]>
Commit | Line | Data |
---|---|---|
129a2fe4 | 1 | /* Output routines for GCC for ARM. |
a941568e | 2 | Copyright (C) 1991, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001, |
25f905c2 | 3 | 2002, 2003, 2004, 2005, 2006, 2007 Free Software Foundation, Inc. |
8cd42699 | 4 | Contributed by Pieter `Tiggr' Schoenmakers (rcpieter@win.tue.nl) |
ad87de1e | 5 | and Martin Simmons (@harleqn.co.uk). |
129a2fe4 | 6 | More major hacks by Richard Earnshaw (rearnsha@arm.com). |
8cd42699 | 7 | |
acf6ed70 | 8 | This file is part of GCC. |
8cd42699 | 9 | |
acf6ed70 | 10 | GCC is free software; you can redistribute it and/or modify it |
11 | under the terms of the GNU General Public License as published | |
12 | by the Free Software Foundation; either version 2, or (at your | |
13 | option) any later version. | |
8cd42699 | 14 | |
acf6ed70 | 15 | GCC is distributed in the hope that it will be useful, but WITHOUT |
16 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY | |
17 | or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public | |
18 | License for more details. | |
8cd42699 | 19 | |
acf6ed70 | 20 | You should have received a copy of the GNU General Public License |
21 | along with GCC; see the file COPYING. If not, write to | |
dbddc6c4 | 22 | the Free Software Foundation, 51 Franklin Street, Fifth Floor, |
23 | Boston, MA 02110-1301, USA. */ | |
9e7454d0 | 24 | |
6019008a | 25 | #include "config.h" |
76676c8e | 26 | #include "system.h" |
805e22b2 | 27 | #include "coretypes.h" |
28 | #include "tm.h" | |
8cd42699 | 29 | #include "rtl.h" |
cffb2a26 | 30 | #include "tree.h" |
a220ccce | 31 | #include "obstack.h" |
8cd42699 | 32 | #include "regs.h" |
33 | #include "hard-reg-set.h" | |
34 | #include "real.h" | |
35 | #include "insn-config.h" | |
36 | #include "conditions.h" | |
8cd42699 | 37 | #include "output.h" |
38 | #include "insn-attr.h" | |
39 | #include "flags.h" | |
9c1b36a9 | 40 | #include "reload.h" |
0a893c29 | 41 | #include "function.h" |
411b0ddb | 42 | #include "expr.h" |
d8fc4d0b | 43 | #include "optabs.h" |
ecd3fcc7 | 44 | #include "toplev.h" |
82ec49ae | 45 | #include "recog.h" |
db80f614 | 46 | #include "ggc.h" |
cffb2a26 | 47 | #include "except.h" |
1fcd08b1 | 48 | #include "c-pragma.h" |
2d9065d3 | 49 | #include "integrate.h" |
78fe751b | 50 | #include "tm_p.h" |
a767736d | 51 | #include "target.h" |
52 | #include "target-def.h" | |
37826745 | 53 | #include "debug.h" |
e60d3615 | 54 | #include "langhooks.h" |
3072d30e | 55 | #include "df.h" |
8cd42699 | 56 | |
cffb2a26 | 57 | /* Forward definitions of types. */ |
58 | typedef struct minipool_node Mnode; | |
59 | typedef struct minipool_fixup Mfix; | |
60 | ||
a12be13c | 61 | const struct attribute_spec arm_attribute_table[]; |
62 | ||
cffb2a26 | 63 | /* Forward function declarations. */ |
f9273c43 | 64 | static arm_stack_offsets *arm_get_frame_offsets (void); |
ebd88a36 | 65 | static void arm_add_gc_roots (void); |
96f57e36 | 66 | static int arm_gen_constant (enum rtx_code, enum machine_mode, rtx, |
67 | HOST_WIDE_INT, rtx, rtx, int, int); | |
ebd88a36 | 68 | static unsigned bit_count (unsigned long); |
69 | static int arm_address_register_rtx_p (rtx, int); | |
b4e8a300 | 70 | static int arm_legitimate_index_p (enum machine_mode, rtx, RTX_CODE, int); |
25f905c2 | 71 | static int thumb2_legitimate_index_p (enum machine_mode, rtx, int); |
72 | static int thumb1_base_register_rtx_p (rtx, enum machine_mode, int); | |
73 | inline static int thumb1_index_register_rtx_p (rtx, int); | |
f9273c43 | 74 | static int thumb_far_jump_used_p (void); |
237533cc | 75 | static bool thumb_force_lr_save (void); |
25f905c2 | 76 | static unsigned long thumb1_compute_save_reg_mask (void); |
ebd88a36 | 77 | static int const_ok_for_op (HOST_WIDE_INT, enum rtx_code); |
ebd88a36 | 78 | static rtx emit_sfm (int, int); |
0858d94e | 79 | static int arm_size_return_regs (void); |
58356836 | 80 | #ifndef AOF_ASSEMBLER |
ebd88a36 | 81 | static bool arm_assemble_integer (rtx, unsigned int, int); |
58356836 | 82 | #endif |
ebd88a36 | 83 | static const char *fp_const_from_val (REAL_VALUE_TYPE *); |
84 | static arm_cc get_arm_condition_code (rtx); | |
ebd88a36 | 85 | static HOST_WIDE_INT int_log2 (HOST_WIDE_INT); |
86 | static rtx is_jump_table (rtx); | |
87 | static const char *output_multi_immediate (rtx *, const char *, const char *, | |
88 | int, HOST_WIDE_INT); | |
ebd88a36 | 89 | static const char *shift_op (rtx, HOST_WIDE_INT *); |
90 | static struct machine_function *arm_init_machine_status (void); | |
4c44712e | 91 | static void thumb_exit (FILE *, int); |
ebd88a36 | 92 | static rtx is_jump_table (rtx); |
93 | static HOST_WIDE_INT get_jump_table_size (rtx); | |
94 | static Mnode *move_minipool_fix_forward_ref (Mnode *, Mnode *, HOST_WIDE_INT); | |
95 | static Mnode *add_minipool_forward_ref (Mfix *); | |
96 | static Mnode *move_minipool_fix_backward_ref (Mnode *, Mnode *, HOST_WIDE_INT); | |
97 | static Mnode *add_minipool_backward_ref (Mfix *); | |
98 | static void assign_minipool_offsets (Mfix *); | |
99 | static void arm_print_value (FILE *, rtx); | |
100 | static void dump_minipool (rtx); | |
101 | static int arm_barrier_cost (rtx); | |
102 | static Mfix *create_fix_barrier (Mfix *, HOST_WIDE_INT); | |
103 | static void push_minipool_barrier (rtx, HOST_WIDE_INT); | |
104 | static void push_minipool_fix (rtx, HOST_WIDE_INT, rtx *, enum machine_mode, | |
105 | rtx); | |
106 | static void arm_reorg (void); | |
107 | static bool note_invalid_constants (rtx, HOST_WIDE_INT, int); | |
ebd88a36 | 108 | static unsigned long arm_compute_save_reg0_reg12_mask (void); |
109 | static unsigned long arm_compute_save_reg_mask (void); | |
110 | static unsigned long arm_isr_value (tree); | |
111 | static unsigned long arm_compute_func_type (void); | |
112 | static tree arm_handle_fndecl_attribute (tree *, tree, tree, int, bool *); | |
113 | static tree arm_handle_isr_attribute (tree *, tree, tree, int, bool *); | |
56826b58 | 114 | #if TARGET_DLLIMPORT_DECL_ATTRIBUTES |
a79b1d93 | 115 | static tree arm_handle_notshared_attribute (tree *, tree, tree, int, bool *); |
56826b58 | 116 | #endif |
ebd88a36 | 117 | static void arm_output_function_epilogue (FILE *, HOST_WIDE_INT); |
118 | static void arm_output_function_prologue (FILE *, HOST_WIDE_INT); | |
25f905c2 | 119 | static void thumb1_output_function_prologue (FILE *, HOST_WIDE_INT); |
ebd88a36 | 120 | static int arm_comp_type_attributes (tree, tree); |
121 | static void arm_set_default_type_attributes (tree); | |
122 | static int arm_adjust_cost (rtx, rtx, rtx, int); | |
ebd88a36 | 123 | static int count_insns_for_constant (HOST_WIDE_INT, int); |
124 | static int arm_get_strip_length (int); | |
125 | static bool arm_function_ok_for_sibcall (tree, tree); | |
126 | static void arm_internal_label (FILE *, const char *, unsigned long); | |
127 | static void arm_output_mi_thunk (FILE *, tree, HOST_WIDE_INT, HOST_WIDE_INT, | |
128 | tree); | |
129 | static int arm_rtx_costs_1 (rtx, enum rtx_code, enum rtx_code); | |
c1877da8 | 130 | static bool arm_size_rtx_costs (rtx, int, int, int *); |
