P (opts
->x_target_flags
)
3060 && arm_arch6
&& (arm_arch_notm
|| arm_arch7
));
3062 else if (opts
->x_unaligned_access
== 1
3063 && !(arm_arch6
&& (arm_arch_notm
|| arm_arch7
)))
3065 warning (0, "target CPU does not support unaligned accesses");
3066 opts
->x_unaligned_access
= 0;
3069 /* Don't warn since it's on by default in -O2. */
3070 if (TARGET_THUMB1_P (opts
->x_target_flags
))
3071 opts
->x_flag_schedule_insns
= 0;
3073 opts
->x_flag_schedule_insns
= to
->x_flag_schedule_insns
;
3075 /* Disable shrink-wrap when optimizing function for size, since it tends to
3076 generate additional returns. */
3077 if (optimize_function_for_size_p (cfun
)
3078 && TARGET_THUMB2_P (opts
->x_target_flags
))
3079 opts
->x_flag_shrink_wrap
= false;
3081 opts
->x_flag_shrink_wrap
= to
->x_flag_shrink_wrap
;
3083 /* In Thumb1 mode, we emit the epilogue in RTL, but the last insn
3084 - epilogue_insns - does not accurately model the corresponding insns
3085 emitted in the asm file. In particular, see the comment in thumb_exit
3086 'Find out how many of the (return) argument registers we can corrupt'.
3087 As a consequence, the epilogue may clobber registers without fipa-ra
3088 finding out about it. Therefore, disable fipa-ra in Thumb1 mode.
3089 TODO: Accurately model clobbers for epilogue_insns and reenable
3091 if (TARGET_THUMB1_P (opts
->x_target_flags
))
3092 opts
->x_flag_ipa_ra
= 0;
3094 opts
->x_flag_ipa_ra
= to
->x_flag_ipa_ra
;
3096 /* Thumb2 inline assembly code should always use unified syntax.
3097 This will apply to ARM and Thumb1 eventually. */
3098 opts
->x_inline_asm_unified
= TARGET_THUMB2_P (opts
->x_target_flags
);
3100 #ifdef SUBTARGET_OVERRIDE_INTERNAL_OPTIONS
3101 SUBTARGET_OVERRIDE_INTERNAL_OPTIONS
;
3105 static sbitmap isa_all_fpubits
;
3106 static sbitmap isa_quirkbits
;
3108 /* Configure a build target TARGET from the user-specified options OPTS and
3109 OPTS_SET. If WARN_COMPATIBLE, emit a diagnostic if both the CPU and
3110 architecture have been specified, but the two are not identical. */
3112 arm_configure_build_target (struct arm_build_target
*target
,
3113 struct cl_target_option
*opts
,
3114 struct gcc_options
*opts_set
,
3115 bool warn_compatible
)
3117 const cpu_option
*arm_selected_tune
= NULL
;
3118 const arch_option
*arm_selected_arch
= NULL
;
3119 const cpu_option
*arm_selected_cpu
= NULL
;
3120 const arm_fpu_desc
*arm_selected_fpu
= NULL
;
3121 const char *tune_opts
= NULL
;
3122 const char *arch_opts
= NULL
;
3123 const char *cpu_opts
= NULL
;
3125 bitmap_clear (target
->isa
);
3126 target
->core_name
= NULL
;
3127 target
->arch_name
= NULL
;
3129 if (opts_set
->x_arm_arch_string
)
3131 arm_selected_arch
= arm_parse_arch_option_name (all_architectures
,
3133 opts
->x_arm_arch_string
);
3134 arch_opts
= strchr (opts
->x_arm_arch_string
, '+');
3137 if (opts_set
->x_arm_cpu_string
)
3139 arm_selected_cpu
= arm_parse_cpu_option_name (all_cores
, "-mcpu",
3140 opts
->x_arm_cpu_string
);
3141 cpu_opts
= strchr (opts
->x_arm_cpu_string
, '+');
3142 arm_selected_tune
= arm_selected_cpu
;
3143 /* If taking the tuning from -mcpu, we don't need to rescan the
3144 options for tuning. */
3147 if (opts_set
->x_arm_tune_string
)
3149 arm_selected_tune
= arm_parse_cpu_option_name (all_cores
, "-mtune",
3150 opts
->x_arm_tune_string
);
3151 tune_opts
= strchr (opts
->x_arm_tune_string
, '+');
3154 if (arm_selected_arch
)
3156 arm_initialize_isa (target
->isa
, arm_selected_arch
->common
.isa_bits
);
3157 arm_parse_option_features (target
->isa
, &arm_selected_arch
->common
,
3160 if (arm_selected_cpu
)
3162 auto_sbitmap
cpu_isa (isa_num_bits
);
3163 auto_sbitmap
isa_delta (isa_num_bits
);
3165 arm_initialize_isa (cpu_isa
, arm_selected_cpu
->common
.isa_bits
);
3166 arm_parse_option_features (cpu_isa
, &arm_selected_cpu
->common
,
3168 bitmap_xor (isa_delta
, cpu_isa
, target
->isa
);
3169 /* Ignore any bits that are quirk bits. */
3170 bitmap_and_compl (isa_delta
, isa_delta
, isa_quirkbits
);
3171 /* Ignore (for now) any bits that might be set by -mfpu. */
3172 bitmap_and_compl (isa_delta
, isa_delta
, isa_all_fpubits
);
3174 if (!bitmap_empty_p (isa_delta
))
3176 if (warn_compatible
)
3177 warning (0, "switch -mcpu=%s conflicts with -march=%s switch",
3178 arm_selected_cpu
->common
.name
,
3179 arm_selected_arch
->common
.name
);
3180 /* -march wins for code generation.
3181 -mcpu wins for default tuning. */
3182 if (!arm_selected_tune
)
3183 arm_selected_tune
= arm_selected_cpu
;
3185 arm_selected_cpu
= all_cores
+ arm_selected_arch
->tune_id
;
3186 target
->arch_name
= arm_selected_arch
->common
.name
;
3190 /* Architecture and CPU are essentially the same.
3191 Prefer the CPU setting. */
3192 arm_selected_arch
= all_architectures
+ arm_selected_cpu
->arch
;
3193 target
->core_name
= arm_selected_cpu
->common
.name
;
3194 /* Copy the CPU's capabilities, so that we inherit the
3195 appropriate extensions and quirks. */
3196 bitmap_copy (target
->isa
, cpu_isa
);
3201 /* Pick a CPU based on the architecture. */
3202 arm_selected_cpu
= all_cores
+ arm_selected_arch
->tune_id
;
3203 target
->arch_name
= arm_selected_arch
->common
.name
;
3204 /* Note: target->core_name is left unset in this path. */
3207 else if (arm_selected_cpu
)
3209 target
->core_name
= arm_selected_cpu
->common
.name
;
3210 arm_initialize_isa (target
->isa
, arm_selected_cpu
->common
.isa_bits
);
3211 arm_parse_option_features (target
->isa
, &arm_selected_cpu
->common
,
3213 arm_selected_arch
= all_architectures
+ arm_selected_cpu
->arch
;
3215 /* If the user did not specify a processor or architecture, choose
3219 const cpu_option
*sel
;
3220 auto_sbitmap
sought_isa (isa_num_bits
);
3221 bitmap_clear (sought_isa
);
3222 auto_sbitmap
default_isa (isa_num_bits
);
3224 arm_selected_cpu
= arm_parse_cpu_option_name (all_cores
, "default CPU",
3225 TARGET_CPU_DEFAULT
);
3226 cpu_opts
= strchr (TARGET_CPU_DEFAULT
, '+');
3227 gcc_assert (arm_selected_cpu
->common
.name
);
3229 /* RWE: All of the selection logic below (to the end of this
3230 'if' clause) looks somewhat suspect. It appears to be mostly
3231 there to support forcing thumb support when the default CPU
3232 does not have thumb (somewhat dubious in terms of what the
3233 user might be expecting). I think it should be removed once
3234 support for the pre-thumb era cores is removed. */
3235 sel
= arm_selected_cpu
;
3236 arm_initialize_isa (default_isa
, sel
->common
.isa_bits
);
3237 arm_parse_option_features (default_isa
, &arm_selected_cpu
->common
,
3240 /* Now check to see if the user has specified any command line
3241 switches that require certain abilities from the cpu. */
3243 if (TARGET_INTERWORK
|| TARGET_THUMB
)
3244 bitmap_set_bit (sought_isa
, isa_bit_thumb
);
3246 /* If there are such requirements and the default CPU does not
3247 satisfy them, we need to run over the complete list of
3248 cores looking for one that is satisfactory. */
3249 if (!bitmap_empty_p (sought_isa
)
3250 && !bitmap_subset_p (sought_isa
, default_isa
))
3252 auto_sbitmap
candidate_isa (isa_num_bits
);
3253 /* We're only interested in a CPU with at least the
3254 capabilities of the default CPU and the required
3255 additional features. */
3256 bitmap_ior (default_isa
, default_isa
, sought_isa
);
3258 /* Try to locate a CPU type that supports all of the abilities
3259 of the default CPU, plus the extra abilities requested by
3261 for (sel
= all_cores
; sel
->common
.name
!= NULL
; sel
++)
3263 arm_initialize_isa (candidate_isa
, sel
->common
.isa_bits
);
3264 /* An exact match? */
3265 if (bitmap_equal_p (default_isa
, candidate_isa
))
3269 if (sel
->common
.name
== NULL
)
3271 unsigned current_bit_count
= isa_num_bits
;
3272 const cpu_option
*best_fit
= NULL
;
3274 /* Ideally we would like to issue an error message here
3275 saying that it was not possible to find a CPU compatible
3276 with the default CPU, but which also supports the command
3277 line options specified by the programmer, and so they
3278 ought to use the -mcpu=<name> command line option to
3279 override the default CPU type.
3281 If we cannot find a CPU that has exactly the
3282 characteristics of the default CPU and the given
3283 command line options we scan the array again looking
3284 for a best match. The best match must have at least
3285 the capabilities of the perfect match. */
3286 for (sel
= all_cores
; sel
->common
.name
!= NULL
; sel
++)
3288 arm_initialize_isa (candidate_isa
, sel
->common
.isa_bits
);
3290 if (bitmap_subset_p (default_isa
, candidate_isa
))
3294 bitmap_and_compl (candidate_isa
, candidate_isa
,
3296 count
= bitmap_popcount (candidate_isa
);
3298 if (count
< current_bit_count
)
3301 current_bit_count
= count
;
3305 gcc_assert (best_fit
);
3309 arm_selected_cpu
= sel
;
3312 /* Now we know the CPU, we can finally initialize the target
3314 target
->core_name
= arm_selected_cpu
->common
.name
;
3315 arm_initialize_isa (target
->isa
, arm_selected_cpu
->common
.isa_bits
);
3316 arm_parse_option_features (target
->isa
, &arm_selected_cpu
->common
,
3318 arm_selected_arch
= all_architectures
+ arm_selected_cpu
->arch
;
3321 gcc_assert (arm_selected_cpu
);
3322 gcc_assert (arm_selected_arch
);
3324 if (opts
->x_arm_fpu_index
!= TARGET_FPU_auto
)
3326 arm_selected_fpu
= &all_fpus
[opts
->x_arm_fpu_index
];
3327 auto_sbitmap
fpu_bits (isa_num_bits
);
3329 arm_initialize_isa (fpu_bits
, arm_selected_fpu
->isa_bits
);
3330 bitmap_and_compl (target
->isa
, target
->isa
, isa_all_fpubits
);
3331 bitmap_ior (target
->isa
, target
->isa
, fpu_bits
);
3334 if (!arm_selected_tune
)
3335 arm_selected_tune
= arm_selected_cpu
;
3336 else /* Validate the features passed to -mtune. */
3337 arm_parse_option_features (NULL
, &arm_selected_tune
->common
, tune_opts
);
3339 const cpu_tune
*tune_data
= &all_tunes
[arm_selected_tune
- all_cores
];
3341 /* Finish initializing the target structure. */
3342 target
->arch_pp_name
= arm_selected_arch
->arch
;
3343 target
->base_arch
= arm_selected_arch
->base_arch
;
3344 target
->profile
= arm_selected_arch
->profile
;
3346 target
->tune_flags
= tune_data
->tune_flags
;
3347 target
->tune
= tune_data
->tune
;
3348 target
->tune_core
= tune_data
->scheduler
;
3349 arm_option_reconfigure_globals ();
3352 /* Fix up any incompatible options that the user has specified. */
3354 arm_option_override (void)
3356 static const enum isa_feature fpu_bitlist
[]
3357 = { ISA_ALL_FPU_INTERNAL
, isa_nobit
};
3358 static const enum isa_feature quirk_bitlist
[] = { ISA_ALL_QUIRKS
, isa_nobit
};
3359 cl_target_option opts
;
3361 isa_quirkbits
= sbitmap_alloc (isa_num_bits
);
3362 arm_initialize_isa (isa_quirkbits
, quirk_bitlist
);
3364 isa_all_fpubits
= sbitmap_alloc (isa_num_bits
);
3365 arm_initialize_isa (isa_all_fpubits
, fpu_bitlist
);
3367 arm_active_target
.isa
= sbitmap_alloc (isa_num_bits
);
3369 if (!global_options_set
.x_arm_fpu_index
)
3374 ok
= opt_enum_arg_to_value (OPT_mfpu_
, FPUTYPE_AUTO
, &fpu_index
,
3377 arm_fpu_index
= (enum fpu_type
) fpu_index
;
3380 cl_target_option_save (&opts
, &global_options
);
3381 arm_configure_build_target (&arm_active_target
, &opts
, &global_options_set
,
3384 #ifdef SUBTARGET_OVERRIDE_OPTIONS
3385 SUBTARGET_OVERRIDE_OPTIONS
;
3388 /* Initialize boolean versions of the architectural flags, for use
3389 in the arm.md file and for enabling feature flags. */
3390 arm_option_reconfigure_globals ();
3392 arm_tune
= arm_active_target
.tune_core
;
3393 tune_flags
= arm_active_target
.tune_flags
;
3394 current_tune
= arm_active_target
.tune
;
3396 /* TBD: Dwarf info for apcs frame is not handled yet. */
3397 if (TARGET_APCS_FRAME
)
3398 flag_shrink_wrap
= false;
3400 if (TARGET_APCS_STACK
&& !TARGET_APCS_FRAME
)
3402 warning (0, "-mapcs-stack-check incompatible with -mno-apcs-frame");
3403 target_flags
|= MASK_APCS_FRAME
;
3406 if (TARGET_POKE_FUNCTION_NAME
)
3407 target_flags
|= MASK_APCS_FRAME
;
3409 if (TARGET_APCS_REENT
&& flag_pic
)
3410 error ("-fpic and -mapcs-reent are incompatible");
3412 if (TARGET_APCS_REENT
)
3413 warning (0, "APCS reentrant code not supported. Ignored");
3415 /* Set up some tuning parameters. */
3416 arm_ld_sched
= (tune_flags
& TF_LDSCHED
) != 0;
3417 arm_tune_strongarm
= (tune_flags
& TF_STRONG
) != 0;
3418 arm_tune_wbuf
= (tune_flags
& TF_WBUF
) != 0;
3419 arm_tune_xscale
= (tune_flags
& TF_XSCALE
) != 0;
3420 arm_tune_cortex_a9
= (arm_tune
== TARGET_CPU_cortexa9
) != 0;
3421 arm_m_profile_small_mul
= (tune_flags
& TF_SMALLMUL
) != 0;
3423 /* For arm2/3 there is no need to do any scheduling if we are doing
3424 software floating-point. */
3425 if (TARGET_SOFT_FLOAT
&& (tune_flags
& TF_NO_MODE32
))
3426 flag_schedule_insns
= flag_schedule_insns_after_reload
= 0;
3428 /* Override the default structure alignment for AAPCS ABI. */
3429 if (!global_options_set
.x_arm_structure_size_boundary
)
3431 if (TARGET_AAPCS_BASED
)
3432 arm_structure_size_boundary
= 8;
3436 warning (0, "option %<-mstructure-size-boundary%> is deprecated");
3438 if (arm_structure_size_boundary
!= 8
3439 && arm_structure_size_boundary
!= 32
3440 && !(ARM_DOUBLEWORD_ALIGN
&& arm_structure_size_boundary
== 64))
3442 if (ARM_DOUBLEWORD_ALIGN
)
3444 "structure size boundary can only be set to 8, 32 or 64");
3446 warning (0, "structure size boundary can only be set to 8 or 32");
3447 arm_structure_size_boundary
3448 = (TARGET_AAPCS_BASED
? 8 : DEFAULT_STRUCTURE_SIZE_BOUNDARY
);
3452 if (TARGET_VXWORKS_RTP
)
3454 if (!global_options_set
.x_arm_pic_data_is_text_relative
)
3455 arm_pic_data_is_text_relative
= 0;
3458 && !arm_pic_data_is_text_relative
3459 && !(global_options_set
.x_target_flags
& MASK_SINGLE_PIC_BASE
))
3460 /* When text & data segments don't have a fixed displacement, the
3461 intended use is with a single, read only, pic base register.
3462 Unless the user explicitly requested not to do that, set
3464 target_flags
|= MASK_SINGLE_PIC_BASE
;
3466 /* If stack checking is disabled, we can use r10 as the PIC register,
3467 which keeps r9 available. The EABI specifies r9 as the PIC register. */
3468 if (flag_pic
&& TARGET_SINGLE_PIC_BASE
)
3470 if (TARGET_VXWORKS_RTP
)
3471 warning (0, "RTP PIC is incompatible with -msingle-pic-base");
3472 arm_pic_register
= (TARGET_APCS_STACK
|| TARGET_AAPCS_BASED
) ? 9 : 10;
3475 if (flag_pic
&& TARGET_VXWORKS_RTP
)
3476 arm_pic_register
= 9;
3478 if (arm_pic_register_string
!= NULL
)
3480 int pic_register
= decode_reg_name (arm_pic_register_string
);
3483 warning (0, "-mpic-register= is useless without -fpic");
3485 /* Prevent the user from choosing an obviously stupid PIC register. */
3486 else if (pic_register
< 0 || call_used_regs
[pic_register
]
3487 || pic_register
== HARD_FRAME_POINTER_REGNUM
3488 || pic_register
== STACK_POINTER_REGNUM
3489 || pic_register
>= PC_REGNUM
3490 || (TARGET_VXWORKS_RTP
3491 && (unsigned int) pic_register
!= arm_pic_register
))
3492 error ("unable to use '%s' for PIC register", arm_pic_register_string
);
3494 arm_pic_register
= pic_register
;
3498 target_word_relocations
= 1;
3500 /* Enable -mfix-cortex-m3-ldrd by default for Cortex-M3 cores. */
3501 if (fix_cm3_ldrd
== 2)
3503 if (bitmap_bit_p (arm_active_target
.isa
, isa_bit_quirk_cm3_ldrd
))
3509 /* Hot/Cold partitioning is not currently supported, since we can't
3510 handle literal pool placement in that case. */
3511 if (flag_reorder_blocks_and_partition
)
3513 inform (input_location
,
3514 "-freorder-blocks-and-partition not supported on this architecture");
3515 flag_reorder_blocks_and_partition
= 0;
3516 flag_reorder_blocks
= 1;
3520 /* Hoisting PIC address calculations more aggressively provides a small,
3521 but measurable, size reduction for PIC code. Therefore, we decrease
3522 the bar for unrestricted expression hoisting to the cost of PIC address
3523 calculation, which is 2 instructions. */
3524 maybe_set_param_value (PARAM_GCSE_UNRESTRICTED_COST
, 2,
3525 global_options
.x_param_values
,
3526 global_options_set
.x_param_values
);
3528 /* ARM EABI defaults to strict volatile bitfields. */
3529 if (TARGET_AAPCS_BASED
&& flag_strict_volatile_bitfields
< 0
3530 && abi_version_at_least(2))
3531 flag_strict_volatile_bitfields
= 1;
3533 /* Enable sw prefetching at -O3 for CPUS that have prefetch, and we
3534 have deemed it beneficial (signified by setting
3535 prefetch.num_slots to 1 or more). */
3536 if (flag_prefetch_loop_arrays
< 0
3539 && current_tune
->prefetch
.num_slots
> 0)
3540 flag_prefetch_loop_arrays
= 1;
3542 /* Set up parameters to be used in prefetching algorithm. Do not
3543 override the defaults unless we are tuning for a core we have
3544 researched values for. */
3545 if (current_tune
->prefetch
.num_slots
> 0)
3546 maybe_set_param_value (PARAM_SIMULTANEOUS_PREFETCHES
,
3547 current_tune
->prefetch
.num_slots
,
3548 global_options
.x_param_values
,
3549 global_options_set
.x_param_values
);
3550 if (current_tune
->prefetch
.l1_cache_line_size
>= 0)
3551 maybe_set_param_value (PARAM_L1_CACHE_LINE_SIZE
,
3552 current_tune
->prefetch
.l1_cache_line_size
,
3553 global_options
.x_param_values
,
3554 global_options_set
.x_param_values
);
3555 if (current_tune
->prefetch
.l1_cache_size
>= 0)
3556 maybe_set_param_value (PARAM_L1_CACHE_SIZE
,
3557 current_tune
->prefetch
.l1_cache_size
,
3558 global_options
.x_param_values
,
3559 global_options_set
.x_param_values
);
3561 /* Use Neon to perform 64-bits operations rather than core
3563 prefer_neon_for_64bits
= current_tune
->prefer_neon_for_64bits
;
3564 if (use_neon_for_64bits
== 1)
3565 prefer_neon_for_64bits
= true;
3567 /* Use the alternative scheduling-pressure algorithm by default. */
3568 maybe_set_param_value (PARAM_SCHED_PRESSURE_ALGORITHM
, SCHED_PRESSURE_MODEL
,
3569 global_options
.x_param_values
,
3570 global_options_set
.x_param_values
);
3572 /* Look through ready list and all of queue for instructions
3573 relevant for L2 auto-prefetcher. */
3574 int param_sched_autopref_queue_depth
;
3576 switch (current_tune
->sched_autopref
)
3578 case tune_params::SCHED_AUTOPREF_OFF
:
3579 param_sched_autopref_queue_depth
= -1;
3582 case tune_params::SCHED_AUTOPREF_RANK
:
3583 param_sched_autopref_queue_depth
= 0;
3586 case tune_params::SCHED_AUTOPREF_FULL
:
3587 param_sched_autopref_queue_depth
= max_insn_queue_index
+ 1;
3594 maybe_set_param_value (PARAM_SCHED_AUTOPREF_QUEUE_DEPTH
,
3595 param_sched_autopref_queue_depth
,
3596 global_options
.x_param_values
,
3597 global_options_set
.x_param_values
);
3599 /* Currently, for slow flash data, we just disable literal pools. We also
3600 disable it for pure-code. */
3601 if (target_slow_flash_data
|| target_pure_code
)
3602 arm_disable_literal_pool
= true;
3604 /* Disable scheduling fusion by default if it's not armv7 processor
3605 or doesn't prefer ldrd/strd. */
3606 if (flag_schedule_fusion
== 2
3607 && (!arm_arch7
|| !current_tune
->prefer_ldrd_strd
))
3608 flag_schedule_fusion
= 0;
3610 /* Need to remember initial options before they are overriden. */
3611 init_optimize
= build_optimization_node (&global_options
);
3613 arm_options_perform_arch_sanity_checks ();
3614 arm_option_override_internal (&global_options
, &global_options_set
);
3615 arm_option_check_internal (&global_options
);
3616 arm_option_params_internal ();
3618 /* Create the default target_options structure. */
3619 target_option_default_node
= target_option_current_node
3620 = build_target_option_node (&global_options
);
3622 /* Register global variables with the garbage collector. */
3623 arm_add_gc_roots ();
3625 /* Init initial mode for testing. */
3626 thumb_flipper
= TARGET_THUMB
;
3630 /* Reconfigure global status flags from the active_target.isa. */
3632 arm_option_reconfigure_globals (void)
3634 sprintf (arm_arch_name
, "__ARM_ARCH_%s__", arm_active_target
.arch_pp_name
);
3635 arm_base_arch
= arm_active_target
.base_arch
;
3637 /* Initialize boolean versions of the architectural flags, for use
3638 in the arm.md file. */
3639 arm_arch4
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_armv4
);
3640 arm_arch4t
= arm_arch4
&& bitmap_bit_p (arm_active_target
.isa
, isa_bit_thumb
);
3641 arm_arch5t
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_armv5t
);
3642 arm_arch5te
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_armv5te
);
3643 arm_arch6
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_armv6
);
3644 arm_arch6k
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_armv6k
);
3645 arm_arch_notm
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_notm
);
3646 arm_arch6m
= arm_arch6
&& !arm_arch_notm
;
3647 arm_arch7
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_armv7
);
3648 arm_arch7em
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_armv7em
);
3649 arm_arch8
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_armv8
);
3650 arm_arch8_1
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_armv8_1
);
3651 arm_arch8_2
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_armv8_2
);
3652 arm_arch_thumb1
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_thumb
);
3653 arm_arch_thumb2
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_thumb2
);
3654 arm_arch_xscale
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_xscale
);
3655 arm_arch_iwmmxt
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_iwmmxt
);
3656 arm_arch_iwmmxt2
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_iwmmxt2
);
3657 arm_arch_thumb_hwdiv
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_tdiv
);
3658 arm_arch_arm_hwdiv
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_adiv
);
3659 arm_arch_crc
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_crc32
);
3660 arm_arch_cmse
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_cmse
);
3661 arm_fp16_inst
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_fp16
);
3662 arm_arch_lpae
= bitmap_bit_p (arm_active_target
.isa
, isa_bit_lpae
);
3665 if (arm_fp16_format
== ARM_FP16_FORMAT_ALTERNATIVE
)
3666 error ("selected fp16 options are incompatible");
3667 arm_fp16_format
= ARM_FP16_FORMAT_IEEE
;
3670 /* And finally, set up some quirks. */
3671 arm_arch_no_volatile_ce
3672 = bitmap_bit_p (arm_active_target
.isa
, isa_bit_quirk_no_volatile_ce
);
3673 arm_arch6kz
= arm_arch6k
&& bitmap_bit_p (arm_active_target
.isa
,
3674 isa_bit_quirk_armv6kz
);
3676 /* Use the cp15 method if it is available. */
3677 if (target_thread_pointer
== TP_AUTO
)
3679 if (arm_arch6k
&& !TARGET_THUMB1
)
3680 target_thread_pointer
= TP_CP15
;
3682 target_thread_pointer
= TP_SOFT
;
3686 /* Perform some validation between the desired architecture and the rest of the
3689 arm_options_perform_arch_sanity_checks (void)
3691 /* V5T code we generate is completely interworking capable, so we turn off
3692 TARGET_INTERWORK here to avoid many tests later on. */
3694 /* XXX However, we must pass the right pre-processor defines to CPP
3695 or GLD can get confused. This is a hack. */
3696 if (TARGET_INTERWORK
)
3697 arm_cpp_interwork
= 1;
3700 target_flags
&= ~MASK_INTERWORK
;
3702 if (TARGET_IWMMXT
&& !ARM_DOUBLEWORD_ALIGN
)
3703 error ("iwmmxt requires an AAPCS compatible ABI for proper operation");
3705 if (TARGET_IWMMXT_ABI
&& !TARGET_IWMMXT
)
3706 error ("iwmmxt abi requires an iwmmxt capable cpu");
3708 /* BPABI targets use linker tricks to allow interworking on cores
3709 without thumb support. */
3710 if (TARGET_INTERWORK
3712 && !bitmap_bit_p (arm_active_target
.isa
, isa_bit_thumb
))
3714 warning (0, "target CPU does not support interworking" );
3715 target_flags
&= ~MASK_INTERWORK
;
3718 /* If soft-float is specified then don't use FPU. */
3719 if (TARGET_SOFT_FLOAT
)
3720 arm_fpu_attr
= FPU_NONE
;
3722 arm_fpu_attr
= FPU_VFP
;
3724 if (TARGET_AAPCS_BASED
)
3726 if (TARGET_CALLER_INTERWORKING
)
3727 error ("AAPCS does not support -mcaller-super-interworking");
3729 if (TARGET_CALLEE_INTERWORKING
)
3730 error ("AAPCS does not support -mcallee-super-interworking");
3733 /* __fp16 support currently assumes the core has ldrh. */
3734 if (!arm_arch4
&& arm_fp16_format
!= ARM_FP16_FORMAT_NONE
)
3735 sorry ("__fp16 and no ldrh");
3737 if (use_cmse
&& !arm_arch_cmse
)
3738 error ("target CPU does not support ARMv8-M Security Extensions");
3740 /* We don't clear D16-D31 VFP registers for cmse_nonsecure_call functions
3741 and ARMv8-M Baseline and Mainline do not allow such configuration. */
3742 if (use_cmse
&& LAST_VFP_REGNUM
> LAST_LO_VFP_REGNUM
)
3743 error ("ARMv8-M Security Extensions incompatible with selected FPU");
3746 if (TARGET_AAPCS_BASED
)
3748 if (arm_abi
== ARM_ABI_IWMMXT
)
3749 arm_pcs_default
= ARM_PCS_AAPCS_IWMMXT
;
3750 else if (TARGET_HARD_FLOAT_ABI
)
3752 arm_pcs_default
= ARM_PCS_AAPCS_VFP
;
3753 if (!bitmap_bit_p (arm_active_target
.isa
, isa_bit_vfpv2
))
3754 error ("-mfloat-abi=hard: selected processor lacks an FPU");
3757 arm_pcs_default
= ARM_PCS_AAPCS
;
3761 if (arm_float_abi
== ARM_FLOAT_ABI_HARD
)
3762 sorry ("-mfloat-abi=hard and VFP");
3764 if (arm_abi
== ARM_ABI_APCS
)
3765 arm_pcs_default
= ARM_PCS_APCS
;
3767 arm_pcs_default
= ARM_PCS_ATPCS
;
3772 arm_add_gc_roots (void)
3774 gcc_obstack_init(&minipool_obstack
);
3775 minipool_startobj
= (char *) obstack_alloc (&minipool_obstack
, 0);
3778 /* A table of known ARM exception types.
3779 For use with the interrupt function attribute. */
3783 const char *const arg
;
3784 const unsigned long return_value
;
3788 static const isr_attribute_arg isr_attribute_args
[] =
3790 { "IRQ", ARM_FT_ISR
},
3791 { "irq", ARM_FT_ISR
},
3792 { "FIQ", ARM_FT_FIQ
},
3793 { "fiq", ARM_FT_FIQ
},
3794 { "ABORT", ARM_FT_ISR
},
3795 { "abort", ARM_FT_ISR
},
3796 { "ABORT", ARM_FT_ISR
},
3797 { "abort", ARM_FT_ISR
},
3798 { "UNDEF", ARM_FT_EXCEPTION
},
3799 { "undef", ARM_FT_EXCEPTION
},
3800 { "SWI", ARM_FT_EXCEPTION
},
3801 { "swi", ARM_FT_EXCEPTION
},
3802 { NULL
, ARM_FT_NORMAL
}
3805 /* Returns the (interrupt) function type of the current
3806 function, or ARM_FT_UNKNOWN if the type cannot be determined. */
3808 static unsigned long
3809 arm_isr_value (tree argument
)
3811 const isr_attribute_arg
* ptr
;
3815 return ARM_FT_NORMAL
| ARM_FT_STACKALIGN
;
3817 /* No argument - default to IRQ. */
3818 if (argument
== NULL_TREE
)
3821 /* Get the value of the argument. */
3822 if (TREE_VALUE (argument
) == NULL_TREE
3823 || TREE_CODE (TREE_VALUE (argument
)) != STRING_CST
)
3824 return ARM_FT_UNKNOWN
;
3826 arg
= TREE_STRING_POINTER (TREE_VALUE (argument
));
3828 /* Check it against the list of known arguments. */
3829 for (ptr
= isr_attribute_args
; ptr
->arg
!= NULL
; ptr
++)
3830 if (streq (arg
, ptr
->arg
))
3831 return ptr
->return_value
;
3833 /* An unrecognized interrupt type. */
3834 return ARM_FT_UNKNOWN
;
3837 /* Computes the type of the current function. */
3839 static unsigned long
3840 arm_compute_func_type (void)
3842 unsigned long type
= ARM_FT_UNKNOWN
;
3846 gcc_assert (TREE_CODE (current_function_decl
) == FUNCTION_DECL
);
3848 /* Decide if the current function is volatile. Such functions
3849 never return, and many memory cycles can be saved by not storing
3850 register values that will never be needed again. This optimization
3851 was added to speed up context switching in a kernel application. */
3853 && (TREE_NOTHROW (current_function_decl
)
3854 || !(flag_unwind_tables
3856 && arm_except_unwind_info (&global_options
) != UI_SJLJ
)))
3857 && TREE_THIS_VOLATILE (current_function_decl
))
3858 type
|= ARM_FT_VOLATILE
;
3860 if (cfun
->static_chain_decl
!= NULL
)
3861 type
|= ARM_FT_NESTED
;
3863 attr
= DECL_ATTRIBUTES (current_function_decl
);
3865 a
= lookup_attribute ("naked", attr
);
3867 type
|= ARM_FT_NAKED
;
3869 a
= lookup_attribute ("isr", attr
);
3871 a
= lookup_attribute ("interrupt", attr
);
3874 type
|= TARGET_INTERWORK
? ARM_FT_INTERWORKED
: ARM_FT_NORMAL
;
3876 type
|= arm_isr_value (TREE_VALUE (a
));
3878 if (lookup_attribute ("cmse_nonsecure_entry", attr
))
3879 type
|= ARM_FT_CMSE_ENTRY
;
3884 /* Returns the type of the current function. */
3887 arm_current_func_type (void)
3889 if (ARM_FUNC_TYPE (cfun
->machine
->func_type
) == ARM_FT_UNKNOWN
)
3890 cfun
->machine
->func_type
= arm_compute_func_type ();
3892 return cfun
->machine
->func_type
;
3896 arm_allocate_stack_slots_for_args (void)
3898 /* Naked functions should not allocate stack slots for arguments. */
3899 return !IS_NAKED (arm_current_func_type ());
3903 arm_warn_func_return (tree decl
)
3905 /* Naked functions are implemented entirely in assembly, including the
3906 return sequence, so suppress warnings about this. */
3907 return lookup_attribute ("naked", DECL_ATTRIBUTES (decl
)) == NULL_TREE
;
3911 /* Output assembler code for a block containing the constant parts
3912 of a trampoline, leaving space for the variable parts.
3914 On the ARM, (if r8 is the static chain regnum, and remembering that
3915 referencing pc adds an offset of 8) the trampoline looks like:
3918 .word static chain value
3919 .word function's address
3920 XXX FIXME: When the trampoline returns, r8 will be clobbered. */
3923 arm_asm_trampoline_template (FILE *f
)
3925 fprintf (f
, "\t.syntax unified\n");
3929 fprintf (f
, "\t.arm\n");
3930 asm_fprintf (f
, "\tldr\t%r, [%r, #0]\n", STATIC_CHAIN_REGNUM
, PC_REGNUM
);
3931 asm_fprintf (f
, "\tldr\t%r, [%r, #0]\n", PC_REGNUM
, PC_REGNUM
);
3933 else if (TARGET_THUMB2
)
3935 fprintf (f
, "\t.thumb\n");
3936 /* The Thumb-2 trampoline is similar to the arm implementation.
3937 Unlike 16-bit Thumb, we enter the stub in thumb mode. */
3938 asm_fprintf (f
, "\tldr.w\t%r, [%r, #4]\n",
3939 STATIC_CHAIN_REGNUM
, PC_REGNUM
);
3940 asm_fprintf (f
, "\tldr.w\t%r, [%r, #4]\n", PC_REGNUM
, PC_REGNUM
);
3944 ASM_OUTPUT_ALIGN (f
, 2);
3945 fprintf (f
, "\t.code\t16\n");
3946 fprintf (f
, ".Ltrampoline_start:\n");
3947 asm_fprintf (f
, "\tpush\t{r0, r1}\n");
3948 asm_fprintf (f
, "\tldr\tr0, [%r, #8]\n", PC_REGNUM
);
3949 asm_fprintf (f
, "\tmov\t%r, r0\n", STATIC_CHAIN_REGNUM
);
3950 asm_fprintf (f
, "\tldr\tr0, [%r, #8]\n", PC_REGNUM
);
3951 asm_fprintf (f
, "\tstr\tr0, [%r, #4]\n", SP_REGNUM
);
3952 asm_fprintf (f
, "\tpop\t{r0, %r}\n", PC_REGNUM
);
3954 assemble_aligned_integer (UNITS_PER_WORD
, const0_rtx
);
3955 assemble_aligned_integer (UNITS_PER_WORD
, const0_rtx
);
3958 /* Emit RTL insns to initialize the variable parts of a trampoline. */
3961 arm_trampoline_init (rtx m_tramp
, tree fndecl
, rtx chain_value
)
3963 rtx fnaddr
, mem
, a_tramp
;
3965 emit_block_move (m_tramp
, assemble_trampoline_template (),
3966 GEN_INT (TRAMPOLINE_SIZE
), BLOCK_OP_NORMAL
);
3968 mem
= adjust_address (m_tramp
, SImode
, TARGET_32BIT
? 8 : 12);
3969 emit_move_insn (mem
, chain_value
);
3971 mem
= adjust_address (m_tramp
, SImode
, TARGET_32BIT
? 12 : 16);
3972 fnaddr
= XEXP (DECL_RTL (fndecl
), 0);
3973 emit_move_insn (mem
, fnaddr
);
3975 a_tramp
= XEXP (m_tramp
, 0);
3976 emit_library_call (gen_rtx_SYMBOL_REF (Pmode
, "__clear_cache"),
3977 LCT_NORMAL
, VOIDmode
, a_tramp
, Pmode
,
3978 plus_constant (Pmode
, a_tramp
, TRAMPOLINE_SIZE
), Pmode
);
3981 /* Thumb trampolines should be entered in thumb mode, so set
3982 the bottom bit of the address. */
3985 arm_trampoline_adjust_address (rtx addr
)
3988 addr
= expand_simple_binop (Pmode
, IOR
, addr
, const1_rtx
,
3989 NULL
, 0, OPTAB_LIB_WIDEN
);
3993 /* Return 1 if it is possible to return using a single instruction.
3994 If SIBLING is non-null, this is a test for a return before a sibling
3995 call. SIBLING is the call insn, so we can examine its register usage. */
3998 use_return_insn (int iscond
, rtx sibling
)
4001 unsigned int func_type
;
4002 unsigned long saved_int_regs
;
4003 unsigned HOST_WIDE_INT stack_adjust
;
4004 arm_stack_offsets
*offsets
;
4006 /* Never use a return instruction before reload has run. */
4007 if (!reload_completed
)
4010 func_type
= arm_current_func_type ();
4012 /* Naked, volatile and stack alignment functions need special
4014 if (func_type
& (ARM_FT_VOLATILE
| ARM_FT_NAKED
| ARM_FT_STACKALIGN
))
4017 /* So do interrupt functions that use the frame pointer and Thumb
4018 interrupt functions. */
4019 if (IS_INTERRUPT (func_type
) && (frame_pointer_needed
|| TARGET_THUMB
))
4022 if (TARGET_LDRD
&& current_tune
->prefer_ldrd_strd
4023 && !optimize_function_for_size_p (cfun
))
4026 offsets
= arm_get_frame_offsets ();
4027 stack_adjust
= offsets
->outgoing_args
- offsets
->saved_regs
;
4029 /* As do variadic functions. */
4030 if (crtl
->args
.pretend_args_size
4031 || cfun
->machine
->uses_anonymous_args
4032 /* Or if the function calls __builtin_eh_return () */
4033 || crtl
->calls_eh_return
4034 /* Or if the function calls alloca */
4035 || cfun
->calls_alloca
4036 /* Or if there is a stack adjustment. However, if the stack pointer
4037 is saved on the stack, we can use a pre-incrementing stack load. */
4038 || !(stack_adjust
== 0 || (TARGET_APCS_FRAME
&& frame_pointer_needed
4039 && stack_adjust
== 4))
4040 /* Or if the static chain register was saved above the frame, under the
4041 assumption that the stack pointer isn't saved on the stack. */
4042 || (!(TARGET_APCS_FRAME
&& frame_pointer_needed
)
4043 && arm_compute_static_chain_stack_bytes() != 0))
4046 saved_int_regs
= offsets
->saved_regs_mask
;
4048 /* Unfortunately, the insn
4050 ldmib sp, {..., sp, ...}
4052 triggers a bug on most SA-110 based devices, such that the stack
4053 pointer won't be correctly restored if the instruction takes a
4054 page fault. We work around this problem by popping r3 along with
4055 the other registers, since that is never slower than executing
4056 another instruction.
4058 We test for !arm_arch5t here, because code for any architecture
4059 less than this could potentially be run on one of the buggy
4061 if (stack_adjust
== 4 && !arm_arch5t
&& TARGET_ARM
)
4063 /* Validate that r3 is a call-clobbered register (always true in
4064 the default abi) ... */
4065 if (!call_used_regs
[3])
4068 /* ... that it isn't being used for a return value ... */
4069 if (arm_size_return_regs () >= (4 * UNITS_PER_WORD
))
4072 /* ... or for a tail-call argument ... */
4075 gcc_assert (CALL_P (sibling
));
4077 if (find_regno_fusage (sibling
, USE
, 3))
4081 /* ... and that there are no call-saved registers in r0-r2
4082 (always true in the default ABI). */
4083 if (saved_int_regs
& 0x7)
4087 /* Can't be done if interworking with Thumb, and any registers have been
4089 if (TARGET_INTERWORK
&& saved_int_regs
!= 0 && !IS_INTERRUPT(func_type
))
4092 /* On StrongARM, conditional returns are expensive if they aren't
4093 taken and multiple registers have been stacked. */
4094 if (iscond
&& arm_tune_strongarm
)
4096 /* Conditional return when just the LR is stored is a simple
4097 conditional-load instruction, that's not expensive. */
4098 if (saved_int_regs
!= 0 && saved_int_regs
!= (1 << LR_REGNUM
))
4102 && arm_pic_register
!= INVALID_REGNUM
4103 && df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM
))
4107 /* ARMv8-M nonsecure entry function need to use bxns to return and thus need
4108 several instructions if anything needs to be popped. */
4109 if (saved_int_regs
&& IS_CMSE_ENTRY (func_type
))
4112 /* If there are saved registers but the LR isn't saved, then we need
4113 two instructions for the return. */
4114 if (saved_int_regs
&& !(saved_int_regs
& (1 << LR_REGNUM
)))
4117 /* Can't be done if any of the VFP regs are pushed,
4118 since this also requires an insn. */
4119 if (TARGET_HARD_FLOAT
)
4120 for (regno
= FIRST_VFP_REGNUM
; regno
<= LAST_VFP_REGNUM
; regno
++)
4121 if (df_regs_ever_live_p (regno
) && !call_used_regs
[regno
])
4124 if (TARGET_REALLY_IWMMXT
)
4125 for (regno
= FIRST_IWMMXT_REGNUM
; regno
<= LAST_IWMMXT_REGNUM
; regno
++)
4126 if (df_regs_ever_live_p (regno
) && ! call_used_regs
[regno
])
4132 /* Return TRUE if we should try to use a simple_return insn, i.e. perform
4133 shrink-wrapping if possible. This is the case if we need to emit a
4134 prologue, which we can test by looking at the offsets. */
4136 use_simple_return_p (void)
4138 arm_stack_offsets
*offsets
;
4140 /* Note this function can be called before or after reload. */
4141 if (!reload_completed
)
4142 arm_compute_frame_layout ();
4144 offsets
= arm_get_frame_offsets ();
4145 return offsets
->outgoing_args
!= 0;
4148 /* Return TRUE if int I is a valid immediate ARM constant. */
4151 const_ok_for_arm (HOST_WIDE_INT i
)
4155 /* For machines with >32 bit HOST_WIDE_INT, the bits above bit 31 must
4156 be all zero, or all one. */
4157 if ((i
& ~(unsigned HOST_WIDE_INT
) 0xffffffff) != 0
4158 && ((i
& ~(unsigned HOST_WIDE_INT
) 0xffffffff)
4159 != ((~(unsigned HOST_WIDE_INT
) 0)
4160 & ~(unsigned HOST_WIDE_INT
) 0xffffffff)))
4163 i
&= (unsigned HOST_WIDE_INT
) 0xffffffff;
4165 /* Fast return for 0 and small values. We must do this for zero, since
4166 the code below can't handle that one case. */
4167 if ((i
& ~(unsigned HOST_WIDE_INT
) 0xff) == 0)
4170 /* Get the number of trailing zeros. */
4171 lowbit
= ffs((int) i
) - 1;
4173 /* Only even shifts are allowed in ARM mode so round down to the
4174 nearest even number. */
4178 if ((i
& ~(((unsigned HOST_WIDE_INT
) 0xff) << lowbit
)) == 0)
4183 /* Allow rotated constants in ARM mode. */
4185 && ((i
& ~0xc000003f) == 0
4186 || (i
& ~0xf000000f) == 0
4187 || (i
& ~0xfc000003) == 0))
4190 else if (TARGET_THUMB2
)
4194 /* Allow repeated patterns 0x00XY00XY or 0xXYXYXYXY. */
4197 if (i
== v
|| i
== (v
| (v
<< 8)))
4200 /* Allow repeated pattern 0xXY00XY00. */
4206 else if (TARGET_HAVE_MOVT
)
4208 /* Thumb-1 Targets with MOVT. */
4218 /* Return true if I is a valid constant for the operation CODE. */
4220 const_ok_for_op (HOST_WIDE_INT i
, enum rtx_code code
)
4222 if (const_ok_for_arm (i
))
4228 /* See if we can use movw. */
4229 if (TARGET_HAVE_MOVT
&& (i
& 0xffff0000) == 0)
4232 /* Otherwise, try mvn. */
4233 return const_ok_for_arm (ARM_SIGN_EXTEND (~i
));
4236 /* See if we can use addw or subw. */
4238 && ((i
& 0xfffff000) == 0
4239 || ((-i
) & 0xfffff000) == 0))
4260 return const_ok_for_arm (ARM_SIGN_EXTEND (-i
));
4262 case MINUS
: /* Should only occur with (MINUS I reg) => rsb */
4268 return const_ok_for_arm (ARM_SIGN_EXTEND (~i
));
4272 return const_ok_for_arm (ARM_SIGN_EXTEND (~i
));
4279 /* Return true if I is a valid di mode constant for the operation CODE. */
4281 const_ok_for_dimode_op (HOST_WIDE_INT i
, enum rtx_code code
)
4283 HOST_WIDE_INT hi_val
= (i
>> 32) & 0xFFFFFFFF;
4284 HOST_WIDE_INT lo_val
= i
& 0xFFFFFFFF;
4285 rtx hi
= GEN_INT (hi_val
);
4286 rtx lo
= GEN_INT (lo_val
);
4296 return (const_ok_for_op (hi_val
, code
) || hi_val
== 0xFFFFFFFF)
4297 && (const_ok_for_op (lo_val
, code
) || lo_val
== 0xFFFFFFFF);
4299 return arm_not_operand (hi
, SImode
) && arm_add_operand (lo
, SImode
);
4306 /* Emit a sequence of insns to handle a large constant.
4307 CODE is the code of the operation required, it can be any of SET, PLUS,
4308 IOR, AND, XOR, MINUS;
4309 MODE is the mode in which the operation is being performed;
4310 VAL is the integer to operate on;
4311 SOURCE is the other operand (a register, or a null-pointer for SET);
4312 SUBTARGETS means it is safe to create scratch registers if that will
4313 either produce a simpler sequence, or we will want to cse the values.
4314 Return value is the number of insns emitted. */
4316 /* ??? Tweak this for thumb2. */
4318 arm_split_constant (enum rtx_code code
, machine_mode mode
, rtx insn
,
4319 HOST_WIDE_INT val
, rtx target
, rtx source
, int subtargets
)
4323 if (insn
&& GET_CODE (PATTERN (insn
)) == COND_EXEC
)
4324 cond
= COND_EXEC_TEST (PATTERN (insn
));
4328 if (subtargets
|| code
== SET
4329 || (REG_P (target
) && REG_P (source
)
4330 && REGNO (target
) != REGNO (source
)))
4332 /* After arm_reorg has been called, we can't fix up expensive
4333 constants by pushing them into memory so we must synthesize
4334 them in-line, regardless of the cost. This is only likely to
4335 be more costly on chips that have load delay slots and we are
4336 compiling without running the scheduler (so no splitting
4337 occurred before the final instruction emission).
4339 Ref: gcc -O1 -mcpu=strongarm gcc.c-torture/compile/980506-2.c
4341 if (!cfun
->machine
->after_arm_reorg
4343 && (arm_gen_constant (code
, mode
, NULL_RTX
, val
, target
, source
,
4345 > (arm_constant_limit (optimize_function_for_size_p (cfun
))
4350 /* Currently SET is the only monadic value for CODE, all
4351 the rest are diadic. */
4352 if (TARGET_USE_MOVT
)
4353 arm_emit_movpair (target
, GEN_INT (val
));
4355 emit_set_insn (target
, GEN_INT (val
));
4361 rtx temp
= subtargets
? gen_reg_rtx (mode
) : target
;
4363 if (TARGET_USE_MOVT
)
4364 arm_emit_movpair (temp
, GEN_INT (val
));
4366 emit_set_insn (temp
, GEN_INT (val
));
4368 /* For MINUS, the value is subtracted from, since we never
4369 have subtraction of a constant. */
4371 emit_set_insn (target
, gen_rtx_MINUS (mode
, temp
, source
));
4373 emit_set_insn (target
,
4374 gen_rtx_fmt_ee (code
, mode
, source
, temp
));
4380 return arm_gen_constant (code
, mode
, cond
, val
, target
, source
, subtargets
,
4384 /* Return a sequence of integers, in RETURN_SEQUENCE that fit into
4385 ARM/THUMB2 immediates, and add up to VAL.
4386 Thr function return value gives the number of insns required. */
4388 optimal_immediate_sequence (enum rtx_code code
, unsigned HOST_WIDE_INT val
,
4389 struct four_ints
*return_sequence
)
4391 int best_consecutive_zeros
= 0;
4395 struct four_ints tmp_sequence
;
4397 /* If we aren't targeting ARM, the best place to start is always at
4398 the bottom, otherwise look more closely. */
4401 for (i
= 0; i
< 32; i
+= 2)
4403 int consecutive_zeros
= 0;
4405 if (!(val
& (3 << i
)))
4407 while ((i
< 32) && !(val
& (3 << i
)))
4409 consecutive_zeros
+= 2;
4412 if (consecutive_zeros
> best_consecutive_zeros
)
4414 best_consecutive_zeros
= consecutive_zeros
;
4415 best_start
= i
- consecutive_zeros
;
4422 /* So long as it won't require any more insns to do so, it's
4423 desirable to emit a small constant (in bits 0...9) in the last
4424 insn. This way there is more chance that it can be combined with
4425 a later addressing insn to form a pre-indexed load or store
4426 operation. Consider:
4428 *((volatile int *)0xe0000100) = 1;
4429 *((volatile int *)0xe0000110) = 2;
4431 We want this to wind up as:
4435 str rB, [rA, #0x100]
4437 str rB, [rA, #0x110]
4439 rather than having to synthesize both large constants from scratch.
4441 Therefore, we calculate how many insns would be required to emit
4442 the constant starting from `best_start', and also starting from
4443 zero (i.e. with bit 31 first to be output). If `best_start' doesn't
4444 yield a shorter sequence, we may as well use zero. */
4445 insns1
= optimal_immediate_sequence_1 (code
, val
, return_sequence
, best_start
);
4447 && ((HOST_WIDE_INT_1U
<< best_start
) < val
))
4449 insns2
= optimal_immediate_sequence_1 (code
, val
, &tmp_sequence
, 0);
4450 if (insns2
<= insns1
)
4452 *return_sequence
= tmp_sequence
;
4460 /* As for optimal_immediate_sequence, but starting at bit-position I. */
4462 optimal_immediate_sequence_1 (enum rtx_code code
, unsigned HOST_WIDE_INT val
,
4463 struct four_ints
*return_sequence
, int i
)
4465 int remainder
= val
& 0xffffffff;
4468 /* Try and find a way of doing the job in either two or three
4471 In ARM mode we can use 8-bit constants, rotated to any 2-bit aligned
4472 location. We start at position I. This may be the MSB, or
4473 optimial_immediate_sequence may have positioned it at the largest block
4474 of zeros that are aligned on a 2-bit boundary. We then fill up the temps,
4475 wrapping around to the top of the word when we drop off the bottom.
4476 In the worst case this code should produce no more than four insns.
4478 In Thumb2 mode, we can use 32/16-bit replicated constants, and 8-bit
4479 constants, shifted to any arbitrary location. We should always start
4484 unsigned int b1
, b2
, b3
, b4
;
4485 unsigned HOST_WIDE_INT result
;
4488 gcc_assert (insns
< 4);
4493 /* First, find the next normal 12/8-bit shifted/rotated immediate. */
4494 if (remainder
& ((TARGET_ARM
? (3 << (i
- 2)) : (1 << (i
- 1)))))
4497 if (i
<= 12 && TARGET_THUMB2
&& code
== PLUS
)
4498 /* We can use addw/subw for the last 12 bits. */
4502 /* Use an 8-bit shifted/rotated immediate. */
4506 result
= remainder
& ((0x0ff << end
)
4507 | ((i
< end
) ? (0xff >> (32 - end
))
4514 /* Arm allows rotates by a multiple of two. Thumb-2 allows
4515 arbitrary shifts. */
4516 i
-= TARGET_ARM
? 2 : 1;
4520 /* Next, see if we can do a better job with a thumb2 replicated
4523 We do it this way around to catch the cases like 0x01F001E0 where
4524 two 8-bit immediates would work, but a replicated constant would
4527 TODO: 16-bit constants that don't clear all the bits, but still win.
4528 TODO: Arithmetic splitting for set/add/sub, rather than bitwise. */
4531 b1
= (remainder
& 0xff000000) >> 24;
4532 b2
= (remainder
& 0x00ff0000) >> 16;
4533 b3
= (remainder
& 0x0000ff00) >> 8;
4534 b4
= remainder
& 0xff;
4538 /* The 8-bit immediate already found clears b1 (and maybe b2),
4539 but must leave b3 and b4 alone. */
4541 /* First try to find a 32-bit replicated constant that clears
4542 almost everything. We can assume that we can't do it in one,
4543 or else we wouldn't be here. */
4544 unsigned int tmp
= b1
& b2
& b3
& b4
;
4545 unsigned int tmp2
= tmp
+ (tmp
<< 8) + (tmp
<< 16)
4547 unsigned int matching_bytes
= (tmp
== b1
) + (tmp
== b2
)
4548 + (tmp
== b3
) + (tmp
== b4
);
4550 && (matching_bytes
>= 3
4551 || (matching_bytes
== 2
4552 && const_ok_for_op (remainder
& ~tmp2
, code
))))
4554 /* At least 3 of the bytes match, and the fourth has at
4555 least as many bits set, or two of the bytes match
4556 and it will only require one more insn to finish. */
4564 /* Second, try to find a 16-bit replicated constant that can
4565 leave three of the bytes clear. If b2 or b4 is already
4566 zero, then we can. If the 8-bit from above would not
4567 clear b2 anyway, then we still win. */
4568 else if (b1
== b3
&& (!b2
|| !b4
4569 || (remainder
& 0x00ff0000 & ~result
)))
4571 result
= remainder
& 0xff00ff00;
4577 /* The 8-bit immediate already found clears b2 (and maybe b3)
4578 and we don't get here unless b1 is alredy clear, but it will
4579 leave b4 unchanged. */
4581 /* If we can clear b2 and b4 at once, then we win, since the
4582 8-bits couldn't possibly reach that far. */
4585 result
= remainder
& 0x00ff00ff;
4591 return_sequence
->i
[insns
++] = result
;
4592 remainder
&= ~result
;
4594 if (code
== SET
|| code
== MINUS
)
4602 /* Emit an instruction with the indicated PATTERN. If COND is
4603 non-NULL, conditionalize the execution of the instruction on COND
4607 emit_constant_insn (rtx cond
, rtx pattern
)
4610 pattern
= gen_rtx_COND_EXEC (VOIDmode
, copy_rtx (cond
), pattern
);
4611 emit_insn (pattern
);
4614 /* As above, but extra parameter GENERATE which, if clear, suppresses
4618 arm_gen_constant (enum rtx_code code
, machine_mode mode
, rtx cond
,
4619 unsigned HOST_WIDE_INT val
, rtx target
, rtx source
,
4620 int subtargets
, int generate
)
4624 int final_invert
= 0;
4626 int set_sign_bit_copies
= 0;
4627 int clear_sign_bit_copies
= 0;
4628 int clear_zero_bit_copies
= 0;
4629 int set_zero_bit_copies
= 0;
4630 int insns
= 0, neg_insns
, inv_insns
;
4631 unsigned HOST_WIDE_INT temp1
, temp2
;
4632 unsigned HOST_WIDE_INT remainder
= val
& 0xffffffff;
4633 struct four_ints
*immediates
;
4634 struct four_ints pos_immediates
, neg_immediates
, inv_immediates
;
4636 /* Find out which operations are safe for a given CODE. Also do a quick
4637 check for degenerate cases; these can occur when DImode operations
4650 if (remainder
== 0xffffffff)
4653 emit_constant_insn (cond
,
4654 gen_rtx_SET (target
,
4655 GEN_INT (ARM_SIGN_EXTEND (val
))));
4661 if (reload_completed
&& rtx_equal_p (target
, source
))
4665 emit_constant_insn (cond
, gen_rtx_SET (target
, source
));
4674 emit_constant_insn (cond
, gen_rtx_SET (target
, const0_rtx
));
4677 if (remainder
== 0xffffffff)
4679 if (reload_completed
&& rtx_equal_p (target
, source
))
4682 emit_constant_insn (cond
, gen_rtx_SET (target
, source
));
4691 if (reload_completed
&& rtx_equal_p (target
, source
))
4694 emit_constant_insn (cond
, gen_rtx_SET (target
, source
));
4698 if (remainder
== 0xffffffff)
4701 emit_constant_insn (cond
,
4702 gen_rtx_SET (target
,
4703 gen_rtx_NOT (mode
, source
)));
4710 /* We treat MINUS as (val - source), since (source - val) is always
4711 passed as (source + (-val)). */
4715 emit_constant_insn (cond
,
4716 gen_rtx_SET (target
,
4717 gen_rtx_NEG (mode
, source
)));
4720 if (const_ok_for_arm (val
))
4723 emit_constant_insn (cond
,
4724 gen_rtx_SET (target
,
4725 gen_rtx_MINUS (mode
, GEN_INT (val
),
4736 /* If we can do it in one insn get out quickly. */
4737 if (const_ok_for_op (val
, code
))
4740 emit_constant_insn (cond
,
4741 gen_rtx_SET (target
,
4743 ? gen_rtx_fmt_ee (code
, mode
, source
,
4749 /* On targets with UXTH/UBFX, we can deal with AND (2^N)-1 in a single
4751 if (code
== AND
&& (i
= exact_log2 (remainder
+ 1)) > 0
4752 && (arm_arch_thumb2
|| (i
== 16 && arm_arch6
&& mode
== SImode
)))
4756 if (mode
== SImode
&& i
== 16)
4757 /* Use UXTH in preference to UBFX, since on Thumb2 it's a
4759 emit_constant_insn (cond
,
4760 gen_zero_extendhisi2
4761 (target
, gen_lowpart (HImode
, source
)));
4763 /* Extz only supports SImode, but we can coerce the operands
4765 emit_constant_insn (cond
,
4766 gen_extzv_t2 (gen_lowpart (SImode
, target
),
4767 gen_lowpart (SImode
, source
),
4768 GEN_INT (i
), const0_rtx
));
4774 /* Calculate a few attributes that may be useful for specific
4776 /* Count number of leading zeros. */
4777 for (i
= 31; i
>= 0; i
--)
4779 if ((remainder
& (1 << i
)) == 0)
4780 clear_sign_bit_copies
++;
4785 /* Count number of leading 1's. */
4786 for (i
= 31; i
>= 0; i
--)
4788 if ((remainder
& (1 << i
)) != 0)
4789 set_sign_bit_copies
++;
4794 /* Count number of trailing zero's. */
4795 for (i
= 0; i
<= 31; i
++)
4797 if ((remainder
& (1 << i
)) == 0)
4798 clear_zero_bit_copies
++;
4803 /* Count number of trailing 1's. */
4804 for (i
= 0; i
<= 31; i
++)
4806 if ((remainder
& (1 << i
)) != 0)
4807 set_zero_bit_copies
++;
4815 /* See if we can do this by sign_extending a constant that is known
4816 to be negative. This is a good, way of doing it, since the shift
4817 may well merge into a subsequent insn. */
4818 if (set_sign_bit_copies
> 1)
4820 if (const_ok_for_arm
4821 (temp1
= ARM_SIGN_EXTEND (remainder
4822 << (set_sign_bit_copies
- 1))))
4826 rtx new_src
= subtargets
? gen_reg_rtx (mode
) : target
;
4827 emit_constant_insn (cond
,
4828 gen_rtx_SET (new_src
, GEN_INT (temp1
)));
4829 emit_constant_insn (cond
,
4830 gen_ashrsi3 (target
, new_src
,
4831 GEN_INT (set_sign_bit_copies
- 1)));
4835 /* For an inverted constant, we will need to set the low bits,
4836 these will be shifted out of harm's way. */
4837 temp1
|= (1 << (set_sign_bit_copies
- 1)) - 1;
4838 if (const_ok_for_arm (~temp1
))
4842 rtx new_src
= subtargets
? gen_reg_rtx (mode
) : target
;
4843 emit_constant_insn (cond
,
4844 gen_rtx_SET (new_src
, GEN_INT (temp1
)));
4845 emit_constant_insn (cond
,
4846 gen_ashrsi3 (target
, new_src
,
4847 GEN_INT (set_sign_bit_copies
- 1)));
4853 /* See if we can calculate the value as the difference between two
4854 valid immediates. */
4855 if (clear_sign_bit_copies
+ clear_zero_bit_copies
<= 16)
4857 int topshift
= clear_sign_bit_copies
& ~1;
4859 temp1
= ARM_SIGN_EXTEND ((remainder
+ (0x00800000 >> topshift
))
4860 & (0xff000000 >> topshift
));
4862 /* If temp1 is zero, then that means the 9 most significant
4863 bits of remainder were 1 and we've caused it to overflow.
4864 When topshift is 0 we don't need to do anything since we
4865 can borrow from 'bit 32'. */
4866 if (temp1
== 0 && topshift
!= 0)
4867 temp1
= 0x80000000 >> (topshift
- 1);
4869 temp2
= ARM_SIGN_EXTEND (temp1
- remainder
);
4871 if (const_ok_for_arm (temp2
))
4875 rtx new_src
= subtargets
? gen_reg_rtx (mode
) : target
;
4876 emit_constant_insn (cond
,
4877 gen_rtx_SET (new_src
, GEN_INT (temp1
)));
4878 emit_constant_insn (cond
,
4879 gen_addsi3 (target
, new_src
,
4887 /* See if we can generate this by setting the bottom (or the top)
4888 16 bits, and then shifting these into the other half of the
4889 word. We only look for the simplest cases, to do more would cost
4890 too much. Be careful, however, not to generate this when the
4891 alternative would take fewer insns. */
4892 if (val
& 0xffff0000)
4894 temp1
= remainder
& 0xffff0000;
4895 temp2
= remainder
& 0x0000ffff;
4897 /* Overlaps outside this range are best done using other methods. */
4898 for (i
= 9; i
< 24; i
++)
4900 if ((((temp2
| (temp2
<< i
)) & 0xffffffff) == remainder
)
4901 && !const_ok_for_arm (temp2
))
4903 rtx new_src
= (subtargets
4904 ? (generate
? gen_reg_rtx (mode
) : NULL_RTX
)
4906 insns
= arm_gen_constant (code
, mode
, cond
, temp2
, new_src
,
4907 source
, subtargets
, generate
);
4915 gen_rtx_ASHIFT (mode
, source
,
4922 /* Don't duplicate cases already considered. */
4923 for (i
= 17; i
< 24; i
++)
4925 if (((temp1
| (temp1
>> i
)) == remainder
)
4926 && !const_ok_for_arm (temp1
))
4928 rtx new_src
= (subtargets
4929 ? (generate
? gen_reg_rtx (mode
) : NULL_RTX
)
4931 insns
= arm_gen_constant (code
, mode
, cond
, temp1
, new_src
,
4932 source
, subtargets
, generate
);
4937 gen_rtx_SET (target
,
4940 gen_rtx_LSHIFTRT (mode
, source
,
4951 /* If we have IOR or XOR, and the constant can be loaded in a
4952 single instruction, and we can find a temporary to put it in,
4953 then this can be done in two instructions instead of 3-4. */
4955 /* TARGET can't be NULL if SUBTARGETS is 0 */
4956 || (reload_completed
&& !reg_mentioned_p (target
, source
)))
4958 if (const_ok_for_arm (ARM_SIGN_EXTEND (~val
)))
4962 rtx sub
= subtargets
? gen_reg_rtx (mode
) : target
;
4964 emit_constant_insn (cond
,
4965 gen_rtx_SET (sub
, GEN_INT (val
)));
4966 emit_constant_insn (cond
,
4967 gen_rtx_SET (target
,
4968 gen_rtx_fmt_ee (code
, mode
,
4979 x = y | constant ( which is composed of set_sign_bit_copies of leading 1s
4980 and the remainder 0s for e.g. 0xfff00000)
4981 x = ~(~(y ashift set_sign_bit_copies) lshiftrt set_sign_bit_copies)
4983 This can be done in 2 instructions by using shifts with mov or mvn.
4988 mvn r0, r0, lsr #12 */
4989 if (set_sign_bit_copies
> 8
4990 && (val
& (HOST_WIDE_INT_M1U
<< (32 - set_sign_bit_copies
))) == val
)
4994 rtx sub
= subtargets
? gen_reg_rtx (mode
) : target
;
4995 rtx shift
= GEN_INT (set_sign_bit_copies
);
5001 gen_rtx_ASHIFT (mode
,
5006 gen_rtx_SET (target
,
5008 gen_rtx_LSHIFTRT (mode
, sub
,
5015 x = y | constant (which has set_zero_bit_copies number of trailing ones).
5017 x = ~((~y lshiftrt set_zero_bit_copies) ashift set_zero_bit_copies).
5019 For eg. r0 = r0 | 0xfff
5024 if (set_zero_bit_copies
> 8
5025 && (remainder
& ((1 << set_zero_bit_copies
) - 1)) == remainder
)
5029 rtx sub
= subtargets
? gen_reg_rtx (mode
) : target
;
5030 rtx shift
= GEN_INT (set_zero_bit_copies
);
5036 gen_rtx_LSHIFTRT (mode
,
5041 gen_rtx_SET (target
,
5043 gen_rtx_ASHIFT (mode
, sub
,
5049 /* This will never be reached for Thumb2 because orn is a valid
5050 instruction. This is for Thumb1 and the ARM 32 bit cases.
5052 x = y | constant (such that ~constant is a valid constant)
5054 x = ~(~y & ~constant).
5056 if (const_ok_for_arm (temp1
= ARM_SIGN_EXTEND (~val
)))
5060 rtx sub
= subtargets
? gen_reg_rtx (mode
) : target
;
5061 emit_constant_insn (cond
,
5063 gen_rtx_NOT (mode
, source
)));
5066 sub
= gen_reg_rtx (mode
);
5067 emit_constant_insn (cond
,
5069 gen_rtx_AND (mode
, source
,
5071 emit_constant_insn (cond
,
5072 gen_rtx_SET (target
,
5073 gen_rtx_NOT (mode
, sub
)));
5080 /* See if two shifts will do 2 or more insn's worth of work. */
5081 if (clear_sign_bit_copies
>= 16 && clear_sign_bit_copies
< 24)
5083 HOST_WIDE_INT shift_mask
= ((0xffffffff
5084 << (32 - clear_sign_bit_copies
))
5087 if ((remainder
| shift_mask
) != 0xffffffff)
5089 HOST_WIDE_INT new_val
5090 = ARM_SIGN_EXTEND (remainder
| shift_mask
);
5094 rtx new_src
= subtargets
? gen_reg_rtx (mode
) : target
;
5095 insns
= arm_gen_constant (AND
, SImode
, cond
, new_val
,
5096 new_src
, source
, subtargets
, 1);
5101 rtx targ
= subtargets
? NULL_RTX
: target
;
5102 insns
= arm_gen_constant (AND
, mode
, cond
, new_val
,
5103 targ
, source
, subtargets
, 0);
5109 rtx new_src
= subtargets
? gen_reg_rtx (mode
) : target
;
5110 rtx shift
= GEN_INT (clear_sign_bit_copies
);
5112 emit_insn (gen_ashlsi3 (new_src
, source
, shift
));
5113 emit_insn (gen_lshrsi3 (target
, new_src
, shift
));
5119 if (clear_zero_bit_copies
>= 16 && clear_zero_bit_copies
< 24)
5121 HOST_WIDE_INT shift_mask
= (1 << clear_zero_bit_copies
) - 1;
5123 if ((remainder
| shift_mask
) != 0xffffffff)
5125 HOST_WIDE_INT new_val
5126 = ARM_SIGN_EXTEND (remainder
| shift_mask
);
5129 rtx new_src
= subtargets
? gen_reg_rtx (mode
) : target
;
5131 insns
= arm_gen_constant (AND
, mode
, cond
, new_val
,
5132 new_src
, source
, subtargets
, 1);
5137 rtx targ
= subtargets
? NULL_RTX
: target
;
5139 insns
= arm_gen_constant (AND
, mode
, cond
, new_val
,
5140 targ
, source
, subtargets
, 0);
5146 rtx new_src
= subtargets
? gen_reg_rtx (mode
) : target
;
5147 rtx shift
= GEN_INT (clear_zero_bit_copies
);
5149 emit_insn (gen_lshrsi3 (new_src
, source
, shift
));
5150 emit_insn (gen_ashlsi3 (target
, new_src
, shift
));
5162 /* Calculate what the instruction sequences would be if we generated it
5163 normally, negated, or inverted. */
5165 /* AND cannot be split into multiple insns, so invert and use BIC. */
5168 insns
= optimal_immediate_sequence (code
, remainder
, &pos_immediates
);
5171 neg_insns
= optimal_immediate_sequence (code
, (-remainder
) & 0xffffffff,
5176 if (can_invert
|| final_invert
)
5177 inv_insns
= optimal_immediate_sequence (code
, remainder
^ 0xffffffff,
5182 immediates
= &pos_immediates
;
5184 /* Is the negated immediate sequence more efficient? */
5185 if (neg_insns
< insns
&& neg_insns
<= inv_insns
)
5188 immediates
= &neg_immediates
;
5193 /* Is the inverted immediate sequence more efficient?
5194 We must allow for an extra NOT instruction for XOR operations, although
5195 there is some chance that the final 'mvn' will get optimized later. */
5196 if ((inv_insns
+ 1) < insns
|| (!final_invert
&& inv_insns
< insns
))
5199 immediates
= &inv_immediates
;
5207 /* Now output the chosen sequence as instructions. */
5210 for (i
= 0; i
< insns
; i
++)
5212 rtx new_src
, temp1_rtx
;
5214 temp1
= immediates
->i
[i
];
5216 if (code
== SET
|| code
== MINUS
)
5217 new_src
= (subtargets
? gen_reg_rtx (mode
) : target
);
5218 else if ((final_invert
|| i
< (insns
- 1)) && subtargets
)
5219 new_src
= gen_reg_rtx (mode
);
5225 else if (can_negate
)
5228 temp1
= trunc_int_for_mode (temp1
, mode
);
5229 temp1_rtx
= GEN_INT (temp1
);
5233 else if (code
== MINUS
)
5234 temp1_rtx
= gen_rtx_MINUS (mode
, temp1_rtx
, source
);
5236 temp1_rtx
= gen_rtx_fmt_ee (code
, mode
, source
, temp1_rtx
);
5238 emit_constant_insn (cond
, gen_rtx_SET (new_src
, temp1_rtx
));
5243 can_negate
= can_invert
;
5247 else if (code
== MINUS
)
5255 emit_constant_insn (cond
, gen_rtx_SET (target
,
5256 gen_rtx_NOT (mode
, source
)));
5263 /* Canonicalize a comparison so that we are more likely to recognize it.
5264 This can be done for a few constant compares, where we can make the
5265 immediate value easier to load. */
5268 arm_canonicalize_comparison (int *code
, rtx
*op0
, rtx
*op1
,
5269 bool op0_preserve_value
)
5272 unsigned HOST_WIDE_INT i
, maxval
;
5274 mode
= GET_MODE (*op0
);
5275 if (mode
== VOIDmode
)
5276 mode
= GET_MODE (*op1
);
5278 maxval
= (HOST_WIDE_INT_1U
<< (GET_MODE_BITSIZE (mode
) - 1)) - 1;
5280 /* For DImode, we have GE/LT/GEU/LTU comparisons. In ARM mode
5281 we can also use cmp/cmpeq for GTU/LEU. GT/LE must be either
5282 reversed or (for constant OP1) adjusted to GE/LT. Similarly
5283 for GTU/LEU in Thumb mode. */
5287 if (*code
== GT
|| *code
== LE
5288 || (!TARGET_ARM
&& (*code
== GTU
|| *code
== LEU
)))
5290 /* Missing comparison. First try to use an available
5292 if (CONST_INT_P (*op1
))
5300 && arm_const_double_by_immediates (GEN_INT (i
+ 1)))
5302 *op1
= GEN_INT (i
+ 1);
5303 *code
= *code
== GT
? GE
: LT
;
5309 if (i
!= ~((unsigned HOST_WIDE_INT
) 0)
5310 && arm_const_double_by_immediates (GEN_INT (i
+ 1)))
5312 *op1
= GEN_INT (i
+ 1);
5313 *code
= *code
== GTU
? GEU
: LTU
;
5322 /* If that did not work, reverse the condition. */
5323 if (!op0_preserve_value
)
5325 std::swap (*op0
, *op1
);
5326 *code
= (int)swap_condition ((enum rtx_code
)*code
);
5332 /* If *op0 is (zero_extend:SI (subreg:QI (reg:SI) 0)) and comparing
5333 with const0_rtx, change it to (and:SI (reg:SI) (const_int 255)),
5334 to facilitate possible combining with a cmp into 'ands'. */
5336 && GET_CODE (*op0
) == ZERO_EXTEND
5337 && GET_CODE (XEXP (*op0
, 0)) == SUBREG
5338 && GET_MODE (XEXP (*op0
, 0)) == QImode
5339 && GET_MODE (SUBREG_REG (XEXP (*op0
, 0))) == SImode
5340 && subreg_lowpart_p (XEXP (*op0
, 0))
5341 && *op1
== const0_rtx
)
5342 *op0
= gen_rtx_AND (SImode
, SUBREG_REG (XEXP (*op0
, 0)),
5345 /* Comparisons smaller than DImode. Only adjust comparisons against
5346 an out-of-range constant. */
5347 if (!CONST_INT_P (*op1
)
5348 || const_ok_for_arm (INTVAL (*op1
))
5349 || const_ok_for_arm (- INTVAL (*op1
)))
5363 && (const_ok_for_arm (i
+ 1) || const_ok_for_arm (-(i
+ 1))))
5365 *op1
= GEN_INT (ARM_SIGN_EXTEND (i
+ 1));
5366 *code
= *code
== GT
? GE
: LT
;
5374 && (const_ok_for_arm (i
- 1) || const_ok_for_arm (-(i
- 1))))
5376 *op1
= GEN_INT (i
- 1);
5377 *code
= *code
== GE
? GT
: LE
;
5384 if (i
!= ~((unsigned HOST_WIDE_INT
) 0)
5385 && (const_ok_for_arm (i
+ 1) || const_ok_for_arm (-(i
+ 1))))
5387 *op1
= GEN_INT (ARM_SIGN_EXTEND (i
+ 1));
5388 *code
= *code
== GTU
? GEU
: LTU
;
5396 && (const_ok_for_arm (i
- 1) || const_ok_for_arm (-(i
- 1))))
5398 *op1
= GEN_INT (i
- 1);
5399 *code
= *code
== GEU
? GTU
: LEU
;
5410 /* Define how to find the value returned by a function. */
5413 arm_function_value(const_tree type
, const_tree func
,
5414 bool outgoing ATTRIBUTE_UNUSED
)
5417 int unsignedp ATTRIBUTE_UNUSED
;
5418 rtx r ATTRIBUTE_UNUSED
;
5420 mode
= TYPE_MODE (type
);
5422 if (TARGET_AAPCS_BASED
)
5423 return aapcs_allocate_return_reg (mode
, type
, func
);
5425 /* Promote integer types. */
5426 if (INTEGRAL_TYPE_P (type
))
5427 mode
= arm_promote_function_mode (type
, mode
, &unsignedp
, func
, 1);
5429 /* Promotes small structs returned in a register to full-word size
5430 for big-endian AAPCS. */
5431 if (arm_return_in_msb (type
))
5433 HOST_WIDE_INT size
= int_size_in_bytes (type
);
5434 if (size
% UNITS_PER_WORD
!= 0)
5436 size
+= UNITS_PER_WORD
- size
% UNITS_PER_WORD
;
5437 mode
= int_mode_for_size (size
* BITS_PER_UNIT
, 0).require ();
5441 return arm_libcall_value_1 (mode
);
5444 /* libcall hashtable helpers. */
5446 struct libcall_hasher
: nofree_ptr_hash
<const rtx_def
>
5448 static inline hashval_t
hash (const rtx_def
*);
5449 static inline bool equal (const rtx_def
*, const rtx_def
*);
5450 static inline void remove (rtx_def
*);
5454 libcall_hasher::equal (const rtx_def
*p1
, const rtx_def
*p2
)
5456 return rtx_equal_p (p1
, p2
);
5460 libcall_hasher::hash (const rtx_def
*p1
)
5462 return hash_rtx (p1
, VOIDmode
, NULL
, NULL
, FALSE
);
5465 typedef hash_table
<libcall_hasher
> libcall_table_type
;
5468 add_libcall (libcall_table_type
*htab
, rtx libcall
)
5470 *htab
->find_slot (libcall
, INSERT
) = libcall
;
5474 arm_libcall_uses_aapcs_base (const_rtx libcall
)
5476 static bool init_done
= false;
5477 static libcall_table_type
*libcall_htab
= NULL
;
5483 libcall_htab
= new libcall_table_type (31);
5484 add_libcall (libcall_htab
,
5485 convert_optab_libfunc (sfloat_optab
, SFmode
, SImode
));
5486 add_libcall (libcall_htab
,
5487 convert_optab_libfunc (sfloat_optab
, DFmode
, SImode
));
5488 add_libcall (libcall_htab
,
5489 convert_optab_libfunc (sfloat_optab
, SFmode
, DImode
));
5490 add_libcall (libcall_htab
,
5491 convert_optab_libfunc (sfloat_optab
, DFmode
, DImode
));
5493 add_libcall (libcall_htab
,
5494 convert_optab_libfunc (ufloat_optab
, SFmode
, SImode
));
5495 add_libcall (libcall_htab
,
5496 convert_optab_libfunc (ufloat_optab
, DFmode
, SImode
));
5497 add_libcall (libcall_htab
,
5498 convert_optab_libfunc (ufloat_optab
, SFmode
, DImode
));
5499 add_libcall (libcall_htab
,
5500 convert_optab_libfunc (ufloat_optab
, DFmode
, DImode
));
5502 add_libcall (libcall_htab
,
5503 convert_optab_libfunc (sext_optab
, SFmode
, HFmode
));
5504 add_libcall (libcall_htab
,
5505 convert_optab_libfunc (trunc_optab
, HFmode
, SFmode
));
5506 add_libcall (libcall_htab
,
5507 convert_optab_libfunc (sfix_optab
, SImode
, DFmode
));
5508 add_libcall (libcall_htab
,
5509 convert_optab_libfunc (ufix_optab
, SImode
, DFmode
));
5510 add_libcall (libcall_htab
,
5511 convert_optab_libfunc (sfix_optab
, DImode
, DFmode
));
5512 add_libcall (libcall_htab
,
5513 convert_optab_libfunc (ufix_optab
, DImode
, DFmode
));
5514 add_libcall (libcall_htab
,
5515 convert_optab_libfunc (sfix_optab
, DImode
, SFmode
));
5516 add_libcall (libcall_htab
,
5517 convert_optab_libfunc (ufix_optab
, DImode
, SFmode
));
5519 /* Values from double-precision helper functions are returned in core
5520 registers if the selected core only supports single-precision
5521 arithmetic, even if we are using the hard-float ABI. The same is
5522 true for single-precision helpers, but we will never be using the
5523 hard-float ABI on a CPU which doesn't support single-precision
5524 operations in hardware. */
5525 add_libcall (libcall_htab
, optab_libfunc (add_optab
, DFmode
));
5526 add_libcall (libcall_htab
, optab_libfunc (sdiv_optab
, DFmode
));
5527 add_libcall (libcall_htab
, optab_libfunc (smul_optab
, DFmode
));
5528 add_libcall (libcall_htab
, optab_libfunc (neg_optab
, DFmode
));
5529 add_libcall (libcall_htab
, optab_libfunc (sub_optab
, DFmode
));
5530 add_libcall (libcall_htab
, optab_libfunc (eq_optab
, DFmode
));
5531 add_libcall (libcall_htab
, optab_libfunc (lt_optab
, DFmode
));
5532 add_libcall (libcall_htab
, optab_libfunc (le_optab
, DFmode
));
5533 add_libcall (libcall_htab
, optab_libfunc (ge_optab
, DFmode
));
5534 add_libcall (libcall_htab
, optab_libfunc (gt_optab
, DFmode
));
5535 add_libcall (libcall_htab
, optab_libfunc (unord_optab
, DFmode
));
5536 add_libcall (libcall_htab
, convert_optab_libfunc (sext_optab
, DFmode
,
5538 add_libcall (libcall_htab
, convert_optab_libfunc (trunc_optab
, SFmode
,
5540 add_libcall (libcall_htab
,
5541 convert_optab_libfunc (trunc_optab
, HFmode
, DFmode
));
5544 return libcall
&& libcall_htab
->find (libcall
) != NULL
;
5548 arm_libcall_value_1 (machine_mode mode
)
5550 if (TARGET_AAPCS_BASED
)
5551 return aapcs_libcall_value (mode
);
5552 else if (TARGET_IWMMXT_ABI
5553 && arm_vector_mode_supported_p (mode
))
5554 return gen_rtx_REG (mode
, FIRST_IWMMXT_REGNUM
);
5556 return gen_rtx_REG (mode
, ARG_REGISTER (1));
5559 /* Define how to find the value returned by a library function
5560 assuming the value has mode MODE. */
5563 arm_libcall_value (machine_mode mode
, const_rtx libcall
)
5565 if (TARGET_AAPCS_BASED
&& arm_pcs_default
!= ARM_PCS_AAPCS
5566 && GET_MODE_CLASS (mode
) == MODE_FLOAT
)
5568 /* The following libcalls return their result in integer registers,
5569 even though they return a floating point value. */
5570 if (arm_libcall_uses_aapcs_base (libcall
))
5571 return gen_rtx_REG (mode
, ARG_REGISTER(1));
5575 return arm_libcall_value_1 (mode
);
5578 /* Implement TARGET_FUNCTION_VALUE_REGNO_P. */
5581 arm_function_value_regno_p (const unsigned int regno
)
5583 if (regno
== ARG_REGISTER (1)
5585 && TARGET_AAPCS_BASED
5586 && TARGET_HARD_FLOAT
5587 && regno
== FIRST_VFP_REGNUM
)
5588 || (TARGET_IWMMXT_ABI
5589 && regno
== FIRST_IWMMXT_REGNUM
))
5595 /* Determine the amount of memory needed to store the possible return
5596 registers of an untyped call. */
5598 arm_apply_result_size (void)
5604 if (TARGET_HARD_FLOAT_ABI
)
5606 if (TARGET_IWMMXT_ABI
)
5613 /* Decide whether TYPE should be returned in memory (true)
5614 or in a register (false). FNTYPE is the type of the function making
5617 arm_return_in_memory (const_tree type
, const_tree fntype
)
5621 size
= int_size_in_bytes (type
); /* Negative if not fixed size. */
5623 if (TARGET_AAPCS_BASED
)
5625 /* Simple, non-aggregate types (ie not including vectors and
5626 complex) are always returned in a register (or registers).
5627 We don't care about which register here, so we can short-cut
5628 some of the detail. */
5629 if (!AGGREGATE_TYPE_P (type
)
5630 && TREE_CODE (type
) != VECTOR_TYPE
5631 && TREE_CODE (type
) != COMPLEX_TYPE
)
5634 /* Any return value that is no larger than one word can be
5636 if (((unsigned HOST_WIDE_INT
) size
) <= UNITS_PER_WORD
)
5639 /* Check any available co-processors to see if they accept the
5640 type as a register candidate (VFP, for example, can return
5641 some aggregates in consecutive registers). These aren't
5642 available if the call is variadic. */
5643 if (aapcs_select_return_coproc (type
, fntype
) >= 0)
5646 /* Vector values should be returned using ARM registers, not
5647 memory (unless they're over 16 bytes, which will break since
5648 we only have four call-clobbered registers to play with). */
5649 if (TREE_CODE (type
) == VECTOR_TYPE
)
5650 return (size
< 0 || size
> (4 * UNITS_PER_WORD
));
5652 /* The rest go in memory. */
5656 if (TREE_CODE (type
) == VECTOR_TYPE
)
5657 return (size
< 0 || size
> (4 * UNITS_PER_WORD
));
5659 if (!AGGREGATE_TYPE_P (type
) &&
5660 (TREE_CODE (type
) != VECTOR_TYPE
))
5661 /* All simple types are returned in registers. */
5664 if (arm_abi
!= ARM_ABI_APCS
)
5666 /* ATPCS and later return aggregate types in memory only if they are
5667 larger than a word (or are variable size). */
5668 return (size
< 0 || size
> UNITS_PER_WORD
);
5671 /* For the arm-wince targets we choose to be compatible with Microsoft's
5672 ARM and Thumb compilers, which always return aggregates in memory. */
5674 /* All structures/unions bigger than one word are returned in memory.
5675 Also catch the case where int_size_in_bytes returns -1. In this case
5676 the aggregate is either huge or of variable size, and in either case
5677 we will want to return it via memory and not in a register. */
5678 if (size
< 0 || size
> UNITS_PER_WORD
)
5681 if (TREE_CODE (type
) == RECORD_TYPE
)
5685 /* For a struct the APCS says that we only return in a register
5686 if the type is 'integer like' and every addressable element
5687 has an offset of zero. For practical purposes this means
5688 that the structure can have at most one non bit-field element
5689 and that this element must be the first one in the structure. */
5691 /* Find the first field, ignoring non FIELD_DECL things which will
5692 have been created by C++. */
5693 for (field
= TYPE_FIELDS (type
);
5694 field
&& TREE_CODE (field
) != FIELD_DECL
;
5695 field
= DECL_CHAIN (field
))
5699 return false; /* An empty structure. Allowed by an extension to ANSI C. */
5701 /* Check that the first field is valid for returning in a register. */
5703 /* ... Floats are not allowed */
5704 if (FLOAT_TYPE_P (TREE_TYPE (field
)))
5707 /* ... Aggregates that are not themselves valid for returning in
5708 a register are not allowed. */
5709 if (arm_return_in_memory (TREE_TYPE (field
), NULL_TREE
))
5712 /* Now check the remaining fields, if any. Only bitfields are allowed,
5713 since they are not addressable. */
5714 for (field
= DECL_CHAIN (field
);
5716 field
= DECL_CHAIN (field
))
5718 if (TREE_CODE (field
) != FIELD_DECL
)
5721 if (!DECL_BIT_FIELD_TYPE (field
))
5728 if (TREE_CODE (type
) == UNION_TYPE
)
5732 /* Unions can be returned in registers if every element is
5733 integral, or can be returned in an integer register. */
5734 for (field
= TYPE_FIELDS (type
);
5736 field
= DECL_CHAIN (field
))
5738 if (TREE_CODE (field
) != FIELD_DECL
)
5741 if (FLOAT_TYPE_P (TREE_TYPE (field
)))
5744 if (arm_return_in_memory (TREE_TYPE (field
), NULL_TREE
))
5750 #endif /* not ARM_WINCE */
5752 /* Return all other types in memory. */
5756 const struct pcs_attribute_arg
5760 } pcs_attribute_args
[] =
5762 {"aapcs", ARM_PCS_AAPCS
},
5763 {"aapcs-vfp", ARM_PCS_AAPCS_VFP
},
5765 /* We could recognize these, but changes would be needed elsewhere
5766 * to implement them. */
5767 {"aapcs-iwmmxt", ARM_PCS_AAPCS_IWMMXT
},
5768 {"atpcs", ARM_PCS_ATPCS
},
5769 {"apcs", ARM_PCS_APCS
},
5771 {NULL
, ARM_PCS_UNKNOWN
}
5775 arm_pcs_from_attribute (tree attr
)
5777 const struct pcs_attribute_arg
*ptr
;
5780 /* Get the value of the argument. */
5781 if (TREE_VALUE (attr
) == NULL_TREE
5782 || TREE_CODE (TREE_VALUE (attr
)) != STRING_CST
)
5783 return ARM_PCS_UNKNOWN
;
5785 arg
= TREE_STRING_POINTER (TREE_VALUE (attr
));
5787 /* Check it against the list of known arguments. */
5788 for (ptr
= pcs_attribute_args
; ptr
->arg
!= NULL
; ptr
++)
5789 if (streq (arg
, ptr
->arg
))
5792 /* An unrecognized interrupt type. */
5793 return ARM_PCS_UNKNOWN
;
5796 /* Get the PCS variant to use for this call. TYPE is the function's type
5797 specification, DECL is the specific declartion. DECL may be null if
5798 the call could be indirect or if this is a library call. */
5800 arm_get_pcs_model (const_tree type
, const_tree decl
)
5802 bool user_convention
= false;
5803 enum arm_pcs user_pcs
= arm_pcs_default
;
5808 attr
= lookup_attribute ("pcs", TYPE_ATTRIBUTES (type
));
5811 user_pcs
= arm_pcs_from_attribute (TREE_VALUE (attr
));
5812 user_convention
= true;
5815 if (TARGET_AAPCS_BASED
)
5817 /* Detect varargs functions. These always use the base rules
5818 (no argument is ever a candidate for a co-processor
5820 bool base_rules
= stdarg_p (type
);
5822 if (user_convention
)
5824 if (user_pcs
> ARM_PCS_AAPCS_LOCAL
)
5825 sorry ("non-AAPCS derived PCS variant");
5826 else if (base_rules
&& user_pcs
!= ARM_PCS_AAPCS
)
5827 error ("variadic functions must use the base AAPCS variant");
5831 return ARM_PCS_AAPCS
;
5832 else if (user_convention
)
5834 else if (decl
&& flag_unit_at_a_time
)
5836 /* Local functions never leak outside this compilation unit,
5837 so we are free to use whatever conventions are
5839 /* FIXME: remove CONST_CAST_TREE when cgraph is constified. */
5840 cgraph_local_info
*i
= cgraph_node::local_info (CONST_CAST_TREE(decl
));
5842 return ARM_PCS_AAPCS_LOCAL
;
5845 else if (user_convention
&& user_pcs
!= arm_pcs_default
)
5846 sorry ("PCS variant");
5848 /* For everything else we use the target's default. */
5849 return arm_pcs_default
;
5854 aapcs_vfp_cum_init (CUMULATIVE_ARGS
*pcum ATTRIBUTE_UNUSED
,
5855 const_tree fntype ATTRIBUTE_UNUSED
,
5856 rtx libcall ATTRIBUTE_UNUSED
,
5857 const_tree fndecl ATTRIBUTE_UNUSED
)
5859 /* Record the unallocated VFP registers. */
5860 pcum
->aapcs_vfp_regs_free
= (1 << NUM_VFP_ARG_REGS
) - 1;
5861 pcum
->aapcs_vfp_reg_alloc
= 0;
5864 /* Walk down the type tree of TYPE counting consecutive base elements.
5865 If *MODEP is VOIDmode, then set it to the first valid floating point
5866 type. If a non-floating point type is found, or if a floating point
5867 type that doesn't match a non-VOIDmode *MODEP is found, then return -1,
5868 otherwise return the count in the sub-tree. */
5870 aapcs_vfp_sub_candidate (const_tree type
, machine_mode
*modep
)
5875 switch (TREE_CODE (type
))
5878 mode
= TYPE_MODE (type
);
5879 if (mode
!= DFmode
&& mode
!= SFmode
&& mode
!= HFmode
)
5882 if (*modep
== VOIDmode
)
5891 mode
= TYPE_MODE (TREE_TYPE (type
));
5892 if (mode
!= DFmode
&& mode
!= SFmode
)
5895 if (*modep
== VOIDmode
)
5904 /* Use V2SImode and V4SImode as representatives of all 64-bit
5905 and 128-bit vector types, whether or not those modes are
5906 supported with the present options. */
5907 size
= int_size_in_bytes (type
);
5920 if (*modep
== VOIDmode
)
5923 /* Vector modes are considered to be opaque: two vectors are
5924 equivalent for the purposes of being homogeneous aggregates
5925 if they are the same size. */
5934 tree index
= TYPE_DOMAIN (type
);
5936 /* Can't handle incomplete types nor sizes that are not
5938 if (!COMPLETE_TYPE_P (type
)
5939 || TREE_CODE (TYPE_SIZE (type
)) != INTEGER_CST
)
5942 count
= aapcs_vfp_sub_candidate (TREE_TYPE (type
), modep
);
5945 || !TYPE_MAX_VALUE (index
)
5946 || !tree_fits_uhwi_p (TYPE_MAX_VALUE (index
))
5947 || !TYPE_MIN_VALUE (index
)
5948 || !tree_fits_uhwi_p (TYPE_MIN_VALUE (index
))
5952 count
*= (1 + tree_to_uhwi (TYPE_MAX_VALUE (index
))
5953 - tree_to_uhwi (TYPE_MIN_VALUE (index
)));
5955 /* There must be no padding. */
5956 if (wi::to_wide (TYPE_SIZE (type
))
5957 != count
* GET_MODE_BITSIZE (*modep
))
5969 /* Can't handle incomplete types nor sizes that are not
5971 if (!COMPLETE_TYPE_P (type
)
5972 || TREE_CODE (TYPE_SIZE (type
)) != INTEGER_CST
)
5975 for (field
= TYPE_FIELDS (type
); field
; field
= DECL_CHAIN (field
))
5977 if (TREE_CODE (field
) != FIELD_DECL
)
5980 sub_count
= aapcs_vfp_sub_candidate (TREE_TYPE (field
), modep
);
5986 /* There must be no padding. */
5987 if (wi::to_wide (TYPE_SIZE (type
))
5988 != count
* GET_MODE_BITSIZE (*modep
))
5995 case QUAL_UNION_TYPE
:
5997 /* These aren't very interesting except in a degenerate case. */
6002 /* Can't handle incomplete types nor sizes that are not
6004 if (!COMPLETE_TYPE_P (type
)
6005 || TREE_CODE (TYPE_SIZE (type
)) != INTEGER_CST
)
6008 for (field
= TYPE_FIELDS (type
); field
; field
= DECL_CHAIN (field
))
6010 if (TREE_CODE (field
) != FIELD_DECL
)
6013 sub_count
= aapcs_vfp_sub_candidate (TREE_TYPE (field
), modep
);
6016 count
= count
> sub_count
? count
: sub_count
;
6019 /* There must be no padding. */
6020 if (wi::to_wide (TYPE_SIZE (type
))
6021 != count
* GET_MODE_BITSIZE (*modep
))
6034 /* Return true if PCS_VARIANT should use VFP registers. */
6036 use_vfp_abi (enum arm_pcs pcs_variant
, bool is_double
)
6038 if (pcs_variant
== ARM_PCS_AAPCS_VFP
)
6040 static bool seen_thumb1_vfp
= false;
6042 if (TARGET_THUMB1
&& !seen_thumb1_vfp
)
6044 sorry ("Thumb-1 hard-float VFP ABI");
6045 /* sorry() is not immediately fatal, so only display this once. */
6046 seen_thumb1_vfp
= true;
6052 if (pcs_variant
!= ARM_PCS_AAPCS_LOCAL
)
6055 return (TARGET_32BIT
&& TARGET_HARD_FLOAT
&&
6056 (TARGET_VFP_DOUBLE
|| !is_double
));
6059 /* Return true if an argument whose type is TYPE, or mode is MODE, is
6060 suitable for passing or returning in VFP registers for the PCS
6061 variant selected. If it is, then *BASE_MODE is updated to contain
6062 a machine mode describing each element of the argument's type and
6063 *COUNT to hold the number of such elements. */
6065 aapcs_vfp_is_call_or_return_candidate (enum arm_pcs pcs_variant
,
6066 machine_mode mode
, const_tree type
,
6067 machine_mode
*base_mode
, int *count
)
6069 machine_mode new_mode
= VOIDmode
;
6071 /* If we have the type information, prefer that to working things
6072 out from the mode. */
6075 int ag_count
= aapcs_vfp_sub_candidate (type
, &new_mode
);
6077 if (ag_count
> 0 && ag_count
<= 4)
6082 else if (GET_MODE_CLASS (mode
) == MODE_FLOAT
6083 || GET_MODE_CLASS (mode
) == MODE_VECTOR_INT
6084 || GET_MODE_CLASS (mode
) == MODE_VECTOR_FLOAT
)
6089 else if (GET_MODE_CLASS (mode
) == MODE_COMPLEX_FLOAT
)
6092 new_mode
= (mode
== DCmode
? DFmode
: SFmode
);
6098 if (!use_vfp_abi (pcs_variant
, ARM_NUM_REGS (new_mode
) > 1))
6101 *base_mode
= new_mode
;
6106 aapcs_vfp_is_return_candidate (enum arm_pcs pcs_variant
,
6107 machine_mode mode
, const_tree type
)
6109 int count ATTRIBUTE_UNUSED
;
6110 machine_mode ag_mode ATTRIBUTE_UNUSED
;
6112 if (!use_vfp_abi (pcs_variant
, false))
6114 return aapcs_vfp_is_call_or_return_candidate (pcs_variant
, mode
, type
,
6119 aapcs_vfp_is_call_candidate (CUMULATIVE_ARGS
*pcum
, machine_mode mode
,
6122 if (!use_vfp_abi (pcum
->pcs_variant
, false))
6125 return aapcs_vfp_is_call_or_return_candidate (pcum
->pcs_variant
, mode
, type
,
6126 &pcum
->aapcs_vfp_rmode
,
6127 &pcum
->aapcs_vfp_rcount
);
6130 /* Implement the allocate field in aapcs_cp_arg_layout. See the comment there
6131 for the behaviour of this function. */
6134 aapcs_vfp_allocate (CUMULATIVE_ARGS
*pcum
, machine_mode mode
,
6135 const_tree type ATTRIBUTE_UNUSED
)
6138 = MAX (GET_MODE_SIZE (pcum
->aapcs_vfp_rmode
), GET_MODE_SIZE (SFmode
));
6139 int shift
= rmode_size
/ GET_MODE_SIZE (SFmode
);
6140 unsigned mask
= (1 << (shift
* pcum
->aapcs_vfp_rcount
)) - 1;
6143 for (regno
= 0; regno
< NUM_VFP_ARG_REGS
; regno
+= shift
)
6144 if (((pcum
->aapcs_vfp_regs_free
>> regno
) & mask
) == mask
)
6146 pcum
->aapcs_vfp_reg_alloc
= mask
<< regno
;
6148 || (mode
== TImode
&& ! TARGET_NEON
)
6149 || ! arm_hard_regno_mode_ok (FIRST_VFP_REGNUM
+ regno
, mode
))
6152 int rcount
= pcum
->aapcs_vfp_rcount
;
6154 machine_mode rmode
= pcum
->aapcs_vfp_rmode
;
6158 /* Avoid using unsupported vector modes. */
6159 if (rmode
== V2SImode
)
6161 else if (rmode
== V4SImode
)
6168 par
= gen_rtx_PARALLEL (mode
, rtvec_alloc (rcount
));
6169 for (i
= 0; i
< rcount
; i
++)
6171 rtx tmp
= gen_rtx_REG (rmode
,
6172 FIRST_VFP_REGNUM
+ regno
+ i
* rshift
);
6173 tmp
= gen_rtx_EXPR_LIST
6175 GEN_INT (i
* GET_MODE_SIZE (rmode
)));
6176 XVECEXP (par
, 0, i
) = tmp
;
6179 pcum
->aapcs_reg
= par
;
6182 pcum
->aapcs_reg
= gen_rtx_REG (mode
, FIRST_VFP_REGNUM
+ regno
);
6188 /* Implement the allocate_return_reg field in aapcs_cp_arg_layout. See the
6189 comment there for the behaviour of this function. */
6192 aapcs_vfp_allocate_return_reg (enum arm_pcs pcs_variant ATTRIBUTE_UNUSED
,
6194 const_tree type ATTRIBUTE_UNUSED
)
6196 if (!use_vfp_abi (pcs_variant
, false))
6200 || (GET_MODE_CLASS (mode
) == MODE_INT
6201 && GET_MODE_SIZE (mode
) >= GET_MODE_SIZE (TImode
)
6205 machine_mode ag_mode
;
6210 aapcs_vfp_is_call_or_return_candidate (pcs_variant
, mode
, type
,
6215 if (ag_mode
== V2SImode
)
6217 else if (ag_mode
== V4SImode
)
6223 shift
= GET_MODE_SIZE(ag_mode
) / GET_MODE_SIZE(SFmode
);
6224 par
= gen_rtx_PARALLEL (mode
, rtvec_alloc (count
));
6225 for (i
= 0; i
< count
; i
++)
6227 rtx tmp
= gen_rtx_REG (ag_mode
, FIRST_VFP_REGNUM
+ i
* shift
);
6228 tmp
= gen_rtx_EXPR_LIST (VOIDmode
, tmp
,
6229 GEN_INT (i
* GET_MODE_SIZE (ag_mode
)));
6230 XVECEXP (par
, 0, i
) = tmp
;
6236 return gen_rtx_REG (mode
, FIRST_VFP_REGNUM
);
6240 aapcs_vfp_advance (CUMULATIVE_ARGS
*pcum ATTRIBUTE_UNUSED
,
6241 machine_mode mode ATTRIBUTE_UNUSED
,
6242 const_tree type ATTRIBUTE_UNUSED
)
6244 pcum
->aapcs_vfp_regs_free
&= ~pcum
->aapcs_vfp_reg_alloc
;
6245 pcum
->aapcs_vfp_reg_alloc
= 0;
6249 #define AAPCS_CP(X) \
6251 aapcs_ ## X ## _cum_init, \
6252 aapcs_ ## X ## _is_call_candidate, \
6253 aapcs_ ## X ## _allocate, \
6254 aapcs_ ## X ## _is_return_candidate, \
6255 aapcs_ ## X ## _allocate_return_reg, \
6256 aapcs_ ## X ## _advance \
6259 /* Table of co-processors that can be used to pass arguments in
6260 registers. Idealy no arugment should be a candidate for more than
6261 one co-processor table entry, but the table is processed in order
6262 and stops after the first match. If that entry then fails to put
6263 the argument into a co-processor register, the argument will go on
6267 /* Initialize co-processor related state in CUMULATIVE_ARGS structure. */
6268 void (*cum_init
) (CUMULATIVE_ARGS
*, const_tree
, rtx
, const_tree
);
6270 /* Return true if an argument of mode MODE (or type TYPE if MODE is
6271 BLKmode) is a candidate for this co-processor's registers; this
6272 function should ignore any position-dependent state in
6273 CUMULATIVE_ARGS and only use call-type dependent information. */
6274 bool (*is_call_candidate
) (CUMULATIVE_ARGS
*, machine_mode
, const_tree
);
6276 /* Return true if the argument does get a co-processor register; it
6277 should set aapcs_reg to an RTX of the register allocated as is
6278 required for a return from FUNCTION_ARG. */
6279 bool (*allocate
) (CUMULATIVE_ARGS
*, machine_mode
, const_tree
);
6281 /* Return true if a result of mode MODE (or type TYPE if MODE is BLKmode) can
6282 be returned in this co-processor's registers. */
6283 bool (*is_return_candidate
) (enum arm_pcs
, machine_mode
, const_tree
);
6285 /* Allocate and return an RTX element to hold the return type of a call. This
6286 routine must not fail and will only be called if is_return_candidate
6287 returned true with the same parameters. */
6288 rtx (*allocate_return_reg
) (enum arm_pcs
, machine_mode
, const_tree
);
6290 /* Finish processing this argument and prepare to start processing
6292 void (*advance
) (CUMULATIVE_ARGS
*, machine_mode
, const_tree
);
6293 } aapcs_cp_arg_layout
[ARM_NUM_COPROC_SLOTS
] =
6301 aapcs_select_call_coproc (CUMULATIVE_ARGS
*pcum
, machine_mode mode
,
6306 for (i
= 0; i
< ARM_NUM_COPROC_SLOTS
; i
++)
6307 if (aapcs_cp_arg_layout
[i
].is_call_candidate (pcum
, mode
, type
))
6314 aapcs_select_return_coproc (const_tree type
, const_tree fntype
)
6316 /* We aren't passed a decl, so we can't check that a call is local.
6317 However, it isn't clear that that would be a win anyway, since it
6318 might limit some tail-calling opportunities. */
6319 enum arm_pcs pcs_variant
;
6323 const_tree fndecl
= NULL_TREE
;
6325 if (TREE_CODE (fntype
) == FUNCTION_DECL
)
6328 fntype
= TREE_TYPE (fntype
);
6331 pcs_variant
= arm_get_pcs_model (fntype
, fndecl
);
6334 pcs_variant
= arm_pcs_default
;
6336 if (pcs_variant
!= ARM_PCS_AAPCS
)
6340 for (i
= 0; i
< ARM_NUM_COPROC_SLOTS
; i
++)
6341 if (aapcs_cp_arg_layout
[i
].is_return_candidate (pcs_variant
,
6350 aapcs_allocate_return_reg (machine_mode mode
, const_tree type
,
6353 /* We aren't passed a decl, so we can't check that a call is local.
6354 However, it isn't clear that that would be a win anyway, since it
6355 might limit some tail-calling opportunities. */
6356 enum arm_pcs pcs_variant
;
6357 int unsignedp ATTRIBUTE_UNUSED
;
6361 const_tree fndecl
= NULL_TREE
;
6363 if (TREE_CODE (fntype
) == FUNCTION_DECL
)
6366 fntype
= TREE_TYPE (fntype
);
6369 pcs_variant
= arm_get_pcs_model (fntype
, fndecl
);
6372 pcs_variant
= arm_pcs_default
;
6374 /* Promote integer types. */
6375 if (type
&& INTEGRAL_TYPE_P (type
))
6376 mode
= arm_promote_function_mode (type
, mode
, &unsignedp
, fntype
, 1);
6378 if (pcs_variant
!= ARM_PCS_AAPCS
)
6382 for (i
= 0; i
< ARM_NUM_COPROC_SLOTS
; i
++)
6383 if (aapcs_cp_arg_layout
[i
].is_return_candidate (pcs_variant
, mode
,
6385 return aapcs_cp_arg_layout
[i
].allocate_return_reg (pcs_variant
,
6389 /* Promotes small structs returned in a register to full-word size
6390 for big-endian AAPCS. */
6391 if (type
&& arm_return_in_msb (type
))
6393 HOST_WIDE_INT size
= int_size_in_bytes (type
);
6394 if (size
% UNITS_PER_WORD
!= 0)
6396 size
+= UNITS_PER_WORD
- size
% UNITS_PER_WORD
;
6397 mode
= int_mode_for_size (size
* BITS_PER_UNIT
, 0).require ();
6401 return gen_rtx_REG (mode
, R0_REGNUM
);
6405 aapcs_libcall_value (machine_mode mode
)
6407 if (BYTES_BIG_ENDIAN
&& ALL_FIXED_POINT_MODE_P (mode
)
6408 && GET_MODE_SIZE (mode
) <= 4)
6411 return aapcs_allocate_return_reg (mode
, NULL_TREE
, NULL_TREE
);
6414 /* Lay out a function argument using the AAPCS rules. The rule
6415 numbers referred to here are those in the AAPCS. */
6417 aapcs_layout_arg (CUMULATIVE_ARGS
*pcum
, machine_mode mode
,
6418 const_tree type
, bool named
)
6423 /* We only need to do this once per argument. */
6424 if (pcum
->aapcs_arg_processed
)
6427 pcum
->aapcs_arg_processed
= true;
6429 /* Special case: if named is false then we are handling an incoming
6430 anonymous argument which is on the stack. */
6434 /* Is this a potential co-processor register candidate? */
6435 if (pcum
->pcs_variant
!= ARM_PCS_AAPCS
)
6437 int slot
= aapcs_select_call_coproc (pcum
, mode
, type
);
6438 pcum
->aapcs_cprc_slot
= slot
;
6440 /* We don't have to apply any of the rules from part B of the
6441 preparation phase, these are handled elsewhere in the
6446 /* A Co-processor register candidate goes either in its own
6447 class of registers or on the stack. */
6448 if (!pcum
->aapcs_cprc_failed
[slot
])
6450 /* C1.cp - Try to allocate the argument to co-processor
6452 if (aapcs_cp_arg_layout
[slot
].allocate (pcum
, mode
, type
))
6455 /* C2.cp - Put the argument on the stack and note that we
6456 can't assign any more candidates in this slot. We also
6457 need to note that we have allocated stack space, so that
6458 we won't later try to split a non-cprc candidate between
6459 core registers and the stack. */
6460 pcum
->aapcs_cprc_failed
[slot
] = true;
6461 pcum
->can_split
= false;
6464 /* We didn't get a register, so this argument goes on the
6466 gcc_assert (pcum
->can_split
== false);
6471 /* C3 - For double-word aligned arguments, round the NCRN up to the
6472 next even number. */
6473 ncrn
= pcum
->aapcs_ncrn
;
6476 int res
= arm_needs_doubleword_align (mode
, type
);
6477 /* Only warn during RTL expansion of call stmts, otherwise we would
6478 warn e.g. during gimplification even on functions that will be
6479 always inlined, and we'd warn multiple times. Don't warn when
6480 called in expand_function_start either, as we warn instead in
6481 arm_function_arg_boundary in that case. */
6482 if (res
< 0 && warn_psabi
&& currently_expanding_gimple_stmt
)
6483 inform (input_location
, "parameter passing for argument of type "
6484 "%qT changed in GCC 7.1", type
);
6489 nregs
= ARM_NUM_REGS2(mode
, type
);
6491 /* Sigh, this test should really assert that nregs > 0, but a GCC
6492 extension allows empty structs and then gives them empty size; it
6493 then allows such a structure to be passed by value. For some of
6494 the code below we have to pretend that such an argument has
6495 non-zero size so that we 'locate' it correctly either in
6496 registers or on the stack. */
6497 gcc_assert (nregs
>= 0);
6499 nregs2
= nregs
? nregs
: 1;
6501 /* C4 - Argument fits entirely in core registers. */
6502 if (ncrn
+ nregs2
<= NUM_ARG_REGS
)
6504 pcum
->aapcs_reg
= gen_rtx_REG (mode
, ncrn
);
6505 pcum
->aapcs_next_ncrn
= ncrn
+ nregs
;
6509 /* C5 - Some core registers left and there are no arguments already
6510 on the stack: split this argument between the remaining core
6511 registers and the stack. */
6512 if (ncrn
< NUM_ARG_REGS
&& pcum
->can_split
)
6514 pcum
->aapcs_reg
= gen_rtx_REG (mode
, ncrn
);
6515 pcum
->aapcs_next_ncrn
= NUM_ARG_REGS
;
6516 pcum
->aapcs_partial
= (NUM_ARG_REGS
- ncrn
) * UNITS_PER_WORD
;
6520 /* C6 - NCRN is set to 4. */
6521 pcum
->aapcs_next_ncrn
= NUM_ARG_REGS
;
6523 /* C7,C8 - arugment goes on the stack. We have nothing to do here. */
6527 /* Initialize a variable CUM of type CUMULATIVE_ARGS
6528 for a call to a function whose data type is FNTYPE.
6529 For a library call, FNTYPE is NULL. */
6531 arm_init_cumulative_args (CUMULATIVE_ARGS
*pcum
, tree fntype
,
6533 tree fndecl ATTRIBUTE_UNUSED
)
6535 /* Long call handling. */
6537 pcum
->pcs_variant
= arm_get_pcs_model (fntype
, fndecl
);
6539 pcum
->pcs_variant
= arm_pcs_default
;
6541 if (pcum
->pcs_variant
<= ARM_PCS_AAPCS_LOCAL
)
6543 if (arm_libcall_uses_aapcs_base (libname
))
6544 pcum
->pcs_variant
= ARM_PCS_AAPCS
;
6546 pcum
->aapcs_ncrn
= pcum
->aapcs_next_ncrn
= 0;
6547 pcum
->aapcs_reg
= NULL_RTX
;
6548 pcum
->aapcs_partial
= 0;
6549 pcum
->aapcs_arg_processed
= false;
6550 pcum
->aapcs_cprc_slot
= -1;
6551 pcum
->can_split
= true;
6553 if (pcum
->pcs_variant
!= ARM_PCS_AAPCS
)
6557 for (i
= 0; i
< ARM_NUM_COPROC_SLOTS
; i
++)
6559 pcum
->aapcs_cprc_failed
[i
] = false;
6560 aapcs_cp_arg_layout
[i
].cum_init (pcum
, fntype
, libname
, fndecl
);
6568 /* On the ARM, the offset starts at 0. */
6570 pcum
->iwmmxt_nregs
= 0;
6571 pcum
->can_split
= true;
6573 /* Varargs vectors are treated the same as long long.
6574 named_count avoids having to change the way arm handles 'named' */
6575 pcum
->named_count
= 0;
6578 if (TARGET_REALLY_IWMMXT
&& fntype
)
6582 for (fn_arg
= TYPE_ARG_TYPES (fntype
);
6584 fn_arg
= TREE_CHAIN (fn_arg
))
6585 pcum
->named_count
+= 1;
6587 if (! pcum
->named_count
)
6588 pcum
->named_count
= INT_MAX
;
6592 /* Return 1 if double word alignment is required for argument passing.
6593 Return -1 if double word alignment used to be required for argument
6594 passing before PR77728 ABI fix, but is not required anymore.
6595 Return 0 if double word alignment is not required and wasn't requried
6598 arm_needs_doubleword_align (machine_mode mode
, const_tree type
)
6601 return GET_MODE_ALIGNMENT (mode
) > PARM_BOUNDARY
;
6603 /* Scalar and vector types: Use natural alignment, i.e. of base type. */
6604 if (!AGGREGATE_TYPE_P (type
))
6605 return TYPE_ALIGN (TYPE_MAIN_VARIANT (type
)) > PARM_BOUNDARY
;
6607 /* Array types: Use member alignment of element type. */
6608 if (TREE_CODE (type
) == ARRAY_TYPE
)
6609 return TYPE_ALIGN (TREE_TYPE (type
)) > PARM_BOUNDARY
;
6612 /* Record/aggregate types: Use greatest member alignment of any member. */
6613 for (tree field
= TYPE_FIELDS (type
); field
; field
= DECL_CHAIN (field
))
6614 if (DECL_ALIGN (field
) > PARM_BOUNDARY
)
6616 if (TREE_CODE (field
) == FIELD_DECL
)
6619 /* Before PR77728 fix, we were incorrectly considering also
6620 other aggregate fields, like VAR_DECLs, TYPE_DECLs etc.
6621 Make sure we can warn about that with -Wpsabi. */
6629 /* Determine where to put an argument to a function.
6630 Value is zero to push the argument on the stack,
6631 or a hard register in which to store the argument.
6633 MODE is the argument's machine mode.
6634 TYPE is the data type of the argument (as a tree).
6635 This is null for libcalls where that information may
6637 CUM is a variable of type CUMULATIVE_ARGS which gives info about
6638 the preceding args and about the function being called.
6639 NAMED is nonzero if this argument is a named parameter
6640 (otherwise it is an extra parameter matching an ellipsis).
6642 On the ARM, normally the first 16 bytes are passed in registers r0-r3; all
6643 other arguments are passed on the stack. If (NAMED == 0) (which happens
6644 only in assign_parms, since TARGET_SETUP_INCOMING_VARARGS is
6645 defined), say it is passed in the stack (function_prologue will
6646 indeed make it pass in the stack if necessary). */
6649 arm_function_arg (cumulative_args_t pcum_v
, machine_mode mode
,
6650 const_tree type
, bool named
)
6652 CUMULATIVE_ARGS
*pcum
= get_cumulative_args (pcum_v
);
6655 /* Handle the special case quickly. Pick an arbitrary value for op2 of
6656 a call insn (op3 of a call_value insn). */
6657 if (mode
== VOIDmode
)
6660 if (pcum
->pcs_variant
<= ARM_PCS_AAPCS_LOCAL
)
6662 aapcs_layout_arg (pcum
, mode
, type
, named
);
6663 return pcum
->aapcs_reg
;
6666 /* Varargs vectors are treated the same as long long.