a2cd141b | 131 | static bool arm_slowmul_rtx_costs (rtx, int, int, int *); |
132 | static bool arm_fastmul_rtx_costs (rtx, int, int, int *); | |
133 | static bool arm_xscale_rtx_costs (rtx, int, int, int *); | |
134 | static bool arm_9e_rtx_costs (rtx, int, int, int *); | |
ebd88a36 | 135 | static int arm_address_cost (rtx); |
136 | static bool arm_memory_load_p (rtx); | |
137 | static bool arm_cirrus_insn_p (rtx); | |
138 | static void cirrus_reorg (rtx); | |
755eb2b4 | 139 | static void arm_init_builtins (void); |
140 | static rtx arm_expand_builtin (tree, rtx, rtx, enum machine_mode, int); | |
141 | static void arm_init_iwmmxt_builtins (void); | |
142 | static rtx safe_vector_operand (rtx, enum machine_mode); | |
143 | static rtx arm_expand_binop_builtin (enum insn_code, tree, rtx); | |
144 | static rtx arm_expand_unop_builtin (enum insn_code, tree, rtx, int); | |
145 | static rtx arm_expand_builtin (tree, rtx, rtx, enum machine_mode, int); | |
96f57e36 | 146 | static void emit_constant_insn (rtx cond, rtx pattern); |
0083a3b9 | 147 | static rtx emit_set_insn (rtx, rtx); |
f054eb3c | 148 | static int arm_arg_partial_bytes (CUMULATIVE_ARGS *, enum machine_mode, |
149 | tree, bool); | |
755eb2b4 | 150 | |
d24bc145 | 151 | #ifdef OBJECT_FORMAT_ELF |
bac7fc85 | 152 | static void arm_elf_asm_constructor (rtx, int) ATTRIBUTE_UNUSED; |
153 | static void arm_elf_asm_destructor (rtx, int) ATTRIBUTE_UNUSED; | |
d24bc145 | 154 | #endif |
7811991d | 155 | #ifndef ARM_PE |
ebd88a36 | 156 | static void arm_encode_section_info (tree, rtx, int); |
7811991d | 157 | #endif |
afe27f3b | 158 | |
159 | static void arm_file_end (void); | |
5fad5123 | 160 | static void arm_file_start (void); |
afe27f3b | 161 | |
67c1e638 | 162 | #ifdef AOF_ASSEMBLER |
ebd88a36 | 163 | static void aof_globalize_label (FILE *, const char *); |
164 | static void aof_dump_imports (FILE *); | |
165 | static void aof_dump_pic_table (FILE *); | |
92c473b8 | 166 | static void aof_file_start (void); |
ebd88a36 | 167 | static void aof_file_end (void); |
2f14b1f9 | 168 | static void aof_asm_init_sections (void); |
67c1e638 | 169 | #endif |
4c66acf1 | 170 | static void arm_setup_incoming_varargs (CUMULATIVE_ARGS *, enum machine_mode, |
171 | tree, int *, int); | |
b981d932 | 172 | static bool arm_pass_by_reference (CUMULATIVE_ARGS *, |
173 | enum machine_mode, tree, bool); | |
72f1246c | 174 | static bool arm_promote_prototypes (tree); |
7faec15f | 175 | static bool arm_default_short_enums (void); |
25ba5be6 | 176 | static bool arm_align_anon_bitfield (void); |
46b5d878 | 177 | static bool arm_return_in_msb (tree); |
178 | static bool arm_must_pass_in_stack (enum machine_mode, tree); | |
1774763d | 179 | #ifdef TARGET_UNWIND_INFO |
180 | static void arm_unwind_emit (FILE *, rtx); | |
181 | static bool arm_output_ttype (rtx); | |
182 | #endif | |
25f905c2 | 183 | static void arm_dwarf_handle_frame_unspec (const char *, rtx, int); |
747af5e7 | 184 | |
c1dc02de | 185 | static tree arm_cxx_guard_type (void); |
186 | static bool arm_cxx_guard_mask_bit (void); | |
600f4be7 | 187 | static tree arm_get_cookie_size (tree); |
188 | static bool arm_cookie_has_size (void); | |
853b7640 | 189 | static bool arm_cxx_cdtor_returns_this (void); |
4fc9a8ec | 190 | static bool arm_cxx_key_method_may_be_inline (void); |
7908506d | 191 | static void arm_cxx_determine_class_data_visibility (tree); |
192 | static bool arm_cxx_class_data_always_comdat (void); | |
46b9ff5e | 193 | static bool arm_cxx_use_aeabi_atexit (void); |
30e9913f | 194 | static void arm_init_libfuncs (void); |
81f6de2f | 195 | static bool arm_handle_option (size_t, const char *, int); |
154ee3b8 | 196 | static void arm_target_help (void); |
c49547c4 | 197 | static unsigned HOST_WIDE_INT arm_shift_truncation_mask (enum machine_mode); |
f655717d | 198 | static bool arm_cannot_copy_insn_p (rtx); |
199 | static bool arm_tls_symbol_p (rtx x); | |
4471a482 | 200 | static void arm_output_dwarf_dtprel (FILE *, int, rtx) ATTRIBUTE_UNUSED; |
f655717d | 201 | |
a767736d | 202 | \f |
203 | /* Initialize the GCC target structure. */ | |
3aa0c315 | 204 | #if TARGET_DLLIMPORT_DECL_ATTRIBUTES |
a12be13c | 205 | #undef TARGET_MERGE_DECL_ATTRIBUTES |
a767736d | 206 | #define TARGET_MERGE_DECL_ATTRIBUTES merge_dllimport_decl_attributes |
207 | #endif | |
a690ede2 | 208 | |
a12be13c | 209 | #undef TARGET_ATTRIBUTE_TABLE |
e3c541f0 | 210 | #define TARGET_ATTRIBUTE_TABLE arm_attribute_table |
a767736d | 211 | |
5fad5123 | 212 | #undef TARGET_ASM_FILE_START |
213 | #define TARGET_ASM_FILE_START arm_file_start | |
afe27f3b | 214 | #undef TARGET_ASM_FILE_END |
215 | #define TARGET_ASM_FILE_END arm_file_end | |
216 | ||
58356836 | 217 | #ifdef AOF_ASSEMBLER |
a12be13c | 218 | #undef TARGET_ASM_BYTE_OP |
58356836 | 219 | #define TARGET_ASM_BYTE_OP "\tDCB\t" |
a12be13c | 220 | #undef TARGET_ASM_ALIGNED_HI_OP |
58356836 | 221 | #define TARGET_ASM_ALIGNED_HI_OP "\tDCW\t" |
a12be13c | 222 | #undef TARGET_ASM_ALIGNED_SI_OP |
58356836 | 223 | #define TARGET_ASM_ALIGNED_SI_OP "\tDCD\t" |
67c1e638 | 224 | #undef TARGET_ASM_GLOBALIZE_LABEL |
225 | #define TARGET_ASM_GLOBALIZE_LABEL aof_globalize_label | |
92c473b8 | 226 | #undef TARGET_ASM_FILE_START |
227 | #define TARGET_ASM_FILE_START aof_file_start | |
f6940372 | 228 | #undef TARGET_ASM_FILE_END |
229 | #define TARGET_ASM_FILE_END aof_file_end | |
58356836 | 230 | #else |
a12be13c | 231 | #undef TARGET_ASM_ALIGNED_SI_OP |
58356836 | 232 | #define TARGET_ASM_ALIGNED_SI_OP NULL |
a12be13c | 233 | #undef TARGET_ASM_INTEGER |
58356836 | 234 | #define TARGET_ASM_INTEGER arm_assemble_integer |
235 | #endif | |
236 | ||
a12be13c | 237 | #undef TARGET_ASM_FUNCTION_PROLOGUE |
17d9b0c3 | 238 | #define TARGET_ASM_FUNCTION_PROLOGUE arm_output_function_prologue |
239 | ||
a12be13c | 240 | #undef TARGET_ASM_FUNCTION_EPILOGUE |
17d9b0c3 | 241 | #define TARGET_ASM_FUNCTION_EPILOGUE arm_output_function_epilogue |
242 | ||
81f6de2f | 243 | #undef TARGET_DEFAULT_TARGET_FLAGS |
244 | #define TARGET_DEFAULT_TARGET_FLAGS (TARGET_DEFAULT | MASK_SCHED_PROLOG) | |
245 | #undef TARGET_HANDLE_OPTION | |
246 | #define TARGET_HANDLE_OPTION arm_handle_option | |
154ee3b8 | 247 | #undef TARGET_HELP |
248 | #define TARGET_HELP arm_target_help | |
81f6de2f | 249 | |
a12be13c | 250 | #undef TARGET_COMP_TYPE_ATTRIBUTES |
f32321da | 251 | #define TARGET_COMP_TYPE_ATTRIBUTES arm_comp_type_attributes |
252 | ||
a12be13c | 253 | #undef TARGET_SET_DEFAULT_TYPE_ATTRIBUTES |
f32321da | 254 | #define TARGET_SET_DEFAULT_TYPE_ATTRIBUTES arm_set_default_type_attributes |
255 | ||
a12be13c | 256 | #undef TARGET_SCHED_ADJUST_COST |
747af5e7 | 257 | #define TARGET_SCHED_ADJUST_COST arm_adjust_cost |
258 | ||
7811991d | 259 | #undef TARGET_ENCODE_SECTION_INFO |
260 | #ifdef ARM_PE | |
261 | #define TARGET_ENCODE_SECTION_INFO arm_pe_encode_section_info | |
262 | #else | |
263 | #define TARGET_ENCODE_SECTION_INFO arm_encode_section_info | |
264 | #endif | |
265 | ||
755eb2b4 | 266 | #undef TARGET_STRIP_NAME_ENCODING |
7b4a38a6 | 267 | #define TARGET_STRIP_NAME_ENCODING arm_strip_name_encoding |
268 | ||
755eb2b4 | 269 | #undef TARGET_ASM_INTERNAL_LABEL |
805e22b2 | 270 | #define TARGET_ASM_INTERNAL_LABEL arm_internal_label |
271 | ||
755eb2b4 | 272 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL |
805e22b2 | 273 | #define TARGET_FUNCTION_OK_FOR_SIBCALL arm_function_ok_for_sibcall |
274 | ||
755eb2b4 | 275 | #undef TARGET_ASM_OUTPUT_MI_THUNK |
6988553d | 276 | #define TARGET_ASM_OUTPUT_MI_THUNK arm_output_mi_thunk |
755eb2b4 | 277 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK |
eb344f43 | 278 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK default_can_output_mi_thunk_no_vcall |
6988553d | 279 | |
a2cd141b | 280 | /* This will be overridden in arm_override_options. */ |
755eb2b4 | 281 | #undef TARGET_RTX_COSTS |
a2cd141b | 282 | #define TARGET_RTX_COSTS arm_slowmul_rtx_costs |
755eb2b4 | 283 | #undef TARGET_ADDRESS_COST |
ec0457a8 | 284 | #define TARGET_ADDRESS_COST arm_address_cost |
fab7adbf | 285 | |
c49547c4 | 286 | #undef TARGET_SHIFT_TRUNCATION_MASK |
287 | #define TARGET_SHIFT_TRUNCATION_MASK arm_shift_truncation_mask | |
9e7454d0 | 288 | #undef TARGET_VECTOR_MODE_SUPPORTED_P |
289 | #define TARGET_VECTOR_MODE_SUPPORTED_P arm_vector_mode_supported_p | |
290 | ||
755eb2b4 | 291 | #undef TARGET_MACHINE_DEPENDENT_REORG |
2efea8c0 | 292 | #define TARGET_MACHINE_DEPENDENT_REORG arm_reorg |
293 | ||
755eb2b4 | 294 | #undef TARGET_INIT_BUILTINS |
295 | #define TARGET_INIT_BUILTINS arm_init_builtins | |
296 | #undef TARGET_EXPAND_BUILTIN | |
297 | #define TARGET_EXPAND_BUILTIN arm_expand_builtin | |
298 | ||
30e9913f | 299 | #undef TARGET_INIT_LIBFUNCS |
300 | #define TARGET_INIT_LIBFUNCS arm_init_libfuncs | |
301 | ||
f2eb2b4d | 302 | #undef TARGET_PROMOTE_FUNCTION_ARGS |
303 | #define TARGET_PROMOTE_FUNCTION_ARGS hook_bool_tree_true | |
19347327 | 304 | #undef TARGET_PROMOTE_FUNCTION_RETURN |
305 | #define TARGET_PROMOTE_FUNCTION_RETURN hook_bool_tree_true | |
f2eb2b4d | 306 | #undef TARGET_PROMOTE_PROTOTYPES |
72f1246c | 307 | #define TARGET_PROMOTE_PROTOTYPES arm_promote_prototypes |
b981d932 | 308 | #undef TARGET_PASS_BY_REFERENCE |
309 | #define TARGET_PASS_BY_REFERENCE arm_pass_by_reference | |
f054eb3c | 310 | #undef TARGET_ARG_PARTIAL_BYTES |
311 | #define TARGET_ARG_PARTIAL_BYTES arm_arg_partial_bytes | |
f2eb2b4d | 312 | |
4c66acf1 | 313 | #undef TARGET_SETUP_INCOMING_VARARGS |
314 | #define TARGET_SETUP_INCOMING_VARARGS arm_setup_incoming_varargs | |
315 | ||
7faec15f | 316 | #undef TARGET_DEFAULT_SHORT_ENUMS |
317 | #define TARGET_DEFAULT_SHORT_ENUMS arm_default_short_enums | |
318 | ||
25ba5be6 | 319 | #undef TARGET_ALIGN_ANON_BITFIELD |
320 | #define TARGET_ALIGN_ANON_BITFIELD arm_align_anon_bitfield | |
321 | ||
5f458503 | 322 | #undef TARGET_NARROW_VOLATILE_BITFIELD |
323 | #define TARGET_NARROW_VOLATILE_BITFIELD hook_bool_void_false | |
324 | ||
c1dc02de | 325 | #undef TARGET_CXX_GUARD_TYPE |
326 | #define TARGET_CXX_GUARD_TYPE arm_cxx_guard_type | |
327 | ||
328 | #undef TARGET_CXX_GUARD_MASK_BIT | |
329 | #define TARGET_CXX_GUARD_MASK_BIT arm_cxx_guard_mask_bit | |
330 | ||
600f4be7 | 331 | #undef TARGET_CXX_GET_COOKIE_SIZE |
332 | #define TARGET_CXX_GET_COOKIE_SIZE arm_get_cookie_size | |
333 | ||
334 | #undef TARGET_CXX_COOKIE_HAS_SIZE | |
335 | #define TARGET_CXX_COOKIE_HAS_SIZE arm_cookie_has_size | |
336 | ||
853b7640 | 337 | #undef TARGET_CXX_CDTOR_RETURNS_THIS |
338 | #define TARGET_CXX_CDTOR_RETURNS_THIS arm_cxx_cdtor_returns_this | |
339 | ||
4fc9a8ec | 340 | #undef TARGET_CXX_KEY_METHOD_MAY_BE_INLINE |
341 | #define TARGET_CXX_KEY_METHOD_MAY_BE_INLINE arm_cxx_key_method_may_be_inline | |
342 | ||
46b9ff5e | 343 | #undef TARGET_CXX_USE_AEABI_ATEXIT |
344 | #define TARGET_CXX_USE_AEABI_ATEXIT arm_cxx_use_aeabi_atexit | |
345 | ||
7908506d | 346 | #undef TARGET_CXX_DETERMINE_CLASS_DATA_VISIBILITY |
347 | #define TARGET_CXX_DETERMINE_CLASS_DATA_VISIBILITY \ | |
348 | arm_cxx_determine_class_data_visibility | |
349 | ||
350 | #undef TARGET_CXX_CLASS_DATA_ALWAYS_COMDAT | |
351 | #define TARGET_CXX_CLASS_DATA_ALWAYS_COMDAT arm_cxx_class_data_always_comdat | |
4fc9a8ec | 352 | |
46b5d878 | 353 | #undef TARGET_RETURN_IN_MSB |
354 | #define TARGET_RETURN_IN_MSB arm_return_in_msb | |
355 | ||
356 | #undef TARGET_MUST_PASS_IN_STACK | |
357 | #define TARGET_MUST_PASS_IN_STACK arm_must_pass_in_stack | |
358 | ||
1774763d | 359 | #ifdef TARGET_UNWIND_INFO |
360 | #undef TARGET_UNWIND_EMIT | |
361 | #define TARGET_UNWIND_EMIT arm_unwind_emit | |
362 | ||
363 | /* EABI unwinding tables use a different format for the typeinfo tables. */ | |
364 | #undef TARGET_ASM_TTYPE | |
365 | #define TARGET_ASM_TTYPE arm_output_ttype | |
366 | ||
367 | #undef TARGET_ARM_EABI_UNWINDER | |
368 | #define TARGET_ARM_EABI_UNWINDER true | |
369 | #endif /* TARGET_UNWIND_INFO */ | |
370 | ||
25f905c2 | 371 | #undef TARGET_DWARF_HANDLE_FRAME_UNSPEC |
372 | #define TARGET_DWARF_HANDLE_FRAME_UNSPEC arm_dwarf_handle_frame_unspec | |
373 | ||
f655717d | 374 | #undef TARGET_CANNOT_COPY_INSN_P |
375 | #define TARGET_CANNOT_COPY_INSN_P arm_cannot_copy_insn_p | |
376 | ||
377 | #ifdef HAVE_AS_TLS | |
378 | #undef TARGET_HAVE_TLS | |
379 | #define TARGET_HAVE_TLS true | |
380 | #endif | |
381 | ||
382 | #undef TARGET_CANNOT_FORCE_CONST_MEM | |
e348ff3e | 383 | #define TARGET_CANNOT_FORCE_CONST_MEM arm_cannot_force_const_mem |
f655717d | 384 | |
eddcdde1 | 385 | #undef TARGET_MANGLE_TYPE |
386 | #define TARGET_MANGLE_TYPE arm_mangle_type | |
387 | ||
4471a482 | 388 | #ifdef HAVE_AS_TLS |
389 | #undef TARGET_ASM_OUTPUT_DWARF_DTPREL | |
390 | #define TARGET_ASM_OUTPUT_DWARF_DTPREL arm_output_dwarf_dtprel | |
391 | #endif | |
392 | ||
57e4bbfb | 393 | struct gcc_target targetm = TARGET_INITIALIZER; |
a767736d | 394 | \f |
a220ccce | 395 | /* Obstack for minipool constant handling. */ |
396 | static struct obstack minipool_obstack; | |
a12be13c | 397 | static char * minipool_startobj; |
a220ccce | 398 | |
a12be13c | 399 | /* The maximum number of insns skipped which |
400 | will be conditionalised if possible. */ | |
78fe751b | 401 | static int max_insns_skipped = 5; |
402 | ||
403 | extern FILE * asm_out_file; | |
404 | ||
66aacf41 | 405 | /* True if we are currently building a constant table. */ |
c9ba03f2 | 406 | int making_const_table; |
407 | ||
2485f877 | 408 | /* Define the information needed to generate branch insns. This is |
66aacf41 | 409 | stored from the compare operation. */ |
9c08d1fa | 410 | rtx arm_compare_op0, arm_compare_op1; |
9c08d1fa | 411 | |
a2cd141b | 412 | /* The processor for which instructions should be scheduled. */ |
413 | enum processor_type arm_tune = arm_none; | |
414 | ||
fa7637bd | 415 | /* The default processor used if not overridden by commandline. */ |
5fad5123 | 416 | static enum processor_type arm_default_cpu = arm_none; |
417 | ||
a2cd141b | 418 | /* Which floating point model to use. */ |
419 | enum arm_fp_model arm_fp_model; | |
411b0ddb | 420 | |
a2cd141b | 421 | /* Which floating point hardware is available. */ |
c7f506fd | 422 | enum fputype arm_fpu_arch; |
f21249d5 | 423 | |
a2cd141b | 424 | /* Which floating point hardware to schedule for. */ |