6667 named_count avoids having to change the way arm handles 'named' */
6668 if (TARGET_IWMMXT_ABI
6669 && arm_vector_mode_supported_p (mode
)
6670 && pcum
->named_count
> pcum
->nargs
+ 1)
6672 if (pcum
->iwmmxt_nregs
<= 9)
6673 return gen_rtx_REG (mode
, pcum
->iwmmxt_nregs
+ FIRST_IWMMXT_REGNUM
);
6676 pcum
->can_split
= false;
6681 /* Put doubleword aligned quantities in even register pairs. */
6682 if ((pcum
->nregs
& 1) && ARM_DOUBLEWORD_ALIGN
)
6684 int res
= arm_needs_doubleword_align (mode
, type
);
6685 if (res
< 0 && warn_psabi
)
6686 inform (input_location
, "parameter passing for argument of type "
6687 "%qT changed in GCC 7.1", type
);
6692 /* Only allow splitting an arg between regs and memory if all preceding
6693 args were allocated to regs. For args passed by reference we only count
6694 the reference pointer. */
6695 if (pcum
->can_split
)
6698 nregs
= ARM_NUM_REGS2 (mode
, type
);
6700 if (!named
|| pcum
->nregs
+ nregs
> NUM_ARG_REGS
)
6703 return gen_rtx_REG (mode
, pcum
->nregs
);
6707 arm_function_arg_boundary (machine_mode mode
, const_tree type
)
6709 if (!ARM_DOUBLEWORD_ALIGN
)
6710 return PARM_BOUNDARY
;
6712 int res
= arm_needs_doubleword_align (mode
, type
);
6713 if (res
< 0 && warn_psabi
)
6714 inform (input_location
, "parameter passing for argument of type %qT "
6715 "changed in GCC 7.1", type
);
6717 return res
> 0 ? DOUBLEWORD_ALIGNMENT
: PARM_BOUNDARY
;
6721 arm_arg_partial_bytes (cumulative_args_t pcum_v
, machine_mode mode
,
6722 tree type
, bool named
)
6724 CUMULATIVE_ARGS
*pcum
= get_cumulative_args (pcum_v
);
6725 int nregs
= pcum
->nregs
;
6727 if (pcum
->pcs_variant
<= ARM_PCS_AAPCS_LOCAL
)
6729 aapcs_layout_arg (pcum
, mode
, type
, named
);
6730 return pcum
->aapcs_partial
;
6733 if (TARGET_IWMMXT_ABI
&& arm_vector_mode_supported_p (mode
))
6736 if (NUM_ARG_REGS
> nregs
6737 && (NUM_ARG_REGS
< nregs
+ ARM_NUM_REGS2 (mode
, type
))
6739 return (NUM_ARG_REGS
- nregs
) * UNITS_PER_WORD
;
6744 /* Update the data in PCUM to advance over an argument
6745 of mode MODE and data type TYPE.
6746 (TYPE is null for libcalls where that information may not be available.) */
6749 arm_function_arg_advance (cumulative_args_t pcum_v
, machine_mode mode
,
6750 const_tree type
, bool named
)
6752 CUMULATIVE_ARGS
*pcum
= get_cumulative_args (pcum_v
);
6754 if (pcum
->pcs_variant
<= ARM_PCS_AAPCS_LOCAL
)
6756 aapcs_layout_arg (pcum
, mode
, type
, named
);
6758 if (pcum
->aapcs_cprc_slot
>= 0)
6760 aapcs_cp_arg_layout
[pcum
->aapcs_cprc_slot
].advance (pcum
, mode
,
6762 pcum
->aapcs_cprc_slot
= -1;
6765 /* Generic stuff. */
6766 pcum
->aapcs_arg_processed
= false;
6767 pcum
->aapcs_ncrn
= pcum
->aapcs_next_ncrn
;
6768 pcum
->aapcs_reg
= NULL_RTX
;
6769 pcum
->aapcs_partial
= 0;
6774 if (arm_vector_mode_supported_p (mode
)
6775 && pcum
->named_count
> pcum
->nargs
6776 && TARGET_IWMMXT_ABI
)
6777 pcum
->iwmmxt_nregs
+= 1;
6779 pcum
->nregs
+= ARM_NUM_REGS2 (mode
, type
);
6783 /* Variable sized types are passed by reference. This is a GCC
6784 extension to the ARM ABI. */
6787 arm_pass_by_reference (cumulative_args_t cum ATTRIBUTE_UNUSED
,
6788 machine_mode mode ATTRIBUTE_UNUSED
,
6789 const_tree type
, bool named ATTRIBUTE_UNUSED
)
6791 return type
&& TREE_CODE (TYPE_SIZE (type
)) != INTEGER_CST
;
6794 /* Encode the current state of the #pragma [no_]long_calls. */
6797 OFF
, /* No #pragma [no_]long_calls is in effect. */
6798 LONG
, /* #pragma long_calls is in effect. */
6799 SHORT
/* #pragma no_long_calls is in effect. */
6802 static arm_pragma_enum arm_pragma_long_calls
= OFF
;
6805 arm_pr_long_calls (struct cpp_reader
* pfile ATTRIBUTE_UNUSED
)
6807 arm_pragma_long_calls
= LONG
;
6811 arm_pr_no_long_calls (struct cpp_reader
* pfile ATTRIBUTE_UNUSED
)
6813 arm_pragma_long_calls
= SHORT
;
6817 arm_pr_long_calls_off (struct cpp_reader
* pfile ATTRIBUTE_UNUSED
)
6819 arm_pragma_long_calls
= OFF
;
6822 /* Handle an attribute requiring a FUNCTION_DECL;
6823 arguments as in struct attribute_spec.handler. */
6825 arm_handle_fndecl_attribute (tree
*node
, tree name
, tree args ATTRIBUTE_UNUSED
,
6826 int flags ATTRIBUTE_UNUSED
, bool *no_add_attrs
)
6828 if (TREE_CODE (*node
) != FUNCTION_DECL
)
6830 warning (OPT_Wattributes
, "%qE attribute only applies to functions",
6832 *no_add_attrs
= true;
6838 /* Handle an "interrupt" or "isr" attribute;
6839 arguments as in struct attribute_spec.handler. */
6841 arm_handle_isr_attribute (tree
*node
, tree name
, tree args
, int flags
,
6846 if (TREE_CODE (*node
) != FUNCTION_DECL
)
6848 warning (OPT_Wattributes
, "%qE attribute only applies to functions",
6850 *no_add_attrs
= true;
6852 /* FIXME: the argument if any is checked for type attributes;
6853 should it be checked for decl ones? */
6857 if (TREE_CODE (*node
) == FUNCTION_TYPE
6858 || TREE_CODE (*node
) == METHOD_TYPE
)
6860 if (arm_isr_value (args
) == ARM_FT_UNKNOWN
)
6862 warning (OPT_Wattributes
, "%qE attribute ignored",
6864 *no_add_attrs
= true;
6867 else if (TREE_CODE (*node
) == POINTER_TYPE
6868 && (TREE_CODE (TREE_TYPE (*node
)) == FUNCTION_TYPE
6869 || TREE_CODE (TREE_TYPE (*node
)) == METHOD_TYPE
)
6870 && arm_isr_value (args
) != ARM_FT_UNKNOWN
)
6872 *node
= build_variant_type_copy (*node
);
6873 TREE_TYPE (*node
) = build_type_attribute_variant
6875 tree_cons (name
, args
, TYPE_ATTRIBUTES (TREE_TYPE (*node
))));
6876 *no_add_attrs
= true;
6880 /* Possibly pass this attribute on from the type to a decl. */
6881 if (flags
& ((int) ATTR_FLAG_DECL_NEXT
6882 | (int) ATTR_FLAG_FUNCTION_NEXT
6883 | (int) ATTR_FLAG_ARRAY_NEXT
))
6885 *no_add_attrs
= true;
6886 return tree_cons (name
, args
, NULL_TREE
);
6890 warning (OPT_Wattributes
, "%qE attribute ignored",
6899 /* Handle a "pcs" attribute; arguments as in struct
6900 attribute_spec.handler. */
6902 arm_handle_pcs_attribute (tree
*node ATTRIBUTE_UNUSED
, tree name
, tree args
,
6903 int flags ATTRIBUTE_UNUSED
, bool *no_add_attrs
)
6905 if (arm_pcs_from_attribute (args
) == ARM_PCS_UNKNOWN
)
6907 warning (OPT_Wattributes
, "%qE attribute ignored", name
);
6908 *no_add_attrs
= true;
6913 #if TARGET_DLLIMPORT_DECL_ATTRIBUTES
6914 /* Handle the "notshared" attribute. This attribute is another way of
6915 requesting hidden visibility. ARM's compiler supports
6916 "__declspec(notshared)"; we support the same thing via an
6920 arm_handle_notshared_attribute (tree
*node
,
6921 tree name ATTRIBUTE_UNUSED
,
6922 tree args ATTRIBUTE_UNUSED
,
6923 int flags ATTRIBUTE_UNUSED
,
6926 tree decl
= TYPE_NAME (*node
);
6930 DECL_VISIBILITY (decl
) = VISIBILITY_HIDDEN
;
6931 DECL_VISIBILITY_SPECIFIED (decl
) = 1;
6932 *no_add_attrs
= false;
6938 /* This function returns true if a function with declaration FNDECL and type
6939 FNTYPE uses the stack to pass arguments or return variables and false
6940 otherwise. This is used for functions with the attributes
6941 'cmse_nonsecure_call' or 'cmse_nonsecure_entry' and this function will issue
6942 diagnostic messages if the stack is used. NAME is the name of the attribute
6946 cmse_func_args_or_return_in_stack (tree fndecl
, tree name
, tree fntype
)
6948 function_args_iterator args_iter
;
6949 CUMULATIVE_ARGS args_so_far_v
;
6950 cumulative_args_t args_so_far
;
6951 bool first_param
= true;
6952 tree arg_type
, prev_arg_type
= NULL_TREE
, ret_type
;
6954 /* Error out if any argument is passed on the stack. */
6955 arm_init_cumulative_args (&args_so_far_v
, fntype
, NULL_RTX
, fndecl
);
6956 args_so_far
= pack_cumulative_args (&args_so_far_v
);
6957 FOREACH_FUNCTION_ARGS (fntype
, arg_type
, args_iter
)
6960 machine_mode arg_mode
= TYPE_MODE (arg_type
);
6962 prev_arg_type
= arg_type
;
6963 if (VOID_TYPE_P (arg_type
))
6967 arm_function_arg_advance (args_so_far
, arg_mode
, arg_type
, true);
6968 arg_rtx
= arm_function_arg (args_so_far
, arg_mode
, arg_type
, true);
6970 || arm_arg_partial_bytes (args_so_far
, arg_mode
, arg_type
, true))
6972 error ("%qE attribute not available to functions with arguments "
6973 "passed on the stack", name
);
6976 first_param
= false;
6979 /* Error out for variadic functions since we cannot control how many
6980 arguments will be passed and thus stack could be used. stdarg_p () is not
6981 used for the checking to avoid browsing arguments twice. */
6982 if (prev_arg_type
!= NULL_TREE
&& !VOID_TYPE_P (prev_arg_type
))
6984 error ("%qE attribute not available to functions with variable number "
6985 "of arguments", name
);
6989 /* Error out if return value is passed on the stack. */
6990 ret_type
= TREE_TYPE (fntype
);
6991 if (arm_return_in_memory (ret_type
, fntype
))
6993 error ("%qE attribute not available to functions that return value on "
7000 /* Called upon detection of the use of the cmse_nonsecure_entry attribute, this
7001 function will check whether the attribute is allowed here and will add the
7002 attribute to the function declaration tree or otherwise issue a warning. */
7005 arm_handle_cmse_nonsecure_entry (tree
*node
, tree name
,
7014 *no_add_attrs
= true;
7015 warning (OPT_Wattributes
, "%qE attribute ignored without -mcmse option.",
7020 /* Ignore attribute for function types. */
7021 if (TREE_CODE (*node
) != FUNCTION_DECL
)
7023 warning (OPT_Wattributes
, "%qE attribute only applies to functions",
7025 *no_add_attrs
= true;
7031 /* Warn for static linkage functions. */
7032 if (!TREE_PUBLIC (fndecl
))
7034 warning (OPT_Wattributes
, "%qE attribute has no effect on functions "
7035 "with static linkage", name
);
7036 *no_add_attrs
= true;
7040 *no_add_attrs
|= cmse_func_args_or_return_in_stack (fndecl
, name
,
7041 TREE_TYPE (fndecl
));
7046 /* Called upon detection of the use of the cmse_nonsecure_call attribute, this
7047 function will check whether the attribute is allowed here and will add the
7048 attribute to the function type tree or otherwise issue a diagnostic. The
7049 reason we check this at declaration time is to only allow the use of the
7050 attribute with declarations of function pointers and not function
7051 declarations. This function checks NODE is of the expected type and issues
7052 diagnostics otherwise using NAME. If it is not of the expected type
7053 *NO_ADD_ATTRS will be set to true. */
7056 arm_handle_cmse_nonsecure_call (tree
*node
, tree name
,
7061 tree decl
= NULL_TREE
, fntype
= NULL_TREE
;
7066 *no_add_attrs
= true;
7067 warning (OPT_Wattributes
, "%qE attribute ignored without -mcmse option.",
7072 if (TREE_CODE (*node
) == VAR_DECL
|| TREE_CODE (*node
) == TYPE_DECL
)
7075 fntype
= TREE_TYPE (decl
);
7078 while (fntype
!= NULL_TREE
&& TREE_CODE (fntype
) == POINTER_TYPE
)
7079 fntype
= TREE_TYPE (fntype
);
7081 if (!decl
|| TREE_CODE (fntype
) != FUNCTION_TYPE
)
7083 warning (OPT_Wattributes
, "%qE attribute only applies to base type of a "
7084 "function pointer", name
);
7085 *no_add_attrs
= true;
7089 *no_add_attrs
|= cmse_func_args_or_return_in_stack (NULL
, name
, fntype
);
7094 /* Prevent trees being shared among function types with and without
7095 cmse_nonsecure_call attribute. */
7096 type
= TREE_TYPE (decl
);
7098 type
= build_distinct_type_copy (type
);
7099 TREE_TYPE (decl
) = type
;
7102 while (TREE_CODE (fntype
) != FUNCTION_TYPE
)
7105 fntype
= TREE_TYPE (fntype
);
7106 fntype
= build_distinct_type_copy (fntype
);
7107 TREE_TYPE (type
) = fntype
;
7110 /* Construct a type attribute and add it to the function type. */
7111 tree attrs
= tree_cons (get_identifier ("cmse_nonsecure_call"), NULL_TREE
,
7112 TYPE_ATTRIBUTES (fntype
));
7113 TYPE_ATTRIBUTES (fntype
) = attrs
;
7117 /* Return 0 if the attributes for two types are incompatible, 1 if they
7118 are compatible, and 2 if they are nearly compatible (which causes a
7119 warning to be generated). */
7121 arm_comp_type_attributes (const_tree type1
, const_tree type2
)
7125 /* Check for mismatch of non-default calling convention. */
7126 if (TREE_CODE (type1
) != FUNCTION_TYPE
)
7129 /* Check for mismatched call attributes. */
7130 l1
= lookup_attribute ("long_call", TYPE_ATTRIBUTES (type1
)) != NULL
;
7131 l2
= lookup_attribute ("long_call", TYPE_ATTRIBUTES (type2
)) != NULL
;
7132 s1
= lookup_attribute ("short_call", TYPE_ATTRIBUTES (type1
)) != NULL
;
7133 s2
= lookup_attribute ("short_call", TYPE_ATTRIBUTES (type2
)) != NULL
;
7135 /* Only bother to check if an attribute is defined. */
7136 if (l1
| l2
| s1
| s2
)
7138 /* If one type has an attribute, the other must have the same attribute. */
7139 if ((l1
!= l2
) || (s1
!= s2
))
7142 /* Disallow mixed attributes. */
7143 if ((l1
& s2
) || (l2
& s1
))
7147 /* Check for mismatched ISR attribute. */
7148 l1
= lookup_attribute ("isr", TYPE_ATTRIBUTES (type1
)) != NULL
;
7150 l1
= lookup_attribute ("interrupt", TYPE_ATTRIBUTES (type1
)) != NULL
;
7151 l2
= lookup_attribute ("isr", TYPE_ATTRIBUTES (type2
)) != NULL
;
7153 l1
= lookup_attribute ("interrupt", TYPE_ATTRIBUTES (type2
)) != NULL
;
7157 l1
= lookup_attribute ("cmse_nonsecure_call",
7158 TYPE_ATTRIBUTES (type1
)) != NULL
;
7159 l2
= lookup_attribute ("cmse_nonsecure_call",
7160 TYPE_ATTRIBUTES (type2
)) != NULL
;
7168 /* Assigns default attributes to newly defined type. This is used to
7169 set short_call/long_call attributes for function types of
7170 functions defined inside corresponding #pragma scopes. */
7172 arm_set_default_type_attributes (tree type
)
7174 /* Add __attribute__ ((long_call)) to all functions, when
7175 inside #pragma long_calls or __attribute__ ((short_call)),
7176 when inside #pragma no_long_calls. */
7177 if (TREE_CODE (type
) == FUNCTION_TYPE
|| TREE_CODE (type
) == METHOD_TYPE
)
7179 tree type_attr_list
, attr_name
;
7180 type_attr_list
= TYPE_ATTRIBUTES (type
);
7182 if (arm_pragma_long_calls
== LONG
)
7183 attr_name
= get_identifier ("long_call");
7184 else if (arm_pragma_long_calls
== SHORT
)
7185 attr_name
= get_identifier ("short_call");
7189 type_attr_list
= tree_cons (attr_name
, NULL_TREE
, type_attr_list
);
7190 TYPE_ATTRIBUTES (type
) = type_attr_list
;
7194 /* Return true if DECL is known to be linked into section SECTION. */
7197 arm_function_in_section_p (tree decl
, section
*section
)
7199 /* We can only be certain about the prevailing symbol definition. */
7200 if (!decl_binds_to_current_def_p (decl
))
7203 /* If DECL_SECTION_NAME is set, assume it is trustworthy. */
7204 if (!DECL_SECTION_NAME (decl
))
7206 /* Make sure that we will not create a unique section for DECL. */
7207 if (flag_function_sections
|| DECL_COMDAT_GROUP (decl
))
7211 return function_section (decl
) == section
;
7214 /* Return nonzero if a 32-bit "long_call" should be generated for
7215 a call from the current function to DECL. We generate a long_call
7218 a. has an __attribute__((long call))
7219 or b. is within the scope of a #pragma long_calls
7220 or c. the -mlong-calls command line switch has been specified
7222 However we do not generate a long call if the function:
7224 d. has an __attribute__ ((short_call))
7225 or e. is inside the scope of a #pragma no_long_calls
7226 or f. is defined in the same section as the current function. */
7229 arm_is_long_call_p (tree decl
)
7234 return TARGET_LONG_CALLS
;
7236 attrs
= TYPE_ATTRIBUTES (TREE_TYPE (decl
));
7237 if (lookup_attribute ("short_call", attrs
))
7240 /* For "f", be conservative, and only cater for cases in which the
7241 whole of the current function is placed in the same section. */
7242 if (!flag_reorder_blocks_and_partition
7243 && TREE_CODE (decl
) == FUNCTION_DECL
7244 && arm_function_in_section_p (decl
, current_function_section ()))
7247 if (lookup_attribute ("long_call", attrs
))
7250 return TARGET_LONG_CALLS
;
7253 /* Return nonzero if it is ok to make a tail-call to DECL. */
7255 arm_function_ok_for_sibcall (tree decl
, tree exp
)
7257 unsigned long func_type
;
7259 if (cfun
->machine
->sibcall_blocked
)
7262 /* Never tailcall something if we are generating code for Thumb-1. */
7266 /* The PIC register is live on entry to VxWorks PLT entries, so we
7267 must make the call before restoring the PIC register. */
7268 if (TARGET_VXWORKS_RTP
&& flag_pic
&& decl
&& !targetm
.binds_local_p (decl
))
7271 /* ??? Cannot tail-call to long calls with APCS frame and VFP, because IP
7272 may be used both as target of the call and base register for restoring
7273 the VFP registers */
7274 if (TARGET_APCS_FRAME
&& TARGET_ARM
7275 && TARGET_HARD_FLOAT
7276 && decl
&& arm_is_long_call_p (decl
))
7279 /* If we are interworking and the function is not declared static
7280 then we can't tail-call it unless we know that it exists in this
7281 compilation unit (since it might be a Thumb routine). */
7282 if (TARGET_INTERWORK
&& decl
&& TREE_PUBLIC (decl
)
7283 && !TREE_ASM_WRITTEN (decl
))
7286 func_type
= arm_current_func_type ();
7287 /* Never tailcall from an ISR routine - it needs a special exit sequence. */
7288 if (IS_INTERRUPT (func_type
))
7291 /* ARMv8-M non-secure entry functions need to return with bxns which is only
7292 generated for entry functions themselves. */
7293 if (IS_CMSE_ENTRY (arm_current_func_type ()))
7296 /* We do not allow ARMv8-M non-secure calls to be turned into sibling calls,
7297 this would complicate matters for later code generation. */
7298 if (TREE_CODE (exp
) == CALL_EXPR
)
7300 tree fntype
= TREE_TYPE (TREE_TYPE (CALL_EXPR_FN (exp
)));
7301 if (lookup_attribute ("cmse_nonsecure_call", TYPE_ATTRIBUTES (fntype
)))
7305 if (!VOID_TYPE_P (TREE_TYPE (DECL_RESULT (cfun
->decl
))))
7307 /* Check that the return value locations are the same. For
7308 example that we aren't returning a value from the sibling in
7309 a VFP register but then need to transfer it to a core
7312 tree decl_or_type
= decl
;
7314 /* If it is an indirect function pointer, get the function type. */
7316 decl_or_type
= TREE_TYPE (TREE_TYPE (CALL_EXPR_FN (exp
)));
7318 a
= arm_function_value (TREE_TYPE (exp
), decl_or_type
, false);
7319 b
= arm_function_value (TREE_TYPE (DECL_RESULT (cfun
->decl
)),
7321 if (!rtx_equal_p (a
, b
))
7325 /* Never tailcall if function may be called with a misaligned SP. */
7326 if (IS_STACKALIGN (func_type
))
7329 /* The AAPCS says that, on bare-metal, calls to unresolved weak
7330 references should become a NOP. Don't convert such calls into
7332 if (TARGET_AAPCS_BASED
7333 && arm_abi
== ARM_ABI_AAPCS
7335 && DECL_WEAK (decl
))
7338 /* We cannot do a tailcall for an indirect call by descriptor if all the
7339 argument registers are used because the only register left to load the
7340 address is IP and it will already contain the static chain. */
7341 if (!decl
&& CALL_EXPR_BY_DESCRIPTOR (exp
) && !flag_trampolines
)
7343 tree fntype
= TREE_TYPE (TREE_TYPE (CALL_EXPR_FN (exp
)));
7344 CUMULATIVE_ARGS cum
;
7345 cumulative_args_t cum_v
;
7347 arm_init_cumulative_args (&cum
, fntype
, NULL_RTX
, NULL_TREE
);
7348 cum_v
= pack_cumulative_args (&cum
);
7350 for (tree t
= TYPE_ARG_TYPES (fntype
); t
; t
= TREE_CHAIN (t
))
7352 tree type
= TREE_VALUE (t
);
7353 if (!VOID_TYPE_P (type
))
7354 arm_function_arg_advance (cum_v
, TYPE_MODE (type
), type
, true);
7357 if (!arm_function_arg (cum_v
, SImode
, integer_type_node
, true))
7361 /* Everything else is ok. */
7366 /* Addressing mode support functions. */
7368 /* Return nonzero if X is a legitimate immediate operand when compiling
7369 for PIC. We know that X satisfies CONSTANT_P and flag_pic is true. */
7371 legitimate_pic_operand_p (rtx x
)
7373 if (GET_CODE (x
) == SYMBOL_REF
7374 || (GET_CODE (x
) == CONST
7375 && GET_CODE (XEXP (x
, 0)) == PLUS
7376 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == SYMBOL_REF
))
7382 /* Record that the current function needs a PIC register. Initialize
7383 cfun->machine->pic_reg if we have not already done so. */
7386 require_pic_register (void)
7388 /* A lot of the logic here is made obscure by the fact that this
7389 routine gets called as part of the rtx cost estimation process.
7390 We don't want those calls to affect any assumptions about the real
7391 function; and further, we can't call entry_of_function() until we
7392 start the real expansion process. */
7393 if (!crtl
->uses_pic_offset_table
)
7395 gcc_assert (can_create_pseudo_p ());
7396 if (arm_pic_register
!= INVALID_REGNUM
7397 && !(TARGET_THUMB1
&& arm_pic_register
> LAST_LO_REGNUM
))
7399 if (!cfun
->machine
->pic_reg
)
7400 cfun
->machine
->pic_reg
= gen_rtx_REG (Pmode
, arm_pic_register
);
7402 /* Play games to avoid marking the function as needing pic
7403 if we are being called as part of the cost-estimation
7405 if (current_ir_type () != IR_GIMPLE
|| currently_expanding_to_rtl
)
7406 crtl
->uses_pic_offset_table
= 1;
7410 rtx_insn
*seq
, *insn
;
7412 if (!cfun
->machine
->pic_reg
)
7413 cfun
->machine
->pic_reg
= gen_reg_rtx (Pmode
);
7415 /* Play games to avoid marking the function as needing pic
7416 if we are being called as part of the cost-estimation
7418 if (current_ir_type () != IR_GIMPLE
|| currently_expanding_to_rtl
)
7420 crtl
->uses_pic_offset_table
= 1;
7423 if (TARGET_THUMB1
&& arm_pic_register
!= INVALID_REGNUM
7424 && arm_pic_register
> LAST_LO_REGNUM
)
7425 emit_move_insn (cfun
->machine
->pic_reg
,
7426 gen_rtx_REG (Pmode
, arm_pic_register
));
7428 arm_load_pic_register (0UL);
7433 for (insn
= seq
; insn
; insn
= NEXT_INSN (insn
))
7435 INSN_LOCATION (insn
) = prologue_location
;
7437 /* We can be called during expansion of PHI nodes, where
7438 we can't yet emit instructions directly in the final
7439 insn stream. Queue the insns on the entry edge, they will
7440 be committed after everything else is expanded. */
7441 insert_insn_on_edge (seq
,
7442 single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun
)));
7449 legitimize_pic_address (rtx orig
, machine_mode mode
, rtx reg
)
7451 if (GET_CODE (orig
) == SYMBOL_REF
7452 || GET_CODE (orig
) == LABEL_REF
)
7456 gcc_assert (can_create_pseudo_p ());
7457 reg
= gen_reg_rtx (Pmode
);
7460 /* VxWorks does not impose a fixed gap between segments; the run-time
7461 gap can be different from the object-file gap. We therefore can't
7462 use GOTOFF unless we are absolutely sure that the symbol is in the
7463 same segment as the GOT. Unfortunately, the flexibility of linker
7464 scripts means that we can't be sure of that in general, so assume
7465 that GOTOFF is never valid on VxWorks. */
7466 /* References to weak symbols cannot be resolved locally: they
7467 may be overridden by a non-weak definition at link time. */
7469 if ((GET_CODE (orig
) == LABEL_REF
7470 || (GET_CODE (orig
) == SYMBOL_REF
7471 && SYMBOL_REF_LOCAL_P (orig
)
7472 && (SYMBOL_REF_DECL (orig
)
7473 ? !DECL_WEAK (SYMBOL_REF_DECL (orig
)) : 1)))
7475 && arm_pic_data_is_text_relative
)
7476 insn
= arm_pic_static_addr (orig
, reg
);
7482 /* If this function doesn't have a pic register, create one now. */
7483 require_pic_register ();
7485 pat
= gen_calculate_pic_address (reg
, cfun
->machine
->pic_reg
, orig
);
7487 /* Make the MEM as close to a constant as possible. */
7488 mem
= SET_SRC (pat
);
7489 gcc_assert (MEM_P (mem
) && !MEM_VOLATILE_P (mem
));
7490 MEM_READONLY_P (mem
) = 1;
7491 MEM_NOTRAP_P (mem
) = 1;
7493 insn
= emit_insn (pat
);
7496 /* Put a REG_EQUAL note on this insn, so that it can be optimized
7498 set_unique_reg_note (insn
, REG_EQUAL
, orig
);
7502 else if (GET_CODE (orig
) == CONST
)
7506 if (GET_CODE (XEXP (orig
, 0)) == PLUS
7507 && XEXP (XEXP (orig
, 0), 0) == cfun
->machine
->pic_reg
)
7510 /* Handle the case where we have: const (UNSPEC_TLS). */
7511 if (GET_CODE (XEXP (orig
, 0)) == UNSPEC
7512 && XINT (XEXP (orig
, 0), 1) == UNSPEC_TLS
)
7515 /* Handle the case where we have:
7516 const (plus (UNSPEC_TLS) (ADDEND)). The ADDEND must be a
7518 if (GET_CODE (XEXP (orig
, 0)) == PLUS
7519 && GET_CODE (XEXP (XEXP (orig
, 0), 0)) == UNSPEC
7520 && XINT (XEXP (XEXP (orig
, 0), 0), 1) == UNSPEC_TLS
)
7522 gcc_assert (CONST_INT_P (XEXP (XEXP (orig
, 0), 1)));
7528 gcc_assert (can_create_pseudo_p ());
7529 reg
= gen_reg_rtx (Pmode
);
7532 gcc_assert (GET_CODE (XEXP (orig
, 0)) == PLUS
);
7534 base
= legitimize_pic_address (XEXP (XEXP (orig
, 0), 0), Pmode
, reg
);
7535 offset
= legitimize_pic_address (XEXP (XEXP (orig
, 0), 1), Pmode
,
7536 base
== reg
? 0 : reg
);
7538 if (CONST_INT_P (offset
))
7540 /* The base register doesn't really matter, we only want to
7541 test the index for the appropriate mode. */
7542 if (!arm_legitimate_index_p (mode
, offset
, SET
, 0))
7544 gcc_assert (can_create_pseudo_p ());
7545 offset
= force_reg (Pmode
, offset
);
7548 if (CONST_INT_P (offset
))
7549 return plus_constant (Pmode
, base
, INTVAL (offset
));
7552 if (GET_MODE_SIZE (mode
) > 4
7553 && (GET_MODE_CLASS (mode
) == MODE_INT
7554 || TARGET_SOFT_FLOAT
))
7556 emit_insn (gen_addsi3 (reg
, base
, offset
));
7560 return gen_rtx_PLUS (Pmode
, base
, offset
);
7567 /* Find a spare register to use during the prolog of a function. */
7570 thumb_find_work_register (unsigned long pushed_regs_mask
)
7574 /* Check the argument registers first as these are call-used. The
7575 register allocation order means that sometimes r3 might be used
7576 but earlier argument registers might not, so check them all. */
7577 for (reg
= LAST_ARG_REGNUM
; reg
>= 0; reg
--)
7578 if (!df_regs_ever_live_p (reg
))
7581 /* Before going on to check the call-saved registers we can try a couple
7582 more ways of deducing that r3 is available. The first is when we are
7583 pushing anonymous arguments onto the stack and we have less than 4
7584 registers worth of fixed arguments(*). In this case r3 will be part of
7585 the variable argument list and so we can be sure that it will be
7586 pushed right at the start of the function. Hence it will be available
7587 for the rest of the prologue.
7588 (*): ie crtl->args.pretend_args_size is greater than 0. */
7589 if (cfun
->machine
->uses_anonymous_args
7590 && crtl
->args
.pretend_args_size
> 0)
7591 return LAST_ARG_REGNUM
;
7593 /* The other case is when we have fixed arguments but less than 4 registers
7594 worth. In this case r3 might be used in the body of the function, but
7595 it is not being used to convey an argument into the function. In theory
7596 we could just check crtl->args.size to see how many bytes are
7597 being passed in argument registers, but it seems that it is unreliable.
7598 Sometimes it will have the value 0 when in fact arguments are being
7599 passed. (See testcase execute/20021111-1.c for an example). So we also
7600 check the args_info.nregs field as well. The problem with this field is
7601 that it makes no allowances for arguments that are passed to the
7602 function but which are not used. Hence we could miss an opportunity
7603 when a function has an unused argument in r3. But it is better to be
7604 safe than to be sorry. */
7605 if (! cfun
->machine
->uses_anonymous_args
7606 && crtl
->args
.size
>= 0
7607 && crtl
->args
.size
<= (LAST_ARG_REGNUM
* UNITS_PER_WORD
)
7608 && (TARGET_AAPCS_BASED
7609 ? crtl
->args
.info
.aapcs_ncrn
< 4
7610 : crtl
->args
.info
.nregs
< 4))
7611 return LAST_ARG_REGNUM
;
7613 /* Otherwise look for a call-saved register that is going to be pushed. */
7614 for (reg
= LAST_LO_REGNUM
; reg
> LAST_ARG_REGNUM
; reg
--)
7615 if (pushed_regs_mask
& (1 << reg
))
7620 /* Thumb-2 can use high regs. */
7621 for (reg
= FIRST_HI_REGNUM
; reg
< 15; reg
++)
7622 if (pushed_regs_mask
& (1 << reg
))
7625 /* Something went wrong - thumb_compute_save_reg_mask()
7626 should have arranged for a suitable register to be pushed. */
7630 static GTY(()) int pic_labelno
;
7632 /* Generate code to load the PIC register. In thumb mode SCRATCH is a
7636 arm_load_pic_register (unsigned long saved_regs ATTRIBUTE_UNUSED
)
7638 rtx l1
, labelno
, pic_tmp
, pic_rtx
, pic_reg
;
7640 if (crtl
->uses_pic_offset_table
== 0 || TARGET_SINGLE_PIC_BASE
)
7643 gcc_assert (flag_pic
);
7645 pic_reg
= cfun
->machine
->pic_reg
;
7646 if (TARGET_VXWORKS_RTP
)
7648 pic_rtx
= gen_rtx_SYMBOL_REF (Pmode
, VXWORKS_GOTT_BASE
);
7649 pic_rtx
= gen_rtx_CONST (Pmode
, pic_rtx
);
7650 emit_insn (gen_pic_load_addr_32bit (pic_reg
, pic_rtx
));
7652 emit_insn (gen_rtx_SET (pic_reg
, gen_rtx_MEM (Pmode
, pic_reg
)));
7654 pic_tmp
= gen_rtx_SYMBOL_REF (Pmode
, VXWORKS_GOTT_INDEX
);
7655 emit_insn (gen_pic_offset_arm (pic_reg
, pic_reg
, pic_tmp
));
7659 /* We use an UNSPEC rather than a LABEL_REF because this label
7660 never appears in the code stream. */
7662 labelno
= GEN_INT (pic_labelno
++);
7663 l1
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, labelno
), UNSPEC_PIC_LABEL
);
7664 l1
= gen_rtx_CONST (VOIDmode
, l1
);
7666 /* On the ARM the PC register contains 'dot + 8' at the time of the
7667 addition, on the Thumb it is 'dot + 4'. */
7668 pic_rtx
= plus_constant (Pmode
, l1
, TARGET_ARM
? 8 : 4);
7669 pic_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, pic_rtx
),
7671 pic_rtx
= gen_rtx_CONST (Pmode
, pic_rtx
);
7675 emit_insn (gen_pic_load_addr_unified (pic_reg
, pic_rtx
, labelno
));
7677 else /* TARGET_THUMB1 */
7679 if (arm_pic_register
!= INVALID_REGNUM
7680 && REGNO (pic_reg
) > LAST_LO_REGNUM
)
7682 /* We will have pushed the pic register, so we should always be
7683 able to find a work register. */
7684 pic_tmp
= gen_rtx_REG (SImode
,
7685 thumb_find_work_register (saved_regs
));
7686 emit_insn (gen_pic_load_addr_thumb1 (pic_tmp
, pic_rtx
));
7687 emit_insn (gen_movsi (pic_offset_table_rtx
, pic_tmp
));
7688 emit_insn (gen_pic_add_dot_plus_four (pic_reg
, pic_reg
, labelno
));
7690 else if (arm_pic_register
!= INVALID_REGNUM
7691 && arm_pic_register
> LAST_LO_REGNUM
7692 && REGNO (pic_reg
) <= LAST_LO_REGNUM
)
7694 emit_insn (gen_pic_load_addr_unified (pic_reg
, pic_rtx
, labelno
));
7695 emit_move_insn (gen_rtx_REG (Pmode
, arm_pic_register
), pic_reg
);
7696 emit_use (gen_rtx_REG (Pmode
, arm_pic_register
));
7699 emit_insn (gen_pic_load_addr_unified (pic_reg
, pic_rtx
, labelno
));
7703 /* Need to emit this whether or not we obey regdecls,
7704 since setjmp/longjmp can cause life info to screw up. */
7708 /* Generate code to load the address of a static var when flag_pic is set. */
7710 arm_pic_static_addr (rtx orig
, rtx reg
)
7712 rtx l1
, labelno
, offset_rtx
;
7714 gcc_assert (flag_pic
);
7716 /* We use an UNSPEC rather than a LABEL_REF because this label
7717 never appears in the code stream. */
7718 labelno
= GEN_INT (pic_labelno
++);
7719 l1
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, labelno
), UNSPEC_PIC_LABEL
);
7720 l1
= gen_rtx_CONST (VOIDmode
, l1
);
7722 /* On the ARM the PC register contains 'dot + 8' at the time of the
7723 addition, on the Thumb it is 'dot + 4'. */
7724 offset_rtx
= plus_constant (Pmode
, l1
, TARGET_ARM
? 8 : 4);
7725 offset_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, orig
, offset_rtx
),
7726 UNSPEC_SYMBOL_OFFSET
);
7727 offset_rtx
= gen_rtx_CONST (Pmode
, offset_rtx
);
7729 return emit_insn (gen_pic_load_addr_unified (reg
, offset_rtx
, labelno
));
7732 /* Return nonzero if X is valid as an ARM state addressing register. */
7734 arm_address_register_rtx_p (rtx x
, int strict_p
)
7744 return ARM_REGNO_OK_FOR_BASE_P (regno
);
7746 return (regno
<= LAST_ARM_REGNUM
7747 || regno
>= FIRST_PSEUDO_REGISTER
7748 || regno
== FRAME_POINTER_REGNUM
7749 || regno
== ARG_POINTER_REGNUM
);
7752 /* Return TRUE if this rtx is the difference of a symbol and a label,
7753 and will reduce to a PC-relative relocation in the object file.
7754 Expressions like this can be left alone when generating PIC, rather
7755 than forced through the GOT. */
7757 pcrel_constant_p (rtx x
)
7759 if (GET_CODE (x
) == MINUS
)
7760 return symbol_mentioned_p (XEXP (x
, 0)) && label_mentioned_p (XEXP (x
, 1));
7765 /* Return true if X will surely end up in an index register after next
7768 will_be_in_index_register (const_rtx x
)
7770 /* arm.md: calculate_pic_address will split this into a register. */
7771 return GET_CODE (x
) == UNSPEC
&& (XINT (x
, 1) == UNSPEC_PIC_SYM
);
7774 /* Return nonzero if X is a valid ARM state address operand. */
7776 arm_legitimate_address_outer_p (machine_mode mode
, rtx x
, RTX_CODE outer
,
7780 enum rtx_code code
= GET_CODE (x
);
7782 if (arm_address_register_rtx_p (x
, strict_p
))
7785 use_ldrd
= (TARGET_LDRD
7786 && (mode
== DImode
|| mode
== DFmode
));
7788 if (code
== POST_INC
|| code
== PRE_DEC
7789 || ((code
== PRE_INC
|| code
== POST_DEC
)
7790 && (use_ldrd
|| GET_MODE_SIZE (mode
) <= 4)))
7791 return arm_address_register_rtx_p (XEXP (x
, 0), strict_p
);
7793 else if ((code
== POST_MODIFY
|| code
== PRE_MODIFY
)
7794 && arm_address_register_rtx_p (XEXP (x
, 0), strict_p
)
7795 && GET_CODE (XEXP (x
, 1)) == PLUS
7796 && rtx_equal_p (XEXP (XEXP (x
, 1), 0), XEXP (x
, 0)))
7798 rtx addend
= XEXP (XEXP (x
, 1), 1);
7800 /* Don't allow ldrd post increment by register because it's hard
7801 to fixup invalid register choices. */
7803 && GET_CODE (x
) == POST_MODIFY
7807 return ((use_ldrd
|| GET_MODE_SIZE (mode
) <= 4)
7808 && arm_legitimate_index_p (mode
, addend
, outer
, strict_p
));
7811 /* After reload constants split into minipools will have addresses
7812 from a LABEL_REF. */
7813 else if (reload_completed
7814 && (code
== LABEL_REF
7816 && GET_CODE (XEXP (x
, 0)) == PLUS
7817 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == LABEL_REF
7818 && CONST_INT_P (XEXP (XEXP (x
, 0), 1)))))
7821 else if (mode
== TImode
|| (TARGET_NEON
&& VALID_NEON_STRUCT_MODE (mode
)))
7824 else if (code
== PLUS
)
7826 rtx xop0
= XEXP (x
, 0);
7827 rtx xop1
= XEXP (x
, 1);
7829 return ((arm_address_register_rtx_p (xop0
, strict_p
)
7830 && ((CONST_INT_P (xop1
)
7831 && arm_legitimate_index_p (mode
, xop1
, outer
, strict_p
))
7832 || (!strict_p
&& will_be_in_index_register (xop1
))))
7833 || (arm_address_register_rtx_p (xop1
, strict_p
)
7834 && arm_legitimate_index_p (mode
, xop0
, outer
, strict_p
)));
7838 /* Reload currently can't handle MINUS, so disable this for now */
7839 else if (GET_CODE (x
) == MINUS
)
7841 rtx xop0
= XEXP (x
, 0);
7842 rtx xop1
= XEXP (x
, 1);
7844 return (arm_address_register_rtx_p (xop0
, strict_p
)
7845 && arm_legitimate_index_p (mode
, xop1
, outer
, strict_p
));
7849 else if (GET_MODE_CLASS (mode
) != MODE_FLOAT
7850 && code
== SYMBOL_REF
7851 && CONSTANT_POOL_ADDRESS_P (x
)
7853 && symbol_mentioned_p (get_pool_constant (x
))
7854 && ! pcrel_constant_p (get_pool_constant (x
))))
7860 /* Return true if we can avoid creating a constant pool entry for x. */
7862 can_avoid_literal_pool_for_label_p (rtx x
)
7864 /* Normally we can assign constant values to target registers without
7865 the help of constant pool. But there are cases we have to use constant
7867 1) assign a label to register.
7868 2) sign-extend a 8bit value to 32bit and then assign to register.
7870 Constant pool access in format:
7871 (set (reg r0) (mem (symbol_ref (".LC0"))))
7872 will cause the use of literal pool (later in function arm_reorg).
7873 So here we mark such format as an invalid format, then the compiler
7874 will adjust it into:
7875 (set (reg r0) (symbol_ref (".LC0")))
7876 (set (reg r0) (mem (reg r0))).
7877 No extra register is required, and (mem (reg r0)) won't cause the use
7878 of literal pools. */
7879 if (arm_disable_literal_pool
&& GET_CODE (x
) == SYMBOL_REF
7880 && CONSTANT_POOL_ADDRESS_P (x
))
7886 /* Return nonzero if X is a valid Thumb-2 address operand. */
7888 thumb2_legitimate_address_p (machine_mode mode
, rtx x
, int strict_p
)
7891 enum rtx_code code
= GET_CODE (x
);
7893 if (arm_address_register_rtx_p (x
, strict_p
))
7896 use_ldrd
= (TARGET_LDRD
7897 && (mode
== DImode
|| mode
== DFmode
));
7899 if (code
== POST_INC
|| code
== PRE_DEC
7900 || ((code
== PRE_INC
|| code
== POST_DEC
)
7901 && (use_ldrd
|| GET_MODE_SIZE (mode
) <= 4)))
7902 return arm_address_register_rtx_p (XEXP (x
, 0), strict_p
);
7904 else if ((code
== POST_MODIFY
|| code
== PRE_MODIFY
)
7905 && arm_address_register_rtx_p (XEXP (x
, 0), strict_p
)
7906 && GET_CODE (XEXP (x
, 1)) == PLUS
7907 && rtx_equal_p (XEXP (XEXP (x
, 1), 0), XEXP (x
, 0)))
7909 /* Thumb-2 only has autoincrement by constant. */
7910 rtx addend
= XEXP (XEXP (x
, 1), 1);
7911 HOST_WIDE_INT offset
;
7913 if (!CONST_INT_P (addend
))
7916 offset
= INTVAL(addend
);
7917 if (GET_MODE_SIZE (mode
) <= 4)
7918 return (offset
> -256 && offset
< 256);
7920 return (use_ldrd
&& offset
> -1024 && offset
< 1024
7921 && (offset
& 3) == 0);
7924 /* After reload constants split into minipools will have addresses
7925 from a LABEL_REF. */
7926 else if (reload_completed
7927 && (code
== LABEL_REF
7929 && GET_CODE (XEXP (x
, 0)) == PLUS
7930 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == LABEL_REF
7931 && CONST_INT_P (XEXP (XEXP (x
, 0), 1)))))
7934 else if (mode
== TImode
|| (TARGET_NEON
&& VALID_NEON_STRUCT_MODE (mode
)))
7937 else if (code
== PLUS
)
7939 rtx xop0
= XEXP (x
, 0);
7940 rtx xop1
= XEXP (x
, 1);
7942 return ((arm_address_register_rtx_p (xop0
, strict_p
)
7943 && (thumb2_legitimate_index_p (mode
, xop1
, strict_p
)
7944 || (!strict_p
&& will_be_in_index_register (xop1
))))
7945 || (arm_address_register_rtx_p (xop1
, strict_p
)
7946 && thumb2_legitimate_index_p (mode
, xop0
, strict_p
)));
7949 else if (can_avoid_literal_pool_for_label_p (x
))
7952 else if (GET_MODE_CLASS (mode
) != MODE_FLOAT
7953 && code
== SYMBOL_REF
7954 && CONSTANT_POOL_ADDRESS_P (x
)
7956 && symbol_mentioned_p (get_pool_constant (x
))
7957 && ! pcrel_constant_p (get_pool_constant (x
))))
7963 /* Return nonzero if INDEX is valid for an address index operand in
7966 arm_legitimate_index_p (machine_mode mode
, rtx index
, RTX_CODE outer
,
7969 HOST_WIDE_INT range
;
7970 enum rtx_code code
= GET_CODE (index
);
7972 /* Standard coprocessor addressing modes. */
7973 if (TARGET_HARD_FLOAT
7974 && (mode
== SFmode
|| mode
== DFmode
))
7975 return (code
== CONST_INT
&& INTVAL (index
) < 1024
7976 && INTVAL (index
) > -1024
7977 && (INTVAL (index
) & 3) == 0);
7979 /* For quad modes, we restrict the constant offset to be slightly less
7980 than what the instruction format permits. We do this because for
7981 quad mode moves, we will actually decompose them into two separate
7982 double-mode reads or writes. INDEX must therefore be a valid
7983 (double-mode) offset and so should INDEX+8. */
7984 if (TARGET_NEON
&& VALID_NEON_QREG_MODE (mode
))
7985 return (code
== CONST_INT
7986 && INTVAL (index
) < 1016
7987 && INTVAL (index
) > -1024
7988 && (INTVAL (index
) & 3) == 0);
7990 /* We have no such constraint on double mode offsets, so we permit the
7991 full range of the instruction format. */
7992 if (TARGET_NEON
&& VALID_NEON_DREG_MODE (mode
))
7993 return (code
== CONST_INT
7994 && INTVAL (index
) < 1024
7995 && INTVAL (index
) > -1024
7996 && (INTVAL (index
) & 3) == 0);
7998 if (TARGET_REALLY_IWMMXT
&& VALID_IWMMXT_REG_MODE (mode
))
7999 return (code
== CONST_INT
8000 && INTVAL (index
) < 1024
8001 && INTVAL (index
) > -1024
8002 && (INTVAL (index
) & 3) == 0);
8004 if (arm_address_register_rtx_p (index
, strict_p
)
8005 && (GET_MODE_SIZE (mode
) <= 4))
8008 if (mode
== DImode
|| mode
== DFmode
)
8010 if (code
== CONST_INT
)
8012 HOST_WIDE_INT val
= INTVAL (index
);
8014 /* Assume we emit ldrd or 2x ldr if !TARGET_LDRD.
8015 If vldr is selected it uses arm_coproc_mem_operand. */
8017 return val
> -256 && val
< 256;
8019 return val
> -4096 && val
< 4092;
8022 return TARGET_LDRD
&& arm_address_register_rtx_p (index
, strict_p
);
8025 if (GET_MODE_SIZE (mode
) <= 4
8029 || (mode
== QImode
&& outer
== SIGN_EXTEND
))))
8033 rtx xiop0
= XEXP (index
, 0);
8034 rtx xiop1
= XEXP (index
, 1);
8036 return ((arm_address_register_rtx_p (xiop0
, strict_p
)
8037 && power_of_two_operand (xiop1
, SImode
))
8038 || (arm_address_register_rtx_p (xiop1
, strict_p
)
8039 && power_of_two_operand (xiop0
, SImode
)));
8041 else if (code
== LSHIFTRT
|| code
== ASHIFTRT
8042 || code
== ASHIFT
|| code
== ROTATERT
)
8044 rtx op
= XEXP (index
, 1);
8046 return (arm_address_register_rtx_p (XEXP (index
, 0), strict_p
)
8049 && INTVAL (op
) <= 31);
8053 /* For ARM v4 we may be doing a sign-extend operation during the
8059 || (outer
== SIGN_EXTEND
&& mode
== QImode
))
8065 range
= (mode
== HImode
|| mode
== HFmode
) ? 4095 : 4096;
8067 return (code
== CONST_INT
8068 && INTVAL (index
) < range
8069 && INTVAL (index
) > -range
);
8072 /* Return true if OP is a valid index scaling factor for Thumb-2 address
8073 index operand. i.e. 1, 2, 4 or 8. */
8075 thumb2_index_mul_operand (rtx op
)
8079 if (!CONST_INT_P (op
))
8083 return (val
== 1 || val
== 2 || val
== 4 || val
== 8);
8086 /* Return nonzero if INDEX is a valid Thumb-2 address index operand. */
8088 thumb2_legitimate_index_p (machine_mode mode
, rtx index
, int strict_p
)
8090 enum rtx_code code
= GET_CODE (index
);
8092 /* ??? Combine arm and thumb2 coprocessor addressing modes. */
8093 /* Standard coprocessor addressing modes. */
8094 if (TARGET_HARD_FLOAT
8095 && (mode
== SFmode
|| mode
== DFmode
))
8096 return (code
== CONST_INT
&& INTVAL (index
) < 1024
8097 /* Thumb-2 allows only > -256 index range for it's core register
8098 load/stores. Since we allow SF/DF in core registers, we have
8099 to use the intersection between -256~4096 (core) and -1024~1024
8101 && INTVAL (index
) > -256
8102 && (INTVAL (index
) & 3) == 0);
8104 if (TARGET_REALLY_IWMMXT
&& VALID_IWMMXT_REG_MODE (mode
))
8106 /* For DImode assume values will usually live in core regs
8107 and only allow LDRD addressing modes. */
8108 if (!TARGET_LDRD
|| mode
!= DImode
)
8109 return (code
== CONST_INT
8110 && INTVAL (index
) < 1024
8111 && INTVAL (index
) > -1024
8112 && (INTVAL (index
) & 3) == 0);
8115 /* For quad modes, we restrict the constant offset to be slightly less
8116 than what the instruction format permits. We do this because for
8117 quad mode moves, we will actually decompose them into two separate
8118 double-mode reads or writes. INDEX must therefore be a valid
8119 (double-mode) offset and so should INDEX+8. */
8120 if (TARGET_NEON
&& VALID_NEON_QREG_MODE (mode
))
8121 return (code
== CONST_INT
8122 && INTVAL (index
) < 1016
8123 && INTVAL (index
) > -1024
8124 && (INTVAL (index
) & 3) == 0);
8126 /* We have no such constraint on double mode offsets, so we permit the
8127 full range of the instruction format. */
8128 if (TARGET_NEON
&& VALID_NEON_DREG_MODE (mode
))
8129 return (code
== CONST_INT
8130 && INTVAL (index
) < 1024
8131 && INTVAL (index
) > -1024
8132 && (INTVAL (index
) & 3) == 0);
8134 if (arm_address_register_rtx_p (index
, strict_p
)
8135 && (GET_MODE_SIZE (mode
) <= 4))
8138 if (mode
== DImode
|| mode
== DFmode
)
8140 if (code
== CONST_INT
)
8142 HOST_WIDE_INT val
= INTVAL (index
);
8143 /* Thumb-2 ldrd only has reg+const addressing modes.
8144 Assume we emit ldrd or 2x ldr if !TARGET_LDRD.
8145 If vldr is selected it uses arm_coproc_mem_operand. */
8147 return IN_RANGE (val
, -1020, 1020) && (val
& 3) == 0;
8149 return IN_RANGE (val
, -255, 4095 - 4);
8157 rtx xiop0
= XEXP (index
, 0);
8158 rtx xiop1
= XEXP (index
, 1);
8160 return ((arm_address_register_rtx_p (xiop0
, strict_p
)
8161 && thumb2_index_mul_operand (xiop1
))
8162 || (arm_address_register_rtx_p (xiop1
, strict_p
)
8163 && thumb2_index_mul_operand (xiop0
)));
8165 else if (code
== ASHIFT
)
8167 rtx op
= XEXP (index
, 1);
8169 return (arm_address_register_rtx_p (XEXP (index
, 0), strict_p
)
8172 && INTVAL (op
) <= 3);
8175 return (code
== CONST_INT
8176 && INTVAL (index
) < 4096
8177 && INTVAL (index
) > -256);
8180 /* Return nonzero if X is valid as a 16-bit Thumb state base register. */
8182 thumb1_base_register_rtx_p (rtx x
, machine_mode mode
, int strict_p
)
8192 return THUMB1_REGNO_MODE_OK_FOR_BASE_P (regno
, mode
);
8194 return (regno
<= LAST_LO_REGNUM
8195 || regno
> LAST_VIRTUAL_REGISTER
8196 || regno
== FRAME_POINTER_REGNUM
8197 || (GET_MODE_SIZE (mode
) >= 4
8198 && (regno
== STACK_POINTER_REGNUM
8199 || regno
>= FIRST_PSEUDO_REGISTER
8200 || x
== hard_frame_pointer_rtx
8201 || x
== arg_pointer_rtx
)));
8204 /* Return nonzero if x is a legitimate index register. This is the case
8205 for any base register that can access a QImode object. */
8207 thumb1_index_register_rtx_p (rtx x
, int strict_p
)
8209 return thumb1_base_register_rtx_p (x
, QImode
, strict_p
);
8212 /* Return nonzero if x is a legitimate 16-bit Thumb-state address.
8214 The AP may be eliminated to either the SP or the FP, so we use the
8215 least common denominator, e.g. SImode, and offsets from 0 to 64.
8217 ??? Verify whether the above is the right approach.
8219 ??? Also, the FP may be eliminated to the SP, so perhaps that
8220 needs special handling also.
8222 ??? Look at how the mips16 port solves this problem. It probably uses
8223 better ways to solve some of these problems.
8225 Although it is not incorrect, we don't accept QImode and HImode
8226 addresses based on the frame pointer or arg pointer until the
8227 reload pass starts. This is so that eliminating such addresses
8228 into stack based ones won't produce impossible code. */
8230 thumb1_legitimate_address_p (machine_mode mode
, rtx x
, int strict_p
)
8232 if (TARGET_HAVE_MOVT
&& can_avoid_literal_pool_for_label_p (x
))
8235 /* ??? Not clear if this is right. Experiment. */
8236 if (GET_MODE_SIZE (mode
) < 4
8237 && !(reload_in_progress
|| reload_completed
)
8238 && (reg_mentioned_p (frame_pointer_rtx
, x
)
8239 || reg_mentioned_p (arg_pointer_rtx
, x
)
8240 || reg_mentioned_p (virtual_incoming_args_rtx
, x
)
8241 || reg_mentioned_p (virtual_outgoing_args_rtx
, x
)
8242 || reg_mentioned_p (virtual_stack_dynamic_rtx
, x
)
8243 || reg_mentioned_p (virtual_stack_vars_rtx
, x
)))
8246 /* Accept any base register. SP only in SImode or larger. */
8247 else if (thumb1_base_register_rtx_p (x
, mode
, strict_p
))
8250 /* This is PC relative data before arm_reorg runs. */
8251 else if (GET_MODE_SIZE (mode
) >= 4 && CONSTANT_P (x
)
8252 && GET_CODE (x
) == SYMBOL_REF
8253 && CONSTANT_POOL_ADDRESS_P (x
) && !flag_pic
)
8256 /* This is PC relative data after arm_reorg runs. */
8257 else if ((GET_MODE_SIZE (mode
) >= 4 || mode
== HFmode
)
8259 && (GET_CODE (x
) == LABEL_REF
8260 || (GET_CODE (x
) == CONST
8261 && GET_CODE (XEXP (x
, 0)) == PLUS
8262 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == LABEL_REF
8263 && CONST_INT_P (XEXP (XEXP (x
, 0), 1)))))
8266 /* Post-inc indexing only supported for SImode and larger. */
8267 else if (GET_CODE (x
) == POST_INC
&& GET_MODE_SIZE (mode
) >= 4
8268 && thumb1_index_register_rtx_p (XEXP (x
, 0), strict_p
))
8271 else if (GET_CODE (x
) == PLUS
)
8273 /* REG+REG address can be any two index registers. */
8274 /* We disallow FRAME+REG addressing since we know that FRAME
8275 will be replaced with STACK, and SP relative addressing only
8276 permits SP+OFFSET. */
8277 if (GET_MODE_SIZE (mode
) <= 4
8278 && XEXP (x
, 0) != frame_pointer_rtx
8279 && XEXP (x
, 1) != frame_pointer_rtx
8280 && thumb1_index_register_rtx_p (XEXP (x
, 0), strict_p
)
8281 && (thumb1_index_register_rtx_p (XEXP (x
, 1), strict_p
)
8282 || (!strict_p
&& will_be_in_index_register (XEXP (x
, 1)))))
8285 /* REG+const has 5-7 bit offset for non-SP registers. */
8286 else if ((thumb1_index_register_rtx_p (XEXP (x
, 0), strict_p
)
8287 || XEXP (x
, 0) == arg_pointer_rtx
)
8288 && CONST_INT_P (XEXP (x
, 1))
8289 && thumb_legitimate_offset_p (mode
, INTVAL (XEXP (x
, 1))))
8292 /* REG+const has 10-bit offset for SP, but only SImode and
8293 larger is supported. */
8294 /* ??? Should probably check for DI/DFmode overflow here
8295 just like GO_IF_LEGITIMATE_OFFSET does. */
8296 else if (REG_P (XEXP (x
, 0))
8297 && REGNO (XEXP (x
, 0)) == STACK_POINTER_REGNUM
8298 && GET_MODE_SIZE (mode
) >= 4
8299 && CONST_INT_P (XEXP (x
, 1))
8300 && INTVAL (XEXP (x
, 1)) >= 0
8301 && INTVAL (XEXP (x
, 1)) + GET_MODE_SIZE (mode
) <= 1024
8302 && (INTVAL (XEXP (x
, 1)) & 3) == 0)
8305 else if (REG_P (XEXP (x
, 0))
8306 && (REGNO (XEXP (x
, 0)) == FRAME_POINTER_REGNUM
8307 || REGNO (XEXP (x
, 0)) == ARG_POINTER_REGNUM
8308 || (REGNO (XEXP (x
, 0)) >= FIRST_VIRTUAL_REGISTER
8309 && REGNO (XEXP (x
, 0))
8310 <= LAST_VIRTUAL_POINTER_REGISTER
))
8311 && GET_MODE_SIZE (mode
) >= 4
8312 && CONST_INT_P (XEXP (x
, 1))
8313 && (INTVAL (XEXP (x
, 1)) & 3) == 0)
8317 else if (GET_MODE_CLASS (mode
) != MODE_FLOAT
8318 && GET_MODE_SIZE (mode
) == 4
8319 && GET_CODE (x
) == SYMBOL_REF
8320 && CONSTANT_POOL_ADDRESS_P (x
)
8322 && symbol_mentioned_p (get_pool_constant (x
))
8323 && ! pcrel_constant_p (get_pool_constant (x
))))
8329 /* Return nonzero if VAL can be used as an offset in a Thumb-state address
8330 instruction of mode MODE. */
8332 thumb_legitimate_offset_p (machine_mode mode
, HOST_WIDE_INT val
)
8334 switch (GET_MODE_SIZE (mode
))
8337 return val
>= 0 && val
< 32;
8340 return val
>= 0 && val
< 64 && (val
& 1) == 0;
8344 && (val
+ GET_MODE_SIZE (mode
)) <= 128
8350 arm_legitimate_address_p (machine_mode mode
, rtx x
, bool strict_p
)
8353 return arm_legitimate_address_outer_p (mode
, x
, SET
, strict_p
);
8354 else if (TARGET_THUMB2
)
8355 return thumb2_legitimate_address_p (mode
, x
, strict_p
);
8356 else /* if (TARGET_THUMB1) */
8357 return thumb1_legitimate_address_p (mode
, x
, strict_p
);
8360 /* Worker function for TARGET_PREFERRED_RELOAD_CLASS.
8362 Given an rtx X being reloaded into a reg required to be
8363 in class CLASS, return the class of reg to actually use.
8364 In general this is just CLASS, but for the Thumb core registers and
8365 immediate constants we prefer a LO_REGS class or a subset. */
8368 arm_preferred_reload_class (rtx x ATTRIBUTE_UNUSED
, reg_class_t rclass
)
8374 if (rclass
== GENERAL_REGS
)
8381 /* Build the SYMBOL_REF for __tls_get_addr. */
8383 static GTY(()) rtx tls_get_addr_libfunc
;
8386 get_tls_get_addr (void)
8388 if (!tls_get_addr_libfunc
)
8389 tls_get_addr_libfunc
= init_one_libfunc ("__tls_get_addr");
8390 return tls_get_addr_libfunc
;
8394 arm_load_tp (rtx target
)
8397 target
= gen_reg_rtx (SImode
);
8401 /* Can return in any reg. */
8402 emit_insn (gen_load_tp_hard (target
));
8406 /* Always returned in r0. Immediately copy the result into a pseudo,
8407 otherwise other uses of r0 (e.g. setting up function arguments) may
8408 clobber the value. */
8412 emit_insn (gen_load_tp_soft ());
8414 tmp
= gen_rtx_REG (SImode
, R0_REGNUM
);
8415 emit_move_insn (target
, tmp
);
8421 load_tls_operand (rtx x
, rtx reg
)
8425 if (reg
== NULL_RTX
)
8426 reg
= gen_reg_rtx (SImode
);
8428 tmp
= gen_rtx_CONST (SImode
, x
);
8430 emit_move_insn (reg
, tmp
);
8436 arm_call_tls_get_addr (rtx x
, rtx reg
, rtx
*valuep
, int reloc
)
8438 rtx label
, labelno
, sum
;
8440 gcc_assert (reloc
!= TLS_DESCSEQ
);
8443 labelno
= GEN_INT (pic_labelno
++);
8444 label
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, labelno
), UNSPEC_PIC_LABEL
);
8445 label
= gen_rtx_CONST (VOIDmode
, label
);
8447 sum
= gen_rtx_UNSPEC (Pmode
,
8448 gen_rtvec (4, x
, GEN_INT (reloc
), label
,
8449 GEN_INT (TARGET_ARM
? 8 : 4)),
8451 reg
= load_tls_operand (sum
, reg
);
8454 emit_insn (gen_pic_add_dot_plus_eight (reg
, reg
, labelno
));
8456 emit_insn (gen_pic_add_dot_plus_four (reg
, reg
, labelno
));
8458 *valuep
= emit_library_call_value (get_tls_get_addr (), NULL_RTX
,
8459 LCT_PURE
, /* LCT_CONST? */
8462 rtx_insn
*insns
= get_insns ();
8469 arm_tls_descseq_addr (rtx x
, rtx reg
)
8471 rtx labelno
= GEN_INT (pic_labelno
++);
8472 rtx label
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, labelno
), UNSPEC_PIC_LABEL
);
8473 rtx sum
= gen_rtx_UNSPEC (Pmode
,
8474 gen_rtvec (4, x
, GEN_INT (TLS_DESCSEQ
),
8475 gen_rtx_CONST (VOIDmode
, label
),
8476 GEN_INT (!TARGET_ARM
)),
8478 rtx reg0
= load_tls_operand (sum
, gen_rtx_REG (SImode
, R0_REGNUM
));
8480 emit_insn (gen_tlscall (x
, labelno
));
8482 reg
= gen_reg_rtx (SImode
);
8484 gcc_assert (REGNO (reg
) != R0_REGNUM
);
8486 emit_move_insn (reg
, reg0
);
8492 legitimize_tls_address (rtx x
, rtx reg
)
8494 rtx dest
, tp
, label
, labelno
, sum
, ret
, eqv
, addend
;
8496 unsigned int model
= SYMBOL_REF_TLS_MODEL (x
);
8500 case TLS_MODEL_GLOBAL_DYNAMIC
:
8501 if (TARGET_GNU2_TLS
)
8503 reg
= arm_tls_descseq_addr (x
, reg
);
8505 tp
= arm_load_tp (NULL_RTX
);
8507 dest
= gen_rtx_PLUS (Pmode
, tp
, reg
);
8511 /* Original scheme */
8512 insns
= arm_call_tls_get_addr (x
, reg
, &ret
, TLS_GD32
);
8513 dest
= gen_reg_rtx (Pmode
);
8514 emit_libcall_block (insns
, dest
, ret
, x
);
8518 case TLS_MODEL_LOCAL_DYNAMIC
:
8519 if (TARGET_GNU2_TLS
)
8521 reg
= arm_tls_descseq_addr (x
, reg
);
8523 tp
= arm_load_tp (NULL_RTX
);
8525 dest
= gen_rtx_PLUS (Pmode
, tp
, reg
);
8529 insns
= arm_call_tls_get_addr (x
, reg
, &ret
, TLS_LDM32
);
8531 /* Attach a unique REG_EQUIV, to allow the RTL optimizers to
8532 share the LDM result with other LD model accesses. */
8533 eqv
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, const1_rtx
),
8535 dest
= gen_reg_rtx (Pmode
);
8536 emit_libcall_block (insns
, dest
, ret
, eqv
);
8538 /* Load the addend. */
8539 addend
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, x
,
8540 GEN_INT (TLS_LDO32
)),
8542 addend
= force_reg (SImode
, gen_rtx_CONST (SImode
, addend
));
8543 dest
= gen_rtx_PLUS (Pmode
, dest
, addend
);
8547 case TLS_MODEL_INITIAL_EXEC
:
8548 labelno
= GEN_INT (pic_labelno
++);
8549 label
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, labelno
), UNSPEC_PIC_LABEL
);
8550 label
= gen_rtx_CONST (VOIDmode
, label
);
8551 sum
= gen_rtx_UNSPEC (Pmode
,
8552 gen_rtvec (4, x
, GEN_INT (TLS_IE32
), label
,
8553 GEN_INT (TARGET_ARM
? 8 : 4)),
8555 reg
= load_tls_operand (sum
, reg
);
8558 emit_insn (gen_tls_load_dot_plus_eight (reg
, reg
, labelno
));
8559 else if (TARGET_THUMB2
)
8560 emit_insn (gen_tls_load_dot_plus_four (reg
, NULL
, reg
, labelno
));
8563 emit_insn (gen_pic_add_dot_plus_four (reg
, reg
, labelno
));
8564 emit_move_insn (reg
, gen_const_mem (SImode
, reg
));
8567 tp
= arm_load_tp (NULL_RTX
);
8569 return gen_rtx_PLUS (Pmode
, tp
, reg
);
8571 case TLS_MODEL_LOCAL_EXEC
:
8572 tp
= arm_load_tp (NULL_RTX
);
8574 reg
= gen_rtx_UNSPEC (Pmode
,
8575 gen_rtvec (2, x
, GEN_INT (TLS_LE32
)),
8577 reg
= force_reg (SImode
, gen_rtx_CONST (SImode
, reg
));
8579 return gen_rtx_PLUS (Pmode
, tp
, reg
);
8586 /* Try machine-dependent ways of modifying an illegitimate address
8587 to be legitimate. If we find one, return the new, valid address. */
8589 arm_legitimize_address (rtx x
, rtx orig_x
, machine_mode mode
)
8591 if (arm_tls_referenced_p (x
))
8595 if (GET_CODE (x
) == CONST
&& GET_CODE (XEXP (x
, 0)) == PLUS
)
8597 addend
= XEXP (XEXP (x
, 0), 1);
8598 x
= XEXP (XEXP (x
, 0), 0);
8601 if (GET_CODE (x
) != SYMBOL_REF
)
8604 gcc_assert (SYMBOL_REF_TLS_MODEL (x
) != 0);
8606 x
= legitimize_tls_address (x
, NULL_RTX
);
8610 x
= gen_rtx_PLUS (SImode
, x
, addend
);
8619 /* TODO: legitimize_address for Thumb2. */
8622 return thumb_legitimize_address (x
, orig_x
, mode
);
8625 if (GET_CODE (x
) == PLUS
)
8627 rtx xop0
= XEXP (x
, 0);
8628 rtx xop1
= XEXP (x
, 1);
8630 if (CONSTANT_P (xop0
) && !symbol_mentioned_p (xop0
))
8631 xop0
= force_reg (SImode
, xop0
);
8633 if (CONSTANT_P (xop1
) && !CONST_INT_P (xop1
)
8634 && !symbol_mentioned_p (xop1
))
8635 xop1
= force_reg (SImode
, xop1
);
8637 if (ARM_BASE_REGISTER_RTX_P (xop0
)
8638 && CONST_INT_P (xop1
))
8640 HOST_WIDE_INT n
, low_n
;
8644 /* VFP addressing modes actually allow greater offsets, but for
8645 now we just stick with the lowest common denominator. */
8646 if (mode
== DImode
|| mode
== DFmode
)
8658 low_n
= ((mode
) == TImode
? 0
8659 : n
>= 0 ? (n
& 0xfff) : -((-n
) & 0xfff));
8663 base_reg
= gen_reg_rtx (SImode
);
8664 val
= force_operand (plus_constant (Pmode
, xop0
, n
), NULL_RTX
);
8665 emit_move_insn (base_reg
, val
);
8666 x
= plus_constant (Pmode
, base_reg
, low_n
);
8668 else if (xop0
!= XEXP (x
, 0) || xop1
!= XEXP (x
, 1))
8669 x
= gen_rtx_PLUS (SImode
, xop0
, xop1
);
8672 /* XXX We don't allow MINUS any more -- see comment in
8673 arm_legitimate_address_outer_p (). */
8674 else if (GET_CODE (x
) == MINUS
)
8676 rtx xop0
= XEXP (x
, 0);
8677 rtx xop1
= XEXP (x
, 1);
8679 if (CONSTANT_P (xop0
))
8680 xop0
= force_reg (SImode
, xop0
);
8682 if (CONSTANT_P (xop1
) && ! symbol_mentioned_p (xop1
))
8683 xop1
= force_reg (SImode
, xop1
);
8685 if (xop0
!= XEXP (x
, 0) || xop1
!= XEXP (x
, 1))
8686 x
= gen_rtx_MINUS (SImode
, xop0
, xop1
);
8689 /* Make sure to take full advantage of the pre-indexed addressing mode
8690 with absolute addresses which often allows for the base register to
8691 be factorized for multiple adjacent memory references, and it might
8692 even allows for the mini pool to be avoided entirely. */
8693 else if (CONST_INT_P (x
) && optimize
> 0)
8696 HOST_WIDE_INT mask
, base
, index
;
8699 /* ldr and ldrb can use a 12-bit index, ldrsb and the rest can only
8700 use a 8-bit index. So let's use a 12-bit index for SImode only and
8701 hope that arm_gen_constant will enable ldrb to use more bits. */
8702 bits
= (mode
== SImode
) ? 12 : 8;
8703 mask
= (1 << bits
) - 1;
8704 base
= INTVAL (x
) & ~mask
;
8705 index
= INTVAL (x
) & mask
;
8706 if (bit_count (base
& 0xffffffff) > (32 - bits
)/2)
8708 /* It'll most probably be more efficient to generate the base
8709 with more bits set and use a negative index instead. */
8713 base_reg
= force_reg (SImode
, GEN_INT (base
));
8714 x
= plus_constant (Pmode
, base_reg
, index
);
8719 /* We need to find and carefully transform any SYMBOL and LABEL
8720 references; so go back to the original address expression. */
8721 rtx new_x
= legitimize_pic_address (orig_x
, mode
, NULL_RTX
);
8723 if (new_x
!= orig_x
)
8731 /* Try machine-dependent ways of modifying an illegitimate Thumb address
8732 to be legitimate. If we find one, return the new, valid address. */
8734 thumb_legitimize_address (rtx x
, rtx orig_x
, machine_mode mode
)
8736 if (GET_CODE (x
) == PLUS
8737 && CONST_INT_P (XEXP (x
, 1))
8738 && (INTVAL (XEXP (x
, 1)) >= 32 * GET_MODE_SIZE (mode
)
8739 || INTVAL (XEXP (x
, 1)) < 0))
8741 rtx xop0
= XEXP (x
, 0);
8742 rtx xop1
= XEXP (x
, 1);
8743 HOST_WIDE_INT offset
= INTVAL (xop1
);
8745 /* Try and fold the offset into a biasing of the base register and
8746 then offsetting that. Don't do this when optimizing for space
8747 since it can cause too many CSEs. */
8748 if (optimize_size
&& offset
>= 0
8749 && offset
< 256 + 31 * GET_MODE_SIZE (mode
))
8751 HOST_WIDE_INT delta
;
8754 delta
= offset
- (256 - GET_MODE_SIZE (mode
));
8755 else if (offset
< 32 * GET_MODE_SIZE (mode
) + 8)
8756 delta
= 31 * GET_MODE_SIZE (mode
);
8758 delta
= offset
& (~31 * GET_MODE_SIZE (mode
));
8760 xop0
= force_operand (plus_constant (Pmode
, xop0
, offset
- delta
),
8762 x
= plus_constant (Pmode
, xop0
, delta
);
8764 else if (offset
< 0 && offset
> -256)
8765 /* Small negative offsets are best done with a subtract before the
8766 dereference, forcing these into a register normally takes two
8768 x
= force_operand (x
, NULL_RTX
);
8771 /* For the remaining cases, force the constant into a register. */
8772 xop1
= force_reg (SImode
, xop1
);
8773 x
= gen_rtx_PLUS (SImode
, xop0
, xop1
);
8776 else if (GET_CODE (x
) == PLUS
8777 && s_register_operand (XEXP (x
, 1), SImode
)
8778 && !s_register_operand (XEXP (x
, 0), SImode
))
8780 rtx xop0
= force_operand (XEXP (x
, 0), NULL_RTX
);
8782 x
= gen_rtx_PLUS (SImode
, xop0
, XEXP (x
, 1));
8787 /* We need to find and carefully transform any SYMBOL and LABEL
8788 references; so go back to the original address expression. */
8789 rtx new_x
= legitimize_pic_address (orig_x
, mode
, NULL_RTX
);
8791 if (new_x
!= orig_x
)
8798 /* Return TRUE if X contains any TLS symbol references. */
8801 arm_tls_referenced_p (rtx x
)
8803 if (! TARGET_HAVE_TLS
)
8806 subrtx_iterator::array_type array
;
8807 FOR_EACH_SUBRTX (iter
, array
, x
, ALL
)
8809 const_rtx x
= *iter
;
8810 if (GET_CODE (x
) == SYMBOL_REF
&& SYMBOL_REF_TLS_MODEL (x
) != 0)
8812 /* ARM currently does not provide relocations to encode TLS variables
8813 into AArch32 instructions, only data, so there is no way to
8814 currently implement these if a literal pool is disabled. */
8815 if (arm_disable_literal_pool
)
8816 sorry ("accessing thread-local storage is not currently supported "
8817 "with -mpure-code or -mslow-flash-data");
8822 /* Don't recurse into UNSPEC_TLS looking for TLS symbols; these are
8823 TLS offsets, not real symbol references. */
8824 if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLS
)
8825 iter
.skip_subrtxes ();
8830 /* Implement TARGET_LEGITIMATE_CONSTANT_P.
8832 On the ARM, allow any integer (invalid ones are removed later by insn
8833 patterns), nice doubles and symbol_refs which refer to the function's
8836 When generating pic allow anything. */
8839 arm_legitimate_constant_p_1 (machine_mode
, rtx x
)
8841 return flag_pic
|| !label_mentioned_p (x
);
8845 thumb_legitimate_constant_p (machine_mode mode ATTRIBUTE_UNUSED
, rtx x
)
8847 /* Splitters for TARGET_USE_MOVT call arm_emit_movpair which creates high
8848 RTX. These RTX must therefore be allowed for Thumb-1 so that when run
8849 for ARMv8-M Baseline or later the result is valid. */
8850 if (TARGET_HAVE_MOVT
&& GET_CODE (x
) == HIGH
)
8853 return (CONST_INT_P (x
)
8854 || CONST_DOUBLE_P (x
)
8855 || CONSTANT_ADDRESS_P (x
)
8856 || (TARGET_HAVE_MOVT
&& GET_CODE (x
) == SYMBOL_REF
)
8861 arm_legitimate_constant_p (machine_mode mode
, rtx x
)
8863 return (!arm_cannot_force_const_mem (mode
, x
)
8865 ? arm_legitimate_constant_p_1 (mode
, x
)
8866 : thumb_legitimate_constant_p (mode
, x
)));
8869 /* Implement TARGET_CANNOT_FORCE_CONST_MEM. */
8872 arm_cannot_force_const_mem (machine_mode mode ATTRIBUTE_UNUSED
, rtx x
)
8876 if (ARM_OFFSETS_MUST_BE_WITHIN_SECTIONS_P
)
8878 split_const (x
, &base
, &offset
);
8879 if (GET_CODE (base
) == SYMBOL_REF
8880 && !offset_within_block_p (base
, INTVAL (offset
)))
8883 return arm_tls_referenced_p (x
);
8886 #define REG_OR_SUBREG_REG(X) \
8888 || (GET_CODE (X) == SUBREG && REG_P (SUBREG_REG (X))))
8890 #define REG_OR_SUBREG_RTX(X) \
8891 (REG_P (X) ? (X) : SUBREG_REG (X))
8894 thumb1_rtx_costs (rtx x
, enum rtx_code code
, enum rtx_code outer
)
8896 machine_mode mode
= GET_MODE (x
);
8905 return (mode
== SImode
) ? COSTS_N_INSNS (1) : COSTS_N_INSNS (2);
8912 return COSTS_N_INSNS (1);
8915 if (arm_arch6m
&& arm_m_profile_small_mul
)
8916 return COSTS_N_INSNS (32);
8918 if (CONST_INT_P (XEXP (x
, 1)))
8921 unsigned HOST_WIDE_INT i
= INTVAL (XEXP (x
, 1));
8928 return COSTS_N_INSNS (2) + cycles
;
8930 return COSTS_N_INSNS (1) + 16;
8933 /* A SET doesn't have a mode, so let's look at the SET_DEST to get
8935 words
= ARM_NUM_INTS (GET_MODE_SIZE (GET_MODE (SET_DEST (x
))));
8936 return (COSTS_N_INSNS (words
)
8937 + 4 * ((MEM_P (SET_SRC (x
)))
8938 + MEM_P (SET_DEST (x
))));
8943 if (UINTVAL (x
) < 256
8944 /* 16-bit constant. */
8945 || (TARGET_HAVE_MOVT
&& !(INTVAL (x
) & 0xffff0000)))
8947 if (thumb_shiftable_const (INTVAL (x
)))
8948 return COSTS_N_INSNS (2);
8949 return COSTS_N_INSNS (3);
8951 else if ((outer
== PLUS
|| outer
== COMPARE
)
8952 && INTVAL (x
) < 256 && INTVAL (x
) > -256)
8954 else if ((outer
== IOR
|| outer
== XOR
|| outer
== AND
)
8955 && INTVAL (x
) < 256 && INTVAL (x
) >= -256)
8956 return COSTS_N_INSNS (1);
8957 else if (outer
== AND
)
8960 /* This duplicates the tests in the andsi3 expander. */
8961 for (i
= 9; i
<= 31; i
++)
8962 if ((HOST_WIDE_INT_1
<< i
) - 1 == INTVAL (x
)
8963 || (HOST_WIDE_INT_1
<< i
) - 1 == ~INTVAL (x
))
8964 return COSTS_N_INSNS (2);
8966 else if (outer
== ASHIFT
|| outer
== ASHIFTRT
8967 || outer
== LSHIFTRT
)
8969 return COSTS_N_INSNS (2);
8975 return COSTS_N_INSNS (3);
8993 /* XXX another guess. */
8994 /* Memory costs quite a lot for the first word, but subsequent words
8995 load at the equivalent of a single insn each. */
8996 return (10 + 4 * ((GET_MODE_SIZE (mode
) - 1) / UNITS_PER_WORD
)
8997 + ((GET_CODE (x
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (x
))
9002 if (GET_CODE (XEXP (x
, 1)) == PC
|| GET_CODE (XEXP (x
, 2)) == PC
)
9008 total
= mode
== DImode
? COSTS_N_INSNS (1) : 0;
9009 total
+= thumb1_rtx_costs (XEXP (x
, 0), GET_CODE (XEXP (x
, 0)), code
);
9015 return total
+ COSTS_N_INSNS (1);
9017 /* Assume a two-shift sequence. Increase the cost slightly so
9018 we prefer actual shifts over an extend operation. */
9019 return total
+ 1 + COSTS_N_INSNS (2);
9026 /* Estimates the size cost of thumb1 instructions.
9027 For now most of the code is copied from thumb1_rtx_costs. We need more
9028 fine grain tuning when we have more related test cases. */
9030 thumb1_size_rtx_costs (rtx x
, enum rtx_code code
, enum rtx_code outer
)
9032 machine_mode mode
= GET_MODE (x
);
9041 return (mode
== SImode
) ? COSTS_N_INSNS (1) : COSTS_N_INSNS (2);
9045 /* Thumb-1 needs two instructions to fulfill shiftadd/shiftsub0/shiftsub1
9046 defined by RTL expansion, especially for the expansion of
9048 if ((GET_CODE (XEXP (x
, 0)) == MULT
9049 && power_of_two_operand (XEXP (XEXP (x
,0),1), SImode
))
9050 || (GET_CODE (XEXP (x
, 1)) == MULT
9051 && power_of_two_operand (XEXP (XEXP (x
, 1), 1), SImode
)))
9052 return COSTS_N_INSNS (2);
9057 return COSTS_N_INSNS (1);
9060 if (CONST_INT_P (XEXP (x
, 1)))
9062 /* Thumb1 mul instruction can't operate on const. We must Load it
9063 into a register first. */
9064 int const_size
= thumb1_size_rtx_costs (XEXP (x
, 1), CONST_INT
, SET
);
9065 /* For the targets which have a very small and high-latency multiply
9066 unit, we prefer to synthesize the mult with up to 5 instructions,
9067 giving a good balance between size and performance. */
9068 if (arm_arch6m
&& arm_m_profile_small_mul
)
9069 return COSTS_N_INSNS (5);
9071 return COSTS_N_INSNS (1) + const_size
;
9073 return COSTS_N_INSNS (1);
9076 /* A SET doesn't have a mode, so let's look at the SET_DEST to get
9078 words
= ARM_NUM_INTS (GET_MODE_SIZE (GET_MODE (SET_DEST (x
))));
9079 cost
= COSTS_N_INSNS (words
);
9080 if (satisfies_constraint_J (SET_SRC (x
))
9081 || satisfies_constraint_K (SET_SRC (x
))
9082 /* Too big an immediate for a 2-byte mov, using MOVT. */
9083 || (CONST_INT_P (SET_SRC (x
))
9084 && UINTVAL (SET_SRC (x
)) >= 256
9086 && satisfies_constraint_j (SET_SRC (x
)))
9087 /* thumb1_movdi_insn. */
9088 || ((words
> 1) && MEM_P (SET_SRC (x
))))
9089 cost
+= COSTS_N_INSNS (1);
9095 if (UINTVAL (x
) < 256)
9096 return COSTS_N_INSNS (1);
9097 /* movw is 4byte long. */
9098 if (TARGET_HAVE_MOVT
&& !(INTVAL (x
) & 0xffff0000))
9099 return COSTS_N_INSNS (2);
9100 /* See split "TARGET_THUMB1 && satisfies_constraint_J". */
9101 if (INTVAL (x
) >= -255 && INTVAL (x
) <= -1)
9102 return COSTS_N_INSNS (2);
9103 /* See split "TARGET_THUMB1 && satisfies_constraint_K". */
9104 if (thumb_shiftable_const (INTVAL (x
)))
9105 return COSTS_N_INSNS (2);
9106 return COSTS_N_INSNS (3);
9108 else if ((outer
== PLUS
|| outer
== COMPARE
)
9109 && INTVAL (x
) < 256 && INTVAL (x
) > -256)
9111 else if ((outer
== IOR
|| outer
== XOR
|| outer
== AND
)
9112 && INTVAL (x
) < 256 && INTVAL (x
) >= -256)
9113 return COSTS_N_INSNS (1);
9114 else if (outer
== AND
)
9117 /* This duplicates the tests in the andsi3 expander. */
9118 for (i
= 9; i
<= 31; i
++)
9119 if ((HOST_WIDE_INT_1
<< i
) - 1 == INTVAL (x
)
9120 || (HOST_WIDE_INT_1
<< i
) - 1 == ~INTVAL (x
))
9121 return COSTS_N_INSNS (2);
9123 else if (outer
== ASHIFT
|| outer
== ASHIFTRT
9124 || outer
== LSHIFTRT
)
9126 return COSTS_N_INSNS (2);
9132 return COSTS_N_INSNS (3);
9146 return COSTS_N_INSNS (1);
9149 return (COSTS_N_INSNS (1)
9151 * ((GET_MODE_SIZE (mode
) - 1) / UNITS_PER_WORD
)
9152 + ((GET_CODE (x
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (x
))
9153 ? COSTS_N_INSNS (1) : 0));
9157 if (GET_CODE (XEXP (x
, 1)) == PC
|| GET_CODE (XEXP (x
, 2)) == PC
)
9162 /* XXX still guessing. */
9163 switch (GET_MODE (XEXP (x
, 0)))
9166 return (1 + (mode
== DImode
? 4 : 0)
9167 + (MEM_P (XEXP (x
, 0)) ? 10 : 0));
9170 return (4 + (mode
== DImode
? 4 : 0)
9171 + (MEM_P (XEXP (x
, 0)) ? 10 : 0));
9174 return (1 + (MEM_P (XEXP (x
, 0)) ? 10 : 0));
9185 /* Helper function for arm_rtx_costs. If the operand is a valid shift
9186 operand, then return the operand that is being shifted. If the shift
9187 is not by a constant, then set SHIFT_REG to point to the operand.
9188 Return NULL if OP is not a shifter operand. */
9190 shifter_op_p (rtx op
, rtx
*shift_reg
)
9192 enum rtx_code code
= GET_CODE (op
);
9194 if (code
== MULT
&& CONST_INT_P (XEXP (op
, 1))
9195 && exact_log2 (INTVAL (XEXP (op
, 1))) > 0)
9196 return XEXP (op
, 0);
9197 else if (code
== ROTATE
&& CONST_INT_P (XEXP (op
, 1)))
9198 return XEXP (op
, 0);
9199 else if (code
== ROTATERT
|| code
== ASHIFT
|| code
== LSHIFTRT
9200 || code
== ASHIFTRT
)
9202 if (!CONST_INT_P (XEXP (op
, 1)))
9203 *shift_reg
= XEXP (op
, 1);
9204 return XEXP (op
, 0);
9211 arm_unspec_cost (rtx x
, enum rtx_code
/* outer_code */, bool speed_p
, int *cost
)
9213 const struct cpu_cost_table
*extra_cost
= current_tune
->insn_extra_cost
;
9214 rtx_code code
= GET_CODE (x
);
9215 gcc_assert (code
== UNSPEC
|| code
== UNSPEC_VOLATILE
);
9217 switch (XINT (x
, 1))
9219 case UNSPEC_UNALIGNED_LOAD
:
9220 /* We can only do unaligned loads into the integer unit, and we can't
9222 *cost
= COSTS_N_INSNS (ARM_NUM_REGS (GET_MODE (x
)));
9224 *cost
+= (ARM_NUM_REGS (GET_MODE (x
)) * extra_cost
->ldst
.load
9225 + extra_cost
->ldst
.load_unaligned
);
9228 *cost
+= arm_address_cost (XEXP (XVECEXP (x
, 0, 0), 0), GET_MODE (x
),
9229 ADDR_SPACE_GENERIC
, speed_p
);
9233 case UNSPEC_UNALIGNED_STORE
:
9234 *cost
= COSTS_N_INSNS (ARM_NUM_REGS (GET_MODE (x
)));
9236 *cost
+= (ARM_NUM_REGS (GET_MODE (x
)) * extra_cost
->ldst
.store
9237 + extra_cost
->ldst
.store_unaligned
);
9239 *cost
+= rtx_cost (XVECEXP (x
, 0, 0), VOIDmode
, UNSPEC
, 0, speed_p
);
9241 *cost
+= arm_address_cost (XEXP (XVECEXP (x
, 0, 0), 0), GET_MODE (x
),
9242 ADDR_SPACE_GENERIC
, speed_p
);
9253 *cost
+= extra_cost
->fp
[GET_MODE (x
) == DFmode
].roundint
;
9257 *cost
= COSTS_N_INSNS (2);
9263 /* Cost of a libcall. We assume one insn per argument, an amount for the
9264 call (one insn for -Os) and then one for processing the result. */
9265 #define LIBCALL_COST(N) COSTS_N_INSNS (N + (speed_p ? 18 : 2))
9267 #define HANDLE_NARROW_SHIFT_ARITH(OP, IDX) \
9270 shift_op = shifter_op_p (XEXP (x, IDX), &shift_reg); \
9271 if (shift_op != NULL \
9272 && arm_rtx_shift_left_p (XEXP (x, IDX))) \
9277 *cost += extra_cost->alu.arith_shift_reg; \
9278 *cost += rtx_cost (shift_reg, GET_MODE (shift_reg), \
9279 ASHIFT, 1, speed_p); \
9282 *cost += extra_cost->alu.arith_shift; \
9284 *cost += (rtx_cost (shift_op, GET_MODE (shift_op), \
9285 ASHIFT, 0, speed_p) \
9286 + rtx_cost (XEXP (x, 1 - IDX), \
9287 GET_MODE (shift_op), \
9294 /* Helper function for arm_rtx_costs_internal. Calculates the cost of a MEM,
9295 considering the costs of the addressing mode and memory access
9298 arm_mem_costs (rtx x
, const struct cpu_cost_table
*extra_cost
,
9299 int *cost
, bool speed_p
)
9301 machine_mode mode
= GET_MODE (x
);
9303 *cost
= COSTS_N_INSNS (1);
9306 && GET_CODE (XEXP (x
, 0)) == PLUS
9307 && will_be_in_index_register (XEXP (XEXP (x
, 0), 1)))
9308 /* This will be split into two instructions. Add the cost of the
9309 additional instruction here. The cost of the memory access is computed
9310 below. See arm.md:calculate_pic_address. */
9311 *cost
+= COSTS_N_INSNS (1);
9313 /* Calculate cost of the addressing mode. */
9316 arm_addr_mode_op op_type
;
9317 switch (GET_CODE (XEXP (x
, 0)))
9321 op_type
= AMO_DEFAULT
;
9324 /* MINUS does not appear in RTL, but the architecture supports it,
9325 so handle this case defensively. */
9328 op_type
= AMO_NO_WB
;
9340 if (VECTOR_MODE_P (mode
))
9341 *cost
+= current_tune
->addr_mode_costs
->vector
[op_type
];
9342 else if (FLOAT_MODE_P (mode
))
9343 *cost
+= current_tune
->addr_mode_costs
->fp
[op_type
];
9345 *cost
+= current_tune
->addr_mode_costs
->integer
[op_type
];
9348 /* Calculate cost of memory access. */
9351 if (FLOAT_MODE_P (mode
))
9353 if (GET_MODE_SIZE (mode
) == 8)
9354 *cost
+= extra_cost
->ldst
.loadd
;
9356 *cost
+= extra_cost
->ldst
.loadf
;
9358 else if (VECTOR_MODE_P (mode
))
9359 *cost
+= extra_cost
->ldst
.loadv
;
9363 if (GET_MODE_SIZE (mode
) == 8)
9364 *cost
+= extra_cost
->ldst
.ldrd
;
9366 *cost
+= extra_cost
->ldst
.load
;
9373 /* RTX costs. Make an estimate of the cost of executing the operation
9374 X, which is contained within an operation with code OUTER_CODE.
9375 SPEED_P indicates whether the cost desired is the performance cost,
9376 or the size cost. The estimate is stored in COST and the return
9377 value is TRUE if the cost calculation is final, or FALSE if the
9378 caller should recurse through the operands of X to add additional
9381 We currently make no attempt to model the size savings of Thumb-2
9382 16-bit instructions. At the normal points in compilation where
9383 this code is called we have no measure of whether the condition
9384 flags are live or not, and thus no realistic way to determine what
9385 the size will eventually be. */
9387 arm_rtx_costs_internal (rtx x
, enum rtx_code code
, enum rtx_code outer_code
,
9388 const struct cpu_cost_table
*extra_cost
,
9389 int *cost
, bool speed_p
)
9391 machine_mode mode
= GET_MODE (x
);
9393 *cost
= COSTS_N_INSNS (1);
9398 *cost
= thumb1_rtx_costs (x
, code
, outer_code
);
9400 *cost
= thumb1_size_rtx_costs (x
, code
, outer_code
);
9408 /* SET RTXs don't have a mode so we get it from the destination. */
9409 mode
= GET_MODE (SET_DEST (x
));
9411 if (REG_P (SET_SRC (x
))
9412 && REG_P (SET_DEST (x
)))
9414 /* Assume that most copies can be done with a single insn,
9415 unless we don't have HW FP, in which case everything
9416 larger than word mode will require two insns. */
9417 *cost
= COSTS_N_INSNS (((!TARGET_HARD_FLOAT
9418 && GET_MODE_SIZE (mode
) > 4)
9421 /* Conditional register moves can be encoded
9422 in 16 bits in Thumb mode. */
9423 if (!speed_p
&& TARGET_THUMB
&& outer_code
== COND_EXEC
)
9429 if (CONST_INT_P (SET_SRC (x
)))
9431 /* Handle CONST_INT here, since the value doesn't have a mode
9432 and we would otherwise be unable to work out the true cost. */
9433 *cost
= rtx_cost (SET_DEST (x
), GET_MODE (SET_DEST (x
)), SET
,
9436 /* Slightly lower the cost of setting a core reg to a constant.
9437 This helps break up chains and allows for better scheduling. */
9438 if (REG_P (SET_DEST (x
))
9439 && REGNO (SET_DEST (x
)) <= LR_REGNUM
)
9442 /* Immediate moves with an immediate in the range [0, 255] can be
9443 encoded in 16 bits in Thumb mode. */
9444 if (!speed_p
&& TARGET_THUMB
&& GET_MODE (x
) == SImode
9445 && INTVAL (x
) >= 0 && INTVAL (x
) <=255)
9447 goto const_int_cost
;
9453 return arm_mem_costs (x
, extra_cost
, cost
, speed_p
);
9457 /* Calculations of LDM costs are complex. We assume an initial cost
9458 (ldm_1st) which will load the number of registers mentioned in
9459 ldm_regs_per_insn_1st registers; then each additional
9460 ldm_regs_per_insn_subsequent registers cost one more insn. The
9461 formula for N regs is thus:
9463 ldm_1st + COSTS_N_INSNS ((max (N - ldm_regs_per_insn_1st, 0)
9464 + ldm_regs_per_insn_subsequent - 1)
9465 / ldm_regs_per_insn_subsequent).