425 | enum fputype arm_fpu_tune; | |
426 | ||
427 | /* Whether to use floating point hardware. */ | |
428 | enum float_abi_type arm_float_abi; | |
429 | ||
f9273c43 | 430 | /* Which ABI to use. */ |
431 | enum arm_abi_type arm_abi; | |
432 | ||
f655717d | 433 | /* Which thread pointer model to use. */ |
434 | enum arm_tp_type target_thread_pointer = TP_AUTO; | |
435 | ||
6cd47762 | 436 | /* Used to parse -mstructure_size_boundary command line option. */ |
8fa0c55d | 437 | int arm_structure_size_boundary = DEFAULT_STRUCTURE_SIZE_BOUNDARY; |
6cd47762 | 438 | |
afe27f3b | 439 | /* Used for Thumb call_via trampolines. */ |
a52972b8 | 440 | rtx thumb_call_via_label[14]; |
afe27f3b | 441 | static int thumb_call_reg_needed; |
442 | ||
82ec49ae | 443 | /* Bit values used to identify processor capabilities. */ |
bba10fb8 | 444 | #define FL_CO_PROC (1 << 0) /* Has external co-processor bus */ |
a2cd141b | 445 | #define FL_ARCH3M (1 << 1) /* Extended multiply */ |
bba10fb8 | 446 | #define FL_MODE26 (1 << 2) /* 26-bit mode support */ |
447 | #define FL_MODE32 (1 << 3) /* 32-bit mode support */ | |
448 | #define FL_ARCH4 (1 << 4) /* Architecture rel 4 */ | |
449 | #define FL_ARCH5 (1 << 5) /* Architecture rel 5 */ | |
450 | #define FL_THUMB (1 << 6) /* Thumb aware */ | |
451 | #define FL_LDSCHED (1 << 7) /* Load scheduling necessary */ | |
452 | #define FL_STRONG (1 << 8) /* StrongARM */ | |
b89db9a7 | 453 | #define FL_ARCH5E (1 << 9) /* DSP extensions to v5 */ |
331beb1a | 454 | #define FL_XSCALE (1 << 10) /* XScale */ |
7d57ec45 | 455 | #define FL_CIRRUS (1 << 11) /* Cirrus/DSP. */ |
a2cd141b | 456 | #define FL_ARCH6 (1 << 12) /* Architecture rel 6. Adds |
2515f61d | 457 | media instructions. */ |
458 | #define FL_VFPV2 (1 << 13) /* Vector Floating Point V2. */ | |
74a71f7d | 459 | #define FL_WBUF (1 << 14) /* Schedule for write buffer ops. |
460 | Note: ARM6 & 7 derivatives only. */ | |
f655717d | 461 | #define FL_ARCH6K (1 << 15) /* Architecture rel 6 K extensions. */ |
25f905c2 | 462 | #define FL_THUMB2 (1 << 16) /* Thumb-2. */ |
463 | #define FL_NOTM (1 << 17) /* Instructions not present in the 'M' | |
464 | profile. */ | |
542d5028 | 465 | #define FL_DIV (1 << 18) /* Hardware divide. */ |
b5a0636d | 466 | #define FL_VFPV3 (1 << 19) /* Vector Floating Point V3. */ |
d98a3884 | 467 | #define FL_NEON (1 << 20) /* Neon instructions. */ |
82ec49ae | 468 | |
a2cd141b | 469 | #define FL_IWMMXT (1 << 29) /* XScale v2 or "Intel Wireless MMX technology". */ |
470 | ||
25f905c2 | 471 | #define FL_FOR_ARCH2 FL_NOTM |
472 | #define FL_FOR_ARCH3 (FL_FOR_ARCH2 | FL_MODE32) | |
e65e4284 | 473 | #define FL_FOR_ARCH3M (FL_FOR_ARCH3 | FL_ARCH3M) |
474 | #define FL_FOR_ARCH4 (FL_FOR_ARCH3M | FL_ARCH4) | |
475 | #define FL_FOR_ARCH4T (FL_FOR_ARCH4 | FL_THUMB) | |
476 | #define FL_FOR_ARCH5 (FL_FOR_ARCH4 | FL_ARCH5) | |
477 | #define FL_FOR_ARCH5T (FL_FOR_ARCH5 | FL_THUMB) | |
478 | #define FL_FOR_ARCH5E (FL_FOR_ARCH5 | FL_ARCH5E) | |
479 | #define FL_FOR_ARCH5TE (FL_FOR_ARCH5E | FL_THUMB) | |
480 | #define FL_FOR_ARCH5TEJ FL_FOR_ARCH5TE | |
481 | #define FL_FOR_ARCH6 (FL_FOR_ARCH5TE | FL_ARCH6) | |
482 | #define FL_FOR_ARCH6J FL_FOR_ARCH6 | |
f655717d | 483 | #define FL_FOR_ARCH6K (FL_FOR_ARCH6 | FL_ARCH6K) |
6b4d6063 | 484 | #define FL_FOR_ARCH6Z FL_FOR_ARCH6 |
f655717d | 485 | #define FL_FOR_ARCH6ZK FL_FOR_ARCH6K |
25f905c2 | 486 | #define FL_FOR_ARCH6T2 (FL_FOR_ARCH6 | FL_THUMB2) |
487 | #define FL_FOR_ARCH7 (FL_FOR_ARCH6T2 &~ FL_NOTM) | |
488 | #define FL_FOR_ARCH7A (FL_FOR_ARCH7 | FL_NOTM) | |
489 | #define FL_FOR_ARCH7R (FL_FOR_ARCH7A | FL_DIV) | |
490 | #define FL_FOR_ARCH7M (FL_FOR_ARCH7 | FL_DIV) | |
e65e4284 | 491 | |
a12be13c | 492 | /* The bits in this mask specify which |
493 | instructions we are allowed to generate. */ | |
a7712927 | 494 | static unsigned long insn_flags = 0; |
cffb2a26 | 495 | |
82ec49ae | 496 | /* The bits in this mask specify which instruction scheduling options should |
a2cd141b | 497 | be used. */ |
a7712927 | 498 | static unsigned long tune_flags = 0; |
82ec49ae | 499 | |
500 | /* The following are used in the arm.md file as equivalents to bits | |
501 | in the above two flag variables. */ | |
502 | ||
a2cd141b | 503 | /* Nonzero if this chip supports the ARM Architecture 3M extensions. */ |
504 | int arm_arch3m = 0; | |
64266206 | 505 | |
66aacf41 | 506 | /* Nonzero if this chip supports the ARM Architecture 4 extensions. */ |
64266206 | 507 | int arm_arch4 = 0; |
508 | ||
f1039640 | 509 | /* Nonzero if this chip supports the ARM Architecture 4t extensions. */ |
510 | int arm_arch4t = 0; | |
511 | ||
66aacf41 | 512 | /* Nonzero if this chip supports the ARM Architecture 5 extensions. */ |
bba10fb8 | 513 | int arm_arch5 = 0; |
514 | ||
e1159bbe | 515 | /* Nonzero if this chip supports the ARM Architecture 5E extensions. */ |
516 | int arm_arch5e = 0; | |
517 | ||
a2cd141b | 518 | /* Nonzero if this chip supports the ARM Architecture 6 extensions. */ |
519 | int arm_arch6 = 0; | |
520 | ||
f655717d | 521 | /* Nonzero if this chip supports the ARM 6K extensions. */ |
522 | int arm_arch6k = 0; | |
523 | ||
25f905c2 | 524 | /* Nonzero if instructions not present in the 'M' profile can be used. */ |
525 | int arm_arch_notm = 0; | |
526 | ||
82ec49ae | 527 | /* Nonzero if this chip can benefit from load scheduling. */ |
9888ad6d | 528 | int arm_ld_sched = 0; |
529 | ||
530 | /* Nonzero if this chip is a StrongARM. */ | |
74a71f7d | 531 | int arm_tune_strongarm = 0; |
9888ad6d | 532 | |
e65e4284 | 533 | /* Nonzero if this chip is a Cirrus variant. */ |
534 | int arm_arch_cirrus = 0; | |
535 | ||
755eb2b4 | 536 | /* Nonzero if this chip supports Intel Wireless MMX technology. */ |
537 | int arm_arch_iwmmxt = 0; | |
538 | ||
331beb1a | 539 | /* Nonzero if this chip is an XScale. */ |
f8e17552 | 540 | int arm_arch_xscale = 0; |
541 | ||
542 | /* Nonzero if tuning for XScale */ | |
543 | int arm_tune_xscale = 0; | |
331beb1a | 544 | |
a60f1840 | 545 | /* Nonzero if we want to tune for stores that access the write-buffer. |
bf2fbc3d | 546 | This typically means an ARM6 or ARM7 with MMU or MPU. */ |
74a71f7d | 547 | int arm_tune_wbuf = 0; |
f21249d5 | 548 | |
1c494086 | 549 | /* Nonzero if generating Thumb instructions. */ |
550 | int thumb_code = 0; | |
551 | ||
80b6339a | 552 | /* Nonzero if we should define __THUMB_INTERWORK__ in the |
9e7454d0 | 553 | preprocessor. |
80b6339a | 554 | XXX This is a bit of a hack, it's intended to help work around |
555 | problems in GLD which doesn't understand that armv5t code is | |
556 | interworking clean. */ | |
557 | int arm_cpp_interwork = 0; | |
558 | ||
25f905c2 | 559 | /* Nonzero if chip supports Thumb 2. */ |
560 | int arm_arch_thumb2; | |
561 | ||
562 | /* Nonzero if chip supports integer division instruction. */ | |
563 | int arm_arch_hwdiv; | |
564 | ||
8cd42699 | 565 | /* In case of a PRE_INC, POST_INC, PRE_DEC, POST_DEC memory reference, we |
566 | must report the mode of the memory reference from PRINT_OPERAND to | |
567 | PRINT_OPERAND_ADDRESS. */ | |
a690ede2 | 568 | enum machine_mode output_memory_reference_mode; |
8cd42699 | 569 | |
69120d26 | 570 | /* The register number to be used for the PIC offset register. */ |
2cb7d577 | 571 | unsigned arm_pic_register = INVALID_REGNUM; |
69120d26 | 572 | |
9c08d1fa | 573 | /* Set to 1 when a return insn is output, this means that the epilogue |
66aacf41 | 574 | is not needed. */ |
cffb2a26 | 575 | int return_used_this_function; |
9c08d1fa | 576 | |
82ec49ae | 577 | /* Set to 1 after arm_reorg has started. Reset to start at the start of |
578 | the next function. */ | |
11de4c7f | 579 | static int after_arm_reorg = 0; |
580 | ||
82ec49ae | 581 | /* The maximum number of insns to be used when loading a constant. */ |
64266206 | 582 | static int arm_constant_limit = 3; |
583 | ||
8cd42699 | 584 | /* For an explanation of these variables, see final_prescan_insn below. */ |
585 | int arm_ccfsm_state; | |
25f905c2 | 586 | /* arm_current_cc is also used for Thumb-2 cond_exec blocks. */ |
1275b964 | 587 | enum arm_cond_code arm_current_cc; |
8cd42699 | 588 | rtx arm_target_insn; |
589 | int arm_target_label; | |
25f905c2 | 590 | /* The number of conditionally executed insns, including the current insn. */ |
591 | int arm_condexec_count = 0; | |
592 | /* A bitmask specifying the patterns for the IT block. | |
593 | Zero means do not output an IT block before this insn. */ | |
594 | int arm_condexec_mask = 0; | |
595 | /* The number of bits used in arm_condexec_mask. */ | |
596 | int arm_condexec_masklen = 0; | |
f6b01912 | 597 | |
598 | /* The condition codes of the ARM, and the inverse function. */ | |
a12be13c | 599 | static const char * const arm_condition_codes[] = |
f6b01912 | 600 | { |
601 | "eq", "ne", "cs", "cc", "mi", "pl", "vs", "vc", | |
602 | "hi", "ls", "ge", "lt", "gt", "le", "al", "nv" | |
603 | }; | |
604 | ||
25f905c2 | 605 | #define ARM_LSL_NAME (TARGET_UNIFIED_ASM ? "lsl" : "asl") |
9888ad6d | 606 | #define streq(string1, string2) (strcmp (string1, string2) == 0) |
25f905c2 | 607 | |
608 | #define THUMB2_WORK_REGS (0xff & ~( (1 << THUMB_HARD_FRAME_POINTER_REGNUM) \ | |
609 | | (1 << SP_REGNUM) | (1 << PC_REGNUM) \ | |
610 | | (1 << PIC_OFFSET_TABLE_REGNUM))) | |
64266206 | 611 | \f |
66aacf41 | 612 | /* Initialization code. */ |
64266206 | 613 | |
64266206 | 614 | struct processors |
615 | { | |
e99c3a1d | 616 | const char *const name; |
a2cd141b | 617 | enum processor_type core; |
e65e4284 | 618 | const char *arch; |
a7712927 | 619 | const unsigned long flags; |
a2cd141b | 620 | bool (* rtx_costs) (rtx, int, int, int *); |
64266206 | 621 | }; |
622 | ||
623 | /* Not all of these give usefully different compilation alternatives, | |
624 | but there is no simple way of generalizing them. */ | |
e99c3a1d | 625 | static const struct processors all_cores[] = |
9888ad6d | 626 | { |
627 | /* ARM Cores */ | |
61b58075 | 628 | #define ARM_CORE(NAME, IDENT, ARCH, FLAGS, COSTS) \ |
629 | {NAME, arm_none, #ARCH, FLAGS | FL_FOR_ARCH##ARCH, arm_##COSTS##_rtx_costs}, | |
a2cd141b | 630 | #include "arm-cores.def" |
631 | #undef ARM_CORE | |
e65e4284 | 632 | {NULL, arm_none, NULL, 0, NULL} |
9888ad6d | 633 | }; |
634 | ||
e99c3a1d | 635 | static const struct processors all_architectures[] = |
64266206 | 636 | { |
9888ad6d | 637 | /* ARM Architectures */ |
a2cd141b | 638 | /* We don't specify rtx_costs here as it will be figured out |
639 | from the core. */ | |
9e7454d0 | 640 | |
e65e4284 | 641 | {"armv2", arm2, "2", FL_CO_PROC | FL_MODE26 | FL_FOR_ARCH2, NULL}, |
642 | {"armv2a", arm2, "2", FL_CO_PROC | FL_MODE26 | FL_FOR_ARCH2, NULL}, | |
643 | {"armv3", arm6, "3", FL_CO_PROC | FL_MODE26 | FL_FOR_ARCH3, NULL}, | |
644 | {"armv3m", arm7m, "3M", FL_CO_PROC | FL_MODE26 | FL_FOR_ARCH3M, NULL}, | |
645 | {"armv4", arm7tdmi, "4", FL_CO_PROC | FL_MODE26 | FL_FOR_ARCH4, NULL}, | |
f21249d5 | 646 | /* Strictly, FL_MODE26 is a permitted option for v4t, but there are no |
647 | implementations that support it, so we will leave it out for now. */ | |
e65e4284 | 648 | {"armv4t", arm7tdmi, "4T", FL_CO_PROC | FL_FOR_ARCH4T, NULL}, |
649 | {"armv5", arm10tdmi, "5", FL_CO_PROC | FL_FOR_ARCH5, NULL}, | |
650 | {"armv5t", arm10tdmi, "5T", FL_CO_PROC | FL_FOR_ARCH5T, NULL}, | |
651 | {"armv5e", arm1026ejs, "5E", FL_CO_PROC | FL_FOR_ARCH5E, NULL}, | |
652 | {"armv5te", arm1026ejs, "5TE", FL_CO_PROC | FL_FOR_ARCH5TE, NULL}, | |
653 | {"armv6", arm1136js, "6", FL_CO_PROC | FL_FOR_ARCH6, NULL}, | |
654 | {"armv6j", arm1136js, "6J", FL_CO_PROC | FL_FOR_ARCH6J, NULL}, | |
6b4d6063 | 655 | {"armv6k", mpcore, "6K", FL_CO_PROC | FL_FOR_ARCH6K, NULL}, |
656 | {"armv6z", arm1176jzs, "6Z", FL_CO_PROC | FL_FOR_ARCH6Z, NULL}, | |
657 | {"armv6zk", arm1176jzs, "6ZK", FL_CO_PROC | FL_FOR_ARCH6ZK, NULL}, | |
25f905c2 | 658 | {"armv6t2", arm1156t2s, "6T2", FL_CO_PROC | FL_FOR_ARCH6T2, NULL}, |
659 | {"armv7", cortexa8, "7", FL_CO_PROC | FL_FOR_ARCH7, NULL}, | |
660 | {"armv7-a", cortexa8, "7A", FL_CO_PROC | FL_FOR_ARCH7A, NULL}, | |
661 | {"armv7-r", cortexr4, "7R", FL_CO_PROC | FL_FOR_ARCH7R, NULL}, | |
662 | {"armv7-m", cortexm3, "7M", FL_CO_PROC | FL_FOR_ARCH7M, NULL}, | |
e65e4284 | 663 | {"ep9312", ep9312, "4T", FL_LDSCHED | FL_CIRRUS | FL_FOR_ARCH4, NULL}, |
664 | {"iwmmxt", iwmmxt, "5TE", FL_LDSCHED | FL_STRONG | FL_FOR_ARCH5TE | FL_XSCALE | FL_IWMMXT , NULL}, | |
665 | {NULL, arm_none, NULL, 0 , NULL} | |
9888ad6d | 666 | }; |
667 | ||
81f6de2f | 668 | struct arm_cpu_select |
669 | { | |
670 | const char * string; | |
671 | const char * name; | |
672 | const struct processors * processors; | |
673 | }; | |
674 | ||
8d232dc7 | 675 | /* This is a magic structure. The 'string' field is magically filled in |
9888ad6d | 676 | with a pointer to the value specified by the user on the command line |
677 | assuming that the user has specified such a value. */ | |
678 | ||
81f6de2f | 679 | static struct arm_cpu_select arm_select[] = |
9888ad6d | 680 | { |
9e7454d0 | 681 | /* string name processors */ |
9888ad6d | 682 | { NULL, "-mcpu=", all_cores }, |
683 | { NULL, "-march=", all_architectures }, | |
684 | { NULL, "-mtune=", all_cores } | |
64266206 | 685 | }; |
686 | ||
c37752cc | 687 | /* Defines representing the indexes into the above table. */ |
688 | #define ARM_OPT_SET_CPU 0 | |
689 | #define ARM_OPT_SET_ARCH 1 | |
690 | #define ARM_OPT_SET_TUNE 2 | |
e65e4284 | 691 | |
15b00d63 | 692 | /* The name of the preprocessor macro to define for this architecture. */ |
e65e4284 | 693 | |
694 | char arm_arch_name[] = "__ARM_ARCH_0UNK__"; | |
695 | ||
a2cd141b | 696 | struct fpu_desc |
697 | { | |
698 | const char * name; | |
699 | enum fputype fpu; | |
700 | }; | |
701 | ||
702 | ||
cb4213cd | 703 | /* Available values for -mfpu=. */ |
a2cd141b | 704 | |
705 | static const struct fpu_desc all_fpus[] = | |
706 | { | |
707 | {"fpa", FPUTYPE_FPA}, | |
708 | {"fpe2", FPUTYPE_FPA_EMU2}, | |
709 | {"fpe3", FPUTYPE_FPA_EMU2}, | |
710 | {"maverick", FPUTYPE_MAVERICK}, | |
b5a0636d | 711 | {"vfp", FPUTYPE_VFP}, |
712 | {"vfp3", FPUTYPE_VFP3}, | |
d98a3884 | 713 | {"neon", FPUTYPE_NEON} |
a2cd141b | 714 | }; |
715 | ||
716 | ||
717 | /* Floating point models used by the different hardware. | |