9467 Additional costs may also be added for addressing. A similar
9468 formula is used for STM. */
9470 bool is_ldm
= load_multiple_operation (x
, SImode
);
9471 bool is_stm
= store_multiple_operation (x
, SImode
);
9473 if (is_ldm
|| is_stm
)
9477 HOST_WIDE_INT nregs
= XVECLEN (x
, 0);
9478 HOST_WIDE_INT regs_per_insn_1st
= is_ldm
9479 ? extra_cost
->ldst
.ldm_regs_per_insn_1st
9480 : extra_cost
->ldst
.stm_regs_per_insn_1st
;
9481 HOST_WIDE_INT regs_per_insn_sub
= is_ldm
9482 ? extra_cost
->ldst
.ldm_regs_per_insn_subsequent
9483 : extra_cost
->ldst
.stm_regs_per_insn_subsequent
;
9485 *cost
+= regs_per_insn_1st
9486 + COSTS_N_INSNS (((MAX (nregs
- regs_per_insn_1st
, 0))
9487 + regs_per_insn_sub
- 1)
9488 / regs_per_insn_sub
);
9497 if (TARGET_HARD_FLOAT
&& GET_MODE_CLASS (mode
) == MODE_FLOAT
9498 && (mode
== SFmode
|| !TARGET_VFP_SINGLE
))
9499 *cost
+= COSTS_N_INSNS (speed_p
9500 ? extra_cost
->fp
[mode
!= SFmode
].div
: 0);
9501 else if (mode
== SImode
&& TARGET_IDIV
)
9502 *cost
+= COSTS_N_INSNS (speed_p
? extra_cost
->mult
[0].idiv
: 0);
9504 *cost
= LIBCALL_COST (2);
9506 /* Make the cost of sdiv more expensive so when both sdiv and udiv are
9507 possible udiv is prefered. */
9508 *cost
+= (code
== DIV
? COSTS_N_INSNS (1) : 0);
9509 return false; /* All arguments must be in registers. */
9512 /* MOD by a power of 2 can be expanded as:
9514 and r0, r0, #(n - 1)
9515 and r1, r1, #(n - 1)
9516 rsbpl r0, r1, #0. */
9517 if (CONST_INT_P (XEXP (x
, 1))
9518 && exact_log2 (INTVAL (XEXP (x
, 1))) > 0
9521 *cost
+= COSTS_N_INSNS (3);
9524 *cost
+= 2 * extra_cost
->alu
.logical
9525 + extra_cost
->alu
.arith
;
9531 /* Make the cost of sdiv more expensive so when both sdiv and udiv are
9532 possible udiv is prefered. */
9533 *cost
= LIBCALL_COST (2) + (code
== MOD
? COSTS_N_INSNS (1) : 0);
9534 return false; /* All arguments must be in registers. */
9537 if (mode
== SImode
&& REG_P (XEXP (x
, 1)))
9539 *cost
+= (COSTS_N_INSNS (1)
9540 + rtx_cost (XEXP (x
, 0), mode
, code
, 0, speed_p
));
9542 *cost
+= extra_cost
->alu
.shift_reg
;
9550 if (mode
== DImode
&& CONST_INT_P (XEXP (x
, 1)))
9552 *cost
+= (COSTS_N_INSNS (2)
9553 + rtx_cost (XEXP (x
, 0), mode
, code
, 0, speed_p
));
9555 *cost
+= 2 * extra_cost
->alu
.shift
;
9556 /* Slightly disparage left shift by 1 at so we prefer adddi3. */
9557 if (code
== ASHIFT
&& XEXP (x
, 1) == CONST1_RTX (SImode
))
9561 else if (mode
== SImode
)
9563 *cost
+= rtx_cost (XEXP (x
, 0), mode
, code
, 0, speed_p
);
9564 /* Slightly disparage register shifts at -Os, but not by much. */
9565 if (!CONST_INT_P (XEXP (x
, 1)))
9566 *cost
+= (speed_p
? extra_cost
->alu
.shift_reg
: 1
9567 + rtx_cost (XEXP (x
, 1), mode
, code
, 1, speed_p
));
9570 else if (GET_MODE_CLASS (mode
) == MODE_INT
9571 && GET_MODE_SIZE (mode
) < 4)
9575 *cost
+= rtx_cost (XEXP (x
, 0), mode
, code
, 0, speed_p
);
9576 /* Slightly disparage register shifts at -Os, but not by
9578 if (!CONST_INT_P (XEXP (x
, 1)))
9579 *cost
+= (speed_p
? extra_cost
->alu
.shift_reg
: 1
9580 + rtx_cost (XEXP (x
, 1), mode
, code
, 1, speed_p
));
9582 else if (code
== LSHIFTRT
|| code
== ASHIFTRT
)
9584 if (arm_arch_thumb2
&& CONST_INT_P (XEXP (x
, 1)))
9586 /* Can use SBFX/UBFX. */
9588 *cost
+= extra_cost
->alu
.bfx
;
9589 *cost
+= rtx_cost (XEXP (x
, 0), mode
, code
, 0, speed_p
);
9593 *cost
+= COSTS_N_INSNS (1);
9594 *cost
+= rtx_cost (XEXP (x
, 0), mode
, code
, 0, speed_p
);
9597 if (CONST_INT_P (XEXP (x
, 1)))
9598 *cost
+= 2 * extra_cost
->alu
.shift
;
9600 *cost
+= (extra_cost
->alu
.shift
9601 + extra_cost
->alu
.shift_reg
);
9604 /* Slightly disparage register shifts. */
9605 *cost
+= !CONST_INT_P (XEXP (x
, 1));
9610 *cost
= COSTS_N_INSNS (2 + !CONST_INT_P (XEXP (x
, 1)));
9611 *cost
+= rtx_cost (XEXP (x
, 0), mode
, code
, 0, speed_p
);
9614 if (CONST_INT_P (XEXP (x
, 1)))
9615 *cost
+= (2 * extra_cost
->alu
.shift
9616 + extra_cost
->alu
.log_shift
);
9618 *cost
+= (extra_cost
->alu
.shift
9619 + extra_cost
->alu
.shift_reg
9620 + extra_cost
->alu
.log_shift_reg
);
9626 *cost
= LIBCALL_COST (2);
9635 *cost
+= extra_cost
->alu
.rev
;
9642 /* No rev instruction available. Look at arm_legacy_rev
9643 and thumb_legacy_rev for the form of RTL used then. */
9646 *cost
+= COSTS_N_INSNS (9);
9650 *cost
+= 6 * extra_cost
->alu
.shift
;
9651 *cost
+= 3 * extra_cost
->alu
.logical
;
9656 *cost
+= COSTS_N_INSNS (4);
9660 *cost
+= 2 * extra_cost
->alu
.shift
;
9661 *cost
+= extra_cost
->alu
.arith_shift
;
9662 *cost
+= 2 * extra_cost
->alu
.logical
;
9670 if (TARGET_HARD_FLOAT
&& GET_MODE_CLASS (mode
) == MODE_FLOAT
9671 && (mode
== SFmode
|| !TARGET_VFP_SINGLE
))
9673 if (GET_CODE (XEXP (x
, 0)) == MULT
9674 || GET_CODE (XEXP (x
, 1)) == MULT
)
9676 rtx mul_op0
, mul_op1
, sub_op
;
9679 *cost
+= extra_cost
->fp
[mode
!= SFmode
].mult_addsub
;
9681 if (GET_CODE (XEXP (x
, 0)) == MULT
)
9683 mul_op0
= XEXP (XEXP (x
, 0), 0);
9684 mul_op1
= XEXP (XEXP (x
, 0), 1);
9685 sub_op
= XEXP (x
, 1);
9689 mul_op0
= XEXP (XEXP (x
, 1), 0);
9690 mul_op1
= XEXP (XEXP (x
, 1), 1);
9691 sub_op
= XEXP (x
, 0);
9694 /* The first operand of the multiply may be optionally
9696 if (GET_CODE (mul_op0
) == NEG
)
9697 mul_op0
= XEXP (mul_op0
, 0);
9699 *cost
+= (rtx_cost (mul_op0
, mode
, code
, 0, speed_p
)
9700 + rtx_cost (mul_op1
, mode
, code
, 0, speed_p
)
9701 + rtx_cost (sub_op
, mode
, code
, 0, speed_p
));
9707 *cost
+= extra_cost
->fp
[mode
!= SFmode
].addsub
;
9713 rtx shift_by_reg
= NULL
;
9717 shift_op
= shifter_op_p (XEXP (x
, 0), &shift_by_reg
);
9718 if (shift_op
== NULL
)
9720 shift_op
= shifter_op_p (XEXP (x
, 1), &shift_by_reg
);
9721 non_shift_op
= XEXP (x
, 0);
9724 non_shift_op
= XEXP (x
, 1);
9726 if (shift_op
!= NULL
)
9728 if (shift_by_reg
!= NULL
)
9731 *cost
+= extra_cost
->alu
.arith_shift_reg
;
9732 *cost
+= rtx_cost (shift_by_reg
, mode
, code
, 0, speed_p
);
9735 *cost
+= extra_cost
->alu
.arith_shift
;
9737 *cost
+= rtx_cost (shift_op
, mode
, code
, 0, speed_p
);
9738 *cost
+= rtx_cost (non_shift_op
, mode
, code
, 0, speed_p
);
9743 && GET_CODE (XEXP (x
, 1)) == MULT
)
9747 *cost
+= extra_cost
->mult
[0].add
;
9748 *cost
+= rtx_cost (XEXP (x
, 0), mode
, MINUS
, 0, speed_p
);
9749 *cost
+= rtx_cost (XEXP (XEXP (x
, 1), 0), mode
, MULT
, 0, speed_p
);
9750 *cost
+= rtx_cost (XEXP (XEXP (x
, 1), 1), mode
, MULT
, 1, speed_p
);
9754 if (CONST_INT_P (XEXP (x
, 0)))
9756 int insns
= arm_gen_constant (MINUS
, SImode
, NULL_RTX
,
9757 INTVAL (XEXP (x
, 0)), NULL_RTX
,
9759 *cost
= COSTS_N_INSNS (insns
);
9761 *cost
+= insns
* extra_cost
->alu
.arith
;
9762 *cost
+= rtx_cost (XEXP (x
, 1), mode
, code
, 1, speed_p
);
9766 *cost
+= extra_cost
->alu
.arith
;
9771 if (GET_MODE_CLASS (mode
) == MODE_INT
9772 && GET_MODE_SIZE (mode
) < 4)
9774 rtx shift_op
, shift_reg
;
9777 /* We check both sides of the MINUS for shifter operands since,
9778 unlike PLUS, it's not commutative. */
9780 HANDLE_NARROW_SHIFT_ARITH (MINUS
, 0);
9781 HANDLE_NARROW_SHIFT_ARITH (MINUS
, 1);
9783 /* Slightly disparage, as we might need to widen the result. */
9786 *cost
+= extra_cost
->alu
.arith
;
9788 if (CONST_INT_P (XEXP (x
, 0)))
9790 *cost
+= rtx_cost (XEXP (x
, 1), mode
, code
, 1, speed_p
);
9799 *cost
+= COSTS_N_INSNS (1);
9801 if (GET_CODE (XEXP (x
, 0)) == ZERO_EXTEND
)
9803 rtx op1
= XEXP (x
, 1);
9806 *cost
+= 2 * extra_cost
->alu
.arith
;
9808 if (GET_CODE (op1
) == ZERO_EXTEND
)
9809 *cost
+= rtx_cost (XEXP (op1
, 0), VOIDmode
, ZERO_EXTEND
,
9812 *cost
+= rtx_cost (op1
, mode
, MINUS
, 1, speed_p
);
9813 *cost
+= rtx_cost (XEXP (XEXP (x
, 0), 0), VOIDmode
, ZERO_EXTEND
,
9817 else if (GET_CODE (XEXP (x
, 0)) == SIGN_EXTEND
)
9820 *cost
+= extra_cost
->alu
.arith
+ extra_cost
->alu
.arith_shift
;
9821 *cost
+= (rtx_cost (XEXP (XEXP (x
, 0), 0), VOIDmode
, SIGN_EXTEND
,
9823 + rtx_cost (XEXP (x
, 1), mode
, MINUS
, 1, speed_p
));
9826 else if (GET_CODE (XEXP (x
, 1)) == ZERO_EXTEND
9827 || GET_CODE (XEXP (x
, 1)) == SIGN_EXTEND
)
9830 *cost
+= (extra_cost
->alu
.arith
9831 + (GET_CODE (XEXP (x
, 1)) == ZERO_EXTEND
9832 ? extra_cost
->alu
.arith
9833 : extra_cost
->alu
.arith_shift
));
9834 *cost
+= (rtx_cost (XEXP (x
, 0), mode
, MINUS
, 0, speed_p
)
9835 + rtx_cost (XEXP (XEXP (x
, 1), 0), VOIDmode
,
9836 GET_CODE (XEXP (x
, 1)), 0, speed_p
));
9841 *cost
+= 2 * extra_cost
->alu
.arith
;
9847 *cost
= LIBCALL_COST (2);
9851 if (TARGET_HARD_FLOAT
&& GET_MODE_CLASS (mode
) == MODE_FLOAT
9852 && (mode
== SFmode
|| !TARGET_VFP_SINGLE
))
9854 if (GET_CODE (XEXP (x
, 0)) == MULT
)
9856 rtx mul_op0
, mul_op1
, add_op
;
9859 *cost
+= extra_cost
->fp
[mode
!= SFmode
].mult_addsub
;
9861 mul_op0
= XEXP (XEXP (x
, 0), 0);
9862 mul_op1
= XEXP (XEXP (x
, 0), 1);
9863 add_op
= XEXP (x
, 1);
9865 *cost
+= (rtx_cost (mul_op0
, mode
, code
, 0, speed_p
)
9866 + rtx_cost (mul_op1
, mode
, code
, 0, speed_p
)
9867 + rtx_cost (add_op
, mode
, code
, 0, speed_p
));
9873 *cost
+= extra_cost
->fp
[mode
!= SFmode
].addsub
;
9876 else if (GET_MODE_CLASS (mode
) == MODE_FLOAT
)
9878 *cost
= LIBCALL_COST (2);
9882 /* Narrow modes can be synthesized in SImode, but the range
9883 of useful sub-operations is limited. Check for shift operations
9884 on one of the operands. Only left shifts can be used in the
9886 if (GET_MODE_CLASS (mode
) == MODE_INT
9887 && GET_MODE_SIZE (mode
) < 4)
9889 rtx shift_op
, shift_reg
;
9892 HANDLE_NARROW_SHIFT_ARITH (PLUS
, 0);
9894 if (CONST_INT_P (XEXP (x
, 1)))
9896 int insns
= arm_gen_constant (PLUS
, SImode
, NULL_RTX
,
9897 INTVAL (XEXP (x
, 1)), NULL_RTX
,
9899 *cost
= COSTS_N_INSNS (insns
);
9901 *cost
+= insns
* extra_cost
->alu
.arith
;
9902 /* Slightly penalize a narrow operation as the result may
9904 *cost
+= 1 + rtx_cost (XEXP (x
, 0), mode
, PLUS
, 0, speed_p
);
9908 /* Slightly penalize a narrow operation as the result may
9912 *cost
+= extra_cost
->alu
.arith
;
9919 rtx shift_op
, shift_reg
;
9922 && (GET_CODE (XEXP (x
, 0)) == ZERO_EXTEND
9923 || GET_CODE (XEXP (x
, 0)) == SIGN_EXTEND
))
9925 /* UXTA[BH] or SXTA[BH]. */
9927 *cost
+= extra_cost
->alu
.extend_arith
;
9928 *cost
+= (rtx_cost (XEXP (XEXP (x
, 0), 0), VOIDmode
, ZERO_EXTEND
,
9930 + rtx_cost (XEXP (x
, 1), mode
, PLUS
, 0, speed_p
));
9935 shift_op
= shifter_op_p (XEXP (x
, 0), &shift_reg
);
9936 if (shift_op
!= NULL
)
9941 *cost
+= extra_cost
->alu
.arith_shift_reg
;
9942 *cost
+= rtx_cost (shift_reg
, mode
, ASHIFT
, 1, speed_p
);
9945 *cost
+= extra_cost
->alu
.arith_shift
;
9947 *cost
+= (rtx_cost (shift_op
, mode
, ASHIFT
, 0, speed_p
)
9948 + rtx_cost (XEXP (x
, 1), mode
, PLUS
, 1, speed_p
));
9951 if (GET_CODE (XEXP (x
, 0)) == MULT
)
9953 rtx mul_op
= XEXP (x
, 0);
9955 if (TARGET_DSP_MULTIPLY
9956 && ((GET_CODE (XEXP (mul_op
, 0)) == SIGN_EXTEND
9957 && (GET_CODE (XEXP (mul_op
, 1)) == SIGN_EXTEND
9958 || (GET_CODE (XEXP (mul_op
, 1)) == ASHIFTRT
9959 && CONST_INT_P (XEXP (XEXP (mul_op
, 1), 1))
9960 && INTVAL (XEXP (XEXP (mul_op
, 1), 1)) == 16)))
9961 || (GET_CODE (XEXP (mul_op
, 0)) == ASHIFTRT
9962 && CONST_INT_P (XEXP (XEXP (mul_op
, 0), 1))
9963 && INTVAL (XEXP (XEXP (mul_op
, 0), 1)) == 16
9964 && (GET_CODE (XEXP (mul_op
, 1)) == SIGN_EXTEND
9965 || (GET_CODE (XEXP (mul_op
, 1)) == ASHIFTRT
9966 && CONST_INT_P (XEXP (XEXP (mul_op
, 1), 1))
9967 && (INTVAL (XEXP (XEXP (mul_op
, 1), 1))
9972 *cost
+= extra_cost
->mult
[0].extend_add
;
9973 *cost
+= (rtx_cost (XEXP (XEXP (mul_op
, 0), 0), mode
,
9974 SIGN_EXTEND
, 0, speed_p
)
9975 + rtx_cost (XEXP (XEXP (mul_op
, 1), 0), mode
,
9976 SIGN_EXTEND
, 0, speed_p
)
9977 + rtx_cost (XEXP (x
, 1), mode
, PLUS
, 1, speed_p
));
9982 *cost
+= extra_cost
->mult
[0].add
;
9983 *cost
+= (rtx_cost (XEXP (mul_op
, 0), mode
, MULT
, 0, speed_p
)
9984 + rtx_cost (XEXP (mul_op
, 1), mode
, MULT
, 1, speed_p
)
9985 + rtx_cost (XEXP (x
, 1), mode
, PLUS
, 1, speed_p
));
9988 if (CONST_INT_P (XEXP (x
, 1)))
9990 int insns
= arm_gen_constant (PLUS
, SImode
, NULL_RTX
,
9991 INTVAL (XEXP (x
, 1)), NULL_RTX
,
9993 *cost
= COSTS_N_INSNS (insns
);
9995 *cost
+= insns
* extra_cost
->alu
.arith
;
9996 *cost
+= rtx_cost (XEXP (x
, 0), mode
, PLUS
, 0, speed_p
);
10000 *cost
+= extra_cost
->alu
.arith
;
10005 if (mode
== DImode
)
10007 if (GET_CODE (XEXP (x
, 0)) == MULT
10008 && ((GET_CODE (XEXP (XEXP (x
, 0), 0)) == ZERO_EXTEND
10009 && GET_CODE (XEXP (XEXP (x
, 0), 1)) == ZERO_EXTEND
)
10010 || (GET_CODE (XEXP (XEXP (x
, 0), 0)) == SIGN_EXTEND
10011 && GET_CODE (XEXP (XEXP (x
, 0), 1)) == SIGN_EXTEND
)))
10014 *cost
+= extra_cost
->mult
[1].extend_add
;
10015 *cost
+= (rtx_cost (XEXP (XEXP (XEXP (x
, 0), 0), 0), mode
,
10016 ZERO_EXTEND
, 0, speed_p
)
10017 + rtx_cost (XEXP (XEXP (XEXP (x
, 0), 1), 0), mode
,
10018 ZERO_EXTEND
, 0, speed_p
)
10019 + rtx_cost (XEXP (x
, 1), mode
, PLUS
, 1, speed_p
));
10023 *cost
+= COSTS_N_INSNS (1);
10025 if (GET_CODE (XEXP (x
, 0)) == ZERO_EXTEND
10026 || GET_CODE (XEXP (x
, 0)) == SIGN_EXTEND
)
10029 *cost
+= (extra_cost
->alu
.arith
10030 + (GET_CODE (XEXP (x
, 0)) == ZERO_EXTEND
10031 ? extra_cost
->alu
.arith
10032 : extra_cost
->alu
.arith_shift
));
10034 *cost
+= (rtx_cost (XEXP (XEXP (x
, 0), 0), VOIDmode
, ZERO_EXTEND
,
10036 + rtx_cost (XEXP (x
, 1), mode
, PLUS
, 1, speed_p
));
10041 *cost
+= 2 * extra_cost
->alu
.arith
;
10046 *cost
= LIBCALL_COST (2);
10049 if (mode
== SImode
&& arm_arch6
&& aarch_rev16_p (x
))
10052 *cost
+= extra_cost
->alu
.rev
;
10056 /* Fall through. */
10057 case AND
: case XOR
:
10058 if (mode
== SImode
)
10060 enum rtx_code subcode
= GET_CODE (XEXP (x
, 0));
10061 rtx op0
= XEXP (x
, 0);
10062 rtx shift_op
, shift_reg
;
10066 || (code
== IOR
&& TARGET_THUMB2
)))
10067 op0
= XEXP (op0
, 0);
10070 shift_op
= shifter_op_p (op0
, &shift_reg
);
10071 if (shift_op
!= NULL
)
10076 *cost
+= extra_cost
->alu
.log_shift_reg
;
10077 *cost
+= rtx_cost (shift_reg
, mode
, ASHIFT
, 1, speed_p
);
10080 *cost
+= extra_cost
->alu
.log_shift
;
10082 *cost
+= (rtx_cost (shift_op
, mode
, ASHIFT
, 0, speed_p
)
10083 + rtx_cost (XEXP (x
, 1), mode
, code
, 1, speed_p
));
10087 if (CONST_INT_P (XEXP (x
, 1)))
10089 int insns
= arm_gen_constant (code
, SImode
, NULL_RTX
,
10090 INTVAL (XEXP (x
, 1)), NULL_RTX
,
10093 *cost
= COSTS_N_INSNS (insns
);
10095 *cost
+= insns
* extra_cost
->alu
.logical
;
10096 *cost
+= rtx_cost (op0
, mode
, code
, 0, speed_p
);
10101 *cost
+= extra_cost
->alu
.logical
;
10102 *cost
+= (rtx_cost (op0
, mode
, code
, 0, speed_p
)
10103 + rtx_cost (XEXP (x
, 1), mode
, code
, 1, speed_p
));
10107 if (mode
== DImode
)
10109 rtx op0
= XEXP (x
, 0);
10110 enum rtx_code subcode
= GET_CODE (op0
);
10112 *cost
+= COSTS_N_INSNS (1);
10116 || (code
== IOR
&& TARGET_THUMB2
)))
10117 op0
= XEXP (op0
, 0);
10119 if (GET_CODE (op0
) == ZERO_EXTEND
)
10122 *cost
+= 2 * extra_cost
->alu
.logical
;
10124 *cost
+= (rtx_cost (XEXP (op0
, 0), VOIDmode
, ZERO_EXTEND
,
10126 + rtx_cost (XEXP (x
, 1), mode
, code
, 0, speed_p
));
10129 else if (GET_CODE (op0
) == SIGN_EXTEND
)
10132 *cost
+= extra_cost
->alu
.logical
+ extra_cost
->alu
.log_shift
;
10134 *cost
+= (rtx_cost (XEXP (op0
, 0), VOIDmode
, SIGN_EXTEND
,
10136 + rtx_cost (XEXP (x
, 1), mode
, code
, 0, speed_p
));
10141 *cost
+= 2 * extra_cost
->alu
.logical
;
10147 *cost
= LIBCALL_COST (2);
10151 if (TARGET_HARD_FLOAT
&& GET_MODE_CLASS (mode
) == MODE_FLOAT
10152 && (mode
== SFmode
|| !TARGET_VFP_SINGLE
))
10154 rtx op0
= XEXP (x
, 0);
10156 if (GET_CODE (op0
) == NEG
&& !flag_rounding_math
)
10157 op0
= XEXP (op0
, 0);
10160 *cost
+= extra_cost
->fp
[mode
!= SFmode
].mult
;
10162 *cost
+= (rtx_cost (op0
, mode
, MULT
, 0, speed_p
)
10163 + rtx_cost (XEXP (x
, 1), mode
, MULT
, 1, speed_p
));
10166 else if (GET_MODE_CLASS (mode
) == MODE_FLOAT
)
10168 *cost
= LIBCALL_COST (2);
10172 if (mode
== SImode
)
10174 if (TARGET_DSP_MULTIPLY
10175 && ((GET_CODE (XEXP (x
, 0)) == SIGN_EXTEND
10176 && (GET_CODE (XEXP (x
, 1)) == SIGN_EXTEND
10177 || (GET_CODE (XEXP (x
, 1)) == ASHIFTRT
10178 && CONST_INT_P (XEXP (XEXP (x
, 1), 1))
10179 && INTVAL (XEXP (XEXP (x
, 1), 1)) == 16)))
10180 || (GET_CODE (XEXP (x
, 0)) == ASHIFTRT
10181 && CONST_INT_P (XEXP (XEXP (x
, 0), 1))
10182 && INTVAL (XEXP (XEXP (x
, 0), 1)) == 16
10183 && (GET_CODE (XEXP (x
, 1)) == SIGN_EXTEND
10184 || (GET_CODE (XEXP (x
, 1)) == ASHIFTRT
10185 && CONST_INT_P (XEXP (XEXP (x
, 1), 1))
10186 && (INTVAL (XEXP (XEXP (x
, 1), 1))
10189 /* SMUL[TB][TB]. */
10191 *cost
+= extra_cost
->mult
[0].extend
;
10192 *cost
+= rtx_cost (XEXP (XEXP (x
, 0), 0), mode
,
10193 SIGN_EXTEND
, 0, speed_p
);
10194 *cost
+= rtx_cost (XEXP (XEXP (x
, 1), 0), mode
,
10195 SIGN_EXTEND
, 1, speed_p
);
10199 *cost
+= extra_cost
->mult
[0].simple
;
10203 if (mode
== DImode
)
10205 if ((GET_CODE (XEXP (x
, 0)) == ZERO_EXTEND
10206 && GET_CODE (XEXP (x
, 1)) == ZERO_EXTEND
)
10207 || (GET_CODE (XEXP (x
, 0)) == SIGN_EXTEND
10208 && GET_CODE (XEXP (x
, 1)) == SIGN_EXTEND
))
10211 *cost
+= extra_cost
->mult
[1].extend
;
10212 *cost
+= (rtx_cost (XEXP (XEXP (x
, 0), 0), VOIDmode
,
10213 ZERO_EXTEND
, 0, speed_p
)
10214 + rtx_cost (XEXP (XEXP (x
, 1), 0), VOIDmode
,
10215 ZERO_EXTEND
, 0, speed_p
));
10219 *cost
= LIBCALL_COST (2);
10224 *cost
= LIBCALL_COST (2);
10228 if (TARGET_HARD_FLOAT
&& GET_MODE_CLASS (mode
) == MODE_FLOAT
10229 && (mode
== SFmode
|| !TARGET_VFP_SINGLE
))
10231 if (GET_CODE (XEXP (x
, 0)) == MULT
)
10234 *cost
= rtx_cost (XEXP (x
, 0), mode
, NEG
, 0, speed_p
);
10239 *cost
+= extra_cost
->fp
[mode
!= SFmode
].neg
;
10243 else if (GET_MODE_CLASS (mode
) == MODE_FLOAT
)
10245 *cost
= LIBCALL_COST (1);
10249 if (mode
== SImode
)
10251 if (GET_CODE (XEXP (x
, 0)) == ABS
)
10253 *cost
+= COSTS_N_INSNS (1);
10254 /* Assume the non-flag-changing variant. */
10256 *cost
+= (extra_cost
->alu
.log_shift
10257 + extra_cost
->alu
.arith_shift
);
10258 *cost
+= rtx_cost (XEXP (XEXP (x
, 0), 0), mode
, ABS
, 0, speed_p
);
10262 if (GET_RTX_CLASS (GET_CODE (XEXP (x
, 0))) == RTX_COMPARE
10263 || GET_RTX_CLASS (GET_CODE (XEXP (x
, 0))) == RTX_COMM_COMPARE
)
10265 *cost
+= COSTS_N_INSNS (1);
10266 /* No extra cost for MOV imm and MVN imm. */
10267 /* If the comparison op is using the flags, there's no further
10268 cost, otherwise we need to add the cost of the comparison. */
10269 if (!(REG_P (XEXP (XEXP (x
, 0), 0))
10270 && REGNO (XEXP (XEXP (x
, 0), 0)) == CC_REGNUM
10271 && XEXP (XEXP (x
, 0), 1) == const0_rtx
))
10273 mode
= GET_MODE (XEXP (XEXP (x
, 0), 0));
10274 *cost
+= (COSTS_N_INSNS (1)
10275 + rtx_cost (XEXP (XEXP (x
, 0), 0), mode
, COMPARE
,
10277 + rtx_cost (XEXP (XEXP (x
, 0), 1), mode
, COMPARE
,
10280 *cost
+= extra_cost
->alu
.arith
;
10286 *cost
+= extra_cost
->alu
.arith
;
10290 if (GET_MODE_CLASS (mode
) == MODE_INT
10291 && GET_MODE_SIZE (mode
) < 4)
10293 /* Slightly disparage, as we might need an extend operation. */
10296 *cost
+= extra_cost
->alu
.arith
;
10300 if (mode
== DImode
)
10302 *cost
+= COSTS_N_INSNS (1);
10304 *cost
+= 2 * extra_cost
->alu
.arith
;
10309 *cost
= LIBCALL_COST (1);
10313 if (mode
== SImode
)
10316 rtx shift_reg
= NULL
;
10318 shift_op
= shifter_op_p (XEXP (x
, 0), &shift_reg
);
10322 if (shift_reg
!= NULL
)
10325 *cost
+= extra_cost
->alu
.log_shift_reg
;
10326 *cost
+= rtx_cost (shift_reg
, mode
, ASHIFT
, 1, speed_p
);
10329 *cost
+= extra_cost
->alu
.log_shift
;
10330 *cost
+= rtx_cost (shift_op
, mode
, ASHIFT
, 0, speed_p
);
10335 *cost
+= extra_cost
->alu
.logical
;
10338 if (mode
== DImode
)
10340 *cost
+= COSTS_N_INSNS (1);
10346 *cost
+= LIBCALL_COST (1);
10351 if (GET_CODE (XEXP (x
, 1)) == PC
|| GET_CODE (XEXP (x
, 2)) == PC
)
10353 *cost
+= COSTS_N_INSNS (3);
10356 int op1cost
= rtx_cost (XEXP (x
, 1), mode
, SET
, 1, speed_p
);
10357 int op2cost
= rtx_cost (XEXP (x
, 2), mode
, SET
, 1, speed_p
);
10359 *cost
= rtx_cost (XEXP (x
, 0), mode
, IF_THEN_ELSE
, 0, speed_p
);
10360 /* Assume that if one arm of the if_then_else is a register,
10361 that it will be tied with the result and eliminate the
10362 conditional insn. */
10363 if (REG_P (XEXP (x
, 1)))
10365 else if (REG_P (XEXP (x
, 2)))
10371 if (extra_cost
->alu
.non_exec_costs_exec
)
10372 *cost
+= op1cost
+ op2cost
+ extra_cost
->alu
.non_exec
;
10374 *cost
+= MAX (op1cost
, op2cost
) + extra_cost
->alu
.non_exec
;
10377 *cost
+= op1cost
+ op2cost
;
10383 if (cc_register (XEXP (x
, 0), VOIDmode
) && XEXP (x
, 1) == const0_rtx
)
10387 machine_mode op0mode
;
10388 /* We'll mostly assume that the cost of a compare is the cost of the
10389 LHS. However, there are some notable exceptions. */
10391 /* Floating point compares are never done as side-effects. */
10392 op0mode
= GET_MODE (XEXP (x
, 0));
10393 if (TARGET_HARD_FLOAT
&& GET_MODE_CLASS (op0mode
) == MODE_FLOAT
10394 && (op0mode
== SFmode
|| !TARGET_VFP_SINGLE
))
10397 *cost
+= extra_cost
->fp
[op0mode
!= SFmode
].compare
;
10399 if (XEXP (x
, 1) == CONST0_RTX (op0mode
))
10401 *cost
+= rtx_cost (XEXP (x
, 0), op0mode
, code
, 0, speed_p
);
10407 else if (GET_MODE_CLASS (op0mode
) == MODE_FLOAT
)
10409 *cost
= LIBCALL_COST (2);
10413 /* DImode compares normally take two insns. */
10414 if (op0mode
== DImode
)
10416 *cost
+= COSTS_N_INSNS (1);
10418 *cost
+= 2 * extra_cost
->alu
.arith
;
10422 if (op0mode
== SImode
)
10427 if (XEXP (x
, 1) == const0_rtx
10428 && !(REG_P (XEXP (x
, 0))
10429 || (GET_CODE (XEXP (x
, 0)) == SUBREG
10430 && REG_P (SUBREG_REG (XEXP (x
, 0))))))
10432 *cost
= rtx_cost (XEXP (x
, 0), op0mode
, COMPARE
, 0, speed_p
);
10434 /* Multiply operations that set the flags are often
10435 significantly more expensive. */
10437 && GET_CODE (XEXP (x
, 0)) == MULT
10438 && !power_of_two_operand (XEXP (XEXP (x
, 0), 1), mode
))
10439 *cost
+= extra_cost
->mult
[0].flag_setting
;
10442 && GET_CODE (XEXP (x
, 0)) == PLUS
10443 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == MULT
10444 && !power_of_two_operand (XEXP (XEXP (XEXP (x
, 0),
10446 *cost
+= extra_cost
->mult
[0].flag_setting
;
10451 shift_op
= shifter_op_p (XEXP (x
, 0), &shift_reg
);
10452 if (shift_op
!= NULL
)
10454 if (shift_reg
!= NULL
)
10456 *cost
+= rtx_cost (shift_reg
, op0mode
, ASHIFT
,
10459 *cost
+= extra_cost
->alu
.arith_shift_reg
;
10462 *cost
+= extra_cost
->alu
.arith_shift
;
10463 *cost
+= rtx_cost (shift_op
, op0mode
, ASHIFT
, 0, speed_p
);
10464 *cost
+= rtx_cost (XEXP (x
, 1), op0mode
, COMPARE
, 1, speed_p
);
10469 *cost
+= extra_cost
->alu
.arith
;
10470 if (CONST_INT_P (XEXP (x
, 1))
10471 && const_ok_for_op (INTVAL (XEXP (x
, 1)), COMPARE
))
10473 *cost
+= rtx_cost (XEXP (x
, 0), op0mode
, COMPARE
, 0, speed_p
);
10481 *cost
= LIBCALL_COST (2);
10504 if (outer_code
== SET
)
10506 /* Is it a store-flag operation? */
10507 if (REG_P (XEXP (x
, 0)) && REGNO (XEXP (x
, 0)) == CC_REGNUM
10508 && XEXP (x
, 1) == const0_rtx
)
10510 /* Thumb also needs an IT insn. */
10511 *cost
+= COSTS_N_INSNS (TARGET_THUMB
? 2 : 1);
10514 if (XEXP (x
, 1) == const0_rtx
)
10519 /* LSR Rd, Rn, #31. */
10521 *cost
+= extra_cost
->alu
.shift
;
10531 *cost
+= COSTS_N_INSNS (1);
10535 /* RSBS T1, Rn, Rn, LSR #31
10537 *cost
+= COSTS_N_INSNS (1);
10539 *cost
+= extra_cost
->alu
.arith_shift
;
10543 /* RSB Rd, Rn, Rn, ASR #1
10544 LSR Rd, Rd, #31. */
10545 *cost
+= COSTS_N_INSNS (1);
10547 *cost
+= (extra_cost
->alu
.arith_shift
10548 + extra_cost
->alu
.shift
);
10554 *cost
+= COSTS_N_INSNS (1);
10556 *cost
+= extra_cost
->alu
.shift
;
10560 /* Remaining cases are either meaningless or would take
10561 three insns anyway. */
10562 *cost
= COSTS_N_INSNS (3);
10565 *cost
+= rtx_cost (XEXP (x
, 0), mode
, code
, 0, speed_p
);
10570 *cost
+= COSTS_N_INSNS (TARGET_THUMB
? 3 : 2);
10571 if (CONST_INT_P (XEXP (x
, 1))
10572 && const_ok_for_op (INTVAL (XEXP (x
, 1)), COMPARE
))
10574 *cost
+= rtx_cost (XEXP (x
, 0), mode
, code
, 0, speed_p
);
10581 /* Not directly inside a set. If it involves the condition code
10582 register it must be the condition for a branch, cond_exec or
10583 I_T_E operation. Since the comparison is performed elsewhere
10584 this is just the control part which has no additional
10586 else if (REG_P (XEXP (x
, 0)) && REGNO (XEXP (x
, 0)) == CC_REGNUM
10587 && XEXP (x
, 1) == const0_rtx
)
10595 if (TARGET_HARD_FLOAT
&& GET_MODE_CLASS (mode
) == MODE_FLOAT
10596 && (mode
== SFmode
|| !TARGET_VFP_SINGLE
))
10599 *cost
+= extra_cost
->fp
[mode
!= SFmode
].neg
;
10603 else if (GET_MODE_CLASS (mode
) == MODE_FLOAT
)
10605 *cost
= LIBCALL_COST (1);
10609 if (mode
== SImode
)
10612 *cost
+= extra_cost
->alu
.log_shift
+ extra_cost
->alu
.arith_shift
;
10616 *cost
= LIBCALL_COST (1);
10620 if ((arm_arch4
|| GET_MODE (XEXP (x
, 0)) == SImode
)
10621 && MEM_P (XEXP (x
, 0)))
10623 if (mode
== DImode
)
10624 *cost
+= COSTS_N_INSNS (1);
10629 if (GET_MODE (XEXP (x
, 0)) == SImode
)
10630 *cost
+= extra_cost
->ldst
.load
;
10632 *cost
+= extra_cost
->ldst
.load_sign_extend
;
10634 if (mode
== DImode
)
10635 *cost
+= extra_cost
->alu
.shift
;
10640 /* Widening from less than 32-bits requires an extend operation. */
10641 if (GET_MODE (XEXP (x
, 0)) != SImode
&& arm_arch6
)
10643 /* We have SXTB/SXTH. */
10644 *cost
+= rtx_cost (XEXP (x
, 0), VOIDmode
, code
, 0, speed_p
);
10646 *cost
+= extra_cost
->alu
.extend
;
10648 else if (GET_MODE (XEXP (x
, 0)) != SImode
)
10650 /* Needs two shifts. */
10651 *cost
+= COSTS_N_INSNS (1);
10652 *cost
+= rtx_cost (XEXP (x
, 0), VOIDmode
, code
, 0, speed_p
);
10654 *cost
+= 2 * extra_cost
->alu
.shift
;
10657 /* Widening beyond 32-bits requires one more insn. */
10658 if (mode
== DImode
)
10660 *cost
+= COSTS_N_INSNS (1);
10662 *cost
+= extra_cost
->alu
.shift
;
10669 || GET_MODE (XEXP (x
, 0)) == SImode
10670 || GET_MODE (XEXP (x
, 0)) == QImode
)
10671 && MEM_P (XEXP (x
, 0)))
10673 *cost
= rtx_cost (XEXP (x
, 0), VOIDmode
, code
, 0, speed_p
);
10675 if (mode
== DImode
)
10676 *cost
+= COSTS_N_INSNS (1); /* No speed penalty. */
10681 /* Widening from less than 32-bits requires an extend operation. */
10682 if (GET_MODE (XEXP (x
, 0)) == QImode
)
10684 /* UXTB can be a shorter instruction in Thumb2, but it might
10685 be slower than the AND Rd, Rn, #255 alternative. When
10686 optimizing for speed it should never be slower to use
10687 AND, and we don't really model 16-bit vs 32-bit insns
10690 *cost
+= extra_cost
->alu
.logical
;
10692 else if (GET_MODE (XEXP (x
, 0)) != SImode
&& arm_arch6
)
10694 /* We have UXTB/UXTH. */
10695 *cost
+= rtx_cost (XEXP (x
, 0), VOIDmode
, code
, 0, speed_p
);
10697 *cost
+= extra_cost
->alu
.extend
;
10699 else if (GET_MODE (XEXP (x
, 0)) != SImode
)
10701 /* Needs two shifts. It's marginally preferable to use
10702 shifts rather than two BIC instructions as the second
10703 shift may merge with a subsequent insn as a shifter
10705 *cost
= COSTS_N_INSNS (2);
10706 *cost
+= rtx_cost (XEXP (x
, 0), VOIDmode
, code
, 0, speed_p
);
10708 *cost
+= 2 * extra_cost
->alu
.shift
;
10711 /* Widening beyond 32-bits requires one more insn. */
10712 if (mode
== DImode
)
10714 *cost
+= COSTS_N_INSNS (1); /* No speed penalty. */
10721 /* CONST_INT has no mode, so we cannot tell for sure how many
10722 insns are really going to be needed. The best we can do is
10723 look at the value passed. If it fits in SImode, then assume
10724 that's the mode it will be used for. Otherwise assume it
10725 will be used in DImode. */
10726 if (INTVAL (x
) == trunc_int_for_mode (INTVAL (x
), SImode
))
10731 /* Avoid blowing up in arm_gen_constant (). */
10732 if (!(outer_code
== PLUS
10733 || outer_code
== AND
10734 || outer_code
== IOR
10735 || outer_code
== XOR
10736 || outer_code
== MINUS
))
10740 if (mode
== SImode
)
10742 *cost
+= COSTS_N_INSNS (arm_gen_constant (outer_code
, SImode
, NULL
,
10743 INTVAL (x
), NULL
, NULL
,
10749 *cost
+= COSTS_N_INSNS (arm_gen_constant
10750 (outer_code
, SImode
, NULL
,
10751 trunc_int_for_mode (INTVAL (x
), SImode
),
10753 + arm_gen_constant (outer_code
, SImode
, NULL
,
10754 INTVAL (x
) >> 32, NULL
,
10766 if (arm_arch_thumb2
&& !flag_pic
)
10767 *cost
+= COSTS_N_INSNS (1);
10769 *cost
+= extra_cost
->ldst
.load
;
10772 *cost
+= COSTS_N_INSNS (1);
10776 *cost
+= COSTS_N_INSNS (1);
10778 *cost
+= extra_cost
->alu
.arith
;
10784 *cost
= COSTS_N_INSNS (4);
10789 if (TARGET_HARD_FLOAT
&& GET_MODE_CLASS (mode
) == MODE_FLOAT
10790 && (mode
== SFmode
|| !TARGET_VFP_SINGLE
))
10792 if (vfp3_const_double_rtx (x
))
10795 *cost
+= extra_cost
->fp
[mode
== DFmode
].fpconst
;
10801 if (mode
== DFmode
)
10802 *cost
+= extra_cost
->ldst
.loadd
;
10804 *cost
+= extra_cost
->ldst
.loadf
;
10807 *cost
+= COSTS_N_INSNS (1 + (mode
== DFmode
));
10811 *cost
= COSTS_N_INSNS (4);
10817 && TARGET_HARD_FLOAT
10818 && (VALID_NEON_DREG_MODE (mode
) || VALID_NEON_QREG_MODE (mode
))
10819 && neon_immediate_valid_for_move (x
, mode
, NULL
, NULL
))
10820 *cost
= COSTS_N_INSNS (1);
10822 *cost
= COSTS_N_INSNS (4);
10827 /* When optimizing for size, we prefer constant pool entries to
10828 MOVW/MOVT pairs, so bump the cost of these slightly. */
10835 *cost
+= extra_cost
->alu
.clz
;
10839 if (XEXP (x
, 1) == const0_rtx
)
10842 *cost
+= extra_cost
->alu
.log_shift
;
10843 *cost
+= rtx_cost (XEXP (x
, 0), mode
, code
, 0, speed_p
);
10846 /* Fall through. */
10850 *cost
+= COSTS_N_INSNS (1);
10854 if (GET_CODE (XEXP (x
, 0)) == ASHIFTRT
10855 && CONST_INT_P (XEXP (XEXP (x
, 0), 1))
10856 && INTVAL (XEXP (XEXP (x
, 0), 1)) == 32
10857 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == MULT
10858 && ((GET_CODE (XEXP (XEXP (XEXP (x
, 0), 0), 0)) == SIGN_EXTEND
10859 && GET_CODE (XEXP (XEXP (XEXP (x
, 0), 0), 1)) == SIGN_EXTEND
)
10860 || (GET_CODE (XEXP (XEXP (XEXP (x
, 0), 0), 0)) == ZERO_EXTEND
10861 && (GET_CODE (XEXP (XEXP (XEXP (x
, 0), 0), 1))
10865 *cost
+= extra_cost
->mult
[1].extend
;
10866 *cost
+= (rtx_cost (XEXP (XEXP (XEXP (x
, 0), 0), 0), VOIDmode
,
10867 ZERO_EXTEND
, 0, speed_p
)
10868 + rtx_cost (XEXP (XEXP (XEXP (x
, 0), 0), 1), VOIDmode
,
10869 ZERO_EXTEND
, 0, speed_p
));
10872 *cost
= LIBCALL_COST (1);
10875 case UNSPEC_VOLATILE
:
10877 return arm_unspec_cost (x
, outer_code
, speed_p
, cost
);
10880 /* Reading the PC is like reading any other register. Writing it
10881 is more expensive, but we take that into account elsewhere. */
10886 /* TODO: Simple zero_extract of bottom bits using AND. */
10887 /* Fall through. */
10891 && CONST_INT_P (XEXP (x
, 1))
10892 && CONST_INT_P (XEXP (x
, 2)))
10895 *cost
+= extra_cost
->alu
.bfx
;
10896 *cost
+= rtx_cost (XEXP (x
, 0), mode
, code
, 0, speed_p
);
10899 /* Without UBFX/SBFX, need to resort to shift operations. */
10900 *cost
+= COSTS_N_INSNS (1);
10902 *cost
+= 2 * extra_cost
->alu
.shift
;
10903 *cost
+= rtx_cost (XEXP (x
, 0), mode
, ASHIFT
, 0, speed_p
);
10907 if (TARGET_HARD_FLOAT
)
10910 *cost
+= extra_cost
->fp
[mode
== DFmode
].widen
;
10912 && GET_MODE (XEXP (x
, 0)) == HFmode
)
10914 /* Pre v8, widening HF->DF is a two-step process, first
10915 widening to SFmode. */
10916 *cost
+= COSTS_N_INSNS (1);
10918 *cost
+= extra_cost
->fp
[0].widen
;
10920 *cost
+= rtx_cost (XEXP (x
, 0), VOIDmode
, code
, 0, speed_p
);
10924 *cost
= LIBCALL_COST (1);
10927 case FLOAT_TRUNCATE
:
10928 if (TARGET_HARD_FLOAT
)
10931 *cost
+= extra_cost
->fp
[mode
== DFmode
].narrow
;
10932 *cost
+= rtx_cost (XEXP (x
, 0), VOIDmode
, code
, 0, speed_p
);
10934 /* Vector modes? */
10936 *cost
= LIBCALL_COST (1);
10940 if (TARGET_32BIT
&& TARGET_HARD_FLOAT
&& TARGET_FMA
)
10942 rtx op0
= XEXP (x
, 0);
10943 rtx op1
= XEXP (x
, 1);
10944 rtx op2
= XEXP (x
, 2);
10947 /* vfms or vfnma. */
10948 if (GET_CODE (op0
) == NEG
)
10949 op0
= XEXP (op0
, 0);
10951 /* vfnms or vfnma. */
10952 if (GET_CODE (op2
) == NEG
)
10953 op2
= XEXP (op2
, 0);
10955 *cost
+= rtx_cost (op0
, mode
, FMA
, 0, speed_p
);
10956 *cost
+= rtx_cost (op1
, mode
, FMA
, 1, speed_p
);
10957 *cost
+= rtx_cost (op2
, mode
, FMA
, 2, speed_p
);
10960 *cost
+= extra_cost
->fp
[mode
==DFmode
].fma
;
10965 *cost
= LIBCALL_COST (3);
10970 if (TARGET_HARD_FLOAT
)
10972 /* The *combine_vcvtf2i reduces a vmul+vcvt into
10973 a vcvt fixed-point conversion. */
10974 if (code
== FIX
&& mode
== SImode
10975 && GET_CODE (XEXP (x
, 0)) == FIX
10976 && GET_MODE (XEXP (x
, 0)) == SFmode
10977 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == MULT
10978 && vfp3_const_double_for_bits (XEXP (XEXP (XEXP (x
, 0), 0), 1))
10982 *cost
+= extra_cost
->fp
[0].toint
;
10984 *cost
+= rtx_cost (XEXP (XEXP (XEXP (x
, 0), 0), 0), mode
,
10989 if (GET_MODE_CLASS (mode
) == MODE_INT
)
10991 mode
= GET_MODE (XEXP (x
, 0));
10993 *cost
+= extra_cost
->fp
[mode
== DFmode
].toint
;
10994 /* Strip of the 'cost' of rounding towards zero. */
10995 if (GET_CODE (XEXP (x
, 0)) == FIX
)
10996 *cost
+= rtx_cost (XEXP (XEXP (x
, 0), 0), mode
, code
,
10999 *cost
+= rtx_cost (XEXP (x
, 0), mode
, code
, 0, speed_p
);
11000 /* ??? Increase the cost to deal with transferring from
11001 FP -> CORE registers? */
11004 else if (GET_MODE_CLASS (mode
) == MODE_FLOAT
11008 *cost
+= extra_cost
->fp
[mode
== DFmode
].roundint
;
11011 /* Vector costs? */
11013 *cost
= LIBCALL_COST (1);
11017 case UNSIGNED_FLOAT
:
11018 if (TARGET_HARD_FLOAT
)
11020 /* ??? Increase the cost to deal with transferring from CORE
11021 -> FP registers? */
11023 *cost
+= extra_cost
->fp
[mode
== DFmode
].fromint
;
11026 *cost
= LIBCALL_COST (1);
11034 /* Just a guess. Guess number of instructions in the asm
11035 plus one insn per input. Always a minimum of COSTS_N_INSNS (1)
11036 though (see PR60663). */
11037 int asm_length
= MAX (1, asm_str_count (ASM_OPERANDS_TEMPLATE (x
)));
11038 int num_operands
= ASM_OPERANDS_INPUT_LENGTH (x
);
11040 *cost
= COSTS_N_INSNS (asm_length
+ num_operands
);
11044 if (mode
!= VOIDmode
)
11045 *cost
= COSTS_N_INSNS (ARM_NUM_REGS (mode
));
11047 *cost
= COSTS_N_INSNS (4); /* Who knows? */
11052 #undef HANDLE_NARROW_SHIFT_ARITH
11054 /* RTX costs entry point. */
11057 arm_rtx_costs (rtx x
, machine_mode mode ATTRIBUTE_UNUSED
, int outer_code
,
11058 int opno ATTRIBUTE_UNUSED
, int *total
, bool speed
)
11061 int code
= GET_CODE (x
);
11062 gcc_assert (current_tune
->insn_extra_cost
);
11064 result
= arm_rtx_costs_internal (x
, (enum rtx_code
) code
,
11065 (enum rtx_code
) outer_code
,
11066 current_tune
->insn_extra_cost
,
11069 if (dump_file
&& arm_verbose_cost
)
11071 print_rtl_single (dump_file
, x
);
11072 fprintf (dump_file
, "\n%s cost: %d (%s)\n", speed
? "Hot" : "Cold",
11073 *total
, result
? "final" : "partial");
11078 /* All address computations that can be done are free, but rtx cost returns
11079 the same for practically all of them. So we weight the different types
11080 of address here in the order (most pref first):
11081 PRE/POST_INC/DEC, SHIFT or NON-INT sum, INT sum, REG, MEM or LABEL. */
11083 arm_arm_address_cost (rtx x
)
11085 enum rtx_code c
= GET_CODE (x
);
11087 if (c
== PRE_INC
|| c
== PRE_DEC
|| c
== POST_INC
|| c
== POST_DEC
)
11089 if (c
== MEM
|| c
== LABEL_REF
|| c
== SYMBOL_REF
)
11094 if (CONST_INT_P (XEXP (x
, 1)))
11097 if (ARITHMETIC_P (XEXP (x
, 0)) || ARITHMETIC_P (XEXP (x
, 1)))
11107 arm_thumb_address_cost (rtx x
)
11109 enum rtx_code c
= GET_CODE (x
);
11114 && REG_P (XEXP (x
, 0))
11115 && CONST_INT_P (XEXP (x
, 1)))
11122 arm_address_cost (rtx x
, machine_mode mode ATTRIBUTE_UNUSED
,
11123 addr_space_t as ATTRIBUTE_UNUSED
, bool speed ATTRIBUTE_UNUSED
)
11125 return TARGET_32BIT
? arm_arm_address_cost (x
) : arm_thumb_address_cost (x
);
11128 /* Adjust cost hook for XScale. */
11130 xscale_sched_adjust_cost (rtx_insn
*insn
, int dep_type
, rtx_insn
*dep
,
11133 /* Some true dependencies can have a higher cost depending
11134 on precisely how certain input operands are used. */
11136 && recog_memoized (insn
) >= 0
11137 && recog_memoized (dep
) >= 0)
11139 int shift_opnum
= get_attr_shift (insn
);
11140 enum attr_type attr_type
= get_attr_type (dep
);
11142 /* If nonzero, SHIFT_OPNUM contains the operand number of a shifted
11143 operand for INSN. If we have a shifted input operand and the
11144 instruction we depend on is another ALU instruction, then we may
11145 have to account for an additional stall. */
11146 if (shift_opnum
!= 0
11147 && (attr_type
== TYPE_ALU_SHIFT_IMM
11148 || attr_type
== TYPE_ALUS_SHIFT_IMM
11149 || attr_type
== TYPE_LOGIC_SHIFT_IMM
11150 || attr_type
== TYPE_LOGICS_SHIFT_IMM
11151 || attr_type
== TYPE_ALU_SHIFT_REG
11152 || attr_type
== TYPE_ALUS_SHIFT_REG
11153 || attr_type
== TYPE_LOGIC_SHIFT_REG
11154 || attr_type
== TYPE_LOGICS_SHIFT_REG
11155 || attr_type
== TYPE_MOV_SHIFT
11156 || attr_type
== TYPE_MVN_SHIFT
11157 || attr_type
== TYPE_MOV_SHIFT_REG
11158 || attr_type
== TYPE_MVN_SHIFT_REG
))
11160 rtx shifted_operand
;
11163 /* Get the shifted operand. */
11164 extract_insn (insn
);
11165 shifted_operand
= recog_data
.operand
[shift_opnum
];
11167 /* Iterate over all the operands in DEP. If we write an operand
11168 that overlaps with SHIFTED_OPERAND, then we have increase the
11169 cost of this dependency. */
11170 extract_insn (dep
);
11171 preprocess_constraints (dep
);
11172 for (opno
= 0; opno
< recog_data
.n_operands
; opno
++)
11174 /* We can ignore strict inputs. */
11175 if (recog_data
.operand_type
[opno
] == OP_IN
)
11178 if (reg_overlap_mentioned_p (recog_data
.operand
[opno
],
11190 /* Adjust cost hook for Cortex A9. */
11192 cortex_a9_sched_adjust_cost (rtx_insn
*insn
, int dep_type
, rtx_insn
*dep
,
11202 case REG_DEP_OUTPUT
:
11203 if (recog_memoized (insn
) >= 0
11204 && recog_memoized (dep
) >= 0)
11206 if (GET_CODE (PATTERN (insn
)) == SET
)
11209 (GET_MODE (SET_DEST (PATTERN (insn
)))) == MODE_FLOAT
11211 (GET_MODE (SET_SRC (PATTERN (insn
)))) == MODE_FLOAT
)
11213 enum attr_type attr_type_insn
= get_attr_type (insn
);
11214 enum attr_type attr_type_dep
= get_attr_type (dep
);
11216 /* By default all dependencies of the form
11219 have an extra latency of 1 cycle because
11220 of the input and output dependency in this
11221 case. However this gets modeled as an true
11222 dependency and hence all these checks. */
11223 if (REG_P (SET_DEST (PATTERN (insn
)))
11224 && reg_set_p (SET_DEST (PATTERN (insn
)), dep
))
11226 /* FMACS is a special case where the dependent
11227 instruction can be issued 3 cycles before
11228 the normal latency in case of an output
11230 if ((attr_type_insn
== TYPE_FMACS
11231 || attr_type_insn
== TYPE_FMACD
)
11232 && (attr_type_dep
== TYPE_FMACS
11233 || attr_type_dep
== TYPE_FMACD
))
11235 if (dep_type
== REG_DEP_OUTPUT
)
11236 *cost
= insn_default_latency (dep
) - 3;
11238 *cost
= insn_default_latency (dep
);
11243 if (dep_type
== REG_DEP_OUTPUT
)
11244 *cost
= insn_default_latency (dep
) + 1;
11246 *cost
= insn_default_latency (dep
);
11256 gcc_unreachable ();
11262 /* Adjust cost hook for FA726TE. */
11264 fa726te_sched_adjust_cost (rtx_insn
*insn
, int dep_type
, rtx_insn
*dep
,
11267 /* For FA726TE, true dependency on CPSR (i.e. set cond followed by predicated)
11268 have penalty of 3. */
11269 if (dep_type
== REG_DEP_TRUE
11270 && recog_memoized (insn
) >= 0
11271 && recog_memoized (dep
) >= 0
11272 && get_attr_conds (dep
) == CONDS_SET
)
11274 /* Use of carry (e.g. 64-bit arithmetic) in ALU: 3-cycle latency. */
11275 if (get_attr_conds (insn
) == CONDS_USE
11276 && get_attr_type (insn
) != TYPE_BRANCH
)
11282 if (GET_CODE (PATTERN (insn
)) == COND_EXEC
11283 || get_attr_conds (insn
) == CONDS_USE
)
11293 /* Implement TARGET_REGISTER_MOVE_COST.
11295 Moves between VFP_REGS and GENERAL_REGS are a single insn, but
11296 it is typically more expensive than a single memory access. We set
11297 the cost to less than two memory accesses so that floating
11298 point to integer conversion does not go through memory. */
11301 arm_register_move_cost (machine_mode mode ATTRIBUTE_UNUSED
,
11302 reg_class_t from
, reg_class_t to
)
11306 if ((IS_VFP_CLASS (from
) && !IS_VFP_CLASS (to
))
11307 || (!IS_VFP_CLASS (from
) && IS_VFP_CLASS (to
)))
11309 else if ((from
== IWMMXT_REGS
&& to
!= IWMMXT_REGS
)
11310 || (from
!= IWMMXT_REGS
&& to
== IWMMXT_REGS
))
11312 else if (from
== IWMMXT_GR_REGS
|| to
== IWMMXT_GR_REGS
)
11319 if (from
== HI_REGS
|| to
== HI_REGS
)
11326 /* Implement TARGET_MEMORY_MOVE_COST. */
11329 arm_memory_move_cost (machine_mode mode
, reg_class_t rclass
,
11330 bool in ATTRIBUTE_UNUSED
)
11336 if (GET_MODE_SIZE (mode
) < 4)
11339 return ((2 * GET_MODE_SIZE (mode
)) * (rclass
== LO_REGS
? 1 : 2));
11343 /* Vectorizer cost model implementation. */
11345 /* Implement targetm.vectorize.builtin_vectorization_cost. */
11347 arm_builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost
,
11349 int misalign ATTRIBUTE_UNUSED
)
11353 switch (type_of_cost
)
11356 return current_tune
->vec_costs
->scalar_stmt_cost
;
11359 return current_tune
->vec_costs
->scalar_load_cost
;
11362 return current_tune
->vec_costs
->scalar_store_cost
;
11365 return current_tune
->vec_costs
->vec_stmt_cost
;
11368 return current_tune
->vec_costs
->vec_align_load_cost
;
11371 return current_tune
->vec_costs
->vec_store_cost
;
11373 case vec_to_scalar
:
11374 return current_tune
->vec_costs
->vec_to_scalar_cost
;
11376 case scalar_to_vec
:
11377 return current_tune
->vec_costs
->scalar_to_vec_cost
;
11379 case unaligned_load
:
11380 case vector_gather_load
:
11381 return current_tune
->vec_costs
->vec_unalign_load_cost
;
11383 case unaligned_store
:
11384 case vector_scatter_store
:
11385 return current_tune
->vec_costs
->vec_unalign_store_cost
;
11387 case cond_branch_taken
:
11388 return current_tune
->vec_costs
->cond_taken_branch_cost
;
11390 case cond_branch_not_taken
:
11391 return current_tune
->vec_costs
->cond_not_taken_branch_cost
;
11394 case vec_promote_demote
:
11395 return current_tune
->vec_costs
->vec_stmt_cost
;
11397 case vec_construct
:
11398 elements
= TYPE_VECTOR_SUBPARTS (vectype
);
11399 return elements
/ 2 + 1;
11402 gcc_unreachable ();
11406 /* Implement targetm.vectorize.add_stmt_cost. */
11409 arm_add_stmt_cost (void *data
, int count
, enum vect_cost_for_stmt kind
,
11410 struct _stmt_vec_info
*stmt_info
, int misalign
,
11411 enum vect_cost_model_location where
)
11413 unsigned *cost
= (unsigned *) data
;
11414 unsigned retval
= 0;
11416 if (flag_vect_cost_model
)
11418 tree vectype
= stmt_info
? stmt_vectype (stmt_info
) : NULL_TREE
;
11419 int stmt_cost
= arm_builtin_vectorization_cost (kind
, vectype
, misalign
);
11421 /* Statements in an inner loop relative to the loop being
11422 vectorized are weighted more heavily. The value here is
11423 arbitrary and could potentially be improved with analysis. */
11424 if (where
== vect_body
&& stmt_info
&& stmt_in_inner_loop_p (stmt_info
))
11425 count
*= 50; /* FIXME. */
11427 retval
= (unsigned) (count
* stmt_cost
);
11428 cost
[where
] += retval
;
11434 /* Return true if and only if this insn can dual-issue only as older. */
11436 cortexa7_older_only (rtx_insn
*insn
)
11438 if (recog_memoized (insn
) < 0)
11441 switch (get_attr_type (insn
))
11443 case TYPE_ALU_DSP_REG
:
11444 case TYPE_ALU_SREG
:
11445 case TYPE_ALUS_SREG
:
11446 case TYPE_LOGIC_REG
:
11447 case TYPE_LOGICS_REG
:
11449 case TYPE_ADCS_REG
:
11454 case TYPE_SHIFT_IMM
:
11455 case TYPE_SHIFT_REG
:
11456 case TYPE_LOAD_BYTE
:
11459 case TYPE_FFARITHS
:
11461 case TYPE_FFARITHD
:
11479 case TYPE_F_STORES
:
11486 /* Return true if and only if this insn can dual-issue as younger. */
11488 cortexa7_younger (FILE *file
, int verbose
, rtx_insn
*insn
)
11490 if (recog_memoized (insn
) < 0)
11493 fprintf (file
, ";; not cortexa7_younger %d\n", INSN_UID (insn
));
11497 switch (get_attr_type (insn
))
11500 case TYPE_ALUS_IMM
:
11501 case TYPE_LOGIC_IMM
:
11502 case TYPE_LOGICS_IMM
:
11507 case TYPE_MOV_SHIFT
:
11508 case TYPE_MOV_SHIFT_REG
:
11518 /* Look for an instruction that can dual issue only as an older
11519 instruction, and move it in front of any instructions that can
11520 dual-issue as younger, while preserving the relative order of all
11521 other instructions in the ready list. This is a hueuristic to help
11522 dual-issue in later cycles, by postponing issue of more flexible
11523 instructions. This heuristic may affect dual issue opportunities
11524 in the current cycle. */
11526 cortexa7_sched_reorder (FILE *file
, int verbose
, rtx_insn
**ready
,
11527 int *n_readyp
, int clock
)
11530 int first_older_only
= -1, first_younger
= -1;
11534 ";; sched_reorder for cycle %d with %d insns in ready list\n",
11538 /* Traverse the ready list from the head (the instruction to issue
11539 first), and looking for the first instruction that can issue as
11540 younger and the first instruction that can dual-issue only as
11542 for (i
= *n_readyp
- 1; i
>= 0; i
--)
11544 rtx_insn
*insn
= ready
[i
];
11545 if (cortexa7_older_only (insn
))
11547 first_older_only
= i
;
11549 fprintf (file
, ";; reorder older found %d\n", INSN_UID (insn
));
11552 else if (cortexa7_younger (file
, verbose
, insn
) && first_younger
== -1)
11556 /* Nothing to reorder because either no younger insn found or insn
11557 that can dual-issue only as older appears before any insn that
11558 can dual-issue as younger. */
11559 if (first_younger
== -1)
11562 fprintf (file
, ";; sched_reorder nothing to reorder as no younger\n");
11566 /* Nothing to reorder because no older-only insn in the ready list. */
11567 if (first_older_only
== -1)
11570 fprintf (file
, ";; sched_reorder nothing to reorder as no older_only\n");
11574 /* Move first_older_only insn before first_younger. */
11576 fprintf (file
, ";; cortexa7_sched_reorder insn %d before %d\n",
11577 INSN_UID(ready
[first_older_only
]),
11578 INSN_UID(ready
[first_younger
]));
11579 rtx_insn
*first_older_only_insn
= ready
[first_older_only
];
11580 for (i
= first_older_only
; i
< first_younger
; i
++)
11582 ready
[i
] = ready
[i
+1];
11585 ready
[i
] = first_older_only_insn
;
11589 /* Implement TARGET_SCHED_REORDER. */
11591 arm_sched_reorder (FILE *file
, int verbose
, rtx_insn
**ready
, int *n_readyp
,
11596 case TARGET_CPU_cortexa7
:
11597 cortexa7_sched_reorder (file
, verbose
, ready
, n_readyp
, clock
);
11600 /* Do nothing for other cores. */
11604 return arm_issue_rate ();
11607 /* This function implements the target macro TARGET_SCHED_ADJUST_COST.
11608 It corrects the value of COST based on the relationship between
11609 INSN and DEP through the dependence LINK. It returns the new
11610 value. There is a per-core adjust_cost hook to adjust scheduler costs
11611 and the per-core hook can choose to completely override the generic
11612 adjust_cost function. Only put bits of code into arm_adjust_cost that
11613 are common across all cores. */
11615 arm_adjust_cost (rtx_insn
*insn
, int dep_type
, rtx_insn
*dep
, int cost
,
11620 /* When generating Thumb-1 code, we want to place flag-setting operations
11621 close to a conditional branch which depends on them, so that we can
11622 omit the comparison. */
11625 && recog_memoized (insn
) == CODE_FOR_cbranchsi4_insn
11626 && recog_memoized (dep
) >= 0
11627 && get_attr_conds (dep
) == CONDS_SET
)
11630 if (current_tune
->sched_adjust_cost
!= NULL
)
11632 if (!current_tune
->sched_adjust_cost (insn
, dep_type
, dep
, &cost
))
11636 /* XXX Is this strictly true? */
11637 if (dep_type
== REG_DEP_ANTI
11638 || dep_type
== REG_DEP_OUTPUT
)
11641 /* Call insns don't incur a stall, even if they follow a load. */
11646 if ((i_pat
= single_set (insn
)) != NULL
11647 && MEM_P (SET_SRC (i_pat
))
11648 && (d_pat
= single_set (dep
)) != NULL
11649 && MEM_P (SET_DEST (d_pat
)))
11651 rtx src_mem
= XEXP (SET_SRC (i_pat
), 0);
11652 /* This is a load after a store, there is no conflict if the load reads
11653 from a cached area. Assume that loads from the stack, and from the
11654 constant pool are cached, and that others will miss. This is a
11657 if ((GET_CODE (src_mem
) == SYMBOL_REF
11658 && CONSTANT_POOL_ADDRESS_P (src_mem
))
11659 || reg_mentioned_p (stack_pointer_rtx
, src_mem
)
11660 || reg_mentioned_p (frame_pointer_rtx
, src_mem
)
11661 || reg_mentioned_p (hard_frame_pointer_rtx
, src_mem
))
11669 arm_max_conditional_execute (void)
11671 return max_insns_skipped
;
11675 arm_default_branch_cost (bool speed_p
, bool predictable_p ATTRIBUTE_UNUSED
)
11678 return (TARGET_THUMB2
&& !speed_p
) ? 1 : 4;
11680 return (optimize
> 0) ? 2 : 0;
11684 arm_cortex_a5_branch_cost (bool speed_p
, bool predictable_p
)
11686 return speed_p
? 0 : arm_default_branch_cost (speed_p
, predictable_p
);
11689 /* Thumb-2 branches are relatively cheap on Cortex-M processors ("1 + P cycles"
11690 on Cortex-M4, where P varies from 1 to 3 according to some criteria), since
11691 sequences of non-executed instructions in IT blocks probably take the same
11692 amount of time as executed instructions (and the IT instruction itself takes
11693 space in icache). This function was experimentally determined to give good
11694 results on a popular embedded benchmark. */
11697 arm_cortex_m_branch_cost (bool speed_p
, bool predictable_p
)
11699 return (TARGET_32BIT
&& speed_p
) ? 1
11700 : arm_default_branch_cost (speed_p
, predictable_p
);
11704 arm_cortex_m7_branch_cost (bool speed_p
, bool predictable_p
)
11706 return speed_p
? 0 : arm_default_branch_cost (speed_p
, predictable_p
);
11709 static bool fp_consts_inited
= false;
11711 static REAL_VALUE_TYPE value_fp0
;
11714 init_fp_table (void)
11718 r
= REAL_VALUE_ATOF ("0", DFmode
);
11720 fp_consts_inited
= true;
11723 /* Return TRUE if rtx X is a valid immediate FP constant. */
11725 arm_const_double_rtx (rtx x
)
11727 const REAL_VALUE_TYPE
*r
;
11729 if (!fp_consts_inited
)
11732 r
= CONST_DOUBLE_REAL_VALUE (x
);
11733 if (REAL_VALUE_MINUS_ZERO (*r
))
11736 if (real_equal (r
, &value_fp0
))
11742 /* VFPv3 has a fairly wide range of representable immediates, formed from
11743 "quarter-precision" floating-point values. These can be evaluated using this
11744 formula (with ^ for exponentiation):
11748 Where 's' is a sign bit (0/1), 'n' and 'r' are integers such that
11749 16 <= n <= 31 and 0 <= r <= 7.
11751 These values are mapped onto an 8-bit integer ABCDEFGH s.t.
11753 - A (most-significant) is the sign bit.
11754 - BCD are the exponent (encoded as r XOR 3).
11755 - EFGH are the mantissa (encoded as n - 16).
11758 /* Return an integer index for a VFPv3 immediate operand X suitable for the
11759 fconst[sd] instruction, or -1 if X isn't suitable. */
11761 vfp3_const_double_index (rtx x
)
11763 REAL_VALUE_TYPE r
, m
;
11764 int sign
, exponent
;
11765 unsigned HOST_WIDE_INT mantissa
, mant_hi
;
11766 unsigned HOST_WIDE_INT mask
;
11767 int point_pos
= 2 * HOST_BITS_PER_WIDE_INT
- 1;
11770 if (!TARGET_VFP3
|| !CONST_DOUBLE_P (x
))
11773 r
= *CONST_DOUBLE_REAL_VALUE (x
);
11775 /* We can't represent these things, so detect them first. */
11776 if (REAL_VALUE_ISINF (r
) || REAL_VALUE_ISNAN (r
) || REAL_VALUE_MINUS_ZERO (r
))
11779 /* Extract sign, exponent and mantissa. */
11780 sign
= REAL_VALUE_NEGATIVE (r
) ? 1 : 0;
11781 r
= real_value_abs (&r
);
11782 exponent
= REAL_EXP (&r
);
11783 /* For the mantissa, we expand into two HOST_WIDE_INTS, apart from the
11784 highest (sign) bit, with a fixed binary point at bit point_pos.
11785 WARNING: If there's ever a VFP version which uses more than 2 * H_W_I - 1
11786 bits for the mantissa, this may fail (low bits would be lost). */
11787 real_ldexp (&m
, &r
, point_pos
- exponent
);
11788 wide_int w
= real_to_integer (&m
, &fail
, HOST_BITS_PER_WIDE_INT
* 2);
11789 mantissa
= w
.elt (0);
11790 mant_hi
= w
.elt (1);
11792 /* If there are bits set in the low part of the mantissa, we can't
11793 represent this value. */
11797 /* Now make it so that mantissa contains the most-significant bits, and move
11798 the point_pos to indicate that the least-significant bits have been
11800 point_pos
-= HOST_BITS_PER_WIDE_INT
;
11801 mantissa
= mant_hi
;
11803 /* We can permit four significant bits of mantissa only, plus a high bit
11804 which is always 1. */
11805 mask
= (HOST_WIDE_INT_1U
<< (point_pos
- 5)) - 1;
11806 if ((mantissa
& mask
) != 0)
11809 /* Now we know the mantissa is in range, chop off the unneeded bits. */
11810 mantissa
>>= point_pos
- 5;
11812 /* The mantissa may be zero. Disallow that case. (It's possible to load the
11813 floating-point immediate zero with Neon using an integer-zero load, but
11814 that case is handled elsewhere.) */
11818 gcc_assert (mantissa
>= 16 && mantissa
<= 31);
11820 /* The value of 5 here would be 4 if GCC used IEEE754-like encoding (where
11821 normalized significands are in the range [1, 2). (Our mantissa is shifted
11822 left 4 places at this point relative to normalized IEEE754 values). GCC
11823 internally uses [0.5, 1) (see real.c), so the exponent returned from
11824 REAL_EXP must be altered. */
11825 exponent
= 5 - exponent
;
11827 if (exponent
< 0 || exponent
> 7)
11830 /* Sign, mantissa and exponent are now in the correct form to plug into the
11831 formula described in the comment above. */
11832 return (sign
<< 7) | ((exponent
^ 3) << 4) | (mantissa
- 16);
11835 /* Return TRUE if rtx X is a valid immediate VFPv3 constant. */
11837 vfp3_const_double_rtx (rtx x
)
11842 return vfp3_const_double_index (x
) != -1;
11845 /* Recognize immediates which can be used in various Neon instructions. Legal
11846 immediates are described by the following table (for VMVN variants, the
11847 bitwise inverse of the constant shown is recognized. In either case, VMOV
11848 is output and the correct instruction to use for a given constant is chosen
11849 by the assembler). The constant shown is replicated across all elements of
11850 the destination vector.
11852 insn elems variant constant (binary)
11853 ---- ----- ------- -----------------
11854 vmov i32 0 00000000 00000000 00000000 abcdefgh
11855 vmov i32 1 00000000 00000000 abcdefgh 00000000
11856 vmov i32 2 00000000 abcdefgh 00000000 00000000
11857 vmov i32 3 abcdefgh 00000000 00000000 00000000
11858 vmov i16 4 00000000 abcdefgh
11859 vmov i16 5 abcdefgh 00000000
11860 vmvn i32 6 00000000 00000000 00000000 abcdefgh
11861 vmvn i32 7 00000000 00000000 abcdefgh 00000000
11862 vmvn i32 8 00000000 abcdefgh 00000000 00000000
11863 vmvn i32 9 abcdefgh 00000000 00000000 00000000
11864 vmvn i16 10 00000000 abcdefgh
11865 vmvn i16 11 abcdefgh 00000000
11866 vmov i32 12 00000000 00000000 abcdefgh 11111111
11867 vmvn i32 13 00000000 00000000 abcdefgh 11111111
11868 vmov i32 14 00000000 abcdefgh 11111111 11111111
11869 vmvn i32 15 00000000 abcdefgh 11111111 11111111
11870 vmov i8 16 abcdefgh
11871 vmov i64 17 aaaaaaaa bbbbbbbb cccccccc dddddddd
11872 eeeeeeee ffffffff gggggggg hhhhhhhh
11873 vmov f32 18 aBbbbbbc defgh000 00000000 00000000
11874 vmov f32 19 00000000 00000000 00000000 00000000
11876 For case 18, B = !b. Representable values are exactly those accepted by
11877 vfp3_const_double_index, but are output as floating-point numbers rather
11880 For case 19, we will change it to vmov.i32 when assembling.
11882 Variants 0-5 (inclusive) may also be used as immediates for the second
11883 operand of VORR/VBIC instructions.
11885 The INVERSE argument causes the bitwise inverse of the given operand to be
11886 recognized instead (used for recognizing legal immediates for the VAND/VORN
11887 pseudo-instructions). If INVERSE is true, the value placed in *MODCONST is
11888 *not* inverted (i.e. the pseudo-instruction forms vand/vorn should still be
11889 output, rather than the real insns vbic/vorr).
11891 INVERSE makes no difference to the recognition of float vectors.
11893 The return value is the variant of immediate as shown in the above table, or
11894 -1 if the given value doesn't match any of the listed patterns.
11897 neon_valid_immediate (rtx op
, machine_mode mode
, int inverse
,
11898 rtx
*modconst
, int *elementwidth
)
11900 #define CHECK(STRIDE, ELSIZE, CLASS, TEST) \
11902 for (i = 0; i < idx; i += (STRIDE)) \
11907 immtype = (CLASS); \
11908 elsize = (ELSIZE); \
11912 unsigned int i
, elsize
= 0, idx
= 0, n_elts
;
11913 unsigned int innersize
;
11914 unsigned char bytes
[16];
11915 int immtype
= -1, matches
;
11916 unsigned int invmask
= inverse
? 0xff : 0;
11917 bool vector
= GET_CODE (op
) == CONST_VECTOR
;
11920 n_elts
= CONST_VECTOR_NUNITS (op
);
11924 if (mode
== VOIDmode
)
11928 innersize
= GET_MODE_UNIT_SIZE (mode
);
11930 /* Vectors of float constants. */
11931 if (GET_MODE_CLASS (mode
) == MODE_VECTOR_FLOAT
)
11933 rtx el0
= CONST_VECTOR_ELT (op
, 0);
11935 if (!vfp3_const_double_rtx (el0
) && el0
!= CONST0_RTX (GET_MODE (el0
)))
11938 /* FP16 vectors cannot be represented. */
11939 if (GET_MODE_INNER (mode
) == HFmode
)
11942 /* All elements in the vector must be the same. Note that 0.0 and -0.0
11943 are distinct in this context. */
11944 if (!const_vec_duplicate_p (op
))
11948 *modconst
= CONST_VECTOR_ELT (op
, 0);
11953 if (el0
== CONST0_RTX (GET_MODE (el0
)))
11959 /* The tricks done in the code below apply for little-endian vector layout.
11960 For big-endian vectors only allow vectors of the form { a, a, a..., a }.
11961 FIXME: Implement logic for big-endian vectors. */
11962 if (BYTES_BIG_ENDIAN
&& vector
&& !const_vec_duplicate_p (op
))
11965 /* Splat vector constant out into a byte vector. */
11966 for (i
= 0; i
< n_elts
; i
++)
11968 rtx el
= vector
? CONST_VECTOR_ELT (op
, i
) : op
;
11969 unsigned HOST_WIDE_INT elpart
;
11971 gcc_assert (CONST_INT_P (el
));
11972 elpart
= INTVAL (el
);
11974 for (unsigned int byte
= 0; byte
< innersize
; byte
++)
11976 bytes
[idx
++] = (elpart
& 0xff) ^ invmask
;
11977 elpart
>>= BITS_PER_UNIT
;
11981 /* Sanity check. */
11982 gcc_assert (idx
== GET_MODE_SIZE (mode
));
11986 CHECK (4, 32, 0, bytes
[i
] == bytes
[0] && bytes
[i
+ 1] == 0
11987 && bytes
[i
+ 2] == 0 && bytes
[i
+ 3] == 0);
11989 CHECK (4, 32, 1, bytes
[i
] == 0 && bytes
[i
+ 1] == bytes
[1]
11990 && bytes
[i
+ 2] == 0 && bytes
[i
+ 3] == 0);
11992 CHECK (4, 32, 2, bytes
[i
] == 0 && bytes
[i
+ 1] == 0
11993 && bytes
[i
+ 2] == bytes
[2] && bytes
[i
+ 3] == 0);
11995 CHECK (4, 32, 3, bytes
[i
] == 0 && bytes
[i
+ 1] == 0
11996 && bytes
[i
+ 2] == 0 && bytes
[i
+ 3] == bytes
[3]);
11998 CHECK (2, 16, 4, bytes
[i
] == bytes
[0] && bytes
[i
+ 1] == 0);
12000 CHECK (2, 16, 5, bytes
[i
] == 0 && bytes
[i
+ 1] == bytes
[1]);
12002 CHECK (4, 32, 6, bytes
[i
] == bytes
[0] && bytes
[i
+ 1] == 0xff
12003 && bytes
[i
+ 2] == 0xff && bytes
[i
+ 3] == 0xff);
12005 CHECK (4, 32, 7, bytes
[i
] == 0xff && bytes
[i
+ 1] == bytes
[1]
12006 && bytes
[i
+ 2] == 0xff && bytes
[i
+ 3] == 0xff);
12008 CHECK (4, 32, 8, bytes
[i
] == 0xff && bytes
[i
+ 1] == 0xff
12009 && bytes
[i
+ 2] == bytes
[2] && bytes
[i
+ 3] == 0xff);
12011 CHECK (4, 32, 9, bytes
[i
] == 0xff && bytes
[i
+ 1] == 0xff
12012 && bytes
[i
+ 2] == 0xff && bytes
[i
+ 3] == bytes
[3]);
12014 CHECK (2, 16, 10, bytes
[i
] == bytes
[0] && bytes
[i
+ 1] == 0xff);
12016 CHECK (2, 16, 11, bytes
[i
] == 0xff && bytes
[i
+ 1] == bytes
[1]);
12018 CHECK (4, 32, 12, bytes
[i
] == 0xff && bytes
[i
+ 1] == bytes
[1]
12019 && bytes
[i
+ 2] == 0 && bytes
[i
+ 3] == 0);
12021 CHECK (4, 32, 13, bytes
[i
] == 0 && bytes
[i
+ 1] == bytes
[1]
12022 && bytes
[i
+ 2] == 0xff && bytes
[i
+ 3] == 0xff);
12024 CHECK (4, 32, 14, bytes
[i
] == 0xff && bytes
[i
+ 1] == 0xff
12025 && bytes
[i
+ 2] == bytes
[2] && bytes
[i
+ 3] == 0);
12027 CHECK (4, 32, 15, bytes
[i
] == 0 && bytes
[i
+ 1] == 0
12028 && bytes
[i
+ 2] == bytes
[2] && bytes
[i
+ 3] == 0xff);
12030 CHECK (1, 8, 16, bytes
[i
] == bytes
[0]);
12032 CHECK (1, 64, 17, (bytes
[i
] == 0 || bytes
[i
] == 0xff)
12033 && bytes
[i
] == bytes
[(i
+ 8) % idx
]);
12041 *elementwidth
= elsize
;
12045 unsigned HOST_WIDE_INT imm
= 0;
12047 /* Un-invert bytes of recognized vector, if necessary. */
12049 for (i
= 0; i
< idx
; i
++)
12050 bytes
[i
] ^= invmask
;
12054 /* FIXME: Broken on 32-bit H_W_I hosts. */
12055 gcc_assert (sizeof (HOST_WIDE_INT
) == 8);
12057 for (i
= 0; i
< 8; i
++)
12058 imm
|= (unsigned HOST_WIDE_INT
) (bytes
[i
] ? 0xff : 0)
12059 << (i
* BITS_PER_UNIT
);
12061 *modconst
= GEN_INT (imm
);
12065 unsigned HOST_WIDE_INT imm
= 0;
12067 for (i
= 0; i
< elsize
/ BITS_PER_UNIT
; i
++)
12068 imm
|= (unsigned HOST_WIDE_INT
) bytes
[i
] << (i
* BITS_PER_UNIT
);
12070 *modconst
= GEN_INT (imm
);
12078 /* Return TRUE if rtx X is legal for use as either a Neon VMOV (or, implicitly,
12079 VMVN) immediate. Write back width per element to *ELEMENTWIDTH (or zero for
12080 float elements), and a modified constant (whatever should be output for a
12081 VMOV) in *MODCONST. */
12084 neon_immediate_valid_for_move (rtx op
, machine_mode mode
,
12085 rtx
*modconst
, int *elementwidth
)
12089 int retval
= neon_valid_immediate (op
, mode
, 0, &tmpconst
, &tmpwidth
);
12095 *modconst
= tmpconst
;
12098 *elementwidth
= tmpwidth
;
12103 /* Return TRUE if rtx X is legal for use in a VORR or VBIC instruction. If
12104 the immediate is valid, write a constant suitable for using as an operand
12105 to VORR/VBIC/VAND/VORN to *MODCONST and the corresponding element width to
12106 *ELEMENTWIDTH. See neon_valid_immediate for description of INVERSE. */
12109 neon_immediate_valid_for_logic (rtx op
, machine_mode mode
, int inverse
,
12110 rtx
*modconst
, int *elementwidth
)
12114 int retval
= neon_valid_immediate (op
, mode
, inverse
, &tmpconst
, &tmpwidth
);
12116 if (retval
< 0 || retval
> 5)
12120 *modconst
= tmpconst
;
12123 *elementwidth
= tmpwidth
;
12128 /* Return TRUE if rtx OP is legal for use in a VSHR or VSHL instruction. If
12129 the immediate is valid, write a constant suitable for using as an operand
12130 to VSHR/VSHL to *MODCONST and the corresponding element width to
12131 *ELEMENTWIDTH. ISLEFTSHIFT is for determine left or right shift,
12132 because they have different limitations. */
12135 neon_immediate_valid_for_shift (rtx op
, machine_mode mode
,
12136 rtx
*modconst
, int *elementwidth
,
12139 unsigned int innersize
= GET_MODE_UNIT_SIZE (mode
);
12140 unsigned int n_elts
= CONST_VECTOR_NUNITS (op
), i
;
12141 unsigned HOST_WIDE_INT last_elt
= 0;
12142 unsigned HOST_WIDE_INT maxshift
;
12144 /* Split vector constant out into a byte vector. */
12145 for (i
= 0; i
< n_elts
; i
++)
12147 rtx el
= CONST_VECTOR_ELT (op
, i
);
12148 unsigned HOST_WIDE_INT elpart
;
12150 if (CONST_INT_P (el
))
12151 elpart
= INTVAL (el
);
12152 else if (CONST_DOUBLE_P (el
))
12155 gcc_unreachable ();
12157 if (i
!= 0 && elpart
!= last_elt
)
12163 /* Shift less than element size. */
12164 maxshift
= innersize
* 8;
12168 /* Left shift immediate value can be from 0 to <size>-1. */
12169 if (last_elt
>= maxshift
)
12174 /* Right shift immediate value can be from 1 to <size>. */
12175 if (last_elt
== 0 || last_elt
> maxshift
)
12180 *elementwidth
= innersize
* 8;
12183 *modconst
= CONST_VECTOR_ELT (op
, 0);
12188 /* Return a string suitable for output of Neon immediate logic operation
12192 neon_output_logic_immediate (const char *mnem
, rtx
*op2
, machine_mode mode
,
12193 int inverse
, int quad
)
12195 int width
, is_valid
;
12196 static char templ
[40];
12198 is_valid
= neon_immediate_valid_for_logic (*op2
, mode
, inverse
, op2
, &width
);
12200 gcc_assert (is_valid
!= 0);
12203 sprintf (templ
, "%s.i%d\t%%q0, %%2", mnem
, width
);
12205 sprintf (templ
, "%s.i%d\t%%P0, %%2", mnem
, width
);
12210 /* Return a string suitable for output of Neon immediate shift operation
12211 (VSHR or VSHL) MNEM. */
12214 neon_output_shift_immediate (const char *mnem
, char sign
, rtx
*op2
,
12215 machine_mode mode
, int quad
,
12218 int width
, is_valid
;
12219 static char templ
[40];
12221 is_valid
= neon_immediate_valid_for_shift (*op2
, mode
, op2
, &width
, isleftshift
);
12222 gcc_assert (is_valid
!= 0);
12225 sprintf (templ
, "%s.%c%d\t%%q0, %%q1, %%2", mnem
, sign
, width
);
12227 sprintf (templ
, "%s.%c%d\t%%P0, %%P1, %%2", mnem
, sign
, width
);
12232 /* Output a sequence of pairwise operations to implement a reduction.
12233 NOTE: We do "too much work" here, because pairwise operations work on two
12234 registers-worth of operands in one go. Unfortunately we can't exploit those
12235 extra calculations to do the full operation in fewer steps, I don't think.
12236 Although all vector elements of the result but the first are ignored, we
12237 actually calculate the same result in each of the elements. An alternative
12238 such as initially loading a vector with zero to use as each of the second
12239 operands would use up an additional register and take an extra instruction,
12240 for no particular gain. */
12243 neon_pairwise_reduce (rtx op0
, rtx op1
, machine_mode mode
,
12244 rtx (*reduc
) (rtx
, rtx
, rtx
))
12246 unsigned int i
, parts
= GET_MODE_SIZE (mode
) / GET_MODE_UNIT_SIZE (mode
);
12249 for (i
= parts
/ 2; i
>= 1; i
/= 2)
12251 rtx dest
= (i
== 1) ? op0
: gen_reg_rtx (mode
);
12252 emit_insn (reduc (dest
, tmpsum
, tmpsum
));
12257 /* If VALS is a vector constant that can be loaded into a register
12258 using VDUP, generate instructions to do so and return an RTX to
12259 assign to the register. Otherwise return NULL_RTX. */
12262 neon_vdup_constant (rtx vals
)
12264 machine_mode mode
= GET_MODE (vals
);
12265 machine_mode inner_mode
= GET_MODE_INNER (mode
);
12268 if (GET_CODE (vals
) != CONST_VECTOR
|| GET_MODE_SIZE (inner_mode
) > 4)
12271 if (!const_vec_duplicate_p (vals
, &x
))
12272 /* The elements are not all the same. We could handle repeating
12273 patterns of a mode larger than INNER_MODE here (e.g. int8x8_t
12274 {0, C, 0, C, 0, C, 0, C} which can be loaded using
12278 /* We can load this constant by using VDUP and a constant in a
12279 single ARM register. This will be cheaper than a vector
12282 x
= copy_to_mode_reg (inner_mode
, x
);
12283 return gen_vec_duplicate (mode
, x
);
12286 /* Generate code to load VALS, which is a PARALLEL containing only
12287 constants (for vec_init) or CONST_VECTOR, efficiently into a
12288 register. Returns an RTX to copy into the register, or NULL_RTX
12289 for a PARALLEL that can not be converted into a CONST_VECTOR. */
12292 neon_make_constant (rtx vals
)
12294 machine_mode mode
= GET_MODE (vals
);
12296 rtx const_vec
= NULL_RTX
;
12297 int n_elts
= GET_MODE_NUNITS (mode
);
12301 if (GET_CODE (vals
) == CONST_VECTOR
)
12303 else if (GET_CODE (vals
) == PARALLEL
)
12305 /* A CONST_VECTOR must contain only CONST_INTs and
12306 CONST_DOUBLEs, but CONSTANT_P allows more (e.g. SYMBOL_REF).
12307 Only store valid constants in a CONST_VECTOR. */
12308 for (i
= 0; i
< n_elts
; ++i
)
12310 rtx x
= XVECEXP (vals
, 0, i
);
12311 if (CONST_INT_P (x
) || CONST_DOUBLE_P (x
))
12314 if (n_const
== n_elts
)
12315 const_vec
= gen_rtx_CONST_VECTOR (mode
, XVEC (vals
, 0));
12318 gcc_unreachable ();
12320 if (const_vec
!= NULL
12321 && neon_immediate_valid_for_move (const_vec
, mode
, NULL
, NULL
))
12322 /* Load using VMOV. On Cortex-A8 this takes one cycle. */
12324 else if ((target
= neon_vdup_constant (vals
)) != NULL_RTX
)
12325 /* Loaded using VDUP. On Cortex-A8 the VDUP takes one NEON
12326 pipeline cycle; creating the constant takes one or two ARM
12327 pipeline cycles. */
12329 else if (const_vec
!= NULL_RTX
)
12330 /* Load from constant pool. On Cortex-A8 this takes two cycles
12331 (for either double or quad vectors). We can not take advantage
12332 of single-cycle VLD1 because we need a PC-relative addressing
12336 /* A PARALLEL containing something not valid inside CONST_VECTOR.
12337 We can not construct an initializer. */
12341 /* Initialize vector TARGET to VALS. */
12344 neon_expand_vector_init (rtx target
, rtx vals
)
12346 machine_mode mode
= GET_MODE (target
);
12347 machine_mode inner_mode
= GET_MODE_INNER (mode
);
12348 int n_elts
= GET_MODE_NUNITS (mode
);
12349 int n_var
= 0, one_var
= -1;
12350 bool all_same
= true;
12354 for (i
= 0; i
< n_elts
; ++i
)
12356 x
= XVECEXP (vals
, 0, i
);
12357 if (!CONSTANT_P (x
))
12358 ++n_var
, one_var
= i
;
12360 if (i
> 0 && !rtx_equal_p (x
, XVECEXP (vals
, 0, 0)))
12366 rtx constant
= neon_make_constant (vals
);
12367 if (constant
!= NULL_RTX
)
12369 emit_move_insn (target
, constant
);
12374 /* Splat a single non-constant element if we can. */
12375 if (all_same
&& GET_MODE_SIZE (inner_mode
) <= 4)
12377 x
= copy_to_mode_reg (inner_mode
, XVECEXP (vals
, 0, 0));
12378 emit_insn (gen_rtx_SET (target
, gen_vec_duplicate (mode
, x
)));
12382 /* One field is non-constant. Load constant then overwrite varying
12383 field. This is more efficient than using the stack. */
12386 rtx copy
= copy_rtx (vals
);
12387 rtx index
= GEN_INT (one_var
);
12389 /* Load constant part of vector, substitute neighboring value for
12390 varying element. */
12391 XVECEXP (copy
, 0, one_var
) = XVECEXP (vals
, 0, (one_var
+ 1) % n_elts
);
12392 neon_expand_vector_init (target
, copy
);
12394 /* Insert variable. */
12395 x
= copy_to_mode_reg (inner_mode
, XVECEXP (vals
, 0, one_var
));
12399 emit_insn (gen_neon_vset_lanev8qi (target
, x
, target
, index
));
12402 emit_insn (gen_neon_vset_lanev16qi (target
, x
, target
, index
));
12405 emit_insn (gen_neon_vset_lanev4hi (target
, x
, target
, index
));
12408 emit_insn (gen_neon_vset_lanev8hi (target
, x
, target
, index
));
12411 emit_insn (gen_neon_vset_lanev2si (target
, x
, target
, index
));
12414 emit_insn (gen_neon_vset_lanev4si (target
, x
, target
, index
));
12417 emit_insn (gen_neon_vset_lanev2sf (target
, x
, target
, index
));
12420 emit_insn (gen_neon_vset_lanev4sf (target
, x
, target
, index
));
12423 emit_insn (gen_neon_vset_lanev2di (target
, x
, target
, index
));
12426 gcc_unreachable ();
12431 /* Construct the vector in memory one field at a time
12432 and load the whole vector. */
12433 mem
= assign_stack_temp (mode
, GET_MODE_SIZE (mode
));
12434 for (i
= 0; i
< n_elts
; i
++)
12435 emit_move_insn (adjust_address_nv (mem
, inner_mode
,
12436 i
* GET_MODE_SIZE (inner_mode
)),
12437 XVECEXP (vals
, 0, i
));
12438 emit_move_insn (target
, mem
);
12441 /* Ensure OPERAND lies between LOW (inclusive) and HIGH (exclusive). Raise
12442 ERR if it doesn't. EXP indicates the source location, which includes the
12443 inlining history for intrinsics. */
12446 bounds_check (rtx operand
, HOST_WIDE_INT low
, HOST_WIDE_INT high
,
12447 const_tree exp
, const char *desc
)
12449 HOST_WIDE_INT lane
;
12451 gcc_assert (CONST_INT_P (operand
));
12453 lane
= INTVAL (operand
);
12455 if (lane
< low
|| lane
>= high
)
12458 error ("%K%s %wd out of range %wd - %wd",
12459 exp
, desc
, lane
, low
, high
- 1);
12461 error ("%s %wd out of range %wd - %wd", desc
, lane
, low
, high
- 1);
12465 /* Bounds-check lanes. */
12468 neon_lane_bounds (rtx operand
, HOST_WIDE_INT low
, HOST_WIDE_INT high
,
12471 bounds_check (operand
, low
, high
, exp
, "lane");
12474 /* Bounds-check constants. */
12477 arm_const_bounds (rtx operand
, HOST_WIDE_INT low
, HOST_WIDE_INT high
)
12479 bounds_check (operand
, low
, high
, NULL_TREE
, "constant");
12483 neon_element_bits (machine_mode mode
)
12485 return GET_MODE_UNIT_BITSIZE (mode
);
12489 /* Predicates for `match_operand' and `match_operator'. */
12491 /* Return TRUE if OP is a valid coprocessor memory address pattern.
12492 WB is true if full writeback address modes are allowed and is false
12493 if limited writeback address modes (POST_INC and PRE_DEC) are
12497 arm_coproc_mem_operand (rtx op
, bool wb
)
12501 /* Reject eliminable registers. */
12502 if (! (reload_in_progress
|| reload_completed
|| lra_in_progress
)
12503 && ( reg_mentioned_p (frame_pointer_rtx
, op
)
12504 || reg_mentioned_p (arg_pointer_rtx
, op
)
12505 || reg_mentioned_p (virtual_incoming_args_rtx
, op
)
12506 || reg_mentioned_p (virtual_outgoing_args_rtx
, op
)
12507 || reg_mentioned_p (virtual_stack_dynamic_rtx
, op
)
12508 || reg_mentioned_p (virtual_stack_vars_rtx
, op
)))
12511 /* Constants are converted into offsets from labels. */
12515 ind
= XEXP (op
, 0);
12517 if (reload_completed
12518 && (GET_CODE (ind
) == LABEL_REF
12519 || (GET_CODE (ind
) == CONST
12520 && GET_CODE (XEXP (ind
, 0)) == PLUS
12521 && GET_CODE (XEXP (XEXP (ind
, 0), 0)) == LABEL_REF
12522 && CONST_INT_P (XEXP (XEXP (ind
, 0), 1)))))
12525 /* Match: (mem (reg)). */
12527 return arm_address_register_rtx_p (ind
, 0);
12529 /* Autoincremment addressing modes. POST_INC and PRE_DEC are
12530 acceptable in any case (subject to verification by
12531 arm_address_register_rtx_p). We need WB to be true to accept
12532 PRE_INC and POST_DEC. */
12533 if (GET_CODE (ind
) == POST_INC
12534 || GET_CODE (ind
) == PRE_DEC
12536 && (GET_CODE (ind
) == PRE_INC
12537 || GET_CODE (ind
) == POST_DEC
)))
12538 return arm_address_register_rtx_p (XEXP (ind
, 0), 0);
12541 && (GET_CODE (ind
) == POST_MODIFY
|| GET_CODE (ind
) == PRE_MODIFY
)
12542 && arm_address_register_rtx_p (XEXP (ind
, 0), 0)
12543 && GET_CODE (XEXP (ind
, 1)) == PLUS
12544 && rtx_equal_p (XEXP (XEXP (ind
, 1), 0), XEXP (ind
, 0)))
12545 ind
= XEXP (ind
, 1);
12550 if (GET_CODE (ind
) == PLUS
12551 && REG_P (XEXP (ind
, 0))
12552 && REG_MODE_OK_FOR_BASE_P (XEXP (ind
, 0), VOIDmode
)
12553 && CONST_INT_P (XEXP (ind
, 1))
12554 && INTVAL (XEXP (ind
, 1)) > -1024
12555 && INTVAL (XEXP (ind
, 1)) < 1024
12556 && (INTVAL (XEXP (ind
, 1)) & 3) == 0)
12562 /* Return TRUE if OP is a memory operand which we can load or store a vector
12563 to/from. TYPE is one of the following values:
12564 0 - Vector load/stor (vldr)
12565 1 - Core registers (ldm)
12566 2 - Element/structure loads (vld1)
12569 neon_vector_mem_operand (rtx op
, int type
, bool strict
)
12573 /* Reject eliminable registers. */
12574 if (strict
&& ! (reload_in_progress
|| reload_completed
)
12575 && (reg_mentioned_p (frame_pointer_rtx
, op
)
12576 || reg_mentioned_p (arg_pointer_rtx
, op
)
12577 || reg_mentioned_p (virtual_incoming_args_rtx
, op
)
12578 || reg_mentioned_p (virtual_outgoing_args_rtx
, op
)
12579 || reg_mentioned_p (virtual_stack_dynamic_rtx
, op
)
12580 || reg_mentioned_p (virtual_stack_vars_rtx
, op
)))
12583 /* Constants are converted into offsets from labels. */
12587 ind
= XEXP (op
, 0);
12589 if (reload_completed
12590 && (GET_CODE (ind
) == LABEL_REF
12591 || (GET_CODE (ind
) == CONST
12592 && GET_CODE (XEXP (ind
, 0)) == PLUS
12593 && GET_CODE (XEXP (XEXP (ind
, 0), 0)) == LABEL_REF
12594 && CONST_INT_P (XEXP (XEXP (ind
, 0), 1)))))
12597 /* Match: (mem (reg)). */
12599 return arm_address_register_rtx_p (ind
, 0);
12601 /* Allow post-increment with Neon registers. */
12602 if ((type
!= 1 && GET_CODE (ind
) == POST_INC
)
12603 || (type
== 0 && GET_CODE (ind
) == PRE_DEC
))
12604 return arm_address_register_rtx_p (XEXP (ind
, 0), 0);
12606 /* Allow post-increment by register for VLDn */
12607 if (type
== 2 && GET_CODE (ind
) == POST_MODIFY
12608 && GET_CODE (XEXP (ind
, 1)) == PLUS
12609 && REG_P (XEXP (XEXP (ind
, 1), 1)))
12616 && GET_CODE (ind
) == PLUS
12617 && REG_P (XEXP (ind
, 0))
12618 && REG_MODE_OK_FOR_BASE_P (XEXP (ind
, 0), VOIDmode
)
12619 && CONST_INT_P (XEXP (ind
, 1))
12620 && INTVAL (XEXP (ind
, 1)) > -1024
12621 /* For quad modes, we restrict the constant offset to be slightly less
12622 than what the instruction format permits. We have no such constraint
12623 on double mode offsets. (This must match arm_legitimate_index_p.) */
12624 && (INTVAL (XEXP (ind
, 1))
12625 < (VALID_NEON_QREG_MODE (GET_MODE (op
))? 1016 : 1024))
12626 && (INTVAL (XEXP (ind
, 1)) & 3) == 0)
12632 /* Return TRUE if OP is a mem suitable for loading/storing a Neon struct
12635 neon_struct_mem_operand (rtx op
)
12639 /* Reject eliminable registers. */
12640 if (! (reload_in_progress
|| reload_completed
)
12641 && ( reg_mentioned_p (frame_pointer_rtx
, op
)
12642 || reg_mentioned_p (arg_pointer_rtx
, op
)
12643 || reg_mentioned_p (virtual_incoming_args_rtx
, op
)
12644 || reg_mentioned_p (virtual_outgoing_args_rtx
, op
)
12645 || reg_mentioned_p (virtual_stack_dynamic_rtx
, op
)
12646 || reg_mentioned_p (virtual_stack_vars_rtx
, op
)))
12649 /* Constants are converted into offsets from labels. */
12653 ind
= XEXP (op
, 0);
12655 if (reload_completed
12656 && (GET_CODE (ind
) == LABEL_REF
12657 || (GET_CODE (ind
) == CONST
12658 && GET_CODE (XEXP (ind
, 0)) == PLUS
12659 && GET_CODE (XEXP (XEXP (ind
, 0), 0)) == LABEL_REF
12660 && CONST_INT_P (XEXP (XEXP (ind
, 0), 1)))))
12663 /* Match: (mem (reg)). */
12665 return arm_address_register_rtx_p (ind
, 0);
12667 /* vldm/vstm allows POST_INC (ia) and PRE_DEC (db). */
12668 if (GET_CODE (ind
) == POST_INC
12669 || GET_CODE (ind
) == PRE_DEC
)
12670 return arm_address_register_rtx_p (XEXP (ind
, 0), 0);
12675 /* Return true if X is a register that will be eliminated later on. */
12677 arm_eliminable_register (rtx x
)
12679 return REG_P (x
) && (REGNO (x
) == FRAME_POINTER_REGNUM
12680 || REGNO (x
) == ARG_POINTER_REGNUM
12681 || (REGNO (x
) >= FIRST_VIRTUAL_REGISTER
12682 && REGNO (x
) <= LAST_VIRTUAL_REGISTER
));
12685 /* Return GENERAL_REGS if a scratch register required to reload x to/from
12686 coprocessor registers. Otherwise return NO_REGS. */
12689 coproc_secondary_reload_class (machine_mode mode
, rtx x
, bool wb
)
12691 if (mode
== HFmode
)
12693 if (!TARGET_NEON_FP16
&& !TARGET_VFP_FP16INST
)
12694 return GENERAL_REGS
;
12695 if (s_register_operand (x
, mode
) || neon_vector_mem_operand (x
, 2, true))
12697 return GENERAL_REGS
;
12700 /* The neon move patterns handle all legitimate vector and struct
12703 && (MEM_P (x
) || GET_CODE (x
) == CONST_VECTOR
)
12704 && (GET_MODE_CLASS (mode
) == MODE_VECTOR_INT
12705 || GET_MODE_CLASS (mode
) == MODE_VECTOR_FLOAT
12706 || VALID_NEON_STRUCT_MODE (mode
)))
12709 if (arm_coproc_mem_operand (x
, wb
) || s_register_operand (x
, mode
))
12712 return GENERAL_REGS
;
12715 /* Values which must be returned in the most-significant end of the return
12719 arm_return_in_msb (const_tree valtype
)
12721 return (TARGET_AAPCS_BASED
12722 && BYTES_BIG_ENDIAN
12723 && (AGGREGATE_TYPE_P (valtype
)
12724 || TREE_CODE (valtype
) == COMPLEX_TYPE
12725 || FIXED_POINT_TYPE_P (valtype
)));
12728 /* Return TRUE if X references a SYMBOL_REF. */
12730 symbol_mentioned_p (rtx x
)
12735 if (GET_CODE (x
) == SYMBOL_REF
)
12738 /* UNSPEC_TLS entries for a symbol include the SYMBOL_REF, but they
12739 are constant offsets, not symbols. */
12740 if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLS
)
12743 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
12745 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
12751 for (j
= XVECLEN (x
, i
) - 1; j
>= 0; j
--)
12752 if (symbol_mentioned_p (XVECEXP (x
, i
, j
)))
12755 else if (fmt
[i
] == 'e' && symbol_mentioned_p (XEXP (x
, i
)))
12762 /* Return TRUE if X references a LABEL_REF. */
12764 label_mentioned_p (rtx x
)
12769 if (GET_CODE (x
) == LABEL_REF
)
12772 /* UNSPEC_TLS entries for a symbol include a LABEL_REF for the referencing
12773 instruction, but they are constant offsets, not symbols. */
12774 if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLS
)
12777 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
12778 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
12784 for (j
= XVECLEN (x
, i
) - 1; j
>= 0; j
--)
12785 if (label_mentioned_p (XVECEXP (x
, i
, j
)))
12788 else if (fmt
[i
] == 'e' && label_mentioned_p (XEXP (x
, i
)))
12796 tls_mentioned_p (rtx x
)
12798 switch (GET_CODE (x
))
12801 return tls_mentioned_p (XEXP (x
, 0));
12804 if (XINT (x
, 1) == UNSPEC_TLS
)
12807 /* Fall through. */
12813 /* Must not copy any rtx that uses a pc-relative address.
12814 Also, disallow copying of load-exclusive instructions that
12815 may appear after splitting of compare-and-swap-style operations
12816 so as to prevent those loops from being transformed away from their
12817 canonical forms (see PR 69904). */
12820 arm_cannot_copy_insn_p (rtx_insn
*insn
)
12822 /* The tls call insn cannot be copied, as it is paired with a data
12824 if (recog_memoized (insn
) == CODE_FOR_tlscall
)
12827 subrtx_iterator::array_type array
;
12828 FOR_EACH_SUBRTX (iter
, array
, PATTERN (insn
), ALL
)
12830 const_rtx x
= *iter
;
12831 if (GET_CODE (x
) == UNSPEC
12832 && (XINT (x
, 1) == UNSPEC_PIC_BASE
12833 || XINT (x
, 1) == UNSPEC_PIC_UNIFIED
))
12837 rtx set
= single_set (insn
);
12840 rtx src
= SET_SRC (set
);
12841 if (GET_CODE (src
) == ZERO_EXTEND
)
12842 src
= XEXP (src
, 0);
12844 /* Catch the load-exclusive and load-acquire operations. */
12845 if (GET_CODE (src
) == UNSPEC_VOLATILE
12846 && (XINT (src
, 1) == VUNSPEC_LL
12847 || XINT (src
, 1) == VUNSPEC_LAX
))
12854 minmax_code (rtx x
)
12856 enum rtx_code code
= GET_CODE (x
);
12869 gcc_unreachable ();
12873 /* Match pair of min/max operators that can be implemented via usat/ssat. */
12876 arm_sat_operator_match (rtx lo_bound
, rtx hi_bound
,
12877 int *mask
, bool *signed_sat
)
12879 /* The high bound must be a power of two minus one. */
12880 int log
= exact_log2 (INTVAL (hi_bound
) + 1);
12884 /* The low bound is either zero (for usat) or one less than the
12885 negation of the high bound (for ssat). */
12886 if (INTVAL (lo_bound
) == 0)
12891 *signed_sat
= false;
12896 if (INTVAL (lo_bound
) == -INTVAL (hi_bound
) - 1)
12901 *signed_sat
= true;
12909 /* Return 1 if memory locations are adjacent. */
12911 adjacent_mem_locations (rtx a
, rtx b
)
12913 /* We don't guarantee to preserve the order of these memory refs. */
12914 if (volatile_refs_p (a
) || volatile_refs_p (b
))
12917 if ((REG_P (XEXP (a
, 0))
12918 || (GET_CODE (XEXP (a
, 0)) == PLUS
12919 && CONST_INT_P (XEXP (XEXP (a
, 0), 1))))
12920 && (REG_P (XEXP (b
, 0))
12921 || (GET_CODE (XEXP (b
, 0)) == PLUS
12922 && CONST_INT_P (XEXP (XEXP (b
, 0), 1)))))
12924 HOST_WIDE_INT val0
= 0, val1
= 0;
12928 if (GET_CODE (XEXP (a
, 0)) == PLUS
)
12930 reg0
= XEXP (XEXP (a
, 0), 0);
12931 val0
= INTVAL (XEXP (XEXP (a
, 0), 1));
12934 reg0
= XEXP (a
, 0);
12936 if (GET_CODE (XEXP (b
, 0)) == PLUS
)
12938 reg1
= XEXP (XEXP (b
, 0), 0);
12939 val1
= INTVAL (XEXP (XEXP (b
, 0), 1));
12942 reg1
= XEXP (b
, 0);
12944 /* Don't accept any offset that will require multiple
12945 instructions to handle, since this would cause the
12946 arith_adjacentmem pattern to output an overlong sequence. */
12947 if (!const_ok_for_op (val0
, PLUS
) || !const_ok_for_op (val1
, PLUS
))
12950 /* Don't allow an eliminable register: register elimination can make
12951 the offset too large. */
12952 if (arm_eliminable_register (reg0
))
12955 val_diff
= val1
- val0
;
12959 /* If the target has load delay slots, then there's no benefit
12960 to using an ldm instruction unless the offset is zero and
12961 we are optimizing for size. */
12962 return (optimize_size
&& (REGNO (reg0
) == REGNO (reg1
))
12963 && (val0
== 0 || val1
== 0 || val0
== 4 || val1
== 4)
12964 && (val_diff
== 4 || val_diff
== -4));
12967 return ((REGNO (reg0
) == REGNO (reg1
))
12968 && (val_diff
== 4 || val_diff
== -4));
12974 /* Return true if OP is a valid load or store multiple operation. LOAD is true
12975 for load operations, false for store operations. CONSECUTIVE is true
12976 if the register numbers in the operation must be consecutive in the register
12977 bank. RETURN_PC is true if value is to be loaded in PC.
12978 The pattern we are trying to match for load is:
12979 [(SET (R_d0) (MEM (PLUS (addr) (offset))))
12980 (SET (R_d1) (MEM (PLUS (addr) (offset + <reg_increment>))))
12983 (SET (R_dn) (MEM (PLUS (addr) (offset + n * <reg_increment>))))
12986 1. If offset is 0, first insn should be (SET (R_d0) (MEM (src_addr))).
12987 2. REGNO (R_d0) < REGNO (R_d1) < ... < REGNO (R_dn).
12988 3. If consecutive is TRUE, then for kth register being loaded,
12989 REGNO (R_dk) = REGNO (R_d0) + k.
12990 The pattern for store is similar. */
12992 ldm_stm_operation_p (rtx op
, bool load
, machine_mode mode
,
12993 bool consecutive
, bool return_pc
)
12995 HOST_WIDE_INT count
= XVECLEN (op
, 0);
12996 rtx reg
, mem
, addr
;
12998 unsigned first_regno
;
12999 HOST_WIDE_INT i
= 1, base
= 0, offset
= 0;
13001 bool addr_reg_in_reglist
= false;
13002 bool update
= false;
13007 /* If not in SImode, then registers must be consecutive
13008 (e.g., VLDM instructions for DFmode). */
13009 gcc_assert ((mode
== SImode
) || consecutive
);
13010 /* Setting return_pc for stores is illegal. */
13011 gcc_assert (!return_pc
|| load
);
13013 /* Set up the increments and the regs per val based on the mode. */
13014 reg_increment
= GET_MODE_SIZE (mode
);
13015 regs_per_val
= reg_increment
/ 4;
13016 offset_adj
= return_pc
? 1 : 0;
13019 || GET_CODE (XVECEXP (op
, 0, offset_adj
)) != SET
13020 || (load
&& !REG_P (SET_DEST (XVECEXP (op
, 0, offset_adj
)))))
13023 /* Check if this is a write-back. */
13024 elt
= XVECEXP (op
, 0, offset_adj
);
13025 if (GET_CODE (SET_SRC (elt
)) == PLUS
)
13031 /* The offset adjustment must be the number of registers being
13032 popped times the size of a single register. */
13033 if (!REG_P (SET_DEST (elt
))
13034 || !REG_P (XEXP (SET_SRC (elt
), 0))
13035 || (REGNO (SET_DEST (elt
)) != REGNO (XEXP (SET_SRC (elt
), 0)))
13036 || !CONST_INT_P (XEXP (SET_SRC (elt
), 1))
13037 || INTVAL (XEXP (SET_SRC (elt
), 1)) !=
13038 ((count
- 1 - offset_adj
) * reg_increment
))
13042 i
= i
+ offset_adj
;
13043 base
= base
+ offset_adj
;
13044 /* Perform a quick check so we don't blow up below. If only one reg is loaded,
13045 success depends on the type: VLDM can do just one reg,
13046 LDM must do at least two. */
13047 if ((count
<= i
) && (mode
== SImode
))
13050 elt
= XVECEXP (op
, 0, i
- 1);
13051 if (GET_CODE (elt
) != SET
)
13056 reg
= SET_DEST (elt
);
13057 mem
= SET_SRC (elt
);
13061 reg
= SET_SRC (elt
);
13062 mem
= SET_DEST (elt
);
13065 if (!REG_P (reg
) || !MEM_P (mem
))
13068 regno
= REGNO (reg
);
13069 first_regno
= regno
;
13070 addr
= XEXP (mem
, 0);
13071 if (GET_CODE (addr
) == PLUS
)
13073 if (!CONST_INT_P (XEXP (addr
, 1)))
13076 offset
= INTVAL (XEXP (addr
, 1));
13077 addr
= XEXP (addr
, 0);
13083 /* Don't allow SP to be loaded unless it is also the base register. It
13084 guarantees that SP is reset correctly when an LDM instruction
13085 is interrupted. Otherwise, we might end up with a corrupt stack. */
13086 if (load
&& (REGNO (reg
) == SP_REGNUM
) && (REGNO (addr
) != SP_REGNUM
))
13089 for (; i
< count
; i
++)
13091 elt
= XVECEXP (op
, 0, i
);
13092 if (GET_CODE (elt
) != SET
)
13097 reg
= SET_DEST (elt
);
13098 mem
= SET_SRC (elt
);
13102 reg
= SET_SRC (elt
);
13103 mem
= SET_DEST (elt
);
13107 || GET_MODE (reg
) != mode
13108 || REGNO (reg
) <= regno
13111 (unsigned int) (first_regno
+ regs_per_val
* (i
- base
))))
13112 /* Don't allow SP to be loaded unless it is also the base register. It
13113 guarantees that SP is reset correctly when an LDM instruction
13114 is interrupted. Otherwise, we might end up with a corrupt stack. */
13115 || (load
&& (REGNO (reg
) == SP_REGNUM
) && (REGNO (addr
) != SP_REGNUM
))
13117 || GET_MODE (mem
) != mode
13118 || ((GET_CODE (XEXP (mem
, 0)) != PLUS
13119 || !rtx_equal_p (XEXP (XEXP (mem
, 0), 0), addr
)
13120 || !CONST_INT_P (XEXP (XEXP (mem
, 0), 1))
13121 || (INTVAL (XEXP (XEXP (mem
, 0), 1)) !=
13122 offset
+ (i
- base
) * reg_increment
))
13123 && (!REG_P (XEXP (mem
, 0))
13124 || offset
+ (i
- base
) * reg_increment
!= 0)))
13127 regno
= REGNO (reg
);
13128 if (regno
== REGNO (addr
))
13129 addr_reg_in_reglist
= true;
13134 if (update
&& addr_reg_in_reglist
)
13137 /* For Thumb-1, address register is always modified - either by write-back
13138 or by explicit load. If the pattern does not describe an update,
13139 then the address register must be in the list of loaded registers. */
13141 return update
|| addr_reg_in_reglist
;
13147 /* Return true iff it would be profitable to turn a sequence of NOPS loads
13148 or stores (depending on IS_STORE) into a load-multiple or store-multiple
13149 instruction. ADD_OFFSET is nonzero if the base address register needs
13150 to be modified with an add instruction before we can use it. */
13153 multiple_operation_profitable_p (bool is_store ATTRIBUTE_UNUSED
,
13154 int nops
, HOST_WIDE_INT add_offset
)
13156 /* For ARM8,9 & StrongARM, 2 ldr instructions are faster than an ldm
13157 if the offset isn't small enough. The reason 2 ldrs are faster
13158 is because these ARMs are able to do more than one cache access
13159 in a single cycle. The ARM9 and StrongARM have Harvard caches,
13160 whilst the ARM8 has a double bandwidth cache. This means that
13161 these cores can do both an instruction fetch and a data fetch in
13162 a single cycle, so the trick of calculating the address into a
13163 scratch register (one of the result regs) and then doing a load
13164 multiple actually becomes slower (and no smaller in code size).
13165 That is the transformation
13167 ldr rd1, [rbase + offset]
13168 ldr rd2, [rbase + offset + 4]
13172 add rd1, rbase, offset
13173 ldmia rd1, {rd1, rd2}
13175 produces worse code -- '3 cycles + any stalls on rd2' instead of
13176 '2 cycles + any stalls on rd2'. On ARMs with only one cache
13177 access per cycle, the first sequence could never complete in less
13178 than 6 cycles, whereas the ldm sequence would only take 5 and
13179 would make better use of sequential accesses if not hitting the
13182 We cheat here and test 'arm_ld_sched' which we currently know to
13183 only be true for the ARM8, ARM9 and StrongARM. If this ever
13184 changes, then the test below needs to be reworked. */
13185 if (nops
== 2 && arm_ld_sched
&& add_offset
!= 0)
13188 /* XScale has load-store double instructions, but they have stricter
13189 alignment requirements than load-store multiple, so we cannot
13192 For XScale ldm requires 2 + NREGS cycles to complete and blocks
13193 the pipeline until completion.
13201 An ldr instruction takes 1-3 cycles, but does not block the
13210 Best case ldr will always win. However, the more ldr instructions
13211 we issue, the less likely we are to be able to schedule them well.
13212 Using ldr instructions also increases code size.
13214 As a compromise, we use ldr for counts of 1 or 2 regs, and ldm
13215 for counts of 3 or 4 regs. */
13216 if (nops
<= 2 && arm_tune_xscale
&& !optimize_size
)
13221 /* Subroutine of load_multiple_sequence and store_multiple_sequence.
13222 Given an array of UNSORTED_OFFSETS, of which there are NOPS, compute
13223 an array ORDER which describes the sequence to use when accessing the
13224 offsets that produces an ascending order. In this sequence, each
13225 offset must be larger by exactly 4 than the previous one. ORDER[0]
13226 must have been filled in with the lowest offset by the caller.
13227 If UNSORTED_REGS is nonnull, it is an array of register numbers that
13228 we use to verify that ORDER produces an ascending order of registers.
13229 Return true if it was possible to construct such an order, false if
13233 compute_offset_order (int nops
, HOST_WIDE_INT
*unsorted_offsets
, int *order
,
13234 int *unsorted_regs
)
13237 for (i
= 1; i
< nops
; i
++)
13241 order
[i
] = order
[i
- 1];
13242 for (j
= 0; j
< nops
; j
++)
13243 if (unsorted_offsets
[j
] == unsorted_offsets
[order
[i
- 1]] + 4)
13245 /* We must find exactly one offset that is higher than the
13246 previous one by 4. */
13247 if (order
[i
] != order
[i
- 1])
13251 if (order
[i
] == order
[i
- 1])
13253 /* The register numbers must be ascending. */
13254 if (unsorted_regs
!= NULL
13255 && unsorted_regs
[order
[i
]] <= unsorted_regs
[order
[i
- 1]])
13261 /* Used to determine in a peephole whether a sequence of load
13262 instructions can be changed into a load-multiple instruction.
13263 NOPS is the number of separate load instructions we are examining. The
13264 first NOPS entries in OPERANDS are the destination registers, the
13265 next NOPS entries are memory operands. If this function is
13266 successful, *BASE is set to the common base register of the memory
13267 accesses; *LOAD_OFFSET is set to the first memory location's offset
13268 from that base register.
13269 REGS is an array filled in with the destination register numbers.
13270 SAVED_ORDER (if nonnull), is an array filled in with an order that maps
13271 insn numbers to an ascending order of stores. If CHECK_REGS is true,
13272 the sequence of registers in REGS matches the loads from ascending memory
13273 locations, and the function verifies that the register numbers are
13274 themselves ascending. If CHECK_REGS is false, the register numbers
13275 are stored in the order they are found in the operands. */
13277 load_multiple_sequence (rtx
*operands
, int nops
, int *regs
, int *saved_order
,
13278 int *base
, HOST_WIDE_INT
*load_offset
, bool check_regs
)
13280 int unsorted_regs
[MAX_LDM_STM_OPS
];
13281 HOST_WIDE_INT unsorted_offsets
[MAX_LDM_STM_OPS
];
13282 int order
[MAX_LDM_STM_OPS
];
13283 rtx base_reg_rtx
= NULL
;
13287 /* Can only handle up to MAX_LDM_STM_OPS insns at present, though could be
13288 easily extended if required. */
13289 gcc_assert (nops
>= 2 && nops
<= MAX_LDM_STM_OPS
);
13291 memset (order
, 0, MAX_LDM_STM_OPS
* sizeof (int));
13293 /* Loop over the operands and check that the memory references are
13294 suitable (i.e. immediate offsets from the same base register). At
13295 the same time, extract the target register, and the memory
13297 for (i
= 0; i
< nops
; i
++)
13302 /* Convert a subreg of a mem into the mem itself. */
13303 if (GET_CODE (operands
[nops
+ i
]) == SUBREG
)
13304 operands
[nops
+ i
] = alter_subreg (operands
+ (nops
+ i
), true);
13306 gcc_assert (MEM_P (operands
[nops
+ i
]));
13308 /* Don't reorder volatile memory references; it doesn't seem worth
13309 looking for the case where the order is ok anyway. */
13310 if (MEM_VOLATILE_P (operands
[nops
+ i
]))
13313 offset
= const0_rtx
;
13315 if ((REG_P (reg
= XEXP (operands
[nops
+ i
], 0))
13316 || (GET_CODE (reg
) == SUBREG
13317 && REG_P (reg
= SUBREG_REG (reg
))))
13318 || (GET_CODE (XEXP (operands
[nops
+ i
], 0)) == PLUS
13319 && ((REG_P (reg
= XEXP (XEXP (operands
[nops
+ i
], 0), 0)))
13320 || (GET_CODE (reg
) == SUBREG
13321 && REG_P (reg
= SUBREG_REG (reg
))))
13322 && (CONST_INT_P (offset
13323 = XEXP (XEXP (operands
[nops
+ i
], 0), 1)))))
13327 base_reg
= REGNO (reg
);
13328 base_reg_rtx
= reg
;
13329 if (TARGET_THUMB1
&& base_reg
> LAST_LO_REGNUM
)
13332 else if (base_reg
!= (int) REGNO (reg
))
13333 /* Not addressed from the same base register. */
13336 unsorted_regs
[i
] = (REG_P (operands
[i
])
13337 ? REGNO (operands
[i
])
13338 : REGNO (SUBREG_REG (operands
[i
])));
13340 /* If it isn't an integer register, or if it overwrites the
13341 base register but isn't the last insn in the list, then
13342 we can't do this. */
13343 if (unsorted_regs
[i
] < 0
13344 || (TARGET_THUMB1
&& unsorted_regs
[i
] > LAST_LO_REGNUM
)
13345 || unsorted_regs
[i
] > 14
13346 || (i
!= nops
- 1 && unsorted_regs
[i
] == base_reg
))
13349 /* Don't allow SP to be loaded unless it is also the base
13350 register. It guarantees that SP is reset correctly when
13351 an LDM instruction is interrupted. Otherwise, we might
13352 end up with a corrupt stack. */
13353 if (unsorted_regs
[i
] == SP_REGNUM
&& base_reg
!= SP_REGNUM
)
13356 unsorted_offsets
[i
] = INTVAL (offset
);
13357 if (i
== 0 || unsorted_offsets
[i
] < unsorted_offsets
[order
[0]])
13361 /* Not a suitable memory address. */
13365 /* All the useful information has now been extracted from the
13366 operands into unsorted_regs and unsorted_offsets; additionally,
13367 order[0] has been set to the lowest offset in the list. Sort
13368 the offsets into order, verifying that they are adjacent, and
13369 check that the register numbers are ascending. */
13370 if (!compute_offset_order (nops
, unsorted_offsets
, order
,
13371 check_regs
? unsorted_regs
: NULL
))
13375 memcpy (saved_order
, order
, sizeof order
);
13381 for (i
= 0; i
< nops
; i
++)
13382 regs
[i
] = unsorted_regs
[check_regs
? order
[i
] : i
];
13384 *load_offset
= unsorted_offsets
[order
[0]];
13388 && !peep2_reg_dead_p (nops
, base_reg_rtx
))
13391 if (unsorted_offsets
[order
[0]] == 0)
13392 ldm_case
= 1; /* ldmia */
13393 else if (TARGET_ARM
&& unsorted_offsets
[order
[0]] == 4)
13394 ldm_case
= 2; /* ldmib */
13395 else if (TARGET_ARM
&& unsorted_offsets
[order
[nops
- 1]] == 0)
13396 ldm_case
= 3; /* ldmda */
13397 else if (TARGET_32BIT
&& unsorted_offsets
[order
[nops
- 1]] == -4)
13398 ldm_case
= 4; /* ldmdb */
13399 else if (const_ok_for_arm (unsorted_offsets
[order
[0]])
13400 || const_ok_for_arm (-unsorted_offsets
[order
[0]]))
13405 if (!multiple_operation_profitable_p (false, nops
,
13407 ? unsorted_offsets
[order
[0]] : 0))
13413 /* Used to determine in a peephole whether a sequence of store instructions can
13414 be changed into a store-multiple instruction.
13415 NOPS is the number of separate store instructions we are examining.
13416 NOPS_TOTAL is the total number of instructions recognized by the peephole
13418 The first NOPS entries in OPERANDS are the source registers, the next
13419 NOPS entries are memory operands. If this function is successful, *BASE is
13420 set to the common base register of the memory accesses; *LOAD_OFFSET is set
13421 to the first memory location's offset from that base register. REGS is an
13422 array filled in with the source register numbers, REG_RTXS (if nonnull) is
13423 likewise filled with the corresponding rtx's.
13424 SAVED_ORDER (if nonnull), is an array filled in with an order that maps insn
13425 numbers to an ascending order of stores.
13426 If CHECK_REGS is true, the sequence of registers in *REGS matches the stores
13427 from ascending memory locations, and the function verifies that the register
13428 numbers are themselves ascending. If CHECK_REGS is false, the register
13429 numbers are stored in the order they are found in the operands. */
13431 store_multiple_sequence (rtx
*operands
, int nops
, int nops_total
,
13432 int *regs
, rtx
*reg_rtxs
, int *saved_order
, int *base
,
13433 HOST_WIDE_INT
*load_offset
, bool check_regs
)
13435 int unsorted_regs
[MAX_LDM_STM_OPS
];
13436 rtx unsorted_reg_rtxs
[MAX_LDM_STM_OPS
];
13437 HOST_WIDE_INT unsorted_offsets
[MAX_LDM_STM_OPS
];
13438 int order
[MAX_LDM_STM_OPS
];
13440 rtx base_reg_rtx
= NULL
;
13443 /* Write back of base register is currently only supported for Thumb 1. */
13444 int base_writeback
= TARGET_THUMB1
;
13446 /* Can only handle up to MAX_LDM_STM_OPS insns at present, though could be
13447 easily extended if required. */
13448 gcc_assert (nops
>= 2 && nops
<= MAX_LDM_STM_OPS
);
13450 memset (order
, 0, MAX_LDM_STM_OPS
* sizeof (int));
13452 /* Loop over the operands and check that the memory references are
13453 suitable (i.e. immediate offsets from the same base register). At
13454 the same time, extract the target register, and the memory
13456 for (i
= 0; i
< nops
; i
++)
13461 /* Convert a subreg of a mem into the mem itself. */
13462 if (GET_CODE (operands
[nops
+ i
]) == SUBREG
)
13463 operands
[nops
+ i
] = alter_subreg (operands
+ (nops
+ i
), true);
13465 gcc_assert (MEM_P (operands
[nops
+ i
]));
13467 /* Don't reorder volatile memory references; it doesn't seem worth
13468 looking for the case where the order is ok anyway. */
13469 if (MEM_VOLATILE_P (operands
[nops
+ i
]))
13472 offset
= const0_rtx
;
13474 if ((REG_P (reg
= XEXP (operands
[nops
+ i
], 0))
13475 || (GET_CODE (reg
) == SUBREG
13476 && REG_P (reg
= SUBREG_REG (reg
))))
13477 || (GET_CODE (XEXP (operands
[nops
+ i
], 0)) == PLUS
13478 && ((REG_P (reg
= XEXP (XEXP (operands
[nops
+ i
], 0), 0)))
13479 || (GET_CODE (reg
) == SUBREG
13480 && REG_P (reg
= SUBREG_REG (reg
))))
13481 && (CONST_INT_P (offset
13482 = XEXP (XEXP (operands
[nops
+ i
], 0), 1)))))
13484 unsorted_reg_rtxs
[i
] = (REG_P (operands
[i
])
13485 ? operands
[i
] : SUBREG_REG (operands
[i
]));
13486 unsorted_regs
[i
] = REGNO (unsorted_reg_rtxs
[i
]);
13490 base_reg
= REGNO (reg
);
13491 base_reg_rtx
= reg
;
13492 if (TARGET_THUMB1
&& base_reg
> LAST_LO_REGNUM
)
13495 else if (base_reg
!= (int) REGNO (reg
))
13496 /* Not addressed from the same base register. */
13499 /* If it isn't an integer register, then we can't do this. */
13500 if (unsorted_regs
[i
] < 0
13501 || (TARGET_THUMB1
&& unsorted_regs
[i
] > LAST_LO_REGNUM
)
13502 /* The effects are unpredictable if the base register is
13503 both updated and stored. */
13504 || (base_writeback
&& unsorted_regs
[i
] == base_reg
)
13505 || (TARGET_THUMB2
&& unsorted_regs
[i
] == SP_REGNUM
)
13506 || unsorted_regs
[i
] > 14)
13509 unsorted_offsets
[i
] = INTVAL (offset
);
13510 if (i
== 0 || unsorted_offsets
[i
] < unsorted_offsets
[order
[0]])
13514 /* Not a suitable memory address. */
13518 /* All the useful information has now been extracted from the
13519 operands into unsorted_regs and unsorted_offsets; additionally,
13520 order[0] has been set to the lowest offset in the list. Sort
13521 the offsets into order, verifying that they are adjacent, and
13522 check that the register numbers are ascending. */
13523 if (!compute_offset_order (nops
, unsorted_offsets
, order
,
13524 check_regs
? unsorted_regs
: NULL
))
13528 memcpy (saved_order
, order
, sizeof order
);
13534 for (i
= 0; i
< nops
; i
++)
13536 regs
[i
] = unsorted_regs
[check_regs
? order
[i
] : i
];
13538 reg_rtxs
[i
] = unsorted_reg_rtxs
[check_regs
? order
[i
] : i
];
13541 *load_offset
= unsorted_offsets
[order
[0]];
13545 && !peep2_reg_dead_p (nops_total
, base_reg_rtx
))
13548 if (unsorted_offsets
[order
[0]] == 0)
13549 stm_case
= 1; /* stmia */
13550 else if (TARGET_ARM
&& unsorted_offsets
[order
[0]] == 4)
13551 stm_case
= 2; /* stmib */
13552 else if (TARGET_ARM
&& unsorted_offsets
[order
[nops
- 1]] == 0)
13553 stm_case
= 3; /* stmda */
13554 else if (TARGET_32BIT
&& unsorted_offsets
[order
[nops
- 1]] == -4)
13555 stm_case
= 4; /* stmdb */
13559 if (!multiple_operation_profitable_p (false, nops
, 0))
13565 /* Routines for use in generating RTL. */
13567 /* Generate a load-multiple instruction. COUNT is the number of loads in
13568 the instruction; REGS and MEMS are arrays containing the operands.
13569 BASEREG is the base register to be used in addressing the memory operands.
13570 WBACK_OFFSET is nonzero if the instruction should update the base
13574 arm_gen_load_multiple_1 (int count
, int *regs
, rtx
*mems
, rtx basereg
,
13575 HOST_WIDE_INT wback_offset
)
13580 if (!multiple_operation_profitable_p (false, count
, 0))
13586 for (i
= 0; i
< count
; i
++)
13587 emit_move_insn (gen_rtx_REG (SImode
, regs
[i
]), mems
[i
]);
13589 if (wback_offset
!= 0)
13590 emit_move_insn (basereg
, plus_constant (Pmode
, basereg
, wback_offset
));
13592 seq
= get_insns ();
13598 result
= gen_rtx_PARALLEL (VOIDmode
,
13599 rtvec_alloc (count
+ (wback_offset
!= 0 ? 1 : 0)));
13600 if (wback_offset
!= 0)
13602 XVECEXP (result
, 0, 0)
13603 = gen_rtx_SET (basereg
, plus_constant (Pmode
, basereg
, wback_offset
));
13608 for (j
= 0; i
< count
; i
++, j
++)
13609 XVECEXP (result
, 0, i
)
13610 = gen_rtx_SET (gen_rtx_REG (SImode
, regs
[j
]), mems
[j
]);
13615 /* Generate a store-multiple instruction. COUNT is the number of stores in
13616 the instruction; REGS and MEMS are arrays containing the operands.
13617 BASEREG is the base register to be used in addressing the memory operands.
13618 WBACK_OFFSET is nonzero if the instruction should update the base
13622 arm_gen_store_multiple_1 (int count
, int *regs
, rtx
*mems
, rtx basereg
,
13623 HOST_WIDE_INT wback_offset
)
13628 if (GET_CODE (basereg
) == PLUS
)
13629 basereg
= XEXP (basereg
, 0);
13631 if (!multiple_operation_profitable_p (false, count
, 0))
13637 for (i
= 0; i
< count
; i
++)
13638 emit_move_insn (mems
[i
], gen_rtx_REG (SImode
, regs
[i
]));
13640 if (wback_offset
!= 0)
13641 emit_move_insn (basereg
, plus_constant (Pmode
, basereg
, wback_offset
));
13643 seq
= get_insns ();
13649 result
= gen_rtx_PARALLEL (VOIDmode
,
13650 rtvec_alloc (count
+ (wback_offset
!= 0 ? 1 : 0)));
13651 if (wback_offset
!= 0)
13653 XVECEXP (result
, 0, 0)
13654 = gen_rtx_SET (basereg
, plus_constant (Pmode
, basereg
, wback_offset
));
13659 for (j
= 0; i
< count
; i
++, j
++)
13660 XVECEXP (result
, 0, i
)
13661 = gen_rtx_SET (mems
[j
], gen_rtx_REG (SImode
, regs
[j
]));
13666 /* Generate either a load-multiple or a store-multiple instruction. This
13667 function can be used in situations where we can start with a single MEM
13668 rtx and adjust its address upwards.
13669 COUNT is the number of operations in the instruction, not counting a
13670 possible update of the base register. REGS is an array containing the
13672 BASEREG is the base register to be used in addressing the memory operands,
13673 which are constructed from BASEMEM.
13674 WRITE_BACK specifies whether the generated instruction should include an
13675 update of the base register.