718 | See fputype in arm.h. */ | |
719 | ||
720 | static const enum fputype fp_model_for_fpu[] = | |
721 | { | |
722 | /* No FP hardware. */ | |
723 | ARM_FP_MODEL_UNKNOWN, /* FPUTYPE_NONE */ | |
724 | ARM_FP_MODEL_FPA, /* FPUTYPE_FPA */ | |
725 | ARM_FP_MODEL_FPA, /* FPUTYPE_FPA_EMU2 */ | |
726 | ARM_FP_MODEL_FPA, /* FPUTYPE_FPA_EMU3 */ | |
727 | ARM_FP_MODEL_MAVERICK, /* FPUTYPE_MAVERICK */ | |
b5a0636d | 728 | ARM_FP_MODEL_VFP, /* FPUTYPE_VFP */ |
d98a3884 | 729 | ARM_FP_MODEL_VFP, /* FPUTYPE_VFP3 */ |
730 | ARM_FP_MODEL_VFP /* FPUTYPE_NEON */ | |
a2cd141b | 731 | }; |
732 | ||
733 | ||
734 | struct float_abi | |
735 | { | |
736 | const char * name; | |
737 | enum float_abi_type abi_type; | |
738 | }; | |
739 | ||
740 | ||
741 | /* Available values for -mfloat-abi=. */ | |
742 | ||
743 | static const struct float_abi all_float_abis[] = | |
744 | { | |
745 | {"soft", ARM_FLOAT_ABI_SOFT}, | |
746 | {"softfp", ARM_FLOAT_ABI_SOFTFP}, | |
747 | {"hard", ARM_FLOAT_ABI_HARD} | |
748 | }; | |
749 | ||
750 | ||
f9273c43 | 751 | struct abi_name |
752 | { | |
753 | const char *name; | |
754 | enum arm_abi_type abi_type; | |
755 | }; | |
756 | ||
757 | ||
758 | /* Available values for -mabi=. */ | |
759 | ||
760 | static const struct abi_name arm_all_abis[] = | |
761 | { | |
762 | {"apcs-gnu", ARM_ABI_APCS}, | |
763 | {"atpcs", ARM_ABI_ATPCS}, | |
764 | {"aapcs", ARM_ABI_AAPCS}, | |
6783d878 | 765 | {"iwmmxt", ARM_ABI_IWMMXT}, |
766 | {"aapcs-linux", ARM_ABI_AAPCS_LINUX} | |
f9273c43 | 767 | }; |
768 | ||
f655717d | 769 | /* Supported TLS relocations. */ |
770 | ||
771 | enum tls_reloc { | |
772 | TLS_GD32, | |
773 | TLS_LDM32, | |
774 | TLS_LDO32, | |
775 | TLS_IE32, | |
776 | TLS_LE32 | |
777 | }; | |
778 | ||
0083a3b9 | 779 | /* Emit an insn that's a simple single-set. Both the operands must be known |
780 | to be valid. */ | |
781 | inline static rtx | |
782 | emit_set_insn (rtx x, rtx y) | |
783 | { | |
784 | return emit_insn (gen_rtx_SET (VOIDmode, x, y)); | |
785 | } | |
786 | ||
a7712927 | 787 | /* Return the number of bits set in VALUE. */ |
788 | static unsigned | |
ebd88a36 | 789 | bit_count (unsigned long value) |
82ec49ae | 790 | { |
cffb2a26 | 791 | unsigned long count = 0; |
9e7454d0 | 792 | |
82ec49ae | 793 | while (value) |
794 | { | |
a7712927 | 795 | count++; |
796 | value &= value - 1; /* Clear the least-significant set bit. */ | |
82ec49ae | 797 | } |
798 | ||
799 | return count; | |
800 | } | |
801 | ||
84cbcde5 | 802 | /* Set up library functions unique to ARM. */ |
30e9913f | 803 | |
804 | static void | |
805 | arm_init_libfuncs (void) | |
806 | { | |
807 | /* There are no special library functions unless we are using the | |
808 | ARM BPABI. */ | |
809 | if (!TARGET_BPABI) | |
810 | return; | |
811 | ||
812 | /* The functions below are described in Section 4 of the "Run-Time | |
813 | ABI for the ARM architecture", Version 1.0. */ | |
814 | ||
815 | /* Double-precision floating-point arithmetic. Table 2. */ | |
816 | set_optab_libfunc (add_optab, DFmode, "__aeabi_dadd"); | |
817 | set_optab_libfunc (sdiv_optab, DFmode, "__aeabi_ddiv"); | |
818 | set_optab_libfunc (smul_optab, DFmode, "__aeabi_dmul"); | |
819 | set_optab_libfunc (neg_optab, DFmode, "__aeabi_dneg"); | |
820 | set_optab_libfunc (sub_optab, DFmode, "__aeabi_dsub"); | |
821 | ||
84cbcde5 | 822 | /* Double-precision comparisons. Table 3. */ |
30e9913f | 823 | set_optab_libfunc (eq_optab, DFmode, "__aeabi_dcmpeq"); |
824 | set_optab_libfunc (ne_optab, DFmode, NULL); | |
825 | set_optab_libfunc (lt_optab, DFmode, "__aeabi_dcmplt"); | |
826 | set_optab_libfunc (le_optab, DFmode, "__aeabi_dcmple"); | |
827 | set_optab_libfunc (ge_optab, DFmode, "__aeabi_dcmpge"); | |
828 | set_optab_libfunc (gt_optab, DFmode, "__aeabi_dcmpgt"); | |
829 | set_optab_libfunc (unord_optab, DFmode, "__aeabi_dcmpun"); | |
830 | ||
831 | /* Single-precision floating-point arithmetic. Table 4. */ | |
832 | set_optab_libfunc (add_optab, SFmode, "__aeabi_fadd"); | |
833 | set_optab_libfunc (sdiv_optab, SFmode, "__aeabi_fdiv"); | |
834 | set_optab_libfunc (smul_optab, SFmode, "__aeabi_fmul"); | |
835 | set_optab_libfunc (neg_optab, SFmode, "__aeabi_fneg"); | |
836 | set_optab_libfunc (sub_optab, SFmode, "__aeabi_fsub"); | |
9e7454d0 | 837 | |
84cbcde5 | 838 | /* Single-precision comparisons. Table 5. */ |
30e9913f | 839 | set_optab_libfunc (eq_optab, SFmode, "__aeabi_fcmpeq"); |
840 | set_optab_libfunc (ne_optab, SFmode, NULL); | |
841 | set_optab_libfunc (lt_optab, SFmode, "__aeabi_fcmplt"); | |
842 | set_optab_libfunc (le_optab, SFmode, "__aeabi_fcmple"); | |
843 | set_optab_libfunc (ge_optab, SFmode, "__aeabi_fcmpge"); | |
844 | set_optab_libfunc (gt_optab, SFmode, "__aeabi_fcmpgt"); | |
845 | set_optab_libfunc (unord_optab, SFmode, "__aeabi_fcmpun"); | |
846 | ||
847 | /* Floating-point to integer conversions. Table 6. */ | |
848 | set_conv_libfunc (sfix_optab, SImode, DFmode, "__aeabi_d2iz"); | |
849 | set_conv_libfunc (ufix_optab, SImode, DFmode, "__aeabi_d2uiz"); | |
850 | set_conv_libfunc (sfix_optab, DImode, DFmode, "__aeabi_d2lz"); | |
851 | set_conv_libfunc (ufix_optab, DImode, DFmode, "__aeabi_d2ulz"); | |
852 | set_conv_libfunc (sfix_optab, SImode, SFmode, "__aeabi_f2iz"); | |
853 | set_conv_libfunc (ufix_optab, SImode, SFmode, "__aeabi_f2uiz"); | |
854 | set_conv_libfunc (sfix_optab, DImode, SFmode, "__aeabi_f2lz"); | |
855 | set_conv_libfunc (ufix_optab, DImode, SFmode, "__aeabi_f2ulz"); | |
856 | ||
857 | /* Conversions between floating types. Table 7. */ | |
858 | set_conv_libfunc (trunc_optab, SFmode, DFmode, "__aeabi_d2f"); | |
859 | set_conv_libfunc (sext_optab, DFmode, SFmode, "__aeabi_f2d"); | |
860 | ||
84cbcde5 | 861 | /* Integer to floating-point conversions. Table 8. */ |
30e9913f | 862 | set_conv_libfunc (sfloat_optab, DFmode, SImode, "__aeabi_i2d"); |
863 | set_conv_libfunc (ufloat_optab, DFmode, SImode, "__aeabi_ui2d"); | |
864 | set_conv_libfunc (sfloat_optab, DFmode, DImode, "__aeabi_l2d"); | |
865 | set_conv_libfunc (ufloat_optab, DFmode, DImode, "__aeabi_ul2d"); | |
866 | set_conv_libfunc (sfloat_optab, SFmode, SImode, "__aeabi_i2f"); | |
867 | set_conv_libfunc (ufloat_optab, SFmode, SImode, "__aeabi_ui2f"); | |
868 | set_conv_libfunc (sfloat_optab, SFmode, DImode, "__aeabi_l2f"); | |
869 | set_conv_libfunc (ufloat_optab, SFmode, DImode, "__aeabi_ul2f"); | |
870 | ||
871 | /* Long long. Table 9. */ | |
872 | set_optab_libfunc (smul_optab, DImode, "__aeabi_lmul"); | |
873 | set_optab_libfunc (sdivmod_optab, DImode, "__aeabi_ldivmod"); | |
874 | set_optab_libfunc (udivmod_optab, DImode, "__aeabi_uldivmod"); | |
875 | set_optab_libfunc (ashl_optab, DImode, "__aeabi_llsl"); | |
876 | set_optab_libfunc (lshr_optab, DImode, "__aeabi_llsr"); | |
877 | set_optab_libfunc (ashr_optab, DImode, "__aeabi_lasr"); | |
878 | set_optab_libfunc (cmp_optab, DImode, "__aeabi_lcmp"); | |
879 | set_optab_libfunc (ucmp_optab, DImode, "__aeabi_ulcmp"); | |
880 | ||
881 | /* Integer (32/32->32) division. \S 4.3.1. */ | |
882 | set_optab_libfunc (sdivmod_optab, SImode, "__aeabi_idivmod"); | |
883 | set_optab_libfunc (udivmod_optab, SImode, "__aeabi_uidivmod"); | |
884 | ||
885 | /* The divmod functions are designed so that they can be used for | |
886 | plain division, even though they return both the quotient and the | |
887 | remainder. The quotient is returned in the usual location (i.e., | |
888 | r0 for SImode, {r0, r1} for DImode), just as would be expected | |
889 | for an ordinary division routine. Because the AAPCS calling | |
890 | conventions specify that all of { r0, r1, r2, r3 } are | |
891 | callee-saved registers, there is no need to tell the compiler | |
892 | explicitly that those registers are clobbered by these | |
893 | routines. */ | |
894 | set_optab_libfunc (sdiv_optab, DImode, "__aeabi_ldivmod"); | |
895 | set_optab_libfunc (udiv_optab, DImode, "__aeabi_uldivmod"); | |
3da35908 | 896 | |
897 | /* For SImode division the ABI provides div-without-mod routines, | |
898 | which are faster. */ | |
899 | set_optab_libfunc (sdiv_optab, SImode, "__aeabi_idiv"); | |
900 | set_optab_libfunc (udiv_optab, SImode, "__aeabi_uidiv"); | |
b027ddbf | 901 | |
902 | /* We don't have mod libcalls. Fortunately gcc knows how to use the | |
903 | divmod libcalls instead. */ | |
904 | set_optab_libfunc (smod_optab, DImode, NULL); | |
905 | set_optab_libfunc (umod_optab, DImode, NULL); | |
906 | set_optab_libfunc (smod_optab, SImode, NULL); | |
907 | set_optab_libfunc (umod_optab, SImode, NULL); | |
30e9913f | 908 | } |
909 | ||
81f6de2f | 910 | /* Implement TARGET_HANDLE_OPTION. */ |
911 | ||
912 | static bool | |
913 | arm_handle_option (size_t code, const char *arg, int value ATTRIBUTE_UNUSED) | |
914 | { | |
915 | switch (code) | |
916 | { | |
81f6de2f | 917 | case OPT_march_: |
918 | arm_select[1].string = arg; | |
919 | return true; | |
920 | ||
921 | case OPT_mcpu_: | |
922 | arm_select[0].string = arg; | |
923 | return true; | |
924 | ||
81f6de2f | 925 | case OPT_mhard_float: |
926 | target_float_abi_name = "hard"; | |
927 | return true; | |
928 | ||
81f6de2f | 929 | case OPT_msoft_float: |
930 | target_float_abi_name = "soft"; | |
931 | return true; | |
932 | ||
81f6de2f | 933 | case OPT_mtune_: |
934 | arm_select[2].string = arg; | |
935 | return true; | |
936 | ||
937 | default: | |
938 | return true; | |
939 | } | |
940 | } | |
941 | ||
154ee3b8 | 942 | static void |
943 | arm_target_help (void) | |
944 | { | |
945 | int i; | |
946 | static int columns = 0; | |
947 | int remaining; | |
948 | ||
949 | /* If we have not done so already, obtain the desired maximum width of | |
950 | the output. Note - this is a duplication of the code at the start of | |
951 | gcc/opts.c:print_specific_help() - the two copies should probably be | |
952 | replaced by a single function. */ | |
953 | if (columns == 0) | |
954 | { | |
955 | const char *p; | |
956 | ||
957 | GET_ENVIRONMENT (p, "COLUMNS"); | |
958 | if (p != NULL) | |
959 | { | |
960 | int value = atoi (p); | |
961 | ||
962 | if (value > 0) | |
963 | columns = value; | |
964 | } | |
965 | ||
966 | if (columns == 0) | |
967 | /* Use a reasonable default. */ | |
968 | columns = 80; | |
969 | } | |
970 | ||
971 | printf (" Known ARM CPUs (for use with the -mcpu= and -mtune= options):\n"); | |
972 | ||
973 | /* The - 2 is because we know that the last entry in the array is NULL. */ | |
974 | i = ARRAY_SIZE (all_cores) - 2; | |
975 | gcc_assert (i > 0); | |
976 | printf (" %s", all_cores[i].name); | |
977 | remaining = columns - (strlen (all_cores[i].name) + 4); | |
978 | gcc_assert (remaining >= 0); | |
979 | ||
980 | while (i--) | |
981 | { | |
982 | int len = strlen (all_cores[i].name); | |
983 | ||
984 | if (remaining > len + 2) | |
985 | { | |
986 | printf (", %s", all_cores[i].name); | |
987 | remaining -= len + 2; | |
988 | } | |
989 | else | |
990 | { | |
991 | if (remaining > 0) | |
992 | printf (","); | |
993 | printf ("\n %s", all_cores[i].name); | |
994 | remaining = columns - (len + 4); | |
995 | } | |
996 | } | |
997 | ||
998 | printf ("\n\n Known ARM architectures (for use with the -march= option):\n"); | |
999 | ||
1000 | i = ARRAY_SIZE (all_architectures) - 2; | |
1001 | gcc_assert (i > 0); | |
1002 | ||
1003 | printf (" %s", all_architectures[i].name); | |
1004 | remaining = columns - (strlen (all_architectures[i].name) + 4); | |
1005 | gcc_assert (remaining >= 0); | |
1006 | ||
1007 | while (i--) | |
1008 | { | |
1009 | int len = strlen (all_architectures[i].name); | |
1010 | ||
1011 | if (remaining > len + 2) | |
1012 | { | |
1013 | printf (", %s", all_architectures[i].name); | |
1014 | remaining -= len + 2; | |
1015 | } | |
1016 | else | |
1017 | { | |
1018 | if (remaining > 0) | |
1019 | printf (","); | |
1020 | printf ("\n %s", all_architectures[i].name); | |
1021 | remaining = columns - (len + 4); | |
1022 | } | |
1023 | } | |
1024 | printf ("\n"); | |
1025 | ||
1026 | } | |
1027 | ||
64266206 | 1028 | /* Fix up any incompatible options that the user has specified. |
1029 | This has now turned into a maze. */ | |
1030 | void | |
ebd88a36 | 1031 | arm_override_options (void) |
64266206 | 1032 | { |
ae175e8f | 1033 | unsigned i; |
c37752cc | 1034 | enum processor_type target_arch_cpu = arm_none; |
a2cd141b | 1035 | |
9888ad6d | 1036 | /* Set up the flags based on the cpu/architecture selected by the user. */ |
3098b2d3 | 1037 | for (i = ARRAY_SIZE (arm_select); i--;) |
9480e6c7 | 1038 | { |
9888ad6d | 1039 | struct arm_cpu_select * ptr = arm_select + i; |
9e7454d0 | 1040 | |
9888ad6d | 1041 | if (ptr->string != NULL && ptr->string[0] != '\0') |
9480e6c7 | 1042 | { |
c9ba03f2 | 1043 | const struct processors * sel; |
9480e6c7 | 1044 | |
215b30b3 | 1045 | for (sel = ptr->processors; sel->name != NULL; sel++) |
9888ad6d | 1046 | if (streq (ptr->string, sel->name)) |
9480e6c7 | 1047 | { |
e65e4284 | 1048 | /* Set the architecture define. */ |
c37752cc | 1049 | if (i != ARM_OPT_SET_TUNE) |
e65e4284 | 1050 | sprintf (arm_arch_name, "__ARM_ARCH_%s__", sel->arch); |
1051 | ||
a2cd141b | 1052 | /* Determine the processor core for which we should |
1053 | tune code-generation. */ | |
1054 | if (/* -mcpu= is a sensible default. */ | |
c37752cc | 1055 | i == ARM_OPT_SET_CPU |
a2cd141b | 1056 | /* -mtune= overrides -mcpu= and -march=. */ |
c37752cc | 1057 | || i == ARM_OPT_SET_TUNE) |
a2cd141b | 1058 | arm_tune = (enum processor_type) (sel - ptr->processors); |
1059 | ||
c37752cc | 1060 | /* Remember the CPU associated with this architecture. |
1061 | If no other option is used to set the CPU type, | |
1062 | we'll use this to guess the most suitable tuning | |
1063 | options. */ | |
1064 | if (i == ARM_OPT_SET_ARCH) | |
1065 | target_arch_cpu = sel->core; | |
a60f1840 | 1066 | |
c37752cc | 1067 | if (i != ARM_OPT_SET_TUNE) |
f21249d5 | 1068 | { |
82ec49ae | 1069 | /* If we have been given an architecture and a processor |
1070 | make sure that they are compatible. We only generate | |
1071 | a warning though, and we prefer the CPU over the | |
66aacf41 | 1072 | architecture. */ |
82ec49ae | 1073 | if (insn_flags != 0 && (insn_flags ^ sel->flags)) |
c3ceba8e | 1074 | warning (0, "switch -mcpu=%s conflicts with -march= switch", |