13676 OFFSETP is used to pass an offset to and from this function; this offset
13677 is not used when constructing the address (instead BASEMEM should have an
13678 appropriate offset in its address), it is used only for setting
13679 MEM_OFFSET. It is updated only if WRITE_BACK is true.*/
13682 arm_gen_multiple_op (bool is_load
, int *regs
, int count
, rtx basereg
,
13683 bool write_back
, rtx basemem
, HOST_WIDE_INT
*offsetp
)
13685 rtx mems
[MAX_LDM_STM_OPS
];
13686 HOST_WIDE_INT offset
= *offsetp
;
13689 gcc_assert (count
<= MAX_LDM_STM_OPS
);
13691 if (GET_CODE (basereg
) == PLUS
)
13692 basereg
= XEXP (basereg
, 0);
13694 for (i
= 0; i
< count
; i
++)
13696 rtx addr
= plus_constant (Pmode
, basereg
, i
* 4);
13697 mems
[i
] = adjust_automodify_address_nv (basemem
, SImode
, addr
, offset
);
13705 return arm_gen_load_multiple_1 (count
, regs
, mems
, basereg
,
13706 write_back
? 4 * count
: 0);
13708 return arm_gen_store_multiple_1 (count
, regs
, mems
, basereg
,
13709 write_back
? 4 * count
: 0);
13713 arm_gen_load_multiple (int *regs
, int count
, rtx basereg
, int write_back
,
13714 rtx basemem
, HOST_WIDE_INT
*offsetp
)
13716 return arm_gen_multiple_op (TRUE
, regs
, count
, basereg
, write_back
, basemem
,
13721 arm_gen_store_multiple (int *regs
, int count
, rtx basereg
, int write_back
,
13722 rtx basemem
, HOST_WIDE_INT
*offsetp
)
13724 return arm_gen_multiple_op (FALSE
, regs
, count
, basereg
, write_back
, basemem
,
13728 /* Called from a peephole2 expander to turn a sequence of loads into an
13729 LDM instruction. OPERANDS are the operands found by the peephole matcher;
13730 NOPS indicates how many separate loads we are trying to combine. SORT_REGS
13731 is true if we can reorder the registers because they are used commutatively
13733 Returns true iff we could generate a new instruction. */
13736 gen_ldm_seq (rtx
*operands
, int nops
, bool sort_regs
)
13738 int regs
[MAX_LDM_STM_OPS
], mem_order
[MAX_LDM_STM_OPS
];
13739 rtx mems
[MAX_LDM_STM_OPS
];
13740 int i
, j
, base_reg
;
13742 HOST_WIDE_INT offset
;
13743 int write_back
= FALSE
;
13747 ldm_case
= load_multiple_sequence (operands
, nops
, regs
, mem_order
,
13748 &base_reg
, &offset
, !sort_regs
);
13754 for (i
= 0; i
< nops
- 1; i
++)
13755 for (j
= i
+ 1; j
< nops
; j
++)
13756 if (regs
[i
] > regs
[j
])
13762 base_reg_rtx
= gen_rtx_REG (Pmode
, base_reg
);
13766 gcc_assert (peep2_reg_dead_p (nops
, base_reg_rtx
));
13767 gcc_assert (ldm_case
== 1 || ldm_case
== 5);
13773 rtx newbase
= TARGET_THUMB1
? base_reg_rtx
: gen_rtx_REG (SImode
, regs
[0]);
13774 emit_insn (gen_addsi3 (newbase
, base_reg_rtx
, GEN_INT (offset
)));
13776 if (!TARGET_THUMB1
)
13777 base_reg_rtx
= newbase
;
13780 for (i
= 0; i
< nops
; i
++)
13782 addr
= plus_constant (Pmode
, base_reg_rtx
, offset
+ i
* 4);
13783 mems
[i
] = adjust_automodify_address_nv (operands
[nops
+ mem_order
[i
]],
13786 emit_insn (arm_gen_load_multiple_1 (nops
, regs
, mems
, base_reg_rtx
,
13787 write_back
? offset
+ i
* 4 : 0));
13791 /* Called from a peephole2 expander to turn a sequence of stores into an
13792 STM instruction. OPERANDS are the operands found by the peephole matcher;
13793 NOPS indicates how many separate stores we are trying to combine.
13794 Returns true iff we could generate a new instruction. */
13797 gen_stm_seq (rtx
*operands
, int nops
)
13800 int regs
[MAX_LDM_STM_OPS
], mem_order
[MAX_LDM_STM_OPS
];
13801 rtx mems
[MAX_LDM_STM_OPS
];
13804 HOST_WIDE_INT offset
;
13805 int write_back
= FALSE
;
13808 bool base_reg_dies
;
13810 stm_case
= store_multiple_sequence (operands
, nops
, nops
, regs
, NULL
,
13811 mem_order
, &base_reg
, &offset
, true);
13816 base_reg_rtx
= gen_rtx_REG (Pmode
, base_reg
);
13818 base_reg_dies
= peep2_reg_dead_p (nops
, base_reg_rtx
);
13821 gcc_assert (base_reg_dies
);
13827 gcc_assert (base_reg_dies
);
13828 emit_insn (gen_addsi3 (base_reg_rtx
, base_reg_rtx
, GEN_INT (offset
)));
13832 addr
= plus_constant (Pmode
, base_reg_rtx
, offset
);
13834 for (i
= 0; i
< nops
; i
++)
13836 addr
= plus_constant (Pmode
, base_reg_rtx
, offset
+ i
* 4);
13837 mems
[i
] = adjust_automodify_address_nv (operands
[nops
+ mem_order
[i
]],
13840 emit_insn (arm_gen_store_multiple_1 (nops
, regs
, mems
, base_reg_rtx
,
13841 write_back
? offset
+ i
* 4 : 0));
13845 /* Called from a peephole2 expander to turn a sequence of stores that are
13846 preceded by constant loads into an STM instruction. OPERANDS are the
13847 operands found by the peephole matcher; NOPS indicates how many
13848 separate stores we are trying to combine; there are 2 * NOPS
13849 instructions in the peephole.
13850 Returns true iff we could generate a new instruction. */
13853 gen_const_stm_seq (rtx
*operands
, int nops
)
13855 int regs
[MAX_LDM_STM_OPS
], sorted_regs
[MAX_LDM_STM_OPS
];
13856 int reg_order
[MAX_LDM_STM_OPS
], mem_order
[MAX_LDM_STM_OPS
];
13857 rtx reg_rtxs
[MAX_LDM_STM_OPS
], orig_reg_rtxs
[MAX_LDM_STM_OPS
];
13858 rtx mems
[MAX_LDM_STM_OPS
];
13861 HOST_WIDE_INT offset
;
13862 int write_back
= FALSE
;
13865 bool base_reg_dies
;
13867 HARD_REG_SET allocated
;
13869 stm_case
= store_multiple_sequence (operands
, nops
, 2 * nops
, regs
, reg_rtxs
,
13870 mem_order
, &base_reg
, &offset
, false);
13875 memcpy (orig_reg_rtxs
, reg_rtxs
, sizeof orig_reg_rtxs
);
13877 /* If the same register is used more than once, try to find a free
13879 CLEAR_HARD_REG_SET (allocated
);
13880 for (i
= 0; i
< nops
; i
++)
13882 for (j
= i
+ 1; j
< nops
; j
++)
13883 if (regs
[i
] == regs
[j
])
13885 rtx t
= peep2_find_free_register (0, nops
* 2,
13886 TARGET_THUMB1
? "l" : "r",
13887 SImode
, &allocated
);
13891 regs
[i
] = REGNO (t
);
13895 /* Compute an ordering that maps the register numbers to an ascending
13898 for (i
= 0; i
< nops
; i
++)
13899 if (regs
[i
] < regs
[reg_order
[0]])
13902 for (i
= 1; i
< nops
; i
++)
13904 int this_order
= reg_order
[i
- 1];
13905 for (j
= 0; j
< nops
; j
++)
13906 if (regs
[j
] > regs
[reg_order
[i
- 1]]
13907 && (this_order
== reg_order
[i
- 1]
13908 || regs
[j
] < regs
[this_order
]))
13910 reg_order
[i
] = this_order
;
13913 /* Ensure that registers that must be live after the instruction end
13914 up with the correct value. */
13915 for (i
= 0; i
< nops
; i
++)
13917 int this_order
= reg_order
[i
];
13918 if ((this_order
!= mem_order
[i
]
13919 || orig_reg_rtxs
[this_order
] != reg_rtxs
[this_order
])
13920 && !peep2_reg_dead_p (nops
* 2, orig_reg_rtxs
[this_order
]))
13924 /* Load the constants. */
13925 for (i
= 0; i
< nops
; i
++)
13927 rtx op
= operands
[2 * nops
+ mem_order
[i
]];
13928 sorted_regs
[i
] = regs
[reg_order
[i
]];
13929 emit_move_insn (reg_rtxs
[reg_order
[i
]], op
);
13932 base_reg_rtx
= gen_rtx_REG (Pmode
, base_reg
);
13934 base_reg_dies
= peep2_reg_dead_p (nops
* 2, base_reg_rtx
);
13937 gcc_assert (base_reg_dies
);
13943 gcc_assert (base_reg_dies
);
13944 emit_insn (gen_addsi3 (base_reg_rtx
, base_reg_rtx
, GEN_INT (offset
)));
13948 addr
= plus_constant (Pmode
, base_reg_rtx
, offset
);
13950 for (i
= 0; i
< nops
; i
++)
13952 addr
= plus_constant (Pmode
, base_reg_rtx
, offset
+ i
* 4);
13953 mems
[i
] = adjust_automodify_address_nv (operands
[nops
+ mem_order
[i
]],
13956 emit_insn (arm_gen_store_multiple_1 (nops
, sorted_regs
, mems
, base_reg_rtx
,
13957 write_back
? offset
+ i
* 4 : 0));
13961 /* Copy a block of memory using plain ldr/str/ldrh/strh instructions, to permit
13962 unaligned copies on processors which support unaligned semantics for those
13963 instructions. INTERLEAVE_FACTOR can be used to attempt to hide load latency
13964 (using more registers) by doing e.g. load/load/store/store for a factor of 2.
13965 An interleave factor of 1 (the minimum) will perform no interleaving.
13966 Load/store multiple are used for aligned addresses where possible. */
13969 arm_block_move_unaligned_straight (rtx dstbase
, rtx srcbase
,
13970 HOST_WIDE_INT length
,
13971 unsigned int interleave_factor
)
13973 rtx
*regs
= XALLOCAVEC (rtx
, interleave_factor
);
13974 int *regnos
= XALLOCAVEC (int, interleave_factor
);
13975 HOST_WIDE_INT block_size_bytes
= interleave_factor
* UNITS_PER_WORD
;
13976 HOST_WIDE_INT i
, j
;
13977 HOST_WIDE_INT remaining
= length
, words
;
13978 rtx halfword_tmp
= NULL
, byte_tmp
= NULL
;
13980 bool src_aligned
= MEM_ALIGN (srcbase
) >= BITS_PER_WORD
;
13981 bool dst_aligned
= MEM_ALIGN (dstbase
) >= BITS_PER_WORD
;
13982 HOST_WIDE_INT srcoffset
, dstoffset
;
13983 HOST_WIDE_INT src_autoinc
, dst_autoinc
;
13986 gcc_assert (interleave_factor
>= 1 && interleave_factor
<= 4);
13988 /* Use hard registers if we have aligned source or destination so we can use
13989 load/store multiple with contiguous registers. */
13990 if (dst_aligned
|| src_aligned
)
13991 for (i
= 0; i
< interleave_factor
; i
++)
13992 regs
[i
] = gen_rtx_REG (SImode
, i
);
13994 for (i
= 0; i
< interleave_factor
; i
++)
13995 regs
[i
] = gen_reg_rtx (SImode
);
13997 dst
= copy_addr_to_reg (XEXP (dstbase
, 0));
13998 src
= copy_addr_to_reg (XEXP (srcbase
, 0));
14000 srcoffset
= dstoffset
= 0;
14002 /* Calls to arm_gen_load_multiple and arm_gen_store_multiple update SRC/DST.
14003 For copying the last bytes we want to subtract this offset again. */
14004 src_autoinc
= dst_autoinc
= 0;
14006 for (i
= 0; i
< interleave_factor
; i
++)
14009 /* Copy BLOCK_SIZE_BYTES chunks. */
14011 for (i
= 0; i
+ block_size_bytes
<= length
; i
+= block_size_bytes
)
14014 if (src_aligned
&& interleave_factor
> 1)
14016 emit_insn (arm_gen_load_multiple (regnos
, interleave_factor
, src
,
14017 TRUE
, srcbase
, &srcoffset
));
14018 src_autoinc
+= UNITS_PER_WORD
* interleave_factor
;
14022 for (j
= 0; j
< interleave_factor
; j
++)
14024 addr
= plus_constant (Pmode
, src
, (srcoffset
+ j
* UNITS_PER_WORD
14026 mem
= adjust_automodify_address (srcbase
, SImode
, addr
,
14027 srcoffset
+ j
* UNITS_PER_WORD
);
14028 emit_insn (gen_unaligned_loadsi (regs
[j
], mem
));
14030 srcoffset
+= block_size_bytes
;
14034 if (dst_aligned
&& interleave_factor
> 1)
14036 emit_insn (arm_gen_store_multiple (regnos
, interleave_factor
, dst
,
14037 TRUE
, dstbase
, &dstoffset
));
14038 dst_autoinc
+= UNITS_PER_WORD
* interleave_factor
;
14042 for (j
= 0; j
< interleave_factor
; j
++)
14044 addr
= plus_constant (Pmode
, dst
, (dstoffset
+ j
* UNITS_PER_WORD
14046 mem
= adjust_automodify_address (dstbase
, SImode
, addr
,
14047 dstoffset
+ j
* UNITS_PER_WORD
);
14048 emit_insn (gen_unaligned_storesi (mem
, regs
[j
]));
14050 dstoffset
+= block_size_bytes
;
14053 remaining
-= block_size_bytes
;
14056 /* Copy any whole words left (note these aren't interleaved with any
14057 subsequent halfword/byte load/stores in the interests of simplicity). */
14059 words
= remaining
/ UNITS_PER_WORD
;
14061 gcc_assert (words
< interleave_factor
);
14063 if (src_aligned
&& words
> 1)
14065 emit_insn (arm_gen_load_multiple (regnos
, words
, src
, TRUE
, srcbase
,
14067 src_autoinc
+= UNITS_PER_WORD
* words
;
14071 for (j
= 0; j
< words
; j
++)
14073 addr
= plus_constant (Pmode
, src
,
14074 srcoffset
+ j
* UNITS_PER_WORD
- src_autoinc
);
14075 mem
= adjust_automodify_address (srcbase
, SImode
, addr
,
14076 srcoffset
+ j
* UNITS_PER_WORD
);
14078 emit_move_insn (regs
[j
], mem
);
14080 emit_insn (gen_unaligned_loadsi (regs
[j
], mem
));
14082 srcoffset
+= words
* UNITS_PER_WORD
;
14085 if (dst_aligned
&& words
> 1)
14087 emit_insn (arm_gen_store_multiple (regnos
, words
, dst
, TRUE
, dstbase
,
14089 dst_autoinc
+= words
* UNITS_PER_WORD
;
14093 for (j
= 0; j
< words
; j
++)
14095 addr
= plus_constant (Pmode
, dst
,
14096 dstoffset
+ j
* UNITS_PER_WORD
- dst_autoinc
);
14097 mem
= adjust_automodify_address (dstbase
, SImode
, addr
,
14098 dstoffset
+ j
* UNITS_PER_WORD
);
14100 emit_move_insn (mem
, regs
[j
]);
14102 emit_insn (gen_unaligned_storesi (mem
, regs
[j
]));
14104 dstoffset
+= words
* UNITS_PER_WORD
;
14107 remaining
-= words
* UNITS_PER_WORD
;
14109 gcc_assert (remaining
< 4);
14111 /* Copy a halfword if necessary. */
14113 if (remaining
>= 2)
14115 halfword_tmp
= gen_reg_rtx (SImode
);
14117 addr
= plus_constant (Pmode
, src
, srcoffset
- src_autoinc
);
14118 mem
= adjust_automodify_address (srcbase
, HImode
, addr
, srcoffset
);
14119 emit_insn (gen_unaligned_loadhiu (halfword_tmp
, mem
));
14121 /* Either write out immediately, or delay until we've loaded the last
14122 byte, depending on interleave factor. */
14123 if (interleave_factor
== 1)
14125 addr
= plus_constant (Pmode
, dst
, dstoffset
- dst_autoinc
);
14126 mem
= adjust_automodify_address (dstbase
, HImode
, addr
, dstoffset
);
14127 emit_insn (gen_unaligned_storehi (mem
,
14128 gen_lowpart (HImode
, halfword_tmp
)));
14129 halfword_tmp
= NULL
;
14137 gcc_assert (remaining
< 2);
14139 /* Copy last byte. */
14141 if ((remaining
& 1) != 0)
14143 byte_tmp
= gen_reg_rtx (SImode
);
14145 addr
= plus_constant (Pmode
, src
, srcoffset
- src_autoinc
);
14146 mem
= adjust_automodify_address (srcbase
, QImode
, addr
, srcoffset
);
14147 emit_move_insn (gen_lowpart (QImode
, byte_tmp
), mem
);
14149 if (interleave_factor
== 1)
14151 addr
= plus_constant (Pmode
, dst
, dstoffset
- dst_autoinc
);
14152 mem
= adjust_automodify_address (dstbase
, QImode
, addr
, dstoffset
);
14153 emit_move_insn (mem
, gen_lowpart (QImode
, byte_tmp
));
14162 /* Store last halfword if we haven't done so already. */
14166 addr
= plus_constant (Pmode
, dst
, dstoffset
- dst_autoinc
);
14167 mem
= adjust_automodify_address (dstbase
, HImode
, addr
, dstoffset
);
14168 emit_insn (gen_unaligned_storehi (mem
,
14169 gen_lowpart (HImode
, halfword_tmp
)));
14173 /* Likewise for last byte. */
14177 addr
= plus_constant (Pmode
, dst
, dstoffset
- dst_autoinc
);
14178 mem
= adjust_automodify_address (dstbase
, QImode
, addr
, dstoffset
);
14179 emit_move_insn (mem
, gen_lowpart (QImode
, byte_tmp
));
14183 gcc_assert (remaining
== 0 && srcoffset
== dstoffset
);
14186 /* From mips_adjust_block_mem:
14188 Helper function for doing a loop-based block operation on memory
14189 reference MEM. Each iteration of the loop will operate on LENGTH
14192 Create a new base register for use within the loop and point it to
14193 the start of MEM. Create a new memory reference that uses this
14194 register. Store them in *LOOP_REG and *LOOP_MEM respectively. */
14197 arm_adjust_block_mem (rtx mem
, HOST_WIDE_INT length
, rtx
*loop_reg
,
14200 *loop_reg
= copy_addr_to_reg (XEXP (mem
, 0));
14202 /* Although the new mem does not refer to a known location,
14203 it does keep up to LENGTH bytes of alignment. */
14204 *loop_mem
= change_address (mem
, BLKmode
, *loop_reg
);
14205 set_mem_align (*loop_mem
, MIN (MEM_ALIGN (mem
), length
* BITS_PER_UNIT
));
14208 /* From mips_block_move_loop:
14210 Move LENGTH bytes from SRC to DEST using a loop that moves BYTES_PER_ITER
14211 bytes at a time. LENGTH must be at least BYTES_PER_ITER. Assume that
14212 the memory regions do not overlap. */
14215 arm_block_move_unaligned_loop (rtx dest
, rtx src
, HOST_WIDE_INT length
,
14216 unsigned int interleave_factor
,
14217 HOST_WIDE_INT bytes_per_iter
)
14219 rtx src_reg
, dest_reg
, final_src
, test
;
14220 HOST_WIDE_INT leftover
;
14222 leftover
= length
% bytes_per_iter
;
14223 length
-= leftover
;
14225 /* Create registers and memory references for use within the loop. */
14226 arm_adjust_block_mem (src
, bytes_per_iter
, &src_reg
, &src
);
14227 arm_adjust_block_mem (dest
, bytes_per_iter
, &dest_reg
, &dest
);
14229 /* Calculate the value that SRC_REG should have after the last iteration of
14231 final_src
= expand_simple_binop (Pmode
, PLUS
, src_reg
, GEN_INT (length
),
14232 0, 0, OPTAB_WIDEN
);
14234 /* Emit the start of the loop. */
14235 rtx_code_label
*label
= gen_label_rtx ();
14236 emit_label (label
);
14238 /* Emit the loop body. */
14239 arm_block_move_unaligned_straight (dest
, src
, bytes_per_iter
,
14240 interleave_factor
);
14242 /* Move on to the next block. */
14243 emit_move_insn (src_reg
, plus_constant (Pmode
, src_reg
, bytes_per_iter
));
14244 emit_move_insn (dest_reg
, plus_constant (Pmode
, dest_reg
, bytes_per_iter
));
14246 /* Emit the loop condition. */
14247 test
= gen_rtx_NE (VOIDmode
, src_reg
, final_src
);
14248 emit_jump_insn (gen_cbranchsi4 (test
, src_reg
, final_src
, label
));
14250 /* Mop up any left-over bytes. */
14252 arm_block_move_unaligned_straight (dest
, src
, leftover
, interleave_factor
);
14255 /* Emit a block move when either the source or destination is unaligned (not
14256 aligned to a four-byte boundary). This may need further tuning depending on
14257 core type, optimize_size setting, etc. */
14260 arm_movmemqi_unaligned (rtx
*operands
)
14262 HOST_WIDE_INT length
= INTVAL (operands
[2]);
14266 bool src_aligned
= MEM_ALIGN (operands
[1]) >= BITS_PER_WORD
;
14267 bool dst_aligned
= MEM_ALIGN (operands
[0]) >= BITS_PER_WORD
;
14268 /* Inlined memcpy using ldr/str/ldrh/strh can be quite big: try to limit
14269 size of code if optimizing for size. We'll use ldm/stm if src_aligned
14270 or dst_aligned though: allow more interleaving in those cases since the
14271 resulting code can be smaller. */
14272 unsigned int interleave_factor
= (src_aligned
|| dst_aligned
) ? 2 : 1;
14273 HOST_WIDE_INT bytes_per_iter
= (src_aligned
|| dst_aligned
) ? 8 : 4;
14276 arm_block_move_unaligned_loop (operands
[0], operands
[1], length
,
14277 interleave_factor
, bytes_per_iter
);
14279 arm_block_move_unaligned_straight (operands
[0], operands
[1], length
,
14280 interleave_factor
);
14284 /* Note that the loop created by arm_block_move_unaligned_loop may be
14285 subject to loop unrolling, which makes tuning this condition a little
14288 arm_block_move_unaligned_loop (operands
[0], operands
[1], length
, 4, 16);
14290 arm_block_move_unaligned_straight (operands
[0], operands
[1], length
, 4);
14297 arm_gen_movmemqi (rtx
*operands
)
14299 HOST_WIDE_INT in_words_to_go
, out_words_to_go
, last_bytes
;
14300 HOST_WIDE_INT srcoffset
, dstoffset
;
14301 rtx src
, dst
, srcbase
, dstbase
;
14302 rtx part_bytes_reg
= NULL
;
14305 if (!CONST_INT_P (operands
[2])
14306 || !CONST_INT_P (operands
[3])
14307 || INTVAL (operands
[2]) > 64)
14310 if (unaligned_access
&& (INTVAL (operands
[3]) & 3) != 0)
14311 return arm_movmemqi_unaligned (operands
);
14313 if (INTVAL (operands
[3]) & 3)
14316 dstbase
= operands
[0];
14317 srcbase
= operands
[1];
14319 dst
= copy_to_mode_reg (SImode
, XEXP (dstbase
, 0));
14320 src
= copy_to_mode_reg (SImode
, XEXP (srcbase
, 0));
14322 in_words_to_go
= ARM_NUM_INTS (INTVAL (operands
[2]));
14323 out_words_to_go
= INTVAL (operands
[2]) / 4;
14324 last_bytes
= INTVAL (operands
[2]) & 3;
14325 dstoffset
= srcoffset
= 0;
14327 if (out_words_to_go
!= in_words_to_go
&& ((in_words_to_go
- 1) & 3) != 0)
14328 part_bytes_reg
= gen_rtx_REG (SImode
, (in_words_to_go
- 1) & 3);
14330 while (in_words_to_go
>= 2)
14332 if (in_words_to_go
> 4)
14333 emit_insn (arm_gen_load_multiple (arm_regs_in_sequence
, 4, src
,
14334 TRUE
, srcbase
, &srcoffset
));
14336 emit_insn (arm_gen_load_multiple (arm_regs_in_sequence
, in_words_to_go
,
14337 src
, FALSE
, srcbase
,
14340 if (out_words_to_go
)
14342 if (out_words_to_go
> 4)
14343 emit_insn (arm_gen_store_multiple (arm_regs_in_sequence
, 4, dst
,
14344 TRUE
, dstbase
, &dstoffset
));
14345 else if (out_words_to_go
!= 1)
14346 emit_insn (arm_gen_store_multiple (arm_regs_in_sequence
,
14347 out_words_to_go
, dst
,
14350 dstbase
, &dstoffset
));
14353 mem
= adjust_automodify_address (dstbase
, SImode
, dst
, dstoffset
);
14354 emit_move_insn (mem
, gen_rtx_REG (SImode
, R0_REGNUM
));
14355 if (last_bytes
!= 0)
14357 emit_insn (gen_addsi3 (dst
, dst
, GEN_INT (4)));
14363 in_words_to_go
-= in_words_to_go
< 4 ? in_words_to_go
: 4;
14364 out_words_to_go
-= out_words_to_go
< 4 ? out_words_to_go
: 4;
14367 /* OUT_WORDS_TO_GO will be zero here if there are byte stores to do. */
14368 if (out_words_to_go
)
14372 mem
= adjust_automodify_address (srcbase
, SImode
, src
, srcoffset
);
14373 sreg
= copy_to_reg (mem
);
14375 mem
= adjust_automodify_address (dstbase
, SImode
, dst
, dstoffset
);
14376 emit_move_insn (mem
, sreg
);
14379 gcc_assert (!in_words_to_go
); /* Sanity check */
14382 if (in_words_to_go
)
14384 gcc_assert (in_words_to_go
> 0);
14386 mem
= adjust_automodify_address (srcbase
, SImode
, src
, srcoffset
);
14387 part_bytes_reg
= copy_to_mode_reg (SImode
, mem
);
14390 gcc_assert (!last_bytes
|| part_bytes_reg
);
14392 if (BYTES_BIG_ENDIAN
&& last_bytes
)
14394 rtx tmp
= gen_reg_rtx (SImode
);
14396 /* The bytes we want are in the top end of the word. */
14397 emit_insn (gen_lshrsi3 (tmp
, part_bytes_reg
,
14398 GEN_INT (8 * (4 - last_bytes
))));
14399 part_bytes_reg
= tmp
;
14403 mem
= adjust_automodify_address (dstbase
, QImode
,
14404 plus_constant (Pmode
, dst
,
14406 dstoffset
+ last_bytes
- 1);
14407 emit_move_insn (mem
, gen_lowpart (QImode
, part_bytes_reg
));
14411 tmp
= gen_reg_rtx (SImode
);
14412 emit_insn (gen_lshrsi3 (tmp
, part_bytes_reg
, GEN_INT (8)));
14413 part_bytes_reg
= tmp
;
14420 if (last_bytes
> 1)
14422 mem
= adjust_automodify_address (dstbase
, HImode
, dst
, dstoffset
);
14423 emit_move_insn (mem
, gen_lowpart (HImode
, part_bytes_reg
));
14427 rtx tmp
= gen_reg_rtx (SImode
);
14428 emit_insn (gen_addsi3 (dst
, dst
, const2_rtx
));
14429 emit_insn (gen_lshrsi3 (tmp
, part_bytes_reg
, GEN_INT (16)));
14430 part_bytes_reg
= tmp
;
14437 mem
= adjust_automodify_address (dstbase
, QImode
, dst
, dstoffset
);
14438 emit_move_insn (mem
, gen_lowpart (QImode
, part_bytes_reg
));
14445 /* Helper for gen_movmem_ldrd_strd. Increase the address of memory rtx
14448 next_consecutive_mem (rtx mem
)
14450 machine_mode mode
= GET_MODE (mem
);
14451 HOST_WIDE_INT offset
= GET_MODE_SIZE (mode
);
14452 rtx addr
= plus_constant (Pmode
, XEXP (mem
, 0), offset
);
14454 return adjust_automodify_address (mem
, mode
, addr
, offset
);
14457 /* Copy using LDRD/STRD instructions whenever possible.
14458 Returns true upon success. */
14460 gen_movmem_ldrd_strd (rtx
*operands
)
14462 unsigned HOST_WIDE_INT len
;
14463 HOST_WIDE_INT align
;
14464 rtx src
, dst
, base
;
14466 bool src_aligned
, dst_aligned
;
14467 bool src_volatile
, dst_volatile
;
14469 gcc_assert (CONST_INT_P (operands
[2]));
14470 gcc_assert (CONST_INT_P (operands
[3]));
14472 len
= UINTVAL (operands
[2]);
14476 /* Maximum alignment we can assume for both src and dst buffers. */
14477 align
= INTVAL (operands
[3]);
14479 if ((!unaligned_access
) && (len
>= 4) && ((align
& 3) != 0))
14482 /* Place src and dst addresses in registers
14483 and update the corresponding mem rtx. */
14485 dst_volatile
= MEM_VOLATILE_P (dst
);
14486 dst_aligned
= MEM_ALIGN (dst
) >= BITS_PER_WORD
;
14487 base
= copy_to_mode_reg (SImode
, XEXP (dst
, 0));
14488 dst
= adjust_automodify_address (dst
, VOIDmode
, base
, 0);
14491 src_volatile
= MEM_VOLATILE_P (src
);
14492 src_aligned
= MEM_ALIGN (src
) >= BITS_PER_WORD
;
14493 base
= copy_to_mode_reg (SImode
, XEXP (src
, 0));
14494 src
= adjust_automodify_address (src
, VOIDmode
, base
, 0);
14496 if (!unaligned_access
&& !(src_aligned
&& dst_aligned
))
14499 if (src_volatile
|| dst_volatile
)
14502 /* If we cannot generate any LDRD/STRD, try to generate LDM/STM. */
14503 if (!(dst_aligned
|| src_aligned
))
14504 return arm_gen_movmemqi (operands
);
14506 /* If the either src or dst is unaligned we'll be accessing it as pairs
14507 of unaligned SImode accesses. Otherwise we can generate DImode
14508 ldrd/strd instructions. */
14509 src
= adjust_address (src
, src_aligned
? DImode
: SImode
, 0);
14510 dst
= adjust_address (dst
, dst_aligned
? DImode
: SImode
, 0);
14515 reg0
= gen_reg_rtx (DImode
);
14516 rtx low_reg
= NULL_RTX
;
14517 rtx hi_reg
= NULL_RTX
;
14519 if (!src_aligned
|| !dst_aligned
)
14521 low_reg
= gen_lowpart (SImode
, reg0
);
14522 hi_reg
= gen_highpart_mode (SImode
, DImode
, reg0
);
14525 emit_move_insn (reg0
, src
);
14528 emit_insn (gen_unaligned_loadsi (low_reg
, src
));
14529 src
= next_consecutive_mem (src
);
14530 emit_insn (gen_unaligned_loadsi (hi_reg
, src
));
14534 emit_move_insn (dst
, reg0
);
14537 emit_insn (gen_unaligned_storesi (dst
, low_reg
));
14538 dst
= next_consecutive_mem (dst
);
14539 emit_insn (gen_unaligned_storesi (dst
, hi_reg
));
14542 src
= next_consecutive_mem (src
);
14543 dst
= next_consecutive_mem (dst
);
14546 gcc_assert (len
< 8);
14549 /* More than a word but less than a double-word to copy. Copy a word. */
14550 reg0
= gen_reg_rtx (SImode
);
14551 src
= adjust_address (src
, SImode
, 0);
14552 dst
= adjust_address (dst
, SImode
, 0);
14554 emit_move_insn (reg0
, src
);
14556 emit_insn (gen_unaligned_loadsi (reg0
, src
));
14559 emit_move_insn (dst
, reg0
);
14561 emit_insn (gen_unaligned_storesi (dst
, reg0
));
14563 src
= next_consecutive_mem (src
);
14564 dst
= next_consecutive_mem (dst
);
14571 /* Copy the remaining bytes. */
14574 dst
= adjust_address (dst
, HImode
, 0);
14575 src
= adjust_address (src
, HImode
, 0);
14576 reg0
= gen_reg_rtx (SImode
);
14578 emit_insn (gen_zero_extendhisi2 (reg0
, src
));
14580 emit_insn (gen_unaligned_loadhiu (reg0
, src
));
14583 emit_insn (gen_movhi (dst
, gen_lowpart(HImode
, reg0
)));
14585 emit_insn (gen_unaligned_storehi (dst
, gen_lowpart (HImode
, reg0
)));
14587 src
= next_consecutive_mem (src
);
14588 dst
= next_consecutive_mem (dst
);
14593 dst
= adjust_address (dst
, QImode
, 0);
14594 src
= adjust_address (src
, QImode
, 0);
14595 reg0
= gen_reg_rtx (QImode
);
14596 emit_move_insn (reg0
, src
);
14597 emit_move_insn (dst
, reg0
);
14601 /* Select a dominance comparison mode if possible for a test of the general
14602 form (OP (COND_OR (X) (Y)) (const_int 0)). We support three forms.
14603 COND_OR == DOM_CC_X_AND_Y => (X && Y)
14604 COND_OR == DOM_CC_NX_OR_Y => ((! X) || Y)
14605 COND_OR == DOM_CC_X_OR_Y => (X || Y)
14606 In all cases OP will be either EQ or NE, but we don't need to know which
14607 here. If we are unable to support a dominance comparison we return
14608 CC mode. This will then fail to match for the RTL expressions that
14609 generate this call. */
14611 arm_select_dominance_cc_mode (rtx x
, rtx y
, HOST_WIDE_INT cond_or
)
14613 enum rtx_code cond1
, cond2
;
14616 /* Currently we will probably get the wrong result if the individual
14617 comparisons are not simple. This also ensures that it is safe to
14618 reverse a comparison if necessary. */
14619 if ((arm_select_cc_mode (cond1
= GET_CODE (x
), XEXP (x
, 0), XEXP (x
, 1))
14621 || (arm_select_cc_mode (cond2
= GET_CODE (y
), XEXP (y
, 0), XEXP (y
, 1))
14625 /* The if_then_else variant of this tests the second condition if the
14626 first passes, but is true if the first fails. Reverse the first
14627 condition to get a true "inclusive-or" expression. */
14628 if (cond_or
== DOM_CC_NX_OR_Y
)
14629 cond1
= reverse_condition (cond1
);
14631 /* If the comparisons are not equal, and one doesn't dominate the other,
14632 then we can't do this. */
14634 && !comparison_dominates_p (cond1
, cond2
)
14635 && (swapped
= 1, !comparison_dominates_p (cond2
, cond1
)))
14639 std::swap (cond1
, cond2
);
14644 if (cond_or
== DOM_CC_X_AND_Y
)
14649 case EQ
: return CC_DEQmode
;
14650 case LE
: return CC_DLEmode
;
14651 case LEU
: return CC_DLEUmode
;
14652 case GE
: return CC_DGEmode
;
14653 case GEU
: return CC_DGEUmode
;
14654 default: gcc_unreachable ();
14658 if (cond_or
== DOM_CC_X_AND_Y
)
14670 gcc_unreachable ();
14674 if (cond_or
== DOM_CC_X_AND_Y
)
14686 gcc_unreachable ();
14690 if (cond_or
== DOM_CC_X_AND_Y
)
14691 return CC_DLTUmode
;
14696 return CC_DLTUmode
;
14698 return CC_DLEUmode
;
14702 gcc_unreachable ();
14706 if (cond_or
== DOM_CC_X_AND_Y
)
14707 return CC_DGTUmode
;
14712 return CC_DGTUmode
;
14714 return CC_DGEUmode
;
14718 gcc_unreachable ();
14721 /* The remaining cases only occur when both comparisons are the
14724 gcc_assert (cond1
== cond2
);
14728 gcc_assert (cond1
== cond2
);
14732 gcc_assert (cond1
== cond2
);
14736 gcc_assert (cond1
== cond2
);
14737 return CC_DLEUmode
;
14740 gcc_assert (cond1
== cond2
);
14741 return CC_DGEUmode
;
14744 gcc_unreachable ();
14749 arm_select_cc_mode (enum rtx_code op
, rtx x
, rtx y
)
14751 /* All floating point compares return CCFP if it is an equality
14752 comparison, and CCFPE otherwise. */
14753 if (GET_MODE_CLASS (GET_MODE (x
)) == MODE_FLOAT
)
14776 gcc_unreachable ();
14780 /* A compare with a shifted operand. Because of canonicalization, the
14781 comparison will have to be swapped when we emit the assembler. */
14782 if (GET_MODE (y
) == SImode
14783 && (REG_P (y
) || (GET_CODE (y
) == SUBREG
))
14784 && (GET_CODE (x
) == ASHIFT
|| GET_CODE (x
) == ASHIFTRT
14785 || GET_CODE (x
) == LSHIFTRT
|| GET_CODE (x
) == ROTATE
14786 || GET_CODE (x
) == ROTATERT
))
14789 /* This operation is performed swapped, but since we only rely on the Z
14790 flag we don't need an additional mode. */
14791 if (GET_MODE (y
) == SImode
14792 && (REG_P (y
) || (GET_CODE (y
) == SUBREG
))
14793 && GET_CODE (x
) == NEG
14794 && (op
== EQ
|| op
== NE
))
14797 /* This is a special case that is used by combine to allow a
14798 comparison of a shifted byte load to be split into a zero-extend
14799 followed by a comparison of the shifted integer (only valid for
14800 equalities and unsigned inequalities). */
14801 if (GET_MODE (x
) == SImode
14802 && GET_CODE (x
) == ASHIFT
14803 && CONST_INT_P (XEXP (x
, 1)) && INTVAL (XEXP (x
, 1)) == 24
14804 && GET_CODE (XEXP (x
, 0)) == SUBREG
14805 && MEM_P (SUBREG_REG (XEXP (x
, 0)))
14806 && GET_MODE (SUBREG_REG (XEXP (x
, 0))) == QImode
14807 && (op
== EQ
|| op
== NE
14808 || op
== GEU
|| op
== GTU
|| op
== LTU
|| op
== LEU
)
14809 && CONST_INT_P (y
))
14812 /* A construct for a conditional compare, if the false arm contains
14813 0, then both conditions must be true, otherwise either condition
14814 must be true. Not all conditions are possible, so CCmode is
14815 returned if it can't be done. */
14816 if (GET_CODE (x
) == IF_THEN_ELSE
14817 && (XEXP (x
, 2) == const0_rtx
14818 || XEXP (x
, 2) == const1_rtx
)
14819 && COMPARISON_P (XEXP (x
, 0))
14820 && COMPARISON_P (XEXP (x
, 1)))
14821 return arm_select_dominance_cc_mode (XEXP (x
, 0), XEXP (x
, 1),
14822 INTVAL (XEXP (x
, 2)));
14824 /* Alternate canonicalizations of the above. These are somewhat cleaner. */
14825 if (GET_CODE (x
) == AND
14826 && (op
== EQ
|| op
== NE
)
14827 && COMPARISON_P (XEXP (x
, 0))
14828 && COMPARISON_P (XEXP (x
, 1)))
14829 return arm_select_dominance_cc_mode (XEXP (x
, 0), XEXP (x
, 1),
14832 if (GET_CODE (x
) == IOR
14833 && (op
== EQ
|| op
== NE
)
14834 && COMPARISON_P (XEXP (x
, 0))
14835 && COMPARISON_P (XEXP (x
, 1)))
14836 return arm_select_dominance_cc_mode (XEXP (x
, 0), XEXP (x
, 1),
14839 /* An operation (on Thumb) where we want to test for a single bit.
14840 This is done by shifting that bit up into the top bit of a
14841 scratch register; we can then branch on the sign bit. */
14843 && GET_MODE (x
) == SImode
14844 && (op
== EQ
|| op
== NE
)
14845 && GET_CODE (x
) == ZERO_EXTRACT
14846 && XEXP (x
, 1) == const1_rtx
)
14849 /* An operation that sets the condition codes as a side-effect, the
14850 V flag is not set correctly, so we can only use comparisons where
14851 this doesn't matter. (For LT and GE we can use "mi" and "pl"
14853 /* ??? Does the ZERO_EXTRACT case really apply to thumb2? */
14854 if (GET_MODE (x
) == SImode
14856 && (op
== EQ
|| op
== NE
|| op
== LT
|| op
== GE
)
14857 && (GET_CODE (x
) == PLUS
|| GET_CODE (x
) == MINUS
14858 || GET_CODE (x
) == AND
|| GET_CODE (x
) == IOR
14859 || GET_CODE (x
) == XOR
|| GET_CODE (x
) == MULT
14860 || GET_CODE (x
) == NOT
|| GET_CODE (x
) == NEG
14861 || GET_CODE (x
) == LSHIFTRT
14862 || GET_CODE (x
) == ASHIFT
|| GET_CODE (x
) == ASHIFTRT
14863 || GET_CODE (x
) == ROTATERT
14864 || (TARGET_32BIT
&& GET_CODE (x
) == ZERO_EXTRACT
)))
14865 return CC_NOOVmode
;
14867 if (GET_MODE (x
) == QImode
&& (op
== EQ
|| op
== NE
))
14870 if (GET_MODE (x
) == SImode
&& (op
== LTU
|| op
== GEU
)
14871 && GET_CODE (x
) == PLUS
14872 && (rtx_equal_p (XEXP (x
, 0), y
) || rtx_equal_p (XEXP (x
, 1), y
)))
14875 if (GET_MODE (x
) == DImode
|| GET_MODE (y
) == DImode
)
14881 /* A DImode comparison against zero can be implemented by
14882 or'ing the two halves together. */
14883 if (y
== const0_rtx
)
14886 /* We can do an equality test in three Thumb instructions. */
14896 /* DImode unsigned comparisons can be implemented by cmp +
14897 cmpeq without a scratch register. Not worth doing in
14908 /* DImode signed and unsigned comparisons can be implemented
14909 by cmp + sbcs with a scratch register, but that does not
14910 set the Z flag - we must reverse GT/LE/GTU/LEU. */
14911 gcc_assert (op
!= EQ
&& op
!= NE
);
14915 gcc_unreachable ();
14919 if (GET_MODE_CLASS (GET_MODE (x
)) == MODE_CC
)
14920 return GET_MODE (x
);
14925 /* X and Y are two things to compare using CODE. Emit the compare insn and
14926 return the rtx for register 0 in the proper mode. FP means this is a
14927 floating point compare: I don't think that it is needed on the arm. */
14929 arm_gen_compare_reg (enum rtx_code code
, rtx x
, rtx y
, rtx scratch
)
14933 int dimode_comparison
= GET_MODE (x
) == DImode
|| GET_MODE (y
) == DImode
;
14935 /* We might have X as a constant, Y as a register because of the predicates
14936 used for cmpdi. If so, force X to a register here. */
14937 if (dimode_comparison
&& !REG_P (x
))
14938 x
= force_reg (DImode
, x
);
14940 mode
= SELECT_CC_MODE (code
, x
, y
);
14941 cc_reg
= gen_rtx_REG (mode
, CC_REGNUM
);
14943 if (dimode_comparison
14944 && mode
!= CC_CZmode
)
14948 /* To compare two non-zero values for equality, XOR them and
14949 then compare against zero. Not used for ARM mode; there
14950 CC_CZmode is cheaper. */
14951 if (mode
== CC_Zmode
&& y
!= const0_rtx
)
14953 gcc_assert (!reload_completed
);
14954 x
= expand_binop (DImode
, xor_optab
, x
, y
, NULL_RTX
, 0, OPTAB_WIDEN
);
14958 /* A scratch register is required. */
14959 if (reload_completed
)
14960 gcc_assert (scratch
!= NULL
&& GET_MODE (scratch
) == SImode
);
14962 scratch
= gen_rtx_SCRATCH (SImode
);
14964 clobber
= gen_rtx_CLOBBER (VOIDmode
, scratch
);
14965 set
= gen_rtx_SET (cc_reg
, gen_rtx_COMPARE (mode
, x
, y
));
14966 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, set
, clobber
)));
14969 emit_set_insn (cc_reg
, gen_rtx_COMPARE (mode
, x
, y
));
14974 /* Generate a sequence of insns that will generate the correct return
14975 address mask depending on the physical architecture that the program
14978 arm_gen_return_addr_mask (void)
14980 rtx reg
= gen_reg_rtx (Pmode
);
14982 emit_insn (gen_return_addr_mask (reg
));
14987 arm_reload_in_hi (rtx
*operands
)
14989 rtx ref
= operands
[1];
14991 HOST_WIDE_INT offset
= 0;
14993 if (GET_CODE (ref
) == SUBREG
)
14995 offset
= SUBREG_BYTE (ref
);
14996 ref
= SUBREG_REG (ref
);
15001 /* We have a pseudo which has been spilt onto the stack; there
15002 are two cases here: the first where there is a simple
15003 stack-slot replacement and a second where the stack-slot is
15004 out of range, or is used as a subreg. */
15005 if (reg_equiv_mem (REGNO (ref
)))
15007 ref
= reg_equiv_mem (REGNO (ref
));
15008 base
= find_replacement (&XEXP (ref
, 0));
15011 /* The slot is out of range, or was dressed up in a SUBREG. */
15012 base
= reg_equiv_address (REGNO (ref
));
15014 /* PR 62554: If there is no equivalent memory location then just move
15015 the value as an SImode register move. This happens when the target
15016 architecture variant does not have an HImode register move. */
15019 gcc_assert (REG_P (operands
[0]));
15020 emit_insn (gen_movsi (gen_rtx_SUBREG (SImode
, operands
[0], 0),
15021 gen_rtx_SUBREG (SImode
, ref
, 0)));
15026 base
= find_replacement (&XEXP (ref
, 0));
15028 /* Handle the case where the address is too complex to be offset by 1. */
15029 if (GET_CODE (base
) == MINUS
15030 || (GET_CODE (base
) == PLUS
&& !CONST_INT_P (XEXP (base
, 1))))
15032 rtx base_plus
= gen_rtx_REG (SImode
, REGNO (operands
[2]) + 1);
15034 emit_set_insn (base_plus
, base
);
15037 else if (GET_CODE (base
) == PLUS
)
15039 /* The addend must be CONST_INT, or we would have dealt with it above. */
15040 HOST_WIDE_INT hi
, lo
;
15042 offset
+= INTVAL (XEXP (base
, 1));
15043 base
= XEXP (base
, 0);
15045 /* Rework the address into a legal sequence of insns. */
15046 /* Valid range for lo is -4095 -> 4095 */
15049 : -((-offset
) & 0xfff));
15051 /* Corner case, if lo is the max offset then we would be out of range
15052 once we have added the additional 1 below, so bump the msb into the
15053 pre-loading insn(s). */
15057 hi
= ((((offset
- lo
) & (HOST_WIDE_INT
) 0xffffffff)
15058 ^ (HOST_WIDE_INT
) 0x80000000)
15059 - (HOST_WIDE_INT
) 0x80000000);
15061 gcc_assert (hi
+ lo
== offset
);
15065 rtx base_plus
= gen_rtx_REG (SImode
, REGNO (operands
[2]) + 1);
15067 /* Get the base address; addsi3 knows how to handle constants
15068 that require more than one insn. */
15069 emit_insn (gen_addsi3 (base_plus
, base
, GEN_INT (hi
)));
15075 /* Operands[2] may overlap operands[0] (though it won't overlap
15076 operands[1]), that's why we asked for a DImode reg -- so we can
15077 use the bit that does not overlap. */
15078 if (REGNO (operands
[2]) == REGNO (operands
[0]))
15079 scratch
= gen_rtx_REG (SImode
, REGNO (operands
[2]) + 1);
15081 scratch
= gen_rtx_REG (SImode
, REGNO (operands
[2]));
15083 emit_insn (gen_zero_extendqisi2 (scratch
,
15084 gen_rtx_MEM (QImode
,
15085 plus_constant (Pmode
, base
,
15087 emit_insn (gen_zero_extendqisi2 (gen_rtx_SUBREG (SImode
, operands
[0], 0),
15088 gen_rtx_MEM (QImode
,
15089 plus_constant (Pmode
, base
,
15091 if (!BYTES_BIG_ENDIAN
)
15092 emit_set_insn (gen_rtx_SUBREG (SImode
, operands
[0], 0),
15093 gen_rtx_IOR (SImode
,
15096 gen_rtx_SUBREG (SImode
, operands
[0], 0),
15100 emit_set_insn (gen_rtx_SUBREG (SImode
, operands
[0], 0),
15101 gen_rtx_IOR (SImode
,
15102 gen_rtx_ASHIFT (SImode
, scratch
,
15104 gen_rtx_SUBREG (SImode
, operands
[0], 0)));
15107 /* Handle storing a half-word to memory during reload by synthesizing as two
15108 byte stores. Take care not to clobber the input values until after we
15109 have moved them somewhere safe. This code assumes that if the DImode
15110 scratch in operands[2] overlaps either the input value or output address
15111 in some way, then that value must die in this insn (we absolutely need
15112 two scratch registers for some corner cases). */
15114 arm_reload_out_hi (rtx
*operands
)
15116 rtx ref
= operands
[0];
15117 rtx outval
= operands
[1];
15119 HOST_WIDE_INT offset
= 0;
15121 if (GET_CODE (ref
) == SUBREG
)
15123 offset
= SUBREG_BYTE (ref
);
15124 ref
= SUBREG_REG (ref
);
15129 /* We have a pseudo which has been spilt onto the stack; there
15130 are two cases here: the first where there is a simple
15131 stack-slot replacement and a second where the stack-slot is
15132 out of range, or is used as a subreg. */
15133 if (reg_equiv_mem (REGNO (ref
)))
15135 ref
= reg_equiv_mem (REGNO (ref
));
15136 base
= find_replacement (&XEXP (ref
, 0));
15139 /* The slot is out of range, or was dressed up in a SUBREG. */
15140 base
= reg_equiv_address (REGNO (ref
));
15142 /* PR 62254: If there is no equivalent memory location then just move
15143 the value as an SImode register move. This happens when the target
15144 architecture variant does not have an HImode register move. */
15147 gcc_assert (REG_P (outval
) || SUBREG_P (outval
));
15149 if (REG_P (outval
))
15151 emit_insn (gen_movsi (gen_rtx_SUBREG (SImode
, ref
, 0),
15152 gen_rtx_SUBREG (SImode
, outval
, 0)));
15154 else /* SUBREG_P (outval) */
15156 if (GET_MODE (SUBREG_REG (outval
)) == SImode
)
15157 emit_insn (gen_movsi (gen_rtx_SUBREG (SImode
, ref
, 0),
15158 SUBREG_REG (outval
)));
15160 /* FIXME: Handle other cases ? */
15161 gcc_unreachable ();
15167 base
= find_replacement (&XEXP (ref
, 0));
15169 scratch
= gen_rtx_REG (SImode
, REGNO (operands
[2]));
15171 /* Handle the case where the address is too complex to be offset by 1. */
15172 if (GET_CODE (base
) == MINUS
15173 || (GET_CODE (base
) == PLUS
&& !CONST_INT_P (XEXP (base
, 1))))
15175 rtx base_plus
= gen_rtx_REG (SImode
, REGNO (operands
[2]) + 1);
15177 /* Be careful not to destroy OUTVAL. */
15178 if (reg_overlap_mentioned_p (base_plus
, outval
))
15180 /* Updating base_plus might destroy outval, see if we can
15181 swap the scratch and base_plus. */
15182 if (!reg_overlap_mentioned_p (scratch
, outval
))
15183 std::swap (scratch
, base_plus
);
15186 rtx scratch_hi
= gen_rtx_REG (HImode
, REGNO (operands
[2]));
15188 /* Be conservative and copy OUTVAL into the scratch now,
15189 this should only be necessary if outval is a subreg
15190 of something larger than a word. */
15191 /* XXX Might this clobber base? I can't see how it can,
15192 since scratch is known to overlap with OUTVAL, and
15193 must be wider than a word. */
15194 emit_insn (gen_movhi (scratch_hi
, outval
));
15195 outval
= scratch_hi
;
15199 emit_set_insn (base_plus
, base
);
15202 else if (GET_CODE (base
) == PLUS
)
15204 /* The addend must be CONST_INT, or we would have dealt with it above. */
15205 HOST_WIDE_INT hi
, lo
;
15207 offset
+= INTVAL (XEXP (base
, 1));
15208 base
= XEXP (base
, 0);
15210 /* Rework the address into a legal sequence of insns. */
15211 /* Valid range for lo is -4095 -> 4095 */
15214 : -((-offset
) & 0xfff));
15216 /* Corner case, if lo is the max offset then we would be out of range
15217 once we have added the additional 1 below, so bump the msb into the
15218 pre-loading insn(s). */
15222 hi
= ((((offset
- lo
) & (HOST_WIDE_INT
) 0xffffffff)
15223 ^ (HOST_WIDE_INT
) 0x80000000)
15224 - (HOST_WIDE_INT
) 0x80000000);
15226 gcc_assert (hi
+ lo
== offset
);
15230 rtx base_plus
= gen_rtx_REG (SImode
, REGNO (operands
[2]) + 1);
15232 /* Be careful not to destroy OUTVAL. */
15233 if (reg_overlap_mentioned_p (base_plus
, outval
))
15235 /* Updating base_plus might destroy outval, see if we
15236 can swap the scratch and base_plus. */
15237 if (!reg_overlap_mentioned_p (scratch
, outval
))
15238 std::swap (scratch
, base_plus
);
15241 rtx scratch_hi
= gen_rtx_REG (HImode
, REGNO (operands
[2]));
15243 /* Be conservative and copy outval into scratch now,
15244 this should only be necessary if outval is a
15245 subreg of something larger than a word. */
15246 /* XXX Might this clobber base? I can't see how it
15247 can, since scratch is known to overlap with
15249 emit_insn (gen_movhi (scratch_hi
, outval
));
15250 outval
= scratch_hi
;
15254 /* Get the base address; addsi3 knows how to handle constants
15255 that require more than one insn. */
15256 emit_insn (gen_addsi3 (base_plus
, base
, GEN_INT (hi
)));
15262 if (BYTES_BIG_ENDIAN
)
15264 emit_insn (gen_movqi (gen_rtx_MEM (QImode
,
15265 plus_constant (Pmode
, base
,
15267 gen_lowpart (QImode
, outval
)));
15268 emit_insn (gen_lshrsi3 (scratch
,
15269 gen_rtx_SUBREG (SImode
, outval
, 0),
15271 emit_insn (gen_movqi (gen_rtx_MEM (QImode
, plus_constant (Pmode
, base
,
15273 gen_lowpart (QImode
, scratch
)));
15277 emit_insn (gen_movqi (gen_rtx_MEM (QImode
, plus_constant (Pmode
, base
,
15279 gen_lowpart (QImode
, outval
)));
15280 emit_insn (gen_lshrsi3 (scratch
,
15281 gen_rtx_SUBREG (SImode
, outval
, 0),
15283 emit_insn (gen_movqi (gen_rtx_MEM (QImode
,
15284 plus_constant (Pmode
, base
,
15286 gen_lowpart (QImode
, scratch
)));
15290 /* Return true if a type must be passed in memory. For AAPCS, small aggregates
15291 (padded to the size of a word) should be passed in a register. */
15294 arm_must_pass_in_stack (machine_mode mode
, const_tree type
)
15296 if (TARGET_AAPCS_BASED
)
15297 return must_pass_in_stack_var_size (mode
, type
);
15299 return must_pass_in_stack_var_size_or_pad (mode
, type
);
15303 /* Implement TARGET_FUNCTION_ARG_PADDING; return PAD_UPWARD if the lowest
15304 byte of a stack argument has useful data. For legacy APCS ABIs we use
15305 the default. For AAPCS based ABIs small aggregate types are placed
15306 in the lowest memory address. */
15308 static pad_direction
15309 arm_function_arg_padding (machine_mode mode
, const_tree type
)
15311 if (!TARGET_AAPCS_BASED
)
15312 return default_function_arg_padding (mode
, type
);
15314 if (type
&& BYTES_BIG_ENDIAN
&& INTEGRAL_TYPE_P (type
))
15315 return PAD_DOWNWARD
;
15321 /* Similarly, for use by BLOCK_REG_PADDING (MODE, TYPE, FIRST).
15322 Return !BYTES_BIG_ENDIAN if the least significant byte of the
15323 register has useful data, and return the opposite if the most
15324 significant byte does. */
15327 arm_pad_reg_upward (machine_mode mode
,
15328 tree type
, int first ATTRIBUTE_UNUSED
)
15330 if (TARGET_AAPCS_BASED
&& BYTES_BIG_ENDIAN
)
15332 /* For AAPCS, small aggregates, small fixed-point types,
15333 and small complex types are always padded upwards. */
15336 if ((AGGREGATE_TYPE_P (type
)
15337 || TREE_CODE (type
) == COMPLEX_TYPE
15338 || FIXED_POINT_TYPE_P (type
))
15339 && int_size_in_bytes (type
) <= 4)
15344 if ((COMPLEX_MODE_P (mode
) || ALL_FIXED_POINT_MODE_P (mode
))
15345 && GET_MODE_SIZE (mode
) <= 4)
15350 /* Otherwise, use default padding. */
15351 return !BYTES_BIG_ENDIAN
;
15354 /* Returns true iff OFFSET is valid for use in an LDRD/STRD instruction,
15355 assuming that the address in the base register is word aligned. */
15357 offset_ok_for_ldrd_strd (HOST_WIDE_INT offset
)
15359 HOST_WIDE_INT max_offset
;
15361 /* Offset must be a multiple of 4 in Thumb mode. */
15362 if (TARGET_THUMB2
&& ((offset
& 3) != 0))
15367 else if (TARGET_ARM
)
15372 return ((offset
<= max_offset
) && (offset
>= -max_offset
));
15375 /* Checks whether the operands are valid for use in an LDRD/STRD instruction.
15376 Assumes that RT, RT2, and RN are REG. This is guaranteed by the patterns.
15377 Assumes that the address in the base register RN is word aligned. Pattern
15378 guarantees that both memory accesses use the same base register,
15379 the offsets are constants within the range, and the gap between the offsets is 4.
15380 If preload complete then check that registers are legal. WBACK indicates whether
15381 address is updated. LOAD indicates whether memory access is load or store. */
15383 operands_ok_ldrd_strd (rtx rt
, rtx rt2
, rtx rn
, HOST_WIDE_INT offset
,
15384 bool wback
, bool load
)
15386 unsigned int t
, t2
, n
;
15388 if (!reload_completed
)
15391 if (!offset_ok_for_ldrd_strd (offset
))
15398 if ((TARGET_THUMB2
)
15399 && ((wback
&& (n
== t
|| n
== t2
))
15400 || (t
== SP_REGNUM
)
15401 || (t
== PC_REGNUM
)
15402 || (t2
== SP_REGNUM
)
15403 || (t2
== PC_REGNUM
)
15404 || (!load
&& (n
== PC_REGNUM
))
15405 || (load
&& (t
== t2
))
15406 /* Triggers Cortex-M3 LDRD errata. */
15407 || (!wback
&& load
&& fix_cm3_ldrd
&& (n
== t
))))
15411 && ((wback
&& (n
== t
|| n
== t2
))
15412 || (t2
== PC_REGNUM
)
15413 || (t
% 2 != 0) /* First destination register is not even. */
15415 /* PC can be used as base register (for offset addressing only),
15416 but it is depricated. */
15417 || (n
== PC_REGNUM
)))
15423 /* Return true if a 64-bit access with alignment ALIGN and with a
15424 constant offset OFFSET from the base pointer is permitted on this
15427 align_ok_ldrd_strd (HOST_WIDE_INT align
, HOST_WIDE_INT offset
)
15429 return (unaligned_access
15430 ? (align
>= BITS_PER_WORD
&& (offset
& 3) == 0)
15431 : (align
>= 2 * BITS_PER_WORD
&& (offset
& 7) == 0));
15434 /* Helper for gen_operands_ldrd_strd. Returns true iff the memory
15435 operand MEM's address contains an immediate offset from the base
15436 register and has no side effects, in which case it sets BASE,
15437 OFFSET and ALIGN accordingly. */
15439 mem_ok_for_ldrd_strd (rtx mem
, rtx
*base
, rtx
*offset
, HOST_WIDE_INT
*align
)
15443 gcc_assert (base
!= NULL
&& offset
!= NULL
);
15445 /* TODO: Handle more general memory operand patterns, such as
15446 PRE_DEC and PRE_INC. */
15448 if (side_effects_p (mem
))
15451 /* Can't deal with subregs. */
15452 if (GET_CODE (mem
) == SUBREG
)
15455 gcc_assert (MEM_P (mem
));
15457 *offset
= const0_rtx
;
15458 *align
= MEM_ALIGN (mem
);
15460 addr
= XEXP (mem
, 0);
15462 /* If addr isn't valid for DImode, then we can't handle it. */
15463 if (!arm_legitimate_address_p (DImode
, addr
,
15464 reload_in_progress
|| reload_completed
))
15472 else if (GET_CODE (addr
) == PLUS
|| GET_CODE (addr
) == MINUS
)
15474 *base
= XEXP (addr
, 0);
15475 *offset
= XEXP (addr
, 1);
15476 return (REG_P (*base
) && CONST_INT_P (*offset
));
15482 /* Called from a peephole2 to replace two word-size accesses with a
15483 single LDRD/STRD instruction. Returns true iff we can generate a
15484 new instruction sequence. That is, both accesses use the same base
15485 register and the gap between constant offsets is 4. This function
15486 may reorder its operands to match ldrd/strd RTL templates.
15487 OPERANDS are the operands found by the peephole matcher;
15488 OPERANDS[0,1] are register operands, and OPERANDS[2,3] are the
15489 corresponding memory operands. LOAD indicaates whether the access
15490 is load or store. CONST_STORE indicates a store of constant
15491 integer values held in OPERANDS[4,5] and assumes that the pattern
15492 is of length 4 insn, for the purpose of checking dead registers.
15493 COMMUTE indicates that register operands may be reordered. */
15495 gen_operands_ldrd_strd (rtx
*operands
, bool load
,
15496 bool const_store
, bool commute
)
15499 HOST_WIDE_INT offsets
[2], offset
, align
[2];
15500 rtx base
= NULL_RTX
;
15501 rtx cur_base
, cur_offset
, tmp
;
15503 HARD_REG_SET regset
;
15505 gcc_assert (!const_store
|| !load
);
15506 /* Check that the memory references are immediate offsets from the
15507 same base register. Extract the base register, the destination
15508 registers, and the corresponding memory offsets. */
15509 for (i
= 0; i
< nops
; i
++)
15511 if (!mem_ok_for_ldrd_strd (operands
[nops
+i
], &cur_base
, &cur_offset
,
15517 else if (REGNO (base
) != REGNO (cur_base
))
15520 offsets
[i
] = INTVAL (cur_offset
);
15521 if (GET_CODE (operands
[i
]) == SUBREG
)
15523 tmp
= SUBREG_REG (operands
[i
]);
15524 gcc_assert (GET_MODE (operands
[i
]) == GET_MODE (tmp
));
15529 /* Make sure there is no dependency between the individual loads. */
15530 if (load
&& REGNO (operands
[0]) == REGNO (base
))
15531 return false; /* RAW */
15533 if (load
&& REGNO (operands
[0]) == REGNO (operands
[1]))
15534 return false; /* WAW */
15536 /* If the same input register is used in both stores
15537 when storing different constants, try to find a free register.
15538 For example, the code
15543 can be transformed into
15547 in Thumb mode assuming that r1 is free.
15548 For ARM mode do the same but only if the starting register
15549 can be made to be even. */
15551 && REGNO (operands
[0]) == REGNO (operands
[1])
15552 && INTVAL (operands
[4]) != INTVAL (operands
[5]))
15556 CLEAR_HARD_REG_SET (regset
);
15557 tmp
= peep2_find_free_register (0, 4, "r", SImode
, ®set
);
15558 if (tmp
== NULL_RTX
)
15561 /* Use the new register in the first load to ensure that
15562 if the original input register is not dead after peephole,
15563 then it will have the correct constant value. */
15566 else if (TARGET_ARM
)
15568 int regno
= REGNO (operands
[0]);
15569 if (!peep2_reg_dead_p (4, operands
[0]))
15571 /* When the input register is even and is not dead after the
15572 pattern, it has to hold the second constant but we cannot
15573 form a legal STRD in ARM mode with this register as the second
15575 if (regno
% 2 == 0)
15578 /* Is regno-1 free? */
15579 SET_HARD_REG_SET (regset
);
15580 CLEAR_HARD_REG_BIT(regset
, regno
- 1);
15581 tmp
= peep2_find_free_register (0, 4, "r", SImode
, ®set
);
15582 if (tmp
== NULL_RTX
)
15589 /* Find a DImode register. */
15590 CLEAR_HARD_REG_SET (regset
);
15591 tmp
= peep2_find_free_register (0, 4, "r", DImode
, ®set
);
15592 if (tmp
!= NULL_RTX
)
15594 operands
[0] = simplify_gen_subreg (SImode
, tmp
, DImode
, 0);
15595 operands
[1] = simplify_gen_subreg (SImode
, tmp
, DImode
, 4);
15599 /* Can we use the input register to form a DI register? */
15600 SET_HARD_REG_SET (regset
);
15601 CLEAR_HARD_REG_BIT(regset
,
15602 regno
% 2 == 0 ? regno
+ 1 : regno
- 1);
15603 tmp
= peep2_find_free_register (0, 4, "r", SImode
, ®set
);
15604 if (tmp
== NULL_RTX
)
15606 operands
[regno
% 2 == 1 ? 0 : 1] = tmp
;
15610 gcc_assert (operands
[0] != NULL_RTX
);
15611 gcc_assert (operands
[1] != NULL_RTX
);
15612 gcc_assert (REGNO (operands
[0]) % 2 == 0);
15613 gcc_assert (REGNO (operands
[1]) == REGNO (operands
[0]) + 1);
15617 /* Make sure the instructions are ordered with lower memory access first. */
15618 if (offsets
[0] > offsets
[1])
15620 gap
= offsets
[0] - offsets
[1];
15621 offset
= offsets
[1];
15623 /* Swap the instructions such that lower memory is accessed first. */
15624 std::swap (operands
[0], operands
[1]);
15625 std::swap (operands
[2], operands
[3]);
15626 std::swap (align
[0], align
[1]);
15628 std::swap (operands
[4], operands
[5]);
15632 gap
= offsets
[1] - offsets
[0];
15633 offset
= offsets
[0];
15636 /* Make sure accesses are to consecutive memory locations. */
15640 if (!align_ok_ldrd_strd (align
[0], offset
))
15643 /* Make sure we generate legal instructions. */
15644 if (operands_ok_ldrd_strd (operands
[0], operands
[1], base
, offset
,
15648 /* In Thumb state, where registers are almost unconstrained, there
15649 is little hope to fix it. */
15653 if (load
&& commute
)
15655 /* Try reordering registers. */
15656 std::swap (operands
[0], operands
[1]);
15657 if (operands_ok_ldrd_strd (operands
[0], operands
[1], base
, offset
,
15664 /* If input registers are dead after this pattern, they can be
15665 reordered or replaced by other registers that are free in the
15666 current pattern. */
15667 if (!peep2_reg_dead_p (4, operands
[0])
15668 || !peep2_reg_dead_p (4, operands
[1]))
15671 /* Try to reorder the input registers. */
15672 /* For example, the code
15677 can be transformed into
15682 if (operands_ok_ldrd_strd (operands
[1], operands
[0], base
, offset
,
15685 std::swap (operands
[0], operands
[1]);
15689 /* Try to find a free DI register. */
15690 CLEAR_HARD_REG_SET (regset
);
15691 add_to_hard_reg_set (®set
, SImode
, REGNO (operands
[0]));
15692 add_to_hard_reg_set (®set
, SImode
, REGNO (operands
[1]));
15695 tmp
= peep2_find_free_register (0, 4, "r", DImode
, ®set
);
15696 if (tmp
== NULL_RTX
)
15699 /* DREG must be an even-numbered register in DImode.
15700 Split it into SI registers. */
15701 operands
[0] = simplify_gen_subreg (SImode
, tmp
, DImode
, 0);
15702 operands
[1] = simplify_gen_subreg (SImode
, tmp
, DImode
, 4);
15703 gcc_assert (operands
[0] != NULL_RTX
);
15704 gcc_assert (operands
[1] != NULL_RTX
);
15705 gcc_assert (REGNO (operands
[0]) % 2 == 0);
15706 gcc_assert (REGNO (operands
[0]) + 1 == REGNO (operands
[1]));
15708 return (operands_ok_ldrd_strd (operands
[0], operands
[1],
15720 /* Print a symbolic form of X to the debug file, F. */
15722 arm_print_value (FILE *f
, rtx x
)
15724 switch (GET_CODE (x
))
15727 fprintf (f
, HOST_WIDE_INT_PRINT_HEX
, INTVAL (x
));
15731 fprintf (f
, "<0x%lx,0x%lx>", (long)XWINT (x
, 2), (long)XWINT (x
, 3));
15739 for (i
= 0; i
< CONST_VECTOR_NUNITS (x
); i
++)
15741 fprintf (f
, HOST_WIDE_INT_PRINT_HEX
, INTVAL (CONST_VECTOR_ELT (x
, i
)));
15742 if (i
< (CONST_VECTOR_NUNITS (x
) - 1))
15750 fprintf (f
, "\"%s\"", XSTR (x
, 0));
15754 fprintf (f
, "`%s'", XSTR (x
, 0));
15758 fprintf (f
, "L%d", INSN_UID (XEXP (x
, 0)));
15762 arm_print_value (f
, XEXP (x
, 0));
15766 arm_print_value (f
, XEXP (x
, 0));
15768 arm_print_value (f
, XEXP (x
, 1));
15776 fprintf (f
, "????");
15781 /* Routines for manipulation of the constant pool. */
15783 /* Arm instructions cannot load a large constant directly into a
15784 register; they have to come from a pc relative load. The constant
15785 must therefore be placed in the addressable range of the pc
15786 relative load. Depending on the precise pc relative load
15787 instruction the range is somewhere between 256 bytes and 4k. This
15788 means that we often have to dump a constant inside a function, and
15789 generate code to branch around it.
15791 It is important to minimize this, since the branches will slow
15792 things down and make the code larger.
15794 Normally we can hide the table after an existing unconditional
15795 branch so that there is no interruption of the flow, but in the
15796 worst case the code looks like this:
15814 We fix this by performing a scan after scheduling, which notices
15815 which instructions need to have their operands fetched from the
15816 constant table and builds the table.
15818 The algorithm starts by building a table of all the constants that
15819 need fixing up and all the natural barriers in the function (places
15820 where a constant table can be dropped without breaking the flow).
15821 For each fixup we note how far the pc-relative replacement will be
15822 able to reach and the offset of the instruction into the function.
15824 Having built the table we then group the fixes together to form
15825 tables that are as large as possible (subject to addressing
15826 constraints) and emit each table of constants after the last
15827 barrier that is within range of all the instructions in the group.
15828 If a group does not contain a barrier, then we forcibly create one
15829 by inserting a jump instruction into the flow. Once the table has
15830 been inserted, the insns are then modified to reference the
15831 relevant entry in the pool.
15833 Possible enhancements to the algorithm (not implemented) are:
15835 1) For some processors and object formats, there may be benefit in
15836 aligning the pools to the start of cache lines; this alignment
15837 would need to be taken into account when calculating addressability
15840 /* These typedefs are located at the start of this file, so that
15841 they can be used in the prototypes there. This comment is to
15842 remind readers of that fact so that the following structures
15843 can be understood more easily.
15845 typedef struct minipool_node Mnode;
15846 typedef struct minipool_fixup Mfix; */
15848 struct minipool_node
15850 /* Doubly linked chain of entries. */
15853 /* The maximum offset into the code that this entry can be placed. While
15854 pushing fixes for forward references, all entries are sorted in order
15855 of increasing max_address. */
15856 HOST_WIDE_INT max_address
;
15857 /* Similarly for an entry inserted for a backwards ref. */
15858 HOST_WIDE_INT min_address
;
15859 /* The number of fixes referencing this entry. This can become zero
15860 if we "unpush" an entry. In this case we ignore the entry when we
15861 come to emit the code. */
15863 /* The offset from the start of the minipool. */
15864 HOST_WIDE_INT offset
;
15865 /* The value in table. */
15867 /* The mode of value. */
15869 /* The size of the value. With iWMMXt enabled
15870 sizes > 4 also imply an alignment of 8-bytes. */
15874 struct minipool_fixup
15878 HOST_WIDE_INT address
;
15884 HOST_WIDE_INT forwards
;
15885 HOST_WIDE_INT backwards
;
15888 /* Fixes less than a word need padding out to a word boundary. */
15889 #define MINIPOOL_FIX_SIZE(mode) \
15890 (GET_MODE_SIZE ((mode)) >= 4 ? GET_MODE_SIZE ((mode)) : 4)
15892 static Mnode
* minipool_vector_head
;
15893 static Mnode
* minipool_vector_tail
;
15894 static rtx_code_label
*minipool_vector_label
;
15895 static int minipool_pad
;
15897 /* The linked list of all minipool fixes required for this function. */
15898 Mfix
* minipool_fix_head
;
15899 Mfix
* minipool_fix_tail
;
15900 /* The fix entry for the current minipool, once it has been placed. */
15901 Mfix
* minipool_barrier
;
15903 #ifndef JUMP_TABLES_IN_TEXT_SECTION
15904 #define JUMP_TABLES_IN_TEXT_SECTION 0
15907 static HOST_WIDE_INT
15908 get_jump_table_size (rtx_jump_table_data
*insn
)
15910 /* ADDR_VECs only take room if read-only data does into the text
15912 if (JUMP_TABLES_IN_TEXT_SECTION
|| readonly_data_section
== text_section
)
15914 rtx body
= PATTERN (insn
);
15915 int elt
= GET_CODE (body
) == ADDR_DIFF_VEC
? 1 : 0;
15916 HOST_WIDE_INT size
;
15917 HOST_WIDE_INT modesize
;
15919 modesize
= GET_MODE_SIZE (GET_MODE (body
));
15920 size
= modesize
* XVECLEN (body
, elt
);
15924 /* Round up size of TBB table to a halfword boundary. */
15925 size
= (size
+ 1) & ~HOST_WIDE_INT_1
;
15928 /* No padding necessary for TBH. */
15931 /* Add two bytes for alignment on Thumb. */
15936 gcc_unreachable ();
15944 /* Return the maximum amount of padding that will be inserted before
15947 static HOST_WIDE_INT
15948 get_label_padding (rtx label
)
15950 HOST_WIDE_INT align
, min_insn_size
;
15952 align
= 1 << label_to_alignment (label
).levels
[0].log
;
15953 min_insn_size
= TARGET_THUMB
? 2 : 4;
15954 return align
> min_insn_size
? align
- min_insn_size
: 0;
15957 /* Move a minipool fix MP from its current location to before MAX_MP.
15958 If MAX_MP is NULL, then MP doesn't need moving, but the addressing
15959 constraints may need updating. */
15961 move_minipool_fix_forward_ref (Mnode
*mp
, Mnode
*max_mp
,
15962 HOST_WIDE_INT max_address
)
15964 /* The code below assumes these are different. */
15965 gcc_assert (mp
!= max_mp
);
15967 if (max_mp
== NULL
)
15969 if (max_address
< mp
->max_address
)
15970 mp
->max_address
= max_address
;
15974 if (max_address
> max_mp
->max_address
- mp
->fix_size
)
15975 mp
->max_address
= max_mp
->max_address
- mp
->fix_size
;
15977 mp
->max_address
= max_address
;
15979 /* Unlink MP from its current position. Since max_mp is non-null,
15980 mp->prev must be non-null. */
15981 mp
->prev
->next
= mp
->next
;
15982 if (mp
->next
!= NULL
)
15983 mp
->next
->prev
= mp
->prev
;
15985 minipool_vector_tail
= mp
->prev
;
15987 /* Re-insert it before MAX_MP. */
15989 mp
->prev
= max_mp
->prev
;
15992 if (mp
->prev
!= NULL
)
15993 mp
->prev
->next
= mp
;
15995 minipool_vector_head
= mp
;
15998 /* Save the new entry. */
16001 /* Scan over the preceding entries and adjust their addresses as
16003 while (mp
->prev
!= NULL
16004 && mp
->prev
->max_address
> mp
->max_address
- mp
->prev
->fix_size
)
16006 mp
->prev
->max_address
= mp
->max_address
- mp
->prev
->fix_size
;
16013 /* Add a constant to the minipool for a forward reference. Returns the
16014 node added or NULL if the constant will not fit in this pool. */
16016 add_minipool_forward_ref (Mfix
*fix
)
16018 /* If set, max_mp is the first pool_entry that has a lower
16019 constraint than the one we are trying to add. */
16020 Mnode
* max_mp
= NULL
;
16021 HOST_WIDE_INT max_address
= fix
->address
+ fix
->forwards
- minipool_pad
;
16024 /* If the minipool starts before the end of FIX->INSN then this FIX
16025 can not be placed into the current pool. Furthermore, adding the
16026 new constant pool entry may cause the pool to start FIX_SIZE bytes
16028 if (minipool_vector_head
&&
16029 (fix
->address
+ get_attr_length (fix
->insn
)
16030 >= minipool_vector_head
->max_address
- fix
->fix_size
))
16033 /* Scan the pool to see if a constant with the same value has
16034 already been added. While we are doing this, also note the
16035 location where we must insert the constant if it doesn't already
16037 for (mp
= minipool_vector_head
; mp
!= NULL
; mp
= mp
->next
)
16039 if (GET_CODE (fix
->value
) == GET_CODE (mp
->value
)
16040 && fix
->mode
== mp
->mode
16041 && (!LABEL_P (fix
->value
)
16042 || (CODE_LABEL_NUMBER (fix
->value
)
16043 == CODE_LABEL_NUMBER (mp
->value
)))
16044 && rtx_equal_p (fix
->value
, mp
->value
))
16046 /* More than one fix references this entry. */
16048 return move_minipool_fix_forward_ref (mp
, max_mp
, max_address
);
16051 /* Note the insertion point if necessary. */
16053 && mp
->max_address
> max_address
)
16056 /* If we are inserting an 8-bytes aligned quantity and
16057 we have not already found an insertion point, then
16058 make sure that all such 8-byte aligned quantities are
16059 placed at the start of the pool. */
16060 if (ARM_DOUBLEWORD_ALIGN
16062 && fix
->fix_size
>= 8
16063 && mp
->fix_size
< 8)
16066 max_address
= mp
->max_address
;
16070 /* The value is not currently in the minipool, so we need to create
16071 a new entry for it. If MAX_MP is NULL, the entry will be put on
16072 the end of the list since the placement is less constrained than
16073 any existing entry. Otherwise, we insert the new fix before
16074 MAX_MP and, if necessary, adjust the constraints on the other
16077 mp
->fix_size
= fix
->fix_size
;
16078 mp
->mode
= fix
->mode
;
16079 mp
->value
= fix
->value
;
16081 /* Not yet required for a backwards ref. */
16082 mp
->min_address
= -65536;
16084 if (max_mp
== NULL
)
16086 mp
->max_address
= max_address
;
16088 mp
->prev
= minipool_vector_tail
;
16090 if (mp
->prev
== NULL
)
16092 minipool_vector_head
= mp
;
16093 minipool_vector_label
= gen_label_rtx ();
16096 mp
->prev
->next
= mp
;
16098 minipool_vector_tail
= mp
;
16102 if (max_address
> max_mp
->max_address
- mp
->fix_size
)
16103 mp
->max_address
= max_mp
->max_address
- mp
->fix_size
;
16105 mp
->max_address
= max_address
;
16108 mp
->prev
= max_mp
->prev
;
16110 if (mp
->prev
!= NULL
)
16111 mp
->prev
->next
= mp
;
16113 minipool_vector_head
= mp
;
16116 /* Save the new entry. */
16119 /* Scan over the preceding entries and adjust their addresses as
16121 while (mp
->prev
!= NULL
16122 && mp
->prev
->max_address
> mp
->max_address
- mp
->prev
->fix_size
)
16124 mp
->prev
->max_address
= mp
->max_address
- mp
->prev
->fix_size
;
16132 move_minipool_fix_backward_ref (Mnode
*mp
, Mnode
*min_mp
,
16133 HOST_WIDE_INT min_address
)
16135 HOST_WIDE_INT offset
;
16137 /* The code below assumes these are different. */
16138 gcc_assert (mp
!= min_mp
);
16140 if (min_mp
== NULL
)
16142 if (min_address
> mp
->min_address
)
16143 mp
->min_address
= min_address
;
16147 /* We will adjust this below if it is too loose. */
16148 mp
->min_address
= min_address
;
16150 /* Unlink MP from its current position. Since min_mp is non-null,
16151 mp->next must be non-null. */
16152 mp
->next
->prev
= mp
->prev
;
16153 if (mp
->prev
!= NULL
)
16154 mp
->prev
->next
= mp
->next
;
16156 minipool_vector_head
= mp
->next
;
16158 /* Reinsert it after MIN_MP. */
16160 mp
->next
= min_mp
->next
;
16162 if (mp
->next
!= NULL
)
16163 mp
->next
->prev
= mp
;
16165 minipool_vector_tail
= mp
;
16171 for (mp
= minipool_vector_head
; mp
!= NULL
; mp
= mp
->next
)
16173 mp
->offset
= offset
;
16174 if (mp
->refcount
> 0)
16175 offset
+= mp
->fix_size
;
16177 if (mp
->next
&& mp
->next
->min_address
< mp
->min_address
+ mp
->fix_size
)
16178 mp
->next
->min_address
= mp
->min_address
+ mp
->fix_size
;
16184 /* Add a constant to the minipool for a backward reference. Returns the
16185 node added or NULL if the constant will not fit in this pool.
16187 Note that the code for insertion for a backwards reference can be
16188 somewhat confusing because the calculated offsets for each fix do
16189 not take into account the size of the pool (which is still under
16192 add_minipool_backward_ref (Mfix
*fix
)
16194 /* If set, min_mp is the last pool_entry that has a lower constraint
16195 than the one we are trying to add. */
16196 Mnode
*min_mp
= NULL
;
16197 /* This can be negative, since it is only a constraint. */
16198 HOST_WIDE_INT min_address
= fix
->address
- fix
->backwards
;
16201 /* If we can't reach the current pool from this insn, or if we can't
16202 insert this entry at the end of the pool without pushing other
16203 fixes out of range, then we don't try. This ensures that we
16204 can't fail later on. */
16205 if (min_address
>= minipool_barrier
->address
16206 || (minipool_vector_tail
->min_address
+ fix
->fix_size
16207 >= minipool_barrier
->address
))
16210 /* Scan the pool to see if a constant with the same value has
16211 already been added. While we are doing this, also note the
16212 location where we must insert the constant if it doesn't already
16214 for (mp
= minipool_vector_tail
; mp
!= NULL
; mp
= mp
->prev
)
16216 if (GET_CODE (fix
->value
) == GET_CODE (mp
->value
)
16217 && fix
->mode
== mp
->mode
16218 && (!LABEL_P (fix
->value
)
16219 || (CODE_LABEL_NUMBER (fix
->value
)
16220 == CODE_LABEL_NUMBER (mp
->value
)))
16221 && rtx_equal_p (fix
->value
, mp
->value
)
16222 /* Check that there is enough slack to move this entry to the
16223 end of the table (this is conservative). */
16224 && (mp
->max_address
16225 > (minipool_barrier
->address
16226 + minipool_vector_tail
->offset
16227 + minipool_vector_tail
->fix_size
)))
16230 return move_minipool_fix_backward_ref (mp
, min_mp
, min_address
);
16233 if (min_mp
!= NULL
)
16234 mp
->min_address
+= fix
->fix_size
;
16237 /* Note the insertion point if necessary. */
16238 if (mp
->min_address
< min_address
)
16240 /* For now, we do not allow the insertion of 8-byte alignment
16241 requiring nodes anywhere but at the start of the pool. */
16242 if (ARM_DOUBLEWORD_ALIGN
16243 && fix
->fix_size
>= 8 && mp
->fix_size
< 8)
16248 else if (mp
->max_address
16249 < minipool_barrier
->address
+ mp
->offset
+ fix
->fix_size
)
16251 /* Inserting before this entry would push the fix beyond
16252 its maximum address (which can happen if we have
16253 re-located a forwards fix); force the new fix to come
16255 if (ARM_DOUBLEWORD_ALIGN
16256 && fix
->fix_size
>= 8 && mp
->fix_size
< 8)
16261 min_address
= mp
->min_address
+ fix
->fix_size
;
16264 /* Do not insert a non-8-byte aligned quantity before 8-byte
16265 aligned quantities. */
16266 else if (ARM_DOUBLEWORD_ALIGN
16267 && fix
->fix_size
< 8
16268 && mp
->fix_size
>= 8)
16271 min_address
= mp
->min_address
+ fix
->fix_size
;
16276 /* We need to create a new entry. */
16278 mp
->fix_size
= fix
->fix_size
;
16279 mp
->mode
= fix
->mode
;
16280 mp
->value
= fix
->value
;
16282 mp
->max_address
= minipool_barrier
->address
+ 65536;
16284 mp
->min_address
= min_address
;
16286 if (min_mp
== NULL
)
16289 mp
->next
= minipool_vector_head
;
16291 if (mp
->next
== NULL
)
16293 minipool_vector_tail
= mp
;
16294 minipool_vector_label
= gen_label_rtx ();
16297 mp
->next
->prev
= mp
;
16299 minipool_vector_head
= mp
;
16303 mp
->next
= min_mp
->next
;
16307 if (mp
->next
!= NULL
)
16308 mp
->next
->prev
= mp
;
16310 minipool_vector_tail
= mp
;
16313 /* Save the new entry. */
16321 /* Scan over the following entries and adjust their offsets. */
16322 while (mp
->next
!= NULL
)
16324 if (mp
->next
->min_address
< mp
->min_address
+ mp
->fix_size
)
16325 mp
->next
->min_address
= mp
->min_address
+ mp
->fix_size
;
16328 mp
->next
->offset
= mp
->offset
+ mp
->fix_size
;
16330 mp
->next
->offset
= mp
->offset
;
16339 assign_minipool_offsets (Mfix
*barrier
)
16341 HOST_WIDE_INT offset
= 0;
16344 minipool_barrier
= barrier
;
16346 for (mp
= minipool_vector_head
; mp
!= NULL
; mp
= mp
->next
)
16348 mp
->offset
= offset
;
16350 if (mp
->refcount
> 0)
16351 offset
+= mp
->fix_size
;
16355 /* Output the literal table */
16357 dump_minipool (rtx_insn
*scan
)
16363 if (ARM_DOUBLEWORD_ALIGN
)
16364 for (mp
= minipool_vector_head
; mp
!= NULL
; mp
= mp
->next
)
16365 if (mp
->refcount
> 0 && mp
->fix_size
>= 8)
16372 fprintf (dump_file
,
16373 ";; Emitting minipool after insn %u; address %ld; align %d (bytes)\n",
16374 INSN_UID (scan
), (unsigned long) minipool_barrier
->address
, align64
? 8 : 4);
16376 scan
= emit_label_after (gen_label_rtx (), scan
);
16377 scan
= emit_insn_after (align64
? gen_align_8 () : gen_align_4 (), scan
);
16378 scan
= emit_label_after (minipool_vector_label
, scan
);
16380 for (mp
= minipool_vector_head
; mp
!= NULL
; mp
= nmp
)
16382 if (mp
->refcount
> 0)
16386 fprintf (dump_file
,
16387 ";; Offset %u, min %ld, max %ld ",
16388 (unsigned) mp
->offset
, (unsigned long) mp
->min_address
,
16389 (unsigned long) mp
->max_address
);
16390 arm_print_value (dump_file
, mp
->value
);
16391 fputc ('\n', dump_file
);
16394 rtx val
= copy_rtx (mp
->value
);
16396 switch (GET_MODE_SIZE (mp
->mode
))
16398 #ifdef HAVE_consttable_1
16400 scan
= emit_insn_after (gen_consttable_1 (val
), scan
);
16404 #ifdef HAVE_consttable_2
16406 scan
= emit_insn_after (gen_consttable_2 (val
), scan
);
16410 #ifdef HAVE_consttable_4
16412 scan
= emit_insn_after (gen_consttable_4 (val
), scan
);
16416 #ifdef HAVE_consttable_8
16418 scan
= emit_insn_after (gen_consttable_8 (val
), scan
);
16422 #ifdef HAVE_consttable_16
16424 scan
= emit_insn_after (gen_consttable_16 (val
), scan
);
16429 gcc_unreachable ();
16437 minipool_vector_head
= minipool_vector_tail
= NULL
;
16438 scan
= emit_insn_after (gen_consttable_end (), scan
);
16439 scan
= emit_barrier_after (scan
);
16442 /* Return the cost of forcibly inserting a barrier after INSN. */
16444 arm_barrier_cost (rtx_insn
*insn
)
16446 /* Basing the location of the pool on the loop depth is preferable,
16447 but at the moment, the basic block information seems to be
16448 corrupt by this stage of the compilation. */
16449 int base_cost
= 50;
16450 rtx_insn
*next
= next_nonnote_insn (insn
);
16452 if (next
!= NULL
&& LABEL_P (next
))
16455 switch (GET_CODE (insn
))
16458 /* It will always be better to place the table before the label, rather
16467 return base_cost
- 10;
16470 return base_cost
+ 10;
16474 /* Find the best place in the insn stream in the range
16475 (FIX->address,MAX_ADDRESS) to forcibly insert a minipool barrier.
16476 Create the barrier by inserting a jump and add a new fix entry for
16479 create_fix_barrier (Mfix
*fix
, HOST_WIDE_INT max_address
)
16481 HOST_WIDE_INT count
= 0;
16482 rtx_barrier
*barrier
;
16483 rtx_insn
*from
= fix
->insn
;
16484 /* The instruction after which we will insert the jump. */
16485 rtx_insn
*selected
= NULL
;
16487 /* The address at which the jump instruction will be placed. */
16488 HOST_WIDE_INT selected_address
;
16490 HOST_WIDE_INT max_count
= max_address
- fix
->address
;
16491 rtx_code_label
*label
= gen_label_rtx ();
16493 selected_cost
= arm_barrier_cost (from
);
16494 selected_address
= fix
->address
;
16496 while (from
&& count
< max_count
)
16498 rtx_jump_table_data
*tmp
;
16501 /* This code shouldn't have been called if there was a natural barrier
16503 gcc_assert (!BARRIER_P (from
));
16505 /* Count the length of this insn. This must stay in sync with the
16506 code that pushes minipool fixes. */
16507 if (LABEL_P (from
))
16508 count
+= get_label_padding (from
);
16510 count
+= get_attr_length (from
);
16512 /* If there is a jump table, add its length. */
16513 if (tablejump_p (from
, NULL
, &tmp
))
16515 count
+= get_jump_table_size (tmp
);
16517 /* Jump tables aren't in a basic block, so base the cost on
16518 the dispatch insn. If we select this location, we will
16519 still put the pool after the table. */
16520 new_cost
= arm_barrier_cost (from
);
16522 if (count
< max_count
16523 && (!selected
|| new_cost
<= selected_cost
))
16526 selected_cost
= new_cost
;
16527 selected_address
= fix
->address
+ count
;
16530 /* Continue after the dispatch table. */
16531 from
= NEXT_INSN (tmp
);
16535 new_cost
= arm_barrier_cost (from
);
16537 if (count
< max_count
16538 && (!selected
|| new_cost
<= selected_cost
))
16541 selected_cost
= new_cost
;
16542 selected_address
= fix
->address
+ count
;
16545 from
= NEXT_INSN (from
);
16548 /* Make sure that we found a place to insert the jump. */
16549 gcc_assert (selected
);
16551 /* Create a new JUMP_INSN that branches around a barrier. */
16552 from
= emit_jump_insn_after (gen_jump (label
), selected
);
16553 JUMP_LABEL (from
) = label
;
16554 barrier
= emit_barrier_after (from
);
16555 emit_label_after (label
, barrier
);
16557 /* Create a minipool barrier entry for the new barrier. */
16558 new_fix
= (Mfix
*) obstack_alloc (&minipool_obstack
, sizeof (* new_fix
));
16559 new_fix
->insn
= barrier
;
16560 new_fix
->address
= selected_address
;
16561 new_fix
->next
= fix
->next
;
16562 fix
->next
= new_fix
;
16567 /* Record that there is a natural barrier in the insn stream at
16570 push_minipool_barrier (rtx_insn
*insn
, HOST_WIDE_INT address
)
16572 Mfix
* fix
= (Mfix
*) obstack_alloc (&minipool_obstack
, sizeof (* fix
));
16575 fix
->address
= address
;
16578 if (minipool_fix_head
!= NULL
)
16579 minipool_fix_tail
->next
= fix
;
16581 minipool_fix_head
= fix
;
16583 minipool_fix_tail
= fix
;
16586 /* Record INSN, which will need fixing up to load a value from the
16587 minipool. ADDRESS is the offset of the insn since the start of the
16588 function; LOC is a pointer to the part of the insn which requires
16589 fixing; VALUE is the constant that must be loaded, which is of type
16592 push_minipool_fix (rtx_insn
*insn
, HOST_WIDE_INT address
, rtx
*loc
,
16593 machine_mode mode
, rtx value
)
16595 gcc_assert (!arm_disable_literal_pool
);
16596 Mfix
* fix
= (Mfix
*) obstack_alloc (&minipool_obstack
, sizeof (* fix
));
16599 fix
->address
= address
;
16602 fix
->fix_size
= MINIPOOL_FIX_SIZE (mode
);
16603 fix
->value
= value
;
16604 fix
->forwards
= get_attr_pool_range (insn
);
16605 fix
->backwards
= get_attr_neg_pool_range (insn
);
16606 fix
->minipool
= NULL
;
16608 /* If an insn doesn't have a range defined for it, then it isn't
16609 expecting to be reworked by this code. Better to stop now than
16610 to generate duff assembly code. */
16611 gcc_assert (fix
->forwards
|| fix
->backwards
);
16613 /* If an entry requires 8-byte alignment then assume all constant pools
16614 require 4 bytes of padding. Trying to do this later on a per-pool
16615 basis is awkward because existing pool entries have to be modified. */
16616 if (ARM_DOUBLEWORD_ALIGN
&& fix
->fix_size
>= 8)
16621 fprintf (dump_file
,
16622 ";; %smode fixup for i%d; addr %lu, range (%ld,%ld): ",
16623 GET_MODE_NAME (mode
),
16624 INSN_UID (insn
), (unsigned long) address
,
16625 -1 * (long)fix
->backwards
, (long)fix
->forwards
);
16626 arm_print_value (dump_file
, fix
->value
);
16627 fprintf (dump_file
, "\n");
16630 /* Add it to the chain of fixes. */
16633 if (minipool_fix_head
!= NULL
)
16634 minipool_fix_tail
->next
= fix
;
16636 minipool_fix_head
= fix
;
16638 minipool_fix_tail
= fix
;
16641 /* Return maximum allowed cost of synthesizing a 64-bit constant VAL inline.
16642 Returns the number of insns needed, or 99 if we always want to synthesize
16645 arm_max_const_double_inline_cost ()
16647 return ((optimize_size
|| arm_ld_sched
) ? 3 : 4);
16650 /* Return the cost of synthesizing a 64-bit constant VAL inline.
16651 Returns the number of insns needed, or 99 if we don't know how to
16654 arm_const_double_inline_cost (rtx val
)
16656 rtx lowpart
, highpart
;
16659 mode
= GET_MODE (val
);
16661 if (mode
== VOIDmode
)
16664 gcc_assert (GET_MODE_SIZE (mode
) == 8);
16666 lowpart
= gen_lowpart (SImode
, val
);
16667 highpart
= gen_highpart_mode (SImode
, mode
, val
);
16669 gcc_assert (CONST_INT_P (lowpart
));
16670 gcc_assert (CONST_INT_P (highpart
));
16672 return (arm_gen_constant (SET
, SImode
, NULL_RTX
, INTVAL (lowpart
),
16673 NULL_RTX
, NULL_RTX
, 0, 0)
16674 + arm_gen_constant (SET
, SImode
, NULL_RTX
, INTVAL (highpart
),
16675 NULL_RTX
, NULL_RTX
, 0, 0));
16678 /* Cost of loading a SImode constant. */
16680 arm_const_inline_cost (enum rtx_code code
, rtx val
)
16682 return arm_gen_constant (code
, SImode
, NULL_RTX
, INTVAL (val
),
16683 NULL_RTX
, NULL_RTX
, 1, 0);
16686 /* Return true if it is worthwhile to split a 64-bit constant into two
16687 32-bit operations. This is the case if optimizing for size, or
16688 if we have load delay slots, or if one 32-bit part can be done with
16689 a single data operation. */
16691 arm_const_double_by_parts (rtx val
)
16693 machine_mode mode
= GET_MODE (val
);
16696 if (optimize_size
|| arm_ld_sched
)
16699 if (mode
== VOIDmode
)
16702 part
= gen_highpart_mode (SImode
, mode
, val
);
16704 gcc_assert (CONST_INT_P (part
));
16706 if (const_ok_for_arm (INTVAL (part
))
16707 || const_ok_for_arm (~INTVAL (part
)))
16710 part
= gen_lowpart (SImode
, val
);
16712 gcc_assert (CONST_INT_P (part
));
16714 if (const_ok_for_arm (INTVAL (part
))
16715 || const_ok_for_arm (~INTVAL (part
)))
16721 /* Return true if it is possible to inline both the high and low parts
16722 of a 64-bit constant into 32-bit data processing instructions. */
16724 arm_const_double_by_immediates (rtx val
)
16726 machine_mode mode
= GET_MODE (val
);
16729 if (mode
== VOIDmode
)
16732 part
= gen_highpart_mode (SImode
, mode
, val
);
16734 gcc_assert (CONST_INT_P (part
));
16736 if (!const_ok_for_arm (INTVAL (part
)))
16739 part
= gen_lowpart (SImode
, val
);
16741 gcc_assert (CONST_INT_P (part
));
16743 if (!const_ok_for_arm (INTVAL (part
)))
16749 /* Scan INSN and note any of its operands that need fixing.
16750 If DO_PUSHES is false we do not actually push any of the fixups
16753 note_invalid_constants (rtx_insn
*insn
, HOST_WIDE_INT address
, int do_pushes
)
16757 extract_constrain_insn (insn
);
16759 if (recog_data
.n_alternatives
== 0)
16762 /* Fill in recog_op_alt with information about the constraints of
16764 preprocess_constraints (insn
);
16766 const operand_alternative
*op_alt
= which_op_alt ();
16767 for (opno
= 0; opno
< recog_data
.n_operands
; opno
++)
16769 /* Things we need to fix can only occur in inputs. */
16770 if (recog_data
.operand_type
[opno
] != OP_IN
)
16773 /* If this alternative is a memory reference, then any mention
16774 of constants in this alternative is really to fool reload
16775 into allowing us to accept one there. We need to fix them up
16776 now so that we output the right code. */
16777 if (op_alt
[opno
].memory_ok
)
16779 rtx op
= recog_data
.operand
[opno
];
16781 if (CONSTANT_P (op
))
16784 push_minipool_fix (insn
, address
, recog_data
.operand_loc
[opno
],
16785 recog_data
.operand_mode
[opno
], op
);
16787 else if (MEM_P (op
)
16788 && GET_CODE (XEXP (op
, 0)) == SYMBOL_REF
16789 && CONSTANT_POOL_ADDRESS_P (XEXP (op
, 0)))
16793 rtx cop
= avoid_constant_pool_reference (op
);
16795 /* Casting the address of something to a mode narrower
16796 than a word can cause avoid_constant_pool_reference()
16797 to return the pool reference itself. That's no good to
16798 us here. Lets just hope that we can use the
16799 constant pool value directly. */
16801 cop
= get_pool_constant (XEXP (op
, 0));
16803 push_minipool_fix (insn
, address
,
16804 recog_data
.operand_loc
[opno
],
16805 recog_data
.operand_mode
[opno
], cop
);
16815 /* This function computes the clear mask and PADDING_BITS_TO_CLEAR for structs
16816 and unions in the context of ARMv8-M Security Extensions. It is used as a
16817 helper function for both 'cmse_nonsecure_call' and 'cmse_nonsecure_entry'
16818 functions. The PADDING_BITS_TO_CLEAR pointer can be the base to either one
16819 or four masks, depending on whether it is being computed for a
16820 'cmse_nonsecure_entry' return value or a 'cmse_nonsecure_call' argument
16821 respectively. The tree for the type of the argument or a field within an
16822 argument is passed in ARG_TYPE, the current register this argument or field
16823 starts in is kept in the pointer REGNO and updated accordingly, the bit this
16824 argument or field starts at is passed in STARTING_BIT and the last used bit
16825 is kept in LAST_USED_BIT which is also updated accordingly. */
16827 static unsigned HOST_WIDE_INT
16828 comp_not_to_clear_mask_str_un (tree arg_type
, int * regno
,
16829 uint32_t * padding_bits_to_clear
,
16830 unsigned starting_bit
, int * last_used_bit
)
16833 unsigned HOST_WIDE_INT not_to_clear_reg_mask
= 0;
16835 if (TREE_CODE (arg_type
) == RECORD_TYPE
)
16837 unsigned current_bit
= starting_bit
;
16839 long int offset
, size
;
16842 field
= TYPE_FIELDS (arg_type
);
16845 /* The offset within a structure is always an offset from
16846 the start of that structure. Make sure we take that into the
16847 calculation of the register based offset that we use here. */
16848 offset
= starting_bit
;
16849 offset
+= TREE_INT_CST_ELT (DECL_FIELD_BIT_OFFSET (field
), 0);
16852 /* This is the actual size of the field, for bitfields this is the
16853 bitfield width and not the container size. */
16854 size
= TREE_INT_CST_ELT (DECL_SIZE (field
), 0);
16856 if (*last_used_bit
!= offset
)
16858 if (offset
< *last_used_bit
)
16860 /* This field's offset is before the 'last_used_bit', that
16861 means this field goes on the next register. So we need to
16862 pad the rest of the current register and increase the
16863 register number. */
16865 mask
= ((uint32_t)-1) - ((uint32_t) 1 << *last_used_bit
);
16868 padding_bits_to_clear
[*regno
] |= mask
;
16869 not_to_clear_reg_mask
|= HOST_WIDE_INT_1U
<< *regno
;
16874 /* Otherwise we pad the bits between the last field's end and
16875 the start of the new field. */
16878 mask
= ((uint32_t)-1) >> (32 - offset
);
16879 mask
-= ((uint32_t) 1 << *last_used_bit
) - 1;
16880 padding_bits_to_clear
[*regno
] |= mask
;
16882 current_bit
= offset
;
16885 /* Calculate further padding bits for inner structs/unions too. */
16886 if (RECORD_OR_UNION_TYPE_P (TREE_TYPE (field
)))
16888 *last_used_bit
= current_bit
;
16889 not_to_clear_reg_mask
16890 |= comp_not_to_clear_mask_str_un (TREE_TYPE (field
), regno
,
16891 padding_bits_to_clear
, offset
,
16896 /* Update 'current_bit' with this field's size. If the
16897 'current_bit' lies in a subsequent register, update 'regno' and
16898 reset 'current_bit' to point to the current bit in that new
16900 current_bit
+= size
;
16901 while (current_bit
>= 32)
16904 not_to_clear_reg_mask
|= HOST_WIDE_INT_1U
<< *regno
;
16907 *last_used_bit
= current_bit
;
16910 field
= TREE_CHAIN (field
);
16912 not_to_clear_reg_mask
|= HOST_WIDE_INT_1U
<< *regno
;
16914 else if (TREE_CODE (arg_type
) == UNION_TYPE
)
16916 tree field
, field_t
;
16917 int i
, regno_t
, field_size
;
16921 uint32_t padding_bits_to_clear_res
[NUM_ARG_REGS
]
16922 = {-1, -1, -1, -1};
16924 /* To compute the padding bits in a union we only consider bits as
16925 padding bits if they are always either a padding bit or fall outside a
16926 fields size for all fields in the union. */
16927 field
= TYPE_FIELDS (arg_type
);
16930 uint32_t padding_bits_to_clear_t
[NUM_ARG_REGS
]
16931 = {0U, 0U, 0U, 0U};
16932 int last_used_bit_t
= *last_used_bit
;
16934 field_t
= TREE_TYPE (field
);
16936 /* If the field's type is either a record or a union make sure to
16937 compute their padding bits too. */
16938 if (RECORD_OR_UNION_TYPE_P (field_t
))
16939 not_to_clear_reg_mask
16940 |= comp_not_to_clear_mask_str_un (field_t
, ®no_t
,
16941 &padding_bits_to_clear_t
[0],
16942 starting_bit
, &last_used_bit_t
);
16945 field_size
= TREE_INT_CST_ELT (DECL_SIZE (field
), 0);
16946 regno_t
= (field_size
/ 32) + *regno
;
16947 last_used_bit_t
= (starting_bit
+ field_size
) % 32;
16950 for (i
= *regno
; i
< regno_t
; i
++)
16952 /* For all but the last register used by this field only keep the
16953 padding bits that were padding bits in this field. */
16954 padding_bits_to_clear_res
[i
] &= padding_bits_to_clear_t
[i
];
16957 /* For the last register, keep all padding bits that were padding
16958 bits in this field and any padding bits that are still valid
16959 as padding bits but fall outside of this field's size. */
16960 mask
= (((uint32_t) -1) - ((uint32_t) 1 << last_used_bit_t
)) + 1;
16961 padding_bits_to_clear_res
[regno_t
]
16962 &= padding_bits_to_clear_t
[regno_t
] | mask
;
16964 /* Update the maximum size of the fields in terms of registers used
16965 ('max_reg') and the 'last_used_bit' in said register. */
16966 if (max_reg
< regno_t
)
16969 max_bit
= last_used_bit_t
;
16971 else if (max_reg
== regno_t
&& max_bit
< last_used_bit_t
)
16972 max_bit
= last_used_bit_t
;
16974 field
= TREE_CHAIN (field
);
16977 /* Update the current padding_bits_to_clear using the intersection of the
16978 padding bits of all the fields. */
16979 for (i
=*regno
; i
< max_reg
; i
++)
16980 padding_bits_to_clear
[i
] |= padding_bits_to_clear_res
[i
];
16982 /* Do not keep trailing padding bits, we do not know yet whether this
16983 is the end of the argument. */
16984 mask
= ((uint32_t) 1 << max_bit
) - 1;
16985 padding_bits_to_clear
[max_reg
]
16986 |= padding_bits_to_clear_res
[max_reg
] & mask
;
16989 *last_used_bit
= max_bit
;
16992 /* This function should only be used for structs and unions. */
16993 gcc_unreachable ();
16995 return not_to_clear_reg_mask
;
16998 /* In the context of ARMv8-M Security Extensions, this function is used for both
16999 'cmse_nonsecure_call' and 'cmse_nonsecure_entry' functions to compute what
17000 registers are used when returning or passing arguments, which is then
17001 returned as a mask. It will also compute a mask to indicate padding/unused
17002 bits for each of these registers, and passes this through the
17003 PADDING_BITS_TO_CLEAR pointer. The tree of the argument type is passed in
17004 ARG_TYPE, the rtl representation of the argument is passed in ARG_RTX and
17005 the starting register used to pass this argument or return value is passed
17006 in REGNO. It makes use of 'comp_not_to_clear_mask_str_un' to compute these
17007 for struct and union types. */
17009 static unsigned HOST_WIDE_INT
17010 compute_not_to_clear_mask (tree arg_type
, rtx arg_rtx
, int regno
,
17011 uint32_t * padding_bits_to_clear
)
17014 int last_used_bit
= 0;
17015 unsigned HOST_WIDE_INT not_to_clear_mask
;
17017 if (RECORD_OR_UNION_TYPE_P (arg_type
))
17020 = comp_not_to_clear_mask_str_un (arg_type
, ®no
,
17021 padding_bits_to_clear
, 0,
17025 /* If the 'last_used_bit' is not zero, that means we are still using a
17026 part of the last 'regno'. In such cases we must clear the trailing
17027 bits. Otherwise we are not using regno and we should mark it as to
17029 if (last_used_bit
!= 0)
17030 padding_bits_to_clear
[regno
]
17031 |= ((uint32_t)-1) - ((uint32_t) 1 << last_used_bit
) + 1;
17033 not_to_clear_mask
&= ~(HOST_WIDE_INT_1U
<< regno
);
17037 not_to_clear_mask
= 0;
17038 /* We are not dealing with structs nor unions. So these arguments may be
17039 passed in floating point registers too. In some cases a BLKmode is
17040 used when returning or passing arguments in multiple VFP registers. */
17041 if (GET_MODE (arg_rtx
) == BLKmode
)
17046 /* This should really only occur when dealing with the hard-float
17048 gcc_assert (TARGET_HARD_FLOAT_ABI
);
17050 for (i
= 0; i
< XVECLEN (arg_rtx
, 0); i
++)
17052 reg
= XEXP (XVECEXP (arg_rtx
, 0, i
), 0);
17053 gcc_assert (REG_P (reg
));
17055 not_to_clear_mask
|= HOST_WIDE_INT_1U
<< REGNO (reg
);
17057 /* If we are dealing with DF mode, make sure we don't
17058 clear either of the registers it addresses. */
17059 arg_regs
= ARM_NUM_REGS (GET_MODE (reg
));
17062 unsigned HOST_WIDE_INT mask
;
17063 mask
= HOST_WIDE_INT_1U
<< (REGNO (reg
) + arg_regs
);
17064 mask
-= HOST_WIDE_INT_1U
<< REGNO (reg
);
17065 not_to_clear_mask
|= mask
;
17071 /* Otherwise we can rely on the MODE to determine how many registers
17072 are being used by this argument. */
17073 int arg_regs
= ARM_NUM_REGS (GET_MODE (arg_rtx
));
17074 not_to_clear_mask
|= HOST_WIDE_INT_1U
<< REGNO (arg_rtx
);
17077 unsigned HOST_WIDE_INT
17078 mask
= HOST_WIDE_INT_1U
<< (REGNO (arg_rtx
) + arg_regs
);
17079 mask
-= HOST_WIDE_INT_1U
<< REGNO (arg_rtx
);
17080 not_to_clear_mask
|= mask
;
17085 return not_to_clear_mask
;
17088 /* Clear registers secret before doing a cmse_nonsecure_call or returning from
17089 a cmse_nonsecure_entry function. TO_CLEAR_BITMAP indicates which registers
17090 are to be fully cleared, using the value in register CLEARING_REG if more
17091 efficient. The PADDING_BITS_LEN entries array PADDING_BITS_TO_CLEAR gives
17092 the bits that needs to be cleared in caller-saved core registers, with
17093 SCRATCH_REG used as a scratch register for that clearing.
17095 NOTE: one of three following assertions must hold:
17096 - SCRATCH_REG is a low register
17097 - CLEARING_REG is in the set of registers fully cleared (ie. its bit is set
17098 in TO_CLEAR_BITMAP)
17099 - CLEARING_REG is a low register. */
17102 cmse_clear_registers (sbitmap to_clear_bitmap
, uint32_t *padding_bits_to_clear
,
17103 int padding_bits_len
, rtx scratch_reg
, rtx clearing_reg
)
17105 bool saved_clearing
= false;
17106 rtx saved_clearing_reg
= NULL_RTX
;
17107 int i
, regno
, clearing_regno
, minregno
= R0_REGNUM
, maxregno
= minregno
- 1;
17109 gcc_assert (arm_arch_cmse
);
17111 if (!bitmap_empty_p (to_clear_bitmap
))
17113 minregno
= bitmap_first_set_bit (to_clear_bitmap
);
17114 maxregno
= bitmap_last_set_bit (to_clear_bitmap
);
17116 clearing_regno
= REGNO (clearing_reg
);
17118 /* Clear padding bits. */
17119 gcc_assert (padding_bits_len
<= NUM_ARG_REGS
);
17120 for (i
= 0, regno
= R0_REGNUM
; i
< padding_bits_len
; i
++, regno
++)
17123 rtx rtx16
, dest
, cleared_reg
= gen_rtx_REG (SImode
, regno
);
17125 if (padding_bits_to_clear
[i
] == 0)
17128 /* If this is a Thumb-1 target and SCRATCH_REG is not a low register, use
17129 CLEARING_REG as scratch. */
17131 && REGNO (scratch_reg
) > LAST_LO_REGNUM
)
17133 /* clearing_reg is not to be cleared, copy its value into scratch_reg
17134 such that we can use clearing_reg to clear the unused bits in the
17136 if ((clearing_regno
> maxregno
17137 || !bitmap_bit_p (to_clear_bitmap
, clearing_regno
))
17138 && !saved_clearing
)
17140 gcc_assert (clearing_regno
<= LAST_LO_REGNUM
);
17141 emit_move_insn (scratch_reg
, clearing_reg
);
17142 saved_clearing
= true;
17143 saved_clearing_reg
= scratch_reg
;
17145 scratch_reg
= clearing_reg
;
17148 /* Fill the lower half of the negated padding_bits_to_clear[i]. */
17149 mask
= (~padding_bits_to_clear
[i
]) & 0xFFFF;
17150 emit_move_insn (scratch_reg
, gen_int_mode (mask
, SImode
));
17152 /* Fill the top half of the negated padding_bits_to_clear[i]. */
17153 mask
= (~padding_bits_to_clear
[i
]) >> 16;
17154 rtx16
= gen_int_mode (16, SImode
);
17155 dest
= gen_rtx_ZERO_EXTRACT (SImode
, scratch_reg
, rtx16
, rtx16
);
17157 emit_insn (gen_rtx_SET (dest
, gen_int_mode (mask
, SImode
)));
17159 emit_insn (gen_andsi3 (cleared_reg
, cleared_reg
, scratch_reg
));
17161 if (saved_clearing
)
17162 emit_move_insn (clearing_reg
, saved_clearing_reg
);
17165 /* Clear full registers. */
17167 /* If not marked for clearing, clearing_reg already does not contain
17169 if (clearing_regno
<= maxregno
17170 && bitmap_bit_p (to_clear_bitmap
, clearing_regno
))
17172 emit_move_insn (clearing_reg
, const0_rtx
);
17173 emit_use (clearing_reg
);
17174 bitmap_clear_bit (to_clear_bitmap
, clearing_regno
);
17177 for (regno
= minregno
; regno
<= maxregno
; regno
++)
17179 if (!bitmap_bit_p (to_clear_bitmap
, regno
))
17182 if (IS_VFP_REGNUM (regno
))
17184 /* If regno is an even vfp register and its successor is also to
17185 be cleared, use vmov. */
17186 if (TARGET_VFP_DOUBLE
17187 && VFP_REGNO_OK_FOR_DOUBLE (regno
)
17188 && bitmap_bit_p (to_clear_bitmap
, regno
+ 1))
17190 emit_move_insn (gen_rtx_REG (DFmode
, regno
),
17191 CONST1_RTX (DFmode
));
17192 emit_use (gen_rtx_REG (DFmode
, regno
));
17197 emit_move_insn (gen_rtx_REG (SFmode
, regno
),
17198 CONST1_RTX (SFmode
));
17199 emit_use (gen_rtx_REG (SFmode
, regno
));
17204 emit_move_insn (gen_rtx_REG (SImode
, regno
), clearing_reg
);
17205 emit_use (gen_rtx_REG (SImode
, regno
));
17210 /* Clears caller saved registers not used to pass arguments before a
17211 cmse_nonsecure_call. Saving, clearing and restoring of callee saved
17212 registers is done in __gnu_cmse_nonsecure_call libcall.
17213 See libgcc/config/arm/cmse_nonsecure_call.S. */
17216 cmse_nonsecure_call_clear_caller_saved (void)
17220 FOR_EACH_BB_FN (bb
, cfun
)
17224 FOR_BB_INSNS (bb
, insn
)
17226 unsigned address_regnum
, regno
, maxregno
=
17227 TARGET_HARD_FLOAT_ABI
? D7_VFP_REGNUM
: NUM_ARG_REGS
- 1;
17228 auto_sbitmap
to_clear_bitmap (maxregno
+ 1);
17230 rtx pat
, call
, unspec
, clearing_reg
, ip_reg
, shift
;
17232 CUMULATIVE_ARGS args_so_far_v
;
17233 cumulative_args_t args_so_far
;
17234 tree arg_type
, fntype
;
17235 bool first_param
= true;
17236 function_args_iterator args_iter
;
17237 uint32_t padding_bits_to_clear
[4] = {0U, 0U, 0U, 0U};
17239 if (!NONDEBUG_INSN_P (insn
))
17242 if (!CALL_P (insn
))
17245 pat
= PATTERN (insn
);
17246 gcc_assert (GET_CODE (pat
) == PARALLEL
&& XVECLEN (pat
, 0) > 0);
17247 call
= XVECEXP (pat
, 0, 0);
17249 /* Get the real call RTX if the insn sets a value, ie. returns. */
17250 if (GET_CODE (call
) == SET
)
17251 call
= SET_SRC (call
);
17253 /* Check if it is a cmse_nonsecure_call. */
17254 unspec
= XEXP (call
, 0);
17255 if (GET_CODE (unspec
) != UNSPEC
17256 || XINT (unspec
, 1) != UNSPEC_NONSECURE_MEM
)
17259 /* Determine the caller-saved registers we need to clear. */
17260 bitmap_clear (to_clear_bitmap
);
17261 bitmap_set_range (to_clear_bitmap
, R0_REGNUM
, NUM_ARG_REGS
);
17263 /* Only look at the caller-saved floating point registers in case of
17264 -mfloat-abi=hard. For -mfloat-abi=softfp we will be using the
17265 lazy store and loads which clear both caller- and callee-saved
17267 if (TARGET_HARD_FLOAT_ABI
)
17269 auto_sbitmap
float_bitmap (maxregno
+ 1);
17271 bitmap_clear (float_bitmap
);
17272 bitmap_set_range (float_bitmap
, FIRST_VFP_REGNUM
,
17273 D7_VFP_REGNUM
- FIRST_VFP_REGNUM
+ 1);
17274 bitmap_ior (to_clear_bitmap
, to_clear_bitmap
, float_bitmap
);
17277 /* Make sure the register used to hold the function address is not
17279 address
= RTVEC_ELT (XVEC (unspec
, 0), 0);
17280 gcc_assert (MEM_P (address
));
17281 gcc_assert (REG_P (XEXP (address
, 0)));
17282 address_regnum
= REGNO (XEXP (address
, 0));
17283 if (address_regnum
< R0_REGNUM
+ NUM_ARG_REGS
)
17284 bitmap_clear_bit (to_clear_bitmap
, address_regnum
);
17286 /* Set basic block of call insn so that df rescan is performed on
17287 insns inserted here. */
17288 set_block_for_insn (insn
, bb
);
17289 df_set_flags (DF_DEFER_INSN_RESCAN
);
17292 /* Make sure the scheduler doesn't schedule other insns beyond
17294 emit_insn (gen_blockage ());
17296 /* Walk through all arguments and clear registers appropriately.
17298 fntype
= TREE_TYPE (MEM_EXPR (address
));
17299 arm_init_cumulative_args (&args_so_far_v
, fntype
, NULL_RTX
,
17301 args_so_far
= pack_cumulative_args (&args_so_far_v
);
17302 FOREACH_FUNCTION_ARGS (fntype
, arg_type
, args_iter
)
17305 uint64_t to_clear_args_mask
;
17306 machine_mode arg_mode
= TYPE_MODE (arg_type
);
17308 if (VOID_TYPE_P (arg_type
))
17312 arm_function_arg_advance (args_so_far
, arg_mode
, arg_type
,
17315 arg_rtx
= arm_function_arg (args_so_far
, arg_mode
, arg_type
,
17317 gcc_assert (REG_P (arg_rtx
));
17319 = compute_not_to_clear_mask (arg_type
, arg_rtx
,
17321 &padding_bits_to_clear
[0]);
17322 if (to_clear_args_mask
)
17324 for (regno
= R0_REGNUM
; regno
<= maxregno
; regno
++)
17326 if (to_clear_args_mask
& (1ULL << regno
))
17327 bitmap_clear_bit (to_clear_bitmap
, regno
);
17331 first_param
= false;
17334 /* We use right shift and left shift to clear the LSB of the address
17335 we jump to instead of using bic, to avoid having to use an extra
17336 register on Thumb-1. */
17337 clearing_reg
= XEXP (address
, 0);
17338 shift
= gen_rtx_LSHIFTRT (SImode
, clearing_reg
, const1_rtx
);
17339 emit_insn (gen_rtx_SET (clearing_reg
, shift
));
17340 shift
= gen_rtx_ASHIFT (SImode
, clearing_reg
, const1_rtx
);
17341 emit_insn (gen_rtx_SET (clearing_reg
, shift
));
17343 /* Clear caller-saved registers that leak before doing a non-secure
17345 ip_reg
= gen_rtx_REG (SImode
, IP_REGNUM
);
17346 cmse_clear_registers (to_clear_bitmap
, padding_bits_to_clear
,
17347 NUM_ARG_REGS
, ip_reg
, clearing_reg
);
17349 seq
= get_insns ();
17351 emit_insn_before (seq
, insn
);
17356 /* Rewrite move insn into subtract of 0 if the condition codes will
17357 be useful in next conditional jump insn. */
17360 thumb1_reorg (void)
17364 FOR_EACH_BB_FN (bb
, cfun
)
17367 rtx cmp
, op0
, op1
, set
= NULL
;
17368 rtx_insn
*prev
, *insn
= BB_END (bb
);
17369 bool insn_clobbered
= false;
17371 while (insn
!= BB_HEAD (bb
) && !NONDEBUG_INSN_P (insn
))
17372 insn
= PREV_INSN (insn
);
17374 /* Find the last cbranchsi4_insn in basic block BB. */
17375 if (insn
== BB_HEAD (bb
)
17376 || INSN_CODE (insn
) != CODE_FOR_cbranchsi4_insn
)
17379 /* Get the register with which we are comparing. */
17380 cmp
= XEXP (SET_SRC (PATTERN (insn
)), 0);
17381 op0
= XEXP (cmp
, 0);
17382 op1
= XEXP (cmp
, 1);
17384 /* Check that comparison is against ZERO. */
17385 if (!CONST_INT_P (op1
) || INTVAL (op1
) != 0)
17388 /* Find the first flag setting insn before INSN in basic block BB. */
17389 gcc_assert (insn
!= BB_HEAD (bb
));
17390 for (prev
= PREV_INSN (insn
);
17392 && prev
!= BB_HEAD (bb
)
17394 || DEBUG_INSN_P (prev
)
17395 || ((set
= single_set (prev
)) != NULL
17396 && get_attr_conds (prev
) == CONDS_NOCOND
)));
17397 prev
= PREV_INSN (prev
))
17399 if (reg_set_p (op0
, prev
))
17400 insn_clobbered
= true;
17403 /* Skip if op0 is clobbered by insn other than prev. */
17404 if (insn_clobbered
)
17410 dest
= SET_DEST (set
);
17411 src
= SET_SRC (set
);
17412 if (!low_register_operand (dest
, SImode
)
17413 || !low_register_operand (src
, SImode
))
17416 /* Rewrite move into subtract of 0 if its operand is compared with ZERO
17417 in INSN. Both src and dest of the move insn are checked. */
17418 if (REGNO (op0
) == REGNO (src
) || REGNO (op0
) == REGNO (dest
))
17420 dest
= copy_rtx (dest
);
17421 src
= copy_rtx (src
);
17422 src
= gen_rtx_MINUS (SImode
, src
, const0_rtx
);
17423 PATTERN (prev
) = gen_rtx_SET (dest
, src
);
17424 INSN_CODE (prev
) = -1;
17425 /* Set test register in INSN to dest. */
17426 XEXP (cmp
, 0) = copy_rtx (dest
);
17427 INSN_CODE (insn
) = -1;
17432 /* Convert instructions to their cc-clobbering variant if possible, since
17433 that allows us to use smaller encodings. */
17436 thumb2_reorg (void)
17441 INIT_REG_SET (&live
);
17443 /* We are freeing block_for_insn in the toplev to keep compatibility
17444 with old MDEP_REORGS that are not CFG based. Recompute it now. */
17445 compute_bb_for_insn ();
17448 enum Convert_Action
{SKIP
, CONV
, SWAP_CONV
};
17450 FOR_EACH_BB_FN (bb
, cfun
)
17452 if ((current_tune
->disparage_flag_setting_t16_encodings
17453 == tune_params::DISPARAGE_FLAGS_ALL
)
17454 && optimize_bb_for_speed_p (bb
))
17458 Convert_Action action
= SKIP
;
17459 Convert_Action action_for_partial_flag_setting
17460 = ((current_tune
->disparage_flag_setting_t16_encodings
17461 != tune_params::DISPARAGE_FLAGS_NEITHER
)
17462 && optimize_bb_for_speed_p (bb
))
17465 COPY_REG_SET (&live
, DF_LR_OUT (bb
));
17466 df_simulate_initialize_backwards (bb
, &live
);
17467 FOR_BB_INSNS_REVERSE (bb
, insn
)
17469 if (NONJUMP_INSN_P (insn
)
17470 && !REGNO_REG_SET_P (&live
, CC_REGNUM
)
17471 && GET_CODE (PATTERN (insn
)) == SET
)
17474 rtx pat
= PATTERN (insn
);
17475 rtx dst
= XEXP (pat
, 0);
17476 rtx src
= XEXP (pat
, 1);
17477 rtx op0
= NULL_RTX
, op1
= NULL_RTX
;
17479 if (UNARY_P (src
) || BINARY_P (src
))
17480 op0
= XEXP (src
, 0);
17482 if (BINARY_P (src
))
17483 op1
= XEXP (src
, 1);
17485 if (low_register_operand (dst
, SImode
))
17487 switch (GET_CODE (src
))
17490 /* Adding two registers and storing the result
17491 in the first source is already a 16-bit
17493 if (rtx_equal_p (dst
, op0
)
17494 && register_operand (op1
, SImode
))
17497 if (low_register_operand (op0
, SImode
))
17499 /* ADDS <Rd>,<Rn>,<Rm> */
17500 if (low_register_operand (op1
, SImode
))
17502 /* ADDS <Rdn>,#<imm8> */
17503 /* SUBS <Rdn>,#<imm8> */
17504 else if (rtx_equal_p (dst
, op0
)
17505 && CONST_INT_P (op1
)
17506 && IN_RANGE (INTVAL (op1
), -255, 255))
17508 /* ADDS <Rd>,<Rn>,#<imm3> */
17509 /* SUBS <Rd>,<Rn>,#<imm3> */
17510 else if (CONST_INT_P (op1
)
17511 && IN_RANGE (INTVAL (op1
), -7, 7))
17514 /* ADCS <Rd>, <Rn> */
17515 else if (GET_CODE (XEXP (src
, 0)) == PLUS
17516 && rtx_equal_p (XEXP (XEXP (src
, 0), 0), dst
)
17517 && low_register_operand (XEXP (XEXP (src
, 0), 1),
17519 && COMPARISON_P (op1
)
17520 && cc_register (XEXP (op1
, 0), VOIDmode
)
17521 && maybe_get_arm_condition_code (op1
) == ARM_CS
17522 && XEXP (op1
, 1) == const0_rtx
)
17527 /* RSBS <Rd>,<Rn>,#0
17528 Not handled here: see NEG below. */
17529 /* SUBS <Rd>,<Rn>,#<imm3>
17531 Not handled here: see PLUS above. */
17532 /* SUBS <Rd>,<Rn>,<Rm> */
17533 if (low_register_operand (op0
, SImode
)
17534 && low_register_operand (op1
, SImode
))
17539 /* MULS <Rdm>,<Rn>,<Rdm>
17540 As an exception to the rule, this is only used
17541 when optimizing for size since MULS is slow on all
17542 known implementations. We do not even want to use
17543 MULS in cold code, if optimizing for speed, so we
17544 test the global flag here. */
17545 if (!optimize_size
)
17547 /* Fall through. */
17551 /* ANDS <Rdn>,<Rm> */
17552 if (rtx_equal_p (dst
, op0
)
17553 && low_register_operand (op1
, SImode
))
17554 action
= action_for_partial_flag_setting
;
17555 else if (rtx_equal_p (dst
, op1
)
17556 && low_register_operand (op0
, SImode
))
17557 action
= action_for_partial_flag_setting
== SKIP
17558 ? SKIP
: SWAP_CONV
;
17564 /* ASRS <Rdn>,<Rm> */
17565 /* LSRS <Rdn>,<Rm> */
17566 /* LSLS <Rdn>,<Rm> */
17567 if (rtx_equal_p (dst
, op0
)
17568 && low_register_operand (op1
, SImode
))
17569 action
= action_for_partial_flag_setting
;
17570 /* ASRS <Rd>,<Rm>,#<imm5> */
17571 /* LSRS <Rd>,<Rm>,#<imm5> */
17572 /* LSLS <Rd>,<Rm>,#<imm5> */
17573 else if (low_register_operand (op0
, SImode
)
17574 && CONST_INT_P (op1
)
17575 && IN_RANGE (INTVAL (op1
), 0, 31))
17576 action
= action_for_partial_flag_setting
;
17580 /* RORS <Rdn>,<Rm> */
17581 if (rtx_equal_p (dst
, op0
)
17582 && low_register_operand (op1
, SImode
))
17583 action
= action_for_partial_flag_setting
;
17587 /* MVNS <Rd>,<Rm> */
17588 if (low_register_operand (op0
, SImode
))
17589 action
= action_for_partial_flag_setting
;
17593 /* NEGS <Rd>,<Rm> (a.k.a RSBS) */
17594 if (low_register_operand (op0
, SImode
))
17599 /* MOVS <Rd>,#<imm8> */
17600 if (CONST_INT_P (src
)
17601 && IN_RANGE (INTVAL (src
), 0, 255))
17602 action
= action_for_partial_flag_setting
;
17606 /* MOVS and MOV<c> with registers have different
17607 encodings, so are not relevant here. */
17615 if (action
!= SKIP
)
17617 rtx ccreg
= gen_rtx_REG (CCmode
, CC_REGNUM
);
17618 rtx clobber
= gen_rtx_CLOBBER (VOIDmode
, ccreg
);
17621 if (action
== SWAP_CONV
)
17623 src
= copy_rtx (src
);
17624 XEXP (src
, 0) = op1
;
17625 XEXP (src
, 1) = op0
;
17626 pat
= gen_rtx_SET (dst
, src
);
17627 vec
= gen_rtvec (2, pat
, clobber
);
17629 else /* action == CONV */
17630 vec
= gen_rtvec (2, pat
, clobber
);
17632 PATTERN (insn
) = gen_rtx_PARALLEL (VOIDmode
, vec
);
17633 INSN_CODE (insn
) = -1;
17637 if (NONDEBUG_INSN_P (insn
))
17638 df_simulate_one_insn_backwards (bb
, insn
, &live
);
17642 CLEAR_REG_SET (&live
);
17645 /* Gcc puts the pool in the wrong place for ARM, since we can only
17646 load addresses a limited distance around the pc. We do some
17647 special munging to move the constant pool values to the correct
17648 point in the code. */
17653 HOST_WIDE_INT address
= 0;
17657 cmse_nonsecure_call_clear_caller_saved ();
17659 /* We cannot run the Thumb passes for thunks because there is no CFG. */
17660 if (cfun
->is_thunk
)
17662 else if (TARGET_THUMB1
)
17664 else if (TARGET_THUMB2
)
17667 /* Ensure all insns that must be split have been split at this point.
17668 Otherwise, the pool placement code below may compute incorrect
17669 insn lengths. Note that when optimizing, all insns have already
17670 been split at this point. */
17672 split_all_insns_noflow ();
17674 /* Make sure we do not attempt to create a literal pool even though it should
17675 no longer be necessary to create any. */
17676 if (arm_disable_literal_pool
)
17679 minipool_fix_head
= minipool_fix_tail
= NULL
;
17681 /* The first insn must always be a note, or the code below won't
17682 scan it properly. */
17683 insn
= get_insns ();
17684 gcc_assert (NOTE_P (insn
));
17687 /* Scan all the insns and record the operands that will need fixing. */
17688 for (insn
= next_nonnote_insn (insn
); insn
; insn
= next_nonnote_insn (insn
))
17690 if (BARRIER_P (insn
))
17691 push_minipool_barrier (insn
, address
);
17692 else if (INSN_P (insn
))
17694 rtx_jump_table_data
*table
;
17696 note_invalid_constants (insn
, address
, true);
17697 address
+= get_attr_length (insn
);
17699 /* If the insn is a vector jump, add the size of the table
17700 and skip the table. */
17701 if (tablejump_p (insn
, NULL
, &table
))
17703 address
+= get_jump_table_size (table
);
17707 else if (LABEL_P (insn
))
17708 /* Add the worst-case padding due to alignment. We don't add
17709 the _current_ padding because the minipool insertions
17710 themselves might change it. */
17711 address
+= get_label_padding (insn
);
17714 fix
= minipool_fix_head
;
17716 /* Now scan the fixups and perform the required changes. */
17721 Mfix
* last_added_fix
;
17722 Mfix
* last_barrier
= NULL
;
17725 /* Skip any further barriers before the next fix. */
17726 while (fix
&& BARRIER_P (fix
->insn
))
17729 /* No more fixes. */
17733 last_added_fix
= NULL
;
17735 for (ftmp
= fix
; ftmp
; ftmp
= ftmp
->next
)
17737 if (BARRIER_P (ftmp
->insn
))
17739 if (ftmp
->address
>= minipool_vector_head
->max_address
)
17742 last_barrier
= ftmp
;
17744 else if ((ftmp
->minipool
= add_minipool_forward_ref (ftmp
)) == NULL
)
17747 last_added_fix
= ftmp
; /* Keep track of the last fix added. */
17750 /* If we found a barrier, drop back to that; any fixes that we
17751 could have reached but come after the barrier will now go in
17752 the next mini-pool. */
17753 if (last_barrier
!= NULL
)
17755 /* Reduce the refcount for those fixes that won't go into this
17757 for (fdel
= last_barrier
->next
;
17758 fdel
&& fdel
!= ftmp
;
17761 fdel
->minipool
->refcount
--;
17762 fdel
->minipool
= NULL
;
17765 ftmp
= last_barrier
;
17769 /* ftmp is first fix that we can't fit into this pool and
17770 there no natural barriers that we could use. Insert a
17771 new barrier in the code somewhere between the previous
17772 fix and this one, and arrange to jump around it. */
17773 HOST_WIDE_INT max_address
;
17775 /* The last item on the list of fixes must be a barrier, so
17776 we can never run off the end of the list of fixes without
17777 last_barrier being set. */
17780 max_address
= minipool_vector_head
->max_address
;
17781 /* Check that there isn't another fix that is in range that
17782 we couldn't fit into this pool because the pool was
17783 already too large: we need to put the pool before such an
17784 instruction. The pool itself may come just after the
17785 fix because create_fix_barrier also allows space for a
17786 jump instruction. */
17787 if (ftmp
->address
< max_address
)
17788 max_address
= ftmp
->address
+ 1;
17790 last_barrier
= create_fix_barrier (last_added_fix
, max_address
);
17793 assign_minipool_offsets (last_barrier
);
17797 if (!BARRIER_P (ftmp
->insn
)
17798 && ((ftmp
->minipool
= add_minipool_backward_ref (ftmp
))
17805 /* Scan over the fixes we have identified for this pool, fixing them
17806 up and adding the constants to the pool itself. */
17807 for (this_fix
= fix
; this_fix
&& ftmp
!= this_fix
;
17808 this_fix
= this_fix
->next
)
17809 if (!BARRIER_P (this_fix
->insn
))
17812 = plus_constant (Pmode
,
17813 gen_rtx_LABEL_REF (VOIDmode
,
17814 minipool_vector_label
),
17815 this_fix
->minipool
->offset
);
17816 *this_fix
->loc
= gen_rtx_MEM (this_fix
->mode
, addr
);
17819 dump_minipool (last_barrier
->insn
);
17823 /* From now on we must synthesize any constants that we can't handle
17824 directly. This can happen if the RTL gets split during final
17825 instruction generation. */
17826 cfun
->machine
->after_arm_reorg
= 1;
17828 /* Free the minipool memory. */
17829 obstack_free (&minipool_obstack
, minipool_startobj
);
17832 /* Routines to output assembly language. */
17834 /* Return string representation of passed in real value. */
17835 static const char *
17836 fp_const_from_val (REAL_VALUE_TYPE
*r
)
17838 if (!fp_consts_inited
)
17841 gcc_assert (real_equal (r
, &value_fp0
));
17845 /* OPERANDS[0] is the entire list of insns that constitute pop,
17846 OPERANDS[1] is the base register, RETURN_PC is true iff return insn
17847 is in the list, UPDATE is true iff the list contains explicit
17848 update of base register. */
17850 arm_output_multireg_pop (rtx
*operands
, bool return_pc
, rtx cond
, bool reverse
,
17856 const char *conditional
;
17857 int num_saves
= XVECLEN (operands
[0], 0);
17858 unsigned int regno
;
17859 unsigned int regno_base
= REGNO (operands
[1]);
17860 bool interrupt_p
= IS_INTERRUPT (arm_current_func_type ());
17863 offset
+= update
? 1 : 0;
17864 offset
+= return_pc
? 1 : 0;
17866 /* Is the base register in the list? */
17867 for (i
= offset
; i
< num_saves
; i
++)
17869 regno
= REGNO (XEXP (XVECEXP (operands
[0], 0, i
), 0));
17870 /* If SP is in the list, then the base register must be SP. */
17871 gcc_assert ((regno
!= SP_REGNUM
) || (regno_base
== SP_REGNUM
));
17872 /* If base register is in the list, there must be no explicit update. */
17873 if (regno
== regno_base
)
17874 gcc_assert (!update
);
17877 conditional
= reverse
? "%?%D0" : "%?%d0";
17878 /* Can't use POP if returning from an interrupt. */
17879 if ((regno_base
== SP_REGNUM
) && update
&& !(interrupt_p
&& return_pc
))
17880 sprintf (pattern
, "pop%s\t{", conditional
);
17883 /* Output ldmfd when the base register is SP, otherwise output ldmia.
17884 It's just a convention, their semantics are identical. */
17885 if (regno_base
== SP_REGNUM
)
17886 sprintf (pattern
, "ldmfd%s\t", conditional
);
17888 sprintf (pattern
, "ldmia%s\t", conditional
);
17890 sprintf (pattern
, "ldm%s\t", conditional
);
17892 strcat (pattern
, reg_names
[regno_base
]);
17894 strcat (pattern
, "!, {");
17896 strcat (pattern
, ", {");
17899 /* Output the first destination register. */
17901 reg_names
[REGNO (XEXP (XVECEXP (operands
[0], 0, offset
), 0))]);
17903 /* Output the rest of the destination registers. */
17904 for (i
= offset
+ 1; i
< num_saves
; i
++)
17906 strcat (pattern
, ", ");
17908 reg_names
[REGNO (XEXP (XVECEXP (operands
[0], 0, i
), 0))]);
17911 strcat (pattern
, "}");
17913 if (interrupt_p
&& return_pc
)
17914 strcat (pattern
, "^");
17916 output_asm_insn (pattern
, &cond
);
17920 /* Output the assembly for a store multiple. */
17923 vfp_output_vstmd (rtx
* operands
)
17929 rtx addr_reg
= REG_P (XEXP (operands
[0], 0))
17930 ? XEXP (operands
[0], 0)
17931 : XEXP (XEXP (operands
[0], 0), 0);
17932 bool push_p
= REGNO (addr_reg
) == SP_REGNUM
;
17935 strcpy (pattern
, "vpush%?.64\t{%P1");
17937 strcpy (pattern
, "vstmdb%?.64\t%m0!, {%P1");
17939 p
= strlen (pattern
);
17941 gcc_assert (REG_P (operands
[1]));
17943 base
= (REGNO (operands
[1]) - FIRST_VFP_REGNUM
) / 2;
17944 for (i
= 1; i
< XVECLEN (operands
[2], 0); i
++)
17946 p
+= sprintf (&pattern
[p
], ", d%d", base
+ i
);
17948 strcpy (&pattern
[p
], "}");
17950 output_asm_insn (pattern
, operands
);
17955 /* Emit RTL to save block of VFP register pairs to the stack. Returns the
17956 number of bytes pushed. */
17959 vfp_emit_fstmd (int base_reg
, int count
)
17966 /* Workaround ARM10 VFPr1 bug. Data corruption can occur when exactly two
17967 register pairs are stored by a store multiple insn. We avoid this
17968 by pushing an extra pair. */
17969 if (count
== 2 && !arm_arch6
)
17971 if (base_reg
== LAST_VFP_REGNUM
- 3)
17976 /* FSTMD may not store more than 16 doubleword registers at once. Split
17977 larger stores into multiple parts (up to a maximum of two, in
17982 /* NOTE: base_reg is an internal register number, so each D register
17984 saved
= vfp_emit_fstmd (base_reg
+ 32, count
- 16);
17985 saved
+= vfp_emit_fstmd (base_reg
, 16);
17989 par
= gen_rtx_PARALLEL (VOIDmode
, rtvec_alloc (count
));
17990 dwarf
= gen_rtx_SEQUENCE (VOIDmode
, rtvec_alloc (count
+ 1));
17992 reg
= gen_rtx_REG (DFmode
, base_reg
);
17995 XVECEXP (par
, 0, 0)
17996 = gen_rtx_SET (gen_frame_mem
17998 gen_rtx_PRE_MODIFY (Pmode
,
18001 (Pmode
, stack_pointer_rtx
,
18004 gen_rtx_UNSPEC (BLKmode
,
18005 gen_rtvec (1, reg
),
18006 UNSPEC_PUSH_MULT
));
18008 tmp
= gen_rtx_SET (stack_pointer_rtx
,
18009 plus_constant (Pmode
, stack_pointer_rtx
, -(count
* 8)));
18010 RTX_FRAME_RELATED_P (tmp
) = 1;
18011 XVECEXP (dwarf
, 0, 0) = tmp
;
18013 tmp
= gen_rtx_SET (gen_frame_mem (DFmode
, stack_pointer_rtx
), reg
);
18014 RTX_FRAME_RELATED_P (tmp
) = 1;
18015 XVECEXP (dwarf
, 0, 1) = tmp
;
18017 for (i
= 1; i
< count
; i
++)
18019 reg
= gen_rtx_REG (DFmode
, base_reg
);
18021 XVECEXP (par
, 0, i
) = gen_rtx_USE (VOIDmode
, reg
);
18023 tmp
= gen_rtx_SET (gen_frame_mem (DFmode
,
18024 plus_constant (Pmode
,
18028 RTX_FRAME_RELATED_P (tmp
) = 1;
18029 XVECEXP (dwarf
, 0, i
+ 1) = tmp
;
18032 par
= emit_insn (par
);
18033 add_reg_note (par
, REG_FRAME_RELATED_EXPR
, dwarf
);
18034 RTX_FRAME_RELATED_P (par
) = 1;
18039 /* Returns true if -mcmse has been passed and the function pointed to by 'addr'
18040 has the cmse_nonsecure_call attribute and returns false otherwise. */
18043 detect_cmse_nonsecure_call (tree addr
)
18048 tree fntype
= TREE_TYPE (addr
);
18049 if (use_cmse
&& lookup_attribute ("cmse_nonsecure_call",
18050 TYPE_ATTRIBUTES (fntype
)))
18056 /* Emit a call instruction with pattern PAT. ADDR is the address of
18057 the call target. */
18060 arm_emit_call_insn (rtx pat
, rtx addr
, bool sibcall
)
18064 insn
= emit_call_insn (pat
);
18066 /* The PIC register is live on entry to VxWorks PIC PLT entries.
18067 If the call might use such an entry, add a use of the PIC register
18068 to the instruction's CALL_INSN_FUNCTION_USAGE. */
18069 if (TARGET_VXWORKS_RTP
18072 && GET_CODE (addr
) == SYMBOL_REF
18073 && (SYMBOL_REF_DECL (addr
)
18074 ? !targetm
.binds_local_p (SYMBOL_REF_DECL (addr
))
18075 : !SYMBOL_REF_LOCAL_P (addr
)))
18077 require_pic_register ();
18078 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), cfun
->machine
->pic_reg
);
18081 if (TARGET_AAPCS_BASED
)
18083 /* For AAPCS, IP and CC can be clobbered by veneers inserted by the
18084 linker. We need to add an IP clobber to allow setting
18085 TARGET_CALL_FUSAGE_CONTAINS_NON_CALLEE_CLOBBERS to true. A CC clobber
18086 is not needed since it's a fixed register. */
18087 rtx
*fusage
= &CALL_INSN_FUNCTION_USAGE (insn
);
18088 clobber_reg (fusage
, gen_rtx_REG (word_mode
, IP_REGNUM
));
18092 /* Output a 'call' insn. */
18094 output_call (rtx
*operands
)
18096 gcc_assert (!arm_arch5t
); /* Patterns should call blx <reg> directly. */
18098 /* Handle calls to lr using ip (which may be clobbered in subr anyway). */
18099 if (REGNO (operands
[0]) == LR_REGNUM
)
18101 operands
[0] = gen_rtx_REG (SImode
, IP_REGNUM
);
18102 output_asm_insn ("mov%?\t%0, %|lr", operands
);
18105 output_asm_insn ("mov%?\t%|lr, %|pc", operands
);
18107 if (TARGET_INTERWORK
|| arm_arch4t
)
18108 output_asm_insn ("bx%?\t%0", operands
);
18110 output_asm_insn ("mov%?\t%|pc, %0", operands
);
18115 /* Output a move from arm registers to arm registers of a long double
18116 OPERANDS[0] is the destination.
18117 OPERANDS[1] is the source. */
18119 output_mov_long_double_arm_from_arm (rtx
*operands
)
18121 /* We have to be careful here because the two might overlap. */
18122 int dest_start
= REGNO (operands
[0]);
18123 int src_start
= REGNO (operands
[1]);
18127 if (dest_start
< src_start
)
18129 for (i
= 0; i
< 3; i
++)
18131 ops
[0] = gen_rtx_REG (SImode
, dest_start
+ i
);
18132 ops
[1] = gen_rtx_REG (SImode
, src_start
+ i
);
18133 output_asm_insn ("mov%?\t%0, %1", ops
);
18138 for (i
= 2; i
>= 0; i
--)
18140 ops
[0] = gen_rtx_REG (SImode
, dest_start
+ i
);
18141 ops
[1] = gen_rtx_REG (SImode
, src_start
+ i
);
18142 output_asm_insn ("mov%?\t%0, %1", ops
);
18150 arm_emit_movpair (rtx dest
, rtx src
)
18152 /* If the src is an immediate, simplify it. */
18153 if (CONST_INT_P (src
))
18155 HOST_WIDE_INT val
= INTVAL (src
);
18156 emit_set_insn (dest
, GEN_INT (val
& 0x0000ffff));
18157 if ((val
>> 16) & 0x0000ffff)
18159 emit_set_insn (gen_rtx_ZERO_EXTRACT (SImode
, dest
, GEN_INT (16),
18161 GEN_INT ((val
>> 16) & 0x0000ffff));
18162 rtx_insn
*insn
= get_last_insn ();
18163 set_unique_reg_note (insn
, REG_EQUAL
, copy_rtx (src
));
18167 emit_set_insn (dest
, gen_rtx_HIGH (SImode
, src
));
18168 emit_set_insn (dest
, gen_rtx_LO_SUM (SImode
, dest
, src
));
18169 rtx_insn
*insn
= get_last_insn ();
18170 set_unique_reg_note (insn
, REG_EQUAL
, copy_rtx (src
));
18173 /* Output a move between double words. It must be REG<-MEM
18176 output_move_double (rtx
*operands
, bool emit
, int *count
)
18178 enum rtx_code code0
= GET_CODE (operands
[0]);
18179 enum rtx_code code1
= GET_CODE (operands
[1]);
18184 /* The only case when this might happen is when
18185 you are looking at the length of a DImode instruction
18186 that has an invalid constant in it. */
18187 if (code0
== REG
&& code1
!= MEM
)
18189 gcc_assert (!emit
);
18196 unsigned int reg0
= REGNO (operands
[0]);
18198 otherops
[0] = gen_rtx_REG (SImode
, 1 + reg0
);
18200 gcc_assert (code1
== MEM
); /* Constraints should ensure this. */
18202 switch (GET_CODE (XEXP (operands
[1], 0)))
18209 && !(fix_cm3_ldrd
&& reg0
== REGNO(XEXP (operands
[1], 0))))
18210 output_asm_insn ("ldrd%?\t%0, [%m1]", operands
);
18212 output_asm_insn ("ldmia%?\t%m1, %M0", operands
);
18217 gcc_assert (TARGET_LDRD
);
18219 output_asm_insn ("ldrd%?\t%0, [%m1, #8]!", operands
);
18226 output_asm_insn ("ldrd%?\t%0, [%m1, #-8]!", operands
);
18228 output_asm_insn ("ldmdb%?\t%m1!, %M0", operands
);
18236 output_asm_insn ("ldrd%?\t%0, [%m1], #8", operands
);
18238 output_asm_insn ("ldmia%?\t%m1!, %M0", operands
);
18243 gcc_assert (TARGET_LDRD
);
18245 output_asm_insn ("ldrd%?\t%0, [%m1], #-8", operands
);
18250 /* Autoicrement addressing modes should never have overlapping
18251 base and destination registers, and overlapping index registers
18252 are already prohibited, so this doesn't need to worry about
18254 otherops
[0] = operands
[0];
18255 otherops
[1] = XEXP (XEXP (XEXP (operands
[1], 0), 1), 0);
18256 otherops
[2] = XEXP (XEXP (XEXP (operands
[1], 0), 1), 1);
18258 if (GET_CODE (XEXP (operands
[1], 0)) == PRE_MODIFY
)
18260 if (reg_overlap_mentioned_p (otherops
[0], otherops
[2]))
18262 /* Registers overlap so split out the increment. */
18265 output_asm_insn ("add%?\t%1, %1, %2", otherops
);
18266 output_asm_insn ("ldrd%?\t%0, [%1] @split", otherops
);
18273 /* Use a single insn if we can.
18274 FIXME: IWMMXT allows offsets larger than ldrd can
18275 handle, fix these up with a pair of ldr. */
18277 || !CONST_INT_P (otherops
[2])
18278 || (INTVAL (otherops
[2]) > -256
18279 && INTVAL (otherops
[2]) < 256))
18282 output_asm_insn ("ldrd%?\t%0, [%1, %2]!", otherops
);
18288 output_asm_insn ("ldr%?\t%0, [%1, %2]!", otherops
);
18289 output_asm_insn ("ldr%?\t%H0, [%1, #4]", otherops
);
18299 /* Use a single insn if we can.
18300 FIXME: IWMMXT allows offsets larger than ldrd can handle,
18301 fix these up with a pair of ldr. */
18303 || !CONST_INT_P (otherops
[2])
18304 || (INTVAL (otherops
[2]) > -256
18305 && INTVAL (otherops
[2]) < 256))
18308 output_asm_insn ("ldrd%?\t%0, [%1], %2", otherops
);
18314 output_asm_insn ("ldr%?\t%H0, [%1, #4]", otherops
);
18315 output_asm_insn ("ldr%?\t%0, [%1], %2", otherops
);
18325 /* We might be able to use ldrd %0, %1 here. However the range is
18326 different to ldr/adr, and it is broken on some ARMv7-M
18327 implementations. */
18328 /* Use the second register of the pair to avoid problematic
18330 otherops
[1] = operands
[1];
18332 output_asm_insn ("adr%?\t%0, %1", otherops
);
18333 operands
[1] = otherops
[0];
18337 output_asm_insn ("ldrd%?\t%0, [%1]", operands
);
18339 output_asm_insn ("ldmia%?\t%1, %M0", operands
);
18346 /* ??? This needs checking for thumb2. */
18348 if (arm_add_operand (XEXP (XEXP (operands
[1], 0), 1),
18349 GET_MODE (XEXP (XEXP (operands
[1], 0), 1))))
18351 otherops
[0] = operands
[0];
18352 otherops
[1] = XEXP (XEXP (operands
[1], 0), 0);
18353 otherops
[2] = XEXP (XEXP (operands
[1], 0), 1);
18355 if (GET_CODE (XEXP (operands
[1], 0)) == PLUS
)
18357 if (CONST_INT_P (otherops
[2]) && !TARGET_LDRD
)
18359 switch ((int) INTVAL (otherops
[2]))
18363 output_asm_insn ("ldmdb%?\t%1, %M0", otherops
);
18369 output_asm_insn ("ldmda%?\t%1, %M0", otherops
);
18375 output_asm_insn ("ldmib%?\t%1, %M0", otherops
);
18379 otherops
[0] = gen_rtx_REG(SImode
, REGNO(operands
[0]) + 1);
18380 operands
[1] = otherops
[0];
18382 && (REG_P (otherops
[2])
18384 || (CONST_INT_P (otherops
[2])
18385 && INTVAL (otherops
[2]) > -256
18386 && INTVAL (otherops
[2]) < 256)))
18388 if (reg_overlap_mentioned_p (operands
[0],
18391 /* Swap base and index registers over to
18392 avoid a conflict. */
18393 std::swap (otherops
[1], otherops
[2]);
18395 /* If both registers conflict, it will usually
18396 have been fixed by a splitter. */
18397 if (reg_overlap_mentioned_p (operands
[0], otherops
[2])
18398 || (fix_cm3_ldrd
&& reg0
== REGNO (otherops
[1])))
18402 output_asm_insn ("add%?\t%0, %1, %2", otherops
);
18403 output_asm_insn ("ldrd%?\t%0, [%1]", operands
);
18410 otherops
[0] = operands
[0];
18412 output_asm_insn ("ldrd%?\t%0, [%1, %2]", otherops
);
18417 if (CONST_INT_P (otherops
[2]))
18421 if (!(const_ok_for_arm (INTVAL (otherops
[2]))))
18422 output_asm_insn ("sub%?\t%0, %1, #%n2", otherops
);
18424 output_asm_insn ("add%?\t%0, %1, %2", otherops
);
18430 output_asm_insn ("add%?\t%0, %1, %2", otherops
);
18436 output_asm_insn ("sub%?\t%0, %1, %2", otherops
);
18443 return "ldrd%?\t%0, [%1]";
18445 return "ldmia%?\t%1, %M0";
18449 otherops
[1] = adjust_address (operands
[1], SImode
, 4);
18450 /* Take care of overlapping base/data reg. */
18451 if (reg_mentioned_p (operands
[0], operands
[1]))
18455 output_asm_insn ("ldr%?\t%0, %1", otherops
);
18456 output_asm_insn ("ldr%?\t%0, %1", operands
);
18466 output_asm_insn ("ldr%?\t%0, %1", operands
);
18467 output_asm_insn ("ldr%?\t%0, %1", otherops
);
18477 /* Constraints should ensure this. */
18478 gcc_assert (code0
== MEM
&& code1
== REG
);
18479 gcc_assert ((REGNO (operands
[1]) != IP_REGNUM
)
18480 || (TARGET_ARM
&& TARGET_LDRD
));
18482 /* For TARGET_ARM the first source register of an STRD
18483 must be even. This is usually the case for double-word
18484 values but user assembly constraints can force an odd
18485 starting register. */
18486 bool allow_strd
= TARGET_LDRD
18487 && !(TARGET_ARM
&& (REGNO (operands
[1]) & 1) == 1);
18488 switch (GET_CODE (XEXP (operands
[0], 0)))
18494 output_asm_insn ("strd%?\t%1, [%m0]", operands
);
18496 output_asm_insn ("stm%?\t%m0, %M1", operands
);
18501 gcc_assert (allow_strd
);
18503 output_asm_insn ("strd%?\t%1, [%m0, #8]!", operands
);
18510 output_asm_insn ("strd%?\t%1, [%m0, #-8]!", operands
);
18512 output_asm_insn ("stmdb%?\t%m0!, %M1", operands
);
18520 output_asm_insn ("strd%?\t%1, [%m0], #8", operands
);
18522 output_asm_insn ("stm%?\t%m0!, %M1", operands
);
18527 gcc_assert (allow_strd
);
18529 output_asm_insn ("strd%?\t%1, [%m0], #-8", operands
);
18534 otherops
[0] = operands
[1];
18535 otherops
[1] = XEXP (XEXP (XEXP (operands
[0], 0), 1), 0);
18536 otherops
[2] = XEXP (XEXP (XEXP (operands
[0], 0), 1), 1);
18538 /* IWMMXT allows offsets larger than strd can handle,
18539 fix these up with a pair of str. */
18541 && CONST_INT_P (otherops
[2])
18542 && (INTVAL(otherops
[2]) <= -256
18543 || INTVAL(otherops
[2]) >= 256))
18545 if (GET_CODE (XEXP (operands
[0], 0)) == PRE_MODIFY
)
18549 output_asm_insn ("str%?\t%0, [%1, %2]!", otherops
);
18550 output_asm_insn ("str%?\t%H0, [%1, #4]", otherops
);
18559 output_asm_insn ("str%?\t%H0, [%1, #4]", otherops
);
18560 output_asm_insn ("str%?\t%0, [%1], %2", otherops
);
18566 else if (GET_CODE (XEXP (operands
[0], 0)) == PRE_MODIFY
)
18569 output_asm_insn ("strd%?\t%0, [%1, %2]!", otherops
);
18574 output_asm_insn ("strd%?\t%0, [%1], %2", otherops
);
18579 otherops
[2] = XEXP (XEXP (operands
[0], 0), 1);
18580 if (CONST_INT_P (otherops
[2]) && !TARGET_LDRD
)
18582 switch ((int) INTVAL (XEXP (XEXP (operands
[0], 0), 1)))
18586 output_asm_insn ("stmdb%?\t%m0, %M1", operands
);
18593 output_asm_insn ("stmda%?\t%m0, %M1", operands
);
18600 output_asm_insn ("stmib%?\t%m0, %M1", operands
);
18605 && (REG_P (otherops
[2])
18607 || (CONST_INT_P (otherops
[2])
18608 && INTVAL (otherops
[2]) > -256
18609 && INTVAL (otherops
[2]) < 256)))
18611 otherops
[0] = operands
[1];
18612 otherops
[1] = XEXP (XEXP (operands
[0], 0), 0);
18614 output_asm_insn ("strd%?\t%0, [%1, %2]", otherops
);
18620 otherops
[0] = adjust_address (operands
[0], SImode
, 4);
18621 otherops
[1] = operands
[1];
18624 output_asm_insn ("str%?\t%1, %0", operands
);
18625 output_asm_insn ("str%?\t%H1, %0", otherops
);
18635 /* Output a move, load or store for quad-word vectors in ARM registers. Only
18636 handles MEMs accepted by neon_vector_mem_operand with TYPE=1. */
18639 output_move_quad (rtx
*operands
)
18641 if (REG_P (operands
[0]))
18643 /* Load, or reg->reg move. */
18645 if (MEM_P (operands
[1]))
18647 switch (GET_CODE (XEXP (operands
[1], 0)))
18650 output_asm_insn ("ldmia%?\t%m1, %M0", operands
);
18655 output_asm_insn ("adr%?\t%0, %1", operands
);
18656 output_asm_insn ("ldmia%?\t%0, %M0", operands
);
18660 gcc_unreachable ();
18668 gcc_assert (REG_P (operands
[1]));
18670 dest
= REGNO (operands
[0]);
18671 src
= REGNO (operands
[1]);
18673 /* This seems pretty dumb, but hopefully GCC won't try to do it
18676 for (i
= 0; i
< 4; i
++)
18678 ops
[0] = gen_rtx_REG (SImode
, dest
+ i
);
18679 ops
[1] = gen_rtx_REG (SImode
, src
+ i
);
18680 output_asm_insn ("mov%?\t%0, %1", ops
);
18683 for (i
= 3; i
>= 0; i
--)
18685 ops
[0] = gen_rtx_REG (SImode
, dest
+ i
);
18686 ops
[1] = gen_rtx_REG (SImode
, src
+ i
);
18687 output_asm_insn ("mov%?\t%0, %1", ops
);
18693 gcc_assert (MEM_P (operands
[0]));
18694 gcc_assert (REG_P (operands
[1]));
18695 gcc_assert (!reg_overlap_mentioned_p (operands
[1], operands
[0]));
18697 switch (GET_CODE (XEXP (operands
[0], 0)))
18700 output_asm_insn ("stm%?\t%m0, %M1", operands
);
18704 gcc_unreachable ();
18711 /* Output a VFP load or store instruction. */
18714 output_move_vfp (rtx
*operands
)
18716 rtx reg
, mem
, addr
, ops
[2];
18717 int load
= REG_P (operands
[0]);
18718 int dp
= GET_MODE_SIZE (GET_MODE (operands
[0])) == 8;
18719 int sp
= (!TARGET_VFP_FP16INST
18720 || GET_MODE_SIZE (GET_MODE (operands
[0])) == 4);
18721 int integer_p
= GET_MODE_CLASS (GET_MODE (operands
[0])) == MODE_INT
;
18726 reg
= operands
[!load
];
18727 mem
= operands
[load
];
18729 mode
= GET_MODE (reg
);
18731 gcc_assert (REG_P (reg
));
18732 gcc_assert (IS_VFP_REGNUM (REGNO (reg
)));
18733 gcc_assert ((mode
== HFmode
&& TARGET_HARD_FLOAT
)
18739 || (TARGET_NEON
&& VALID_NEON_DREG_MODE (mode
)));
18740 gcc_assert (MEM_P (mem
));
18742 addr
= XEXP (mem
, 0);
18744 switch (GET_CODE (addr
))
18747 templ
= "v%smdb%%?.%s\t%%0!, {%%%s1}%s";
18748 ops
[0] = XEXP (addr
, 0);
18753 templ
= "v%smia%%?.%s\t%%0!, {%%%s1}%s";
18754 ops
[0] = XEXP (addr
, 0);
18759 templ
= "v%sr%%?.%s\t%%%s0, %%1%s";
18765 sprintf (buff
, templ
,
18766 load
? "ld" : "st",
18767 dp
? "64" : sp
? "32" : "16",
18769 integer_p
? "\t%@ int" : "");
18770 output_asm_insn (buff
, ops
);
18775 /* Output a Neon double-word or quad-word load or store, or a load
18776 or store for larger structure modes.
18778 WARNING: The ordering of elements is weird in big-endian mode,
18779 because the EABI requires that vectors stored in memory appear
18780 as though they were stored by a VSTM, as required by the EABI.
18781 GCC RTL defines element ordering based on in-memory order.
18782 This can be different from the architectural ordering of elements
18783 within a NEON register. The intrinsics defined in arm_neon.h use the
18784 NEON register element ordering, not the GCC RTL element ordering.
18786 For example, the in-memory ordering of a big-endian a quadword
18787 vector with 16-bit elements when stored from register pair {d0,d1}
18788 will be (lowest address first, d0[N] is NEON register element N):
18790 [d0[3], d0[2], d0[1], d0[0], d1[7], d1[6], d1[5], d1[4]]
18792 When necessary, quadword registers (dN, dN+1) are moved to ARM
18793 registers from rN in the order:
18795 dN -> (rN+1, rN), dN+1 -> (rN+3, rN+2)
18797 So that STM/LDM can be used on vectors in ARM registers, and the
18798 same memory layout will result as if VSTM/VLDM were used.
18800 Instead of VSTM/VLDM we prefer to use VST1.64/VLD1.64 where
18801 possible, which allows use of appropriate alignment tags.
18802 Note that the choice of "64" is independent of the actual vector
18803 element size; this size simply ensures that the behavior is
18804 equivalent to VSTM/VLDM in both little-endian and big-endian mode.
18806 Due to limitations of those instructions, use of VST1.64/VLD1.64
18807 is not possible if:
18808 - the address contains PRE_DEC, or
18809 - the mode refers to more than 4 double-word registers
18811 In those cases, it would be possible to replace VSTM/VLDM by a
18812 sequence of instructions; this is not currently implemented since
18813 this is not certain to actually improve performance. */
18816 output_move_neon (rtx
*operands
)
18818 rtx reg
, mem
, addr
, ops
[2];
18819 int regno
, nregs
, load
= REG_P (operands
[0]);
18824 reg
= operands
[!load
];
18825 mem
= operands
[load
];
18827 mode
= GET_MODE (reg
);
18829 gcc_assert (REG_P (reg
));
18830 regno
= REGNO (reg
);
18831 nregs
= REG_NREGS (reg
) / 2;
18832 gcc_assert (VFP_REGNO_OK_FOR_DOUBLE (regno
)
18833 || NEON_REGNO_OK_FOR_QUAD (regno
));
18834 gcc_assert (VALID_NEON_DREG_MODE (mode
)
18835 || VALID_NEON_QREG_MODE (mode
)
18836 || VALID_NEON_STRUCT_MODE (mode
));
18837 gcc_assert (MEM_P (mem
));
18839 addr
= XEXP (mem
, 0);
18841 /* Strip off const from addresses like (const (plus (...))). */
18842 if (GET_CODE (addr
) == CONST
&& GET_CODE (XEXP (addr
, 0)) == PLUS
)
18843 addr
= XEXP (addr
, 0);
18845 switch (GET_CODE (addr
))
18848 /* We have to use vldm / vstm for too-large modes. */
18851 templ
= "v%smia%%?\t%%0!, %%h1";
18852 ops
[0] = XEXP (addr
, 0);
18856 templ
= "v%s1.64\t%%h1, %%A0";
18863 /* We have to use vldm / vstm in this case, since there is no
18864 pre-decrement form of the vld1 / vst1 instructions. */
18865 templ
= "v%smdb%%?\t%%0!, %%h1";
18866 ops
[0] = XEXP (addr
, 0);
18871 /* FIXME: Not currently enabled in neon_vector_mem_operand. */
18872 gcc_unreachable ();
18875 /* We have to use vldm / vstm for too-large modes. */
18879 templ
= "v%smia%%?\t%%m0, %%h1";
18881 templ
= "v%s1.64\t%%h1, %%A0";
18887 /* Fall through. */
18893 for (i
= 0; i
< nregs
; i
++)
18895 /* We're only using DImode here because it's a convenient size. */
18896 ops
[0] = gen_rtx_REG (DImode
, REGNO (reg
) + 2 * i
);
18897 ops
[1] = adjust_address (mem
, DImode
, 8 * i
);
18898 if (reg_overlap_mentioned_p (ops
[0], mem
))
18900 gcc_assert (overlap
== -1);
18905 sprintf (buff
, "v%sr%%?\t%%P0, %%1", load
? "ld" : "st");
18906 output_asm_insn (buff
, ops
);
18911 ops
[0] = gen_rtx_REG (DImode
, REGNO (reg
) + 2 * overlap
);
18912 ops
[1] = adjust_address (mem
, SImode
, 8 * overlap
);
18913 sprintf (buff
, "v%sr%%?\t%%P0, %%1", load
? "ld" : "st");
18914 output_asm_insn (buff
, ops
);
18921 gcc_unreachable ();
18924 sprintf (buff
, templ
, load
? "ld" : "st");
18925 output_asm_insn (buff
, ops
);
18930 /* Compute and return the length of neon_mov<mode>, where <mode> is
18931 one of VSTRUCT modes: EI, OI, CI or XI. */
18933 arm_attr_length_move_neon (rtx_insn
*insn
)
18935 rtx reg
, mem
, addr
;
18939 extract_insn_cached (insn
);
18941 if (REG_P (recog_data
.operand
[0]) && REG_P (recog_data
.operand
[1]))
18943 mode
= GET_MODE (recog_data
.operand
[0]);
18954 gcc_unreachable ();
18958 load
= REG_P (recog_data
.operand
[0]);
18959 reg
= recog_data
.operand
[!load
];
18960 mem
= recog_data
.operand
[load
];
18962 gcc_assert (MEM_P (mem
));
18964 addr
= XEXP (mem
, 0);
18966 /* Strip off const from addresses like (const (plus (...))). */
18967 if (GET_CODE (addr
) == CONST
&& GET_CODE (XEXP (addr
, 0)) == PLUS
)
18968 addr
= XEXP (addr
, 0);
18970 if (GET_CODE (addr
) == LABEL_REF
|| GET_CODE (addr
) == PLUS
)
18972 int insns
= REG_NREGS (reg
) / 2;
18979 /* Return nonzero if the offset in the address is an immediate. Otherwise,
18983 arm_address_offset_is_imm (rtx_insn
*insn
)
18987 extract_insn_cached (insn
);
18989 if (REG_P (recog_data
.operand
[0]))
18992 mem
= recog_data
.operand
[0];
18994 gcc_assert (MEM_P (mem
));
18996 addr
= XEXP (mem
, 0);
18999 || (GET_CODE (addr
) == PLUS
19000 && REG_P (XEXP (addr
, 0))
19001 && CONST_INT_P (XEXP (addr
, 1))))
19007 /* Output an ADD r, s, #n where n may be too big for one instruction.
19008 If adding zero to one register, output nothing. */
19010 output_add_immediate (rtx
*operands
)
19012 HOST_WIDE_INT n
= INTVAL (operands
[2]);
19014 if (n
!= 0 || REGNO (operands
[0]) != REGNO (operands
[1]))
19017 output_multi_immediate (operands
,
19018 "sub%?\t%0, %1, %2", "sub%?\t%0, %0, %2", 2,
19021 output_multi_immediate (operands
,
19022 "add%?\t%0, %1, %2", "add%?\t%0, %0, %2", 2,
19029 /* Output a multiple immediate operation.
19030 OPERANDS is the vector of operands referred to in the output patterns.
19031 INSTR1 is the output pattern to use for the first constant.
19032 INSTR2 is the output pattern to use for subsequent constants.
19033 IMMED_OP is the index of the constant slot in OPERANDS.
19034 N is the constant value. */
19035 static const char *
19036 output_multi_immediate (rtx
*operands
, const char *instr1
, const char *instr2
,
19037 int immed_op
, HOST_WIDE_INT n
)
19039 #if HOST_BITS_PER_WIDE_INT > 32
19045 /* Quick and easy output. */
19046 operands
[immed_op
] = const0_rtx
;
19047 output_asm_insn (instr1
, operands
);
19052 const char * instr
= instr1
;
19054 /* Note that n is never zero here (which would give no output). */
19055 for (i
= 0; i
< 32; i
+= 2)
19059 operands
[immed_op
] = GEN_INT (n
& (255 << i
));
19060 output_asm_insn (instr
, operands
);
19070 /* Return the name of a shifter operation. */
19071 static const char *
19072 arm_shift_nmem(enum rtx_code code
)
19077 return ARM_LSL_NAME
;
19093 /* Return the appropriate ARM instruction for the operation code.
19094 The returned result should not be overwritten. OP is the rtx of the
19095 operation. SHIFT_FIRST_ARG is TRUE if the first argument of the operator
19098 arithmetic_instr (rtx op
, int shift_first_arg
)
19100 switch (GET_CODE (op
))
19106 return shift_first_arg
? "rsb" : "sub";
19121 return arm_shift_nmem(GET_CODE(op
));
19124 gcc_unreachable ();
19128 /* Ensure valid constant shifts and return the appropriate shift mnemonic
19129 for the operation code. The returned result should not be overwritten.
19130 OP is the rtx code of the shift.
19131 On exit, *AMOUNTP will be -1 if the shift is by a register, or a constant
19133 static const char *
19134 shift_op (rtx op
, HOST_WIDE_INT
*amountp
)
19137 enum rtx_code code
= GET_CODE (op
);
19142 if (!CONST_INT_P (XEXP (op
, 1)))
19144 output_operand_lossage ("invalid shift operand");
19149 *amountp
= 32 - INTVAL (XEXP (op
, 1));
19157 mnem
= arm_shift_nmem(code
);
19158 if (CONST_INT_P (XEXP (op
, 1)))
19160 *amountp
= INTVAL (XEXP (op
, 1));
19162 else if (REG_P (XEXP (op
, 1)))
19169 output_operand_lossage ("invalid shift operand");
19175 /* We never have to worry about the amount being other than a
19176 power of 2, since this case can never be reloaded from a reg. */
19177 if (!CONST_INT_P (XEXP (op
, 1)))
19179 output_operand_lossage ("invalid shift operand");
19183 *amountp
= INTVAL (XEXP (op
, 1)) & 0xFFFFFFFF;
19185 /* Amount must be a power of two. */
19186 if (*amountp
& (*amountp
- 1))
19188 output_operand_lossage ("invalid shift operand");
19192 *amountp
= exact_log2 (*amountp
);
19193 gcc_assert (IN_RANGE (*amountp
, 0, 31));
19194 return ARM_LSL_NAME
;
19197 output_operand_lossage ("invalid shift operand");
19201 /* This is not 100% correct, but follows from the desire to merge
19202 multiplication by a power of 2 with the recognizer for a
19203 shift. >=32 is not a valid shift for "lsl", so we must try and
19204 output a shift that produces the correct arithmetical result.
19205 Using lsr #32 is identical except for the fact that the carry bit
19206 is not set correctly if we set the flags; but we never use the
19207 carry bit from such an operation, so we can ignore that. */
19208 if (code
== ROTATERT
)
19209 /* Rotate is just modulo 32. */
19211 else if (*amountp
!= (*amountp
& 31))
19213 if (code
== ASHIFT
)
19218 /* Shifts of 0 are no-ops. */
19225 /* Output a .ascii pseudo-op, keeping track of lengths. This is
19226 because /bin/as is horribly restrictive. The judgement about
19227 whether or not each character is 'printable' (and can be output as
19228 is) or not (and must be printed with an octal escape) must be made
19229 with reference to the *host* character set -- the situation is
19230 similar to that discussed in the comments above pp_c_char in
19231 c-pretty-print.c. */
19233 #define MAX_ASCII_LEN 51
19236 output_ascii_pseudo_op (FILE *stream
, const unsigned char *p
, int len
)
19239 int len_so_far
= 0;
19241 fputs ("\t.ascii\t\"", stream
);
19243 for (i
= 0; i
< len
; i
++)
19247 if (len_so_far
>= MAX_ASCII_LEN
)
19249 fputs ("\"\n\t.ascii\t\"", stream
);
19255 if (c
== '\\' || c
== '\"')
19257 putc ('\\', stream
);
19265 fprintf (stream
, "\\%03o", c
);
19270 fputs ("\"\n", stream
);
19273 /* Whether a register is callee saved or not. This is necessary because high
19274 registers are marked as caller saved when optimizing for size on Thumb-1
19275 targets despite being callee saved in order to avoid using them. */
19276 #define callee_saved_reg_p(reg) \
19277 (!call_used_regs[reg] \
19278 || (TARGET_THUMB1 && optimize_size \
19279 && reg >= FIRST_HI_REGNUM && reg <= LAST_HI_REGNUM))
19281 /* Compute the register save mask for registers 0 through 12
19282 inclusive. This code is used by arm_compute_save_core_reg_mask (). */
19284 static unsigned long
19285 arm_compute_save_reg0_reg12_mask (void)
19287 unsigned long func_type
= arm_current_func_type ();
19288 unsigned long save_reg_mask
= 0;
19291 if (IS_INTERRUPT (func_type
))
19293 unsigned int max_reg
;
19294 /* Interrupt functions must not corrupt any registers,
19295 even call clobbered ones. If this is a leaf function
19296 we can just examine the registers used by the RTL, but
19297 otherwise we have to assume that whatever function is
19298 called might clobber anything, and so we have to save
19299 all the call-clobbered registers as well. */
19300 if (ARM_FUNC_TYPE (func_type
) == ARM_FT_FIQ
)
19301 /* FIQ handlers have registers r8 - r12 banked, so
19302 we only need to check r0 - r7, Normal ISRs only
19303 bank r14 and r15, so we must check up to r12.
19304 r13 is the stack pointer which is always preserved,
19305 so we do not need to consider it here. */
19310 for (reg
= 0; reg
<= max_reg
; reg
++)
19311 if (df_regs_ever_live_p (reg
)
19312 || (! crtl
->is_leaf
&& call_used_regs
[reg
]))
19313 save_reg_mask
|= (1 << reg
);
19315 /* Also save the pic base register if necessary. */
19317 && !TARGET_SINGLE_PIC_BASE
19318 && arm_pic_register
!= INVALID_REGNUM
19319 && crtl
->uses_pic_offset_table
)
19320 save_reg_mask
|= 1 << PIC_OFFSET_TABLE_REGNUM
;
19322 else if (IS_VOLATILE(func_type
))
19324 /* For noreturn functions we historically omitted register saves
19325 altogether. However this really messes up debugging. As a
19326 compromise save just the frame pointers. Combined with the link
19327 register saved elsewhere this should be sufficient to get
19329 if (frame_pointer_needed
)
19330 save_reg_mask
|= 1 << HARD_FRAME_POINTER_REGNUM
;
19331 if (df_regs_ever_live_p (ARM_HARD_FRAME_POINTER_REGNUM
))
19332 save_reg_mask
|= 1 << ARM_HARD_FRAME_POINTER_REGNUM
;
19333 if (df_regs_ever_live_p (THUMB_HARD_FRAME_POINTER_REGNUM
))
19334 save_reg_mask
|= 1 << THUMB_HARD_FRAME_POINTER_REGNUM
;
19338 /* In the normal case we only need to save those registers
19339 which are call saved and which are used by this function. */
19340 for (reg
= 0; reg
<= 11; reg
++)
19341 if (df_regs_ever_live_p (reg
) && callee_saved_reg_p (reg
))
19342 save_reg_mask
|= (1 << reg
);
19344 /* Handle the frame pointer as a special case. */
19345 if (frame_pointer_needed
)
19346 save_reg_mask
|= 1 << HARD_FRAME_POINTER_REGNUM
;
19348 /* If we aren't loading the PIC register,
19349 don't stack it even though it may be live. */
19351 && !TARGET_SINGLE_PIC_BASE
19352 && arm_pic_register
!= INVALID_REGNUM
19353 && (df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM
)
19354 || crtl
->uses_pic_offset_table
))
19355 save_reg_mask
|= 1 << PIC_OFFSET_TABLE_REGNUM
;
19357 /* The prologue will copy SP into R0, so save it. */
19358 if (IS_STACKALIGN (func_type
))
19359 save_reg_mask
|= 1;
19362 /* Save registers so the exception handler can modify them. */
19363 if (crtl
->calls_eh_return
)
19369 reg
= EH_RETURN_DATA_REGNO (i
);
19370 if (reg
== INVALID_REGNUM
)
19372 save_reg_mask
|= 1 << reg
;
19376 return save_reg_mask
;
19379 /* Return true if r3 is live at the start of the function. */
19382 arm_r3_live_at_start_p (void)
19384 /* Just look at cfg info, which is still close enough to correct at this
19385 point. This gives false positives for broken functions that might use
19386 uninitialized data that happens to be allocated in r3, but who cares? */
19387 return REGNO_REG_SET_P (df_get_live_out (ENTRY_BLOCK_PTR_FOR_FN (cfun
)), 3);
19390 /* Compute the number of bytes used to store the static chain register on the
19391 stack, above the stack frame. We need to know this accurately to get the
19392 alignment of the rest of the stack frame correct. */
19395 arm_compute_static_chain_stack_bytes (void)
19397 /* Once the value is updated from the init value of -1, do not
19399 if (cfun
->machine
->static_chain_stack_bytes
!= -1)
19400 return cfun
->machine
->static_chain_stack_bytes
;
19402 /* See the defining assertion in arm_expand_prologue. */
19403 if (IS_NESTED (arm_current_func_type ())
19404 && ((TARGET_APCS_FRAME
&& frame_pointer_needed
&& TARGET_ARM
)
19405 || ((flag_stack_check
== STATIC_BUILTIN_STACK_CHECK
19406 || flag_stack_clash_protection
)
19407 && !df_regs_ever_live_p (LR_REGNUM
)))
19408 && arm_r3_live_at_start_p ()
19409 && crtl
->args
.pretend_args_size
== 0)
19415 /* Compute a bit mask of which core registers need to be
19416 saved on the stack for the current function.
19417 This is used by arm_compute_frame_layout, which may add extra registers. */
19419 static unsigned long
19420 arm_compute_save_core_reg_mask (void)
19422 unsigned int save_reg_mask
= 0;
19423 unsigned long func_type
= arm_current_func_type ();
19426 if (IS_NAKED (func_type
))
19427 /* This should never really happen. */
19430 /* If we are creating a stack frame, then we must save the frame pointer,
19431 IP (which will hold the old stack pointer), LR and the PC. */
19432 if (TARGET_APCS_FRAME
&& frame_pointer_needed
&& TARGET_ARM
)
19434 (1 << ARM_HARD_FRAME_POINTER_REGNUM
)
19437 | (1 << PC_REGNUM
);
19439 save_reg_mask
|= arm_compute_save_reg0_reg12_mask ();
19441 /* Decide if we need to save the link register.
19442 Interrupt routines have their own banked link register,
19443 so they never need to save it.
19444 Otherwise if we do not use the link register we do not need to save
19445 it. If we are pushing other registers onto the stack however, we
19446 can save an instruction in the epilogue by pushing the link register
19447 now and then popping it back into the PC. This incurs extra memory
19448 accesses though, so we only do it when optimizing for size, and only
19449 if we know that we will not need a fancy return sequence. */
19450 if (df_regs_ever_live_p (LR_REGNUM
)
19453 && ARM_FUNC_TYPE (func_type
) == ARM_FT_NORMAL
19454 && !crtl
->tail_call_emit
19455 && !crtl
->calls_eh_return
))
19456 save_reg_mask
|= 1 << LR_REGNUM
;
19458 if (cfun
->machine
->lr_save_eliminated
)
19459 save_reg_mask
&= ~ (1 << LR_REGNUM
);
19461 if (TARGET_REALLY_IWMMXT
19462 && ((bit_count (save_reg_mask
)
19463 + ARM_NUM_INTS (crtl
->args
.pretend_args_size
+
19464 arm_compute_static_chain_stack_bytes())
19467 /* The total number of registers that are going to be pushed
19468 onto the stack is odd. We need to ensure that the stack
19469 is 64-bit aligned before we start to save iWMMXt registers,
19470 and also before we start to create locals. (A local variable
19471 might be a double or long long which we will load/store using
19472 an iWMMXt instruction). Therefore we need to push another
19473 ARM register, so that the stack will be 64-bit aligned. We
19474 try to avoid using the arg registers (r0 -r3) as they might be
19475 used to pass values in a tail call. */
19476 for (reg
= 4; reg
<= 12; reg
++)
19477 if ((save_reg_mask
& (1 << reg
)) == 0)
19481 save_reg_mask
|= (1 << reg
);
19484 cfun
->machine
->sibcall_blocked
= 1;
19485 save_reg_mask
|= (1 << 3);
19489 /* We may need to push an additional register for use initializing the
19490 PIC base register. */
19491 if (TARGET_THUMB2
&& IS_NESTED (func_type
) && flag_pic
19492 && (save_reg_mask
& THUMB2_WORK_REGS
) == 0)
19494 reg
= thumb_find_work_register (1 << 4);
19495 if (!call_used_regs
[reg
])
19496 save_reg_mask
|= (1 << reg
);
19499 return save_reg_mask
;
19502 /* Compute a bit mask of which core registers need to be
19503 saved on the stack for the current function. */
19504 static unsigned long
19505 thumb1_compute_save_core_reg_mask (void)
19507 unsigned long mask
;
19511 for (reg
= 0; reg
< 12; reg
++)
19512 if (df_regs_ever_live_p (reg
) && callee_saved_reg_p (reg
))
19515 /* Handle the frame pointer as a special case. */
19516 if (frame_pointer_needed
)
19517 mask
|= 1 << HARD_FRAME_POINTER_REGNUM
;
19520 && !TARGET_SINGLE_PIC_BASE
19521 && arm_pic_register
!= INVALID_REGNUM
19522 && crtl
->uses_pic_offset_table
)
19523 mask
|= 1 << PIC_OFFSET_TABLE_REGNUM
;
19525 /* See if we might need r11 for calls to _interwork_r11_call_via_rN(). */
19526 if (!frame_pointer_needed
&& CALLER_INTERWORKING_SLOT_SIZE
> 0)
19527 mask
|= 1 << ARM_HARD_FRAME_POINTER_REGNUM
;
19529 /* LR will also be pushed if any lo regs are pushed. */
19530 if (mask
& 0xff || thumb_force_lr_save ())
19531 mask
|= (1 << LR_REGNUM
);
19533 /* Make sure we have a low work register if we need one.
19534 We will need one if we are going to push a high register,
19535 but we are not currently intending to push a low register. */
19536 if ((mask
& 0xff) == 0
19537 && ((mask
& 0x0f00) || TARGET_BACKTRACE
))
19539 /* Use thumb_find_work_register to choose which register
19540 we will use. If the register is live then we will
19541 have to push it. Use LAST_LO_REGNUM as our fallback
19542 choice for the register to select. */
19543 reg
= thumb_find_work_register (1 << LAST_LO_REGNUM
);
19544 /* Make sure the register returned by thumb_find_work_register is
19545 not part of the return value. */
19546 if (reg
* UNITS_PER_WORD
<= (unsigned) arm_size_return_regs ())
19547 reg
= LAST_LO_REGNUM
;
19549 if (callee_saved_reg_p (reg
))
19553 /* The 504 below is 8 bytes less than 512 because there are two possible
19554 alignment words. We can't tell here if they will be present or not so we
19555 have to play it safe and assume that they are. */
19556 if ((CALLER_INTERWORKING_SLOT_SIZE
+
19557 ROUND_UP_WORD (get_frame_size ()) +
19558 crtl
->outgoing_args_size
) >= 504)
19560 /* This is the same as the code in thumb1_expand_prologue() which
19561 determines which register to use for stack decrement. */
19562 for (reg
= LAST_ARG_REGNUM
+ 1; reg
<= LAST_LO_REGNUM
; reg
++)
19563 if (mask
& (1 << reg
))
19566 if (reg
> LAST_LO_REGNUM
)
19568 /* Make sure we have a register available for stack decrement. */
19569 mask
|= 1 << LAST_LO_REGNUM
;
19577 /* Return the number of bytes required to save VFP registers. */
19579 arm_get_vfp_saved_size (void)
19581 unsigned int regno
;
19586 /* Space for saved VFP registers. */
19587 if (TARGET_HARD_FLOAT
)
19590 for (regno
= FIRST_VFP_REGNUM
;
19591 regno
< LAST_VFP_REGNUM
;
19594 if ((!df_regs_ever_live_p (regno
) || call_used_regs
[regno
])
19595 && (!df_regs_ever_live_p (regno
+ 1) || call_used_regs
[regno
+ 1]))
19599 /* Workaround ARM10 VFPr1 bug. */
19600 if (count
== 2 && !arm_arch6
)
19602 saved
+= count
* 8;
19611 if (count
== 2 && !arm_arch6
)
19613 saved
+= count
* 8;
19620 /* Generate a function exit sequence. If REALLY_RETURN is false, then do
19621 everything bar the final return instruction. If simple_return is true,
19622 then do not output epilogue, because it has already been emitted in RTL.
19624 Note: do not forget to update length attribute of corresponding insn pattern
19625 when changing assembly output (eg. length attribute of
19626 thumb2_cmse_entry_return when updating Armv8-M Mainline Security Extensions
19627 register clearing sequences). */
19629 output_return_instruction (rtx operand
, bool really_return
, bool reverse
,
19630 bool simple_return
)
19632 char conditional
[10];
19635 unsigned long live_regs_mask
;
19636 unsigned long func_type
;
19637 arm_stack_offsets
*offsets
;
19639 func_type
= arm_current_func_type ();
19641 if (IS_NAKED (func_type
))
19644 if (IS_VOLATILE (func_type
) && TARGET_ABORT_NORETURN
)
19646 /* If this function was declared non-returning, and we have
19647 found a tail call, then we have to trust that the called
19648 function won't return. */
19653 /* Otherwise, trap an attempted return by aborting. */
19655 ops
[1] = gen_rtx_SYMBOL_REF (Pmode
, NEED_PLT_RELOC
? "abort(PLT)"
19657 assemble_external_libcall (ops
[1]);
19658 output_asm_insn (reverse
? "bl%D0\t%a1" : "bl%d0\t%a1", ops
);
19664 gcc_assert (!cfun
->calls_alloca
|| really_return
);
19666 sprintf (conditional
, "%%?%%%c0", reverse
? 'D' : 'd');
19668 cfun
->machine
->return_used_this_function
= 1;
19670 offsets
= arm_get_frame_offsets ();
19671 live_regs_mask
= offsets
->saved_regs_mask
;
19673 if (!simple_return
&& live_regs_mask
)
19675 const char * return_reg
;
19677 /* If we do not have any special requirements for function exit
19678 (e.g. interworking) then we can load the return address
19679 directly into the PC. Otherwise we must load it into LR. */
19681 && !IS_CMSE_ENTRY (func_type
)
19682 && (IS_INTERRUPT (func_type
) || !TARGET_INTERWORK
))
19683 return_reg
= reg_names
[PC_REGNUM
];
19685 return_reg
= reg_names
[LR_REGNUM
];
19687 if ((live_regs_mask
& (1 << IP_REGNUM
)) == (1 << IP_REGNUM
))
19689 /* There are three possible reasons for the IP register
19690 being saved. 1) a stack frame was created, in which case
19691 IP contains the old stack pointer, or 2) an ISR routine
19692 corrupted it, or 3) it was saved to align the stack on
19693 iWMMXt. In case 1, restore IP into SP, otherwise just
19695 if (frame_pointer_needed
)
19697 live_regs_mask
&= ~ (1 << IP_REGNUM
);
19698 live_regs_mask
|= (1 << SP_REGNUM
);
19701 gcc_assert (IS_INTERRUPT (func_type
) || TARGET_REALLY_IWMMXT
);
19704 /* On some ARM architectures it is faster to use LDR rather than
19705 LDM to load a single register. On other architectures, the
19706 cost is the same. In 26 bit mode, or for exception handlers,
19707 we have to use LDM to load the PC so that the CPSR is also
19709 for (reg
= 0; reg
<= LAST_ARM_REGNUM
; reg
++)
19710 if (live_regs_mask
== (1U << reg
))
19713 if (reg
<= LAST_ARM_REGNUM
19714 && (reg
!= LR_REGNUM
19716 || ! IS_INTERRUPT (func_type
)))
19718 sprintf (instr
, "ldr%s\t%%|%s, [%%|sp], #4", conditional
,
19719 (reg
== LR_REGNUM
) ? return_reg
: reg_names
[reg
]);
19726 /* Generate the load multiple instruction to restore the
19727 registers. Note we can get here, even if
19728 frame_pointer_needed is true, but only if sp already
19729 points to the base of the saved core registers. */
19730 if (live_regs_mask
& (1 << SP_REGNUM
))
19732 unsigned HOST_WIDE_INT stack_adjust
;
19734 stack_adjust
= offsets
->outgoing_args
- offsets
->saved_regs
;
19735 gcc_assert (stack_adjust
== 0 || stack_adjust
== 4);
19737 if (stack_adjust
&& arm_arch5t
&& TARGET_ARM
)
19738 sprintf (instr
, "ldmib%s\t%%|sp, {", conditional
);
19741 /* If we can't use ldmib (SA110 bug),
19742 then try to pop r3 instead. */
19744 live_regs_mask
|= 1 << 3;
19746 sprintf (instr
, "ldmfd%s\t%%|sp, {", conditional
);
19749 /* For interrupt returns we have to use an LDM rather than
19750 a POP so that we can use the exception return variant. */
19751 else if (IS_INTERRUPT (func_type
))
19752 sprintf (instr
, "ldmfd%s\t%%|sp!, {", conditional
);
19754 sprintf (instr
, "pop%s\t{", conditional
);
19756 p
= instr
+ strlen (instr
);
19758 for (reg
= 0; reg
<= SP_REGNUM
; reg
++)
19759 if (live_regs_mask
& (1 << reg
))
19761 int l
= strlen (reg_names
[reg
]);
19767 memcpy (p
, ", ", 2);
19771 memcpy (p
, "%|", 2);
19772 memcpy (p
+ 2, reg_names
[reg
], l
);
19776 if (live_regs_mask
& (1 << LR_REGNUM
))
19778 sprintf (p
, "%s%%|%s}", first
? "" : ", ", return_reg
);
19779 /* If returning from an interrupt, restore the CPSR. */
19780 if (IS_INTERRUPT (func_type
))
19787 output_asm_insn (instr
, & operand
);
19789 /* See if we need to generate an extra instruction to
19790 perform the actual function return. */
19792 && func_type
!= ARM_FT_INTERWORKED
19793 && (live_regs_mask
& (1 << LR_REGNUM
)) != 0)
19795 /* The return has already been handled
19796 by loading the LR into the PC. */
19803 switch ((int) ARM_FUNC_TYPE (func_type
))
19807 /* ??? This is wrong for unified assembly syntax. */
19808 sprintf (instr
, "sub%ss\t%%|pc, %%|lr, #4", conditional
);
19811 case ARM_FT_INTERWORKED
:
19812 gcc_assert (arm_arch5t
|| arm_arch4t
);
19813 sprintf (instr
, "bx%s\t%%|lr", conditional
);
19816 case ARM_FT_EXCEPTION
:
19817 /* ??? This is wrong for unified assembly syntax. */
19818 sprintf (instr
, "mov%ss\t%%|pc, %%|lr", conditional
);
19822 if (IS_CMSE_ENTRY (func_type
))
19824 /* Check if we have to clear the 'GE bits' which is only used if
19825 parallel add and subtraction instructions are available. */
19826 if (TARGET_INT_SIMD
)
19827 snprintf (instr
, sizeof (instr
),
19828 "msr%s\tAPSR_nzcvqg, %%|lr", conditional
);
19830 snprintf (instr
, sizeof (instr
),
19831 "msr%s\tAPSR_nzcvq, %%|lr", conditional
);
19833 output_asm_insn (instr
, & operand
);
19834 if (TARGET_HARD_FLOAT
&& !TARGET_THUMB1
)
19836 /* Clear the cumulative exception-status bits (0-4,7) and the
19837 condition code bits (28-31) of the FPSCR. We need to
19838 remember to clear the first scratch register used (IP) and
19839 save and restore the second (r4). */
19840 snprintf (instr
, sizeof (instr
), "push\t{%%|r4}");
19841 output_asm_insn (instr
, & operand
);
19842 snprintf (instr
, sizeof (instr
), "vmrs\t%%|ip, fpscr");
19843 output_asm_insn (instr
, & operand
);
19844 snprintf (instr
, sizeof (instr
), "movw\t%%|r4, #65376");
19845 output_asm_insn (instr
, & operand
);
19846 snprintf (instr
, sizeof (instr
), "movt\t%%|r4, #4095");
19847 output_asm_insn (instr
, & operand
);
19848 snprintf (instr
, sizeof (instr
), "and\t%%|ip, %%|r4");
19849 output_asm_insn (instr
, & operand
);
19850 snprintf (instr
, sizeof (instr
), "vmsr\tfpscr, %%|ip");
19851 output_asm_insn (instr
, & operand
);
19852 snprintf (instr
, sizeof (instr
), "pop\t{%%|r4}");
19853 output_asm_insn (instr
, & operand
);
19854 snprintf (instr
, sizeof (instr
), "mov\t%%|ip, %%|lr");
19855 output_asm_insn (instr
, & operand
);
19857 snprintf (instr
, sizeof (instr
), "bxns\t%%|lr");
19859 /* Use bx if it's available. */
19860 else if (arm_arch5t
|| arm_arch4t
)
19861 sprintf (instr
, "bx%s\t%%|lr", conditional
);
19863 sprintf (instr
, "mov%s\t%%|pc, %%|lr", conditional
);
19867 output_asm_insn (instr
, & operand
);
19873 /* Output in FILE asm statements needed to declare the NAME of the function
19874 defined by its DECL node. */
19877 arm_asm_declare_function_name (FILE *file
, const char *name
, tree decl
)
19879 size_t cmse_name_len
;
19880 char *cmse_name
= 0;
19881 char cmse_prefix
[] = "__acle_se_";
19883 /* When compiling with ARMv8-M Security Extensions enabled, we should print an
19884 extra function label for each function with the 'cmse_nonsecure_entry'
19885 attribute. This extra function label should be prepended with
19886 '__acle_se_', telling the linker that it needs to create secure gateway
19887 veneers for this function. */
19888 if (use_cmse
&& lookup_attribute ("cmse_nonsecure_entry",
19889 DECL_ATTRIBUTES (decl
)))
19891 cmse_name_len
= sizeof (cmse_prefix
) + strlen (name
);
19892 cmse_name
= XALLOCAVEC (char, cmse_name_len
);
19893 snprintf (cmse_name
, cmse_name_len
, "%s%s", cmse_prefix
, name
);
19894 targetm
.asm_out
.globalize_label (file
, cmse_name
);
19896 ARM_DECLARE_FUNCTION_NAME (file
, cmse_name
, decl
);
19897 ASM_OUTPUT_TYPE_DIRECTIVE (file
, cmse_name
, "function");
19900 ARM_DECLARE_FUNCTION_NAME (file
, name
, decl
);
19901 ASM_OUTPUT_TYPE_DIRECTIVE (file
, name
, "function");
19902 ASM_DECLARE_RESULT (file
, DECL_RESULT (decl
));
19903 ASM_OUTPUT_LABEL (file
, name
);
19906 ASM_OUTPUT_LABEL (file
, cmse_name
);
19908 ARM_OUTPUT_FN_UNWIND (file
, TRUE
);
19911 /* Write the function name into the code section, directly preceding
19912 the function prologue.
19914 Code will be output similar to this:
19916 .ascii "arm_poke_function_name", 0
19919 .word 0xff000000 + (t1 - t0)
19920 arm_poke_function_name
19922 stmfd sp!, {fp, ip, lr, pc}
19925 When performing a stack backtrace, code can inspect the value
19926 of 'pc' stored at 'fp' + 0. If the trace function then looks
19927 at location pc - 12 and the top 8 bits are set, then we know
19928 that there is a function name embedded immediately preceding this
19929 location and has length ((pc[-3]) & 0xff000000).
19931 We assume that pc is declared as a pointer to an unsigned long.
19933 It is of no benefit to output the function name if we are assembling
19934 a leaf function. These function types will not contain a stack
19935 backtrace structure, therefore it is not possible to determine the
19938 arm_poke_function_name (FILE *stream
, const char *name
)
19940 unsigned long alignlength
;
19941 unsigned long length
;
19944 length
= strlen (name
) + 1;
19945 alignlength
= ROUND_UP_WORD (length
);
19947 ASM_OUTPUT_ASCII (stream
, name
, length
);
19948 ASM_OUTPUT_ALIGN (stream
, 2);
19949 x
= GEN_INT ((unsigned HOST_WIDE_INT
) 0xff000000 + alignlength
);
19950 assemble_aligned_integer (UNITS_PER_WORD
, x
);
19953 /* Place some comments into the assembler stream
19954 describing the current function. */
19956 arm_output_function_prologue (FILE *f
)
19958 unsigned long func_type
;
19960 /* Sanity check. */
19961 gcc_assert (!arm_ccfsm_state
&& !arm_target_insn
);
19963 func_type
= arm_current_func_type ();
19965 switch ((int) ARM_FUNC_TYPE (func_type
))
19968 case ARM_FT_NORMAL
:
19970 case ARM_FT_INTERWORKED
:
19971 asm_fprintf (f
, "\t%@ Function supports interworking.\n");
19974 asm_fprintf (f
, "\t%@ Interrupt Service Routine.\n");
19977 asm_fprintf (f
, "\t%@ Fast Interrupt Service Routine.\n");
19979 case ARM_FT_EXCEPTION
:
19980 asm_fprintf (f
, "\t%@ ARM Exception Handler.\n");
19984 if (IS_NAKED (func_type
))
19985 asm_fprintf (f
, "\t%@ Naked Function: prologue and epilogue provided by programmer.\n");
19987 if (IS_VOLATILE (func_type
))
19988 asm_fprintf (f
, "\t%@ Volatile: function does not return.\n");
19990 if (IS_NESTED (func_type
))
19991 asm_fprintf (f
, "\t%@ Nested: function declared inside another function.\n");
19992 if (IS_STACKALIGN (func_type
))
19993 asm_fprintf (f
, "\t%@ Stack Align: May be called with mis-aligned SP.\n");
19994 if (IS_CMSE_ENTRY (func_type
))
19995 asm_fprintf (f
, "\t%@ Non-secure entry function: called from non-secure code.\n");
19997 asm_fprintf (f
, "\t%@ args = %wd, pretend = %d, frame = %wd\n",
19998 (HOST_WIDE_INT
) crtl
->args
.size
,
19999 crtl
->args
.pretend_args_size
,
20000 (HOST_WIDE_INT
) get_frame_size ());
20002 asm_fprintf (f
, "\t%@ frame_needed = %d, uses_anonymous_args = %d\n",
20003 frame_pointer_needed
,
20004 cfun
->machine
->uses_anonymous_args
);
20006 if (cfun
->machine
->lr_save_eliminated
)
20007 asm_fprintf (f
, "\t%@ link register save eliminated.\n");
20009 if (crtl
->calls_eh_return
)
20010 asm_fprintf (f
, "\t@ Calls __builtin_eh_return.\n");
20015 arm_output_function_epilogue (FILE *)
20017 arm_stack_offsets
*offsets
;
20023 /* Emit any call-via-reg trampolines that are needed for v4t support
20024 of call_reg and call_value_reg type insns. */
20025 for (regno
= 0; regno
< LR_REGNUM
; regno
++)
20027 rtx label
= cfun
->machine
->call_via
[regno
];
20031 switch_to_section (function_section (current_function_decl
));
20032 targetm
.asm_out
.internal_label (asm_out_file
, "L",
20033 CODE_LABEL_NUMBER (label
));
20034 asm_fprintf (asm_out_file
, "\tbx\t%r\n", regno
);
20038 /* ??? Probably not safe to set this here, since it assumes that a
20039 function will be emitted as assembly immediately after we generate
20040 RTL for it. This does not happen for inline functions. */
20041 cfun
->machine
->return_used_this_function
= 0;
20043 else /* TARGET_32BIT */
20045 /* We need to take into account any stack-frame rounding. */
20046 offsets
= arm_get_frame_offsets ();
20048 gcc_assert (!use_return_insn (FALSE
, NULL
)
20049 || (cfun
->machine
->return_used_this_function
!= 0)
20050 || offsets
->saved_regs
== offsets
->outgoing_args
20051 || frame_pointer_needed
);
20055 /* Generate and emit a sequence of insns equivalent to PUSH, but using
20056 STR and STRD. If an even number of registers are being pushed, one
20057 or more STRD patterns are created for each register pair. If an
20058 odd number of registers are pushed, emit an initial STR followed by
20059 as many STRD instructions as are needed. This works best when the
20060 stack is initially 64-bit aligned (the normal case), since it
20061 ensures that each STRD is also 64-bit aligned. */
20063 thumb2_emit_strd_push (unsigned long saved_regs_mask
)
20068 rtx par
= NULL_RTX
;
20069 rtx dwarf
= NULL_RTX
;
20073 num_regs
= bit_count (saved_regs_mask
);
20075 /* Must be at least one register to save, and can't save SP or PC. */
20076 gcc_assert (num_regs
> 0 && num_regs
<= 14);
20077 gcc_assert (!(saved_regs_mask
& (1 << SP_REGNUM
)));
20078 gcc_assert (!(saved_regs_mask
& (1 << PC_REGNUM
)));
20080 /* Create sequence for DWARF info. All the frame-related data for
20081 debugging is held in this wrapper. */
20082 dwarf
= gen_rtx_SEQUENCE (VOIDmode
, rtvec_alloc (num_regs
+ 1));
20084 /* Describe the stack adjustment. */
20085 tmp
= gen_rtx_SET (stack_pointer_rtx
,
20086 plus_constant (Pmode
, stack_pointer_rtx
, -4 * num_regs
));
20087 RTX_FRAME_RELATED_P (tmp
) = 1;
20088 XVECEXP (dwarf
, 0, 0) = tmp
;
20090 /* Find the first register. */
20091 for (regno
= 0; (saved_regs_mask
& (1 << regno
)) == 0; regno
++)
20096 /* If there's an odd number of registers to push. Start off by
20097 pushing a single register. This ensures that subsequent strd
20098 operations are dword aligned (assuming that SP was originally
20099 64-bit aligned). */
20100 if ((num_regs
& 1) != 0)
20102 rtx reg
, mem
, insn
;
20104 reg
= gen_rtx_REG (SImode
, regno
);
20106 mem
= gen_frame_mem (Pmode
, gen_rtx_PRE_DEC (Pmode
,
20107 stack_pointer_rtx
));
20109 mem
= gen_frame_mem (Pmode
,
20111 (Pmode
, stack_pointer_rtx
,
20112 plus_constant (Pmode
, stack_pointer_rtx
,
20115 tmp
= gen_rtx_SET (mem
, reg
);
20116 RTX_FRAME_RELATED_P (tmp
) = 1;
20117 insn
= emit_insn (tmp
);
20118 RTX_FRAME_RELATED_P (insn
) = 1;
20119 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
, dwarf
);
20120 tmp
= gen_rtx_SET (gen_frame_mem (Pmode
, stack_pointer_rtx
), reg
);
20121 RTX_FRAME_RELATED_P (tmp
) = 1;
20124 XVECEXP (dwarf
, 0, i
) = tmp
;
20128 while (i
< num_regs
)
20129 if (saved_regs_mask
& (1 << regno
))
20131 rtx reg1
, reg2
, mem1
, mem2
;
20132 rtx tmp0
, tmp1
, tmp2
;
20135 /* Find the register to pair with this one. */
20136 for (regno2
= regno
+ 1; (saved_regs_mask
& (1 << regno2
)) == 0;
20140 reg1
= gen_rtx_REG (SImode
, regno
);
20141 reg2
= gen_rtx_REG (SImode
, regno2
);
20148 mem1
= gen_frame_mem (Pmode
, plus_constant (Pmode
,
20151 mem2
= gen_frame_mem (Pmode
, plus_constant (Pmode
,
20153 -4 * (num_regs
- 1)));
20154 tmp0
= gen_rtx_SET (stack_pointer_rtx
,
20155 plus_constant (Pmode
, stack_pointer_rtx
,
20157 tmp1
= gen_rtx_SET (mem1
, reg1
);
20158 tmp2
= gen_rtx_SET (mem2
, reg2
);
20159 RTX_FRAME_RELATED_P (tmp0
) = 1;
20160 RTX_FRAME_RELATED_P (tmp1
) = 1;
20161 RTX_FRAME_RELATED_P (tmp2
) = 1;
20162 par
= gen_rtx_PARALLEL (VOIDmode
, rtvec_alloc (3));
20163 XVECEXP (par
, 0, 0) = tmp0
;
20164 XVECEXP (par
, 0, 1) = tmp1
;
20165 XVECEXP (par
, 0, 2) = tmp2
;
20166 insn
= emit_insn (par
);
20167 RTX_FRAME_RELATED_P (insn
) = 1;
20168 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
, dwarf
);
20172 mem1
= gen_frame_mem (Pmode
, plus_constant (Pmode
,
20175 mem2
= gen_frame_mem (Pmode
, plus_constant (Pmode
,
20178 tmp1
= gen_rtx_SET (mem1
, reg1
);
20179 tmp2
= gen_rtx_SET (mem2
, reg2
);
20180 RTX_FRAME_RELATED_P (tmp1
) = 1;
20181 RTX_FRAME_RELATED_P (tmp2
) = 1;
20182 par
= gen_rtx_PARALLEL (VOIDmode
, rtvec_alloc (2));
20183 XVECEXP (par
, 0, 0) = tmp1
;
20184 XVECEXP (par
, 0, 1) = tmp2
;
20188 /* Create unwind information. This is an approximation. */
20189 tmp1
= gen_rtx_SET (gen_frame_mem (Pmode
,
20190 plus_constant (Pmode
,
20194 tmp2
= gen_rtx_SET (gen_frame_mem (Pmode
,
20195 plus_constant (Pmode
,
20200 RTX_FRAME_RELATED_P (tmp1
) = 1;
20201 RTX_FRAME_RELATED_P (tmp2
) = 1;
20202 XVECEXP (dwarf
, 0, i
+ 1) = tmp1
;
20203 XVECEXP (dwarf
, 0, i
+ 2) = tmp2
;
20205 regno
= regno2
+ 1;
20213 /* STRD in ARM mode requires consecutive registers. This function emits STRD
20214 whenever possible, otherwise it emits single-word stores. The first store
20215 also allocates stack space for all saved registers, using writeback with
20216 post-addressing mode. All other stores use offset addressing. If no STRD
20217 can be emitted, this function emits a sequence of single-word stores,
20218 and not an STM as before, because single-word stores provide more freedom
20219 scheduling and can be turned into an STM by peephole optimizations. */
20221 arm_emit_strd_push (unsigned long saved_regs_mask
)
20224 int i
, j
, dwarf_index
= 0;
20226 rtx dwarf
= NULL_RTX
;
20227 rtx insn
= NULL_RTX
;
20230 /* TODO: A more efficient code can be emitted by changing the
20231 layout, e.g., first push all pairs that can use STRD to keep the
20232 stack aligned, and then push all other registers. */
20233 for (i
= 0; i
<= LAST_ARM_REGNUM
; i
++)
20234 if (saved_regs_mask
& (1 << i
))
20237 gcc_assert (!(saved_regs_mask
& (1 << SP_REGNUM
)));
20238 gcc_assert (!(saved_regs_mask
& (1 << PC_REGNUM
)));
20239 gcc_assert (num_regs
> 0);
20241 /* Create sequence for DWARF info. */
20242 dwarf
= gen_rtx_SEQUENCE (VOIDmode
, rtvec_alloc (num_regs
+ 1));
20244 /* For dwarf info, we generate explicit stack update. */
20245 tmp
= gen_rtx_SET (stack_pointer_rtx
,
20246 plus_constant (Pmode
, stack_pointer_rtx
, -4 * num_regs
));
20247 RTX_FRAME_RELATED_P (tmp
) = 1;
20248 XVECEXP (dwarf
, 0, dwarf_index
++) = tmp
;
20250 /* Save registers. */
20251 offset
= - 4 * num_regs
;
20253 while (j
<= LAST_ARM_REGNUM
)
20254 if (saved_regs_mask
& (1 << j
))
20257 && (saved_regs_mask
& (1 << (j
+ 1))))
20259 /* Current register and previous register form register pair for
20260 which STRD can be generated. */
20263 /* Allocate stack space for all saved registers. */
20264 tmp
= plus_constant (Pmode
, stack_pointer_rtx
, offset
);
20265 tmp
= gen_rtx_PRE_MODIFY (Pmode
, stack_pointer_rtx
, tmp
);
20266 mem
= gen_frame_mem (DImode
, tmp
);
20269 else if (offset
> 0)
20270 mem
= gen_frame_mem (DImode
,
20271 plus_constant (Pmode
,
20275 mem
= gen_frame_mem (DImode
, stack_pointer_rtx
);
20277 tmp
= gen_rtx_SET (mem
, gen_rtx_REG (DImode
, j
));
20278 RTX_FRAME_RELATED_P (tmp
) = 1;
20279 tmp
= emit_insn (tmp
);
20281 /* Record the first store insn. */
20282 if (dwarf_index
== 1)
20285 /* Generate dwarf info. */
20286 mem
= gen_frame_mem (SImode
,
20287 plus_constant (Pmode
,
20290 tmp
= gen_rtx_SET (mem
, gen_rtx_REG (SImode
, j
));
20291 RTX_FRAME_RELATED_P (tmp
) = 1;
20292 XVECEXP (dwarf
, 0, dwarf_index
++) = tmp
;
20294 mem
= gen_frame_mem (SImode
,
20295 plus_constant (Pmode
,
20298 tmp
= gen_rtx_SET (mem
, gen_rtx_REG (SImode
, j
+ 1));
20299 RTX_FRAME_RELATED_P (tmp
) = 1;
20300 XVECEXP (dwarf
, 0, dwarf_index
++) = tmp
;
20307 /* Emit a single word store. */
20310 /* Allocate stack space for all saved registers. */
20311 tmp
= plus_constant (Pmode
, stack_pointer_rtx
, offset
);
20312 tmp
= gen_rtx_PRE_MODIFY (Pmode
, stack_pointer_rtx
, tmp
);
20313 mem
= gen_frame_mem (SImode
, tmp
);
20316 else if (offset
> 0)
20317 mem
= gen_frame_mem (SImode
,
20318 plus_constant (Pmode
,
20322 mem
= gen_frame_mem (SImode
, stack_pointer_rtx
);
20324 tmp
= gen_rtx_SET (mem
, gen_rtx_REG (SImode
, j
));
20325 RTX_FRAME_RELATED_P (tmp
) = 1;
20326 tmp
= emit_insn (tmp
);
20328 /* Record the first store insn. */
20329 if (dwarf_index
== 1)
20332 /* Generate dwarf info. */
20333 mem
= gen_frame_mem (SImode
,
20334 plus_constant(Pmode
,
20337 tmp
= gen_rtx_SET (mem
, gen_rtx_REG (SImode
, j
));
20338 RTX_FRAME_RELATED_P (tmp
) = 1;
20339 XVECEXP (dwarf
, 0, dwarf_index
++) = tmp
;
20348 /* Attach dwarf info to the first insn we generate. */
20349 gcc_assert (insn
!= NULL_RTX
);
20350 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
, dwarf
);
20351 RTX_FRAME_RELATED_P (insn
) = 1;
20354 /* Generate and emit an insn that we will recognize as a push_multi.
20355 Unfortunately, since this insn does not reflect very well the actual
20356 semantics of the operation, we need to annotate the insn for the benefit
20357 of DWARF2 frame unwind information. DWARF_REGS_MASK is a subset of
20358 MASK for registers that should be annotated for DWARF2 frame unwind
20361 emit_multi_reg_push (unsigned long mask
, unsigned long dwarf_regs_mask
)
20364 int num_dwarf_regs
= 0;
20368 int dwarf_par_index
;
20371 /* We don't record the PC in the dwarf frame information. */
20372 dwarf_regs_mask
&= ~(1 << PC_REGNUM
);
20374 for (i
= 0; i
<= LAST_ARM_REGNUM
; i
++)
20376 if (mask
& (1 << i
))
20378 if (dwarf_regs_mask
& (1 << i
))
20382 gcc_assert (num_regs
&& num_regs
<= 16);
20383 gcc_assert ((dwarf_regs_mask
& ~mask
) == 0);
20385 /* For the body of the insn we are going to generate an UNSPEC in
20386 parallel with several USEs. This allows the insn to be recognized
20387 by the push_multi pattern in the arm.md file.
20389 The body of the insn looks something like this:
20392 (set (mem:BLK (pre_modify:SI (reg:SI sp)
20393 (const_int:SI <num>)))
20394 (unspec:BLK [(reg:SI r4)] UNSPEC_PUSH_MULT))
20400 For the frame note however, we try to be more explicit and actually
20401 show each register being stored into the stack frame, plus a (single)
20402 decrement of the stack pointer. We do it this way in order to be
20403 friendly to the stack unwinding code, which only wants to see a single
20404 stack decrement per instruction. The RTL we generate for the note looks
20405 something like this:
20408 (set (reg:SI sp) (plus:SI (reg:SI sp) (const_int -20)))
20409 (set (mem:SI (reg:SI sp)) (reg:SI r4))
20410 (set (mem:SI (plus:SI (reg:SI sp) (const_int 4))) (reg:SI XX))
20411 (set (mem:SI (plus:SI (reg:SI sp) (const_int 8))) (reg:SI YY))
20415 FIXME:: In an ideal world the PRE_MODIFY would not exist and
20416 instead we'd have a parallel expression detailing all
20417 the stores to the various memory addresses so that debug
20418 information is more up-to-date. Remember however while writing
20419 this to take care of the constraints with the push instruction.
20421 Note also that this has to be taken care of for the VFP registers.
20423 For more see PR43399. */
20425 par
= gen_rtx_PARALLEL (VOIDmode
, rtvec_alloc (num_regs
));
20426 dwarf
= gen_rtx_SEQUENCE (VOIDmode
, rtvec_alloc (num_dwarf_regs
+ 1));
20427 dwarf_par_index
= 1;
20429 for (i
= 0; i
<= LAST_ARM_REGNUM
; i
++)
20431 if (mask
& (1 << i
))
20433 reg
= gen_rtx_REG (SImode
, i
);
20435 XVECEXP (par
, 0, 0)
20436 = gen_rtx_SET (gen_frame_mem
20438 gen_rtx_PRE_MODIFY (Pmode
,
20441 (Pmode
, stack_pointer_rtx
,
20444 gen_rtx_UNSPEC (BLKmode
,
20445 gen_rtvec (1, reg
),
20446 UNSPEC_PUSH_MULT
));
20448 if (dwarf_regs_mask
& (1 << i
))
20450 tmp
= gen_rtx_SET (gen_frame_mem (SImode
, stack_pointer_rtx
),
20452 RTX_FRAME_RELATED_P (tmp
) = 1;
20453 XVECEXP (dwarf
, 0, dwarf_par_index
++) = tmp
;
20460 for (j
= 1, i
++; j
< num_regs
; i
++)
20462 if (mask
& (1 << i
))
20464 reg
= gen_rtx_REG (SImode
, i
);
20466 XVECEXP (par
, 0, j
) = gen_rtx_USE (VOIDmode
, reg
);
20468 if (dwarf_regs_mask
& (1 << i
))
20471 = gen_rtx_SET (gen_frame_mem
20473 plus_constant (Pmode
, stack_pointer_rtx
,
20476 RTX_FRAME_RELATED_P (tmp
) = 1;
20477 XVECEXP (dwarf
, 0, dwarf_par_index
++) = tmp
;
20484 par
= emit_insn (par
);
20486 tmp
= gen_rtx_SET (stack_pointer_rtx
,
20487 plus_constant (Pmode
, stack_pointer_rtx
, -4 * num_regs
));
20488 RTX_FRAME_RELATED_P (tmp
) = 1;
20489 XVECEXP (dwarf
, 0, 0) = tmp
;
20491 add_reg_note (par
, REG_FRAME_RELATED_EXPR
, dwarf
);
20496 /* Add a REG_CFA_ADJUST_CFA REG note to INSN.
20497 SIZE is the offset to be adjusted.
20498 DEST and SRC might be stack_pointer_rtx or hard_frame_pointer_rtx. */
20500 arm_add_cfa_adjust_cfa_note (rtx insn
, int size
, rtx dest
, rtx src
)
20504 RTX_FRAME_RELATED_P (insn
) = 1;
20505 dwarf
= gen_rtx_SET (dest
, plus_constant (Pmode
, src
, size
));
20506 add_reg_note (insn
, REG_CFA_ADJUST_CFA
, dwarf
);
20509 /* Generate and emit an insn pattern that we will recognize as a pop_multi.
20510 SAVED_REGS_MASK shows which registers need to be restored.
20512 Unfortunately, since this insn does not reflect very well the actual
20513 semantics of the operation, we need to annotate the insn for the benefit
20514 of DWARF2 frame unwind information. */
20516 arm_emit_multi_reg_pop (unsigned long saved_regs_mask
)
20521 rtx dwarf
= NULL_RTX
;
20523 bool return_in_pc
= saved_regs_mask
& (1 << PC_REGNUM
);
20527 offset_adj
= return_in_pc
? 1 : 0;
20528 for (i
= 0; i
<= LAST_ARM_REGNUM
; i
++)
20529 if (saved_regs_mask
& (1 << i
))
20532 gcc_assert (num_regs
&& num_regs
<= 16);
20534 /* If SP is in reglist, then we don't emit SP update insn. */
20535 emit_update
= (saved_regs_mask
& (1 << SP_REGNUM
)) ? 0 : 1;
20537 /* The parallel needs to hold num_regs SETs
20538 and one SET for the stack update. */
20539 par
= gen_rtx_PARALLEL (VOIDmode
, rtvec_alloc (num_regs
+ emit_update
+ offset_adj
));
20542 XVECEXP (par
, 0, 0) = ret_rtx
;
20546 /* Increment the stack pointer, based on there being
20547 num_regs 4-byte registers to restore. */
20548 tmp
= gen_rtx_SET (stack_pointer_rtx
,
20549 plus_constant (Pmode
,
20552 RTX_FRAME_RELATED_P (tmp
) = 1;
20553 XVECEXP (par
, 0, offset_adj
) = tmp
;
20556 /* Now restore every reg, which may include PC. */
20557 for (j
= 0, i
= 0; j
< num_regs
; i
++)
20558 if (saved_regs_mask
& (1 << i
))
20560 reg
= gen_rtx_REG (SImode
, i
);
20561 if ((num_regs
== 1) && emit_update
&& !return_in_pc
)
20563 /* Emit single load with writeback. */
20564 tmp
= gen_frame_mem (SImode
,
20565 gen_rtx_POST_INC (Pmode
,
20566 stack_pointer_rtx
));
20567 tmp
= emit_insn (gen_rtx_SET (reg
, tmp
));
20568 REG_NOTES (tmp
) = alloc_reg_note (REG_CFA_RESTORE
, reg
, dwarf
);
20572 tmp
= gen_rtx_SET (reg
,
20575 plus_constant (Pmode
, stack_pointer_rtx
, 4 * j
)));
20576 RTX_FRAME_RELATED_P (tmp
) = 1;
20577 XVECEXP (par
, 0, j
+ emit_update
+ offset_adj
) = tmp
;
20579 /* We need to maintain a sequence for DWARF info too. As dwarf info
20580 should not have PC, skip PC. */
20581 if (i
!= PC_REGNUM
)
20582 dwarf
= alloc_reg_note (REG_CFA_RESTORE
, reg
, dwarf
);
20588 par
= emit_jump_insn (par
);
20590 par
= emit_insn (par
);
20592 REG_NOTES (par
) = dwarf
;
20594 arm_add_cfa_adjust_cfa_note (par
, UNITS_PER_WORD
* num_regs
,
20595 stack_pointer_rtx
, stack_pointer_rtx
);
20598 /* Generate and emit an insn pattern that we will recognize as a pop_multi
20599 of NUM_REGS consecutive VFP regs, starting at FIRST_REG.
20601 Unfortunately, since this insn does not reflect very well the actual
20602 semantics of the operation, we need to annotate the insn for the benefit
20603 of DWARF2 frame unwind information. */
20605 arm_emit_vfp_multi_reg_pop (int first_reg
, int num_regs
, rtx base_reg
)
20609 rtx dwarf
= NULL_RTX
;
20612 gcc_assert (num_regs
&& num_regs
<= 32);
20614 /* Workaround ARM10 VFPr1 bug. */
20615 if (num_regs
== 2 && !arm_arch6
)
20617 if (first_reg
== 15)
20623 /* We can emit at most 16 D-registers in a single pop_multi instruction, and
20624 there could be up to 32 D-registers to restore.
20625 If there are more than 16 D-registers, make two recursive calls,
20626 each of which emits one pop_multi instruction. */
20629 arm_emit_vfp_multi_reg_pop (first_reg
, 16, base_reg
);
20630 arm_emit_vfp_multi_reg_pop (first_reg
+ 16, num_regs
- 16, base_reg
);
20634 /* The parallel needs to hold num_regs SETs
20635 and one SET for the stack update. */
20636 par
= gen_rtx_PARALLEL (VOIDmode
, rtvec_alloc (num_regs
+ 1));
20638 /* Increment the stack pointer, based on there being
20639 num_regs 8-byte registers to restore. */
20640 tmp
= gen_rtx_SET (base_reg
, plus_constant (Pmode
, base_reg
, 8 * num_regs
));
20641 RTX_FRAME_RELATED_P (tmp
) = 1;
20642 XVECEXP (par
, 0, 0) = tmp
;
20644 /* Now show every reg that will be restored, using a SET for each. */
20645 for (j
= 0, i
=first_reg
; j
< num_regs
; i
+= 2)
20647 reg
= gen_rtx_REG (DFmode
, i
);
20649 tmp
= gen_rtx_SET (reg
,
20652 plus_constant (Pmode
, base_reg
, 8 * j
)));
20653 RTX_FRAME_RELATED_P (tmp
) = 1;
20654 XVECEXP (par
, 0, j
+ 1) = tmp
;
20656 dwarf
= alloc_reg_note (REG_CFA_RESTORE
, reg
, dwarf
);
20661 par
= emit_insn (par
);
20662 REG_NOTES (par
) = dwarf
;
20664 /* Make sure cfa doesn't leave with IP_REGNUM to allow unwinding fron FP. */
20665 if (REGNO (base_reg
) == IP_REGNUM
)
20667 RTX_FRAME_RELATED_P (par
) = 1;
20668 add_reg_note (par
, REG_CFA_DEF_CFA
, hard_frame_pointer_rtx
);
20671 arm_add_cfa_adjust_cfa_note (par
, 2 * UNITS_PER_WORD
* num_regs
,
20672 base_reg
, base_reg
);
20675 /* Generate and emit a pattern that will be recognized as LDRD pattern. If even
20676 number of registers are being popped, multiple LDRD patterns are created for
20677 all register pairs. If odd number of registers are popped, last register is
20678 loaded by using LDR pattern. */
20680 thumb2_emit_ldrd_pop (unsigned long saved_regs_mask
)
20684 rtx par
= NULL_RTX
;
20685 rtx dwarf
= NULL_RTX
;
20686 rtx tmp
, reg
, tmp1
;
20687 bool return_in_pc
= saved_regs_mask
& (1 << PC_REGNUM
);
20689 for (i
= 0; i
<= LAST_ARM_REGNUM
; i
++)
20690 if (saved_regs_mask
& (1 << i
))
20693 gcc_assert (num_regs
&& num_regs
<= 16);
20695 /* We cannot generate ldrd for PC. Hence, reduce the count if PC is
20696 to be popped. So, if num_regs is even, now it will become odd,
20697 and we can generate pop with PC. If num_regs is odd, it will be
20698 even now, and ldr with return can be generated for PC. */
20702 gcc_assert (!(saved_regs_mask
& (1 << SP_REGNUM
)));
20704 /* Var j iterates over all the registers to gather all the registers in
20705 saved_regs_mask. Var i gives index of saved registers in stack frame.
20706 A PARALLEL RTX of register-pair is created here, so that pattern for
20707 LDRD can be matched. As PC is always last register to be popped, and
20708 we have already decremented num_regs if PC, we don't have to worry
20709 about PC in this loop. */
20710 for (i
= 0, j
= 0; i
< (num_regs
- (num_regs
% 2)); j
++)
20711 if (saved_regs_mask
& (1 << j
))
20713 /* Create RTX for memory load. */
20714 reg
= gen_rtx_REG (SImode
, j
);
20715 tmp
= gen_rtx_SET (reg
,
20716 gen_frame_mem (SImode
,
20717 plus_constant (Pmode
,
20718 stack_pointer_rtx
, 4 * i
)));
20719 RTX_FRAME_RELATED_P (tmp
) = 1;
20723 /* When saved-register index (i) is even, the RTX to be emitted is
20724 yet to be created. Hence create it first. The LDRD pattern we
20725 are generating is :
20726 [ (SET (reg_t0) (MEM (PLUS (SP) (NUM))))
20727 (SET (reg_t1) (MEM (PLUS (SP) (NUM + 4)))) ]
20728 where target registers need not be consecutive. */
20729 par
= gen_rtx_PARALLEL (VOIDmode
, rtvec_alloc (2));
20733 /* ith register is added in PARALLEL RTX. If i is even, the reg_i is
20734 added as 0th element and if i is odd, reg_i is added as 1st element
20735 of LDRD pattern shown above. */
20736 XVECEXP (par
, 0, (i
% 2)) = tmp
;
20737 dwarf
= alloc_reg_note (REG_CFA_RESTORE
, reg
, dwarf
);
20741 /* When saved-register index (i) is odd, RTXs for both the registers
20742 to be loaded are generated in above given LDRD pattern, and the
20743 pattern can be emitted now. */
20744 par
= emit_insn (par
);
20745 REG_NOTES (par
) = dwarf
;
20746 RTX_FRAME_RELATED_P (par
) = 1;
20752 /* If the number of registers pushed is odd AND return_in_pc is false OR
20753 number of registers are even AND return_in_pc is true, last register is
20754 popped using LDR. It can be PC as well. Hence, adjust the stack first and
20755 then LDR with post increment. */
20757 /* Increment the stack pointer, based on there being
20758 num_regs 4-byte registers to restore. */
20759 tmp
= gen_rtx_SET (stack_pointer_rtx
,
20760 plus_constant (Pmode
, stack_pointer_rtx
, 4 * i
));
20761 RTX_FRAME_RELATED_P (tmp
) = 1;
20762 tmp
= emit_insn (tmp
);
20765 arm_add_cfa_adjust_cfa_note (tmp
, UNITS_PER_WORD
* i
,
20766 stack_pointer_rtx
, stack_pointer_rtx
);
20771 if (((num_regs
% 2) == 1 && !return_in_pc
)
20772 || ((num_regs
% 2) == 0 && return_in_pc
))
20774 /* Scan for the single register to be popped. Skip until the saved
20775 register is found. */
20776 for (; (saved_regs_mask
& (1 << j
)) == 0; j
++);
20778 /* Gen LDR with post increment here. */
20779 tmp1
= gen_rtx_MEM (SImode
,
20780 gen_rtx_POST_INC (SImode
,
20781 stack_pointer_rtx
));
20782 set_mem_alias_set (tmp1
, get_frame_alias_set ());
20784 reg
= gen_rtx_REG (SImode
, j
);
20785 tmp
= gen_rtx_SET (reg
, tmp1
);
20786 RTX_FRAME_RELATED_P (tmp
) = 1;
20787 dwarf
= alloc_reg_note (REG_CFA_RESTORE
, reg
, dwarf
);
20791 /* If return_in_pc, j must be PC_REGNUM. */
20792 gcc_assert (j
== PC_REGNUM
);
20793 par
= gen_rtx_PARALLEL (VOIDmode
, rtvec_alloc (2));
20794 XVECEXP (par
, 0, 0) = ret_rtx
;
20795 XVECEXP (par
, 0, 1) = tmp
;
20796 par
= emit_jump_insn (par
);
20800 par
= emit_insn (tmp
);
20801 REG_NOTES (par
) = dwarf
;
20802 arm_add_cfa_adjust_cfa_note (par
, UNITS_PER_WORD
,
20803 stack_pointer_rtx
, stack_pointer_rtx
);
20807 else if ((num_regs
% 2) == 1 && return_in_pc
)
20809 /* There are 2 registers to be popped. So, generate the pattern
20810 pop_multiple_with_stack_update_and_return to pop in PC. */
20811 arm_emit_multi_reg_pop (saved_regs_mask
& (~((1 << j
) - 1)));
20817 /* LDRD in ARM mode needs consecutive registers as operands. This function
20818 emits LDRD whenever possible, otherwise it emits single-word loads. It uses
20819 offset addressing and then generates one separate stack udpate. This provides
20820 more scheduling freedom, compared to writeback on every load. However,
20821 if the function returns using load into PC directly
20822 (i.e., if PC is in SAVED_REGS_MASK), the stack needs to be updated
20823 before the last load. TODO: Add a peephole optimization to recognize
20824 the new epilogue sequence as an LDM instruction whenever possible. TODO: Add
20825 peephole optimization to merge the load at stack-offset zero
20826 with the stack update instruction using load with writeback
20827 in post-index addressing mode. */
20829 arm_emit_ldrd_pop (unsigned long saved_regs_mask
)
20833 rtx par
= NULL_RTX
;
20834 rtx dwarf
= NULL_RTX
;
20837 /* Restore saved registers. */
20838 gcc_assert (!((saved_regs_mask
& (1 << SP_REGNUM
))));
20840 while (j
<= LAST_ARM_REGNUM
)
20841 if (saved_regs_mask
& (1 << j
))
20844 && (saved_regs_mask
& (1 << (j
+ 1)))
20845 && (j
+ 1) != PC_REGNUM
)
20847 /* Current register and next register form register pair for which
20848 LDRD can be generated. PC is always the last register popped, and
20849 we handle it separately. */
20851 mem
= gen_frame_mem (DImode
,
20852 plus_constant (Pmode
,
20856 mem
= gen_frame_mem (DImode
, stack_pointer_rtx
);
20858 tmp
= gen_rtx_SET (gen_rtx_REG (DImode
, j
), mem
);
20859 tmp
= emit_insn (tmp
);
20860 RTX_FRAME_RELATED_P (tmp
) = 1;
20862 /* Generate dwarf info. */
20864 dwarf
= alloc_reg_note (REG_CFA_RESTORE
,
20865 gen_rtx_REG (SImode
, j
),
20867 dwarf
= alloc_reg_note (REG_CFA_RESTORE
,
20868 gen_rtx_REG (SImode
, j
+ 1),
20871 REG_NOTES (tmp
) = dwarf
;
20876 else if (j
!= PC_REGNUM
)
20878 /* Emit a single word load. */
20880 mem
= gen_frame_mem (SImode
,
20881 plus_constant (Pmode
,
20885 mem
= gen_frame_mem (SImode
, stack_pointer_rtx
);
20887 tmp
= gen_rtx_SET (gen_rtx_REG (SImode
, j
), mem
);
20888 tmp
= emit_insn (tmp
);
20889 RTX_FRAME_RELATED_P (tmp
) = 1;
20891 /* Generate dwarf info. */
20892 REG_NOTES (tmp
) = alloc_reg_note (REG_CFA_RESTORE
,
20893 gen_rtx_REG (SImode
, j
),
20899 else /* j == PC_REGNUM */
20905 /* Update the stack. */
20908 tmp
= gen_rtx_SET (stack_pointer_rtx
,
20909 plus_constant (Pmode
,
20912 tmp
= emit_insn (tmp
);
20913 arm_add_cfa_adjust_cfa_note (tmp
, offset
,
20914 stack_pointer_rtx
, stack_pointer_rtx
);
20918 if (saved_regs_mask
& (1 << PC_REGNUM
))
20920 /* Only PC is to be popped. */
20921 par
= gen_rtx_PARALLEL (VOIDmode
, rtvec_alloc (2));
20922 XVECEXP (par
, 0, 0) = ret_rtx
;
20923 tmp
= gen_rtx_SET (gen_rtx_REG (SImode
, PC_REGNUM
),
20924 gen_frame_mem (SImode
,
20925 gen_rtx_POST_INC (SImode
,
20926 stack_pointer_rtx
)));
20927 RTX_FRAME_RELATED_P (tmp
) = 1;
20928 XVECEXP (par
, 0, 1) = tmp
;
20929 par
= emit_jump_insn (par
);
20931 /* Generate dwarf info. */
20932 dwarf
= alloc_reg_note (REG_CFA_RESTORE
,
20933 gen_rtx_REG (SImode
, PC_REGNUM
),
20935 REG_NOTES (par
) = dwarf
;
20936 arm_add_cfa_adjust_cfa_note (par
, UNITS_PER_WORD
,
20937 stack_pointer_rtx
, stack_pointer_rtx
);
20941 /* Calculate the size of the return value that is passed in registers. */
20943 arm_size_return_regs (void)
20947 if (crtl
->return_rtx
!= 0)
20948 mode
= GET_MODE (crtl
->return_rtx
);
20950 mode
= DECL_MODE (DECL_RESULT (current_function_decl
));
20952 return GET_MODE_SIZE (mode
);
20955 /* Return true if the current function needs to save/restore LR. */
20957 thumb_force_lr_save (void)
20959 return !cfun
->machine
->lr_save_eliminated
20961 || thumb_far_jump_used_p ()
20962 || df_regs_ever_live_p (LR_REGNUM
));
20965 /* We do not know if r3 will be available because
20966 we do have an indirect tailcall happening in this
20967 particular case. */
20969 is_indirect_tailcall_p (rtx call
)
20971 rtx pat
= PATTERN (call
);
20973 /* Indirect tail call. */
20974 pat
= XVECEXP (pat
, 0, 0);
20975 if (GET_CODE (pat
) == SET
)
20976 pat
= SET_SRC (pat
);
20978 pat
= XEXP (XEXP (pat
, 0), 0);
20979 return REG_P (pat
);
20982 /* Return true if r3 is used by any of the tail call insns in the
20983 current function. */
20985 any_sibcall_could_use_r3 (void)
20990 if (!crtl
->tail_call_emit
)
20992 FOR_EACH_EDGE (e
, ei
, EXIT_BLOCK_PTR_FOR_FN (cfun
)->preds
)
20993 if (e
->flags
& EDGE_SIBCALL
)
20995 rtx_insn
*call
= BB_END (e
->src
);
20996 if (!CALL_P (call
))
20997 call
= prev_nonnote_nondebug_insn (call
);
20998 gcc_assert (CALL_P (call
) && SIBLING_CALL_P (call
));
20999 if (find_regno_fusage (call
, USE
, 3)
21000 || is_indirect_tailcall_p (call
))
21007 /* Compute the distance from register FROM to register TO.
21008 These can be the arg pointer (26), the soft frame pointer (25),
21009 the stack pointer (13) or the hard frame pointer (11).
21010 In thumb mode r7 is used as the soft frame pointer, if needed.
21011 Typical stack layout looks like this:
21013 old stack pointer -> | |
21016 | | saved arguments for
21017 | | vararg functions
21020 hard FP & arg pointer -> | | \
21028 soft frame pointer -> | | /
21033 locals base pointer -> | | /
21038 current stack pointer -> | | /
21041 For a given function some or all of these stack components
21042 may not be needed, giving rise to the possibility of
21043 eliminating some of the registers.
21045 The values returned by this function must reflect the behavior
21046 of arm_expand_prologue () and arm_compute_save_core_reg_mask ().
21048 The sign of the number returned reflects the direction of stack
21049 growth, so the values are positive for all eliminations except
21050 from the soft frame pointer to the hard frame pointer.
21052 SFP may point just inside the local variables block to ensure correct
21056 /* Return cached stack offsets. */
21058 static arm_stack_offsets
*
21059 arm_get_frame_offsets (void)
21061 struct arm_stack_offsets
*offsets
;
21063 offsets
= &cfun
->machine
->stack_offsets
;
21069 /* Calculate stack offsets. These are used to calculate register elimination
21070 offsets and in prologue/epilogue code. Also calculates which registers
21071 should be saved. */
21074 arm_compute_frame_layout (void)
21076 struct arm_stack_offsets
*offsets
;
21077 unsigned long func_type
;
21080 HOST_WIDE_INT frame_size
;
21083 offsets
= &cfun
->machine
->stack_offsets
;
21085 /* Initially this is the size of the local variables. It will translated
21086 into an offset once we have determined the size of preceding data. */
21087 frame_size
= ROUND_UP_WORD (get_frame_size ());
21089 /* Space for variadic functions. */
21090 offsets
->saved_args
= crtl
->args
.pretend_args_size
;
21092 /* In Thumb mode this is incorrect, but never used. */
21094 = (offsets
->saved_args
21095 + arm_compute_static_chain_stack_bytes ()
21096 + (frame_pointer_needed
? 4 : 0));
21100 unsigned int regno
;
21102 offsets
->saved_regs_mask
= arm_compute_save_core_reg_mask ();
21103 core_saved
= bit_count (offsets
->saved_regs_mask
) * 4;
21104 saved
= core_saved
;
21106 /* We know that SP will be doubleword aligned on entry, and we must
21107 preserve that condition at any subroutine call. We also require the
21108 soft frame pointer to be doubleword aligned. */
21110 if (TARGET_REALLY_IWMMXT
)
21112 /* Check for the call-saved iWMMXt registers. */
21113 for (regno
= FIRST_IWMMXT_REGNUM
;
21114 regno
<= LAST_IWMMXT_REGNUM
;
21116 if (df_regs_ever_live_p (regno
) && ! call_used_regs
[regno
])
21120 func_type
= arm_current_func_type ();
21121 /* Space for saved VFP registers. */
21122 if (! IS_VOLATILE (func_type
)
21123 && TARGET_HARD_FLOAT
)
21124 saved
+= arm_get_vfp_saved_size ();
21126 else /* TARGET_THUMB1 */
21128 offsets
->saved_regs_mask
= thumb1_compute_save_core_reg_mask ();
21129 core_saved
= bit_count (offsets
->saved_regs_mask
) * 4;
21130 saved
= core_saved
;
21131 if (TARGET_BACKTRACE
)
21135 /* Saved registers include the stack frame. */
21136 offsets
->saved_regs
21137 = offsets
->saved_args
+ arm_compute_static_chain_stack_bytes () + saved
;
21138 offsets
->soft_frame
= offsets
->saved_regs
+ CALLER_INTERWORKING_SLOT_SIZE
;
21140 /* A leaf function does not need any stack alignment if it has nothing
21142 if (crtl
->is_leaf
&& frame_size
== 0
21143 /* However if it calls alloca(), we have a dynamically allocated
21144 block of BIGGEST_ALIGNMENT on stack, so still do stack alignment. */
21145 && ! cfun
->calls_alloca
)
21147 offsets
->outgoing_args
= offsets
->soft_frame
;
21148 offsets
->locals_base
= offsets
->soft_frame
;
21152 /* Ensure SFP has the correct alignment. */
21153 if (ARM_DOUBLEWORD_ALIGN
21154 && (offsets
->soft_frame
& 7))
21156 offsets
->soft_frame
+= 4;
21157 /* Try to align stack by pushing an extra reg. Don't bother doing this
21158 when there is a stack frame as the alignment will be rolled into
21159 the normal stack adjustment. */
21160 if (frame_size
+ crtl
->outgoing_args_size
== 0)
21164 /* Register r3 is caller-saved. Normally it does not need to be
21165 saved on entry by the prologue. However if we choose to save
21166 it for padding then we may confuse the compiler into thinking
21167 a prologue sequence is required when in fact it is not. This
21168 will occur when shrink-wrapping if r3 is used as a scratch
21169 register and there are no other callee-saved writes.
21171 This situation can be avoided when other callee-saved registers
21172 are available and r3 is not mandatory if we choose a callee-saved
21173 register for padding. */
21174 bool prefer_callee_reg_p
= false;
21176 /* If it is safe to use r3, then do so. This sometimes
21177 generates better code on Thumb-2 by avoiding the need to
21178 use 32-bit push/pop instructions. */
21179 if (! any_sibcall_could_use_r3 ()
21180 && arm_size_return_regs () <= 12
21181 && (offsets
->saved_regs_mask
& (1 << 3)) == 0
21183 || !(TARGET_LDRD
&& current_tune
->prefer_ldrd_strd
)))
21186 if (!TARGET_THUMB2
)
21187 prefer_callee_reg_p
= true;
21190 || prefer_callee_reg_p
)
21192 for (i
= 4; i
<= (TARGET_THUMB1
? LAST_LO_REGNUM
: 11); i
++)
21194 /* Avoid fixed registers; they may be changed at
21195 arbitrary times so it's unsafe to restore them
21196 during the epilogue. */
21198 && (offsets
->saved_regs_mask
& (1 << i
)) == 0)
21208 offsets
->saved_regs
+= 4;
21209 offsets
->saved_regs_mask
|= (1 << reg
);
21214 offsets
->locals_base
= offsets
->soft_frame
+ frame_size
;
21215 offsets
->outgoing_args
= (offsets
->locals_base
21216 + crtl
->outgoing_args_size
);
21218 if (ARM_DOUBLEWORD_ALIGN
)
21220 /* Ensure SP remains doubleword aligned. */
21221 if (offsets
->outgoing_args
& 7)
21222 offsets
->outgoing_args
+= 4;
21223 gcc_assert (!(offsets
->outgoing_args
& 7));
21228 /* Calculate the relative offsets for the different stack pointers. Positive
21229 offsets are in the direction of stack growth. */
21232 arm_compute_initial_elimination_offset (unsigned int from
, unsigned int to
)
21234 arm_stack_offsets
*offsets
;
21236 offsets
= arm_get_frame_offsets ();
21238 /* OK, now we have enough information to compute the distances.
21239 There must be an entry in these switch tables for each pair
21240 of registers in ELIMINABLE_REGS, even if some of the entries
21241 seem to be redundant or useless. */
21244 case ARG_POINTER_REGNUM
:
21247 case THUMB_HARD_FRAME_POINTER_REGNUM
:
21250 case FRAME_POINTER_REGNUM
:
21251 /* This is the reverse of the soft frame pointer
21252 to hard frame pointer elimination below. */
21253 return offsets
->soft_frame
- offsets
->saved_args
;
21255 case ARM_HARD_FRAME_POINTER_REGNUM
:
21256 /* This is only non-zero in the case where the static chain register
21257 is stored above the frame. */
21258 return offsets
->frame
- offsets
->saved_args
- 4;
21260 case STACK_POINTER_REGNUM
:
21261 /* If nothing has been pushed on the stack at all
21262 then this will return -4. This *is* correct! */
21263 return offsets
->outgoing_args
- (offsets
->saved_args
+ 4);
21266 gcc_unreachable ();
21268 gcc_unreachable ();
21270 case FRAME_POINTER_REGNUM
:
21273 case THUMB_HARD_FRAME_POINTER_REGNUM
:
21276 case ARM_HARD_FRAME_POINTER_REGNUM
:
21277 /* The hard frame pointer points to the top entry in the
21278 stack frame. The soft frame pointer to the bottom entry
21279 in the stack frame. If there is no stack frame at all,
21280 then they are identical. */
21282 return offsets
->frame
- offsets
->soft_frame
;
21284 case STACK_POINTER_REGNUM
:
21285 return offsets
->outgoing_args
- offsets
->soft_frame
;
21288 gcc_unreachable ();
21290 gcc_unreachable ();
21293 /* You cannot eliminate from the stack pointer.
21294 In theory you could eliminate from the hard frame
21295 pointer to the stack pointer, but this will never
21296 happen, since if a stack frame is not needed the
21297 hard frame pointer will never be used. */
21298 gcc_unreachable ();
21302 /* Given FROM and TO register numbers, say whether this elimination is
21303 allowed. Frame pointer elimination is automatically handled.
21305 All eliminations are permissible. Note that ARG_POINTER_REGNUM and
21306 HARD_FRAME_POINTER_REGNUM are in fact the same thing. If we need a frame
21307 pointer, we must eliminate FRAME_POINTER_REGNUM into
21308 HARD_FRAME_POINTER_REGNUM and not into STACK_POINTER_REGNUM or
21309 ARG_POINTER_REGNUM. */
21312 arm_can_eliminate (const int from
, const int to
)
21314 return ((to
== FRAME_POINTER_REGNUM
&& from
== ARG_POINTER_REGNUM
) ? false :
21315 (to
== STACK_POINTER_REGNUM
&& frame_pointer_needed
) ? false :
21316 (to
== ARM_HARD_FRAME_POINTER_REGNUM
&& TARGET_THUMB
) ? false :
21317 (to
== THUMB_HARD_FRAME_POINTER_REGNUM
&& TARGET_ARM
) ? false :
21321 /* Emit RTL to save coprocessor registers on function entry. Returns the
21322 number of bytes pushed. */
21325 arm_save_coproc_regs(void)
21327 int saved_size
= 0;
21329 unsigned start_reg
;
21332 for (reg
= LAST_IWMMXT_REGNUM
; reg
>= FIRST_IWMMXT_REGNUM
; reg
--)
21333 if (df_regs_ever_live_p (reg
) && ! call_used_regs
[reg
])
21335 insn
= gen_rtx_PRE_DEC (Pmode
, stack_pointer_rtx
);
21336 insn
= gen_rtx_MEM (V2SImode
, insn
);
21337 insn
= emit_set_insn (insn
, gen_rtx_REG (V2SImode
, reg
));
21338 RTX_FRAME_RELATED_P (insn
) = 1;
21342 if (TARGET_HARD_FLOAT
)
21344 start_reg
= FIRST_VFP_REGNUM
;
21346 for (reg
= FIRST_VFP_REGNUM
; reg
< LAST_VFP_REGNUM
; reg
+= 2)
21348 if ((!df_regs_ever_live_p (reg
) || call_used_regs
[reg
])
21349 && (!df_regs_ever_live_p (reg
+ 1) || call_used_regs
[reg
+ 1]))
21351 if (start_reg
!= reg
)
21352 saved_size
+= vfp_emit_fstmd (start_reg
,
21353 (reg
- start_reg
) / 2);
21354 start_reg
= reg
+ 2;
21357 if (start_reg
!= reg
)
21358 saved_size
+= vfp_emit_fstmd (start_reg
,
21359 (reg
- start_reg
) / 2);
21365 /* Set the Thumb frame pointer from the stack pointer. */
21368 thumb_set_frame_pointer (arm_stack_offsets
*offsets
)
21370 HOST_WIDE_INT amount
;
21373 amount
= offsets
->outgoing_args
- offsets
->locals_base
;
21375 insn
= emit_insn (gen_addsi3 (hard_frame_pointer_rtx
,
21376 stack_pointer_rtx
, GEN_INT (amount
)));
21379 emit_insn (gen_movsi (hard_frame_pointer_rtx
, GEN_INT (amount
)));
21380 /* Thumb-2 RTL patterns expect sp as the first input. Thumb-1
21381 expects the first two operands to be the same. */
21384 insn
= emit_insn (gen_addsi3 (hard_frame_pointer_rtx
,
21386 hard_frame_pointer_rtx
));
21390 insn
= emit_insn (gen_addsi3 (hard_frame_pointer_rtx
,
21391 hard_frame_pointer_rtx
,
21392 stack_pointer_rtx
));
21394 dwarf
= gen_rtx_SET (hard_frame_pointer_rtx
,
21395 plus_constant (Pmode
, stack_pointer_rtx
, amount
));
21396 RTX_FRAME_RELATED_P (dwarf
) = 1;
21397 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
, dwarf
);
21400 RTX_FRAME_RELATED_P (insn
) = 1;
21403 struct scratch_reg
{
21408 /* Return a short-lived scratch register for use as a 2nd scratch register on
21409 function entry after the registers are saved in the prologue. This register
21410 must be released by means of release_scratch_register_on_entry. IP is not
21411 considered since it is always used as the 1st scratch register if available.
21413 REGNO1 is the index number of the 1st scratch register and LIVE_REGS is the
21414 mask of live registers. */
21417 get_scratch_register_on_entry (struct scratch_reg
*sr
, unsigned int regno1
,
21418 unsigned long live_regs
)
21424 if (regno1
!= LR_REGNUM
&& (live_regs
& (1 << LR_REGNUM
)) != 0)
21430 for (i
= 4; i
< 11; i
++)
21431 if (regno1
!= i
&& (live_regs
& (1 << i
)) != 0)
21439 /* If IP is used as the 1st scratch register for a nested function,
21440 then either r3 wasn't available or is used to preserve IP. */
21441 if (regno1
== IP_REGNUM
&& IS_NESTED (arm_current_func_type ()))
21443 regno
= (regno1
== 3 ? 2 : 3);
21445 = REGNO_REG_SET_P (df_get_live_out (ENTRY_BLOCK_PTR_FOR_FN (cfun
)),
21450 sr
->reg
= gen_rtx_REG (SImode
, regno
);
21453 rtx addr
= gen_rtx_PRE_DEC (Pmode
, stack_pointer_rtx
);
21454 rtx insn
= emit_set_insn (gen_frame_mem (SImode
, addr
), sr
->reg
);
21455 rtx x
= gen_rtx_SET (stack_pointer_rtx
,
21456 plus_constant (Pmode
, stack_pointer_rtx
, -4));
21457 RTX_FRAME_RELATED_P (insn
) = 1;
21458 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
, x
);
21462 /* Release a scratch register obtained from the preceding function. */
21465 release_scratch_register_on_entry (struct scratch_reg
*sr
)
21469 rtx addr
= gen_rtx_POST_INC (Pmode
, stack_pointer_rtx
);
21470 rtx insn
= emit_set_insn (sr
->reg
, gen_frame_mem (SImode
, addr
));
21471 rtx x
= gen_rtx_SET (stack_pointer_rtx
,
21472 plus_constant (Pmode
, stack_pointer_rtx
, 4));
21473 RTX_FRAME_RELATED_P (insn
) = 1;
21474 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
, x
);
21478 #define PROBE_INTERVAL (1 << STACK_CHECK_PROBE_INTERVAL_EXP)
21480 #if PROBE_INTERVAL > 4096
21481 #error Cannot use indexed addressing mode for stack probing
21484 /* Emit code to probe a range of stack addresses from FIRST to FIRST+SIZE,
21485 inclusive. These are offsets from the current stack pointer. REGNO1
21486 is the index number of the 1st scratch register and LIVE_REGS is the
21487 mask of live registers. */
21490 arm_emit_probe_stack_range (HOST_WIDE_INT first
, HOST_WIDE_INT size
,
21491 unsigned int regno1
, unsigned long live_regs
)
21493 rtx reg1
= gen_rtx_REG (Pmode
, regno1
);
21495 /* See if we have a constant small number of probes to generate. If so,
21496 that's the easy case. */
21497 if (size
<= PROBE_INTERVAL
)
21499 emit_move_insn (reg1
, GEN_INT (first
+ PROBE_INTERVAL
));
21500 emit_set_insn (reg1
, gen_rtx_MINUS (Pmode
, stack_pointer_rtx
, reg1
));
21501 emit_stack_probe (plus_constant (Pmode
, reg1
, PROBE_INTERVAL
- size
));
21504 /* The run-time loop is made up of 10 insns in the generic case while the
21505 compile-time loop is made up of 4+2*(n-2) insns for n # of intervals. */
21506 else if (size
<= 5 * PROBE_INTERVAL
)
21508 HOST_WIDE_INT i
, rem
;
21510 emit_move_insn (reg1
, GEN_INT (first
+ PROBE_INTERVAL
));
21511 emit_set_insn (reg1
, gen_rtx_MINUS (Pmode
, stack_pointer_rtx
, reg1
));
21512 emit_stack_probe (reg1
);
21514 /* Probe at FIRST + N * PROBE_INTERVAL for values of N from 2 until
21515 it exceeds SIZE. If only two probes are needed, this will not
21516 generate any code. Then probe at FIRST + SIZE. */
21517 for (i
= 2 * PROBE_INTERVAL
; i
< size
; i
+= PROBE_INTERVAL
)
21519 emit_set_insn (reg1
, plus_constant (Pmode
, reg1
, -PROBE_INTERVAL
));
21520 emit_stack_probe (reg1
);
21523 rem
= size
- (i
- PROBE_INTERVAL
);
21524 if (rem
> 4095 || (TARGET_THUMB2
&& rem
> 255))
21526 emit_set_insn (reg1
, plus_constant (Pmode
, reg1
, -PROBE_INTERVAL
));
21527 emit_stack_probe (plus_constant (Pmode
, reg1
, PROBE_INTERVAL
- rem
));
21530 emit_stack_probe (plus_constant (Pmode
, reg1
, -rem
));
21533 /* Otherwise, do the same as above, but in a loop. Note that we must be
21534 extra careful with variables wrapping around because we might be at
21535 the very top (or the very bottom) of the address space and we have
21536 to be able to handle this case properly; in particular, we use an
21537 equality test for the loop condition. */
21540 HOST_WIDE_INT rounded_size
;
21541 struct scratch_reg sr
;
21543 get_scratch_register_on_entry (&sr
, regno1
, live_regs
);
21545 emit_move_insn (reg1
, GEN_INT (first
));
21548 /* Step 1: round SIZE to the previous multiple of the interval. */
21550 rounded_size
= size
& -PROBE_INTERVAL
;
21551 emit_move_insn (sr
.reg
, GEN_INT (rounded_size
));
21554 /* Step 2: compute initial and final value of the loop counter. */
21556 /* TEST_ADDR = SP + FIRST. */
21557 emit_set_insn (reg1
, gen_rtx_MINUS (Pmode
, stack_pointer_rtx
, reg1
));
21559 /* LAST_ADDR = SP + FIRST + ROUNDED_SIZE. */
21560 emit_set_insn (sr
.reg
, gen_rtx_MINUS (Pmode
, reg1
, sr
.reg
));
21563 /* Step 3: the loop
21567 TEST_ADDR = TEST_ADDR + PROBE_INTERVAL
21570 while (TEST_ADDR != LAST_ADDR)
21572 probes at FIRST + N * PROBE_INTERVAL for values of N from 1
21573 until it is equal to ROUNDED_SIZE. */
21575 emit_insn (gen_probe_stack_range (reg1
, reg1
, sr
.reg
));
21578 /* Step 4: probe at FIRST + SIZE if we cannot assert at compile-time
21579 that SIZE is equal to ROUNDED_SIZE. */
21581 if (size
!= rounded_size
)
21583 HOST_WIDE_INT rem
= size
- rounded_size
;
21585 if (rem
> 4095 || (TARGET_THUMB2
&& rem
> 255))
21587 emit_set_insn (sr
.reg
,
21588 plus_constant (Pmode
, sr
.reg
, -PROBE_INTERVAL
));
21589 emit_stack_probe (plus_constant (Pmode
, sr
.reg
,
21590 PROBE_INTERVAL
- rem
));
21593 emit_stack_probe (plus_constant (Pmode
, sr
.reg
, -rem
));
21596 release_scratch_register_on_entry (&sr
);
21599 /* Make sure nothing is scheduled before we are done. */
21600 emit_insn (gen_blockage ());
21603 /* Probe a range of stack addresses from REG1 to REG2 inclusive. These are
21604 absolute addresses. */
21607 output_probe_stack_range (rtx reg1
, rtx reg2
)
21609 static int labelno
= 0;
21613 ASM_GENERATE_INTERNAL_LABEL (loop_lab
, "LPSRL", labelno
++);
21616 ASM_OUTPUT_INTERNAL_LABEL (asm_out_file
, loop_lab
);
21618 /* TEST_ADDR = TEST_ADDR + PROBE_INTERVAL. */
21620 xops
[1] = GEN_INT (PROBE_INTERVAL
);
21621 output_asm_insn ("sub\t%0, %0, %1", xops
);
21623 /* Probe at TEST_ADDR. */
21624 output_asm_insn ("str\tr0, [%0, #0]", xops
);
21626 /* Test if TEST_ADDR == LAST_ADDR. */
21628 output_asm_insn ("cmp\t%0, %1", xops
);
21631 fputs ("\tbne\t", asm_out_file
);
21632 assemble_name_raw (asm_out_file
, loop_lab
);
21633 fputc ('\n', asm_out_file
);
21638 /* Generate the prologue instructions for entry into an ARM or Thumb-2
21641 arm_expand_prologue (void)
21646 unsigned long live_regs_mask
;
21647 unsigned long func_type
;
21649 int saved_pretend_args
= 0;
21650 int saved_regs
= 0;
21651 unsigned HOST_WIDE_INT args_to_push
;
21652 HOST_WIDE_INT size
;
21653 arm_stack_offsets
*offsets
;
21656 func_type
= arm_current_func_type ();
21658 /* Naked functions don't have prologues. */
21659 if (IS_NAKED (func_type
))
21661 if (flag_stack_usage_info
)
21662 current_function_static_stack_size
= 0;
21666 /* Make a copy of c_f_p_a_s as we may need to modify it locally. */
21667 args_to_push
= crtl
->args
.pretend_args_size
;
21669 /* Compute which register we will have to save onto the stack. */
21670 offsets
= arm_get_frame_offsets ();
21671 live_regs_mask
= offsets
->saved_regs_mask
;
21673 ip_rtx
= gen_rtx_REG (SImode
, IP_REGNUM
);
21675 if (IS_STACKALIGN (func_type
))
21679 /* Handle a word-aligned stack pointer. We generate the following:
21684 <save and restore r0 in normal prologue/epilogue>
21688 The unwinder doesn't need to know about the stack realignment.
21689 Just tell it we saved SP in r0. */
21690 gcc_assert (TARGET_THUMB2
&& !arm_arch_notm
&& args_to_push
== 0);
21692 r0
= gen_rtx_REG (SImode
, R0_REGNUM
);
21693 r1
= gen_rtx_REG (SImode
, R1_REGNUM
);
21695 insn
= emit_insn (gen_movsi (r0
, stack_pointer_rtx
));
21696 RTX_FRAME_RELATED_P (insn
) = 1;
21697 add_reg_note (insn
, REG_CFA_REGISTER
, NULL
);
21699 emit_insn (gen_andsi3 (r1
, r0
, GEN_INT (~(HOST_WIDE_INT
)7)));
21701 /* ??? The CFA changes here, which may cause GDB to conclude that it
21702 has entered a different function. That said, the unwind info is
21703 correct, individually, before and after this instruction because
21704 we've described the save of SP, which will override the default
21705 handling of SP as restoring from the CFA. */
21706 emit_insn (gen_movsi (stack_pointer_rtx
, r1
));
21709 /* Let's compute the static_chain_stack_bytes required and store it. Right
21710 now the value must be -1 as stored by arm_init_machine_status (). */
21711 cfun
->machine
->static_chain_stack_bytes
21712 = arm_compute_static_chain_stack_bytes ();
21714 /* The static chain register is the same as the IP register. If it is
21715 clobbered when creating the frame, we need to save and restore it. */
21716 clobber_ip
= IS_NESTED (func_type
)
21717 && ((TARGET_APCS_FRAME
&& frame_pointer_needed
&& TARGET_ARM
)
21718 || ((flag_stack_check
== STATIC_BUILTIN_STACK_CHECK
21719 || flag_stack_clash_protection
)
21720 && !df_regs_ever_live_p (LR_REGNUM
)
21721 && arm_r3_live_at_start_p ()));
21723 /* Find somewhere to store IP whilst the frame is being created.
21724 We try the following places in order:
21726 1. The last argument register r3 if it is available.
21727 2. A slot on the stack above the frame if there are no
21728 arguments to push onto the stack.
21729 3. Register r3 again, after pushing the argument registers
21730 onto the stack, if this is a varargs function.
21731 4. The last slot on the stack created for the arguments to
21732 push, if this isn't a varargs function.
21734 Note - we only need to tell the dwarf2 backend about the SP
21735 adjustment in the second variant; the static chain register
21736 doesn't need to be unwound, as it doesn't contain a value
21737 inherited from the caller. */
21740 if (!arm_r3_live_at_start_p ())
21741 insn
= emit_set_insn (gen_rtx_REG (SImode
, 3), ip_rtx
);
21742 else if (args_to_push
== 0)
21746 gcc_assert(arm_compute_static_chain_stack_bytes() == 4);
21749 addr
= gen_rtx_PRE_DEC (Pmode
, stack_pointer_rtx
);
21750 insn
= emit_set_insn (gen_frame_mem (SImode
, addr
), ip_rtx
);
21753 /* Just tell the dwarf backend that we adjusted SP. */
21754 dwarf
= gen_rtx_SET (stack_pointer_rtx
,
21755 plus_constant (Pmode
, stack_pointer_rtx
,
21757 RTX_FRAME_RELATED_P (insn
) = 1;
21758 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
, dwarf
);
21762 /* Store the args on the stack. */
21763 if (cfun
->machine
->uses_anonymous_args
)
21765 insn
= emit_multi_reg_push ((0xf0 >> (args_to_push
/ 4)) & 0xf,
21766 (0xf0 >> (args_to_push
/ 4)) & 0xf);
21767 emit_set_insn (gen_rtx_REG (SImode
, 3), ip_rtx
);
21768 saved_pretend_args
= 1;
21774 if (args_to_push
== 4)
21775 addr
= gen_rtx_PRE_DEC (Pmode
, stack_pointer_rtx
);
21777 addr
= gen_rtx_PRE_MODIFY (Pmode
, stack_pointer_rtx
,
21778 plus_constant (Pmode
,
21782 insn
= emit_set_insn (gen_frame_mem (SImode
, addr
), ip_rtx
);
21784 /* Just tell the dwarf backend that we adjusted SP. */
21785 dwarf
= gen_rtx_SET (stack_pointer_rtx
,
21786 plus_constant (Pmode
, stack_pointer_rtx
,
21788 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
, dwarf
);
21791 RTX_FRAME_RELATED_P (insn
) = 1;
21792 fp_offset
= args_to_push
;
21797 if (TARGET_APCS_FRAME
&& frame_pointer_needed
&& TARGET_ARM
)
21799 if (IS_INTERRUPT (func_type
))
21801 /* Interrupt functions must not corrupt any registers.
21802 Creating a frame pointer however, corrupts the IP
21803 register, so we must push it first. */
21804 emit_multi_reg_push (1 << IP_REGNUM
, 1 << IP_REGNUM
);
21806 /* Do not set RTX_FRAME_RELATED_P on this insn.
21807 The dwarf stack unwinding code only wants to see one
21808 stack decrement per function, and this is not it. If
21809 this instruction is labeled as being part of the frame
21810 creation sequence then dwarf2out_frame_debug_expr will
21811 die when it encounters the assignment of IP to FP
21812 later on, since the use of SP here establishes SP as
21813 the CFA register and not IP.
21815 Anyway this instruction is not really part of the stack
21816 frame creation although it is part of the prologue. */
21819 insn
= emit_set_insn (ip_rtx
,
21820 plus_constant (Pmode
, stack_pointer_rtx
,
21822 RTX_FRAME_RELATED_P (insn
) = 1;
21827 /* Push the argument registers, or reserve space for them. */
21828 if (cfun
->machine
->uses_anonymous_args
)
21829 insn
= emit_multi_reg_push
21830 ((0xf0 >> (args_to_push
/ 4)) & 0xf,
21831 (0xf0 >> (args_to_push
/ 4)) & 0xf);
21834 (gen_addsi3 (stack_pointer_rtx
, stack_pointer_rtx
,
21835 GEN_INT (- args_to_push
)));
21836 RTX_FRAME_RELATED_P (insn
) = 1;
21839 /* If this is an interrupt service routine, and the link register
21840 is going to be pushed, and we're not generating extra
21841 push of IP (needed when frame is needed and frame layout if apcs),
21842 subtracting four from LR now will mean that the function return
21843 can be done with a single instruction. */
21844 if ((func_type
== ARM_FT_ISR
|| func_type
== ARM_FT_FIQ
)
21845 && (live_regs_mask
& (1 << LR_REGNUM
)) != 0
21846 && !(frame_pointer_needed
&& TARGET_APCS_FRAME
)
21849 rtx lr
= gen_rtx_REG (SImode
, LR_REGNUM
);
21851 emit_set_insn (lr
, plus_constant (SImode
, lr
, -4));
21854 if (live_regs_mask
)
21856 unsigned long dwarf_regs_mask
= live_regs_mask
;
21858 saved_regs
+= bit_count (live_regs_mask
) * 4;
21859 if (optimize_size
&& !frame_pointer_needed
21860 && saved_regs
== offsets
->saved_regs
- offsets
->saved_args
)
21862 /* If no coprocessor registers are being pushed and we don't have
21863 to worry about a frame pointer then push extra registers to
21864 create the stack frame. This is done in a way that does not
21865 alter the frame layout, so is independent of the epilogue. */
21869 while (n
< 8 && (live_regs_mask
& (1 << n
)) == 0)
21871 frame
= offsets
->outgoing_args
- (offsets
->saved_args
+ saved_regs
);
21872 if (frame
&& n
* 4 >= frame
)
21875 live_regs_mask
|= (1 << n
) - 1;
21876 saved_regs
+= frame
;
21881 && current_tune
->prefer_ldrd_strd
21882 && !optimize_function_for_size_p (cfun
))
21884 gcc_checking_assert (live_regs_mask
== dwarf_regs_mask
);
21886 thumb2_emit_strd_push (live_regs_mask
);
21887 else if (TARGET_ARM
21888 && !TARGET_APCS_FRAME
21889 && !IS_INTERRUPT (func_type
))
21890 arm_emit_strd_push (live_regs_mask
);
21893 insn
= emit_multi_reg_push (live_regs_mask
, live_regs_mask
);
21894 RTX_FRAME_RELATED_P (insn
) = 1;
21899 insn
= emit_multi_reg_push (live_regs_mask
, dwarf_regs_mask
);
21900 RTX_FRAME_RELATED_P (insn
) = 1;
21904 if (! IS_VOLATILE (func_type
))
21905 saved_regs
+= arm_save_coproc_regs ();
21907 if (frame_pointer_needed
&& TARGET_ARM
)
21909 /* Create the new frame pointer. */
21910 if (TARGET_APCS_FRAME
)
21912 insn
= GEN_INT (-(4 + args_to_push
+ fp_offset
));
21913 insn
= emit_insn (gen_addsi3 (hard_frame_pointer_rtx
, ip_rtx
, insn
));
21914 RTX_FRAME_RELATED_P (insn
) = 1;
21918 insn
= GEN_INT (saved_regs
- (4 + fp_offset
));
21919 insn
= emit_insn (gen_addsi3 (hard_frame_pointer_rtx
,
21920 stack_pointer_rtx
, insn
));
21921 RTX_FRAME_RELATED_P (insn
) = 1;
21925 size
= offsets
->outgoing_args
- offsets
->saved_args
;
21926 if (flag_stack_usage_info
)
21927 current_function_static_stack_size
= size
;
21929 /* If this isn't an interrupt service routine and we have a frame, then do
21930 stack checking. We use IP as the first scratch register, except for the
21931 non-APCS nested functions if LR or r3 are available (see clobber_ip). */
21932 if (!IS_INTERRUPT (func_type
)
21933 && (flag_stack_check
== STATIC_BUILTIN_STACK_CHECK
21934 || flag_stack_clash_protection
))
21936 unsigned int regno
;
21938 if (!IS_NESTED (func_type
) || clobber_ip
)
21940 else if (df_regs_ever_live_p (LR_REGNUM
))
21945 if (crtl
->is_leaf
&& !cfun
->calls_alloca
)
21947 if (size
> PROBE_INTERVAL
&& size
> get_stack_check_protect ())
21948 arm_emit_probe_stack_range (get_stack_check_protect (),
21949 size
- get_stack_check_protect (),
21950 regno
, live_regs_mask
);
21953 arm_emit_probe_stack_range (get_stack_check_protect (), size
,
21954 regno
, live_regs_mask
);
21957 /* Recover the static chain register. */
21960 if (!arm_r3_live_at_start_p () || saved_pretend_args
)
21961 insn
= gen_rtx_REG (SImode
, 3);
21964 insn
= plus_constant (Pmode
, hard_frame_pointer_rtx
, 4);
21965 insn
= gen_frame_mem (SImode
, insn
);
21967 emit_set_insn (ip_rtx
, insn
);
21968 emit_insn (gen_force_register_use (ip_rtx
));
21971 if (offsets
->outgoing_args
!= offsets
->saved_args
+ saved_regs
)
21973 /* This add can produce multiple insns for a large constant, so we
21974 need to get tricky. */
21975 rtx_insn
*last
= get_last_insn ();
21977 amount
= GEN_INT (offsets
->saved_args
+ saved_regs
21978 - offsets
->outgoing_args
);
21980 insn
= emit_insn (gen_addsi3 (stack_pointer_rtx
, stack_pointer_rtx
,
21984 last
= last
? NEXT_INSN (last
) : get_insns ();
21985 RTX_FRAME_RELATED_P (last
) = 1;
21987 while (last
!= insn
);
21989 /* If the frame pointer is needed, emit a special barrier that
21990 will prevent the scheduler from moving stores to the frame
21991 before the stack adjustment. */
21992 if (frame_pointer_needed
)
21993 emit_insn (gen_stack_tie (stack_pointer_rtx
,
21994 hard_frame_pointer_rtx
));
21998 if (frame_pointer_needed
&& TARGET_THUMB2
)
21999 thumb_set_frame_pointer (offsets
);
22001 if (flag_pic
&& arm_pic_register
!= INVALID_REGNUM
)
22003 unsigned long mask
;
22005 mask
= live_regs_mask
;
22006 mask
&= THUMB2_WORK_REGS
;
22007 if (!IS_NESTED (func_type
))
22008 mask
|= (1 << IP_REGNUM
);
22009 arm_load_pic_register (mask
);
22012 /* If we are profiling, make sure no instructions are scheduled before
22013 the call to mcount. Similarly if the user has requested no
22014 scheduling in the prolog. Similarly if we want non-call exceptions
22015 using the EABI unwinder, to prevent faulting instructions from being
22016 swapped with a stack adjustment. */
22017 if (crtl
->profile
|| !TARGET_SCHED_PROLOG
22018 || (arm_except_unwind_info (&global_options
) == UI_TARGET
22019 && cfun
->can_throw_non_call_exceptions
))
22020 emit_insn (gen_blockage ());
22022 /* If the link register is being kept alive, with the return address in it,
22023 then make sure that it does not get reused by the ce2 pass. */
22024 if ((live_regs_mask
& (1 << LR_REGNUM
)) == 0)
22025 cfun
->machine
->lr_save_eliminated
= 1;
22028 /* Print condition code to STREAM. Helper function for arm_print_operand. */
22030 arm_print_condition (FILE *stream
)
22032 if (arm_ccfsm_state
== 3 || arm_ccfsm_state
== 4)
22034 /* Branch conversion is not implemented for Thumb-2. */
22037 output_operand_lossage ("predicated Thumb instruction");
22040 if (current_insn_predicate
!= NULL
)
22042 output_operand_lossage
22043 ("predicated instruction in conditional sequence");
22047 fputs (arm_condition_codes
[arm_current_cc
], stream
);
22049 else if (current_insn_predicate
)
22051 enum arm_cond_code code
;
22055 output_operand_lossage ("predicated Thumb instruction");
22059 code
= get_arm_condition_code (current_insn_predicate
);
22060 fputs (arm_condition_codes
[code
], stream
);
22065 /* Globally reserved letters: acln
22066 Puncutation letters currently used: @_|?().!#
22067 Lower case letters currently used: bcdefhimpqtvwxyz
22068 Upper case letters currently used: ABCDFGHJKLMNOPQRSTU
22069 Letters previously used, but now deprecated/obsolete: sVWXYZ.
22071 Note that the global reservation for 'c' is only for CONSTANT_ADDRESS_P.
22073 If CODE is 'd', then the X is a condition operand and the instruction
22074 should only be executed if the condition is true.
22075 if CODE is 'D', then the X is a condition operand and the instruction
22076 should only be executed if the condition is false: however, if the mode
22077 of the comparison is CCFPEmode, then always execute the instruction -- we
22078 do this because in these circumstances !GE does not necessarily imply LT;
22079 in these cases the instruction pattern will take care to make sure that
22080 an instruction containing %d will follow, thereby undoing the effects of
22081 doing this instruction unconditionally.
22082 If CODE is 'N' then X is a floating point operand that must be negated
22084 If CODE is 'B' then output a bitwise inverted value of X (a const int).
22085 If X is a REG and CODE is `M', output a ldm/stm style multi-reg. */
22087 arm_print_operand (FILE *stream
, rtx x
, int code
)
22092 fputs (ASM_COMMENT_START
, stream
);
22096 fputs (user_label_prefix
, stream
);
22100 fputs (REGISTER_PREFIX
, stream
);
22104 arm_print_condition (stream
);
22108 /* The current condition code for a condition code setting instruction.
22109 Preceded by 's' in unified syntax, otherwise followed by 's'. */
22110 fputc('s', stream
);
22111 arm_print_condition (stream
);
22115 /* If the instruction is conditionally executed then print
22116 the current condition code, otherwise print 's'. */
22117 gcc_assert (TARGET_THUMB2
);
22118 if (current_insn_predicate
)
22119 arm_print_condition (stream
);
22121 fputc('s', stream
);
22124 /* %# is a "break" sequence. It doesn't output anything, but is used to
22125 separate e.g. operand numbers from following text, if that text consists
22126 of further digits which we don't want to be part of the operand
22134 r
= real_value_negate (CONST_DOUBLE_REAL_VALUE (x
));
22135 fprintf (stream
, "%s", fp_const_from_val (&r
));
22139 /* An integer or symbol address without a preceding # sign. */
22141 switch (GET_CODE (x
))
22144 fprintf (stream
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
));
22148 output_addr_const (stream
, x
);
22152 if (GET_CODE (XEXP (x
, 0)) == PLUS
22153 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == SYMBOL_REF
)
22155 output_addr_const (stream
, x
);
22158 /* Fall through. */
22161 output_operand_lossage ("Unsupported operand for code '%c'", code
);
22165 /* An integer that we want to print in HEX. */
22167 switch (GET_CODE (x
))
22170 fprintf (stream
, "#" HOST_WIDE_INT_PRINT_HEX
, INTVAL (x
));
22174 output_operand_lossage ("Unsupported operand for code '%c'", code
);
22179 if (CONST_INT_P (x
))
22182 val
= ARM_SIGN_EXTEND (~INTVAL (x
));
22183 fprintf (stream
, HOST_WIDE_INT_PRINT_DEC
, val
);
22187 putc ('~', stream
);
22188 output_addr_const (stream
, x
);
22193 /* Print the log2 of a CONST_INT. */
22197 if (!CONST_INT_P (x
)
22198 || (val
= exact_log2 (INTVAL (x
) & 0xffffffff)) < 0)
22199 output_operand_lossage ("Unsupported operand for code '%c'", code
);
22201 fprintf (stream
, "#" HOST_WIDE_INT_PRINT_DEC
, val
);
22206 /* The low 16 bits of an immediate constant. */
22207 fprintf (stream
, HOST_WIDE_INT_PRINT_DEC
, INTVAL(x
) & 0xffff);
22211 fprintf (stream
, "%s", arithmetic_instr (x
, 1));
22215 fprintf (stream
, "%s", arithmetic_instr (x
, 0));
22223 shift
= shift_op (x
, &val
);
22227 fprintf (stream
, ", %s ", shift
);
22229 arm_print_operand (stream
, XEXP (x
, 1), 0);
22231 fprintf (stream
, "#" HOST_WIDE_INT_PRINT_DEC
, val
);
22236 /* An explanation of the 'Q', 'R' and 'H' register operands:
22238 In a pair of registers containing a DI or DF value the 'Q'
22239 operand returns the register number of the register containing
22240 the least significant part of the value. The 'R' operand returns
22241 the register number of the register containing the most
22242 significant part of the value.
22244 The 'H' operand returns the higher of the two register numbers.
22245 On a run where WORDS_BIG_ENDIAN is true the 'H' operand is the
22246 same as the 'Q' operand, since the most significant part of the
22247 value is held in the lower number register. The reverse is true
22248 on systems where WORDS_BIG_ENDIAN is false.
22250 The purpose of these operands is to distinguish between cases
22251 where the endian-ness of the values is important (for example
22252 when they are added together), and cases where the endian-ness
22253 is irrelevant, but the order of register operations is important.
22254 For example when loading a value from memory into a register
22255 pair, the endian-ness does not matter. Provided that the value
22256 from the lower memory address is put into the lower numbered
22257 register, and the value from the higher address is put into the
22258 higher numbered register, the load will work regardless of whether
22259 the value being loaded is big-wordian or little-wordian. The
22260 order of the two register loads can matter however, if the address
22261 of the memory location is actually held in one of the registers
22262 being overwritten by the load.
22264 The 'Q' and 'R' constraints are also available for 64-bit
22267 if (CONST_INT_P (x
) || CONST_DOUBLE_P (x
))
22269 rtx part
= gen_lowpart (SImode
, x
);
22270 fprintf (stream
, "#" HOST_WIDE_INT_PRINT_DEC
, INTVAL (part
));
22274 if (!REG_P (x
) || REGNO (x
) > LAST_ARM_REGNUM
)
22276 output_operand_lossage ("invalid operand for code '%c'", code
);
22280 asm_fprintf (stream
, "%r", REGNO (x
) + (WORDS_BIG_ENDIAN
? 1 : 0));
22284 if (CONST_INT_P (x
) || CONST_DOUBLE_P (x
))
22286 machine_mode mode
= GET_MODE (x
);
22289 if (mode
== VOIDmode
)
22291 part
= gen_highpart_mode (SImode
, mode
, x
);
22292 fprintf (stream
, "#" HOST_WIDE_INT_PRINT_DEC
, INTVAL (part
));
22296 if (!REG_P (x
) || REGNO (x
) > LAST_ARM_REGNUM
)
22298 output_operand_lossage ("invalid operand for code '%c'", code
);
22302 asm_fprintf (stream
, "%r", REGNO (x
) + (WORDS_BIG_ENDIAN
? 0 : 1));
22306 if (!REG_P (x
) || REGNO (x
) > LAST_ARM_REGNUM
)
22308 output_operand_lossage ("invalid operand for code '%c'", code
);
22312 asm_fprintf (stream
, "%r", REGNO (x
) + 1);
22316 if (!REG_P (x
) || REGNO (x
) > LAST_ARM_REGNUM
)
22318 output_operand_lossage ("invalid operand for code '%c'", code
);
22322 asm_fprintf (stream
, "%r", REGNO (x
) + (WORDS_BIG_ENDIAN
? 3 : 2));
22326 if (!REG_P (x
) || REGNO (x
) > LAST_ARM_REGNUM
)
22328 output_operand_lossage ("invalid operand for code '%c'", code
);
22332 asm_fprintf (stream
, "%r", REGNO (x
) + (WORDS_BIG_ENDIAN
? 2 : 3));
22336 asm_fprintf (stream
, "%r",
22337 REG_P (XEXP (x
, 0))
22338 ? REGNO (XEXP (x
, 0)) : REGNO (XEXP (XEXP (x
, 0), 0)));
22342 asm_fprintf (stream
, "{%r-%r}",
22344 REGNO (x
) + ARM_NUM_REGS (GET_MODE (x
)) - 1);
22347 /* Like 'M', but writing doubleword vector registers, for use by Neon
22351 int regno
= (REGNO (x
) - FIRST_VFP_REGNUM
) / 2;
22352 int numregs
= ARM_NUM_REGS (GET_MODE (x
)) / 2;
22354 asm_fprintf (stream
, "{d%d}", regno
);
22356 asm_fprintf (stream
, "{d%d-d%d}", regno
, regno
+ numregs
- 1);
22361 /* CONST_TRUE_RTX means always -- that's the default. */
22362 if (x
== const_true_rtx
)
22365 if (!COMPARISON_P (x
))
22367 output_operand_lossage ("invalid operand for code '%c'", code
);
22371 fputs (arm_condition_codes
[get_arm_condition_code (x
)],
22376 /* CONST_TRUE_RTX means not always -- i.e. never. We shouldn't ever
22377 want to do that. */
22378 if (x
== const_true_rtx
)
22380 output_operand_lossage ("instruction never executed");
22383 if (!COMPARISON_P (x
))
22385 output_operand_lossage ("invalid operand for code '%c'", code
);
22389 fputs (arm_condition_codes
[ARM_INVERSE_CONDITION_CODE
22390 (get_arm_condition_code (x
))],
22400 /* Former Maverick support, removed after GCC-4.7. */
22401 output_operand_lossage ("obsolete Maverick format code '%c'", code
);
22406 || REGNO (x
) < FIRST_IWMMXT_GR_REGNUM
22407 || REGNO (x
) > LAST_IWMMXT_GR_REGNUM
)
22408 /* Bad value for wCG register number. */
22410 output_operand_lossage ("invalid operand for code '%c'", code
);
22415 fprintf (stream
, "%d", REGNO (x
) - FIRST_IWMMXT_GR_REGNUM
);
22418 /* Print an iWMMXt control register name. */
22420 if (!CONST_INT_P (x
)
22422 || INTVAL (x
) >= 16)
22423 /* Bad value for wC register number. */
22425 output_operand_lossage ("invalid operand for code '%c'", code
);
22431 static const char * wc_reg_names
[16] =
22433 "wCID", "wCon", "wCSSF", "wCASF",
22434 "wC4", "wC5", "wC6", "wC7",
22435 "wCGR0", "wCGR1", "wCGR2", "wCGR3",
22436 "wC12", "wC13", "wC14", "wC15"
22439 fputs (wc_reg_names
[INTVAL (x
)], stream
);
22443 /* Print the high single-precision register of a VFP double-precision
22447 machine_mode mode
= GET_MODE (x
);
22450 if (GET_MODE_SIZE (mode
) != 8 || !REG_P (x
))
22452 output_operand_lossage ("invalid operand for code '%c'", code
);
22457 if (!VFP_REGNO_OK_FOR_DOUBLE (regno
))
22459 output_operand_lossage ("invalid operand for code '%c'", code
);
22463 fprintf (stream
, "s%d", regno
- FIRST_VFP_REGNUM
+ 1);
22467 /* Print a VFP/Neon double precision or quad precision register name. */
22471 machine_mode mode
= GET_MODE (x
);
22472 int is_quad
= (code
== 'q');
22475 if (GET_MODE_SIZE (mode
) != (is_quad
? 16 : 8))
22477 output_operand_lossage ("invalid operand for code '%c'", code
);
22482 || !IS_VFP_REGNUM (REGNO (x
)))
22484 output_operand_lossage ("invalid operand for code '%c'", code
);
22489 if ((is_quad
&& !NEON_REGNO_OK_FOR_QUAD (regno
))
22490 || (!is_quad
&& !VFP_REGNO_OK_FOR_DOUBLE (regno
)))
22492 output_operand_lossage ("invalid operand for code '%c'", code
);
22496 fprintf (stream
, "%c%d", is_quad
? 'q' : 'd',
22497 (regno
- FIRST_VFP_REGNUM
) >> (is_quad
? 2 : 1));
22501 /* These two codes print the low/high doubleword register of a Neon quad
22502 register, respectively. For pair-structure types, can also print
22503 low/high quadword registers. */
22507 machine_mode mode
= GET_MODE (x
);
22510 if ((GET_MODE_SIZE (mode
) != 16
22511 && GET_MODE_SIZE (mode
) != 32) || !REG_P (x
))
22513 output_operand_lossage ("invalid operand for code '%c'", code
);
22518 if (!NEON_REGNO_OK_FOR_QUAD (regno
))
22520 output_operand_lossage ("invalid operand for code '%c'", code
);
22524 if (GET_MODE_SIZE (mode
) == 16)
22525 fprintf (stream
, "d%d", ((regno
- FIRST_VFP_REGNUM
) >> 1)
22526 + (code
== 'f' ? 1 : 0));
22528 fprintf (stream
, "q%d", ((regno
- FIRST_VFP_REGNUM
) >> 2)
22529 + (code
== 'f' ? 1 : 0));
22533 /* Print a VFPv3 floating-point constant, represented as an integer
22537 int index
= vfp3_const_double_index (x
);
22538 gcc_assert (index
!= -1);
22539 fprintf (stream
, "%d", index
);
22543 /* Print bits representing opcode features for Neon.
22545 Bit 0 is 1 for signed, 0 for unsigned. Floats count as signed
22546 and polynomials as unsigned.
22548 Bit 1 is 1 for floats and polynomials, 0 for ordinary integers.
22550 Bit 2 is 1 for rounding functions, 0 otherwise. */
22552 /* Identify the type as 's', 'u', 'p' or 'f'. */
22555 HOST_WIDE_INT bits
= INTVAL (x
);
22556 fputc ("uspf"[bits
& 3], stream
);
22560 /* Likewise, but signed and unsigned integers are both 'i'. */
22563 HOST_WIDE_INT bits
= INTVAL (x
);
22564 fputc ("iipf"[bits
& 3], stream
);
22568 /* As for 'T', but emit 'u' instead of 'p'. */
22571 HOST_WIDE_INT bits
= INTVAL (x
);
22572 fputc ("usuf"[bits
& 3], stream
);
22576 /* Bit 2: rounding (vs none). */
22579 HOST_WIDE_INT bits
= INTVAL (x
);
22580 fputs ((bits
& 4) != 0 ? "r" : "", stream
);
22584 /* Memory operand for vld1/vst1 instruction. */
22588 bool postinc
= FALSE
;
22589 rtx postinc_reg
= NULL
;
22590 unsigned align
, memsize
, align_bits
;
22592 gcc_assert (MEM_P (x
));
22593 addr
= XEXP (x
, 0);
22594 if (GET_CODE (addr
) == POST_INC
)
22597 addr
= XEXP (addr
, 0);
22599 if (GET_CODE (addr
) == POST_MODIFY
)
22601 postinc_reg
= XEXP( XEXP (addr
, 1), 1);
22602 addr
= XEXP (addr
, 0);
22604 asm_fprintf (stream
, "[%r", REGNO (addr
));
22606 /* We know the alignment of this access, so we can emit a hint in the
22607 instruction (for some alignments) as an aid to the memory subsystem
22609 align
= MEM_ALIGN (x
) >> 3;
22610 memsize
= MEM_SIZE (x
);
22612 /* Only certain alignment specifiers are supported by the hardware. */
22613 if (memsize
== 32 && (align
% 32) == 0)
22615 else if ((memsize
== 16 || memsize
== 32) && (align
% 16) == 0)
22617 else if (memsize
>= 8 && (align
% 8) == 0)
22622 if (align_bits
!= 0)
22623 asm_fprintf (stream
, ":%d", align_bits
);
22625 asm_fprintf (stream
, "]");
22628 fputs("!", stream
);
22630 asm_fprintf (stream
, ", %r", REGNO (postinc_reg
));
22638 gcc_assert (MEM_P (x
));
22639 addr
= XEXP (x
, 0);
22640 gcc_assert (REG_P (addr
));
22641 asm_fprintf (stream
, "[%r]", REGNO (addr
));
22645 /* Translate an S register number into a D register number and element index. */
22648 machine_mode mode
= GET_MODE (x
);
22651 if (GET_MODE_SIZE (mode
) != 4 || !REG_P (x
))
22653 output_operand_lossage ("invalid operand for code '%c'", code
);
22658 if (!VFP_REGNO_OK_FOR_SINGLE (regno
))
22660 output_operand_lossage ("invalid operand for code '%c'", code
);
22664 regno
= regno
- FIRST_VFP_REGNUM
;
22665 fprintf (stream
, "d%d[%d]", regno
/ 2, regno
% 2);
22670 gcc_assert (CONST_DOUBLE_P (x
));
22672 result
= vfp3_const_double_for_fract_bits (x
);
22674 result
= vfp3_const_double_for_bits (x
);
22675 fprintf (stream
, "#%d", result
);
22678 /* Register specifier for vld1.16/vst1.16. Translate the S register
22679 number into a D register number and element index. */
22682 machine_mode mode
= GET_MODE (x
);
22685 if (GET_MODE_SIZE (mode
) != 2 || !REG_P (x
))
22687 output_operand_lossage ("invalid operand for code '%c'", code
);
22692 if (!VFP_REGNO_OK_FOR_SINGLE (regno
))
22694 output_operand_lossage ("invalid operand for code '%c'", code
);
22698 regno
= regno
- FIRST_VFP_REGNUM
;
22699 fprintf (stream
, "d%d[%d]", regno
/2, ((regno
% 2) ? 2 : 0));
22706 output_operand_lossage ("missing operand");
22710 switch (GET_CODE (x
))
22713 asm_fprintf (stream
, "%r", REGNO (x
));
22717 output_address (GET_MODE (x
), XEXP (x
, 0));
22723 real_to_decimal (fpstr
, CONST_DOUBLE_REAL_VALUE (x
),
22724 sizeof (fpstr
), 0, 1);
22725 fprintf (stream
, "#%s", fpstr
);
22730 gcc_assert (GET_CODE (x
) != NEG
);
22731 fputc ('#', stream
);
22732 if (GET_CODE (x
) == HIGH
)
22734 fputs (":lower16:", stream
);
22738 output_addr_const (stream
, x
);
22744 /* Target hook for printing a memory address. */
22746 arm_print_operand_address (FILE *stream
, machine_mode mode
, rtx x
)
22750 int is_minus
= GET_CODE (x
) == MINUS
;
22753 asm_fprintf (stream
, "[%r]", REGNO (x
));
22754 else if (GET_CODE (x
) == PLUS
|| is_minus
)
22756 rtx base
= XEXP (x
, 0);
22757 rtx index
= XEXP (x
, 1);
22758 HOST_WIDE_INT offset
= 0;
22760 || (REG_P (index
) && REGNO (index
) == SP_REGNUM
))
22762 /* Ensure that BASE is a register. */
22763 /* (one of them must be). */
22764 /* Also ensure the SP is not used as in index register. */
22765 std::swap (base
, index
);
22767 switch (GET_CODE (index
))
22770 offset
= INTVAL (index
);
22773 asm_fprintf (stream
, "[%r, #%wd]",
22774 REGNO (base
), offset
);
22778 asm_fprintf (stream
, "[%r, %s%r]",
22779 REGNO (base
), is_minus
? "-" : "",
22789 asm_fprintf (stream
, "[%r, %s%r",
22790 REGNO (base
), is_minus
? "-" : "",
22791 REGNO (XEXP (index
, 0)));
22792 arm_print_operand (stream
, index
, 'S');
22793 fputs ("]", stream
);
22798 gcc_unreachable ();
22801 else if (GET_CODE (x
) == PRE_INC
|| GET_CODE (x
) == POST_INC
22802 || GET_CODE (x
) == PRE_DEC
|| GET_CODE (x
) == POST_DEC
)
22804 gcc_assert (REG_P (XEXP (x
, 0)));
22806 if (GET_CODE (x
) == PRE_DEC
|| GET_CODE (x
) == PRE_INC
)
22807 asm_fprintf (stream
, "[%r, #%s%d]!",
22808 REGNO (XEXP (x
, 0)),
22809 GET_CODE (x
) == PRE_DEC
? "-" : "",
22810 GET_MODE_SIZE (mode
));
22812 asm_fprintf (stream
, "[%r], #%s%d",
22813 REGNO (XEXP (x
, 0)),
22814 GET_CODE (x
) == POST_DEC
? "-" : "",
22815 GET_MODE_SIZE (mode
));
22817 else if (GET_CODE (x
) == PRE_MODIFY
)
22819 asm_fprintf (stream
, "[%r, ", REGNO (XEXP (x
, 0)));
22820 if (CONST_INT_P (XEXP (XEXP (x
, 1), 1)))
22821 asm_fprintf (stream
, "#%wd]!",
22822 INTVAL (XEXP (XEXP (x
, 1), 1)));
22824 asm_fprintf (stream
, "%r]!",
22825 REGNO (XEXP (XEXP (x
, 1), 1)));
22827 else if (GET_CODE (x
) == POST_MODIFY
)
22829 asm_fprintf (stream
, "[%r], ", REGNO (XEXP (x
, 0)));
22830 if (CONST_INT_P (XEXP (XEXP (x
, 1), 1)))
22831 asm_fprintf (stream
, "#%wd",
22832 INTVAL (XEXP (XEXP (x
, 1), 1)));
22834 asm_fprintf (stream
, "%r",
22835 REGNO (XEXP (XEXP (x
, 1), 1)));
22837 else output_addr_const (stream
, x
);
22842 asm_fprintf (stream
, "[%r]", REGNO (x
));
22843 else if (GET_CODE (x
) == POST_INC
)
22844 asm_fprintf (stream
, "%r!", REGNO (XEXP (x
, 0)));
22845 else if (GET_CODE (x
) == PLUS
)
22847 gcc_assert (REG_P (XEXP (x
, 0)));
22848 if (CONST_INT_P (XEXP (x
, 1)))
22849 asm_fprintf (stream
, "[%r, #%wd]",
22850 REGNO (XEXP (x
, 0)),
22851 INTVAL (XEXP (x
, 1)));
22853 asm_fprintf (stream
, "[%r, %r]",
22854 REGNO (XEXP (x
, 0)),
22855 REGNO (XEXP (x
, 1)));
22858 output_addr_const (stream
, x
);
22862 /* Target hook for indicating whether a punctuation character for
22863 TARGET_PRINT_OPERAND is valid. */
22865 arm_print_operand_punct_valid_p (unsigned char code
)
22867 return (code
== '@' || code
== '|' || code
== '.'
22868 || code
== '(' || code
== ')' || code
== '#'
22869 || (TARGET_32BIT
&& (code
== '?'))
22870 || (TARGET_THUMB2
&& (code
== '!'))
22871 || (TARGET_THUMB
&& (code
== '_')));
22874 /* Target hook for assembling integer objects. The ARM version needs to
22875 handle word-sized values specially. */
22877 arm_assemble_integer (rtx x
, unsigned int size
, int aligned_p
)
22881 if (size
== UNITS_PER_WORD
&& aligned_p
)
22883 fputs ("\t.word\t", asm_out_file
);
22884 output_addr_const (asm_out_file
, x
);
22886 /* Mark symbols as position independent. We only do this in the
22887 .text segment, not in the .data segment. */
22888 if (NEED_GOT_RELOC
&& flag_pic
&& making_const_table
&&
22889 (GET_CODE (x
) == SYMBOL_REF
|| GET_CODE (x
) == LABEL_REF
))
22891 /* See legitimize_pic_address for an explanation of the
22892 TARGET_VXWORKS_RTP check. */
22893 /* References to weak symbols cannot be resolved locally:
22894 they may be overridden by a non-weak definition at link
22896 if (!arm_pic_data_is_text_relative
22897 || (GET_CODE (x
) == SYMBOL_REF
22898 && (!SYMBOL_REF_LOCAL_P (x
)
22899 || (SYMBOL_REF_DECL (x
)
22900 ? DECL_WEAK (SYMBOL_REF_DECL (x
)) : 0))))
22901 fputs ("(GOT)", asm_out_file
);
22903 fputs ("(GOTOFF)", asm_out_file
);
22905 fputc ('\n', asm_out_file
);
22909 mode
= GET_MODE (x
);
22911 if (arm_vector_mode_supported_p (mode
))
22915 gcc_assert (GET_CODE (x
) == CONST_VECTOR
);
22917 units
= CONST_VECTOR_NUNITS (x
);
22918 size
= GET_MODE_UNIT_SIZE (mode
);
22920 if (GET_MODE_CLASS (mode
) == MODE_VECTOR_INT
)
22921 for (i
= 0; i
< units
; i
++)
22923 rtx elt
= CONST_VECTOR_ELT (x
, i
);
22925 (elt
, size
, i
== 0 ? BIGGEST_ALIGNMENT
: size
* BITS_PER_UNIT
, 1);
22928 for (i
= 0; i
< units
; i
++)
22930 rtx elt
= CONST_VECTOR_ELT (x
, i
);
22932 (*CONST_DOUBLE_REAL_VALUE (elt
),
22933 as_a
<scalar_float_mode
> (GET_MODE_INNER (mode
)),
22934 i
== 0 ? BIGGEST_ALIGNMENT
: size
* BITS_PER_UNIT
);
22940 return default_assemble_integer (x
, size
, aligned_p
);
22944 arm_elf_asm_cdtor (rtx symbol
, int priority
, bool is_ctor
)
22948 if (!TARGET_AAPCS_BASED
)
22951 default_named_section_asm_out_constructor
22952 : default_named_section_asm_out_destructor
) (symbol
, priority
);
22956 /* Put these in the .init_array section, using a special relocation. */
22957 if (priority
!= DEFAULT_INIT_PRIORITY
)
22960 sprintf (buf
, "%s.%.5u",
22961 is_ctor
? ".init_array" : ".fini_array",
22963 s
= get_section (buf
, SECTION_WRITE
| SECTION_NOTYPE
, NULL_TREE
);
22970 switch_to_section (s
);
22971 assemble_align (POINTER_SIZE
);
22972 fputs ("\t.word\t", asm_out_file
);
22973 output_addr_const (asm_out_file
, symbol
);
22974 fputs ("(target1)\n", asm_out_file
);
22977 /* Add a function to the list of static constructors. */
22980 arm_elf_asm_constructor (rtx symbol
, int priority
)
22982 arm_elf_asm_cdtor (symbol
, priority
, /*is_ctor=*/true);
22985 /* Add a function to the list of static destructors. */
22988 arm_elf_asm_destructor (rtx symbol
, int priority
)
22990 arm_elf_asm_cdtor (symbol
, priority
, /*is_ctor=*/false);
22993 /* A finite state machine takes care of noticing whether or not instructions
22994 can be conditionally executed, and thus decrease execution time and code
22995 size by deleting branch instructions. The fsm is controlled by
22996 final_prescan_insn, and controls the actions of ASM_OUTPUT_OPCODE. */
22998 /* The state of the fsm controlling condition codes are:
22999 0: normal, do nothing special
23000 1: make ASM_OUTPUT_OPCODE not output this instruction
23001 2: make ASM_OUTPUT_OPCODE not output this instruction
23002 3: make instructions conditional
23003 4: make instructions conditional
23005 State transitions (state->state by whom under condition):
23006 0 -> 1 final_prescan_insn if the `target' is a label
23007 0 -> 2 final_prescan_insn if the `target' is an unconditional branch
23008 1 -> 3 ASM_OUTPUT_OPCODE after not having output the conditional branch
23009 2 -> 4 ASM_OUTPUT_OPCODE after not having output the conditional branch
23010 3 -> 0 (*targetm.asm_out.internal_label) if the `target' label is reached
23011 (the target label has CODE_LABEL_NUMBER equal to arm_target_label).
23012 4 -> 0 final_prescan_insn if the `target' unconditional branch is reached
23013 (the target insn is arm_target_insn).
23015 If the jump clobbers the conditions then we use states 2 and 4.
23017 A similar thing can be done with conditional return insns.
23019 XXX In case the `target' is an unconditional branch, this conditionalising
23020 of the instructions always reduces code size, but not always execution
23021 time. But then, I want to reduce the code size to somewhere near what
23022 /bin/cc produces. */
23024 /* In addition to this, state is maintained for Thumb-2 COND_EXEC
23025 instructions. When a COND_EXEC instruction is seen the subsequent
23026 instructions are scanned so that multiple conditional instructions can be
23027 combined into a single IT block. arm_condexec_count and arm_condexec_mask
23028 specify the length and true/false mask for the IT block. These will be
23029 decremented/zeroed by arm_asm_output_opcode as the insns are output. */
23031 /* Returns the index of the ARM condition code string in
23032 `arm_condition_codes', or ARM_NV if the comparison is invalid.
23033 COMPARISON should be an rtx like `(eq (...) (...))'. */
23036 maybe_get_arm_condition_code (rtx comparison
)
23038 machine_mode mode
= GET_MODE (XEXP (comparison
, 0));
23039 enum arm_cond_code code
;
23040 enum rtx_code comp_code
= GET_CODE (comparison
);
23042 if (GET_MODE_CLASS (mode
) != MODE_CC
)
23043 mode
= SELECT_CC_MODE (comp_code
, XEXP (comparison
, 0),
23044 XEXP (comparison
, 1));
23048 case E_CC_DNEmode
: code
= ARM_NE
; goto dominance
;
23049 case E_CC_DEQmode
: code
= ARM_EQ
; goto dominance
;
23050 case E_CC_DGEmode
: code
= ARM_GE
; goto dominance
;
23051 case E_CC_DGTmode
: code
= ARM_GT
; goto dominance
;
23052 case E_CC_DLEmode
: code
= ARM_LE
; goto dominance
;
23053 case E_CC_DLTmode
: code
= ARM_LT
; goto dominance
;
23054 case E_CC_DGEUmode
: code
= ARM_CS
; goto dominance
;
23055 case E_CC_DGTUmode
: code
= ARM_HI
; goto dominance
;
23056 case E_CC_DLEUmode
: code
= ARM_LS
; goto dominance
;
23057 case E_CC_DLTUmode
: code
= ARM_CC
;
23060 if (comp_code
== EQ
)
23061 return ARM_INVERSE_CONDITION_CODE (code
);
23062 if (comp_code
== NE
)
23066 case E_CC_NOOVmode
:
23069 case NE
: return ARM_NE
;
23070 case EQ
: return ARM_EQ
;
23071 case GE
: return ARM_PL
;
23072 case LT
: return ARM_MI
;
23073 default: return ARM_NV
;
23079 case NE
: return ARM_NE
;
23080 case EQ
: return ARM_EQ
;
23081 default: return ARM_NV
;
23087 case NE
: return ARM_MI
;
23088 case EQ
: return ARM_PL
;
23089 default: return ARM_NV
;
23094 /* We can handle all cases except UNEQ and LTGT. */
23097 case GE
: return ARM_GE
;
23098 case GT
: return ARM_GT
;
23099 case LE
: return ARM_LS
;
23100 case LT
: return ARM_MI
;
23101 case NE
: return ARM_NE
;
23102 case EQ
: return ARM_EQ
;
23103 case ORDERED
: return ARM_VC
;
23104 case UNORDERED
: return ARM_VS
;
23105 case UNLT
: return ARM_LT
;
23106 case UNLE
: return ARM_LE
;
23107 case UNGT
: return ARM_HI
;
23108 case UNGE
: return ARM_PL
;
23109 /* UNEQ and LTGT do not have a representation. */
23110 case UNEQ
: /* Fall through. */
23111 case LTGT
: /* Fall through. */
23112 default: return ARM_NV
;
23118 case NE
: return ARM_NE
;
23119 case EQ
: return ARM_EQ
;
23120 case GE
: return ARM_LE
;
23121 case GT
: return ARM_LT
;
23122 case LE
: return ARM_GE
;
23123 case LT
: return ARM_GT
;
23124 case GEU
: return ARM_LS
;
23125 case GTU
: return ARM_CC
;
23126 case LEU
: return ARM_CS
;
23127 case LTU
: return ARM_HI
;
23128 default: return ARM_NV
;
23134 case LTU
: return ARM_CS
;
23135 case GEU
: return ARM_CC
;
23136 case NE
: return ARM_CS
;
23137 case EQ
: return ARM_CC
;
23138 default: return ARM_NV
;
23144 case NE
: return ARM_NE
;
23145 case EQ
: return ARM_EQ
;
23146 case GEU
: return ARM_CS
;
23147 case GTU
: return ARM_HI
;
23148 case LEU
: return ARM_LS
;
23149 case LTU
: return ARM_CC
;
23150 default: return ARM_NV
;
23156 case GE
: return ARM_GE
;
23157 case LT
: return ARM_LT
;
23158 case GEU
: return ARM_CS
;
23159 case LTU
: return ARM_CC
;
23160 default: return ARM_NV
;
23166 case NE
: return ARM_VS
;
23167 case EQ
: return ARM_VC
;
23168 default: return ARM_NV
;
23174 case NE
: return ARM_NE
;
23175 case EQ
: return ARM_EQ
;
23176 case GE
: return ARM_GE
;
23177 case GT
: return ARM_GT
;
23178 case LE
: return ARM_LE
;
23179 case LT
: return ARM_LT
;
23180 case GEU
: return ARM_CS
;
23181 case GTU
: return ARM_HI
;
23182 case LEU
: return ARM_LS
;
23183 case LTU
: return ARM_CC
;
23184 default: return ARM_NV
;
23187 default: gcc_unreachable ();
23191 /* Like maybe_get_arm_condition_code, but never return ARM_NV. */
23192 static enum arm_cond_code
23193 get_arm_condition_code (rtx comparison
)
23195 enum arm_cond_code code
= maybe_get_arm_condition_code (comparison
);
23196 gcc_assert (code
!= ARM_NV
);
23200 /* Implement TARGET_FIXED_CONDITION_CODE_REGS. We only have condition
23201 code registers when not targetting Thumb1. The VFP condition register
23202 only exists when generating hard-float code. */
23204 arm_fixed_condition_code_regs (unsigned int *p1
, unsigned int *p2
)
23210 *p2
= TARGET_HARD_FLOAT
? VFPCC_REGNUM
: INVALID_REGNUM
;
23214 /* Tell arm_asm_output_opcode to output IT blocks for conditionally executed
23217 thumb2_final_prescan_insn (rtx_insn
*insn
)
23219 rtx_insn
*first_insn
= insn
;
23220 rtx body
= PATTERN (insn
);
23222 enum arm_cond_code code
;
23227 /* max_insns_skipped in the tune was already taken into account in the
23228 cost model of ifcvt pass when generating COND_EXEC insns. At this stage
23229 just emit the IT blocks as we can. It does not make sense to split
23231 max
= MAX_INSN_PER_IT_BLOCK
;
23233 /* Remove the previous insn from the count of insns to be output. */
23234 if (arm_condexec_count
)
23235 arm_condexec_count
--;
23237 /* Nothing to do if we are already inside a conditional block. */
23238 if (arm_condexec_count
)
23241 if (GET_CODE (body
) != COND_EXEC
)
23244 /* Conditional jumps are implemented directly. */
23248 predicate
= COND_EXEC_TEST (body
);
23249 arm_current_cc
= get_arm_condition_code (predicate
);
23251 n
= get_attr_ce_count (insn
);
23252 arm_condexec_count
= 1;
23253 arm_condexec_mask
= (1 << n
) - 1;
23254 arm_condexec_masklen
= n
;
23255 /* See if subsequent instructions can be combined into the same block. */
23258 insn
= next_nonnote_insn (insn
);
23260 /* Jumping into the middle of an IT block is illegal, so a label or
23261 barrier terminates the block. */
23262 if (!NONJUMP_INSN_P (insn
) && !JUMP_P (insn
))
23265 body
= PATTERN (insn
);
23266 /* USE and CLOBBER aren't really insns, so just skip them. */
23267 if (GET_CODE (body
) == USE
23268 || GET_CODE (body
) == CLOBBER
)
23271 /* ??? Recognize conditional jumps, and combine them with IT blocks. */
23272 if (GET_CODE (body
) != COND_EXEC
)
23274 /* Maximum number of conditionally executed instructions in a block. */
23275 n
= get_attr_ce_count (insn
);
23276 if (arm_condexec_masklen
+ n
> max
)
23279 predicate
= COND_EXEC_TEST (body
);
23280 code
= get_arm_condition_code (predicate
);
23281 mask
= (1 << n
) - 1;
23282 if (arm_current_cc
== code
)
23283 arm_condexec_mask
|= (mask
<< arm_condexec_masklen
);
23284 else if (arm_current_cc
!= ARM_INVERSE_CONDITION_CODE(code
))
23287 arm_condexec_count
++;
23288 arm_condexec_masklen
+= n
;
23290 /* A jump must be the last instruction in a conditional block. */
23294 /* Restore recog_data (getting the attributes of other insns can
23295 destroy this array, but final.c assumes that it remains intact
23296 across this call). */
23297 extract_constrain_insn_cached (first_insn
);
23301 arm_final_prescan_insn (rtx_insn
*insn
)
23303 /* BODY will hold the body of INSN. */
23304 rtx body
= PATTERN (insn
);
23306 /* This will be 1 if trying to repeat the trick, and things need to be
23307 reversed if it appears to fail. */
23310 /* If we start with a return insn, we only succeed if we find another one. */
23311 int seeking_return
= 0;
23312 enum rtx_code return_code
= UNKNOWN
;
23314 /* START_INSN will hold the insn from where we start looking. This is the
23315 first insn after the following code_label if REVERSE is true. */
23316 rtx_insn
*start_insn
= insn
;
23318 /* If in state 4, check if the target branch is reached, in order to
23319 change back to state 0. */
23320 if (arm_ccfsm_state
== 4)
23322 if (insn
== arm_target_insn
)
23324 arm_target_insn
= NULL
;
23325 arm_ccfsm_state
= 0;
23330 /* If in state 3, it is possible to repeat the trick, if this insn is an
23331 unconditional branch to a label, and immediately following this branch
23332 is the previous target label which is only used once, and the label this
23333 branch jumps to is not too far off. */
23334 if (arm_ccfsm_state
== 3)
23336 if (simplejump_p (insn
))
23338 start_insn
= next_nonnote_insn (start_insn
);
23339 if (BARRIER_P (start_insn
))
23341 /* XXX Isn't this always a barrier? */
23342 start_insn
= next_nonnote_insn (start_insn
);
23344 if (LABEL_P (start_insn
)
23345 && CODE_LABEL_NUMBER (start_insn
) == arm_target_label
23346 && LABEL_NUSES (start_insn
) == 1)
23351 else if (ANY_RETURN_P (body
))
23353 start_insn
= next_nonnote_insn (start_insn
);
23354 if (BARRIER_P (start_insn
))
23355 start_insn
= next_nonnote_insn (start_insn
);
23356 if (LABEL_P (start_insn
)
23357 && CODE_LABEL_NUMBER (start_insn
) == arm_target_label
23358 && LABEL_NUSES (start_insn
) == 1)
23361 seeking_return
= 1;
23362 return_code
= GET_CODE (body
);
23371 gcc_assert (!arm_ccfsm_state
|| reverse
);
23372 if (!JUMP_P (insn
))
23375 /* This jump might be paralleled with a clobber of the condition codes
23376 the jump should always come first */
23377 if (GET_CODE (body
) == PARALLEL
&& XVECLEN (body
, 0) > 0)
23378 body
= XVECEXP (body
, 0, 0);
23381 || (GET_CODE (body
) == SET
&& GET_CODE (SET_DEST (body
)) == PC
23382 && GET_CODE (SET_SRC (body
)) == IF_THEN_ELSE
))
23385 int fail
= FALSE
, succeed
= FALSE
;
23386 /* Flag which part of the IF_THEN_ELSE is the LABEL_REF. */
23387 int then_not_else
= TRUE
;
23388 rtx_insn
*this_insn
= start_insn
;
23391 /* Register the insn jumped to. */
23394 if (!seeking_return
)
23395 label
= XEXP (SET_SRC (body
), 0);
23397 else if (GET_CODE (XEXP (SET_SRC (body
), 1)) == LABEL_REF
)
23398 label
= XEXP (XEXP (SET_SRC (body
), 1), 0);
23399 else if (GET_CODE (XEXP (SET_SRC (body
), 2)) == LABEL_REF
)
23401 label
= XEXP (XEXP (SET_SRC (body
), 2), 0);
23402 then_not_else
= FALSE
;
23404 else if (ANY_RETURN_P (XEXP (SET_SRC (body
), 1)))
23406 seeking_return
= 1;
23407 return_code
= GET_CODE (XEXP (SET_SRC (body
), 1));
23409 else if (ANY_RETURN_P (XEXP (SET_SRC (body
), 2)))
23411 seeking_return
= 1;
23412 return_code
= GET_CODE (XEXP (SET_SRC (body
), 2));
23413 then_not_else
= FALSE
;
23416 gcc_unreachable ();
23418 /* See how many insns this branch skips, and what kind of insns. If all
23419 insns are okay, and the label or unconditional branch to the same
23420 label is not too far away, succeed. */
23421 for (insns_skipped
= 0;
23422 !fail
&& !succeed
&& insns_skipped
++ < max_insns_skipped
;)
23426 this_insn
= next_nonnote_insn (this_insn
);
23430 switch (GET_CODE (this_insn
))
23433 /* Succeed if it is the target label, otherwise fail since
23434 control falls in from somewhere else. */
23435 if (this_insn
== label
)
23437 arm_ccfsm_state
= 1;
23445 /* Succeed if the following insn is the target label.
23447 If return insns are used then the last insn in a function
23448 will be a barrier. */
23449 this_insn
= next_nonnote_insn (this_insn
);
23450 if (this_insn
&& this_insn
== label
)
23452 arm_ccfsm_state
= 1;
23460 /* The AAPCS says that conditional calls should not be
23461 used since they make interworking inefficient (the
23462 linker can't transform BL<cond> into BLX). That's
23463 only a problem if the machine has BLX. */
23470 /* Succeed if the following insn is the target label, or
23471 if the following two insns are a barrier and the
23473 this_insn
= next_nonnote_insn (this_insn
);
23474 if (this_insn
&& BARRIER_P (this_insn
))
23475 this_insn
= next_nonnote_insn (this_insn
);
23477 if (this_insn
&& this_insn
== label
23478 && insns_skipped
< max_insns_skipped
)
23480 arm_ccfsm_state
= 1;
23488 /* If this is an unconditional branch to the same label, succeed.
23489 If it is to another label, do nothing. If it is conditional,
23491 /* XXX Probably, the tests for SET and the PC are
23494 scanbody
= PATTERN (this_insn
);
23495 if (GET_CODE (scanbody
) == SET
23496 && GET_CODE (SET_DEST (scanbody
)) == PC
)
23498 if (GET_CODE (SET_SRC (scanbody
)) == LABEL_REF
23499 && XEXP (SET_SRC (scanbody
), 0) == label
&& !reverse
)
23501 arm_ccfsm_state
= 2;
23504 else if (GET_CODE (SET_SRC (scanbody
)) == IF_THEN_ELSE
)
23507 /* Fail if a conditional return is undesirable (e.g. on a
23508 StrongARM), but still allow this if optimizing for size. */
23509 else if (GET_CODE (scanbody
) == return_code
23510 && !use_return_insn (TRUE
, NULL
)
23513 else if (GET_CODE (scanbody
) == return_code
)
23515 arm_ccfsm_state
= 2;
23518 else if (GET_CODE (scanbody
) == PARALLEL
)
23520 switch (get_attr_conds (this_insn
))
23530 fail
= TRUE
; /* Unrecognized jump (e.g. epilogue). */
23535 /* Instructions using or affecting the condition codes make it
23537 scanbody
= PATTERN (this_insn
);
23538 if (!(GET_CODE (scanbody
) == SET
23539 || GET_CODE (scanbody
) == PARALLEL
)
23540 || get_attr_conds (this_insn
) != CONDS_NOCOND
)
23550 if ((!seeking_return
) && (arm_ccfsm_state
== 1 || reverse
))
23551 arm_target_label
= CODE_LABEL_NUMBER (label
);
23554 gcc_assert (seeking_return
|| arm_ccfsm_state
== 2);
23556 while (this_insn
&& GET_CODE (PATTERN (this_insn
)) == USE
)
23558 this_insn
= next_nonnote_insn (this_insn
);
23559 gcc_assert (!this_insn
23560 || (!BARRIER_P (this_insn
)
23561 && !LABEL_P (this_insn
)));
23565 /* Oh, dear! we ran off the end.. give up. */
23566 extract_constrain_insn_cached (insn
);
23567 arm_ccfsm_state
= 0;
23568 arm_target_insn
= NULL
;
23571 arm_target_insn
= this_insn
;
23574 /* If REVERSE is true, ARM_CURRENT_CC needs to be inverted from
23577 arm_current_cc
= get_arm_condition_code (XEXP (SET_SRC (body
), 0));
23579 if (reverse
|| then_not_else
)
23580 arm_current_cc
= ARM_INVERSE_CONDITION_CODE (arm_current_cc
);
23583 /* Restore recog_data (getting the attributes of other insns can
23584 destroy this array, but final.c assumes that it remains intact
23585 across this call. */
23586 extract_constrain_insn_cached (insn
);
23590 /* Output IT instructions. */
23592 thumb2_asm_output_opcode (FILE * stream
)
23597 if (arm_condexec_mask
)
23599 for (n
= 0; n
< arm_condexec_masklen
; n
++)
23600 buff
[n
] = (arm_condexec_mask
& (1 << n
)) ? 't' : 'e';
23602 asm_fprintf(stream
, "i%s\t%s\n\t", buff
,
23603 arm_condition_codes
[arm_current_cc
]);
23604 arm_condexec_mask
= 0;
23608 /* Implement TARGET_HARD_REGNO_NREGS. On the ARM core regs are
23609 UNITS_PER_WORD bytes wide. */
23610 static unsigned int
23611 arm_hard_regno_nregs (unsigned int regno
, machine_mode mode
)
23614 && regno
> PC_REGNUM
23615 && regno
!= FRAME_POINTER_REGNUM
23616 && regno
!= ARG_POINTER_REGNUM
23617 && !IS_VFP_REGNUM (regno
))
23620 return ARM_NUM_REGS (mode
);
23623 /* Implement TARGET_HARD_REGNO_MODE_OK. */
23625 arm_hard_regno_mode_ok (unsigned int regno
, machine_mode mode
)
23627 if (GET_MODE_CLASS (mode
) == MODE_CC
)
23628 return (regno
== CC_REGNUM
23629 || (TARGET_HARD_FLOAT
23630 && regno
== VFPCC_REGNUM
));
23632 if (regno
== CC_REGNUM
&& GET_MODE_CLASS (mode
) != MODE_CC
)
23636 /* For the Thumb we only allow values bigger than SImode in
23637 registers 0 - 6, so that there is always a second low
23638 register available to hold the upper part of the value.
23639 We probably we ought to ensure that the register is the
23640 start of an even numbered register pair. */
23641 return (ARM_NUM_REGS (mode
) < 2) || (regno
< LAST_LO_REGNUM
);
23643 if (TARGET_HARD_FLOAT
&& IS_VFP_REGNUM (regno
))
23645 if (mode
== SFmode
|| mode
== SImode
)
23646 return VFP_REGNO_OK_FOR_SINGLE (regno
);
23648 if (mode
== DFmode
)
23649 return VFP_REGNO_OK_FOR_DOUBLE (regno
);
23651 if (mode
== HFmode
)
23652 return VFP_REGNO_OK_FOR_SINGLE (regno
);
23654 /* VFP registers can hold HImode values. */
23655 if (mode
== HImode
)
23656 return VFP_REGNO_OK_FOR_SINGLE (regno
);
23659 return (VALID_NEON_DREG_MODE (mode
) && VFP_REGNO_OK_FOR_DOUBLE (regno
))
23660 || (VALID_NEON_QREG_MODE (mode
)
23661 && NEON_REGNO_OK_FOR_QUAD (regno
))
23662 || (mode
== TImode
&& NEON_REGNO_OK_FOR_NREGS (regno
, 2))
23663 || (mode
== EImode
&& NEON_REGNO_OK_FOR_NREGS (regno
, 3))
23664 || (mode
== OImode
&& NEON_REGNO_OK_FOR_NREGS (regno
, 4))
23665 || (mode
== CImode
&& NEON_REGNO_OK_FOR_NREGS (regno
, 6))
23666 || (mode
== XImode
&& NEON_REGNO_OK_FOR_NREGS (regno
, 8));
23671 if (TARGET_REALLY_IWMMXT
)
23673 if (IS_IWMMXT_GR_REGNUM (regno
))
23674 return mode
== SImode
;
23676 if (IS_IWMMXT_REGNUM (regno
))
23677 return VALID_IWMMXT_REG_MODE (mode
);
23680 /* We allow almost any value to be stored in the general registers.
23681 Restrict doubleword quantities to even register pairs in ARM state
23682 so that we can use ldrd. Do not allow very large Neon structure
23683 opaque modes in general registers; they would use too many. */
23684 if (regno
<= LAST_ARM_REGNUM
)
23686 if (ARM_NUM_REGS (mode
) > 4)
23692 return !(TARGET_LDRD
&& GET_MODE_SIZE (mode
) > 4 && (regno
& 1) != 0);
23695 if (regno
== FRAME_POINTER_REGNUM
23696 || regno
== ARG_POINTER_REGNUM
)
23697 /* We only allow integers in the fake hard registers. */
23698 return GET_MODE_CLASS (mode
) == MODE_INT
;
23703 /* Implement TARGET_MODES_TIEABLE_P. */
23706 arm_modes_tieable_p (machine_mode mode1
, machine_mode mode2
)
23708 if (GET_MODE_CLASS (mode1
) == GET_MODE_CLASS (mode2
))
23711 /* We specifically want to allow elements of "structure" modes to
23712 be tieable to the structure. This more general condition allows
23713 other rarer situations too. */
23715 && (VALID_NEON_DREG_MODE (mode1
)
23716 || VALID_NEON_QREG_MODE (mode1
)
23717 || VALID_NEON_STRUCT_MODE (mode1
))
23718 && (VALID_NEON_DREG_MODE (mode2
)
23719 || VALID_NEON_QREG_MODE (mode2
)
23720 || VALID_NEON_STRUCT_MODE (mode2
)))
23726 /* For efficiency and historical reasons LO_REGS, HI_REGS and CC_REGS are
23727 not used in arm mode. */
23730 arm_regno_class (int regno
)
23732 if (regno
== PC_REGNUM
)
23737 if (regno
== STACK_POINTER_REGNUM
)
23739 if (regno
== CC_REGNUM
)
23746 if (TARGET_THUMB2
&& regno
< 8)
23749 if ( regno
<= LAST_ARM_REGNUM
23750 || regno
== FRAME_POINTER_REGNUM
23751 || regno
== ARG_POINTER_REGNUM
)
23752 return TARGET_THUMB2
? HI_REGS
: GENERAL_REGS
;
23754 if (regno
== CC_REGNUM
|| regno
== VFPCC_REGNUM
)
23755 return TARGET_THUMB2
? CC_REG
: NO_REGS
;
23757 if (IS_VFP_REGNUM (regno
))
23759 if (regno
<= D7_VFP_REGNUM
)
23760 return VFP_D0_D7_REGS
;
23761 else if (regno
<= LAST_LO_VFP_REGNUM
)
23762 return VFP_LO_REGS
;
23764 return VFP_HI_REGS
;
23767 if (IS_IWMMXT_REGNUM (regno
))
23768 return IWMMXT_REGS
;
23770 if (IS_IWMMXT_GR_REGNUM (regno
))
23771 return IWMMXT_GR_REGS
;
23776 /* Handle a special case when computing the offset
23777 of an argument from the frame pointer. */
23779 arm_debugger_arg_offset (int value
, rtx addr
)
23783 /* We are only interested if dbxout_parms() failed to compute the offset. */
23787 /* We can only cope with the case where the address is held in a register. */
23791 /* If we are using the frame pointer to point at the argument, then
23792 an offset of 0 is correct. */
23793 if (REGNO (addr
) == (unsigned) HARD_FRAME_POINTER_REGNUM
)
23796 /* If we are using the stack pointer to point at the
23797 argument, then an offset of 0 is correct. */
23798 /* ??? Check this is consistent with thumb2 frame layout. */
23799 if ((TARGET_THUMB
|| !frame_pointer_needed
)
23800 && REGNO (addr
) == SP_REGNUM
)
23803 /* Oh dear. The argument is pointed to by a register rather
23804 than being held in a register, or being stored at a known
23805 offset from the frame pointer. Since GDB only understands
23806 those two kinds of argument we must translate the address
23807 held in the register into an offset from the frame pointer.
23808 We do this by searching through the insns for the function
23809 looking to see where this register gets its value. If the
23810 register is initialized from the frame pointer plus an offset
23811 then we are in luck and we can continue, otherwise we give up.
23813 This code is exercised by producing debugging information
23814 for a function with arguments like this:
23816 double func (double a, double b, int c, double d) {return d;}
23818 Without this code the stab for parameter 'd' will be set to
23819 an offset of 0 from the frame pointer, rather than 8. */
23821 /* The if() statement says:
23823 If the insn is a normal instruction
23824 and if the insn is setting the value in a register
23825 and if the register being set is the register holding the address of the argument
23826 and if the address is computing by an addition
23827 that involves adding to a register
23828 which is the frame pointer
23833 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
23835 if ( NONJUMP_INSN_P (insn
)
23836 && GET_CODE (PATTERN (insn
)) == SET
23837 && REGNO (XEXP (PATTERN (insn
), 0)) == REGNO (addr
)
23838 && GET_CODE (XEXP (PATTERN (insn
), 1)) == PLUS
23839 && REG_P (XEXP (XEXP (PATTERN (insn
), 1), 0))
23840 && REGNO (XEXP (XEXP (PATTERN (insn
), 1), 0)) == (unsigned) HARD_FRAME_POINTER_REGNUM
23841 && CONST_INT_P (XEXP (XEXP (PATTERN (insn
), 1), 1))
23844 value
= INTVAL (XEXP (XEXP (PATTERN (insn
), 1), 1));
23853 warning (0, "unable to compute real location of stacked parameter");
23854 value
= 8; /* XXX magic hack */
23860 /* Implement TARGET_PROMOTED_TYPE. */
23863 arm_promoted_type (const_tree t
)
23865 if (SCALAR_FLOAT_TYPE_P (t
)
23866 && TYPE_PRECISION (t
) == 16
23867 && TYPE_MAIN_VARIANT (t
) == arm_fp16_type_node
)
23868 return float_type_node
;
23872 /* Implement TARGET_SCALAR_MODE_SUPPORTED_P.
23873 This simply adds HFmode as a supported mode; even though we don't
23874 implement arithmetic on this type directly, it's supported by
23875 optabs conversions, much the way the double-word arithmetic is
23876 special-cased in the default hook. */
23879 arm_scalar_mode_supported_p (scalar_mode mode
)
23881 if (mode
== HFmode
)
23882 return (arm_fp16_format
!= ARM_FP16_FORMAT_NONE
);
23883 else if (ALL_FIXED_POINT_MODE_P (mode
))
23886 return default_scalar_mode_supported_p (mode
);
23889 /* Set the value of FLT_EVAL_METHOD.
23890 ISO/IEC TS 18661-3 defines two values that we'd like to make use of:
23892 0: evaluate all operations and constants, whose semantic type has at
23893 most the range and precision of type float, to the range and
23894 precision of float; evaluate all other operations and constants to
23895 the range and precision of the semantic type;
23897 N, where _FloatN is a supported interchange floating type
23898 evaluate all operations and constants, whose semantic type has at
23899 most the range and precision of _FloatN type, to the range and
23900 precision of the _FloatN type; evaluate all other operations and
23901 constants to the range and precision of the semantic type;
23903 If we have the ARMv8.2-A extensions then we support _Float16 in native
23904 precision, so we should set this to 16. Otherwise, we support the type,
23905 but want to evaluate expressions in float precision, so set this to
23908 static enum flt_eval_method
23909 arm_excess_precision (enum excess_precision_type type
)
23913 case EXCESS_PRECISION_TYPE_FAST
:
23914 case EXCESS_PRECISION_TYPE_STANDARD
:
23915 /* We can calculate either in 16-bit range and precision or
23916 32-bit range and precision. Make that decision based on whether
23917 we have native support for the ARMv8.2-A 16-bit floating-point
23918 instructions or not. */
23919 return (TARGET_VFP_FP16INST
23920 ? FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16
23921 : FLT_EVAL_METHOD_PROMOTE_TO_FLOAT
);
23922 case EXCESS_PRECISION_TYPE_IMPLICIT
:
23923 return FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16
;
23925 gcc_unreachable ();
23927 return FLT_EVAL_METHOD_UNPREDICTABLE
;
23931 /* Implement TARGET_FLOATN_MODE. Make very sure that we don't provide
23932 _Float16 if we are using anything other than ieee format for 16-bit
23933 floating point. Otherwise, punt to the default implementation. */
23934 static opt_scalar_float_mode
23935 arm_floatn_mode (int n
, bool extended
)
23937 if (!extended
&& n
== 16)
23939 if (arm_fp16_format
== ARM_FP16_FORMAT_IEEE
)
23941 return opt_scalar_float_mode ();
23944 return default_floatn_mode (n
, extended
);
23948 /* Set up OPERANDS for a register copy from SRC to DEST, taking care
23949 not to early-clobber SRC registers in the process.
23951 We assume that the operands described by SRC and DEST represent a
23952 decomposed copy of OPERANDS[1] into OPERANDS[0]. COUNT is the
23953 number of components into which the copy has been decomposed. */
23955 neon_disambiguate_copy (rtx
*operands
, rtx
*dest
, rtx
*src
, unsigned int count
)
23959 if (!reg_overlap_mentioned_p (operands
[0], operands
[1])
23960 || REGNO (operands
[0]) < REGNO (operands
[1]))
23962 for (i
= 0; i
< count
; i
++)
23964 operands
[2 * i
] = dest
[i
];
23965 operands
[2 * i
+ 1] = src
[i
];
23970 for (i
= 0; i
< count
; i
++)
23972 operands
[2 * i
] = dest
[count
- i
- 1];
23973 operands
[2 * i
+ 1] = src
[count
- i
- 1];
23978 /* Split operands into moves from op[1] + op[2] into op[0]. */
23981 neon_split_vcombine (rtx operands
[3])
23983 unsigned int dest
= REGNO (operands
[0]);
23984 unsigned int src1
= REGNO (operands
[1]);
23985 unsigned int src2
= REGNO (operands
[2]);
23986 machine_mode halfmode
= GET_MODE (operands
[1]);
23987 unsigned int halfregs
= REG_NREGS (operands
[1]);
23988 rtx destlo
, desthi
;
23990 if (src1
== dest
&& src2
== dest
+ halfregs
)
23992 /* No-op move. Can't split to nothing; emit something. */
23993 emit_note (NOTE_INSN_DELETED
);
23997 /* Preserve register attributes for variable tracking. */
23998 destlo
= gen_rtx_REG_offset (operands
[0], halfmode
, dest
, 0);
23999 desthi
= gen_rtx_REG_offset (operands
[0], halfmode
, dest
+ halfregs
,
24000 GET_MODE_SIZE (halfmode
));
24002 /* Special case of reversed high/low parts. Use VSWP. */
24003 if (src2
== dest
&& src1
== dest
+ halfregs
)
24005 rtx x
= gen_rtx_SET (destlo
, operands
[1]);
24006 rtx y
= gen_rtx_SET (desthi
, operands
[2]);
24007 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, x
, y
)));
24011 if (!reg_overlap_mentioned_p (operands
[2], destlo
))
24013 /* Try to avoid unnecessary moves if part of the result
24014 is in the right place already. */
24016 emit_move_insn (destlo
, operands
[1]);
24017 if (src2
!= dest
+ halfregs
)
24018 emit_move_insn (desthi
, operands
[2]);
24022 if (src2
!= dest
+ halfregs
)
24023 emit_move_insn (desthi
, operands
[2]);
24025 emit_move_insn (destlo
, operands
[1]);
24029 /* Return the number (counting from 0) of
24030 the least significant set bit in MASK. */
24033 number_of_first_bit_set (unsigned mask
)
24035 return ctz_hwi (mask
);
24038 /* Like emit_multi_reg_push, but allowing for a different set of
24039 registers to be described as saved. MASK is the set of registers
24040 to be saved; REAL_REGS is the set of registers to be described as
24041 saved. If REAL_REGS is 0, only describe the stack adjustment. */
24044 thumb1_emit_multi_reg_push (unsigned long mask
, unsigned long real_regs
)
24046 unsigned long regno
;
24047 rtx par
[10], tmp
, reg
;
24051 /* Build the parallel of the registers actually being stored. */
24052 for (i
= 0; mask
; ++i
, mask
&= mask
- 1)
24054 regno
= ctz_hwi (mask
);
24055 reg
= gen_rtx_REG (SImode
, regno
);
24058 tmp
= gen_rtx_UNSPEC (BLKmode
, gen_rtvec (1, reg
), UNSPEC_PUSH_MULT
);
24060 tmp
= gen_rtx_USE (VOIDmode
, reg
);
24065 tmp
= plus_constant (Pmode
, stack_pointer_rtx
, -4 * i
);
24066 tmp
= gen_rtx_PRE_MODIFY (Pmode
, stack_pointer_rtx
, tmp
);
24067 tmp
= gen_frame_mem (BLKmode
, tmp
);
24068 tmp
= gen_rtx_SET (tmp
, par
[0]);
24071 tmp
= gen_rtx_PARALLEL (VOIDmode
, gen_rtvec_v (i
, par
));
24072 insn
= emit_insn (tmp
);
24074 /* Always build the stack adjustment note for unwind info. */
24075 tmp
= plus_constant (Pmode
, stack_pointer_rtx
, -4 * i
);
24076 tmp
= gen_rtx_SET (stack_pointer_rtx
, tmp
);
24079 /* Build the parallel of the registers recorded as saved for unwind. */
24080 for (j
= 0; real_regs
; ++j
, real_regs
&= real_regs
- 1)
24082 regno
= ctz_hwi (real_regs
);
24083 reg
= gen_rtx_REG (SImode
, regno
);
24085 tmp
= plus_constant (Pmode
, stack_pointer_rtx
, j
* 4);
24086 tmp
= gen_frame_mem (SImode
, tmp
);
24087 tmp
= gen_rtx_SET (tmp
, reg
);
24088 RTX_FRAME_RELATED_P (tmp
) = 1;
24096 RTX_FRAME_RELATED_P (par
[0]) = 1;
24097 tmp
= gen_rtx_SEQUENCE (VOIDmode
, gen_rtvec_v (j
+ 1, par
));
24100 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
, tmp
);
24105 /* Emit code to push or pop registers to or from the stack. F is the
24106 assembly file. MASK is the registers to pop. */
24108 thumb_pop (FILE *f
, unsigned long mask
)
24111 int lo_mask
= mask
& 0xFF;
24115 if (lo_mask
== 0 && (mask
& (1 << PC_REGNUM
)))
24117 /* Special case. Do not generate a POP PC statement here, do it in
24119 thumb_exit (f
, -1);
24123 fprintf (f
, "\tpop\t{");
24125 /* Look at the low registers first. */
24126 for (regno
= 0; regno
<= LAST_LO_REGNUM
; regno
++, lo_mask
>>= 1)
24130 asm_fprintf (f
, "%r", regno
);
24132 if ((lo_mask
& ~1) != 0)
24137 if (mask
& (1 << PC_REGNUM
))
24139 /* Catch popping the PC. */
24140 if (TARGET_INTERWORK
|| TARGET_BACKTRACE
|| crtl
->calls_eh_return
24141 || IS_CMSE_ENTRY (arm_current_func_type ()))
24143 /* The PC is never poped directly, instead
24144 it is popped into r3 and then BX is used. */
24145 fprintf (f
, "}\n");
24147 thumb_exit (f
, -1);
24156 asm_fprintf (f
, "%r", PC_REGNUM
);
24160 fprintf (f
, "}\n");
24163 /* Generate code to return from a thumb function.
24164 If 'reg_containing_return_addr' is -1, then the return address is
24165 actually on the stack, at the stack pointer.
24167 Note: do not forget to update length attribute of corresponding insn pattern
24168 when changing assembly output (eg. length attribute of epilogue_insns when
24169 updating Armv8-M Baseline Security Extensions register clearing
24172 thumb_exit (FILE *f
, int reg_containing_return_addr
)
24174 unsigned regs_available_for_popping
;
24175 unsigned regs_to_pop
;
24177 unsigned available
;
24181 int restore_a4
= FALSE
;
24183 /* Compute the registers we need to pop. */
24187 if (reg_containing_return_addr
== -1)
24189 regs_to_pop
|= 1 << LR_REGNUM
;
24193 if (TARGET_BACKTRACE
)
24195 /* Restore the (ARM) frame pointer and stack pointer. */
24196 regs_to_pop
|= (1 << ARM_HARD_FRAME_POINTER_REGNUM
) | (1 << SP_REGNUM
);
24200 /* If there is nothing to pop then just emit the BX instruction and
24202 if (pops_needed
== 0)
24204 if (crtl
->calls_eh_return
)
24205 asm_fprintf (f
, "\tadd\t%r, %r\n", SP_REGNUM
, ARM_EH_STACKADJ_REGNUM
);
24207 if (IS_CMSE_ENTRY (arm_current_func_type ()))
24209 asm_fprintf (f
, "\tmsr\tAPSR_nzcvq, %r\n",
24210 reg_containing_return_addr
);
24211 asm_fprintf (f
, "\tbxns\t%r\n", reg_containing_return_addr
);
24214 asm_fprintf (f
, "\tbx\t%r\n", reg_containing_return_addr
);
24217 /* Otherwise if we are not supporting interworking and we have not created
24218 a backtrace structure and the function was not entered in ARM mode then
24219 just pop the return address straight into the PC. */
24220 else if (!TARGET_INTERWORK
24221 && !TARGET_BACKTRACE
24222 && !is_called_in_ARM_mode (current_function_decl
)
24223 && !crtl
->calls_eh_return
24224 && !IS_CMSE_ENTRY (arm_current_func_type ()))
24226 asm_fprintf (f
, "\tpop\t{%r}\n", PC_REGNUM
);
24230 /* Find out how many of the (return) argument registers we can corrupt. */
24231 regs_available_for_popping
= 0;
24233 /* If returning via __builtin_eh_return, the bottom three registers
24234 all contain information needed for the return. */
24235 if (crtl
->calls_eh_return
)
24239 /* If we can deduce the registers used from the function's
24240 return value. This is more reliable that examining
24241 df_regs_ever_live_p () because that will be set if the register is
24242 ever used in the function, not just if the register is used
24243 to hold a return value. */
24245 if (crtl
->return_rtx
!= 0)
24246 mode
= GET_MODE (crtl
->return_rtx
);
24248 mode
= DECL_MODE (DECL_RESULT (current_function_decl
));
24250 size
= GET_MODE_SIZE (mode
);
24254 /* In a void function we can use any argument register.
24255 In a function that returns a structure on the stack
24256 we can use the second and third argument registers. */
24257 if (mode
== VOIDmode
)
24258 regs_available_for_popping
=
24259 (1 << ARG_REGISTER (1))
24260 | (1 << ARG_REGISTER (2))
24261 | (1 << ARG_REGISTER (3));
24263 regs_available_for_popping
=
24264 (1 << ARG_REGISTER (2))
24265 | (1 << ARG_REGISTER (3));
24267 else if (size
<= 4)
24268 regs_available_for_popping
=
24269 (1 << ARG_REGISTER (2))
24270 | (1 << ARG_REGISTER (3));
24271 else if (size
<= 8)
24272 regs_available_for_popping
=
24273 (1 << ARG_REGISTER (3));
24276 /* Match registers to be popped with registers into which we pop them. */
24277 for (available
= regs_available_for_popping
,
24278 required
= regs_to_pop
;
24279 required
!= 0 && available
!= 0;
24280 available
&= ~(available
& - available
),
24281 required
&= ~(required
& - required
))
24284 /* If we have any popping registers left over, remove them. */
24286 regs_available_for_popping
&= ~available
;
24288 /* Otherwise if we need another popping register we can use
24289 the fourth argument register. */
24290 else if (pops_needed
)
24292 /* If we have not found any free argument registers and
24293 reg a4 contains the return address, we must move it. */
24294 if (regs_available_for_popping
== 0
24295 && reg_containing_return_addr
== LAST_ARG_REGNUM
)
24297 asm_fprintf (f
, "\tmov\t%r, %r\n", LR_REGNUM
, LAST_ARG_REGNUM
);
24298 reg_containing_return_addr
= LR_REGNUM
;
24300 else if (size
> 12)
24302 /* Register a4 is being used to hold part of the return value,
24303 but we have dire need of a free, low register. */
24306 asm_fprintf (f
, "\tmov\t%r, %r\n",IP_REGNUM
, LAST_ARG_REGNUM
);
24309 if (reg_containing_return_addr
!= LAST_ARG_REGNUM
)
24311 /* The fourth argument register is available. */
24312 regs_available_for_popping
|= 1 << LAST_ARG_REGNUM
;
24318 /* Pop as many registers as we can. */
24319 thumb_pop (f
, regs_available_for_popping
);
24321 /* Process the registers we popped. */
24322 if (reg_containing_return_addr
== -1)
24324 /* The return address was popped into the lowest numbered register. */
24325 regs_to_pop
&= ~(1 << LR_REGNUM
);
24327 reg_containing_return_addr
=
24328 number_of_first_bit_set (regs_available_for_popping
);
24330 /* Remove this register for the mask of available registers, so that
24331 the return address will not be corrupted by further pops. */
24332 regs_available_for_popping
&= ~(1 << reg_containing_return_addr
);
24335 /* If we popped other registers then handle them here. */
24336 if (regs_available_for_popping
)
24340 /* Work out which register currently contains the frame pointer. */
24341 frame_pointer
= number_of_first_bit_set (regs_available_for_popping
);
24343 /* Move it into the correct place. */
24344 asm_fprintf (f
, "\tmov\t%r, %r\n",
24345 ARM_HARD_FRAME_POINTER_REGNUM
, frame_pointer
);
24347 /* (Temporarily) remove it from the mask of popped registers. */
24348 regs_available_for_popping
&= ~(1 << frame_pointer
);
24349 regs_to_pop
&= ~(1 << ARM_HARD_FRAME_POINTER_REGNUM
);
24351 if (regs_available_for_popping
)
24355 /* We popped the stack pointer as well,
24356 find the register that contains it. */
24357 stack_pointer
= number_of_first_bit_set (regs_available_for_popping
);
24359 /* Move it into the stack register. */
24360 asm_fprintf (f
, "\tmov\t%r, %r\n", SP_REGNUM
, stack_pointer
);
24362 /* At this point we have popped all necessary registers, so
24363 do not worry about restoring regs_available_for_popping
24364 to its correct value:
24366 assert (pops_needed == 0)
24367 assert (regs_available_for_popping == (1 << frame_pointer))
24368 assert (regs_to_pop == (1 << STACK_POINTER)) */
24372 /* Since we have just move the popped value into the frame
24373 pointer, the popping register is available for reuse, and
24374 we know that we still have the stack pointer left to pop. */
24375 regs_available_for_popping
|= (1 << frame_pointer
);
24379 /* If we still have registers left on the stack, but we no longer have
24380 any registers into which we can pop them, then we must move the return
24381 address into the link register and make available the register that
24383 if (regs_available_for_popping
== 0 && pops_needed
> 0)
24385 regs_available_for_popping
|= 1 << reg_containing_return_addr
;
24387 asm_fprintf (f
, "\tmov\t%r, %r\n", LR_REGNUM
,
24388 reg_containing_return_addr
);
24390 reg_containing_return_addr
= LR_REGNUM
;
24393 /* If we have registers left on the stack then pop some more.
24394 We know that at most we will want to pop FP and SP. */
24395 if (pops_needed
> 0)
24400 thumb_pop (f
, regs_available_for_popping
);
24402 /* We have popped either FP or SP.
24403 Move whichever one it is into the correct register. */
24404 popped_into
= number_of_first_bit_set (regs_available_for_popping
);
24405 move_to
= number_of_first_bit_set (regs_to_pop
);
24407 asm_fprintf (f
, "\tmov\t%r, %r\n", move_to
, popped_into
);
24411 /* If we still have not popped everything then we must have only
24412 had one register available to us and we are now popping the SP. */
24413 if (pops_needed
> 0)
24417 thumb_pop (f
, regs_available_for_popping
);
24419 popped_into
= number_of_first_bit_set (regs_available_for_popping
);
24421 asm_fprintf (f
, "\tmov\t%r, %r\n", SP_REGNUM
, popped_into
);
24423 assert (regs_to_pop == (1 << STACK_POINTER))
24424 assert (pops_needed == 1)
24428 /* If necessary restore the a4 register. */
24431 if (reg_containing_return_addr
!= LR_REGNUM
)
24433 asm_fprintf (f
, "\tmov\t%r, %r\n", LR_REGNUM
, LAST_ARG_REGNUM
);
24434 reg_containing_return_addr
= LR_REGNUM
;
24437 asm_fprintf (f
, "\tmov\t%r, %r\n", LAST_ARG_REGNUM
, IP_REGNUM
);
24440 if (crtl
->calls_eh_return
)
24441 asm_fprintf (f
, "\tadd\t%r, %r\n", SP_REGNUM
, ARM_EH_STACKADJ_REGNUM
);
24443 /* Return to caller. */
24444 if (IS_CMSE_ENTRY (arm_current_func_type ()))
24446 /* This is for the cases where LR is not being used to contain the return
24447 address. It may therefore contain information that we might not want
24448 to leak, hence it must be cleared. The value in R0 will never be a
24449 secret at this point, so it is safe to use it, see the clearing code
24450 in 'cmse_nonsecure_entry_clear_before_return'. */
24451 if (reg_containing_return_addr
!= LR_REGNUM
)
24452 asm_fprintf (f
, "\tmov\tlr, r0\n");
24454 asm_fprintf (f
, "\tmsr\tAPSR_nzcvq, %r\n", reg_containing_return_addr
);
24455 asm_fprintf (f
, "\tbxns\t%r\n", reg_containing_return_addr
);
24458 asm_fprintf (f
, "\tbx\t%r\n", reg_containing_return_addr
);
24461 /* Scan INSN just before assembler is output for it.
24462 For Thumb-1, we track the status of the condition codes; this
24463 information is used in the cbranchsi4_insn pattern. */
24465 thumb1_final_prescan_insn (rtx_insn
*insn
)
24467 if (flag_print_asm_name
)
24468 asm_fprintf (asm_out_file
, "%@ 0x%04x\n",
24469 INSN_ADDRESSES (INSN_UID (insn
)));
24470 /* Don't overwrite the previous setter when we get to a cbranch. */
24471 if (INSN_CODE (insn
) != CODE_FOR_cbranchsi4_insn
)
24473 enum attr_conds conds
;
24475 if (cfun
->machine
->thumb1_cc_insn
)
24477 if (modified_in_p (cfun
->machine
->thumb1_cc_op0
, insn
)
24478 || modified_in_p (cfun
->machine
->thumb1_cc_op1
, insn
))
24481 conds
= get_attr_conds (insn
);
24482 if (conds
== CONDS_SET
)
24484 rtx set
= single_set (insn
);
24485 cfun
->machine
->thumb1_cc_insn
= insn
;
24486 cfun
->machine
->thumb1_cc_op0
= SET_DEST (set
);
24487 cfun
->machine
->thumb1_cc_op1
= const0_rtx
;
24488 cfun
->machine
->thumb1_cc_mode
= CC_NOOVmode
;
24489 if (INSN_CODE (insn
) == CODE_FOR_thumb1_subsi3_insn
)
24491 rtx src1
= XEXP (SET_SRC (set
), 1);
24492 if (src1
== const0_rtx
)
24493 cfun
->machine
->thumb1_cc_mode
= CCmode
;
24495 else if (REG_P (SET_DEST (set
)) && REG_P (SET_SRC (set
)))
24497 /* Record the src register operand instead of dest because
24498 cprop_hardreg pass propagates src. */
24499 cfun
->machine
->thumb1_cc_op0
= SET_SRC (set
);
24502 else if (conds
!= CONDS_NOCOND
)
24503 cfun
->machine
->thumb1_cc_insn
= NULL_RTX
;
24506 /* Check if unexpected far jump is used. */
24507 if (cfun
->machine
->lr_save_eliminated
24508 && get_attr_far_jump (insn
) == FAR_JUMP_YES
)
24509 internal_error("Unexpected thumb1 far jump");
24513 thumb_shiftable_const (unsigned HOST_WIDE_INT val
)
24515 unsigned HOST_WIDE_INT mask
= 0xff;
24518 val
= val
& (unsigned HOST_WIDE_INT
)0xffffffffu
;
24519 if (val
== 0) /* XXX */
24522 for (i
= 0; i
< 25; i
++)
24523 if ((val
& (mask
<< i
)) == val
)
24529 /* Returns nonzero if the current function contains,
24530 or might contain a far jump. */
24532 thumb_far_jump_used_p (void)
24535 bool far_jump
= false;
24536 unsigned int func_size
= 0;
24538 /* If we have already decided that far jumps may be used,
24539 do not bother checking again, and always return true even if
24540 it turns out that they are not being used. Once we have made
24541 the decision that far jumps are present (and that hence the link
24542 register will be pushed onto the stack) we cannot go back on it. */
24543 if (cfun
->machine
->far_jump_used
)
24546 /* If this function is not being called from the prologue/epilogue
24547 generation code then it must be being called from the
24548 INITIAL_ELIMINATION_OFFSET macro. */
24549 if (!(ARM_DOUBLEWORD_ALIGN
|| reload_completed
))
24551 /* In this case we know that we are being asked about the elimination
24552 of the arg pointer register. If that register is not being used,
24553 then there are no arguments on the stack, and we do not have to
24554 worry that a far jump might force the prologue to push the link
24555 register, changing the stack offsets. In this case we can just
24556 return false, since the presence of far jumps in the function will
24557 not affect stack offsets.
24559 If the arg pointer is live (or if it was live, but has now been
24560 eliminated and so set to dead) then we do have to test to see if
24561 the function might contain a far jump. This test can lead to some
24562 false negatives, since before reload is completed, then length of
24563 branch instructions is not known, so gcc defaults to returning their
24564 longest length, which in turn sets the far jump attribute to true.
24566 A false negative will not result in bad code being generated, but it
24567 will result in a needless push and pop of the link register. We
24568 hope that this does not occur too often.
24570 If we need doubleword stack alignment this could affect the other
24571 elimination offsets so we can't risk getting it wrong. */
24572 if (df_regs_ever_live_p (ARG_POINTER_REGNUM
))
24573 cfun
->machine
->arg_pointer_live
= 1;
24574 else if (!cfun
->machine
->arg_pointer_live
)
24578 /* We should not change far_jump_used during or after reload, as there is
24579 no chance to change stack frame layout. */
24580 if (reload_in_progress
|| reload_completed
)
24583 /* Check to see if the function contains a branch
24584 insn with the far jump attribute set. */
24585 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
24587 if (JUMP_P (insn
) && get_attr_far_jump (insn
) == FAR_JUMP_YES
)
24591 func_size
+= get_attr_length (insn
);
24594 /* Attribute far_jump will always be true for thumb1 before
24595 shorten_branch pass. So checking far_jump attribute before
24596 shorten_branch isn't much useful.
24598 Following heuristic tries to estimate more accurately if a far jump
24599 may finally be used. The heuristic is very conservative as there is
24600 no chance to roll-back the decision of not to use far jump.
24602 Thumb1 long branch offset is -2048 to 2046. The worst case is each
24603 2-byte insn is associated with a 4 byte constant pool. Using
24604 function size 2048/3 as the threshold is conservative enough. */
24607 if ((func_size
* 3) >= 2048)
24609 /* Record the fact that we have decided that
24610 the function does use far jumps. */
24611 cfun
->machine
->far_jump_used
= 1;
24619 /* Return nonzero if FUNC must be entered in ARM mode. */
24621 is_called_in_ARM_mode (tree func
)
24623 gcc_assert (TREE_CODE (func
) == FUNCTION_DECL
);
24625 /* Ignore the problem about functions whose address is taken. */
24626 if (TARGET_CALLEE_INTERWORKING
&& TREE_PUBLIC (func
))
24630 return lookup_attribute ("interfacearm", DECL_ATTRIBUTES (func
)) != NULL_TREE
;
24636 /* Given the stack offsets and register mask in OFFSETS, decide how
24637 many additional registers to push instead of subtracting a constant
24638 from SP. For epilogues the principle is the same except we use pop.
24639 FOR_PROLOGUE indicates which we're generating. */
24641 thumb1_extra_regs_pushed (arm_stack_offsets
*offsets
, bool for_prologue
)
24643 HOST_WIDE_INT amount
;
24644 unsigned long live_regs_mask
= offsets
->saved_regs_mask
;
24645 /* Extract a mask of the ones we can give to the Thumb's push/pop
24647 unsigned long l_mask
= live_regs_mask
& (for_prologue
? 0x40ff : 0xff);
24648 /* Then count how many other high registers will need to be pushed. */
24649 unsigned long high_regs_pushed
= bit_count (live_regs_mask
& 0x0f00);
24650 int n_free
, reg_base
, size
;
24652 if (!for_prologue
&& frame_pointer_needed
)
24653 amount
= offsets
->locals_base
- offsets
->saved_regs
;
24655 amount
= offsets
->outgoing_args
- offsets
->saved_regs
;
24657 /* If the stack frame size is 512 exactly, we can save one load
24658 instruction, which should make this a win even when optimizing
24660 if (!optimize_size
&& amount
!= 512)
24663 /* Can't do this if there are high registers to push. */
24664 if (high_regs_pushed
!= 0)
24667 /* Shouldn't do it in the prologue if no registers would normally
24668 be pushed at all. In the epilogue, also allow it if we'll have
24669 a pop insn for the PC. */
24672 || TARGET_BACKTRACE
24673 || (live_regs_mask
& 1 << LR_REGNUM
) == 0
24674 || TARGET_INTERWORK
24675 || crtl
->args
.pretend_args_size
!= 0))
24678 /* Don't do this if thumb_expand_prologue wants to emit instructions
24679 between the push and the stack frame allocation. */
24681 && ((flag_pic
&& arm_pic_register
!= INVALID_REGNUM
)
24682 || (!frame_pointer_needed
&& CALLER_INTERWORKING_SLOT_SIZE
> 0)))
24689 size
= arm_size_return_regs ();
24690 reg_base
= ARM_NUM_INTS (size
);
24691 live_regs_mask
>>= reg_base
;
24694 while (reg_base
+ n_free
< 8 && !(live_regs_mask
& 1)
24695 && (for_prologue
|| call_used_regs
[reg_base
+ n_free
]))
24697 live_regs_mask
>>= 1;
24703 gcc_assert (amount
/ 4 * 4 == amount
);
24705 if (amount
>= 512 && (amount
- n_free
* 4) < 512)
24706 return (amount
- 508) / 4;
24707 if (amount
<= n_free
* 4)
24712 /* The bits which aren't usefully expanded as rtl. */
24714 thumb1_unexpanded_epilogue (void)
24716 arm_stack_offsets
*offsets
;
24718 unsigned long live_regs_mask
= 0;
24719 int high_regs_pushed
= 0;
24721 int had_to_push_lr
;
24724 if (cfun
->machine
->return_used_this_function
!= 0)
24727 if (IS_NAKED (arm_current_func_type ()))
24730 offsets
= arm_get_frame_offsets ();
24731 live_regs_mask
= offsets
->saved_regs_mask
;
24732 high_regs_pushed
= bit_count (live_regs_mask
& 0x0f00);
24734 /* If we can deduce the registers used from the function's return value.
24735 This is more reliable that examining df_regs_ever_live_p () because that
24736 will be set if the register is ever used in the function, not just if
24737 the register is used to hold a return value. */
24738 size
= arm_size_return_regs ();
24740 extra_pop
= thumb1_extra_regs_pushed (offsets
, false);
24743 unsigned long extra_mask
= (1 << extra_pop
) - 1;
24744 live_regs_mask
|= extra_mask
<< ARM_NUM_INTS (size
);
24747 /* The prolog may have pushed some high registers to use as
24748 work registers. e.g. the testsuite file:
24749 gcc/testsuite/gcc/gcc.c-torture/execute/complex-2.c
24750 compiles to produce:
24751 push {r4, r5, r6, r7, lr}
24755 as part of the prolog. We have to undo that pushing here. */
24757 if (high_regs_pushed
)
24759 unsigned long mask
= live_regs_mask
& 0xff;
24762 /* The available low registers depend on the size of the value we are
24770 /* Oh dear! We have no low registers into which we can pop
24773 ("no low registers available for popping high registers");
24775 for (next_hi_reg
= 8; next_hi_reg
< 13; next_hi_reg
++)
24776 if (live_regs_mask
& (1 << next_hi_reg
))
24779 while (high_regs_pushed
)
24781 /* Find lo register(s) into which the high register(s) can
24783 for (regno
= 0; regno
<= LAST_LO_REGNUM
; regno
++)
24785 if (mask
& (1 << regno
))
24786 high_regs_pushed
--;
24787 if (high_regs_pushed
== 0)
24791 mask
&= (2 << regno
) - 1; /* A noop if regno == 8 */
24793 /* Pop the values into the low register(s). */
24794 thumb_pop (asm_out_file
, mask
);
24796 /* Move the value(s) into the high registers. */
24797 for (regno
= 0; regno
<= LAST_LO_REGNUM
; regno
++)
24799 if (mask
& (1 << regno
))
24801 asm_fprintf (asm_out_file
, "\tmov\t%r, %r\n", next_hi_reg
,
24804 for (next_hi_reg
++; next_hi_reg
< 13; next_hi_reg
++)
24805 if (live_regs_mask
& (1 << next_hi_reg
))
24810 live_regs_mask
&= ~0x0f00;
24813 had_to_push_lr
= (live_regs_mask
& (1 << LR_REGNUM
)) != 0;
24814 live_regs_mask
&= 0xff;
24816 if (crtl
->args
.pretend_args_size
== 0 || TARGET_BACKTRACE
)
24818 /* Pop the return address into the PC. */
24819 if (had_to_push_lr
)
24820 live_regs_mask
|= 1 << PC_REGNUM
;
24822 /* Either no argument registers were pushed or a backtrace
24823 structure was created which includes an adjusted stack
24824 pointer, so just pop everything. */
24825 if (live_regs_mask
)
24826 thumb_pop (asm_out_file
, live_regs_mask
);
24828 /* We have either just popped the return address into the
24829 PC or it is was kept in LR for the entire function.
24830 Note that thumb_pop has already called thumb_exit if the
24831 PC was in the list. */
24832 if (!had_to_push_lr
)
24833 thumb_exit (asm_out_file
, LR_REGNUM
);
24837 /* Pop everything but the return address. */
24838 if (live_regs_mask
)
24839 thumb_pop (asm_out_file
, live_regs_mask
);
24841 if (had_to_push_lr
)
24845 /* We have no free low regs, so save one. */
24846 asm_fprintf (asm_out_file
, "\tmov\t%r, %r\n", IP_REGNUM
,
24850 /* Get the return address into a temporary register. */
24851 thumb_pop (asm_out_file
, 1 << LAST_ARG_REGNUM
);
24855 /* Move the return address to lr. */
24856 asm_fprintf (asm_out_file
, "\tmov\t%r, %r\n", LR_REGNUM
,
24858 /* Restore the low register. */
24859 asm_fprintf (asm_out_file
, "\tmov\t%r, %r\n", LAST_ARG_REGNUM
,
24864 regno
= LAST_ARG_REGNUM
;
24869 /* Remove the argument registers that were pushed onto the stack. */
24870 asm_fprintf (asm_out_file
, "\tadd\t%r, %r, #%d\n",
24871 SP_REGNUM
, SP_REGNUM
,
24872 crtl
->args
.pretend_args_size
);
24874 thumb_exit (asm_out_file
, regno
);
24880 /* Functions to save and restore machine-specific function data. */
24881 static struct machine_function
*
24882 arm_init_machine_status (void)
24884 struct machine_function
*machine
;
24885 machine
= ggc_cleared_alloc
<machine_function
> ();
24887 #if ARM_FT_UNKNOWN != 0
24888 machine
->func_type
= ARM_FT_UNKNOWN
;
24890 machine
->static_chain_stack_bytes
= -1;
24894 /* Return an RTX indicating where the return address to the
24895 calling function can be found. */
24897 arm_return_addr (int count
, rtx frame ATTRIBUTE_UNUSED
)
24902 return get_hard_reg_initial_val (Pmode
, LR_REGNUM
);
24905 /* Do anything needed before RTL is emitted for each function. */
24907 arm_init_expanders (void)
24909 /* Arrange to initialize and mark the machine per-function status. */
24910 init_machine_status
= arm_init_machine_status
;
24912 /* This is to stop the combine pass optimizing away the alignment
24913 adjustment of va_arg. */
24914 /* ??? It is claimed that this should not be necessary. */
24916 mark_reg_pointer (arg_pointer_rtx
, PARM_BOUNDARY
);
24919 /* Check that FUNC is called with a different mode. */
24922 arm_change_mode_p (tree func
)
24924 if (TREE_CODE (func
) != FUNCTION_DECL
)
24927 tree callee_tree
= DECL_FUNCTION_SPECIFIC_TARGET (func
);
24930 callee_tree
= target_option_default_node
;
24932 struct cl_target_option
*callee_opts
= TREE_TARGET_OPTION (callee_tree
);
24933 int flags
= callee_opts
->x_target_flags
;
24935 return (TARGET_THUMB_P (flags
) != TARGET_THUMB
);
24938 /* Like arm_compute_initial_elimination offset. Simpler because there
24939 isn't an ABI specified frame pointer for Thumb. Instead, we set it
24940 to point at the base of the local variables after static stack
24941 space for a function has been allocated. */
24944 thumb_compute_initial_elimination_offset (unsigned int from
, unsigned int to
)
24946 arm_stack_offsets
*offsets
;
24948 offsets
= arm_get_frame_offsets ();
24952 case ARG_POINTER_REGNUM
:
24955 case STACK_POINTER_REGNUM
:
24956 return offsets
->outgoing_args
- offsets
->saved_args
;
24958 case FRAME_POINTER_REGNUM
:
24959 return offsets
->soft_frame
- offsets
->saved_args
;
24961 case ARM_HARD_FRAME_POINTER_REGNUM
:
24962 return offsets
->saved_regs
- offsets
->saved_args
;
24964 case THUMB_HARD_FRAME_POINTER_REGNUM
:
24965 return offsets
->locals_base
- offsets
->saved_args
;
24968 gcc_unreachable ();
24972 case FRAME_POINTER_REGNUM
:
24975 case STACK_POINTER_REGNUM
:
24976 return offsets
->outgoing_args
- offsets
->soft_frame
;
24978 case ARM_HARD_FRAME_POINTER_REGNUM
:
24979 return offsets
->saved_regs
- offsets
->soft_frame
;
24981 case THUMB_HARD_FRAME_POINTER_REGNUM
:
24982 return offsets
->locals_base
- offsets
->soft_frame
;
24985 gcc_unreachable ();
24990 gcc_unreachable ();
24994 /* Generate the function's prologue. */
24997 thumb1_expand_prologue (void)
25001 HOST_WIDE_INT amount
;
25002 HOST_WIDE_INT size
;
25003 arm_stack_offsets
*offsets
;
25004 unsigned long func_type
;
25006 unsigned long live_regs_mask
;
25007 unsigned long l_mask
;
25008 unsigned high_regs_pushed
= 0;
25009 bool lr_needs_saving
;
25011 func_type
= arm_current_func_type ();
25013 /* Naked functions don't have prologues. */
25014 if (IS_NAKED (func_type
))
25016 if (flag_stack_usage_info
)
25017 current_function_static_stack_size
= 0;
25021 if (IS_INTERRUPT (func_type
))
25023 error ("interrupt Service Routines cannot be coded in Thumb mode");
25027 if (is_called_in_ARM_mode (current_function_decl
))
25028 emit_insn (gen_prologue_thumb1_interwork ());
25030 offsets
= arm_get_frame_offsets ();
25031 live_regs_mask
= offsets
->saved_regs_mask
;
25032 lr_needs_saving
= live_regs_mask
& (1 << LR_REGNUM
);
25034 /* Extract a mask of the ones we can give to the Thumb's push instruction. */
25035 l_mask
= live_regs_mask
& 0x40ff;
25036 /* Then count how many other high registers will need to be pushed. */
25037 high_regs_pushed
= bit_count (live_regs_mask
& 0x0f00);
25039 if (crtl
->args
.pretend_args_size
)
25041 rtx x
= GEN_INT (-crtl
->args
.pretend_args_size
);
25043 if (cfun
->machine
->uses_anonymous_args
)
25045 int num_pushes
= ARM_NUM_INTS (crtl
->args
.pretend_args_size
);
25046 unsigned long mask
;
25048 mask
= 1ul << (LAST_ARG_REGNUM
+ 1);
25049 mask
-= 1ul << (LAST_ARG_REGNUM
+ 1 - num_pushes
);
25051 insn
= thumb1_emit_multi_reg_push (mask
, 0);
25055 insn
= emit_insn (gen_addsi3 (stack_pointer_rtx
,
25056 stack_pointer_rtx
, x
));
25058 RTX_FRAME_RELATED_P (insn
) = 1;
25061 if (TARGET_BACKTRACE
)
25063 HOST_WIDE_INT offset
= 0;
25064 unsigned work_register
;
25065 rtx work_reg
, x
, arm_hfp_rtx
;
25067 /* We have been asked to create a stack backtrace structure.
25068 The code looks like this:
25072 0 sub SP, #16 Reserve space for 4 registers.
25073 2 push {R7} Push low registers.
25074 4 add R7, SP, #20 Get the stack pointer before the push.
25075 6 str R7, [SP, #8] Store the stack pointer
25076 (before reserving the space).
25077 8 mov R7, PC Get hold of the start of this code + 12.
25078 10 str R7, [SP, #16] Store it.
25079 12 mov R7, FP Get hold of the current frame pointer.
25080 14 str R7, [SP, #4] Store it.
25081 16 mov R7, LR Get hold of the current return address.
25082 18 str R7, [SP, #12] Store it.
25083 20 add R7, SP, #16 Point at the start of the
25084 backtrace structure.
25085 22 mov FP, R7 Put this value into the frame pointer. */
25087 work_register
= thumb_find_work_register (live_regs_mask
);
25088 work_reg
= gen_rtx_REG (SImode
, work_register
);
25089 arm_hfp_rtx
= gen_rtx_REG (SImode
, ARM_HARD_FRAME_POINTER_REGNUM
);
25091 insn
= emit_insn (gen_addsi3 (stack_pointer_rtx
,
25092 stack_pointer_rtx
, GEN_INT (-16)));
25093 RTX_FRAME_RELATED_P (insn
) = 1;
25097 insn
= thumb1_emit_multi_reg_push (l_mask
, l_mask
);
25098 RTX_FRAME_RELATED_P (insn
) = 1;
25099 lr_needs_saving
= false;
25101 offset
= bit_count (l_mask
) * UNITS_PER_WORD
;
25104 x
= GEN_INT (offset
+ 16 + crtl
->args
.pretend_args_size
);
25105 emit_insn (gen_addsi3 (work_reg
, stack_pointer_rtx
, x
));
25107 x
= plus_constant (Pmode
, stack_pointer_rtx
, offset
+ 4);
25108 x
= gen_frame_mem (SImode
, x
);
25109 emit_move_insn (x
, work_reg
);
25111 /* Make sure that the instruction fetching the PC is in the right place
25112 to calculate "start of backtrace creation code + 12". */
25113 /* ??? The stores using the common WORK_REG ought to be enough to
25114 prevent the scheduler from doing anything weird. Failing that
25115 we could always move all of the following into an UNSPEC_VOLATILE. */
25118 x
= gen_rtx_REG (SImode
, PC_REGNUM
);
25119 emit_move_insn (work_reg
, x
);
25121 x
= plus_constant (Pmode
, stack_pointer_rtx
, offset
+ 12);
25122 x
= gen_frame_mem (SImode
, x
);
25123 emit_move_insn (x
, work_reg
);
25125 emit_move_insn (work_reg
, arm_hfp_rtx
);
25127 x
= plus_constant (Pmode
, stack_pointer_rtx
, offset
);
25128 x
= gen_frame_mem (SImode
, x
);
25129 emit_move_insn (x
, work_reg
);
25133 emit_move_insn (work_reg
, arm_hfp_rtx
);
25135 x
= plus_constant (Pmode
, stack_pointer_rtx
, offset
);
25136 x
= gen_frame_mem (SImode
, x
);
25137 emit_move_insn (x
, work_reg
);
25139 x
= gen_rtx_REG (SImode
, PC_REGNUM
);
25140 emit_move_insn (work_reg
, x
);
25142 x
= plus_constant (Pmode
, stack_pointer_rtx
, offset
+ 12);
25143 x
= gen_frame_mem (SImode
, x
);
25144 emit_move_insn (x
, work_reg
);
25147 x
= gen_rtx_REG (SImode
, LR_REGNUM
);
25148 emit_move_insn (work_reg
, x
);
25150 x
= plus_constant (Pmode
, stack_pointer_rtx
, offset
+ 8);
25151 x
= gen_frame_mem (SImode
, x
);
25152 emit_move_insn (x
, work_reg
);
25154 x
= GEN_INT (offset
+ 12);
25155 emit_insn (gen_addsi3 (work_reg
, stack_pointer_rtx
, x
));
25157 emit_move_insn (arm_hfp_rtx
, work_reg
);
25159 /* Optimization: If we are not pushing any low registers but we are going
25160 to push some high registers then delay our first push. This will just
25161 be a push of LR and we can combine it with the push of the first high
25163 else if ((l_mask
& 0xff) != 0
25164 || (high_regs_pushed
== 0 && lr_needs_saving
))
25166 unsigned long mask
= l_mask
;
25167 mask
|= (1 << thumb1_extra_regs_pushed (offsets
, true)) - 1;
25168 insn
= thumb1_emit_multi_reg_push (mask
, mask
);
25169 RTX_FRAME_RELATED_P (insn
) = 1;
25170 lr_needs_saving
= false;
25173 if (high_regs_pushed
)
25175 unsigned pushable_regs
;
25176 unsigned next_hi_reg
;
25177 unsigned arg_regs_num
= TARGET_AAPCS_BASED
? crtl
->args
.info
.aapcs_ncrn
25178 : crtl
->args
.info
.nregs
;
25179 unsigned arg_regs_mask
= (1 << arg_regs_num
) - 1;
25181 for (next_hi_reg
= 12; next_hi_reg
> LAST_LO_REGNUM
; next_hi_reg
--)
25182 if (live_regs_mask
& (1 << next_hi_reg
))
25185 /* Here we need to mask out registers used for passing arguments
25186 even if they can be pushed. This is to avoid using them to stash the high
25187 registers. Such kind of stash may clobber the use of arguments. */
25188 pushable_regs
= l_mask
& (~arg_regs_mask
);
25189 if (lr_needs_saving
)
25190 pushable_regs
&= ~(1 << LR_REGNUM
);
25192 if (pushable_regs
== 0)
25193 pushable_regs
= 1 << thumb_find_work_register (live_regs_mask
);
25195 while (high_regs_pushed
> 0)
25197 unsigned long real_regs_mask
= 0;
25198 unsigned long push_mask
= 0;
25200 for (regno
= LR_REGNUM
; regno
>= 0; regno
--)
25202 if (pushable_regs
& (1 << regno
))
25204 emit_move_insn (gen_rtx_REG (SImode
, regno
),
25205 gen_rtx_REG (SImode
, next_hi_reg
));
25207 high_regs_pushed
--;
25208 real_regs_mask
|= (1 << next_hi_reg
);
25209 push_mask
|= (1 << regno
);
25211 if (high_regs_pushed
)
25213 for (next_hi_reg
--; next_hi_reg
> LAST_LO_REGNUM
;
25215 if (live_regs_mask
& (1 << next_hi_reg
))
25223 /* If we had to find a work register and we have not yet
25224 saved the LR then add it to the list of regs to push. */
25225 if (lr_needs_saving
)
25227 push_mask
|= 1 << LR_REGNUM
;
25228 real_regs_mask
|= 1 << LR_REGNUM
;
25229 lr_needs_saving
= false;
25232 insn
= thumb1_emit_multi_reg_push (push_mask
, real_regs_mask
);
25233 RTX_FRAME_RELATED_P (insn
) = 1;
25237 /* Load the pic register before setting the frame pointer,
25238 so we can use r7 as a temporary work register. */
25239 if (flag_pic
&& arm_pic_register
!= INVALID_REGNUM
)
25240 arm_load_pic_register (live_regs_mask
);
25242 if (!frame_pointer_needed
&& CALLER_INTERWORKING_SLOT_SIZE
> 0)
25243 emit_move_insn (gen_rtx_REG (Pmode
, ARM_HARD_FRAME_POINTER_REGNUM
),
25244 stack_pointer_rtx
);
25246 size
= offsets
->outgoing_args
- offsets
->saved_args
;
25247 if (flag_stack_usage_info
)
25248 current_function_static_stack_size
= size
;
25250 /* If we have a frame, then do stack checking. FIXME: not implemented. */
25251 if ((flag_stack_check
== STATIC_BUILTIN_STACK_CHECK
25252 || flag_stack_clash_protection
)
25254 sorry ("-fstack-check=specific for Thumb-1");
25256 amount
= offsets
->outgoing_args
- offsets
->saved_regs
;
25257 amount
-= 4 * thumb1_extra_regs_pushed (offsets
, true);
25262 insn
= emit_insn (gen_addsi3 (stack_pointer_rtx
, stack_pointer_rtx
,
25263 GEN_INT (- amount
)));
25264 RTX_FRAME_RELATED_P (insn
) = 1;
25270 /* The stack decrement is too big for an immediate value in a single
25271 insn. In theory we could issue multiple subtracts, but after
25272 three of them it becomes more space efficient to place the full
25273 value in the constant pool and load into a register. (Also the
25274 ARM debugger really likes to see only one stack decrement per
25275 function). So instead we look for a scratch register into which
25276 we can load the decrement, and then we subtract this from the
25277 stack pointer. Unfortunately on the thumb the only available
25278 scratch registers are the argument registers, and we cannot use
25279 these as they may hold arguments to the function. Instead we
25280 attempt to locate a call preserved register which is used by this
25281 function. If we can find one, then we know that it will have
25282 been pushed at the start of the prologue and so we can corrupt
25284 for (regno
= LAST_ARG_REGNUM
+ 1; regno
<= LAST_LO_REGNUM
; regno
++)
25285 if (live_regs_mask
& (1 << regno
))
25288 gcc_assert(regno
<= LAST_LO_REGNUM
);
25290 reg
= gen_rtx_REG (SImode
, regno
);
25292 emit_insn (gen_movsi (reg
, GEN_INT (- amount
)));
25294 insn
= emit_insn (gen_addsi3 (stack_pointer_rtx
,
25295 stack_pointer_rtx
, reg
));
25297 dwarf
= gen_rtx_SET (stack_pointer_rtx
,
25298 plus_constant (Pmode
, stack_pointer_rtx
,
25300 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
, dwarf
);
25301 RTX_FRAME_RELATED_P (insn
) = 1;
25305 if (frame_pointer_needed
)
25306 thumb_set_frame_pointer (offsets
);
25308 /* If we are profiling, make sure no instructions are scheduled before
25309 the call to mcount. Similarly if the user has requested no
25310 scheduling in the prolog. Similarly if we want non-call exceptions
25311 using the EABI unwinder, to prevent faulting instructions from being
25312 swapped with a stack adjustment. */
25313 if (crtl
->profile
|| !TARGET_SCHED_PROLOG
25314 || (arm_except_unwind_info (&global_options
) == UI_TARGET
25315 && cfun
->can_throw_non_call_exceptions
))
25316 emit_insn (gen_blockage ());
25318 cfun
->machine
->lr_save_eliminated
= !thumb_force_lr_save ();
25319 if (live_regs_mask
& 0xff)
25320 cfun
->machine
->lr_save_eliminated
= 0;
25323 /* Clear caller saved registers not used to pass return values and leaked
25324 condition flags before exiting a cmse_nonsecure_entry function. */
25327 cmse_nonsecure_entry_clear_before_return (void)
25329 int regno
, maxregno
= TARGET_HARD_FLOAT
? LAST_VFP_REGNUM
: IP_REGNUM
;
25330 uint32_t padding_bits_to_clear
= 0;
25331 auto_sbitmap
to_clear_bitmap (maxregno
+ 1);
25332 rtx r1_reg
, result_rtl
, clearing_reg
= NULL_RTX
;
25335 bitmap_clear (to_clear_bitmap
);
25336 bitmap_set_range (to_clear_bitmap
, R0_REGNUM
, NUM_ARG_REGS
);
25337 bitmap_set_bit (to_clear_bitmap
, IP_REGNUM
);
25339 /* If we are not dealing with -mfloat-abi=soft we will need to clear VFP
25341 if (TARGET_HARD_FLOAT
)
25343 int float_bits
= D7_VFP_REGNUM
- FIRST_VFP_REGNUM
+ 1;
25345 bitmap_set_range (to_clear_bitmap
, FIRST_VFP_REGNUM
, float_bits
);
25347 /* Make sure we don't clear the two scratch registers used to clear the
25348 relevant FPSCR bits in output_return_instruction. */
25349 emit_use (gen_rtx_REG (SImode
, IP_REGNUM
));
25350 bitmap_clear_bit (to_clear_bitmap
, IP_REGNUM
);
25351 emit_use (gen_rtx_REG (SImode
, 4));
25352 bitmap_clear_bit (to_clear_bitmap
, 4);
25355 /* If the user has defined registers to be caller saved, these are no longer
25356 restored by the function before returning and must thus be cleared for
25357 security purposes. */
25358 for (regno
= NUM_ARG_REGS
; regno
<= maxregno
; regno
++)
25360 /* We do not touch registers that can be used to pass arguments as per
25361 the AAPCS, since these should never be made callee-saved by user
25363 if (IN_RANGE (regno
, FIRST_VFP_REGNUM
, D7_VFP_REGNUM
))
25365 if (IN_RANGE (regno
, IP_REGNUM
, PC_REGNUM
))
25367 if (call_used_regs
[regno
])
25368 bitmap_set_bit (to_clear_bitmap
, regno
);
25371 /* Make sure we do not clear the registers used to return the result in. */
25372 result_type
= TREE_TYPE (DECL_RESULT (current_function_decl
));
25373 if (!VOID_TYPE_P (result_type
))
25375 uint64_t to_clear_return_mask
;
25376 result_rtl
= arm_function_value (result_type
, current_function_decl
, 0);
25378 /* No need to check that we return in registers, because we don't
25379 support returning on stack yet. */
25380 gcc_assert (REG_P (result_rtl
));
25381 to_clear_return_mask
25382 = compute_not_to_clear_mask (result_type
, result_rtl
, 0,
25383 &padding_bits_to_clear
);
25384 if (to_clear_return_mask
)
25386 gcc_assert ((unsigned) maxregno
< sizeof (long long) * __CHAR_BIT__
);
25387 for (regno
= R0_REGNUM
; regno
<= maxregno
; regno
++)
25389 if (to_clear_return_mask
& (1ULL << regno
))
25390 bitmap_clear_bit (to_clear_bitmap
, regno
);
25395 if (padding_bits_to_clear
!= 0)
25397 int to_clear_bitmap_size
= SBITMAP_SIZE ((sbitmap
) to_clear_bitmap
);
25398 auto_sbitmap
to_clear_arg_regs_bitmap (to_clear_bitmap_size
);
25400 /* Padding_bits_to_clear is not 0 so we know we are dealing with
25401 returning a composite type, which only uses r0. Let's make sure that
25402 r1-r3 is cleared too. */
25403 bitmap_clear (to_clear_arg_regs_bitmap
);
25404 bitmap_set_range (to_clear_arg_regs_bitmap
, R1_REGNUM
, NUM_ARG_REGS
- 1);
25405 gcc_assert (bitmap_subset_p (to_clear_arg_regs_bitmap
, to_clear_bitmap
));
25408 /* Clear full registers that leak before returning. */
25409 clearing_reg
= gen_rtx_REG (SImode
, TARGET_THUMB1
? R0_REGNUM
: LR_REGNUM
);
25410 r1_reg
= gen_rtx_REG (SImode
, R0_REGNUM
+ 1);
25411 cmse_clear_registers (to_clear_bitmap
, &padding_bits_to_clear
, 1, r1_reg
,
25415 /* Generate pattern *pop_multiple_with_stack_update_and_return if single
25416 POP instruction can be generated. LR should be replaced by PC. All
25417 the checks required are already done by USE_RETURN_INSN (). Hence,
25418 all we really need to check here is if single register is to be
25419 returned, or multiple register return. */
25421 thumb2_expand_return (bool simple_return
)
25424 unsigned long saved_regs_mask
;
25425 arm_stack_offsets
*offsets
;
25427 offsets
= arm_get_frame_offsets ();
25428 saved_regs_mask
= offsets
->saved_regs_mask
;
25430 for (i
= 0, num_regs
= 0; i
<= LAST_ARM_REGNUM
; i
++)
25431 if (saved_regs_mask
& (1 << i
))
25434 if (!simple_return
&& saved_regs_mask
)
25436 /* TODO: Verify that this path is never taken for cmse_nonsecure_entry
25437 functions or adapt code to handle according to ACLE. This path should
25438 not be reachable for cmse_nonsecure_entry functions though we prefer
25439 to assert it for now to ensure that future code changes do not silently
25440 change this behavior. */
25441 gcc_assert (!IS_CMSE_ENTRY (arm_current_func_type ()));
25444 rtx par
= gen_rtx_PARALLEL (VOIDmode
, rtvec_alloc (2));
25445 rtx reg
= gen_rtx_REG (SImode
, PC_REGNUM
);
25446 rtx addr
= gen_rtx_MEM (SImode
,
25447 gen_rtx_POST_INC (SImode
,
25448 stack_pointer_rtx
));
25449 set_mem_alias_set (addr
, get_frame_alias_set ());
25450 XVECEXP (par
, 0, 0) = ret_rtx
;
25451 XVECEXP (par
, 0, 1) = gen_rtx_SET (reg
, addr
);
25452 RTX_FRAME_RELATED_P (XVECEXP (par
, 0, 1)) = 1;
25453 emit_jump_insn (par
);
25457 saved_regs_mask
&= ~ (1 << LR_REGNUM
);
25458 saved_regs_mask
|= (1 << PC_REGNUM
);
25459 arm_emit_multi_reg_pop (saved_regs_mask
);
25464 if (IS_CMSE_ENTRY (arm_current_func_type ()))
25465 cmse_nonsecure_entry_clear_before_return ();
25466 emit_jump_insn (simple_return_rtx
);
25471 thumb1_expand_epilogue (void)
25473 HOST_WIDE_INT amount
;
25474 arm_stack_offsets
*offsets
;
25477 /* Naked functions don't have prologues. */
25478 if (IS_NAKED (arm_current_func_type ()))
25481 offsets
= arm_get_frame_offsets ();
25482 amount
= offsets
->outgoing_args
- offsets
->saved_regs
;
25484 if (frame_pointer_needed
)
25486 emit_insn (gen_movsi (stack_pointer_rtx
, hard_frame_pointer_rtx
));
25487 amount
= offsets
->locals_base
- offsets
->saved_regs
;
25489 amount
-= 4 * thumb1_extra_regs_pushed (offsets
, false);
25491 gcc_assert (amount
>= 0);
25494 emit_insn (gen_blockage ());
25497 emit_insn (gen_addsi3 (stack_pointer_rtx
, stack_pointer_rtx
,
25498 GEN_INT (amount
)));
25501 /* r3 is always free in the epilogue. */
25502 rtx reg
= gen_rtx_REG (SImode
, LAST_ARG_REGNUM
);
25504 emit_insn (gen_movsi (reg
, GEN_INT (amount
)));
25505 emit_insn (gen_addsi3 (stack_pointer_rtx
, stack_pointer_rtx
, reg
));
25509 /* Emit a USE (stack_pointer_rtx), so that
25510 the stack adjustment will not be deleted. */
25511 emit_insn (gen_force_register_use (stack_pointer_rtx
));
25513 if (crtl
->profile
|| !TARGET_SCHED_PROLOG
)
25514 emit_insn (gen_blockage ());
25516 /* Emit a clobber for each insn that will be restored in the epilogue,
25517 so that flow2 will get register lifetimes correct. */
25518 for (regno
= 0; regno
< 13; regno
++)
25519 if (df_regs_ever_live_p (regno
) && !call_used_regs
[regno
])
25520 emit_clobber (gen_rtx_REG (SImode
, regno
));
25522 if (! df_regs_ever_live_p (LR_REGNUM
))
25523 emit_use (gen_rtx_REG (SImode
, LR_REGNUM
));
25525 /* Clear all caller-saved regs that are not used to return. */
25526 if (IS_CMSE_ENTRY (arm_current_func_type ()))
25527 cmse_nonsecure_entry_clear_before_return ();
25530 /* Epilogue code for APCS frame. */
25532 arm_expand_epilogue_apcs_frame (bool really_return
)
25534 unsigned long func_type
;
25535 unsigned long saved_regs_mask
;
25538 int floats_from_frame
= 0;
25539 arm_stack_offsets
*offsets
;
25541 gcc_assert (TARGET_APCS_FRAME
&& frame_pointer_needed
&& TARGET_ARM
);
25542 func_type
= arm_current_func_type ();
25544 /* Get frame offsets for ARM. */
25545 offsets
= arm_get_frame_offsets ();
25546 saved_regs_mask
= offsets
->saved_regs_mask
;
25548 /* Find the offset of the floating-point save area in the frame. */
25550 = (offsets
->saved_args
25551 + arm_compute_static_chain_stack_bytes ()
25554 /* Compute how many core registers saved and how far away the floats are. */
25555 for (i
= 0; i
<= LAST_ARM_REGNUM
; i
++)
25556 if (saved_regs_mask
& (1 << i
))
25559 floats_from_frame
+= 4;
25562 if (TARGET_HARD_FLOAT
)
25565 rtx ip_rtx
= gen_rtx_REG (SImode
, IP_REGNUM
);
25567 /* The offset is from IP_REGNUM. */
25568 int saved_size
= arm_get_vfp_saved_size ();
25569 if (saved_size
> 0)
25572 floats_from_frame
+= saved_size
;
25573 insn
= emit_insn (gen_addsi3 (ip_rtx
,
25574 hard_frame_pointer_rtx
,
25575 GEN_INT (-floats_from_frame
)));
25576 arm_add_cfa_adjust_cfa_note (insn
, -floats_from_frame
,
25577 ip_rtx
, hard_frame_pointer_rtx
);
25580 /* Generate VFP register multi-pop. */
25581 start_reg
= FIRST_VFP_REGNUM
;
25583 for (i
= FIRST_VFP_REGNUM
; i
< LAST_VFP_REGNUM
; i
+= 2)
25584 /* Look for a case where a reg does not need restoring. */
25585 if ((!df_regs_ever_live_p (i
) || call_used_regs
[i
])
25586 && (!df_regs_ever_live_p (i
+ 1)
25587 || call_used_regs
[i
+ 1]))
25589 if (start_reg
!= i
)
25590 arm_emit_vfp_multi_reg_pop (start_reg
,
25591 (i
- start_reg
) / 2,
25592 gen_rtx_REG (SImode
,
25597 /* Restore the remaining regs that we have discovered (or possibly
25598 even all of them, if the conditional in the for loop never
25600 if (start_reg
!= i
)
25601 arm_emit_vfp_multi_reg_pop (start_reg
,
25602 (i
- start_reg
) / 2,
25603 gen_rtx_REG (SImode
, IP_REGNUM
));
25608 /* The frame pointer is guaranteed to be non-double-word aligned, as
25609 it is set to double-word-aligned old_stack_pointer - 4. */
25611 int lrm_count
= (num_regs
% 2) ? (num_regs
+ 2) : (num_regs
+ 1);
25613 for (i
= LAST_IWMMXT_REGNUM
; i
>= FIRST_IWMMXT_REGNUM
; i
--)
25614 if (df_regs_ever_live_p (i
) && !call_used_regs
[i
])
25616 rtx addr
= gen_frame_mem (V2SImode
,
25617 plus_constant (Pmode
, hard_frame_pointer_rtx
,
25619 insn
= emit_insn (gen_movsi (gen_rtx_REG (V2SImode
, i
), addr
));
25620 REG_NOTES (insn
) = alloc_reg_note (REG_CFA_RESTORE
,
25621 gen_rtx_REG (V2SImode
, i
),
25627 /* saved_regs_mask should contain IP which contains old stack pointer
25628 at the time of activation creation. Since SP and IP are adjacent registers,
25629 we can restore the value directly into SP. */
25630 gcc_assert (saved_regs_mask
& (1 << IP_REGNUM
));
25631 saved_regs_mask
&= ~(1 << IP_REGNUM
);
25632 saved_regs_mask
|= (1 << SP_REGNUM
);
25634 /* There are two registers left in saved_regs_mask - LR and PC. We
25635 only need to restore LR (the return address), but to
25636 save time we can load it directly into PC, unless we need a
25637 special function exit sequence, or we are not really returning. */
25639 && ARM_FUNC_TYPE (func_type
) == ARM_FT_NORMAL
25640 && !crtl
->calls_eh_return
)
25641 /* Delete LR from the register mask, so that LR on
25642 the stack is loaded into the PC in the register mask. */
25643 saved_regs_mask
&= ~(1 << LR_REGNUM
);
25645 saved_regs_mask
&= ~(1 << PC_REGNUM
);
25647 num_regs
= bit_count (saved_regs_mask
);
25648 if ((offsets
->outgoing_args
!= (1 + num_regs
)) || cfun
->calls_alloca
)
25651 emit_insn (gen_blockage ());
25652 /* Unwind the stack to just below the saved registers. */
25653 insn
= emit_insn (gen_addsi3 (stack_pointer_rtx
,
25654 hard_frame_pointer_rtx
,
25655 GEN_INT (- 4 * num_regs
)));
25657 arm_add_cfa_adjust_cfa_note (insn
, - 4 * num_regs
,
25658 stack_pointer_rtx
, hard_frame_pointer_rtx
);
25661 arm_emit_multi_reg_pop (saved_regs_mask
);
25663 if (IS_INTERRUPT (func_type
))
25665 /* Interrupt handlers will have pushed the
25666 IP onto the stack, so restore it now. */
25668 rtx addr
= gen_rtx_MEM (SImode
,
25669 gen_rtx_POST_INC (SImode
,
25670 stack_pointer_rtx
));
25671 set_mem_alias_set (addr
, get_frame_alias_set ());
25672 insn
= emit_insn (gen_movsi (gen_rtx_REG (SImode
, IP_REGNUM
), addr
));
25673 REG_NOTES (insn
) = alloc_reg_note (REG_CFA_RESTORE
,
25674 gen_rtx_REG (SImode
, IP_REGNUM
),
25678 if (!really_return
|| (saved_regs_mask
& (1 << PC_REGNUM
)))
25681 if (crtl
->calls_eh_return
)
25682 emit_insn (gen_addsi3 (stack_pointer_rtx
,
25684 gen_rtx_REG (SImode
, ARM_EH_STACKADJ_REGNUM
)));
25686 if (IS_STACKALIGN (func_type
))
25687 /* Restore the original stack pointer. Before prologue, the stack was
25688 realigned and the original stack pointer saved in r0. For details,
25689 see comment in arm_expand_prologue. */
25690 emit_insn (gen_movsi (stack_pointer_rtx
, gen_rtx_REG (SImode
, R0_REGNUM
)));
25692 emit_jump_insn (simple_return_rtx
);
25695 /* Generate RTL to represent ARM epilogue. Really_return is true if the
25696 function is not a sibcall. */
25698 arm_expand_epilogue (bool really_return
)
25700 unsigned long func_type
;
25701 unsigned long saved_regs_mask
;
25705 arm_stack_offsets
*offsets
;
25707 func_type
= arm_current_func_type ();
25709 /* Naked functions don't have epilogue. Hence, generate return pattern, and
25710 let output_return_instruction take care of instruction emission if any. */
25711 if (IS_NAKED (func_type
)
25712 || (IS_VOLATILE (func_type
) && TARGET_ABORT_NORETURN
))
25715 emit_jump_insn (simple_return_rtx
);
25719 /* If we are throwing an exception, then we really must be doing a
25720 return, so we can't tail-call. */
25721 gcc_assert (!crtl
->calls_eh_return
|| really_return
);
25723 if (TARGET_APCS_FRAME
&& frame_pointer_needed
&& TARGET_ARM
)
25725 arm_expand_epilogue_apcs_frame (really_return
);
25729 /* Get frame offsets for ARM. */
25730 offsets
= arm_get_frame_offsets ();
25731 saved_regs_mask
= offsets
->saved_regs_mask
;
25732 num_regs
= bit_count (saved_regs_mask
);
25734 if (frame_pointer_needed
)
25737 /* Restore stack pointer if necessary. */
25740 /* In ARM mode, frame pointer points to first saved register.
25741 Restore stack pointer to last saved register. */
25742 amount
= offsets
->frame
- offsets
->saved_regs
;
25744 /* Force out any pending memory operations that reference stacked data
25745 before stack de-allocation occurs. */
25746 emit_insn (gen_blockage ());
25747 insn
= emit_insn (gen_addsi3 (stack_pointer_rtx
,
25748 hard_frame_pointer_rtx
,
25749 GEN_INT (amount
)));
25750 arm_add_cfa_adjust_cfa_note (insn
, amount
,
25752 hard_frame_pointer_rtx
);
25754 /* Emit USE(stack_pointer_rtx) to ensure that stack adjustment is not
25756 emit_insn (gen_force_register_use (stack_pointer_rtx
));
25760 /* In Thumb-2 mode, the frame pointer points to the last saved
25762 amount
= offsets
->locals_base
- offsets
->saved_regs
;
25765 insn
= emit_insn (gen_addsi3 (hard_frame_pointer_rtx
,
25766 hard_frame_pointer_rtx
,
25767 GEN_INT (amount
)));
25768 arm_add_cfa_adjust_cfa_note (insn
, amount
,
25769 hard_frame_pointer_rtx
,
25770 hard_frame_pointer_rtx
);
25773 /* Force out any pending memory operations that reference stacked data
25774 before stack de-allocation occurs. */
25775 emit_insn (gen_blockage ());
25776 insn
= emit_insn (gen_movsi (stack_pointer_rtx
,
25777 hard_frame_pointer_rtx
));
25778 arm_add_cfa_adjust_cfa_note (insn
, 0,
25780 hard_frame_pointer_rtx
);
25781 /* Emit USE(stack_pointer_rtx) to ensure that stack adjustment is not
25783 emit_insn (gen_force_register_use (stack_pointer_rtx
));
25788 /* Pop off outgoing args and local frame to adjust stack pointer to
25789 last saved register. */
25790 amount
= offsets
->outgoing_args
- offsets
->saved_regs
;
25794 /* Force out any pending memory operations that reference stacked data
25795 before stack de-allocation occurs. */
25796 emit_insn (gen_blockage ());
25797 tmp
= emit_insn (gen_addsi3 (stack_pointer_rtx
,
25799 GEN_INT (amount
)));
25800 arm_add_cfa_adjust_cfa_note (tmp
, amount
,
25801 stack_pointer_rtx
, stack_pointer_rtx
);
25802 /* Emit USE(stack_pointer_rtx) to ensure that stack adjustment is
25804 emit_insn (gen_force_register_use (stack_pointer_rtx
));
25808 if (TARGET_HARD_FLOAT
)
25810 /* Generate VFP register multi-pop. */
25811 int end_reg
= LAST_VFP_REGNUM
+ 1;
25813 /* Scan the registers in reverse order. We need to match
25814 any groupings made in the prologue and generate matching
25815 vldm operations. The need to match groups is because,
25816 unlike pop, vldm can only do consecutive regs. */
25817 for (i
= LAST_VFP_REGNUM
- 1; i
>= FIRST_VFP_REGNUM
; i
-= 2)
25818 /* Look for a case where a reg does not need restoring. */
25819 if ((!df_regs_ever_live_p (i
) || call_used_regs
[i
])
25820 && (!df_regs_ever_live_p (i
+ 1)
25821 || call_used_regs
[i
+ 1]))
25823 /* Restore the regs discovered so far (from reg+2 to
25825 if (end_reg
> i
+ 2)
25826 arm_emit_vfp_multi_reg_pop (i
+ 2,
25827 (end_reg
- (i
+ 2)) / 2,
25828 stack_pointer_rtx
);
25832 /* Restore the remaining regs that we have discovered (or possibly
25833 even all of them, if the conditional in the for loop never
25835 if (end_reg
> i
+ 2)
25836 arm_emit_vfp_multi_reg_pop (i
+ 2,
25837 (end_reg
- (i
+ 2)) / 2,
25838 stack_pointer_rtx
);
25842 for (i
= FIRST_IWMMXT_REGNUM
; i
<= LAST_IWMMXT_REGNUM
; i
++)
25843 if (df_regs_ever_live_p (i
) && !call_used_regs
[i
])
25846 rtx addr
= gen_rtx_MEM (V2SImode
,
25847 gen_rtx_POST_INC (SImode
,
25848 stack_pointer_rtx
));
25849 set_mem_alias_set (addr
, get_frame_alias_set ());
25850 insn
= emit_insn (gen_movsi (gen_rtx_REG (V2SImode
, i
), addr
));
25851 REG_NOTES (insn
) = alloc_reg_note (REG_CFA_RESTORE
,
25852 gen_rtx_REG (V2SImode
, i
),
25854 arm_add_cfa_adjust_cfa_note (insn
, UNITS_PER_WORD
,
25855 stack_pointer_rtx
, stack_pointer_rtx
);
25858 if (saved_regs_mask
)
25861 bool return_in_pc
= false;
25863 if (ARM_FUNC_TYPE (func_type
) != ARM_FT_INTERWORKED
25864 && (TARGET_ARM
|| ARM_FUNC_TYPE (func_type
) == ARM_FT_NORMAL
)
25865 && !IS_CMSE_ENTRY (func_type
)
25866 && !IS_STACKALIGN (func_type
)
25868 && crtl
->args
.pretend_args_size
== 0
25869 && saved_regs_mask
& (1 << LR_REGNUM
)
25870 && !crtl
->calls_eh_return
)
25872 saved_regs_mask
&= ~(1 << LR_REGNUM
);
25873 saved_regs_mask
|= (1 << PC_REGNUM
);
25874 return_in_pc
= true;
25877 if (num_regs
== 1 && (!IS_INTERRUPT (func_type
) || !return_in_pc
))
25879 for (i
= 0; i
<= LAST_ARM_REGNUM
; i
++)
25880 if (saved_regs_mask
& (1 << i
))
25882 rtx addr
= gen_rtx_MEM (SImode
,
25883 gen_rtx_POST_INC (SImode
,
25884 stack_pointer_rtx
));
25885 set_mem_alias_set (addr
, get_frame_alias_set ());
25887 if (i
== PC_REGNUM
)
25889 insn
= gen_rtx_PARALLEL (VOIDmode
, rtvec_alloc (2));
25890 XVECEXP (insn
, 0, 0) = ret_rtx
;
25891 XVECEXP (insn
, 0, 1) = gen_rtx_SET (gen_rtx_REG (SImode
, i
),
25893 RTX_FRAME_RELATED_P (XVECEXP (insn
, 0, 1)) = 1;
25894 insn
= emit_jump_insn (insn
);
25898 insn
= emit_insn (gen_movsi (gen_rtx_REG (SImode
, i
),
25900 REG_NOTES (insn
) = alloc_reg_note (REG_CFA_RESTORE
,
25901 gen_rtx_REG (SImode
, i
),
25903 arm_add_cfa_adjust_cfa_note (insn
, UNITS_PER_WORD
,
25905 stack_pointer_rtx
);
25912 && current_tune
->prefer_ldrd_strd
25913 && !optimize_function_for_size_p (cfun
))
25916 thumb2_emit_ldrd_pop (saved_regs_mask
);
25917 else if (TARGET_ARM
&& !IS_INTERRUPT (func_type
))
25918 arm_emit_ldrd_pop (saved_regs_mask
);
25920 arm_emit_multi_reg_pop (saved_regs_mask
);
25923 arm_emit_multi_reg_pop (saved_regs_mask
);
25931 = crtl
->args
.pretend_args_size
+ arm_compute_static_chain_stack_bytes();
25935 rtx dwarf
= NULL_RTX
;
25937 emit_insn (gen_addsi3 (stack_pointer_rtx
,
25939 GEN_INT (amount
)));
25941 RTX_FRAME_RELATED_P (tmp
) = 1;
25943 if (cfun
->machine
->uses_anonymous_args
)
25945 /* Restore pretend args. Refer arm_expand_prologue on how to save
25946 pretend_args in stack. */
25947 int num_regs
= crtl
->args
.pretend_args_size
/ 4;
25948 saved_regs_mask
= (0xf0 >> num_regs
) & 0xf;
25949 for (j
= 0, i
= 0; j
< num_regs
; i
++)
25950 if (saved_regs_mask
& (1 << i
))
25952 rtx reg
= gen_rtx_REG (SImode
, i
);
25953 dwarf
= alloc_reg_note (REG_CFA_RESTORE
, reg
, dwarf
);
25956 REG_NOTES (tmp
) = dwarf
;
25958 arm_add_cfa_adjust_cfa_note (tmp
, amount
,
25959 stack_pointer_rtx
, stack_pointer_rtx
);
25962 /* Clear all caller-saved regs that are not used to return. */
25963 if (IS_CMSE_ENTRY (arm_current_func_type ()))
25965 /* CMSE_ENTRY always returns. */
25966 gcc_assert (really_return
);
25967 cmse_nonsecure_entry_clear_before_return ();
25970 if (!really_return
)
25973 if (crtl
->calls_eh_return
)
25974 emit_insn (gen_addsi3 (stack_pointer_rtx
,
25976 gen_rtx_REG (SImode
, ARM_EH_STACKADJ_REGNUM
)));
25978 if (IS_STACKALIGN (func_type
))
25979 /* Restore the original stack pointer. Before prologue, the stack was
25980 realigned and the original stack pointer saved in r0. For details,
25981 see comment in arm_expand_prologue. */
25982 emit_insn (gen_movsi (stack_pointer_rtx
, gen_rtx_REG (SImode
, R0_REGNUM
)));
25984 emit_jump_insn (simple_return_rtx
);
25987 /* Implementation of insn prologue_thumb1_interwork. This is the first
25988 "instruction" of a function called in ARM mode. Swap to thumb mode. */
25991 thumb1_output_interwork (void)
25994 FILE *f
= asm_out_file
;
25996 gcc_assert (MEM_P (DECL_RTL (current_function_decl
)));
25997 gcc_assert (GET_CODE (XEXP (DECL_RTL (current_function_decl
), 0))
25999 name
= XSTR (XEXP (DECL_RTL (current_function_decl
), 0), 0);
26001 /* Generate code sequence to switch us into Thumb mode. */
26002 /* The .code 32 directive has already been emitted by
26003 ASM_DECLARE_FUNCTION_NAME. */
26004 asm_fprintf (f
, "\torr\t%r, %r, #1\n", IP_REGNUM
, PC_REGNUM
);
26005 asm_fprintf (f
, "\tbx\t%r\n", IP_REGNUM
);
26007 /* Generate a label, so that the debugger will notice the
26008 change in instruction sets. This label is also used by
26009 the assembler to bypass the ARM code when this function
26010 is called from a Thumb encoded function elsewhere in the
26011 same file. Hence the definition of STUB_NAME here must
26012 agree with the definition in gas/config/tc-arm.c. */
26014 #define STUB_NAME ".real_start_of"
26016 fprintf (f
, "\t.code\t16\n");
26018 if (arm_dllexport_name_p (name
))
26019 name
= arm_strip_name_encoding (name
);
26021 asm_fprintf (f
, "\t.globl %s%U%s\n", STUB_NAME
, name
);
26022 fprintf (f
, "\t.thumb_func\n");
26023 asm_fprintf (f
, "%s%U%s:\n", STUB_NAME
, name
);
26028 /* Handle the case of a double word load into a low register from
26029 a computed memory address. The computed address may involve a
26030 register which is overwritten by the load. */
26032 thumb_load_double_from_address (rtx
*operands
)
26040 gcc_assert (REG_P (operands
[0]));
26041 gcc_assert (MEM_P (operands
[1]));
26043 /* Get the memory address. */
26044 addr
= XEXP (operands
[1], 0);
26046 /* Work out how the memory address is computed. */
26047 switch (GET_CODE (addr
))
26050 operands
[2] = adjust_address (operands
[1], SImode
, 4);
26052 if (REGNO (operands
[0]) == REGNO (addr
))
26054 output_asm_insn ("ldr\t%H0, %2", operands
);
26055 output_asm_insn ("ldr\t%0, %1", operands
);
26059 output_asm_insn ("ldr\t%0, %1", operands
);
26060 output_asm_insn ("ldr\t%H0, %2", operands
);
26065 /* Compute <address> + 4 for the high order load. */
26066 operands
[2] = adjust_address (operands
[1], SImode
, 4);
26068 output_asm_insn ("ldr\t%0, %1", operands
);
26069 output_asm_insn ("ldr\t%H0, %2", operands
);
26073 arg1
= XEXP (addr
, 0);
26074 arg2
= XEXP (addr
, 1);
26076 if (CONSTANT_P (arg1
))
26077 base
= arg2
, offset
= arg1
;
26079 base
= arg1
, offset
= arg2
;
26081 gcc_assert (REG_P (base
));
26083 /* Catch the case of <address> = <reg> + <reg> */
26084 if (REG_P (offset
))
26086 int reg_offset
= REGNO (offset
);
26087 int reg_base
= REGNO (base
);
26088 int reg_dest
= REGNO (operands
[0]);
26090 /* Add the base and offset registers together into the
26091 higher destination register. */
26092 asm_fprintf (asm_out_file
, "\tadd\t%r, %r, %r",
26093 reg_dest
+ 1, reg_base
, reg_offset
);
26095 /* Load the lower destination register from the address in
26096 the higher destination register. */
26097 asm_fprintf (asm_out_file
, "\tldr\t%r, [%r, #0]",
26098 reg_dest
, reg_dest
+ 1);
26100 /* Load the higher destination register from its own address
26102 asm_fprintf (asm_out_file
, "\tldr\t%r, [%r, #4]",
26103 reg_dest
+ 1, reg_dest
+ 1);
26107 /* Compute <address> + 4 for the high order load. */
26108 operands
[2] = adjust_address (operands
[1], SImode
, 4);
26110 /* If the computed address is held in the low order register
26111 then load the high order register first, otherwise always
26112 load the low order register first. */
26113 if (REGNO (operands
[0]) == REGNO (base
))
26115 output_asm_insn ("ldr\t%H0, %2", operands
);
26116 output_asm_insn ("ldr\t%0, %1", operands
);
26120 output_asm_insn ("ldr\t%0, %1", operands
);
26121 output_asm_insn ("ldr\t%H0, %2", operands
);
26127 /* With no registers to worry about we can just load the value
26129 operands
[2] = adjust_address (operands
[1], SImode
, 4);
26131 output_asm_insn ("ldr\t%H0, %2", operands
);
26132 output_asm_insn ("ldr\t%0, %1", operands
);
26136 gcc_unreachable ();
26143 thumb_output_move_mem_multiple (int n
, rtx
*operands
)
26148 if (REGNO (operands
[4]) > REGNO (operands
[5]))
26149 std::swap (operands
[4], operands
[5]);
26151 output_asm_insn ("ldmia\t%1!, {%4, %5}", operands
);
26152 output_asm_insn ("stmia\t%0!, {%4, %5}", operands
);
26156 if (REGNO (operands
[4]) > REGNO (operands
[5]))
26157 std::swap (operands
[4], operands
[5]);
26158 if (REGNO (operands
[5]) > REGNO (operands
[6]))
26159 std::swap (operands
[5], operands
[6]);
26160 if (REGNO (operands
[4]) > REGNO (operands
[5]))
26161 std::swap (operands
[4], operands
[5]);
26163 output_asm_insn ("ldmia\t%1!, {%4, %5, %6}", operands
);
26164 output_asm_insn ("stmia\t%0!, {%4, %5, %6}", operands
);
26168 gcc_unreachable ();
26174 /* Output a call-via instruction for thumb state. */
26176 thumb_call_via_reg (rtx reg
)
26178 int regno
= REGNO (reg
);
26181 gcc_assert (regno
< LR_REGNUM
);
26183 /* If we are in the normal text section we can use a single instance
26184 per compilation unit. If we are doing function sections, then we need
26185 an entry per section, since we can't rely on reachability. */
26186 if (in_section
== text_section
)
26188 thumb_call_reg_needed
= 1;
26190 if (thumb_call_via_label
[regno
] == NULL
)
26191 thumb_call_via_label
[regno
] = gen_label_rtx ();
26192 labelp
= thumb_call_via_label
+ regno
;
26196 if (cfun
->machine
->call_via
[regno
] == NULL
)
26197 cfun
->machine
->call_via
[regno
] = gen_label_rtx ();
26198 labelp
= cfun
->machine
->call_via
+ regno
;
26201 output_asm_insn ("bl\t%a0", labelp
);
26205 /* Routines for generating rtl. */
26207 thumb_expand_movmemqi (rtx
*operands
)
26209 rtx out
= copy_to_mode_reg (SImode
, XEXP (operands
[0], 0));
26210 rtx in
= copy_to_mode_reg (SImode
, XEXP (operands
[1], 0));
26211 HOST_WIDE_INT len
= INTVAL (operands
[2]);
26212 HOST_WIDE_INT offset
= 0;
26216 emit_insn (gen_movmem12b (out
, in
, out
, in
));
26222 emit_insn (gen_movmem8b (out
, in
, out
, in
));
26228 rtx reg
= gen_reg_rtx (SImode
);
26229 emit_insn (gen_movsi (reg
, gen_rtx_MEM (SImode
, in
)));
26230 emit_insn (gen_movsi (gen_rtx_MEM (SImode
, out
), reg
));
26237 rtx reg
= gen_reg_rtx (HImode
);
26238 emit_insn (gen_movhi (reg
, gen_rtx_MEM (HImode
,
26239 plus_constant (Pmode
, in
,
26241 emit_insn (gen_movhi (gen_rtx_MEM (HImode
, plus_constant (Pmode
, out
,
26250 rtx reg
= gen_reg_rtx (QImode
);
26251 emit_insn (gen_movqi (reg
, gen_rtx_MEM (QImode
,
26252 plus_constant (Pmode
, in
,
26254 emit_insn (gen_movqi (gen_rtx_MEM (QImode
, plus_constant (Pmode
, out
,
26261 thumb_reload_out_hi (rtx
*operands
)
26263 emit_insn (gen_thumb_movhi_clobber (operands
[0], operands
[1], operands
[2]));
26266 /* Return the length of a function name prefix
26267 that starts with the character 'c'. */
26269 arm_get_strip_length (int c
)
26273 ARM_NAME_ENCODING_LENGTHS
26278 /* Return a pointer to a function's name with any
26279 and all prefix encodings stripped from it. */
26281 arm_strip_name_encoding (const char *name
)
26285 while ((skip
= arm_get_strip_length (* name
)))
26291 /* If there is a '*' anywhere in the name's prefix, then
26292 emit the stripped name verbatim, otherwise prepend an
26293 underscore if leading underscores are being used. */
26295 arm_asm_output_labelref (FILE *stream
, const char *name
)
26300 while ((skip
= arm_get_strip_length (* name
)))
26302 verbatim
|= (*name
== '*');
26307 fputs (name
, stream
);
26309 asm_fprintf (stream
, "%U%s", name
);
26312 /* This function is used to emit an EABI tag and its associated value.
26313 We emit the numerical value of the tag in case the assembler does not
26314 support textual tags. (Eg gas prior to 2.20). If requested we include
26315 the tag name in a comment so that anyone reading the assembler output
26316 will know which tag is being set.
26318 This function is not static because arm-c.c needs it too. */
26321 arm_emit_eabi_attribute (const char *name
, int num
, int val
)
26323 asm_fprintf (asm_out_file
, "\t.eabi_attribute %d, %d", num
, val
);
26324 if (flag_verbose_asm
|| flag_debug_asm
)
26325 asm_fprintf (asm_out_file
, "\t%s %s", ASM_COMMENT_START
, name
);
26326 asm_fprintf (asm_out_file
, "\n");
26329 /* This function is used to print CPU tuning information as comment
26330 in assembler file. Pointers are not printed for now. */
26333 arm_print_tune_info (void)
26335 asm_fprintf (asm_out_file
, "\t" ASM_COMMENT_START
".tune parameters\n");
26336 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
"constant_limit:\t%d\n",
26337 current_tune
->constant_limit
);
26338 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
26339 "max_insns_skipped:\t%d\n", current_tune
->max_insns_skipped
);
26340 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
26341 "prefetch.num_slots:\t%d\n", current_tune
->prefetch
.num_slots
);
26342 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
26343 "prefetch.l1_cache_size:\t%d\n",
26344 current_tune
->prefetch
.l1_cache_size
);
26345 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
26346 "prefetch.l1_cache_line_size:\t%d\n",
26347 current_tune
->prefetch
.l1_cache_line_size
);
26348 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
26349 "prefer_constant_pool:\t%d\n",
26350 (int) current_tune
->prefer_constant_pool
);
26351 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
26352 "branch_cost:\t(s:speed, p:predictable)\n");
26353 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
"\t\ts&p\tcost\n");
26354 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
"\t\t00\t%d\n",
26355 current_tune
->branch_cost (false, false));
26356 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
"\t\t01\t%d\n",
26357 current_tune
->branch_cost (false, true));
26358 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
"\t\t10\t%d\n",
26359 current_tune
->branch_cost (true, false));
26360 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
"\t\t11\t%d\n",
26361 current_tune
->branch_cost (true, true));
26362 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
26363 "prefer_ldrd_strd:\t%d\n",
26364 (int) current_tune
->prefer_ldrd_strd
);
26365 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
26366 "logical_op_non_short_circuit:\t[%d,%d]\n",
26367 (int) current_tune
->logical_op_non_short_circuit_thumb
,
26368 (int) current_tune
->logical_op_non_short_circuit_arm
);
26369 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
26370 "prefer_neon_for_64bits:\t%d\n",
26371 (int) current_tune
->prefer_neon_for_64bits
);
26372 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
26373 "disparage_flag_setting_t16_encodings:\t%d\n",
26374 (int) current_tune
->disparage_flag_setting_t16_encodings
);
26375 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
26376 "string_ops_prefer_neon:\t%d\n",
26377 (int) current_tune
->string_ops_prefer_neon
);
26378 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
26379 "max_insns_inline_memset:\t%d\n",
26380 current_tune
->max_insns_inline_memset
);
26381 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
"fusible_ops:\t%u\n",
26382 current_tune
->fusible_ops
);
26383 asm_fprintf (asm_out_file
, "\t\t" ASM_COMMENT_START
"sched_autopref:\t%d\n",
26384 (int) current_tune
->sched_autopref
);
26387 /* Print .arch and .arch_extension directives corresponding to the
26388 current architecture configuration. */
26390 arm_print_asm_arch_directives ()
26392 const arch_option
*arch
26393 = arm_parse_arch_option_name (all_architectures
, "-march",
26394 arm_active_target
.arch_name
);
26395 auto_sbitmap
opt_bits (isa_num_bits
);
26399 asm_fprintf (asm_out_file
, "\t.arch %s\n", arm_active_target
.arch_name
);
26400 arm_last_printed_arch_string
= arm_active_target
.arch_name
;
26401 if (!arch
->common
.extensions
)
26404 for (const struct cpu_arch_extension
*opt
= arch
->common
.extensions
;
26410 arm_initialize_isa (opt_bits
, opt
->isa_bits
);
26412 /* If every feature bit of this option is set in the target
26413 ISA specification, print out the option name. However,
26414 don't print anything if all the bits are part of the
26415 FPU specification. */
26416 if (bitmap_subset_p (opt_bits
, arm_active_target
.isa
)
26417 && !bitmap_subset_p (opt_bits
, isa_all_fpubits
))
26418 asm_fprintf (asm_out_file
, "\t.arch_extension %s\n", opt
->name
);
26424 arm_file_start (void)
26430 /* We don't have a specified CPU. Use the architecture to
26433 Note: it might be better to do this unconditionally, then the
26434 assembler would not need to know about all new CPU names as
26436 if (!arm_active_target
.core_name
)
26438 /* armv7ve doesn't support any extensions. */
26439 if (strcmp (arm_active_target
.arch_name
, "armv7ve") == 0)
26441 /* Keep backward compatability for assemblers
26442 which don't support armv7ve. */
26443 asm_fprintf (asm_out_file
, "\t.arch armv7-a\n");
26444 asm_fprintf (asm_out_file
, "\t.arch_extension virt\n");
26445 asm_fprintf (asm_out_file
, "\t.arch_extension idiv\n");
26446 asm_fprintf (asm_out_file
, "\t.arch_extension sec\n");
26447 asm_fprintf (asm_out_file
, "\t.arch_extension mp\n");
26448 arm_last_printed_arch_string
= "armv7ve";
26451 arm_print_asm_arch_directives ();
26453 else if (strncmp (arm_active_target
.core_name
, "generic", 7) == 0)
26455 asm_fprintf (asm_out_file
, "\t.arch %s\n",
26456 arm_active_target
.core_name
+ 8);
26457 arm_last_printed_arch_string
= arm_active_target
.core_name
+ 8;
26461 const char* truncated_name
26462 = arm_rewrite_selected_cpu (arm_active_target
.core_name
);
26463 asm_fprintf (asm_out_file
, "\t.cpu %s\n", truncated_name
);
26466 if (print_tune_info
)
26467 arm_print_tune_info ();
26469 if (! TARGET_SOFT_FLOAT
)
26471 if (TARGET_HARD_FLOAT
&& TARGET_VFP_SINGLE
)
26472 arm_emit_eabi_attribute ("Tag_ABI_HardFP_use", 27, 1);
26474 if (TARGET_HARD_FLOAT_ABI
)
26475 arm_emit_eabi_attribute ("Tag_ABI_VFP_args", 28, 1);
26478 /* Some of these attributes only apply when the corresponding features
26479 are used. However we don't have any easy way of figuring this out.
26480 Conservatively record the setting that would have been used. */
26482 if (flag_rounding_math
)
26483 arm_emit_eabi_attribute ("Tag_ABI_FP_rounding", 19, 1);
26485 if (!flag_unsafe_math_optimizations
)
26487 arm_emit_eabi_attribute ("Tag_ABI_FP_denormal", 20, 1);
26488 arm_emit_eabi_attribute ("Tag_ABI_FP_exceptions", 21, 1);
26490 if (flag_signaling_nans
)
26491 arm_emit_eabi_attribute ("Tag_ABI_FP_user_exceptions", 22, 1);
26493 arm_emit_eabi_attribute ("Tag_ABI_FP_number_model", 23,
26494 flag_finite_math_only
? 1 : 3);
26496 arm_emit_eabi_attribute ("Tag_ABI_align8_needed", 24, 1);
26497 arm_emit_eabi_attribute ("Tag_ABI_align8_preserved", 25, 1);
26498 arm_emit_eabi_attribute ("Tag_ABI_enum_size", 26,
26499 flag_short_enums
? 1 : 2);
26501 /* Tag_ABI_optimization_goals. */
26504 else if (optimize
>= 2)
26510 arm_emit_eabi_attribute ("Tag_ABI_optimization_goals", 30, val
);
26512 arm_emit_eabi_attribute ("Tag_CPU_unaligned_access", 34,
26515 if (arm_fp16_format
)
26516 arm_emit_eabi_attribute ("Tag_ABI_FP_16bit_format", 38,
26517 (int) arm_fp16_format
);
26519 if (arm_lang_output_object_attributes_hook
)
26520 arm_lang_output_object_attributes_hook();
26523 default_file_start ();
26527 arm_file_end (void)
26531 if (NEED_INDICATE_EXEC_STACK
)
26532 /* Add .note.GNU-stack. */
26533 file_end_indicate_exec_stack ();
26535 if (! thumb_call_reg_needed
)
26538 switch_to_section (text_section
);
26539 asm_fprintf (asm_out_file
, "\t.code 16\n");
26540 ASM_OUTPUT_ALIGN (asm_out_file
, 1);
26542 for (regno
= 0; regno
< LR_REGNUM
; regno
++)
26544 rtx label
= thumb_call_via_label
[regno
];
26548 targetm
.asm_out
.internal_label (asm_out_file
, "L",
26549 CODE_LABEL_NUMBER (label
));
26550 asm_fprintf (asm_out_file
, "\tbx\t%r\n", regno
);
26556 /* Symbols in the text segment can be accessed without indirecting via the
26557 constant pool; it may take an extra binary operation, but this is still
26558 faster than indirecting via memory. Don't do this when not optimizing,
26559 since we won't be calculating al of the offsets necessary to do this
26563 arm_encode_section_info (tree decl
, rtx rtl
, int first
)
26565 if (optimize
> 0 && TREE_CONSTANT (decl
))
26566 SYMBOL_REF_FLAG (XEXP (rtl
, 0)) = 1;
26568 default_encode_section_info (decl
, rtl
, first
);
26570 #endif /* !ARM_PE */
26573 arm_internal_label (FILE *stream
, const char *prefix
, unsigned long labelno
)
26575 if (arm_ccfsm_state
== 3 && (unsigned) arm_target_label
== labelno
26576 && !strcmp (prefix
, "L"))
26578 arm_ccfsm_state
= 0;
26579 arm_target_insn
= NULL
;
26581 default_internal_label (stream
, prefix
, labelno
);
26584 /* Output code to add DELTA to the first argument, and then jump
26585 to FUNCTION. Used for C++ multiple inheritance. */
26588 arm_thumb1_mi_thunk (FILE *file
, tree
, HOST_WIDE_INT delta
,
26589 HOST_WIDE_INT
, tree function
)
26591 static int thunk_label
= 0;
26594 int mi_delta
= delta
;
26595 const char *const mi_op
= mi_delta
< 0 ? "sub" : "add";
26597 int this_regno
= (aggregate_value_p (TREE_TYPE (TREE_TYPE (function
)), function
)
26600 mi_delta
= - mi_delta
;
26602 final_start_function (emit_barrier (), file
, 1);
26606 int labelno
= thunk_label
++;
26607 ASM_GENERATE_INTERNAL_LABEL (label
, "LTHUMBFUNC", labelno
);
26608 /* Thunks are entered in arm mode when available. */
26609 if (TARGET_THUMB1_ONLY
)
26611 /* push r3 so we can use it as a temporary. */
26612 /* TODO: Omit this save if r3 is not used. */
26613 fputs ("\tpush {r3}\n", file
);
26614 fputs ("\tldr\tr3, ", file
);
26618 fputs ("\tldr\tr12, ", file
);
26620 assemble_name (file
, label
);
26621 fputc ('\n', file
);
26624 /* If we are generating PIC, the ldr instruction below loads
26625 "(target - 7) - .LTHUNKPCn" into r12. The pc reads as
26626 the address of the add + 8, so we have:
26628 r12 = (target - 7) - .LTHUNKPCn + (.LTHUNKPCn + 8)
26631 Note that we have "+ 1" because some versions of GNU ld
26632 don't set the low bit of the result for R_ARM_REL32
26633 relocations against thumb function symbols.
26634 On ARMv6M this is +4, not +8. */
26635 ASM_GENERATE_INTERNAL_LABEL (labelpc
, "LTHUNKPC", labelno
);
26636 assemble_name (file
, labelpc
);
26637 fputs (":\n", file
);
26638 if (TARGET_THUMB1_ONLY
)
26640 /* This is 2 insns after the start of the thunk, so we know it
26641 is 4-byte aligned. */
26642 fputs ("\tadd\tr3, pc, r3\n", file
);
26643 fputs ("\tmov r12, r3\n", file
);
26646 fputs ("\tadd\tr12, pc, r12\n", file
);
26648 else if (TARGET_THUMB1_ONLY
)
26649 fputs ("\tmov r12, r3\n", file
);
26651 if (TARGET_THUMB1_ONLY
)
26653 if (mi_delta
> 255)
26655 fputs ("\tldr\tr3, ", file
);
26656 assemble_name (file
, label
);
26657 fputs ("+4\n", file
);
26658 asm_fprintf (file
, "\t%ss\t%r, %r, r3\n",
26659 mi_op
, this_regno
, this_regno
);
26661 else if (mi_delta
!= 0)
26663 /* Thumb1 unified syntax requires s suffix in instruction name when
26664 one of the operands is immediate. */
26665 asm_fprintf (file
, "\t%ss\t%r, %r, #%d\n",
26666 mi_op
, this_regno
, this_regno
,
26672 /* TODO: Use movw/movt for large constants when available. */
26673 while (mi_delta
!= 0)
26675 if ((mi_delta
& (3 << shift
)) == 0)
26679 asm_fprintf (file
, "\t%s\t%r, %r, #%d\n",
26680 mi_op
, this_regno
, this_regno
,
26681 mi_delta
& (0xff << shift
));
26682 mi_delta
&= ~(0xff << shift
);
26689 if (TARGET_THUMB1_ONLY
)
26690 fputs ("\tpop\t{r3}\n", file
);
26692 fprintf (file
, "\tbx\tr12\n");
26693 ASM_OUTPUT_ALIGN (file
, 2);
26694 assemble_name (file
, label
);
26695 fputs (":\n", file
);
26698 /* Output ".word .LTHUNKn-[3,7]-.LTHUNKPCn". */
26699 rtx tem
= XEXP (DECL_RTL (function
), 0);
26700 /* For TARGET_THUMB1_ONLY the thunk is in Thumb mode, so the PC
26701 pipeline offset is four rather than eight. Adjust the offset
26703 tem
= plus_constant (GET_MODE (tem
), tem
,
26704 TARGET_THUMB1_ONLY
? -3 : -7);
26705 tem
= gen_rtx_MINUS (GET_MODE (tem
),
26707 gen_rtx_SYMBOL_REF (Pmode
,
26708 ggc_strdup (labelpc
)));
26709 assemble_integer (tem
, 4, BITS_PER_WORD
, 1);
26712 /* Output ".word .LTHUNKn". */
26713 assemble_integer (XEXP (DECL_RTL (function
), 0), 4, BITS_PER_WORD
, 1);
26715 if (TARGET_THUMB1_ONLY
&& mi_delta
> 255)
26716 assemble_integer (GEN_INT(mi_delta
), 4, BITS_PER_WORD
, 1);
26720 fputs ("\tb\t", file
);
26721 assemble_name (file
, XSTR (XEXP (DECL_RTL (function
), 0), 0));
26722 if (NEED_PLT_RELOC
)
26723 fputs ("(PLT)", file
);
26724 fputc ('\n', file
);
26727 final_end_function ();
26730 /* MI thunk handling for TARGET_32BIT. */
26733 arm32_output_mi_thunk (FILE *file
, tree
, HOST_WIDE_INT delta
,
26734 HOST_WIDE_INT vcall_offset
, tree function
)
26736 const bool long_call_p
= arm_is_long_call_p (function
);
26738 /* On ARM, this_regno is R0 or R1 depending on
26739 whether the function returns an aggregate or not.
26741 int this_regno
= (aggregate_value_p (TREE_TYPE (TREE_TYPE (function
)),
26743 ? R1_REGNUM
: R0_REGNUM
);
26745 rtx temp
= gen_rtx_REG (Pmode
, IP_REGNUM
);
26746 rtx this_rtx
= gen_rtx_REG (Pmode
, this_regno
);
26747 reload_completed
= 1;
26748 emit_note (NOTE_INSN_PROLOGUE_END
);
26750 /* Add DELTA to THIS_RTX. */
26752 arm_split_constant (PLUS
, Pmode
, NULL_RTX
,
26753 delta
, this_rtx
, this_rtx
, false);
26755 /* Add *(*THIS_RTX + VCALL_OFFSET) to THIS_RTX. */
26756 if (vcall_offset
!= 0)
26758 /* Load *THIS_RTX. */
26759 emit_move_insn (temp
, gen_rtx_MEM (Pmode
, this_rtx
));
26760 /* Compute *THIS_RTX + VCALL_OFFSET. */
26761 arm_split_constant (PLUS
, Pmode
, NULL_RTX
, vcall_offset
, temp
, temp
,
26763 /* Compute *(*THIS_RTX + VCALL_OFFSET). */
26764 emit_move_insn (temp
, gen_rtx_MEM (Pmode
, temp
));
26765 emit_insn (gen_add3_insn (this_rtx
, this_rtx
, temp
));
26768 /* Generate a tail call to the target function. */
26769 if (!TREE_USED (function
))
26771 assemble_external (function
);
26772 TREE_USED (function
) = 1;
26774 rtx funexp
= XEXP (DECL_RTL (function
), 0);
26777 emit_move_insn (temp
, funexp
);
26780 funexp
= gen_rtx_MEM (FUNCTION_MODE
, funexp
);
26781 rtx_insn
*insn
= emit_call_insn (gen_sibcall (funexp
, const0_rtx
, NULL_RTX
));
26782 SIBLING_CALL_P (insn
) = 1;
26785 /* Indirect calls require a bit of fixup in PIC mode. */
26788 split_all_insns_noflow ();
26792 insn
= get_insns ();
26793 shorten_branches (insn
);
26794 final_start_function (insn
, file
, 1);
26795 final (insn
, file
, 1);
26796 final_end_function ();
26798 /* Stop pretending this is a post-reload pass. */
26799 reload_completed
= 0;
26802 /* Output code to add DELTA to the first argument, and then jump
26803 to FUNCTION. Used for C++ multiple inheritance. */
26806 arm_output_mi_thunk (FILE *file
, tree thunk
, HOST_WIDE_INT delta
,
26807 HOST_WIDE_INT vcall_offset
, tree function
)
26810 arm32_output_mi_thunk (file
, thunk
, delta
, vcall_offset
, function
);
26812 arm_thumb1_mi_thunk (file
, thunk
, delta
, vcall_offset
, function
);
26816 arm_emit_vector_const (FILE *file
, rtx x
)
26819 const char * pattern
;
26821 gcc_assert (GET_CODE (x
) == CONST_VECTOR
);
26823 switch (GET_MODE (x
))
26825 case E_V2SImode
: pattern
= "%08x"; break;
26826 case E_V4HImode
: pattern
= "%04x"; break;
26827 case E_V8QImode
: pattern
= "%02x"; break;
26828 default: gcc_unreachable ();
26831 fprintf (file
, "0x");
26832 for (i
= CONST_VECTOR_NUNITS (x
); i
--;)
26836 element
= CONST_VECTOR_ELT (x
, i
);
26837 fprintf (file
, pattern
, INTVAL (element
));
26843 /* Emit a fp16 constant appropriately padded to occupy a 4-byte word.
26844 HFmode constant pool entries are actually loaded with ldr. */
26846 arm_emit_fp16_const (rtx c
)
26850 bits
= real_to_target (NULL
, CONST_DOUBLE_REAL_VALUE (c
), HFmode
);
26851 if (WORDS_BIG_ENDIAN
)
26852 assemble_zeros (2);
26853 assemble_integer (GEN_INT (bits
), 2, BITS_PER_WORD
, 1);
26854 if (!WORDS_BIG_ENDIAN
)
26855 assemble_zeros (2);
26859 arm_output_load_gr (rtx
*operands
)
26866 if (!MEM_P (operands
[1])
26867 || GET_CODE (sum
= XEXP (operands
[1], 0)) != PLUS
26868 || !REG_P (reg
= XEXP (sum
, 0))
26869 || !CONST_INT_P (offset
= XEXP (sum
, 1))
26870 || ((INTVAL (offset
) < 1024) && (INTVAL (offset
) > -1024)))
26871 return "wldrw%?\t%0, %1";
26873 /* Fix up an out-of-range load of a GR register. */
26874 output_asm_insn ("str%?\t%0, [sp, #-4]!\t@ Start of GR load expansion", & reg
);
26875 wcgr
= operands
[0];
26877 output_asm_insn ("ldr%?\t%0, %1", operands
);
26879 operands
[0] = wcgr
;
26881 output_asm_insn ("tmcr%?\t%0, %1", operands
);
26882 output_asm_insn ("ldr%?\t%0, [sp], #4\t@ End of GR load expansion", & reg
);
26887 /* Worker function for TARGET_SETUP_INCOMING_VARARGS.
26889 On the ARM, PRETEND_SIZE is set in order to have the prologue push the last
26890 named arg and all anonymous args onto the stack.
26891 XXX I know the prologue shouldn't be pushing registers, but it is faster
26895 arm_setup_incoming_varargs (cumulative_args_t pcum_v
,
26899 int second_time ATTRIBUTE_UNUSED
)
26901 CUMULATIVE_ARGS
*pcum
= get_cumulative_args (pcum_v
);
26904 cfun
->machine
->uses_anonymous_args
= 1;
26905 if (pcum
->pcs_variant
<= ARM_PCS_AAPCS_LOCAL
)
26907 nregs
= pcum
->aapcs_ncrn
;
26910 int res
= arm_needs_doubleword_align (mode
, type
);
26911 if (res
< 0 && warn_psabi
)
26912 inform (input_location
, "parameter passing for argument of "
26913 "type %qT changed in GCC 7.1", type
);
26919 nregs
= pcum
->nregs
;
26921 if (nregs
< NUM_ARG_REGS
)
26922 *pretend_size
= (NUM_ARG_REGS
- nregs
) * UNITS_PER_WORD
;
26925 /* We can't rely on the caller doing the proper promotion when
26926 using APCS or ATPCS. */
26929 arm_promote_prototypes (const_tree t ATTRIBUTE_UNUSED
)
26931 return !TARGET_AAPCS_BASED
;
26934 static machine_mode
26935 arm_promote_function_mode (const_tree type ATTRIBUTE_UNUSED
,
26937 int *punsignedp ATTRIBUTE_UNUSED
,
26938 const_tree fntype ATTRIBUTE_UNUSED
,
26939 int for_return ATTRIBUTE_UNUSED
)
26941 if (GET_MODE_CLASS (mode
) == MODE_INT
26942 && GET_MODE_SIZE (mode
) < 4)
26950 arm_default_short_enums (void)
26952 return ARM_DEFAULT_SHORT_ENUMS
;
26956 /* AAPCS requires that anonymous bitfields affect structure alignment. */
26959 arm_align_anon_bitfield (void)
26961 return TARGET_AAPCS_BASED
;
26965 /* The generic C++ ABI says 64-bit (long long). The EABI says 32-bit. */
26968 arm_cxx_guard_type (void)
26970 return TARGET_AAPCS_BASED
? integer_type_node
: long_long_integer_type_node
;
26974 /* The EABI says test the least significant bit of a guard variable. */
26977 arm_cxx_guard_mask_bit (void)
26979 return TARGET_AAPCS_BASED
;
26983 /* The EABI specifies that all array cookies are 8 bytes long. */
26986 arm_get_cookie_size (tree type
)
26990 if (!TARGET_AAPCS_BASED
)
26991 return default_cxx_get_cookie_size (type
);
26993 size
= build_int_cst (sizetype
, 8);
26998 /* The EABI says that array cookies should also contain the element size. */
27001 arm_cookie_has_size (void)
27003 return TARGET_AAPCS_BASED
;
27007 /* The EABI says constructors and destructors should return a pointer to
27008 the object constructed/destroyed. */
27011 arm_cxx_cdtor_returns_this (void)
27013 return TARGET_AAPCS_BASED
;
27016 /* The EABI says that an inline function may never be the key
27020 arm_cxx_key_method_may_be_inline (void)
27022 return !TARGET_AAPCS_BASED
;
27026 arm_cxx_determine_class_data_visibility (tree decl
)
27028 if (!TARGET_AAPCS_BASED
27029 || !TARGET_DLLIMPORT_DECL_ATTRIBUTES
)
27032 /* In general, \S 3.2.5.5 of the ARM EABI requires that class data
27033 is exported. However, on systems without dynamic vague linkage,
27034 \S 3.2.5.6 says that COMDAT class data has hidden linkage. */
27035 if (!TARGET_ARM_DYNAMIC_VAGUE_LINKAGE_P
&& DECL_COMDAT (decl
))
27036 DECL_VISIBILITY (decl
) = VISIBILITY_HIDDEN
;
27038 DECL_VISIBILITY (decl
) = VISIBILITY_DEFAULT
;
27039 DECL_VISIBILITY_SPECIFIED (decl
) = 1;
27043 arm_cxx_class_data_always_comdat (void)
27045 /* \S 3.2.5.4 of the ARM C++ ABI says that class data only have
27046 vague linkage if the class has no key function. */
27047 return !TARGET_AAPCS_BASED
;
27051 /* The EABI says __aeabi_atexit should be used to register static
27055 arm_cxx_use_aeabi_atexit (void)
27057 return TARGET_AAPCS_BASED
;
27062 arm_set_return_address (rtx source
, rtx scratch
)
27064 arm_stack_offsets
*offsets
;
27065 HOST_WIDE_INT delta
;
27067 unsigned long saved_regs
;
27069 offsets
= arm_get_frame_offsets ();
27070 saved_regs
= offsets
->saved_regs_mask
;
27072 if ((saved_regs
& (1 << LR_REGNUM
)) == 0)
27073 emit_move_insn (gen_rtx_REG (Pmode
, LR_REGNUM
), source
);
27076 if (frame_pointer_needed
)
27077 addr
= plus_constant (Pmode
, hard_frame_pointer_rtx
, -4);
27080 /* LR will be the first saved register. */
27081 delta
= offsets
->outgoing_args
- (offsets
->frame
+ 4);
27086 emit_insn (gen_addsi3 (scratch
, stack_pointer_rtx
,
27087 GEN_INT (delta
& ~4095)));
27092 addr
= stack_pointer_rtx
;
27094 addr
= plus_constant (Pmode
, addr
, delta
);
27097 /* The store needs to be marked to prevent DSE from deleting
27098 it as dead if it is based on fp. */
27099 mem
= gen_frame_mem (Pmode
, addr
);
27100 MEM_VOLATILE_P (mem
) = true;
27101 emit_move_insn (mem
, source
);
27107 thumb_set_return_address (rtx source
, rtx scratch
)
27109 arm_stack_offsets
*offsets
;
27110 HOST_WIDE_INT delta
;
27111 HOST_WIDE_INT limit
;
27114 unsigned long mask
;
27118 offsets
= arm_get_frame_offsets ();
27119 mask
= offsets
->saved_regs_mask
;
27120 if (mask
& (1 << LR_REGNUM
))
27123 /* Find the saved regs. */
27124 if (frame_pointer_needed
)
27126 delta
= offsets
->soft_frame
- offsets
->saved_args
;
27127 reg
= THUMB_HARD_FRAME_POINTER_REGNUM
;
27133 delta
= offsets
->outgoing_args
- offsets
->saved_args
;
27136 /* Allow for the stack frame. */
27137 if (TARGET_THUMB1
&& TARGET_BACKTRACE
)
27139 /* The link register is always the first saved register. */
27142 /* Construct the address. */
27143 addr
= gen_rtx_REG (SImode
, reg
);
27146 emit_insn (gen_movsi (scratch
, GEN_INT (delta
)));
27147 emit_insn (gen_addsi3 (scratch
, scratch
, stack_pointer_rtx
));
27151 addr
= plus_constant (Pmode
, addr
, delta
);
27153 /* The store needs to be marked to prevent DSE from deleting
27154 it as dead if it is based on fp. */
27155 mem
= gen_frame_mem (Pmode
, addr
);
27156 MEM_VOLATILE_P (mem
) = true;
27157 emit_move_insn (mem
, source
);
27160 emit_move_insn (gen_rtx_REG (Pmode
, LR_REGNUM
), source
);
27163 /* Implements target hook vector_mode_supported_p. */
27165 arm_vector_mode_supported_p (machine_mode mode
)
27167 /* Neon also supports V2SImode, etc. listed in the clause below. */
27168 if (TARGET_NEON
&& (mode
== V2SFmode
|| mode
== V4SImode
|| mode
== V8HImode
27169 || mode
== V4HFmode
|| mode
== V16QImode
|| mode
== V4SFmode
27170 || mode
== V2DImode
|| mode
== V8HFmode
))
27173 if ((TARGET_NEON
|| TARGET_IWMMXT
)
27174 && ((mode
== V2SImode
)
27175 || (mode
== V4HImode
)
27176 || (mode
== V8QImode
)))
27179 if (TARGET_INT_SIMD
&& (mode
== V4UQQmode
|| mode
== V4QQmode
27180 || mode
== V2UHQmode
|| mode
== V2HQmode
|| mode
== V2UHAmode
27181 || mode
== V2HAmode
))
27187 /* Implements target hook array_mode_supported_p. */
27190 arm_array_mode_supported_p (machine_mode mode
,
27191 unsigned HOST_WIDE_INT nelems
)
27193 /* We don't want to enable interleaved loads and stores for BYTES_BIG_ENDIAN
27194 for now, as the lane-swapping logic needs to be extended in the expanders.
27195 See PR target/82518. */
27196 if (TARGET_NEON
&& !BYTES_BIG_ENDIAN
27197 && (VALID_NEON_DREG_MODE (mode
) || VALID_NEON_QREG_MODE (mode
))
27198 && (nelems
>= 2 && nelems
<= 4))
27204 /* Use the option -mvectorize-with-neon-double to override the use of quardword
27205 registers when autovectorizing for Neon, at least until multiple vector
27206 widths are supported properly by the middle-end. */
27208 static machine_mode
27209 arm_preferred_simd_mode (scalar_mode mode
)
27215 return TARGET_NEON_VECTORIZE_DOUBLE
? V2SFmode
: V4SFmode
;
27217 return TARGET_NEON_VECTORIZE_DOUBLE
? V2SImode
: V4SImode
;
27219 return TARGET_NEON_VECTORIZE_DOUBLE
? V4HImode
: V8HImode
;
27221 return TARGET_NEON_VECTORIZE_DOUBLE
? V8QImode
: V16QImode
;
27223 if (!TARGET_NEON_VECTORIZE_DOUBLE
)
27230 if (TARGET_REALLY_IWMMXT
)
27246 /* Implement TARGET_CLASS_LIKELY_SPILLED_P.
27248 We need to define this for LO_REGS on Thumb-1. Otherwise we can end up
27249 using r0-r4 for function arguments, r7 for the stack frame and don't have
27250 enough left over to do doubleword arithmetic. For Thumb-2 all the
27251 potentially problematic instructions accept high registers so this is not
27252 necessary. Care needs to be taken to avoid adding new Thumb-2 patterns
27253 that require many low registers. */
27255 arm_class_likely_spilled_p (reg_class_t rclass
)
27257 if ((TARGET_THUMB1
&& rclass
== LO_REGS
)
27258 || rclass
== CC_REG
)
27264 /* Implements target hook small_register_classes_for_mode_p. */
27266 arm_small_register_classes_for_mode_p (machine_mode mode ATTRIBUTE_UNUSED
)
27268 return TARGET_THUMB1
;
27271 /* Implement TARGET_SHIFT_TRUNCATION_MASK. SImode shifts use normal
27272 ARM insns and therefore guarantee that the shift count is modulo 256.
27273 DImode shifts (those implemented by lib1funcs.S or by optabs.c)
27274 guarantee no particular behavior for out-of-range counts. */
27276 static unsigned HOST_WIDE_INT
27277 arm_shift_truncation_mask (machine_mode mode
)
27279 return mode
== SImode
? 255 : 0;
27283 /* Map internal gcc register numbers to DWARF2 register numbers. */
27286 arm_dbx_register_number (unsigned int regno
)
27291 if (IS_VFP_REGNUM (regno
))
27293 /* See comment in arm_dwarf_register_span. */
27294 if (VFP_REGNO_OK_FOR_SINGLE (regno
))
27295 return 64 + regno
- FIRST_VFP_REGNUM
;
27297 return 256 + (regno
- FIRST_VFP_REGNUM
) / 2;
27300 if (IS_IWMMXT_GR_REGNUM (regno
))
27301 return 104 + regno
- FIRST_IWMMXT_GR_REGNUM
;
27303 if (IS_IWMMXT_REGNUM (regno
))
27304 return 112 + regno
- FIRST_IWMMXT_REGNUM
;
27306 return DWARF_FRAME_REGISTERS
;
27309 /* Dwarf models VFPv3 registers as 32 64-bit registers.
27310 GCC models tham as 64 32-bit registers, so we need to describe this to
27311 the DWARF generation code. Other registers can use the default. */
27313 arm_dwarf_register_span (rtx rtl
)
27321 regno
= REGNO (rtl
);
27322 if (!IS_VFP_REGNUM (regno
))
27325 /* XXX FIXME: The EABI defines two VFP register ranges:
27326 64-95: Legacy VFPv2 numbering for S0-S31 (obsolescent)
27328 The recommended encoding for S0-S31 is a DW_OP_bit_piece of the
27329 corresponding D register. Until GDB supports this, we shall use the
27330 legacy encodings. We also use these encodings for D0-D15 for
27331 compatibility with older debuggers. */
27332 mode
= GET_MODE (rtl
);
27333 if (GET_MODE_SIZE (mode
) < 8)
27336 if (VFP_REGNO_OK_FOR_SINGLE (regno
))
27338 nregs
= GET_MODE_SIZE (mode
) / 4;
27339 for (i
= 0; i
< nregs
; i
+= 2)
27340 if (TARGET_BIG_END
)
27342 parts
[i
] = gen_rtx_REG (SImode
, regno
+ i
+ 1);
27343 parts
[i
+ 1] = gen_rtx_REG (SImode
, regno
+ i
);
27347 parts
[i
] = gen_rtx_REG (SImode
, regno
+ i
);
27348 parts
[i
+ 1] = gen_rtx_REG (SImode
, regno
+ i
+ 1);
27353 nregs
= GET_MODE_SIZE (mode
) / 8;
27354 for (i
= 0; i
< nregs
; i
++)
27355 parts
[i
] = gen_rtx_REG (DImode
, regno
+ i
);
27358 return gen_rtx_PARALLEL (VOIDmode
, gen_rtvec_v (nregs
, parts
));
27361 #if ARM_UNWIND_INFO
27362 /* Emit unwind directives for a store-multiple instruction or stack pointer
27363 push during alignment.
27364 These should only ever be generated by the function prologue code, so
27365 expect them to have a particular form.
27366 The store-multiple instruction sometimes pushes pc as the last register,
27367 although it should not be tracked into unwind information, or for -Os
27368 sometimes pushes some dummy registers before first register that needs
27369 to be tracked in unwind information; such dummy registers are there just
27370 to avoid separate stack adjustment, and will not be restored in the
27374 arm_unwind_emit_sequence (FILE * asm_out_file
, rtx p
)
27377 HOST_WIDE_INT offset
;
27378 HOST_WIDE_INT nregs
;
27382 unsigned padfirst
= 0, padlast
= 0;
27385 e
= XVECEXP (p
, 0, 0);
27386 gcc_assert (GET_CODE (e
) == SET
);
27388 /* First insn will adjust the stack pointer. */
27389 gcc_assert (GET_CODE (e
) == SET
27390 && REG_P (SET_DEST (e
))
27391 && REGNO (SET_DEST (e
)) == SP_REGNUM
27392 && GET_CODE (SET_SRC (e
)) == PLUS
);
27394 offset
= -INTVAL (XEXP (SET_SRC (e
), 1));
27395 nregs
= XVECLEN (p
, 0) - 1;
27396 gcc_assert (nregs
);
27398 reg
= REGNO (SET_SRC (XVECEXP (p
, 0, 1)));
27401 /* For -Os dummy registers can be pushed at the beginning to
27402 avoid separate stack pointer adjustment. */
27403 e
= XVECEXP (p
, 0, 1);
27404 e
= XEXP (SET_DEST (e
), 0);
27405 if (GET_CODE (e
) == PLUS
)
27406 padfirst
= INTVAL (XEXP (e
, 1));
27407 gcc_assert (padfirst
== 0 || optimize_size
);
27408 /* The function prologue may also push pc, but not annotate it as it is
27409 never restored. We turn this into a stack pointer adjustment. */
27410 e
= XVECEXP (p
, 0, nregs
);
27411 e
= XEXP (SET_DEST (e
), 0);
27412 if (GET_CODE (e
) == PLUS
)
27413 padlast
= offset
- INTVAL (XEXP (e
, 1)) - 4;
27415 padlast
= offset
- 4;
27416 gcc_assert (padlast
== 0 || padlast
== 4);
27418 fprintf (asm_out_file
, "\t.pad #4\n");
27420 fprintf (asm_out_file
, "\t.save {");
27422 else if (IS_VFP_REGNUM (reg
))
27425 fprintf (asm_out_file
, "\t.vsave {");
27428 /* Unknown register type. */
27429 gcc_unreachable ();
27431 /* If the stack increment doesn't match the size of the saved registers,
27432 something has gone horribly wrong. */
27433 gcc_assert (offset
== padfirst
+ nregs
* reg_size
+ padlast
);
27437 /* The remaining insns will describe the stores. */
27438 for (i
= 1; i
<= nregs
; i
++)
27440 /* Expect (set (mem <addr>) (reg)).
27441 Where <addr> is (reg:SP) or (plus (reg:SP) (const_int)). */
27442 e
= XVECEXP (p
, 0, i
);
27443 gcc_assert (GET_CODE (e
) == SET
27444 && MEM_P (SET_DEST (e
))
27445 && REG_P (SET_SRC (e
)));
27447 reg
= REGNO (SET_SRC (e
));
27448 gcc_assert (reg
>= lastreg
);
27451 fprintf (asm_out_file
, ", ");
27452 /* We can't use %r for vfp because we need to use the
27453 double precision register names. */
27454 if (IS_VFP_REGNUM (reg
))
27455 asm_fprintf (asm_out_file
, "d%d", (reg
- FIRST_VFP_REGNUM
) / 2);
27457 asm_fprintf (asm_out_file
, "%r", reg
);
27461 /* Check that the addresses are consecutive. */
27462 e
= XEXP (SET_DEST (e
), 0);
27463 if (GET_CODE (e
) == PLUS
)
27464 gcc_assert (REG_P (XEXP (e
, 0))
27465 && REGNO (XEXP (e
, 0)) == SP_REGNUM
27466 && CONST_INT_P (XEXP (e
, 1))
27467 && offset
== INTVAL (XEXP (e
, 1)));
27471 && REGNO (e
) == SP_REGNUM
);
27472 offset
+= reg_size
;
27475 fprintf (asm_out_file
, "}\n");
27477 fprintf (asm_out_file
, "\t.pad #%d\n", padfirst
);
27480 /* Emit unwind directives for a SET. */
27483 arm_unwind_emit_set (FILE * asm_out_file
, rtx p
)
27491 switch (GET_CODE (e0
))
27494 /* Pushing a single register. */
27495 if (GET_CODE (XEXP (e0
, 0)) != PRE_DEC
27496 || !REG_P (XEXP (XEXP (e0
, 0), 0))
27497 || REGNO (XEXP (XEXP (e0
, 0), 0)) != SP_REGNUM
)
27500 asm_fprintf (asm_out_file
, "\t.save ");
27501 if (IS_VFP_REGNUM (REGNO (e1
)))
27502 asm_fprintf(asm_out_file
, "{d%d}\n",
27503 (REGNO (e1
) - FIRST_VFP_REGNUM
) / 2);
27505 asm_fprintf(asm_out_file
, "{%r}\n", REGNO (e1
));
27509 if (REGNO (e0
) == SP_REGNUM
)
27511 /* A stack increment. */
27512 if (GET_CODE (e1
) != PLUS
27513 || !REG_P (XEXP (e1
, 0))
27514 || REGNO (XEXP (e1
, 0)) != SP_REGNUM
27515 || !CONST_INT_P (XEXP (e1
, 1)))
27518 asm_fprintf (asm_out_file
, "\t.pad #%wd\n",
27519 -INTVAL (XEXP (e1
, 1)));
27521 else if (REGNO (e0
) == HARD_FRAME_POINTER_REGNUM
)
27523 HOST_WIDE_INT offset
;
27525 if (GET_CODE (e1
) == PLUS
)
27527 if (!REG_P (XEXP (e1
, 0))
27528 || !CONST_INT_P (XEXP (e1
, 1)))
27530 reg
= REGNO (XEXP (e1
, 0));
27531 offset
= INTVAL (XEXP (e1
, 1));
27532 asm_fprintf (asm_out_file
, "\t.setfp %r, %r, #%wd\n",
27533 HARD_FRAME_POINTER_REGNUM
, reg
,
27536 else if (REG_P (e1
))
27539 asm_fprintf (asm_out_file
, "\t.setfp %r, %r\n",
27540 HARD_FRAME_POINTER_REGNUM
, reg
);
27545 else if (REG_P (e1
) && REGNO (e1
) == SP_REGNUM
)
27547 /* Move from sp to reg. */
27548 asm_fprintf (asm_out_file
, "\t.movsp %r\n", REGNO (e0
));
27550 else if (GET_CODE (e1
) == PLUS
27551 && REG_P (XEXP (e1
, 0))
27552 && REGNO (XEXP (e1
, 0)) == SP_REGNUM
27553 && CONST_INT_P (XEXP (e1
, 1)))
27555 /* Set reg to offset from sp. */
27556 asm_fprintf (asm_out_file
, "\t.movsp %r, #%d\n",
27557 REGNO (e0
), (int)INTVAL(XEXP (e1
, 1)));
27569 /* Emit unwind directives for the given insn. */
27572 arm_unwind_emit (FILE * asm_out_file
, rtx_insn
*insn
)
27575 bool handled_one
= false;
27577 if (arm_except_unwind_info (&global_options
) != UI_TARGET
)
27580 if (!(flag_unwind_tables
|| crtl
->uses_eh_lsda
)
27581 && (TREE_NOTHROW (current_function_decl
)
27582 || crtl
->all_throwers_are_sibcalls
))
27585 if (NOTE_P (insn
) || !RTX_FRAME_RELATED_P (insn
))
27588 for (note
= REG_NOTES (insn
); note
; note
= XEXP (note
, 1))
27590 switch (REG_NOTE_KIND (note
))
27592 case REG_FRAME_RELATED_EXPR
:
27593 pat
= XEXP (note
, 0);
27596 case REG_CFA_REGISTER
:
27597 pat
= XEXP (note
, 0);
27600 pat
= PATTERN (insn
);
27601 if (GET_CODE (pat
) == PARALLEL
)
27602 pat
= XVECEXP (pat
, 0, 0);
27605 /* Only emitted for IS_STACKALIGN re-alignment. */
27610 src
= SET_SRC (pat
);
27611 dest
= SET_DEST (pat
);
27613 gcc_assert (src
== stack_pointer_rtx
);
27614 reg
= REGNO (dest
);
27615 asm_fprintf (asm_out_file
, "\t.unwind_raw 0, 0x%x @ vsp = r%d\n",
27618 handled_one
= true;
27621 /* The INSN is generated in epilogue. It is set as RTX_FRAME_RELATED_P
27622 to get correct dwarf information for shrink-wrap. We should not
27623 emit unwind information for it because these are used either for
27624 pretend arguments or notes to adjust sp and restore registers from
27626 case REG_CFA_DEF_CFA
:
27627 case REG_CFA_ADJUST_CFA
:
27628 case REG_CFA_RESTORE
:
27631 case REG_CFA_EXPRESSION
:
27632 case REG_CFA_OFFSET
:
27633 /* ??? Only handling here what we actually emit. */
27634 gcc_unreachable ();
27642 pat
= PATTERN (insn
);
27645 switch (GET_CODE (pat
))
27648 arm_unwind_emit_set (asm_out_file
, pat
);
27652 /* Store multiple. */
27653 arm_unwind_emit_sequence (asm_out_file
, pat
);
27662 /* Output a reference from a function exception table to the type_info
27663 object X. The EABI specifies that the symbol should be relocated by
27664 an R_ARM_TARGET2 relocation. */
27667 arm_output_ttype (rtx x
)
27669 fputs ("\t.word\t", asm_out_file
);
27670 output_addr_const (asm_out_file
, x
);
27671 /* Use special relocations for symbol references. */
27672 if (!CONST_INT_P (x
))
27673 fputs ("(TARGET2)", asm_out_file
);
27674 fputc ('\n', asm_out_file
);
27679 /* Implement TARGET_ASM_EMIT_EXCEPT_PERSONALITY. */
27682 arm_asm_emit_except_personality (rtx personality
)
27684 fputs ("\t.personality\t", asm_out_file
);
27685 output_addr_const (asm_out_file
, personality
);
27686 fputc ('\n', asm_out_file
);
27688 #endif /* ARM_UNWIND_INFO */
27690 /* Implement TARGET_ASM_INITIALIZE_SECTIONS. */
27693 arm_asm_init_sections (void)
27695 #if ARM_UNWIND_INFO
27696 exception_section
= get_unnamed_section (0, output_section_asm_op
,
27698 #endif /* ARM_UNWIND_INFO */
27700 #ifdef OBJECT_FORMAT_ELF
27701 if (target_pure_code
)
27702 text_section
->unnamed
.data
= "\t.section .text,\"0x20000006\",%progbits";
27706 /* Output unwind directives for the start/end of a function. */
27709 arm_output_fn_unwind (FILE * f
, bool prologue
)
27711 if (arm_except_unwind_info (&global_options
) != UI_TARGET
)
27715 fputs ("\t.fnstart\n", f
);
27718 /* If this function will never be unwound, then mark it as such.
27719 The came condition is used in arm_unwind_emit to suppress
27720 the frame annotations. */
27721 if (!(flag_unwind_tables
|| crtl
->uses_eh_lsda
)
27722 && (TREE_NOTHROW (current_function_decl
)
27723 || crtl
->all_throwers_are_sibcalls
))
27724 fputs("\t.cantunwind\n", f
);
27726 fputs ("\t.fnend\n", f
);
27731 arm_emit_tls_decoration (FILE *fp
, rtx x
)
27733 enum tls_reloc reloc
;
27736 val
= XVECEXP (x
, 0, 0);
27737 reloc
= (enum tls_reloc
) INTVAL (XVECEXP (x
, 0, 1));
27739 output_addr_const (fp
, val
);
27744 fputs ("(tlsgd)", fp
);
27747 fputs ("(tlsldm)", fp
);
27750 fputs ("(tlsldo)", fp
);
27753 fputs ("(gottpoff)", fp
);
27756 fputs ("(tpoff)", fp
);
27759 fputs ("(tlsdesc)", fp
);
27762 gcc_unreachable ();
27771 fputs (" + (. - ", fp
);
27772 output_addr_const (fp
, XVECEXP (x
, 0, 2));
27773 /* For DESCSEQ the 3rd operand encodes thumbness, and is added */
27774 fputs (reloc
== TLS_DESCSEQ
? " + " : " - ", fp
);
27775 output_addr_const (fp
, XVECEXP (x
, 0, 3));
27785 /* ARM implementation of TARGET_ASM_OUTPUT_DWARF_DTPREL. */
27788 arm_output_dwarf_dtprel (FILE *file
, int size
, rtx x
)
27790 gcc_assert (size
== 4);
27791 fputs ("\t.word\t", file
);
27792 output_addr_const (file
, x
);
27793 fputs ("(tlsldo)", file
);
27796 /* Implement TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA. */
27799 arm_output_addr_const_extra (FILE *fp
, rtx x
)
27801 if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLS
)
27802 return arm_emit_tls_decoration (fp
, x
);
27803 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_PIC_LABEL
)
27806 int labelno
= INTVAL (XVECEXP (x
, 0, 0));
27808 ASM_GENERATE_INTERNAL_LABEL (label
, "LPIC", labelno
);
27809 assemble_name_raw (fp
, label
);
27813 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_GOTSYM_OFF
)
27815 assemble_name (fp
, "_GLOBAL_OFFSET_TABLE_");
27819 output_addr_const (fp
, XVECEXP (x
, 0, 0));
27823 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_SYMBOL_OFFSET
)
27825 output_addr_const (fp
, XVECEXP (x
, 0, 0));
27829 output_addr_const (fp
, XVECEXP (x
, 0, 1));
27833 else if (GET_CODE (x
) == CONST_VECTOR
)
27834 return arm_emit_vector_const (fp
, x
);
27839 /* Output assembly for a shift instruction.
27840 SET_FLAGS determines how the instruction modifies the condition codes.
27841 0 - Do not set condition codes.
27842 1 - Set condition codes.
27843 2 - Use smallest instruction. */
27845 arm_output_shift(rtx
* operands
, int set_flags
)
27848 static const char flag_chars
[3] = {'?', '.', '!'};
27853 c
= flag_chars
[set_flags
];
27854 shift
= shift_op(operands
[3], &val
);
27858 operands
[2] = GEN_INT(val
);
27859 sprintf (pattern
, "%s%%%c\t%%0, %%1, %%2", shift
, c
);
27862 sprintf (pattern
, "mov%%%c\t%%0, %%1", c
);
27864 output_asm_insn (pattern
, operands
);
27868 /* Output assembly for a WMMX immediate shift instruction. */
27870 arm_output_iwmmxt_shift_immediate (const char *insn_name
, rtx
*operands
, bool wror_or_wsra
)
27872 int shift
= INTVAL (operands
[2]);
27874 machine_mode opmode
= GET_MODE (operands
[0]);
27876 gcc_assert (shift
>= 0);
27878 /* If the shift value in the register versions is > 63 (for D qualifier),
27879 31 (for W qualifier) or 15 (for H qualifier). */
27880 if (((opmode
== V4HImode
) && (shift
> 15))
27881 || ((opmode
== V2SImode
) && (shift
> 31))
27882 || ((opmode
== DImode
) && (shift
> 63)))
27886 sprintf (templ
, "%s\t%%0, %%1, #%d", insn_name
, 32);
27887 output_asm_insn (templ
, operands
);
27888 if (opmode
== DImode
)
27890 sprintf (templ
, "%s\t%%0, %%0, #%d", insn_name
, 32);
27891 output_asm_insn (templ
, operands
);
27896 /* The destination register will contain all zeros. */
27897 sprintf (templ
, "wzero\t%%0");
27898 output_asm_insn (templ
, operands
);
27903 if ((opmode
== DImode
) && (shift
> 32))
27905 sprintf (templ
, "%s\t%%0, %%1, #%d", insn_name
, 32);
27906 output_asm_insn (templ
, operands
);
27907 sprintf (templ
, "%s\t%%0, %%0, #%d", insn_name
, shift
- 32);
27908 output_asm_insn (templ
, operands
);
27912 sprintf (templ
, "%s\t%%0, %%1, #%d", insn_name
, shift
);
27913 output_asm_insn (templ
, operands
);
27918 /* Output assembly for a WMMX tinsr instruction. */
27920 arm_output_iwmmxt_tinsr (rtx
*operands
)
27922 int mask
= INTVAL (operands
[3]);
27925 int units
= mode_nunits
[GET_MODE (operands
[0])];
27926 gcc_assert ((mask
& (mask
- 1)) == 0);
27927 for (i
= 0; i
< units
; ++i
)
27929 if ((mask
& 0x01) == 1)
27935 gcc_assert (i
< units
);
27937 switch (GET_MODE (operands
[0]))
27940 sprintf (templ
, "tinsrb%%?\t%%0, %%2, #%d", i
);
27943 sprintf (templ
, "tinsrh%%?\t%%0, %%2, #%d", i
);
27946 sprintf (templ
, "tinsrw%%?\t%%0, %%2, #%d", i
);
27949 gcc_unreachable ();
27952 output_asm_insn (templ
, operands
);
27957 /* Output a Thumb-1 casesi dispatch sequence. */
27959 thumb1_output_casesi (rtx
*operands
)
27961 rtx diff_vec
= PATTERN (NEXT_INSN (as_a
<rtx_insn
*> (operands
[0])));
27963 gcc_assert (GET_CODE (diff_vec
) == ADDR_DIFF_VEC
);
27965 switch (GET_MODE(diff_vec
))
27968 return (ADDR_DIFF_VEC_FLAGS (diff_vec
).offset_unsigned
?
27969 "bl\t%___gnu_thumb1_case_uqi" : "bl\t%___gnu_thumb1_case_sqi");
27971 return (ADDR_DIFF_VEC_FLAGS (diff_vec
).offset_unsigned
?
27972 "bl\t%___gnu_thumb1_case_uhi" : "bl\t%___gnu_thumb1_case_shi");
27974 return "bl\t%___gnu_thumb1_case_si";
27976 gcc_unreachable ();
27980 /* Output a Thumb-2 casesi instruction. */
27982 thumb2_output_casesi (rtx
*operands
)
27984 rtx diff_vec
= PATTERN (NEXT_INSN (as_a
<rtx_insn
*> (operands
[2])));
27986 gcc_assert (GET_CODE (diff_vec
) == ADDR_DIFF_VEC
);
27988 output_asm_insn ("cmp\t%0, %1", operands
);
27989 output_asm_insn ("bhi\t%l3", operands
);
27990 switch (GET_MODE(diff_vec
))
27993 return "tbb\t[%|pc, %0]";
27995 return "tbh\t[%|pc, %0, lsl #1]";
27999 output_asm_insn ("adr\t%4, %l2", operands
);
28000 output_asm_insn ("ldr\t%5, [%4, %0, lsl #2]", operands
);
28001 output_asm_insn ("add\t%4, %4, %5", operands
);
28006 output_asm_insn ("adr\t%4, %l2", operands
);
28007 return "ldr\t%|pc, [%4, %0, lsl #2]";
28010 gcc_unreachable ();
28014 /* Implement TARGET_SCHED_ISSUE_RATE. Lookup the issue rate in the
28015 per-core tuning structs. */
28017 arm_issue_rate (void)
28019 return current_tune
->issue_rate
;
28022 /* Return how many instructions should scheduler lookahead to choose the
28025 arm_first_cycle_multipass_dfa_lookahead (void)
28027 int issue_rate
= arm_issue_rate ();
28029 return issue_rate
> 1 && !sched_fusion
? issue_rate
: 0;
28032 /* Enable modeling of L2 auto-prefetcher. */
28034 arm_first_cycle_multipass_dfa_lookahead_guard (rtx_insn
*insn
, int ready_index
)
28036 return autopref_multipass_dfa_lookahead_guard (insn
, ready_index
);
28040 arm_mangle_type (const_tree type
)
28042 /* The ARM ABI documents (10th October 2008) say that "__va_list"
28043 has to be managled as if it is in the "std" namespace. */
28044 if (TARGET_AAPCS_BASED
28045 && lang_hooks
.types_compatible_p (CONST_CAST_TREE (type
), va_list_type
))
28046 return "St9__va_list";
28048 /* Half-precision float. */
28049 if (TREE_CODE (type
) == REAL_TYPE
&& TYPE_PRECISION (type
) == 16)
28052 /* Try mangling as a Neon type, TYPE_NAME is non-NULL if this is a
28054 if (TYPE_NAME (type
) != NULL
)
28055 return arm_mangle_builtin_type (type
);
28057 /* Use the default mangling. */
28061 /* Order of allocation of core registers for Thumb: this allocation is
28062 written over the corresponding initial entries of the array
28063 initialized with REG_ALLOC_ORDER. We allocate all low registers
28064 first. Saving and restoring a low register is usually cheaper than
28065 using a call-clobbered high register. */
28067 static const int thumb_core_reg_alloc_order
[] =
28069 3, 2, 1, 0, 4, 5, 6, 7,
28070 12, 14, 8, 9, 10, 11
28073 /* Adjust register allocation order when compiling for Thumb. */
28076 arm_order_regs_for_local_alloc (void)
28078 const int arm_reg_alloc_order
[] = REG_ALLOC_ORDER
;
28079 memcpy(reg_alloc_order
, arm_reg_alloc_order
, sizeof (reg_alloc_order
));
28081 memcpy (reg_alloc_order
, thumb_core_reg_alloc_order
,
28082 sizeof (thumb_core_reg_alloc_order
));
28085 /* Implement TARGET_FRAME_POINTER_REQUIRED. */
28088 arm_frame_pointer_required (void)
28090 if (SUBTARGET_FRAME_POINTER_REQUIRED
)
28093 /* If the function receives nonlocal gotos, it needs to save the frame
28094 pointer in the nonlocal_goto_save_area object. */
28095 if (cfun
->has_nonlocal_label
)
28098 /* The frame pointer is required for non-leaf APCS frames. */
28099 if (TARGET_ARM
&& TARGET_APCS_FRAME
&& !crtl
->is_leaf
)
28102 /* If we are probing the stack in the prologue, we will have a faulting
28103 instruction prior to the stack adjustment and this requires a frame
28104 pointer if we want to catch the exception using the EABI unwinder. */
28105 if (!IS_INTERRUPT (arm_current_func_type ())
28106 && (flag_stack_check
== STATIC_BUILTIN_STACK_CHECK
28107 || flag_stack_clash_protection
)
28108 && arm_except_unwind_info (&global_options
) == UI_TARGET
28109 && cfun
->can_throw_non_call_exceptions
)
28111 HOST_WIDE_INT size
= get_frame_size ();
28113 /* That's irrelevant if there is no stack adjustment. */
28117 /* That's relevant only if there is a stack probe. */
28118 if (crtl
->is_leaf
&& !cfun
->calls_alloca
)
28120 /* We don't have the final size of the frame so adjust. */
28121 size
+= 32 * UNITS_PER_WORD
;
28122 if (size
> PROBE_INTERVAL
&& size
> get_stack_check_protect ())
28132 /* Only thumb1 can't support conditional execution, so return true if
28133 the target is not thumb1. */
28135 arm_have_conditional_execution (void)
28137 return !TARGET_THUMB1
;
28140 /* The AAPCS sets the maximum alignment of a vector to 64 bits. */
28141 static HOST_WIDE_INT
28142 arm_vector_alignment (const_tree type
)
28144 HOST_WIDE_INT align
= tree_to_shwi (TYPE_SIZE (type
));
28146 if (TARGET_AAPCS_BASED
)
28147 align
= MIN (align
, 64);
28153 arm_autovectorize_vector_sizes (vector_sizes
*sizes
)
28155 if (!TARGET_NEON_VECTORIZE_DOUBLE
)
28157 sizes
->safe_push (16);
28158 sizes
->safe_push (8);
28163 arm_vector_alignment_reachable (const_tree type
, bool is_packed
)
28165 /* Vectors which aren't in packed structures will not be less aligned than
28166 the natural alignment of their element type, so this is safe. */
28167 if (TARGET_NEON
&& !BYTES_BIG_ENDIAN
&& unaligned_access
)
28170 return default_builtin_vector_alignment_reachable (type
, is_packed
);
28174 arm_builtin_support_vector_misalignment (machine_mode mode
,
28175 const_tree type
, int misalignment
,
28178 if (TARGET_NEON
&& !BYTES_BIG_ENDIAN
&& unaligned_access
)
28180 HOST_WIDE_INT align
= TYPE_ALIGN_UNIT (type
);
28185 /* If the misalignment is unknown, we should be able to handle the access
28186 so long as it is not to a member of a packed data structure. */
28187 if (misalignment
== -1)
28190 /* Return true if the misalignment is a multiple of the natural alignment
28191 of the vector's element type. This is probably always going to be
28192 true in practice, since we've already established that this isn't a
28194 return ((misalignment
% align
) == 0);
28197 return default_builtin_support_vector_misalignment (mode
, type
, misalignment
,
28202 arm_conditional_register_usage (void)
28206 if (TARGET_THUMB1
&& optimize_size
)
28208 /* When optimizing for size on Thumb-1, it's better not
28209 to use the HI regs, because of the overhead of
28211 for (regno
= FIRST_HI_REGNUM
; regno
<= LAST_HI_REGNUM
; ++regno
)
28212 fixed_regs
[regno
] = call_used_regs
[regno
] = 1;
28215 /* The link register can be clobbered by any branch insn,
28216 but we have no way to track that at present, so mark
28217 it as unavailable. */
28219 fixed_regs
[LR_REGNUM
] = call_used_regs
[LR_REGNUM
] = 1;
28221 if (TARGET_32BIT
&& TARGET_HARD_FLOAT
)
28223 /* VFPv3 registers are disabled when earlier VFP
28224 versions are selected due to the definition of
28225 LAST_VFP_REGNUM. */
28226 for (regno
= FIRST_VFP_REGNUM
;
28227 regno
<= LAST_VFP_REGNUM
; ++ regno
)
28229 fixed_regs
[regno
] = 0;
28230 call_used_regs
[regno
] = regno
< FIRST_VFP_REGNUM
+ 16
28231 || regno
>= FIRST_VFP_REGNUM
+ 32;
28235 if (TARGET_REALLY_IWMMXT
)
28237 regno
= FIRST_IWMMXT_GR_REGNUM
;
28238 /* The 2002/10/09 revision of the XScale ABI has wCG0
28239 and wCG1 as call-preserved registers. The 2002/11/21
28240 revision changed this so that all wCG registers are
28241 scratch registers. */
28242 for (regno
= FIRST_IWMMXT_GR_REGNUM
;
28243 regno
<= LAST_IWMMXT_GR_REGNUM
; ++ regno
)
28244 fixed_regs
[regno
] = 0;
28245 /* The XScale ABI has wR0 - wR9 as scratch registers,
28246 the rest as call-preserved registers. */
28247 for (regno
= FIRST_IWMMXT_REGNUM
;
28248 regno
<= LAST_IWMMXT_REGNUM
; ++ regno
)
28250 fixed_regs
[regno
] = 0;
28251 call_used_regs
[regno
] = regno
< FIRST_IWMMXT_REGNUM
+ 10;
28255 if ((unsigned) PIC_OFFSET_TABLE_REGNUM
!= INVALID_REGNUM
)
28257 fixed_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
28258 call_used_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
28260 else if (TARGET_APCS_STACK
)
28262 fixed_regs
[10] = 1;
28263 call_used_regs
[10] = 1;
28265 /* -mcaller-super-interworking reserves r11 for calls to
28266 _interwork_r11_call_via_rN(). Making the register global
28267 is an easy way of ensuring that it remains valid for all
28269 if (TARGET_APCS_FRAME
|| TARGET_CALLER_INTERWORKING
28270 || TARGET_TPCS_FRAME
|| TARGET_TPCS_LEAF_FRAME
)
28272 fixed_regs
[ARM_HARD_FRAME_POINTER_REGNUM
] = 1;
28273 call_used_regs
[ARM_HARD_FRAME_POINTER_REGNUM
] = 1;
28274 if (TARGET_CALLER_INTERWORKING
)
28275 global_regs
[ARM_HARD_FRAME_POINTER_REGNUM
] = 1;
28277 SUBTARGET_CONDITIONAL_REGISTER_USAGE
28281 arm_preferred_rename_class (reg_class_t rclass
)
28283 /* Thumb-2 instructions using LO_REGS may be smaller than instructions
28284 using GENERIC_REGS. During register rename pass, we prefer LO_REGS,
28285 and code size can be reduced. */
28286 if (TARGET_THUMB2
&& rclass
== GENERAL_REGS
)
28292 /* Compute the attribute "length" of insn "*push_multi".
28293 So this function MUST be kept in sync with that insn pattern. */
28295 arm_attr_length_push_multi(rtx parallel_op
, rtx first_op
)
28297 int i
, regno
, hi_reg
;
28298 int num_saves
= XVECLEN (parallel_op
, 0);
28308 regno
= REGNO (first_op
);
28309 /* For PUSH/STM under Thumb2 mode, we can use 16-bit encodings if the register
28310 list is 8-bit. Normally this means all registers in the list must be
28311 LO_REGS, that is (R0 -R7). If any HI_REGS used, then we must use 32-bit
28312 encodings. There is one exception for PUSH that LR in HI_REGS can be used
28313 with 16-bit encoding. */
28314 hi_reg
= (REGNO_REG_CLASS (regno
) == HI_REGS
) && (regno
!= LR_REGNUM
);
28315 for (i
= 1; i
< num_saves
&& !hi_reg
; i
++)
28317 regno
= REGNO (XEXP (XVECEXP (parallel_op
, 0, i
), 0));
28318 hi_reg
|= (REGNO_REG_CLASS (regno
) == HI_REGS
) && (regno
!= LR_REGNUM
);
28326 /* Compute the attribute "length" of insn. Currently, this function is used
28327 for "*load_multiple_with_writeback", "*pop_multiple_with_return" and
28328 "*pop_multiple_with_writeback_and_return". OPERANDS is the toplevel PARALLEL
28329 rtx, RETURN_PC is true if OPERANDS contains return insn. WRITE_BACK_P is
28330 true if OPERANDS contains insn which explicit updates base register. */
28333 arm_attr_length_pop_multi (rtx
*operands
, bool return_pc
, bool write_back_p
)
28342 rtx parallel_op
= operands
[0];
28343 /* Initialize to elements number of PARALLEL. */
28344 unsigned indx
= XVECLEN (parallel_op
, 0) - 1;
28345 /* Initialize the value to base register. */
28346 unsigned regno
= REGNO (operands
[1]);
28347 /* Skip return and write back pattern.
28348 We only need register pop pattern for later analysis. */
28349 unsigned first_indx
= 0;
28350 first_indx
+= return_pc
? 1 : 0;
28351 first_indx
+= write_back_p
? 1 : 0;
28353 /* A pop operation can be done through LDM or POP. If the base register is SP
28354 and if it's with write back, then a LDM will be alias of POP. */
28355 bool pop_p
= (regno
== SP_REGNUM
&& write_back_p
);
28356 bool ldm_p
= !pop_p
;
28358 /* Check base register for LDM. */
28359 if (ldm_p
&& REGNO_REG_CLASS (regno
) == HI_REGS
)
28362 /* Check each register in the list. */
28363 for (; indx
>= first_indx
; indx
--)
28365 regno
= REGNO (XEXP (XVECEXP (parallel_op
, 0, indx
), 0));
28366 /* For POP, PC in HI_REGS can be used with 16-bit encoding. See similar
28367 comment in arm_attr_length_push_multi. */
28368 if (REGNO_REG_CLASS (regno
) == HI_REGS
28369 && (regno
!= PC_REGNUM
|| ldm_p
))
28376 /* Compute the number of instructions emitted by output_move_double. */
28378 arm_count_output_move_double_insns (rtx
*operands
)
28382 /* output_move_double may modify the operands array, so call it
28383 here on a copy of the array. */
28384 ops
[0] = operands
[0];
28385 ops
[1] = operands
[1];
28386 output_move_double (ops
, false, &count
);
28391 vfp3_const_double_for_fract_bits (rtx operand
)
28393 REAL_VALUE_TYPE r0
;
28395 if (!CONST_DOUBLE_P (operand
))
28398 r0
= *CONST_DOUBLE_REAL_VALUE (operand
);
28399 if (exact_real_inverse (DFmode
, &r0
)
28400 && !REAL_VALUE_NEGATIVE (r0
))
28402 if (exact_real_truncate (DFmode
, &r0
))
28404 HOST_WIDE_INT value
= real_to_integer (&r0
);
28405 value
= value
& 0xffffffff;
28406 if ((value
!= 0) && ( (value
& (value
- 1)) == 0))
28408 int ret
= exact_log2 (value
);
28409 gcc_assert (IN_RANGE (ret
, 0, 31));
28417 /* If X is a CONST_DOUBLE with a value that is a power of 2 whose
28418 log2 is in [1, 32], return that log2. Otherwise return -1.
28419 This is used in the patterns for vcvt.s32.f32 floating-point to
28420 fixed-point conversions. */
28423 vfp3_const_double_for_bits (rtx x
)
28425 const REAL_VALUE_TYPE
*r
;
28427 if (!CONST_DOUBLE_P (x
))
28430 r
= CONST_DOUBLE_REAL_VALUE (x
);
28432 if (REAL_VALUE_NEGATIVE (*r
)
28433 || REAL_VALUE_ISNAN (*r
)
28434 || REAL_VALUE_ISINF (*r
)
28435 || !real_isinteger (r
, SFmode
))
28438 HOST_WIDE_INT hwint
= exact_log2 (real_to_integer (r
));
28440 /* The exact_log2 above will have returned -1 if this is
28441 not an exact log2. */
28442 if (!IN_RANGE (hwint
, 1, 32))
28449 /* Emit a memory barrier around an atomic sequence according to MODEL. */
28452 arm_pre_atomic_barrier (enum memmodel model
)
28454 if (need_atomic_barrier_p (model
, true))
28455 emit_insn (gen_memory_barrier ());
28459 arm_post_atomic_barrier (enum memmodel model
)
28461 if (need_atomic_barrier_p (model
, false))
28462 emit_insn (gen_memory_barrier ());
28465 /* Emit the load-exclusive and store-exclusive instructions.
28466 Use acquire and release versions if necessary. */
28469 arm_emit_load_exclusive (machine_mode mode
, rtx rval
, rtx mem
, bool acq
)
28471 rtx (*gen
) (rtx
, rtx
);
28477 case E_QImode
: gen
= gen_arm_load_acquire_exclusiveqi
; break;
28478 case E_HImode
: gen
= gen_arm_load_acquire_exclusivehi
; break;
28479 case E_SImode
: gen
= gen_arm_load_acquire_exclusivesi
; break;
28480 case E_DImode
: gen
= gen_arm_load_acquire_exclusivedi
; break;
28482 gcc_unreachable ();
28489 case E_QImode
: gen
= gen_arm_load_exclusiveqi
; break;
28490 case E_HImode
: gen
= gen_arm_load_exclusivehi
; break;
28491 case E_SImode
: gen
= gen_arm_load_exclusivesi
; break;
28492 case E_DImode
: gen
= gen_arm_load_exclusivedi
; break;
28494 gcc_unreachable ();
28498 emit_insn (gen (rval
, mem
));
28502 arm_emit_store_exclusive (machine_mode mode
, rtx bval
, rtx rval
,
28505 rtx (*gen
) (rtx
, rtx
, rtx
);
28511 case E_QImode
: gen
= gen_arm_store_release_exclusiveqi
; break;
28512 case E_HImode
: gen
= gen_arm_store_release_exclusivehi
; break;
28513 case E_SImode
: gen
= gen_arm_store_release_exclusivesi
; break;
28514 case E_DImode
: gen
= gen_arm_store_release_exclusivedi
; break;
28516 gcc_unreachable ();
28523 case E_QImode
: gen
= gen_arm_store_exclusiveqi
; break;
28524 case E_HImode
: gen
= gen_arm_store_exclusivehi
; break;
28525 case E_SImode
: gen
= gen_arm_store_exclusivesi
; break;
28526 case E_DImode
: gen
= gen_arm_store_exclusivedi
; break;
28528 gcc_unreachable ();
28532 emit_insn (gen (bval
, rval
, mem
));
28535 /* Mark the previous jump instruction as unlikely. */
28538 emit_unlikely_jump (rtx insn
)
28540 rtx_insn
*jump
= emit_jump_insn (insn
);
28541 add_reg_br_prob_note (jump
, profile_probability::very_unlikely ());
28544 /* Expand a compare and swap pattern. */
28547 arm_expand_compare_and_swap (rtx operands
[])
28549 rtx bval
, bdst
, rval
, mem
, oldval
, newval
, is_weak
, mod_s
, mod_f
, x
;
28551 rtx (*gen
) (rtx
, rtx
, rtx
, rtx
, rtx
, rtx
, rtx
, rtx
);
28553 bval
= operands
[0];
28554 rval
= operands
[1];
28556 oldval
= operands
[3];
28557 newval
= operands
[4];
28558 is_weak
= operands
[5];
28559 mod_s
= operands
[6];
28560 mod_f
= operands
[7];
28561 mode
= GET_MODE (mem
);
28563 /* Normally the succ memory model must be stronger than fail, but in the
28564 unlikely event of fail being ACQUIRE and succ being RELEASE we need to
28565 promote succ to ACQ_REL so that we don't lose the acquire semantics. */
28567 if (TARGET_HAVE_LDACQ
28568 && is_mm_acquire (memmodel_from_int (INTVAL (mod_f
)))
28569 && is_mm_release (memmodel_from_int (INTVAL (mod_s
))))
28570 mod_s
= GEN_INT (MEMMODEL_ACQ_REL
);
28576 /* For narrow modes, we're going to perform the comparison in SImode,
28577 so do the zero-extension now. */
28578 rval
= gen_reg_rtx (SImode
);
28579 oldval
= convert_modes (SImode
, mode
, oldval
, true);
28583 /* Force the value into a register if needed. We waited until after
28584 the zero-extension above to do this properly. */
28585 if (!arm_add_operand (oldval
, SImode
))
28586 oldval
= force_reg (SImode
, oldval
);
28590 if (!cmpdi_operand (oldval
, mode
))
28591 oldval
= force_reg (mode
, oldval
);
28595 gcc_unreachable ();
28602 case E_QImode
: gen
= gen_atomic_compare_and_swapt1qi_1
; break;
28603 case E_HImode
: gen
= gen_atomic_compare_and_swapt1hi_1
; break;
28604 case E_SImode
: gen
= gen_atomic_compare_and_swapt1si_1
; break;
28605 case E_DImode
: gen
= gen_atomic_compare_and_swapt1di_1
; break;
28607 gcc_unreachable ();
28614 case E_QImode
: gen
= gen_atomic_compare_and_swap32qi_1
; break;
28615 case E_HImode
: gen
= gen_atomic_compare_and_swap32hi_1
; break;
28616 case E_SImode
: gen
= gen_atomic_compare_and_swap32si_1
; break;
28617 case E_DImode
: gen
= gen_atomic_compare_and_swap32di_1
; break;
28619 gcc_unreachable ();
28623 bdst
= TARGET_THUMB1
? bval
: gen_rtx_REG (CC_Zmode
, CC_REGNUM
);
28624 emit_insn (gen (bdst
, rval
, mem
, oldval
, newval
, is_weak
, mod_s
, mod_f
));
28626 if (mode
== QImode
|| mode
== HImode
)
28627 emit_move_insn (operands
[1], gen_lowpart (mode
, rval
));
28629 /* In all cases, we arrange for success to be signaled by Z set.
28630 This arrangement allows for the boolean result to be used directly
28631 in a subsequent branch, post optimization. For Thumb-1 targets, the
28632 boolean negation of the result is also stored in bval because Thumb-1
28633 backend lacks dependency tracking for CC flag due to flag-setting not
28634 being represented at RTL level. */
28636 emit_insn (gen_cstoresi_eq0_thumb1 (bval
, bdst
));
28639 x
= gen_rtx_EQ (SImode
, bdst
, const0_rtx
);
28640 emit_insn (gen_rtx_SET (bval
, x
));
28644 /* Split a compare and swap pattern. It is IMPLEMENTATION DEFINED whether
28645 another memory store between the load-exclusive and store-exclusive can
28646 reset the monitor from Exclusive to Open state. This means we must wait
28647 until after reload to split the pattern, lest we get a register spill in
28648 the middle of the atomic sequence. Success of the compare and swap is
28649 indicated by the Z flag set for 32bit targets and by neg_bval being zero
28650 for Thumb-1 targets (ie. negation of the boolean value returned by
28651 atomic_compare_and_swapmode standard pattern in operand 0). */
28654 arm_split_compare_and_swap (rtx operands
[])
28656 rtx rval
, mem
, oldval
, newval
, neg_bval
, mod_s_rtx
;
28658 enum memmodel mod_s
, mod_f
;
28660 rtx_code_label
*label1
, *label2
;
28663 rval
= operands
[1];
28665 oldval
= operands
[3];
28666 newval
= operands
[4];
28667 is_weak
= (operands
[5] != const0_rtx
);
28668 mod_s_rtx
= operands
[6];
28669 mod_s
= memmodel_from_int (INTVAL (mod_s_rtx
));
28670 mod_f
= memmodel_from_int (INTVAL (operands
[7]));
28671 neg_bval
= TARGET_THUMB1
? operands
[0] : operands
[8];
28672 mode
= GET_MODE (mem
);
28674 bool is_armv8_sync
= arm_arch8
&& is_mm_sync (mod_s
);
28676 bool use_acquire
= TARGET_HAVE_LDACQ
&& aarch_mm_needs_acquire (mod_s_rtx
);
28677 bool use_release
= TARGET_HAVE_LDACQ
&& aarch_mm_needs_release (mod_s_rtx
);
28679 /* For ARMv8, the load-acquire is too weak for __sync memory orders. Instead,
28680 a full barrier is emitted after the store-release. */
28682 use_acquire
= false;
28684 /* Checks whether a barrier is needed and emits one accordingly. */
28685 if (!(use_acquire
|| use_release
))
28686 arm_pre_atomic_barrier (mod_s
);
28691 label1
= gen_label_rtx ();
28692 emit_label (label1
);
28694 label2
= gen_label_rtx ();
28696 arm_emit_load_exclusive (mode
, rval
, mem
, use_acquire
);
28698 /* Z is set to 0 for 32bit targets (resp. rval set to 1) if oldval != rval,
28699 as required to communicate with arm_expand_compare_and_swap. */
28702 cond
= arm_gen_compare_reg (NE
, rval
, oldval
, neg_bval
);
28703 x
= gen_rtx_NE (VOIDmode
, cond
, const0_rtx
);
28704 x
= gen_rtx_IF_THEN_ELSE (VOIDmode
, x
,
28705 gen_rtx_LABEL_REF (Pmode
, label2
), pc_rtx
);
28706 emit_unlikely_jump (gen_rtx_SET (pc_rtx
, x
));
28710 emit_move_insn (neg_bval
, const1_rtx
);
28711 cond
= gen_rtx_NE (VOIDmode
, rval
, oldval
);
28712 if (thumb1_cmpneg_operand (oldval
, SImode
))
28713 emit_unlikely_jump (gen_cbranchsi4_scratch (neg_bval
, rval
, oldval
,
28716 emit_unlikely_jump (gen_cbranchsi4_insn (cond
, rval
, oldval
, label2
));
28719 arm_emit_store_exclusive (mode
, neg_bval
, mem
, newval
, use_release
);
28721 /* Weak or strong, we want EQ to be true for success, so that we
28722 match the flags that we got from the compare above. */
28725 cond
= gen_rtx_REG (CCmode
, CC_REGNUM
);
28726 x
= gen_rtx_COMPARE (CCmode
, neg_bval
, const0_rtx
);
28727 emit_insn (gen_rtx_SET (cond
, x
));
28732 /* Z is set to boolean value of !neg_bval, as required to communicate
28733 with arm_expand_compare_and_swap. */
28734 x
= gen_rtx_NE (VOIDmode
, neg_bval
, const0_rtx
);
28735 emit_unlikely_jump (gen_cbranchsi4 (x
, neg_bval
, const0_rtx
, label1
));
28738 if (!is_mm_relaxed (mod_f
))
28739 emit_label (label2
);
28741 /* Checks whether a barrier is needed and emits one accordingly. */
28743 || !(use_acquire
|| use_release
))
28744 arm_post_atomic_barrier (mod_s
);
28746 if (is_mm_relaxed (mod_f
))
28747 emit_label (label2
);
28750 /* Split an atomic operation pattern. Operation is given by CODE and is one
28751 of PLUS, MINUS, IOR, XOR, SET (for an exchange operation) or NOT (for a nand
28752 operation). Operation is performed on the content at MEM and on VALUE
28753 following the memory model MODEL_RTX. The content at MEM before and after
28754 the operation is returned in OLD_OUT and NEW_OUT respectively while the
28755 success of the operation is returned in COND. Using a scratch register or
28756 an operand register for these determines what result is returned for that
28760 arm_split_atomic_op (enum rtx_code code
, rtx old_out
, rtx new_out
, rtx mem
,
28761 rtx value
, rtx model_rtx
, rtx cond
)
28763 enum memmodel model
= memmodel_from_int (INTVAL (model_rtx
));
28764 machine_mode mode
= GET_MODE (mem
);
28765 machine_mode wmode
= (mode
== DImode
? DImode
: SImode
);
28766 rtx_code_label
*label
;
28767 bool all_low_regs
, bind_old_new
;
28770 bool is_armv8_sync
= arm_arch8
&& is_mm_sync (model
);
28772 bool use_acquire
= TARGET_HAVE_LDACQ
&& aarch_mm_needs_acquire (model_rtx
);
28773 bool use_release
= TARGET_HAVE_LDACQ
&& aarch_mm_needs_release (model_rtx
);
28775 /* For ARMv8, a load-acquire is too weak for __sync memory orders. Instead,
28776 a full barrier is emitted after the store-release. */
28778 use_acquire
= false;
28780 /* Checks whether a barrier is needed and emits one accordingly. */
28781 if (!(use_acquire
|| use_release
))
28782 arm_pre_atomic_barrier (model
);
28784 label
= gen_label_rtx ();
28785 emit_label (label
);
28788 new_out
= gen_lowpart (wmode
, new_out
);
28790 old_out
= gen_lowpart (wmode
, old_out
);
28793 value
= simplify_gen_subreg (wmode
, value
, mode
, 0);
28795 arm_emit_load_exclusive (mode
, old_out
, mem
, use_acquire
);
28797 /* Does the operation require destination and first operand to use the same
28798 register? This is decided by register constraints of relevant insn
28799 patterns in thumb1.md. */
28800 gcc_assert (!new_out
|| REG_P (new_out
));
28801 all_low_regs
= REG_P (value
) && REGNO_REG_CLASS (REGNO (value
)) == LO_REGS
28802 && new_out
&& REGNO_REG_CLASS (REGNO (new_out
)) == LO_REGS
28803 && REGNO_REG_CLASS (REGNO (old_out
)) == LO_REGS
;
28808 && (code
!= PLUS
|| (!all_low_regs
&& !satisfies_constraint_L (value
))));
28810 /* We want to return the old value while putting the result of the operation
28811 in the same register as the old value so copy the old value over to the
28812 destination register and use that register for the operation. */
28813 if (old_out
&& bind_old_new
)
28815 emit_move_insn (new_out
, old_out
);
28826 x
= gen_rtx_AND (wmode
, old_out
, value
);
28827 emit_insn (gen_rtx_SET (new_out
, x
));
28828 x
= gen_rtx_NOT (wmode
, new_out
);
28829 emit_insn (gen_rtx_SET (new_out
, x
));
28833 if (CONST_INT_P (value
))
28835 value
= GEN_INT (-INTVAL (value
));
28841 if (mode
== DImode
)
28843 /* DImode plus/minus need to clobber flags. */
28844 /* The adddi3 and subdi3 patterns are incorrectly written so that
28845 they require matching operands, even when we could easily support
28846 three operands. Thankfully, this can be fixed up post-splitting,
28847 as the individual add+adc patterns do accept three operands and
28848 post-reload cprop can make these moves go away. */
28849 emit_move_insn (new_out
, old_out
);
28851 x
= gen_adddi3 (new_out
, new_out
, value
);
28853 x
= gen_subdi3 (new_out
, new_out
, value
);
28860 x
= gen_rtx_fmt_ee (code
, wmode
, old_out
, value
);
28861 emit_insn (gen_rtx_SET (new_out
, x
));
28865 arm_emit_store_exclusive (mode
, cond
, mem
, gen_lowpart (mode
, new_out
),
28868 x
= gen_rtx_NE (VOIDmode
, cond
, const0_rtx
);
28869 emit_unlikely_jump (gen_cbranchsi4 (x
, cond
, const0_rtx
, label
));
28871 /* Checks whether a barrier is needed and emits one accordingly. */
28873 || !(use_acquire
|| use_release
))
28874 arm_post_atomic_barrier (model
);
28877 #define MAX_VECT_LEN 16
28879 struct expand_vec_perm_d
28881 rtx target
, op0
, op1
;
28882 vec_perm_indices perm
;
28883 machine_mode vmode
;
28888 /* Generate a variable permutation. */
28891 arm_expand_vec_perm_1 (rtx target
, rtx op0
, rtx op1
, rtx sel
)
28893 machine_mode vmode
= GET_MODE (target
);
28894 bool one_vector_p
= rtx_equal_p (op0
, op1
);
28896 gcc_checking_assert (vmode
== V8QImode
|| vmode
== V16QImode
);
28897 gcc_checking_assert (GET_MODE (op0
) == vmode
);
28898 gcc_checking_assert (GET_MODE (op1
) == vmode
);
28899 gcc_checking_assert (GET_MODE (sel
) == vmode
);
28900 gcc_checking_assert (TARGET_NEON
);
28904 if (vmode
== V8QImode
)
28905 emit_insn (gen_neon_vtbl1v8qi (target
, op0
, sel
));
28907 emit_insn (gen_neon_vtbl1v16qi (target
, op0
, sel
));
28913 if (vmode
== V8QImode
)
28915 pair
= gen_reg_rtx (V16QImode
);
28916 emit_insn (gen_neon_vcombinev8qi (pair
, op0
, op1
));
28917 pair
= gen_lowpart (TImode
, pair
);
28918 emit_insn (gen_neon_vtbl2v8qi (target
, pair
, sel
));
28922 pair
= gen_reg_rtx (OImode
);
28923 emit_insn (gen_neon_vcombinev16qi (pair
, op0
, op1
));
28924 emit_insn (gen_neon_vtbl2v16qi (target
, pair
, sel
));
28930 arm_expand_vec_perm (rtx target
, rtx op0
, rtx op1
, rtx sel
)
28932 machine_mode vmode
= GET_MODE (target
);
28933 unsigned int nelt
= GET_MODE_NUNITS (vmode
);
28934 bool one_vector_p
= rtx_equal_p (op0
, op1
);
28937 /* TODO: ARM's VTBL indexing is little-endian. In order to handle GCC's
28938 numbering of elements for big-endian, we must reverse the order. */
28939 gcc_checking_assert (!BYTES_BIG_ENDIAN
);
28941 /* The VTBL instruction does not use a modulo index, so we must take care
28942 of that ourselves. */
28943 mask
= GEN_INT (one_vector_p
? nelt
- 1 : 2 * nelt
- 1);
28944 mask
= gen_const_vec_duplicate (vmode
, mask
);
28945 sel
= expand_simple_binop (vmode
, AND
, sel
, mask
, NULL
, 0, OPTAB_LIB_WIDEN
);
28947 arm_expand_vec_perm_1 (target
, op0
, op1
, sel
);
28950 /* Map lane ordering between architectural lane order, and GCC lane order,
28951 taking into account ABI. See comment above output_move_neon for details. */
28954 neon_endian_lane_map (machine_mode mode
, int lane
)
28956 if (BYTES_BIG_ENDIAN
)
28958 int nelems
= GET_MODE_NUNITS (mode
);
28959 /* Reverse lane order. */
28960 lane
= (nelems
- 1 - lane
);
28961 /* Reverse D register order, to match ABI. */
28962 if (GET_MODE_SIZE (mode
) == 16)
28963 lane
= lane
^ (nelems
/ 2);
28968 /* Some permutations index into pairs of vectors, this is a helper function
28969 to map indexes into those pairs of vectors. */
28972 neon_pair_endian_lane_map (machine_mode mode
, int lane
)
28974 int nelem
= GET_MODE_NUNITS (mode
);
28975 if (BYTES_BIG_ENDIAN
)
28977 neon_endian_lane_map (mode
, lane
& (nelem
- 1)) + (lane
& nelem
);
28981 /* Generate or test for an insn that supports a constant permutation. */
28983 /* Recognize patterns for the VUZP insns. */
28986 arm_evpc_neon_vuzp (struct expand_vec_perm_d
*d
)
28988 unsigned int i
, odd
, mask
, nelt
= d
->perm
.length ();
28989 rtx out0
, out1
, in0
, in1
;
28990 rtx (*gen
)(rtx
, rtx
, rtx
, rtx
);
28994 if (GET_MODE_UNIT_SIZE (d
->vmode
) >= 8)
28997 /* arm_expand_vec_perm_const_1 () helpfully swaps the operands for the
28998 big endian pattern on 64 bit vectors, so we correct for that. */
28999 swap_nelt
= BYTES_BIG_ENDIAN
&& !d
->one_vector_p
29000 && GET_MODE_SIZE (d
->vmode
) == 8 ? nelt
: 0;
29002 first_elem
= d
->perm
[neon_endian_lane_map (d
->vmode
, 0)] ^ swap_nelt
;
29004 if (first_elem
== neon_endian_lane_map (d
->vmode
, 0))
29006 else if (first_elem
== neon_endian_lane_map (d
->vmode
, 1))
29010 mask
= (d
->one_vector_p
? nelt
- 1 : 2 * nelt
- 1);
29012 for (i
= 0; i
< nelt
; i
++)
29015 (neon_pair_endian_lane_map (d
->vmode
, i
) * 2 + odd
) & mask
;
29016 if ((d
->perm
[i
] ^ swap_nelt
) != neon_pair_endian_lane_map (d
->vmode
, elt
))
29026 case E_V16QImode
: gen
= gen_neon_vuzpv16qi_internal
; break;
29027 case E_V8QImode
: gen
= gen_neon_vuzpv8qi_internal
; break;
29028 case E_V8HImode
: gen
= gen_neon_vuzpv8hi_internal
; break;
29029 case E_V4HImode
: gen
= gen_neon_vuzpv4hi_internal
; break;
29030 case E_V8HFmode
: gen
= gen_neon_vuzpv8hf_internal
; break;
29031 case E_V4HFmode
: gen
= gen_neon_vuzpv4hf_internal
; break;
29032 case E_V4SImode
: gen
= gen_neon_vuzpv4si_internal
; break;
29033 case E_V2SImode
: gen
= gen_neon_vuzpv2si_internal
; break;
29034 case E_V2SFmode
: gen
= gen_neon_vuzpv2sf_internal
; break;
29035 case E_V4SFmode
: gen
= gen_neon_vuzpv4sf_internal
; break;
29037 gcc_unreachable ();
29042 if (swap_nelt
!= 0)
29043 std::swap (in0
, in1
);
29046 out1
= gen_reg_rtx (d
->vmode
);
29048 std::swap (out0
, out1
);
29050 emit_insn (gen (out0
, in0
, in1
, out1
));
29054 /* Recognize patterns for the VZIP insns. */
29057 arm_evpc_neon_vzip (struct expand_vec_perm_d
*d
)
29059 unsigned int i
, high
, mask
, nelt
= d
->perm
.length ();
29060 rtx out0
, out1
, in0
, in1
;
29061 rtx (*gen
)(rtx
, rtx
, rtx
, rtx
);
29065 if (GET_MODE_UNIT_SIZE (d
->vmode
) >= 8)
29068 is_swapped
= BYTES_BIG_ENDIAN
;
29070 first_elem
= d
->perm
[neon_endian_lane_map (d
->vmode
, 0) ^ is_swapped
];
29073 if (first_elem
== neon_endian_lane_map (d
->vmode
, high
))
29075 else if (first_elem
== neon_endian_lane_map (d
->vmode
, 0))
29079 mask
= (d
->one_vector_p
? nelt
- 1 : 2 * nelt
- 1);
29081 for (i
= 0; i
< nelt
/ 2; i
++)
29084 neon_pair_endian_lane_map (d
->vmode
, i
+ high
) & mask
;
29085 if (d
->perm
[neon_pair_endian_lane_map (d
->vmode
, 2 * i
+ is_swapped
)]
29089 neon_pair_endian_lane_map (d
->vmode
, i
+ nelt
+ high
) & mask
;
29090 if (d
->perm
[neon_pair_endian_lane_map (d
->vmode
, 2 * i
+ !is_swapped
)]
29101 case E_V16QImode
: gen
= gen_neon_vzipv16qi_internal
; break;
29102 case E_V8QImode
: gen
= gen_neon_vzipv8qi_internal
; break;
29103 case E_V8HImode
: gen
= gen_neon_vzipv8hi_internal
; break;
29104 case E_V4HImode
: gen
= gen_neon_vzipv4hi_internal
; break;
29105 case E_V8HFmode
: gen
= gen_neon_vzipv8hf_internal
; break;
29106 case E_V4HFmode
: gen
= gen_neon_vzipv4hf_internal
; break;
29107 case E_V4SImode
: gen
= gen_neon_vzipv4si_internal
; break;
29108 case E_V2SImode
: gen
= gen_neon_vzipv2si_internal
; break;
29109 case E_V2SFmode
: gen
= gen_neon_vzipv2sf_internal
; break;
29110 case E_V4SFmode
: gen
= gen_neon_vzipv4sf_internal
; break;
29112 gcc_unreachable ();
29118 std::swap (in0
, in1
);
29121 out1
= gen_reg_rtx (d
->vmode
);
29123 std::swap (out0
, out1
);
29125 emit_insn (gen (out0
, in0
, in1
, out1
));
29129 /* Recognize patterns for the VREV insns. */
29132 arm_evpc_neon_vrev (struct expand_vec_perm_d
*d
)
29134 unsigned int i
, j
, diff
, nelt
= d
->perm
.length ();
29135 rtx (*gen
)(rtx
, rtx
);
29137 if (!d
->one_vector_p
)
29146 case E_V16QImode
: gen
= gen_neon_vrev64v16qi
; break;
29147 case E_V8QImode
: gen
= gen_neon_vrev64v8qi
; break;
29155 case E_V16QImode
: gen
= gen_neon_vrev32v16qi
; break;
29156 case E_V8QImode
: gen
= gen_neon_vrev32v8qi
; break;
29157 case E_V8HImode
: gen
= gen_neon_vrev64v8hi
; break;
29158 case E_V4HImode
: gen
= gen_neon_vrev64v4hi
; break;
29159 case E_V8HFmode
: gen
= gen_neon_vrev64v8hf
; break;
29160 case E_V4HFmode
: gen
= gen_neon_vrev64v4hf
; break;
29168 case E_V16QImode
: gen
= gen_neon_vrev16v16qi
; break;
29169 case E_V8QImode
: gen
= gen_neon_vrev16v8qi
; break;
29170 case E_V8HImode
: gen
= gen_neon_vrev32v8hi
; break;
29171 case E_V4HImode
: gen
= gen_neon_vrev32v4hi
; break;
29172 case E_V4SImode
: gen
= gen_neon_vrev64v4si
; break;
29173 case E_V2SImode
: gen
= gen_neon_vrev64v2si
; break;
29174 case E_V4SFmode
: gen
= gen_neon_vrev64v4sf
; break;
29175 case E_V2SFmode
: gen
= gen_neon_vrev64v2sf
; break;
29184 for (i
= 0; i
< nelt
; i
+= diff
+ 1)
29185 for (j
= 0; j
<= diff
; j
+= 1)
29187 /* This is guaranteed to be true as the value of diff
29188 is 7, 3, 1 and we should have enough elements in the
29189 queue to generate this. Getting a vector mask with a
29190 value of diff other than these values implies that
29191 something is wrong by the time we get here. */
29192 gcc_assert (i
+ j
< nelt
);
29193 if (d
->perm
[i
+ j
] != i
+ diff
- j
)
29201 emit_insn (gen (d
->target
, d
->op0
));
29205 /* Recognize patterns for the VTRN insns. */
29208 arm_evpc_neon_vtrn (struct expand_vec_perm_d
*d
)
29210 unsigned int i
, odd
, mask
, nelt
= d
->perm
.length ();
29211 rtx out0
, out1
, in0
, in1
;
29212 rtx (*gen
)(rtx
, rtx
, rtx
, rtx
);
29214 if (GET_MODE_UNIT_SIZE (d
->vmode
) >= 8)
29217 /* Note that these are little-endian tests. Adjust for big-endian later. */
29218 if (d
->perm
[0] == 0)
29220 else if (d
->perm
[0] == 1)
29224 mask
= (d
->one_vector_p
? nelt
- 1 : 2 * nelt
- 1);
29226 for (i
= 0; i
< nelt
; i
+= 2)
29228 if (d
->perm
[i
] != i
+ odd
)
29230 if (d
->perm
[i
+ 1] != ((i
+ nelt
+ odd
) & mask
))
29240 case E_V16QImode
: gen
= gen_neon_vtrnv16qi_internal
; break;
29241 case E_V8QImode
: gen
= gen_neon_vtrnv8qi_internal
; break;
29242 case E_V8HImode
: gen
= gen_neon_vtrnv8hi_internal
; break;
29243 case E_V4HImode
: gen
= gen_neon_vtrnv4hi_internal
; break;
29244 case E_V8HFmode
: gen
= gen_neon_vtrnv8hf_internal
; break;
29245 case E_V4HFmode
: gen
= gen_neon_vtrnv4hf_internal
; break;
29246 case E_V4SImode
: gen
= gen_neon_vtrnv4si_internal
; break;
29247 case E_V2SImode
: gen
= gen_neon_vtrnv2si_internal
; break;
29248 case E_V2SFmode
: gen
= gen_neon_vtrnv2sf_internal
; break;
29249 case E_V4SFmode
: gen
= gen_neon_vtrnv4sf_internal
; break;
29251 gcc_unreachable ();
29256 if (BYTES_BIG_ENDIAN
)
29258 std::swap (in0
, in1
);
29263 out1
= gen_reg_rtx (d
->vmode
);
29265 std::swap (out0
, out1
);
29267 emit_insn (gen (out0
, in0
, in1
, out1
));
29271 /* Recognize patterns for the VEXT insns. */
29274 arm_evpc_neon_vext (struct expand_vec_perm_d
*d
)
29276 unsigned int i
, nelt
= d
->perm
.length ();
29277 rtx (*gen
) (rtx
, rtx
, rtx
, rtx
);
29280 unsigned int location
;
29282 unsigned int next
= d
->perm
[0] + 1;
29284 /* TODO: Handle GCC's numbering of elements for big-endian. */
29285 if (BYTES_BIG_ENDIAN
)
29288 /* Check if the extracted indexes are increasing by one. */
29289 for (i
= 1; i
< nelt
; next
++, i
++)
29291 /* If we hit the most significant element of the 2nd vector in
29292 the previous iteration, no need to test further. */
29293 if (next
== 2 * nelt
)
29296 /* If we are operating on only one vector: it could be a
29297 rotation. If there are only two elements of size < 64, let
29298 arm_evpc_neon_vrev catch it. */
29299 if (d
->one_vector_p
&& (next
== nelt
))
29301 if ((nelt
== 2) && (d
->vmode
!= V2DImode
))
29307 if (d
->perm
[i
] != next
)
29311 location
= d
->perm
[0];
29315 case E_V16QImode
: gen
= gen_neon_vextv16qi
; break;
29316 case E_V8QImode
: gen
= gen_neon_vextv8qi
; break;
29317 case E_V4HImode
: gen
= gen_neon_vextv4hi
; break;
29318 case E_V8HImode
: gen
= gen_neon_vextv8hi
; break;
29319 case E_V2SImode
: gen
= gen_neon_vextv2si
; break;
29320 case E_V4SImode
: gen
= gen_neon_vextv4si
; break;
29321 case E_V4HFmode
: gen
= gen_neon_vextv4hf
; break;
29322 case E_V8HFmode
: gen
= gen_neon_vextv8hf
; break;
29323 case E_V2SFmode
: gen
= gen_neon_vextv2sf
; break;
29324 case E_V4SFmode
: gen
= gen_neon_vextv4sf
; break;
29325 case E_V2DImode
: gen
= gen_neon_vextv2di
; break;
29334 offset
= GEN_INT (location
);
29335 emit_insn (gen (d
->target
, d
->op0
, d
->op1
, offset
));
29339 /* The NEON VTBL instruction is a fully variable permuation that's even
29340 stronger than what we expose via VEC_PERM_EXPR. What it doesn't do
29341 is mask the index operand as VEC_PERM_EXPR requires. Therefore we
29342 can do slightly better by expanding this as a constant where we don't
29343 have to apply a mask. */
29346 arm_evpc_neon_vtbl (struct expand_vec_perm_d
*d
)
29348 rtx rperm
[MAX_VECT_LEN
], sel
;
29349 machine_mode vmode
= d
->vmode
;
29350 unsigned int i
, nelt
= d
->perm
.length ();
29352 /* TODO: ARM's VTBL indexing is little-endian. In order to handle GCC's
29353 numbering of elements for big-endian, we must reverse the order. */
29354 if (BYTES_BIG_ENDIAN
)
29360 /* Generic code will try constant permutation twice. Once with the
29361 original mode and again with the elements lowered to QImode.
29362 So wait and don't do the selector expansion ourselves. */
29363 if (vmode
!= V8QImode
&& vmode
!= V16QImode
)
29366 for (i
= 0; i
< nelt
; ++i
)
29367 rperm
[i
] = GEN_INT (d
->perm
[i
]);
29368 sel
= gen_rtx_CONST_VECTOR (vmode
, gen_rtvec_v (nelt
, rperm
));
29369 sel
= force_reg (vmode
, sel
);
29371 arm_expand_vec_perm_1 (d
->target
, d
->op0
, d
->op1
, sel
);
29376 arm_expand_vec_perm_const_1 (struct expand_vec_perm_d
*d
)
29378 /* Check if the input mask matches vext before reordering the
29381 if (arm_evpc_neon_vext (d
))
29384 /* The pattern matching functions above are written to look for a small
29385 number to begin the sequence (0, 1, N/2). If we begin with an index
29386 from the second operand, we can swap the operands. */
29387 unsigned int nelt
= d
->perm
.length ();
29388 if (d
->perm
[0] >= nelt
)
29390 d
->perm
.rotate_inputs (1);
29391 std::swap (d
->op0
, d
->op1
);
29396 if (arm_evpc_neon_vuzp (d
))
29398 if (arm_evpc_neon_vzip (d
))
29400 if (arm_evpc_neon_vrev (d
))
29402 if (arm_evpc_neon_vtrn (d
))
29404 return arm_evpc_neon_vtbl (d
);
29409 /* Implement TARGET_VECTORIZE_VEC_PERM_CONST. */
29412 arm_vectorize_vec_perm_const (machine_mode vmode
, rtx target
, rtx op0
, rtx op1
,
29413 const vec_perm_indices
&sel
)
29415 struct expand_vec_perm_d d
;
29416 int i
, nelt
, which
;
29418 if (!VALID_NEON_DREG_MODE (vmode
) && !VALID_NEON_QREG_MODE (vmode
))
29426 gcc_assert (VECTOR_MODE_P (d
.vmode
));
29427 d
.testing_p
= !target
;
29429 nelt
= GET_MODE_NUNITS (d
.vmode
);
29430 for (i
= which
= 0; i
< nelt
; ++i
)
29432 int ei
= sel
[i
] & (2 * nelt
- 1);
29433 which
|= (ei
< nelt
? 1 : 2);
29442 d
.one_vector_p
= false;
29443 if (d
.testing_p
|| !rtx_equal_p (op0
, op1
))
29446 /* The elements of PERM do not suggest that only the first operand
29447 is used, but both operands are identical. Allow easier matching
29448 of the permutation by folding the permutation into the single
29453 d
.one_vector_p
= true;
29458 d
.one_vector_p
= true;
29462 d
.perm
.new_vector (sel
.encoding (), d
.one_vector_p
? 1 : 2, nelt
);
29465 return arm_expand_vec_perm_const_1 (&d
);
29467 d
.target
= gen_raw_REG (d
.vmode
, LAST_VIRTUAL_REGISTER
+ 1);
29468 d
.op1
= d
.op0
= gen_raw_REG (d
.vmode
, LAST_VIRTUAL_REGISTER
+ 2);
29469 if (!d
.one_vector_p
)
29470 d
.op1
= gen_raw_REG (d
.vmode
, LAST_VIRTUAL_REGISTER
+ 3);
29473 bool ret
= arm_expand_vec_perm_const_1 (&d
);
29480 arm_autoinc_modes_ok_p (machine_mode mode
, enum arm_auto_incmodes code
)
29482 /* If we are soft float and we do not have ldrd
29483 then all auto increment forms are ok. */
29484 if (TARGET_SOFT_FLOAT
&& (TARGET_LDRD
|| GET_MODE_SIZE (mode
) <= 4))
29489 /* Post increment and Pre Decrement are supported for all
29490 instruction forms except for vector forms. */
29493 if (VECTOR_MODE_P (mode
))
29495 if (code
!= ARM_PRE_DEC
)
29505 /* Without LDRD and mode size greater than
29506 word size, there is no point in auto-incrementing
29507 because ldm and stm will not have these forms. */
29508 if (!TARGET_LDRD
&& GET_MODE_SIZE (mode
) > 4)
29511 /* Vector and floating point modes do not support
29512 these auto increment forms. */
29513 if (FLOAT_MODE_P (mode
) || VECTOR_MODE_P (mode
))
29526 /* The default expansion of general 64-bit shifts in core-regs is suboptimal,
29527 on ARM, since we know that shifts by negative amounts are no-ops.
29528 Additionally, the default expansion code is not available or suitable
29529 for post-reload insn splits (this can occur when the register allocator
29530 chooses not to do a shift in NEON).
29532 This function is used in both initial expand and post-reload splits, and
29533 handles all kinds of 64-bit shifts.
29535 Input requirements:
29536 - It is safe for the input and output to be the same register, but
29537 early-clobber rules apply for the shift amount and scratch registers.
29538 - Shift by register requires both scratch registers. In all other cases
29539 the scratch registers may be NULL.
29540 - Ashiftrt by a register also clobbers the CC register. */
29542 arm_emit_coreregs_64bit_shift (enum rtx_code code
, rtx out
, rtx in
,
29543 rtx amount
, rtx scratch1
, rtx scratch2
)
29545 rtx out_high
= gen_highpart (SImode
, out
);
29546 rtx out_low
= gen_lowpart (SImode
, out
);
29547 rtx in_high
= gen_highpart (SImode
, in
);
29548 rtx in_low
= gen_lowpart (SImode
, in
);
29551 in = the register pair containing the input value.
29552 out = the destination register pair.
29553 up = the high- or low-part of each pair.
29554 down = the opposite part to "up".
29555 In a shift, we can consider bits to shift from "up"-stream to
29556 "down"-stream, so in a left-shift "up" is the low-part and "down"
29557 is the high-part of each register pair. */
29559 rtx out_up
= code
== ASHIFT
? out_low
: out_high
;
29560 rtx out_down
= code
== ASHIFT
? out_high
: out_low
;
29561 rtx in_up
= code
== ASHIFT
? in_low
: in_high
;
29562 rtx in_down
= code
== ASHIFT
? in_high
: in_low
;
29564 gcc_assert (code
== ASHIFT
|| code
== ASHIFTRT
|| code
== LSHIFTRT
);
29566 && (REG_P (out
) || GET_CODE (out
) == SUBREG
)
29567 && GET_MODE (out
) == DImode
);
29569 && (REG_P (in
) || GET_CODE (in
) == SUBREG
)
29570 && GET_MODE (in
) == DImode
);
29572 && (((REG_P (amount
) || GET_CODE (amount
) == SUBREG
)
29573 && GET_MODE (amount
) == SImode
)
29574 || CONST_INT_P (amount
)));
29575 gcc_assert (scratch1
== NULL
29576 || (GET_CODE (scratch1
) == SCRATCH
)
29577 || (GET_MODE (scratch1
) == SImode
29578 && REG_P (scratch1
)));
29579 gcc_assert (scratch2
== NULL
29580 || (GET_CODE (scratch2
) == SCRATCH
)
29581 || (GET_MODE (scratch2
) == SImode
29582 && REG_P (scratch2
)));
29583 gcc_assert (!REG_P (out
) || !REG_P (amount
)
29584 || !HARD_REGISTER_P (out
)
29585 || (REGNO (out
) != REGNO (amount
)
29586 && REGNO (out
) + 1 != REGNO (amount
)));
29588 /* Macros to make following code more readable. */
29589 #define SUB_32(DEST,SRC) \
29590 gen_addsi3 ((DEST), (SRC), GEN_INT (-32))
29591 #define RSB_32(DEST,SRC) \
29592 gen_subsi3 ((DEST), GEN_INT (32), (SRC))
29593 #define SUB_S_32(DEST,SRC) \
29594 gen_addsi3_compare0 ((DEST), (SRC), \
29596 #define SET(DEST,SRC) \
29597 gen_rtx_SET ((DEST), (SRC))
29598 #define SHIFT(CODE,SRC,AMOUNT) \
29599 gen_rtx_fmt_ee ((CODE), SImode, (SRC), (AMOUNT))
29600 #define LSHIFT(CODE,SRC,AMOUNT) \
29601 gen_rtx_fmt_ee ((CODE) == ASHIFT ? ASHIFT : LSHIFTRT, \
29602 SImode, (SRC), (AMOUNT))
29603 #define REV_LSHIFT(CODE,SRC,AMOUNT) \
29604 gen_rtx_fmt_ee ((CODE) == ASHIFT ? LSHIFTRT : ASHIFT, \
29605 SImode, (SRC), (AMOUNT))
29607 gen_rtx_IOR (SImode, (A), (B))
29608 #define BRANCH(COND,LABEL) \
29609 gen_arm_cond_branch ((LABEL), \
29610 gen_rtx_ ## COND (CCmode, cc_reg, \
29614 /* Shifts by register and shifts by constant are handled separately. */
29615 if (CONST_INT_P (amount
))
29617 /* We have a shift-by-constant. */
29619 /* First, handle out-of-range shift amounts.
29620 In both cases we try to match the result an ARM instruction in a
29621 shift-by-register would give. This helps reduce execution
29622 differences between optimization levels, but it won't stop other
29623 parts of the compiler doing different things. This is "undefined
29624 behavior, in any case. */
29625 if (INTVAL (amount
) <= 0)
29626 emit_insn (gen_movdi (out
, in
));
29627 else if (INTVAL (amount
) >= 64)
29629 if (code
== ASHIFTRT
)
29631 rtx const31_rtx
= GEN_INT (31);
29632 emit_insn (SET (out_down
, SHIFT (code
, in_up
, const31_rtx
)));
29633 emit_insn (SET (out_up
, SHIFT (code
, in_up
, const31_rtx
)));
29636 emit_insn (gen_movdi (out
, const0_rtx
));
29639 /* Now handle valid shifts. */
29640 else if (INTVAL (amount
) < 32)
29642 /* Shifts by a constant less than 32. */
29643 rtx reverse_amount
= GEN_INT (32 - INTVAL (amount
));
29645 /* Clearing the out register in DImode first avoids lots
29646 of spilling and results in less stack usage.
29647 Later this redundant insn is completely removed.
29648 Do that only if "in" and "out" are different registers. */
29649 if (REG_P (out
) && REG_P (in
) && REGNO (out
) != REGNO (in
))
29650 emit_insn (SET (out
, const0_rtx
));
29651 emit_insn (SET (out_down
, LSHIFT (code
, in_down
, amount
)));
29652 emit_insn (SET (out_down
,
29653 ORR (REV_LSHIFT (code
, in_up
, reverse_amount
),
29655 emit_insn (SET (out_up
, SHIFT (code
, in_up
, amount
)));
29659 /* Shifts by a constant greater than 31. */
29660 rtx adj_amount
= GEN_INT (INTVAL (amount
) - 32);
29662 if (REG_P (out
) && REG_P (in
) && REGNO (out
) != REGNO (in
))
29663 emit_insn (SET (out
, const0_rtx
));
29664 emit_insn (SET (out_down
, SHIFT (code
, in_up
, adj_amount
)));
29665 if (code
== ASHIFTRT
)
29666 emit_insn (gen_ashrsi3 (out_up
, in_up
,
29669 emit_insn (SET (out_up
, const0_rtx
));
29674 /* We have a shift-by-register. */
29675 rtx cc_reg
= gen_rtx_REG (CC_NOOVmode
, CC_REGNUM
);
29677 /* This alternative requires the scratch registers. */
29678 gcc_assert (scratch1
&& REG_P (scratch1
));
29679 gcc_assert (scratch2
&& REG_P (scratch2
));
29681 /* We will need the values "amount-32" and "32-amount" later.
29682 Swapping them around now allows the later code to be more general. */
29686 emit_insn (SUB_32 (scratch1
, amount
));
29687 emit_insn (RSB_32 (scratch2
, amount
));
29690 emit_insn (RSB_32 (scratch1
, amount
));
29691 /* Also set CC = amount > 32. */
29692 emit_insn (SUB_S_32 (scratch2
, amount
));
29695 emit_insn (RSB_32 (scratch1
, amount
));
29696 emit_insn (SUB_32 (scratch2
, amount
));
29699 gcc_unreachable ();
29702 /* Emit code like this:
29705 out_down = in_down << amount;
29706 out_down = (in_up << (amount - 32)) | out_down;
29707 out_down = ((unsigned)in_up >> (32 - amount)) | out_down;
29708 out_up = in_up << amount;
29711 out_down = in_down >> amount;
29712 out_down = (in_up << (32 - amount)) | out_down;
29714 out_down = ((signed)in_up >> (amount - 32)) | out_down;
29715 out_up = in_up << amount;
29718 out_down = in_down >> amount;
29719 out_down = (in_up << (32 - amount)) | out_down;
29721 out_down = ((unsigned)in_up >> (amount - 32)) | out_down;
29722 out_up = in_up << amount;
29724 The ARM and Thumb2 variants are the same but implemented slightly
29725 differently. If this were only called during expand we could just
29726 use the Thumb2 case and let combine do the right thing, but this
29727 can also be called from post-reload splitters. */
29729 emit_insn (SET (out_down
, LSHIFT (code
, in_down
, amount
)));
29731 if (!TARGET_THUMB2
)
29733 /* Emit code for ARM mode. */
29734 emit_insn (SET (out_down
,
29735 ORR (SHIFT (ASHIFT
, in_up
, scratch1
), out_down
)));
29736 if (code
== ASHIFTRT
)
29738 rtx_code_label
*done_label
= gen_label_rtx ();
29739 emit_jump_insn (BRANCH (LT
, done_label
));
29740 emit_insn (SET (out_down
, ORR (SHIFT (ASHIFTRT
, in_up
, scratch2
),
29742 emit_label (done_label
);
29745 emit_insn (SET (out_down
, ORR (SHIFT (LSHIFTRT
, in_up
, scratch2
),
29750 /* Emit code for Thumb2 mode.
29751 Thumb2 can't do shift and or in one insn. */
29752 emit_insn (SET (scratch1
, SHIFT (ASHIFT
, in_up
, scratch1
)));
29753 emit_insn (gen_iorsi3 (out_down
, out_down
, scratch1
));
29755 if (code
== ASHIFTRT
)
29757 rtx_code_label
*done_label
= gen_label_rtx ();
29758 emit_jump_insn (BRANCH (LT
, done_label
));
29759 emit_insn (SET (scratch2
, SHIFT (ASHIFTRT
, in_up
, scratch2
)));
29760 emit_insn (SET (out_down
, ORR (out_down
, scratch2
)));
29761 emit_label (done_label
);
29765 emit_insn (SET (scratch2
, SHIFT (LSHIFTRT
, in_up
, scratch2
)));
29766 emit_insn (gen_iorsi3 (out_down
, out_down
, scratch2
));
29770 emit_insn (SET (out_up
, SHIFT (code
, in_up
, amount
)));
29784 /* Returns true if the pattern is a valid symbolic address, which is either a
29785 symbol_ref or (symbol_ref + addend).
29787 According to the ARM ELF ABI, the initial addend of REL-type relocations
29788 processing MOVW and MOVT instructions is formed by interpreting the 16-bit
29789 literal field of the instruction as a 16-bit signed value in the range
29790 -32768 <= A < 32768. */
29793 arm_valid_symbolic_address_p (rtx addr
)
29795 rtx xop0
, xop1
= NULL_RTX
;
29798 if (target_word_relocations
)
29801 if (GET_CODE (tmp
) == SYMBOL_REF
|| GET_CODE (tmp
) == LABEL_REF
)
29804 /* (const (plus: symbol_ref const_int)) */
29805 if (GET_CODE (addr
) == CONST
)
29806 tmp
= XEXP (addr
, 0);
29808 if (GET_CODE (tmp
) == PLUS
)
29810 xop0
= XEXP (tmp
, 0);
29811 xop1
= XEXP (tmp
, 1);
29813 if (GET_CODE (xop0
) == SYMBOL_REF
&& CONST_INT_P (xop1
))
29814 return IN_RANGE (INTVAL (xop1
), -0x8000, 0x7fff);
29820 /* Returns true if a valid comparison operation and makes
29821 the operands in a form that is valid. */
29823 arm_validize_comparison (rtx
*comparison
, rtx
* op1
, rtx
* op2
)
29825 enum rtx_code code
= GET_CODE (*comparison
);
29827 machine_mode mode
= (GET_MODE (*op1
) == VOIDmode
)
29828 ? GET_MODE (*op2
) : GET_MODE (*op1
);
29830 gcc_assert (GET_MODE (*op1
) != VOIDmode
|| GET_MODE (*op2
) != VOIDmode
);
29832 if (code
== UNEQ
|| code
== LTGT
)
29835 code_int
= (int)code
;
29836 arm_canonicalize_comparison (&code_int
, op1
, op2
, 0);
29837 PUT_CODE (*comparison
, (enum rtx_code
)code_int
);
29842 if (!arm_add_operand (*op1
, mode
))
29843 *op1
= force_reg (mode
, *op1
);
29844 if (!arm_add_operand (*op2
, mode
))
29845 *op2
= force_reg (mode
, *op2
);
29849 if (!cmpdi_operand (*op1
, mode
))
29850 *op1
= force_reg (mode
, *op1
);
29851 if (!cmpdi_operand (*op2
, mode
))
29852 *op2
= force_reg (mode
, *op2
);
29856 if (!TARGET_VFP_FP16INST
)
29858 /* FP16 comparisons are done in SF mode. */
29860 *op1
= convert_to_mode (mode
, *op1
, 1);
29861 *op2
= convert_to_mode (mode
, *op2
, 1);
29862 /* Fall through. */
29865 if (!vfp_compare_operand (*op1
, mode
))
29866 *op1
= force_reg (mode
, *op1
);
29867 if (!vfp_compare_operand (*op2
, mode
))
29868 *op2
= force_reg (mode
, *op2
);
29878 /* Maximum number of instructions to set block of memory. */
29880 arm_block_set_max_insns (void)
29882 if (optimize_function_for_size_p (cfun
))
29885 return current_tune
->max_insns_inline_memset
;
29888 /* Return TRUE if it's profitable to set block of memory for
29889 non-vectorized case. VAL is the value to set the memory
29890 with. LENGTH is the number of bytes to set. ALIGN is the
29891 alignment of the destination memory in bytes. UNALIGNED_P
29892 is TRUE if we can only set the memory with instructions
29893 meeting alignment requirements. USE_STRD_P is TRUE if we
29894 can use strd to set the memory. */
29896 arm_block_set_non_vect_profit_p (rtx val
,
29897 unsigned HOST_WIDE_INT length
,
29898 unsigned HOST_WIDE_INT align
,
29899 bool unaligned_p
, bool use_strd_p
)
29902 /* For leftovers in bytes of 0-7, we can set the memory block using
29903 strb/strh/str with minimum instruction number. */
29904 const int leftover
[8] = {0, 1, 1, 2, 1, 2, 2, 3};
29908 num
= arm_const_inline_cost (SET
, val
);
29909 num
+= length
/ align
+ length
% align
;
29911 else if (use_strd_p
)
29913 num
= arm_const_double_inline_cost (val
);
29914 num
+= (length
>> 3) + leftover
[length
& 7];
29918 num
= arm_const_inline_cost (SET
, val
);
29919 num
+= (length
>> 2) + leftover
[length
& 3];
29922 /* We may be able to combine last pair STRH/STRB into a single STR
29923 by shifting one byte back. */
29924 if (unaligned_access
&& length
> 3 && (length
& 3) == 3)
29927 return (num
<= arm_block_set_max_insns ());
29930 /* Return TRUE if it's profitable to set block of memory for
29931 vectorized case. LENGTH is the number of bytes to set.
29932 ALIGN is the alignment of destination memory in bytes.
29933 MODE is the vector mode used to set the memory. */
29935 arm_block_set_vect_profit_p (unsigned HOST_WIDE_INT length
,
29936 unsigned HOST_WIDE_INT align
,
29940 bool unaligned_p
= ((align
& 3) != 0);
29941 unsigned int nelt
= GET_MODE_NUNITS (mode
);
29943 /* Instruction loading constant value. */
29945 /* Instructions storing the memory. */
29946 num
+= (length
+ nelt
- 1) / nelt
;
29947 /* Instructions adjusting the address expression. Only need to
29948 adjust address expression if it's 4 bytes aligned and bytes
29949 leftover can only be stored by mis-aligned store instruction. */
29950 if (!unaligned_p
&& (length
& 3) != 0)
29953 /* Store the first 16 bytes using vst1:v16qi for the aligned case. */
29954 if (!unaligned_p
&& mode
== V16QImode
)
29957 return (num
<= arm_block_set_max_insns ());
29960 /* Set a block of memory using vectorization instructions for the
29961 unaligned case. We fill the first LENGTH bytes of the memory
29962 area starting from DSTBASE with byte constant VALUE. ALIGN is
29963 the alignment requirement of memory. Return TRUE if succeeded. */
29965 arm_block_set_unaligned_vect (rtx dstbase
,
29966 unsigned HOST_WIDE_INT length
,
29967 unsigned HOST_WIDE_INT value
,
29968 unsigned HOST_WIDE_INT align
)
29970 unsigned int i
, nelt_v16
, nelt_v8
, nelt_mode
;
29973 rtx (*gen_func
) (rtx
, rtx
);
29975 unsigned HOST_WIDE_INT v
= value
;
29976 unsigned int offset
= 0;
29977 gcc_assert ((align
& 0x3) != 0);
29978 nelt_v8
= GET_MODE_NUNITS (V8QImode
);
29979 nelt_v16
= GET_MODE_NUNITS (V16QImode
);
29980 if (length
>= nelt_v16
)
29983 gen_func
= gen_movmisalignv16qi
;
29988 gen_func
= gen_movmisalignv8qi
;
29990 nelt_mode
= GET_MODE_NUNITS (mode
);
29991 gcc_assert (length
>= nelt_mode
);
29992 /* Skip if it isn't profitable. */
29993 if (!arm_block_set_vect_profit_p (length
, align
, mode
))
29996 dst
= copy_addr_to_reg (XEXP (dstbase
, 0));
29997 mem
= adjust_automodify_address (dstbase
, mode
, dst
, offset
);
29999 v
= sext_hwi (v
, BITS_PER_WORD
);
30001 reg
= gen_reg_rtx (mode
);
30002 val_vec
= gen_const_vec_duplicate (mode
, GEN_INT (v
));
30003 /* Emit instruction loading the constant value. */
30004 emit_move_insn (reg
, val_vec
);
30006 /* Handle nelt_mode bytes in a vector. */
30007 for (i
= 0; (i
+ nelt_mode
<= length
); i
+= nelt_mode
)
30009 emit_insn ((*gen_func
) (mem
, reg
));
30010 if (i
+ 2 * nelt_mode
<= length
)
30012 emit_insn (gen_add2_insn (dst
, GEN_INT (nelt_mode
)));
30013 offset
+= nelt_mode
;
30014 mem
= adjust_automodify_address (dstbase
, mode
, dst
, offset
);
30018 /* If there are not less than nelt_v8 bytes leftover, we must be in
30020 gcc_assert ((i
+ nelt_v8
) > length
|| mode
== V16QImode
);
30022 /* Handle (8, 16) bytes leftover. */
30023 if (i
+ nelt_v8
< length
)
30025 emit_insn (gen_add2_insn (dst
, GEN_INT (length
- i
)));
30026 offset
+= length
- i
;
30027 mem
= adjust_automodify_address (dstbase
, mode
, dst
, offset
);
30029 /* We are shifting bytes back, set the alignment accordingly. */
30030 if ((length
& 1) != 0 && align
>= 2)
30031 set_mem_align (mem
, BITS_PER_UNIT
);
30033 emit_insn (gen_movmisalignv16qi (mem
, reg
));
30035 /* Handle (0, 8] bytes leftover. */
30036 else if (i
< length
&& i
+ nelt_v8
>= length
)
30038 if (mode
== V16QImode
)
30039 reg
= gen_lowpart (V8QImode
, reg
);
30041 emit_insn (gen_add2_insn (dst
, GEN_INT ((length
- i
)
30042 + (nelt_mode
- nelt_v8
))));
30043 offset
+= (length
- i
) + (nelt_mode
- nelt_v8
);
30044 mem
= adjust_automodify_address (dstbase
, V8QImode
, dst
, offset
);
30046 /* We are shifting bytes back, set the alignment accordingly. */
30047 if ((length
& 1) != 0 && align
>= 2)
30048 set_mem_align (mem
, BITS_PER_UNIT
);
30050 emit_insn (gen_movmisalignv8qi (mem
, reg
));
30056 /* Set a block of memory using vectorization instructions for the
30057 aligned case. We fill the first LENGTH bytes of the memory area
30058 starting from DSTBASE with byte constant VALUE. ALIGN is the
30059 alignment requirement of memory. Return TRUE if succeeded. */
30061 arm_block_set_aligned_vect (rtx dstbase
,
30062 unsigned HOST_WIDE_INT length
,
30063 unsigned HOST_WIDE_INT value
,
30064 unsigned HOST_WIDE_INT align
)
30066 unsigned int i
, nelt_v8
, nelt_v16
, nelt_mode
;
30067 rtx dst
, addr
, mem
;
30070 unsigned int offset
= 0;
30072 gcc_assert ((align
& 0x3) == 0);
30073 nelt_v8
= GET_MODE_NUNITS (V8QImode
);
30074 nelt_v16
= GET_MODE_NUNITS (V16QImode
);
30075 if (length
>= nelt_v16
&& unaligned_access
&& !BYTES_BIG_ENDIAN
)
30080 nelt_mode
= GET_MODE_NUNITS (mode
);
30081 gcc_assert (length
>= nelt_mode
);
30082 /* Skip if it isn't profitable. */
30083 if (!arm_block_set_vect_profit_p (length
, align
, mode
))
30086 dst
= copy_addr_to_reg (XEXP (dstbase
, 0));
30088 reg
= gen_reg_rtx (mode
);
30089 val_vec
= gen_const_vec_duplicate (mode
, gen_int_mode (value
, QImode
));
30090 /* Emit instruction loading the constant value. */
30091 emit_move_insn (reg
, val_vec
);
30094 /* Handle first 16 bytes specially using vst1:v16qi instruction. */
30095 if (mode
== V16QImode
)
30097 mem
= adjust_automodify_address (dstbase
, mode
, dst
, offset
);
30098 emit_insn (gen_movmisalignv16qi (mem
, reg
));
30100 /* Handle (8, 16) bytes leftover using vst1:v16qi again. */
30101 if (i
+ nelt_v8
< length
&& i
+ nelt_v16
> length
)
30103 emit_insn (gen_add2_insn (dst
, GEN_INT (length
- nelt_mode
)));
30104 offset
+= length
- nelt_mode
;
30105 mem
= adjust_automodify_address (dstbase
, mode
, dst
, offset
);
30106 /* We are shifting bytes back, set the alignment accordingly. */
30107 if ((length
& 0x3) == 0)
30108 set_mem_align (mem
, BITS_PER_UNIT
* 4);
30109 else if ((length
& 0x1) == 0)
30110 set_mem_align (mem
, BITS_PER_UNIT
* 2);
30112 set_mem_align (mem
, BITS_PER_UNIT
);
30114 emit_insn (gen_movmisalignv16qi (mem
, reg
));
30117 /* Fall through for bytes leftover. */
30119 nelt_mode
= GET_MODE_NUNITS (mode
);
30120 reg
= gen_lowpart (V8QImode
, reg
);
30123 /* Handle 8 bytes in a vector. */
30124 for (; (i
+ nelt_mode
<= length
); i
+= nelt_mode
)
30126 addr
= plus_constant (Pmode
, dst
, i
);
30127 mem
= adjust_automodify_address (dstbase
, mode
, addr
, offset
+ i
);
30128 emit_move_insn (mem
, reg
);
30131 /* Handle single word leftover by shifting 4 bytes back. We can
30132 use aligned access for this case. */
30133 if (i
+ UNITS_PER_WORD
== length
)
30135 addr
= plus_constant (Pmode
, dst
, i
- UNITS_PER_WORD
);
30136 offset
+= i
- UNITS_PER_WORD
;
30137 mem
= adjust_automodify_address (dstbase
, mode
, addr
, offset
);
30138 /* We are shifting 4 bytes back, set the alignment accordingly. */
30139 if (align
> UNITS_PER_WORD
)
30140 set_mem_align (mem
, BITS_PER_UNIT
* UNITS_PER_WORD
);
30142 emit_move_insn (mem
, reg
);
30144 /* Handle (0, 4), (4, 8) bytes leftover by shifting bytes back.
30145 We have to use unaligned access for this case. */
30146 else if (i
< length
)
30148 emit_insn (gen_add2_insn (dst
, GEN_INT (length
- nelt_mode
)));
30149 offset
+= length
- nelt_mode
;
30150 mem
= adjust_automodify_address (dstbase
, mode
, dst
, offset
);
30151 /* We are shifting bytes back, set the alignment accordingly. */
30152 if ((length
& 1) == 0)
30153 set_mem_align (mem
, BITS_PER_UNIT
* 2);
30155 set_mem_align (mem
, BITS_PER_UNIT
);
30157 emit_insn (gen_movmisalignv8qi (mem
, reg
));
30163 /* Set a block of memory using plain strh/strb instructions, only
30164 using instructions allowed by ALIGN on processor. We fill the
30165 first LENGTH bytes of the memory area starting from DSTBASE
30166 with byte constant VALUE. ALIGN is the alignment requirement
30169 arm_block_set_unaligned_non_vect (rtx dstbase
,
30170 unsigned HOST_WIDE_INT length
,
30171 unsigned HOST_WIDE_INT value
,
30172 unsigned HOST_WIDE_INT align
)
30175 rtx dst
, addr
, mem
;
30176 rtx val_exp
, val_reg
, reg
;
30178 HOST_WIDE_INT v
= value
;
30180 gcc_assert (align
== 1 || align
== 2);
30183 v
|= (value
<< BITS_PER_UNIT
);
30185 v
= sext_hwi (v
, BITS_PER_WORD
);
30186 val_exp
= GEN_INT (v
);
30187 /* Skip if it isn't profitable. */
30188 if (!arm_block_set_non_vect_profit_p (val_exp
, length
,
30189 align
, true, false))
30192 dst
= copy_addr_to_reg (XEXP (dstbase
, 0));
30193 mode
= (align
== 2 ? HImode
: QImode
);
30194 val_reg
= force_reg (SImode
, val_exp
);
30195 reg
= gen_lowpart (mode
, val_reg
);
30197 for (i
= 0; (i
+ GET_MODE_SIZE (mode
) <= length
); i
+= GET_MODE_SIZE (mode
))
30199 addr
= plus_constant (Pmode
, dst
, i
);
30200 mem
= adjust_automodify_address (dstbase
, mode
, addr
, i
);
30201 emit_move_insn (mem
, reg
);
30204 /* Handle single byte leftover. */
30205 if (i
+ 1 == length
)
30207 reg
= gen_lowpart (QImode
, val_reg
);
30208 addr
= plus_constant (Pmode
, dst
, i
);
30209 mem
= adjust_automodify_address (dstbase
, QImode
, addr
, i
);
30210 emit_move_insn (mem
, reg
);
30214 gcc_assert (i
== length
);
30218 /* Set a block of memory using plain strd/str/strh/strb instructions,
30219 to permit unaligned copies on processors which support unaligned
30220 semantics for those instructions. We fill the first LENGTH bytes
30221 of the memory area starting from DSTBASE with byte constant VALUE.
30222 ALIGN is the alignment requirement of memory. */
30224 arm_block_set_aligned_non_vect (rtx dstbase
,
30225 unsigned HOST_WIDE_INT length
,
30226 unsigned HOST_WIDE_INT value
,
30227 unsigned HOST_WIDE_INT align
)
30230 rtx dst
, addr
, mem
;
30231 rtx val_exp
, val_reg
, reg
;
30232 unsigned HOST_WIDE_INT v
;
30235 use_strd_p
= (length
>= 2 * UNITS_PER_WORD
&& (align
& 3) == 0
30236 && TARGET_LDRD
&& current_tune
->prefer_ldrd_strd
);
30238 v
= (value
| (value
<< 8) | (value
<< 16) | (value
<< 24));
30239 if (length
< UNITS_PER_WORD
)
30240 v
&= (0xFFFFFFFF >> (UNITS_PER_WORD
- length
) * BITS_PER_UNIT
);
30243 v
|= (v
<< BITS_PER_WORD
);
30245 v
= sext_hwi (v
, BITS_PER_WORD
);
30247 val_exp
= GEN_INT (v
);
30248 /* Skip if it isn't profitable. */
30249 if (!arm_block_set_non_vect_profit_p (val_exp
, length
,
30250 align
, false, use_strd_p
))
30255 /* Try without strd. */
30256 v
= (v
>> BITS_PER_WORD
);
30257 v
= sext_hwi (v
, BITS_PER_WORD
);
30258 val_exp
= GEN_INT (v
);
30259 use_strd_p
= false;
30260 if (!arm_block_set_non_vect_profit_p (val_exp
, length
,
30261 align
, false, use_strd_p
))
30266 dst
= copy_addr_to_reg (XEXP (dstbase
, 0));
30267 /* Handle double words using strd if possible. */
30270 val_reg
= force_reg (DImode
, val_exp
);
30272 for (; (i
+ 8 <= length
); i
+= 8)
30274 addr
= plus_constant (Pmode
, dst
, i
);
30275 mem
= adjust_automodify_address (dstbase
, DImode
, addr
, i
);
30276 emit_move_insn (mem
, reg
);
30280 val_reg
= force_reg (SImode
, val_exp
);
30282 /* Handle words. */
30283 reg
= (use_strd_p
? gen_lowpart (SImode
, val_reg
) : val_reg
);
30284 for (; (i
+ 4 <= length
); i
+= 4)
30286 addr
= plus_constant (Pmode
, dst
, i
);
30287 mem
= adjust_automodify_address (dstbase
, SImode
, addr
, i
);
30288 if ((align
& 3) == 0)
30289 emit_move_insn (mem
, reg
);
30291 emit_insn (gen_unaligned_storesi (mem
, reg
));
30294 /* Merge last pair of STRH and STRB into a STR if possible. */
30295 if (unaligned_access
&& i
> 0 && (i
+ 3) == length
)
30297 addr
= plus_constant (Pmode
, dst
, i
- 1);
30298 mem
= adjust_automodify_address (dstbase
, SImode
, addr
, i
- 1);
30299 /* We are shifting one byte back, set the alignment accordingly. */
30300 if ((align
& 1) == 0)
30301 set_mem_align (mem
, BITS_PER_UNIT
);
30303 /* Most likely this is an unaligned access, and we can't tell at
30304 compilation time. */
30305 emit_insn (gen_unaligned_storesi (mem
, reg
));
30309 /* Handle half word leftover. */
30310 if (i
+ 2 <= length
)
30312 reg
= gen_lowpart (HImode
, val_reg
);
30313 addr
= plus_constant (Pmode
, dst
, i
);
30314 mem
= adjust_automodify_address (dstbase
, HImode
, addr
, i
);
30315 if ((align
& 1) == 0)
30316 emit_move_insn (mem
, reg
);
30318 emit_insn (gen_unaligned_storehi (mem
, reg
));
30323 /* Handle single byte leftover. */
30324 if (i
+ 1 == length
)
30326 reg
= gen_lowpart (QImode
, val_reg
);
30327 addr
= plus_constant (Pmode
, dst
, i
);
30328 mem
= adjust_automodify_address (dstbase
, QImode
, addr
, i
);
30329 emit_move_insn (mem
, reg
);
30335 /* Set a block of memory using vectorization instructions for both
30336 aligned and unaligned cases. We fill the first LENGTH bytes of
30337 the memory area starting from DSTBASE with byte constant VALUE.
30338 ALIGN is the alignment requirement of memory. */
30340 arm_block_set_vect (rtx dstbase
,
30341 unsigned HOST_WIDE_INT length
,
30342 unsigned HOST_WIDE_INT value
,
30343 unsigned HOST_WIDE_INT align
)
30345 /* Check whether we need to use unaligned store instruction. */
30346 if (((align
& 3) != 0 || (length
& 3) != 0)
30347 /* Check whether unaligned store instruction is available. */
30348 && (!unaligned_access
|| BYTES_BIG_ENDIAN
))
30351 if ((align
& 3) == 0)
30352 return arm_block_set_aligned_vect (dstbase
, length
, value
, align
);
30354 return arm_block_set_unaligned_vect (dstbase
, length
, value
, align
);
30357 /* Expand string store operation. Firstly we try to do that by using
30358 vectorization instructions, then try with ARM unaligned access and
30359 double-word store if profitable. OPERANDS[0] is the destination,
30360 OPERANDS[1] is the number of bytes, operands[2] is the value to
30361 initialize the memory, OPERANDS[3] is the known alignment of the
30364 arm_gen_setmem (rtx
*operands
)
30366 rtx dstbase
= operands
[0];
30367 unsigned HOST_WIDE_INT length
;
30368 unsigned HOST_WIDE_INT value
;
30369 unsigned HOST_WIDE_INT align
;
30371 if (!CONST_INT_P (operands
[2]) || !CONST_INT_P (operands
[1]))
30374 length
= UINTVAL (operands
[1]);
30378 value
= (UINTVAL (operands
[2]) & 0xFF);
30379 align
= UINTVAL (operands
[3]);
30380 if (TARGET_NEON
&& length
>= 8
30381 && current_tune
->string_ops_prefer_neon
30382 && arm_block_set_vect (dstbase
, length
, value
, align
))
30385 if (!unaligned_access
&& (align
& 3) != 0)
30386 return arm_block_set_unaligned_non_vect (dstbase
, length
, value
, align
);
30388 return arm_block_set_aligned_non_vect (dstbase
, length
, value
, align
);
30393 arm_macro_fusion_p (void)
30395 return current_tune
->fusible_ops
!= tune_params::FUSE_NOTHING
;
30398 /* Return true if the two back-to-back sets PREV_SET, CURR_SET are suitable
30399 for MOVW / MOVT macro fusion. */
30402 arm_sets_movw_movt_fusible_p (rtx prev_set
, rtx curr_set
)
30404 /* We are trying to fuse
30405 movw imm / movt imm
30406 instructions as a group that gets scheduled together. */
30408 rtx set_dest
= SET_DEST (curr_set
);
30410 if (GET_MODE (set_dest
) != SImode
)
30413 /* We are trying to match:
30414 prev (movw) == (set (reg r0) (const_int imm16))
30415 curr (movt) == (set (zero_extract (reg r0)
30418 (const_int imm16_1))
30420 prev (movw) == (set (reg r1)
30421 (high (symbol_ref ("SYM"))))
30422 curr (movt) == (set (reg r0)
30424 (symbol_ref ("SYM")))) */
30426 if (GET_CODE (set_dest
) == ZERO_EXTRACT
)
30428 if (CONST_INT_P (SET_SRC (curr_set
))
30429 && CONST_INT_P (SET_SRC (prev_set
))
30430 && REG_P (XEXP (set_dest
, 0))
30431 && REG_P (SET_DEST (prev_set
))
30432 && REGNO (XEXP (set_dest
, 0)) == REGNO (SET_DEST (prev_set
)))
30436 else if (GET_CODE (SET_SRC (curr_set
)) == LO_SUM
30437 && REG_P (SET_DEST (curr_set
))
30438 && REG_P (SET_DEST (prev_set
))
30439 && GET_CODE (SET_SRC (prev_set
)) == HIGH
30440 && REGNO (SET_DEST (curr_set
)) == REGNO (SET_DEST (prev_set
)))
30447 aarch_macro_fusion_pair_p (rtx_insn
* prev
, rtx_insn
* curr
)
30449 rtx prev_set
= single_set (prev
);
30450 rtx curr_set
= single_set (curr
);
30456 if (any_condjump_p (curr
))
30459 if (!arm_macro_fusion_p ())
30462 if (current_tune
->fusible_ops
& tune_params::FUSE_AES_AESMC
30463 && aarch_crypto_can_dual_issue (prev
, curr
))
30466 if (current_tune
->fusible_ops
& tune_params::FUSE_MOVW_MOVT
30467 && arm_sets_movw_movt_fusible_p (prev_set
, curr_set
))
30473 /* Return true iff the instruction fusion described by OP is enabled. */
30475 arm_fusion_enabled_p (tune_params::fuse_ops op
)
30477 return current_tune
->fusible_ops
& op
;
30480 /* Implement TARGET_SCHED_CAN_SPECULATE_INSN. Return true if INSN can be
30481 scheduled for speculative execution. Reject the long-running division
30482 and square-root instructions. */
30485 arm_sched_can_speculate_insn (rtx_insn
*insn
)
30487 switch (get_attr_type (insn
))
30495 case TYPE_NEON_FP_SQRT_S
:
30496 case TYPE_NEON_FP_SQRT_D
:
30497 case TYPE_NEON_FP_SQRT_S_Q
:
30498 case TYPE_NEON_FP_SQRT_D_Q
:
30499 case TYPE_NEON_FP_DIV_S
:
30500 case TYPE_NEON_FP_DIV_D
:
30501 case TYPE_NEON_FP_DIV_S_Q
:
30502 case TYPE_NEON_FP_DIV_D_Q
:
30509 /* Implement the TARGET_ASAN_SHADOW_OFFSET hook. */
30511 static unsigned HOST_WIDE_INT
30512 arm_asan_shadow_offset (void)
30514 return HOST_WIDE_INT_1U
<< 29;
30518 /* This is a temporary fix for PR60655. Ideally we need
30519 to handle most of these cases in the generic part but
30520 currently we reject minus (..) (sym_ref). We try to
30521 ameliorate the case with minus (sym_ref1) (sym_ref2)
30522 where they are in the same section. */
30525 arm_const_not_ok_for_debug_p (rtx p
)
30527 tree decl_op0
= NULL
;
30528 tree decl_op1
= NULL
;
30530 if (GET_CODE (p
) == UNSPEC
)
30532 if (GET_CODE (p
) == MINUS
)
30534 if (GET_CODE (XEXP (p
, 1)) == SYMBOL_REF
)
30536 decl_op1
= SYMBOL_REF_DECL (XEXP (p
, 1));
30538 && GET_CODE (XEXP (p
, 0)) == SYMBOL_REF
30539 && (decl_op0
= SYMBOL_REF_DECL (XEXP (p
, 0))))
30541 if ((VAR_P (decl_op1
)
30542 || TREE_CODE (decl_op1
) == CONST_DECL
)
30543 && (VAR_P (decl_op0
)
30544 || TREE_CODE (decl_op0
) == CONST_DECL
))
30545 return (get_variable_section (decl_op1
, false)
30546 != get_variable_section (decl_op0
, false));
30548 if (TREE_CODE (decl_op1
) == LABEL_DECL
30549 && TREE_CODE (decl_op0
) == LABEL_DECL
)
30550 return (DECL_CONTEXT (decl_op1
)
30551 != DECL_CONTEXT (decl_op0
));
30561 /* return TRUE if x is a reference to a value in a constant pool */
30563 arm_is_constant_pool_ref (rtx x
)
30566 && GET_CODE (XEXP (x
, 0)) == SYMBOL_REF
30567 && CONSTANT_POOL_ADDRESS_P (XEXP (x
, 0)));
30570 /* Remember the last target of arm_set_current_function. */
30571 static GTY(()) tree arm_previous_fndecl
;
30573 /* Restore or save the TREE_TARGET_GLOBALS from or to NEW_TREE. */
30576 save_restore_target_globals (tree new_tree
)
30578 /* If we have a previous state, use it. */
30579 if (TREE_TARGET_GLOBALS (new_tree
))
30580 restore_target_globals (TREE_TARGET_GLOBALS (new_tree
));
30581 else if (new_tree
== target_option_default_node
)
30582 restore_target_globals (&default_target_globals
);
30585 /* Call target_reinit and save the state for TARGET_GLOBALS. */
30586 TREE_TARGET_GLOBALS (new_tree
) = save_target_globals_default_opts ();
30589 arm_option_params_internal ();
30592 /* Invalidate arm_previous_fndecl. */
30595 arm_reset_previous_fndecl (void)
30597 arm_previous_fndecl
= NULL_TREE
;
30600 /* Establish appropriate back-end context for processing the function
30601 FNDECL. The argument might be NULL to indicate processing at top
30602 level, outside of any function scope. */
30605 arm_set_current_function (tree fndecl
)
30607 if (!fndecl
|| fndecl
== arm_previous_fndecl
)
30610 tree old_tree
= (arm_previous_fndecl
30611 ? DECL_FUNCTION_SPECIFIC_TARGET (arm_previous_fndecl
)
30614 tree new_tree
= DECL_FUNCTION_SPECIFIC_TARGET (fndecl
);
30616 /* If current function has no attributes but previous one did,
30617 use the default node. */
30618 if (! new_tree
&& old_tree
)
30619 new_tree
= target_option_default_node
;
30621 /* If nothing to do return. #pragma GCC reset or #pragma GCC pop to
30622 the default have been handled by save_restore_target_globals from
30623 arm_pragma_target_parse. */
30624 if (old_tree
== new_tree
)
30627 arm_previous_fndecl
= fndecl
;
30629 /* First set the target options. */
30630 cl_target_option_restore (&global_options
, TREE_TARGET_OPTION (new_tree
));
30632 save_restore_target_globals (new_tree
);
30635 /* Implement TARGET_OPTION_PRINT. */
30638 arm_option_print (FILE *file
, int indent
, struct cl_target_option
*ptr
)
30640 int flags
= ptr
->x_target_flags
;
30641 const char *fpu_name
;
30643 fpu_name
= (ptr
->x_arm_fpu_index
== TARGET_FPU_auto
30644 ? "auto" : all_fpus
[ptr
->x_arm_fpu_index
].name
);
30646 fprintf (file
, "%*sselected isa %s\n", indent
, "",
30647 TARGET_THUMB2_P (flags
) ? "thumb2" :
30648 TARGET_THUMB_P (flags
) ? "thumb1" :
30651 if (ptr
->x_arm_arch_string
)
30652 fprintf (file
, "%*sselected architecture %s\n", indent
, "",
30653 ptr
->x_arm_arch_string
);
30655 if (ptr
->x_arm_cpu_string
)
30656 fprintf (file
, "%*sselected CPU %s\n", indent
, "",
30657 ptr
->x_arm_cpu_string
);
30659 if (ptr
->x_arm_tune_string
)
30660 fprintf (file
, "%*sselected tune %s\n", indent
, "",
30661 ptr
->x_arm_tune_string
);
30663 fprintf (file
, "%*sselected fpu %s\n", indent
, "", fpu_name
);
30666 /* Hook to determine if one function can safely inline another. */
30669 arm_can_inline_p (tree caller
, tree callee
)
30671 tree caller_tree
= DECL_FUNCTION_SPECIFIC_TARGET (caller
);
30672 tree callee_tree
= DECL_FUNCTION_SPECIFIC_TARGET (callee
);
30673 bool can_inline
= true;
30675 struct cl_target_option
*caller_opts
30676 = TREE_TARGET_OPTION (caller_tree
? caller_tree
30677 : target_option_default_node
);
30679 struct cl_target_option
*callee_opts
30680 = TREE_TARGET_OPTION (callee_tree
? callee_tree
30681 : target_option_default_node
);
30683 if (callee_opts
== caller_opts
)
30686 /* Callee's ISA features should be a subset of the caller's. */
30687 struct arm_build_target caller_target
;
30688 struct arm_build_target callee_target
;
30689 caller_target
.isa
= sbitmap_alloc (isa_num_bits
);
30690 callee_target
.isa
= sbitmap_alloc (isa_num_bits
);
30692 arm_configure_build_target (&caller_target
, caller_opts
, &global_options_set
,
30694 arm_configure_build_target (&callee_target
, callee_opts
, &global_options_set
,
30696 if (!bitmap_subset_p (callee_target
.isa
, caller_target
.isa
))
30697 can_inline
= false;
30699 sbitmap_free (caller_target
.isa
);
30700 sbitmap_free (callee_target
.isa
);
30702 /* OK to inline between different modes.
30703 Function with mode specific instructions, e.g using asm,
30704 must be explicitly protected with noinline. */
30708 /* Hook to fix function's alignment affected by target attribute. */
30711 arm_relayout_function (tree fndecl
)
30713 if (DECL_USER_ALIGN (fndecl
))
30716 tree callee_tree
= DECL_FUNCTION_SPECIFIC_TARGET (fndecl
);
30719 callee_tree
= target_option_default_node
;
30721 struct cl_target_option
*opts
= TREE_TARGET_OPTION (callee_tree
);
30724 FUNCTION_ALIGNMENT (FUNCTION_BOUNDARY_P (opts
->x_target_flags
)));
30727 /* Inner function to process the attribute((target(...))), take an argument and
30728 set the current options from the argument. If we have a list, recursively
30729 go over the list. */
30732 arm_valid_target_attribute_rec (tree args
, struct gcc_options
*opts
)
30734 if (TREE_CODE (args
) == TREE_LIST
)
30738 for (; args
; args
= TREE_CHAIN (args
))
30739 if (TREE_VALUE (args
)
30740 && !arm_valid_target_attribute_rec (TREE_VALUE (args
), opts
))
30745 else if (TREE_CODE (args
) != STRING_CST
)
30747 error ("attribute %<target%> argument not a string");
30751 char *argstr
= ASTRDUP (TREE_STRING_POINTER (args
));
30754 while ((q
= strtok (argstr
, ",")) != NULL
)
30756 while (ISSPACE (*q
)) ++q
;
30759 if (!strncmp (q
, "thumb", 5))
30760 opts
->x_target_flags
|= MASK_THUMB
;
30762 else if (!strncmp (q
, "arm", 3))
30763 opts
->x_target_flags
&= ~MASK_THUMB
;
30765 else if (!strncmp (q
, "fpu=", 4))
30768 if (! opt_enum_arg_to_value (OPT_mfpu_
, q
+4,
30769 &fpu_index
, CL_TARGET
))
30771 error ("invalid fpu for target attribute or pragma %qs", q
);
30774 if (fpu_index
== TARGET_FPU_auto
)
30776 /* This doesn't really make sense until we support
30777 general dynamic selection of the architecture and all
30779 sorry ("auto fpu selection not currently permitted here");
30782 opts
->x_arm_fpu_index
= (enum fpu_type
) fpu_index
;
30784 else if (!strncmp (q
, "arch=", 5))
30787 const arch_option
*arm_selected_arch
30788 = arm_parse_arch_option_name (all_architectures
, "arch", arch
);
30790 if (!arm_selected_arch
)
30792 error ("invalid architecture for target attribute or pragma %qs",
30797 opts
->x_arm_arch_string
= xstrndup (arch
, strlen (arch
));
30799 else if (q
[0] == '+')
30801 opts
->x_arm_arch_string
30802 = xasprintf ("%s%s", opts
->x_arm_arch_string
, q
);
30806 error ("unknown target attribute or pragma %qs", q
);
30814 /* Return a TARGET_OPTION_NODE tree of the target options listed or NULL. */
30817 arm_valid_target_attribute_tree (tree args
, struct gcc_options
*opts
,
30818 struct gcc_options
*opts_set
)
30820 struct cl_target_option cl_opts
;
30822 if (!arm_valid_target_attribute_rec (args
, opts
))
30825 cl_target_option_save (&cl_opts
, opts
);
30826 arm_configure_build_target (&arm_active_target
, &cl_opts
, opts_set
, false);
30827 arm_option_check_internal (opts
);
30828 /* Do any overrides, such as global options arch=xxx.
30829 We do this since arm_active_target was overridden. */
30830 arm_option_reconfigure_globals ();
30831 arm_options_perform_arch_sanity_checks ();
30832 arm_option_override_internal (opts
, opts_set
);
30834 return build_target_option_node (opts
);
30838 add_attribute (const char * mode
, tree
*attributes
)
30840 size_t len
= strlen (mode
);
30841 tree value
= build_string (len
, mode
);
30843 TREE_TYPE (value
) = build_array_type (char_type_node
,
30844 build_index_type (size_int (len
)));
30846 *attributes
= tree_cons (get_identifier ("target"),
30847 build_tree_list (NULL_TREE
, value
),
30851 /* For testing. Insert thumb or arm modes alternatively on functions. */
30854 arm_insert_attributes (tree fndecl
, tree
* attributes
)
30858 if (! TARGET_FLIP_THUMB
)
30861 if (TREE_CODE (fndecl
) != FUNCTION_DECL
|| DECL_EXTERNAL(fndecl
)
30862 || fndecl_built_in_p (fndecl
) || DECL_ARTIFICIAL (fndecl
))
30865 /* Nested definitions must inherit mode. */
30866 if (current_function_decl
)
30868 mode
= TARGET_THUMB
? "thumb" : "arm";
30869 add_attribute (mode
, attributes
);
30873 /* If there is already a setting don't change it. */
30874 if (lookup_attribute ("target", *attributes
) != NULL
)
30877 mode
= thumb_flipper
? "thumb" : "arm";
30878 add_attribute (mode
, attributes
);
30880 thumb_flipper
= !thumb_flipper
;
30883 /* Hook to validate attribute((target("string"))). */
30886 arm_valid_target_attribute_p (tree fndecl
, tree
ARG_UNUSED (name
),
30887 tree args
, int ARG_UNUSED (flags
))
30890 struct gcc_options func_options
;
30891 tree cur_tree
, new_optimize
;
30892 gcc_assert ((fndecl
!= NULL_TREE
) && (args
!= NULL_TREE
));
30894 /* Get the optimization options of the current function. */
30895 tree func_optimize
= DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl
);
30897 /* If the function changed the optimization levels as well as setting target
30898 options, start with the optimizations specified. */
30899 if (!func_optimize
)
30900 func_optimize
= optimization_default_node
;
30902 /* Init func_options. */
30903 memset (&func_options
, 0, sizeof (func_options
));
30904 init_options_struct (&func_options
, NULL
);
30905 lang_hooks
.init_options_struct (&func_options
);
30907 /* Initialize func_options to the defaults. */
30908 cl_optimization_restore (&func_options
,
30909 TREE_OPTIMIZATION (func_optimize
));
30911 cl_target_option_restore (&func_options
,
30912 TREE_TARGET_OPTION (target_option_default_node
));
30914 /* Set func_options flags with new target mode. */
30915 cur_tree
= arm_valid_target_attribute_tree (args
, &func_options
,
30916 &global_options_set
);
30918 if (cur_tree
== NULL_TREE
)
30921 new_optimize
= build_optimization_node (&func_options
);
30923 DECL_FUNCTION_SPECIFIC_TARGET (fndecl
) = cur_tree
;
30925 DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl
) = new_optimize
;
30927 finalize_options_struct (&func_options
);
30932 /* Match an ISA feature bitmap to a named FPU. We always use the
30933 first entry that exactly matches the feature set, so that we
30934 effectively canonicalize the FPU name for the assembler. */
30936 arm_identify_fpu_from_isa (sbitmap isa
)
30938 auto_sbitmap
fpubits (isa_num_bits
);
30939 auto_sbitmap
cand_fpubits (isa_num_bits
);
30941 bitmap_and (fpubits
, isa
, isa_all_fpubits
);
30943 /* If there are no ISA feature bits relating to the FPU, we must be
30944 doing soft-float. */
30945 if (bitmap_empty_p (fpubits
))
30948 for (unsigned int i
= 0; i
< TARGET_FPU_auto
; i
++)
30950 arm_initialize_isa (cand_fpubits
, all_fpus
[i
].isa_bits
);
30951 if (bitmap_equal_p (fpubits
, cand_fpubits
))
30952 return all_fpus
[i
].name
;
30954 /* We must find an entry, or things have gone wrong. */
30955 gcc_unreachable ();
30958 /* Implement ASM_DECLARE_FUNCTION_NAME. Output the ISA features used
30959 by the function fndecl. */
30961 arm_declare_function_name (FILE *stream
, const char *name
, tree decl
)
30963 tree target_parts
= DECL_FUNCTION_SPECIFIC_TARGET (decl
);
30965 struct cl_target_option
*targ_options
;
30967 targ_options
= TREE_TARGET_OPTION (target_parts
);
30969 targ_options
= TREE_TARGET_OPTION (target_option_current_node
);
30970 gcc_assert (targ_options
);
30972 /* Only update the assembler .arch string if it is distinct from the last
30973 such string we printed. arch_to_print is set conditionally in case
30974 targ_options->x_arm_arch_string is NULL which can be the case
30975 when cc1 is invoked directly without passing -march option. */
30976 std::string arch_to_print
;
30977 if (targ_options
->x_arm_arch_string
)
30978 arch_to_print
= targ_options
->x_arm_arch_string
;
30980 if (arch_to_print
!= arm_last_printed_arch_string
)
30982 std::string arch_name
30983 = arch_to_print
.substr (0, arch_to_print
.find ("+"));
30984 asm_fprintf (asm_out_file
, "\t.arch %s\n", arch_name
.c_str ());
30985 const arch_option
*arch
30986 = arm_parse_arch_option_name (all_architectures
, "-march",
30987 targ_options
->x_arm_arch_string
);
30988 auto_sbitmap
opt_bits (isa_num_bits
);
30991 if (arch
->common
.extensions
)
30993 for (const struct cpu_arch_extension
*opt
= arch
->common
.extensions
;
30999 arm_initialize_isa (opt_bits
, opt
->isa_bits
);
31000 if (bitmap_subset_p (opt_bits
, arm_active_target
.isa
)
31001 && !bitmap_subset_p (opt_bits
, isa_all_fpubits
))
31002 asm_fprintf (asm_out_file
, "\t.arch_extension %s\n",
31008 arm_last_printed_arch_string
= arch_to_print
;
31011 fprintf (stream
, "\t.syntax unified\n");
31015 if (is_called_in_ARM_mode (decl
)
31016 || (TARGET_THUMB1
&& !TARGET_THUMB1_ONLY
31017 && cfun
->is_thunk
))
31018 fprintf (stream
, "\t.code 32\n");
31019 else if (TARGET_THUMB1
)
31020 fprintf (stream
, "\t.code\t16\n\t.thumb_func\n");
31022 fprintf (stream
, "\t.thumb\n\t.thumb_func\n");
31025 fprintf (stream
, "\t.arm\n");
31027 std::string fpu_to_print
31028 = TARGET_SOFT_FLOAT
31029 ? "softvfp" : arm_identify_fpu_from_isa (arm_active_target
.isa
);
31031 if (fpu_to_print
!= arm_last_printed_arch_string
)
31033 asm_fprintf (asm_out_file
, "\t.fpu %s\n", fpu_to_print
.c_str ());
31034 arm_last_printed_fpu_string
= fpu_to_print
;
31037 if (TARGET_POKE_FUNCTION_NAME
)
31038 arm_poke_function_name (stream
, (const char *) name
);
31041 /* If MEM is in the form of [base+offset], extract the two parts
31042 of address and set to BASE and OFFSET, otherwise return false
31043 after clearing BASE and OFFSET. */
31046 extract_base_offset_in_addr (rtx mem
, rtx
*base
, rtx
*offset
)
31050 gcc_assert (MEM_P (mem
));
31052 addr
= XEXP (mem
, 0);
31054 /* Strip off const from addresses like (const (addr)). */
31055 if (GET_CODE (addr
) == CONST
)
31056 addr
= XEXP (addr
, 0);
31058 if (GET_CODE (addr
) == REG
)
31061 *offset
= const0_rtx
;
31065 if (GET_CODE (addr
) == PLUS
31066 && GET_CODE (XEXP (addr
, 0)) == REG
31067 && CONST_INT_P (XEXP (addr
, 1)))
31069 *base
= XEXP (addr
, 0);
31070 *offset
= XEXP (addr
, 1);
31075 *offset
= NULL_RTX
;
31080 /* If INSN is a load or store of address in the form of [base+offset],
31081 extract the two parts and set to BASE and OFFSET. IS_LOAD is set
31082 to TRUE if it's a load. Return TRUE if INSN is such an instruction,
31083 otherwise return FALSE. */
31086 fusion_load_store (rtx_insn
*insn
, rtx
*base
, rtx
*offset
, bool *is_load
)
31090 gcc_assert (INSN_P (insn
));
31091 x
= PATTERN (insn
);
31092 if (GET_CODE (x
) != SET
)
31096 dest
= SET_DEST (x
);
31097 if (GET_CODE (src
) == REG
&& GET_CODE (dest
) == MEM
)
31100 extract_base_offset_in_addr (dest
, base
, offset
);
31102 else if (GET_CODE (src
) == MEM
&& GET_CODE (dest
) == REG
)
31105 extract_base_offset_in_addr (src
, base
, offset
);
31110 return (*base
!= NULL_RTX
&& *offset
!= NULL_RTX
);
31113 /* Implement the TARGET_SCHED_FUSION_PRIORITY hook.
31115 Currently we only support to fuse ldr or str instructions, so FUSION_PRI
31116 and PRI are only calculated for these instructions. For other instruction,
31117 FUSION_PRI and PRI are simply set to MAX_PRI. In the future, other kind
31118 instruction fusion can be supported by returning different priorities.
31120 It's important that irrelevant instructions get the largest FUSION_PRI. */
31123 arm_sched_fusion_priority (rtx_insn
*insn
, int max_pri
,
31124 int *fusion_pri
, int *pri
)
31130 gcc_assert (INSN_P (insn
));
31133 if (!fusion_load_store (insn
, &base
, &offset
, &is_load
))
31140 /* Load goes first. */
31142 *fusion_pri
= tmp
- 1;
31144 *fusion_pri
= tmp
- 2;
31148 /* INSN with smaller base register goes first. */
31149 tmp
-= ((REGNO (base
) & 0xff) << 20);
31151 /* INSN with smaller offset goes first. */
31152 off_val
= (int)(INTVAL (offset
));
31154 tmp
-= (off_val
& 0xfffff);
31156 tmp
+= ((- off_val
) & 0xfffff);
31163 /* Construct and return a PARALLEL RTX vector with elements numbering the
31164 lanes of either the high (HIGH == TRUE) or low (HIGH == FALSE) half of
31165 the vector - from the perspective of the architecture. This does not
31166 line up with GCC's perspective on lane numbers, so we end up with
31167 different masks depending on our target endian-ness. The diagram
31168 below may help. We must draw the distinction when building masks
31169 which select one half of the vector. An instruction selecting
31170 architectural low-lanes for a big-endian target, must be described using
31171 a mask selecting GCC high-lanes.
31173 Big-Endian Little-Endian
31175 GCC 0 1 2 3 3 2 1 0
31176 | x | x | x | x | | x | x | x | x |
31177 Architecture 3 2 1 0 3 2 1 0
31179 Low Mask: { 2, 3 } { 0, 1 }
31180 High Mask: { 0, 1 } { 2, 3 }
31184 arm_simd_vect_par_cnst_half (machine_mode mode
, bool high
)
31186 int nunits
= GET_MODE_NUNITS (mode
);
31187 rtvec v
= rtvec_alloc (nunits
/ 2);
31188 int high_base
= nunits
/ 2;
31194 if (BYTES_BIG_ENDIAN
)
31195 base
= high
? low_base
: high_base
;
31197 base
= high
? high_base
: low_base
;
31199 for (i
= 0; i
< nunits
/ 2; i
++)
31200 RTVEC_ELT (v
, i
) = GEN_INT (base
+ i
);
31202 t1
= gen_rtx_PARALLEL (mode
, v
);
31206 /* Check OP for validity as a PARALLEL RTX vector with elements
31207 numbering the lanes of either the high (HIGH == TRUE) or low lanes,
31208 from the perspective of the architecture. See the diagram above
31209 arm_simd_vect_par_cnst_half_p for more details. */
31212 arm_simd_check_vect_par_cnst_half_p (rtx op
, machine_mode mode
,
31215 rtx ideal
= arm_simd_vect_par_cnst_half (mode
, high
);
31216 HOST_WIDE_INT count_op
= XVECLEN (op
, 0);
31217 HOST_WIDE_INT count_ideal
= XVECLEN (ideal
, 0);
31220 if (!VECTOR_MODE_P (mode
))
31223 if (count_op
!= count_ideal
)
31226 for (i
= 0; i
< count_ideal
; i
++)
31228 rtx elt_op
= XVECEXP (op
, 0, i
);
31229 rtx elt_ideal
= XVECEXP (ideal
, 0, i
);
31231 if (!CONST_INT_P (elt_op
)
31232 || INTVAL (elt_ideal
) != INTVAL (elt_op
))
31238 /* Can output mi_thunk for all cases except for non-zero vcall_offset
31241 arm_can_output_mi_thunk (const_tree
, HOST_WIDE_INT
, HOST_WIDE_INT vcall_offset
,
31244 /* For now, we punt and not handle this for TARGET_THUMB1. */
31245 if (vcall_offset
&& TARGET_THUMB1
)
31248 /* Otherwise ok. */
31252 /* Generate RTL for a conditional branch with rtx comparison CODE in
31253 mode CC_MODE. The destination of the unlikely conditional branch
31257 arm_gen_unlikely_cbranch (enum rtx_code code
, machine_mode cc_mode
,
31261 x
= gen_rtx_fmt_ee (code
, VOIDmode
,
31262 gen_rtx_REG (cc_mode
, CC_REGNUM
),
31265 x
= gen_rtx_IF_THEN_ELSE (VOIDmode
, x
,
31266 gen_rtx_LABEL_REF (VOIDmode
, label_ref
),
31268 emit_unlikely_jump (gen_rtx_SET (pc_rtx
, x
));
31271 /* Implement the TARGET_ASM_ELF_FLAGS_NUMERIC hook.
31273 For pure-code sections there is no letter code for this attribute, so
31274 output all the section flags numerically when this is needed. */
31277 arm_asm_elf_flags_numeric (unsigned int flags
, unsigned int *num
)
31280 if (flags
& SECTION_ARM_PURECODE
)
31284 if (!(flags
& SECTION_DEBUG
))
31286 if (flags
& SECTION_EXCLUDE
)
31287 *num
|= 0x80000000;
31288 if (flags
& SECTION_WRITE
)
31290 if (flags
& SECTION_CODE
)
31292 if (flags
& SECTION_MERGE
)
31294 if (flags
& SECTION_STRINGS
)
31296 if (flags
& SECTION_TLS
)
31298 if (HAVE_COMDAT_GROUP
&& (flags
& SECTION_LINKONCE
))
31307 /* Implement the TARGET_ASM_FUNCTION_SECTION hook.
31309 If pure-code is passed as an option, make sure all functions are in
31310 sections that have the SHF_ARM_PURECODE attribute. */
31313 arm_function_section (tree decl
, enum node_frequency freq
,
31314 bool startup
, bool exit
)
31316 const char * section_name
;
31319 if (!decl
|| TREE_CODE (decl
) != FUNCTION_DECL
)
31320 return default_function_section (decl
, freq
, startup
, exit
);
31322 if (!target_pure_code
)
31323 return default_function_section (decl
, freq
, startup
, exit
);
31326 section_name
= DECL_SECTION_NAME (decl
);
31328 /* If a function is not in a named section then it falls under the 'default'
31329 text section, also known as '.text'. We can preserve previous behavior as
31330 the default text section already has the SHF_ARM_PURECODE section
31334 section
*default_sec
= default_function_section (decl
, freq
, startup
,
31337 /* If default_sec is not null, then it must be a special section like for
31338 example .text.startup. We set the pure-code attribute and return the
31339 same section to preserve existing behavior. */
31341 default_sec
->common
.flags
|= SECTION_ARM_PURECODE
;
31342 return default_sec
;
31345 /* Otherwise look whether a section has already been created with
31347 sec
= get_named_section (decl
, section_name
, 0);
31349 /* If that is not the case passing NULL as the section's name to
31350 'get_named_section' will create a section with the declaration's
31352 sec
= get_named_section (decl
, NULL
, 0);
31354 /* Set the SHF_ARM_PURECODE attribute. */
31355 sec
->common
.flags
|= SECTION_ARM_PURECODE
;
31360 /* Implements the TARGET_SECTION_FLAGS hook.
31362 If DECL is a function declaration and pure-code is passed as an option
31363 then add the SFH_ARM_PURECODE attribute to the section flags. NAME is the
31364 section's name and RELOC indicates whether the declarations initializer may
31365 contain runtime relocations. */
31367 static unsigned int
31368 arm_elf_section_type_flags (tree decl
, const char *name
, int reloc
)
31370 unsigned int flags
= default_section_type_flags (decl
, name
, reloc
);
31372 if (decl
&& TREE_CODE (decl
) == FUNCTION_DECL
&& target_pure_code
)
31373 flags
|= SECTION_ARM_PURECODE
;
31378 /* Generate call to __aeabi_[mode]divmod (op0, op1). */
31381 arm_expand_divmod_libfunc (rtx libfunc
, machine_mode mode
,
31383 rtx
*quot_p
, rtx
*rem_p
)
31385 if (mode
== SImode
)
31386 gcc_assert (!TARGET_IDIV
);
31388 scalar_int_mode libval_mode
31389 = smallest_int_mode_for_size (2 * GET_MODE_BITSIZE (mode
));
31391 rtx libval
= emit_library_call_value (libfunc
, NULL_RTX
, LCT_CONST
,
31393 op0
, GET_MODE (op0
),
31394 op1
, GET_MODE (op1
));
31396 rtx quotient
= simplify_gen_subreg (mode
, libval
, libval_mode
, 0);
31397 rtx remainder
= simplify_gen_subreg (mode
, libval
, libval_mode
,
31398 GET_MODE_SIZE (mode
));
31400 gcc_assert (quotient
);
31401 gcc_assert (remainder
);
31403 *quot_p
= quotient
;
31404 *rem_p
= remainder
;
31407 /* This function checks for the availability of the coprocessor builtin passed
31408 in BUILTIN for the current target. Returns true if it is available and
31409 false otherwise. If a BUILTIN is passed for which this function has not
31410 been implemented it will cause an exception. */
31413 arm_coproc_builtin_available (enum unspecv builtin
)
31415 /* None of these builtins are available in Thumb mode if the target only
31416 supports Thumb-1. */
31434 case VUNSPEC_LDC2L
:
31436 case VUNSPEC_STC2L
:
31439 /* Only present in ARMv5*, ARMv6 (but not ARMv6-M), ARMv7* and
31446 /* Only present in ARMv5TE, ARMv6 (but not ARMv6-M), ARMv7* and
31448 if (arm_arch6
|| arm_arch5te
)
31451 case VUNSPEC_MCRR2
:
31452 case VUNSPEC_MRRC2
:
31457 gcc_unreachable ();
31462 /* This function returns true if OP is a valid memory operand for the ldc and
31463 stc coprocessor instructions and false otherwise. */
31466 arm_coproc_ldc_stc_legitimate_address (rtx op
)
31468 HOST_WIDE_INT range
;
31469 /* Has to be a memory operand. */
31475 /* We accept registers. */
31479 switch GET_CODE (op
)
31483 /* Or registers with an offset. */
31484 if (!REG_P (XEXP (op
, 0)))
31489 /* The offset must be an immediate though. */
31490 if (!CONST_INT_P (op
))
31493 range
= INTVAL (op
);
31495 /* Within the range of [-1020,1020]. */
31496 if (!IN_RANGE (range
, -1020, 1020))
31499 /* And a multiple of 4. */
31500 return (range
% 4) == 0;
31506 return REG_P (XEXP (op
, 0));
31508 gcc_unreachable ();
31513 /* Implement TARGET_CAN_CHANGE_MODE_CLASS.
31515 In VFPv1, VFP registers could only be accessed in the mode they were
31516 set, so subregs would be invalid there. However, we don't support
31517 VFPv1 at the moment, and the restriction was lifted in VFPv2.
31519 In big-endian mode, modes greater than word size (i.e. DFmode) are stored in
31520 VFP registers in little-endian order. We can't describe that accurately to
31521 GCC, so avoid taking subregs of such values.
31523 The only exception is going from a 128-bit to a 64-bit type. In that
31524 case the data layout happens to be consistent for big-endian, so we
31525 explicitly allow that case. */
31528 arm_can_change_mode_class (machine_mode from
, machine_mode to
,
31529 reg_class_t rclass
)
31532 && !(GET_MODE_SIZE (from
) == 16 && GET_MODE_SIZE (to
) == 8)
31533 && (GET_MODE_SIZE (from
) > UNITS_PER_WORD
31534 || GET_MODE_SIZE (to
) > UNITS_PER_WORD
)
31535 && reg_classes_intersect_p (VFP_REGS
, rclass
))
31540 /* Implement TARGET_CONSTANT_ALIGNMENT. Make strings word-aligned so
31541 strcpy from constants will be faster. */
31543 static HOST_WIDE_INT
31544 arm_constant_alignment (const_tree exp
, HOST_WIDE_INT align
)
31546 unsigned int factor
= (TARGET_THUMB
|| ! arm_tune_xscale
? 1 : 2);
31547 if (TREE_CODE (exp
) == STRING_CST
&& !optimize_size
)
31548 return MAX (align
, BITS_PER_WORD
* factor
);
31552 /* Emit a speculation barrier on target architectures that do not have
31553 DSB/ISB directly. Such systems probably don't need a barrier
31554 themselves, but if the code is ever run on a later architecture, it
31555 might become a problem. */
31557 arm_emit_speculation_barrier_function ()
31559 emit_library_call (speculation_barrier_libfunc
, LCT_NORMAL
, VOIDmode
);
31563 namespace selftest
{
31565 /* Scan the static data tables generated by parsecpu.awk looking for
31566 potential issues with the data. We primarily check for
31567 inconsistencies in the option extensions at present (extensions
31568 that duplicate others but aren't marked as aliases). Furthermore,
31569 for correct canonicalization later options must never be a subset
31570 of an earlier option. Any extension should also only specify other
31571 feature bits and never an architecture bit. The architecture is inferred
31572 from the declaration of the extension. */
31574 arm_test_cpu_arch_data (void)
31576 const arch_option
*arch
;
31577 const cpu_option
*cpu
;
31578 auto_sbitmap
target_isa (isa_num_bits
);
31579 auto_sbitmap
isa1 (isa_num_bits
);
31580 auto_sbitmap
isa2 (isa_num_bits
);
31582 for (arch
= all_architectures
; arch
->common
.name
!= NULL
; ++arch
)
31584 const cpu_arch_extension
*ext1
, *ext2
;
31586 if (arch
->common
.extensions
== NULL
)
31589 arm_initialize_isa (target_isa
, arch
->common
.isa_bits
);
31591 for (ext1
= arch
->common
.extensions
; ext1
->name
!= NULL
; ++ext1
)
31596 arm_initialize_isa (isa1
, ext1
->isa_bits
);
31597 for (ext2
= ext1
+ 1; ext2
->name
!= NULL
; ++ext2
)
31599 if (ext2
->alias
|| ext1
->remove
!= ext2
->remove
)
31602 arm_initialize_isa (isa2
, ext2
->isa_bits
);
31603 /* If the option is a subset of the parent option, it doesn't
31604 add anything and so isn't useful. */
31605 ASSERT_TRUE (!bitmap_subset_p (isa2
, isa1
));
31607 /* If the extension specifies any architectural bits then
31608 disallow it. Extensions should only specify feature bits. */
31609 ASSERT_TRUE (!bitmap_intersect_p (isa2
, target_isa
));
31614 for (cpu
= all_cores
; cpu
->common
.name
!= NULL
; ++cpu
)
31616 const cpu_arch_extension
*ext1
, *ext2
;
31618 if (cpu
->common
.extensions
== NULL
)
31621 arm_initialize_isa (target_isa
, arch
->common
.isa_bits
);
31623 for (ext1
= cpu
->common
.extensions
; ext1
->name
!= NULL
; ++ext1
)
31628 arm_initialize_isa (isa1
, ext1
->isa_bits
);
31629 for (ext2
= ext1
+ 1; ext2
->name
!= NULL
; ++ext2
)
31631 if (ext2
->alias
|| ext1
->remove
!= ext2
->remove
)
31634 arm_initialize_isa (isa2
, ext2
->isa_bits
);
31635 /* If the option is a subset of the parent option, it doesn't
31636 add anything and so isn't useful. */
31637 ASSERT_TRUE (!bitmap_subset_p (isa2
, isa1
));
31639 /* If the extension specifies any architectural bits then
31640 disallow it. Extensions should only specify feature bits. */
31641 ASSERT_TRUE (!bitmap_intersect_p (isa2
, target_isa
));
31647 /* Scan the static data tables generated by parsecpu.awk looking for
31648 potential issues with the data. Here we check for consistency between the
31649 fpu bits, in particular we check that ISA_ALL_FPU_INTERNAL does not contain
31650 a feature bit that is not defined by any FPU flag. */
31652 arm_test_fpu_data (void)
31654 auto_sbitmap
isa_all_fpubits (isa_num_bits
);
31655 auto_sbitmap
fpubits (isa_num_bits
);
31656 auto_sbitmap
tmpset (isa_num_bits
);
31658 static const enum isa_feature fpu_bitlist
[]
31659 = { ISA_ALL_FPU_INTERNAL
, isa_nobit
};
31660 arm_initialize_isa (isa_all_fpubits
, fpu_bitlist
);
31662 for (unsigned int i
= 0; i
< TARGET_FPU_auto
; i
++)
31664 arm_initialize_isa (fpubits
, all_fpus
[i
].isa_bits
);
31665 bitmap_and_compl (tmpset
, isa_all_fpubits
, fpubits
);
31666 bitmap_clear (isa_all_fpubits
);
31667 bitmap_copy (isa_all_fpubits
, tmpset
);
31670 if (!bitmap_empty_p (isa_all_fpubits
))
31672 fprintf (stderr
, "Error: found feature bits in the ALL_FPU_INTERAL"
31673 " group that are not defined by any FPU.\n"
31674 " Check your arm-cpus.in.\n");
31675 ASSERT_TRUE (bitmap_empty_p (isa_all_fpubits
));
31680 arm_run_selftests (void)
31682 arm_test_cpu_arch_data ();
31683 arm_test_fpu_data ();
31685 } /* Namespace selftest. */
31687 #undef TARGET_RUN_TARGET_SELFTESTS
31688 #define TARGET_RUN_TARGET_SELFTESTS selftest::arm_run_selftests
31689 #endif /* CHECKING_P */
31691 struct gcc_target targetm
= TARGET_INITIALIZER
;
31693 #include "gt-arm.h"