82ec49ae | 1075 | ptr->string); |
9e7454d0 | 1076 | |
82ec49ae | 1077 | insn_flags = sel->flags; |
f21249d5 | 1078 | } |
9e7454d0 | 1079 | |
9480e6c7 | 1080 | break; |
1081 | } | |
1082 | ||
1083 | if (sel->name == NULL) | |
1084 | error ("bad value (%s) for %s switch", ptr->string, ptr->name); | |
1085 | } | |
1086 | } | |
9e7454d0 | 1087 | |
c37752cc | 1088 | /* Guess the tuning options from the architecture if necessary. */ |
1089 | if (arm_tune == arm_none) | |
1090 | arm_tune = target_arch_cpu; | |
1091 | ||
9888ad6d | 1092 | /* If the user did not specify a processor, choose one for them. */ |
82ec49ae | 1093 | if (insn_flags == 0) |
9888ad6d | 1094 | { |
e99c3a1d | 1095 | const struct processors * sel; |
82ec49ae | 1096 | unsigned int sought; |
e65e4284 | 1097 | enum processor_type cpu; |
82ec49ae | 1098 | |
e65e4284 | 1099 | cpu = TARGET_CPU_DEFAULT; |
1100 | if (cpu == arm_none) | |
1101 | { | |
1102 | #ifdef SUBTARGET_CPU_DEFAULT | |
1103 | /* Use the subtarget default CPU if none was specified by | |
1104 | configure. */ | |
1105 | cpu = SUBTARGET_CPU_DEFAULT; | |
1106 | #endif | |
1107 | /* Default to ARM6. */ | |
1108 | if (cpu == arm_none) | |
1109 | cpu = arm6; | |
1110 | } | |
1111 | sel = &all_cores[cpu]; | |
82ec49ae | 1112 | |
1113 | insn_flags = sel->flags; | |
a2cd141b | 1114 | |
82ec49ae | 1115 | /* Now check to see if the user has specified some command line |
1116 | switch that require certain abilities from the cpu. */ | |
1117 | sought = 0; | |
9e7454d0 | 1118 | |
cffb2a26 | 1119 | if (TARGET_INTERWORK || TARGET_THUMB) |
9888ad6d | 1120 | { |
82ec49ae | 1121 | sought |= (FL_THUMB | FL_MODE32); |
9e7454d0 | 1122 | |
cffb2a26 | 1123 | /* There are no ARM processors that support both APCS-26 and |
82ec49ae | 1124 | interworking. Therefore we force FL_MODE26 to be removed |
1125 | from insn_flags here (if it was set), so that the search | |
1126 | below will always be able to find a compatible processor. */ | |
215b30b3 | 1127 | insn_flags &= ~FL_MODE26; |
9888ad6d | 1128 | } |
9e7454d0 | 1129 | |
82ec49ae | 1130 | if (sought != 0 && ((sought & insn_flags) != sought)) |
9888ad6d | 1131 | { |
82ec49ae | 1132 | /* Try to locate a CPU type that supports all of the abilities |
1133 | of the default CPU, plus the extra abilities requested by | |
1134 | the user. */ | |
215b30b3 | 1135 | for (sel = all_cores; sel->name != NULL; sel++) |
82ec49ae | 1136 | if ((sel->flags & sought) == (sought | insn_flags)) |
9888ad6d | 1137 | break; |
1138 | ||
1139 | if (sel->name == NULL) | |
82ec49ae | 1140 | { |
a7712927 | 1141 | unsigned current_bit_count = 0; |
e99c3a1d | 1142 | const struct processors * best_fit = NULL; |
9e7454d0 | 1143 | |
82ec49ae | 1144 | /* Ideally we would like to issue an error message here |
1145 | saying that it was not possible to find a CPU compatible | |
1146 | with the default CPU, but which also supports the command | |
1147 | line options specified by the programmer, and so they | |
1148 | ought to use the -mcpu=<name> command line option to | |
1149 | override the default CPU type. | |
1150 | ||
c1a66faf | 1151 | If we cannot find a cpu that has both the |
1152 | characteristics of the default cpu and the given | |
1153 | command line options we scan the array again looking | |
1154 | for a best match. */ | |
215b30b3 | 1155 | for (sel = all_cores; sel->name != NULL; sel++) |
82ec49ae | 1156 | if ((sel->flags & sought) == sought) |
1157 | { | |
a7712927 | 1158 | unsigned count; |
82ec49ae | 1159 | |
1160 | count = bit_count (sel->flags & insn_flags); | |
1161 | ||
1162 | if (count >= current_bit_count) | |
1163 | { | |
1164 | best_fit = sel; | |
1165 | current_bit_count = count; | |
1166 | } | |
1167 | } | |
9888ad6d | 1168 | |
ed29c566 | 1169 | gcc_assert (best_fit); |
1170 | sel = best_fit; | |
82ec49ae | 1171 | } |
1172 | ||
1173 | insn_flags = sel->flags; | |
9888ad6d | 1174 | } |
e65e4284 | 1175 | sprintf (arm_arch_name, "__ARM_ARCH_%s__", sel->arch); |
5fad5123 | 1176 | arm_default_cpu = (enum processor_type) (sel - all_cores); |
a2cd141b | 1177 | if (arm_tune == arm_none) |
5fad5123 | 1178 | arm_tune = arm_default_cpu; |
9888ad6d | 1179 | } |
9e7454d0 | 1180 | |
a2cd141b | 1181 | /* The processor for which we should tune should now have been |
1182 | chosen. */ | |
ed29c566 | 1183 | gcc_assert (arm_tune != arm_none); |
9e7454d0 | 1184 | |
a2cd141b | 1185 | tune_flags = all_cores[(int)arm_tune].flags; |
c1877da8 | 1186 | if (optimize_size) |
1187 | targetm.rtx_costs = arm_size_rtx_costs; | |
1188 | else | |
1189 | targetm.rtx_costs = all_cores[(int)arm_tune].rtx_costs; | |
9f468b67 | 1190 | |
9888ad6d | 1191 | /* Make sure that the processor choice does not conflict with any of the |
1192 | other command line choices. */ | |
25f905c2 | 1193 | if (TARGET_ARM && !(insn_flags & FL_NOTM)) |
1194 | error ("target CPU does not support ARM mode"); | |
1195 | ||
96576951 | 1196 | if (TARGET_INTERWORK && !(insn_flags & FL_THUMB)) |
9888ad6d | 1197 | { |
c3ceba8e | 1198 | warning (0, "target CPU does not support interworking" ); |
81f6de2f | 1199 | target_flags &= ~MASK_INTERWORK; |
9888ad6d | 1200 | } |
9e7454d0 | 1201 | |
cffb2a26 | 1202 | if (TARGET_THUMB && !(insn_flags & FL_THUMB)) |
1203 | { | |
c3ceba8e | 1204 | warning (0, "target CPU does not support THUMB instructions"); |
81f6de2f | 1205 | target_flags &= ~MASK_THUMB; |
cffb2a26 | 1206 | } |
1207 | ||
1208 | if (TARGET_APCS_FRAME && TARGET_THUMB) | |
1209 | { | |
c3ceba8e | 1210 | /* warning (0, "ignoring -mapcs-frame because -mthumb was used"); */ |
81f6de2f | 1211 | target_flags &= ~MASK_APCS_FRAME; |
cffb2a26 | 1212 | } |
331beb1a | 1213 | |
551836e5 | 1214 | /* Callee super interworking implies thumb interworking. Adding |
1215 | this to the flags here simplifies the logic elsewhere. */ | |
1216 | if (TARGET_THUMB && TARGET_CALLEE_INTERWORKING) | |
1217 | target_flags |= MASK_INTERWORK; | |
1218 | ||
cffb2a26 | 1219 | /* TARGET_BACKTRACE calls leaf_function_p, which causes a crash if done |
1220 | from here where no function is being compiled currently. */ | |
81f6de2f | 1221 | if ((TARGET_TPCS_FRAME || TARGET_TPCS_LEAF_FRAME) && TARGET_ARM) |
c3ceba8e | 1222 | warning (0, "enabling backtrace support is only meaningful when compiling for the Thumb"); |
cffb2a26 | 1223 | |
1224 | if (TARGET_ARM && TARGET_CALLEE_INTERWORKING) | |
c3ceba8e | 1225 | warning (0, "enabling callee interworking support is only meaningful when compiling for the Thumb"); |
cffb2a26 | 1226 | |
1227 | if (TARGET_ARM && TARGET_CALLER_INTERWORKING) | |
c3ceba8e | 1228 | warning (0, "enabling caller interworking support is only meaningful when compiling for the Thumb"); |
cffb2a26 | 1229 | |
215b30b3 | 1230 | if (TARGET_APCS_STACK && !TARGET_APCS_FRAME) |
9888ad6d | 1231 | { |
c3ceba8e | 1232 | warning (0, "-mapcs-stack-check incompatible with -mno-apcs-frame"); |
81f6de2f | 1233 | target_flags |= MASK_APCS_FRAME; |
9888ad6d | 1234 | } |
9e7454d0 | 1235 | |
64266206 | 1236 | if (TARGET_POKE_FUNCTION_NAME) |
81f6de2f | 1237 | target_flags |= MASK_APCS_FRAME; |
9e7454d0 | 1238 | |
64266206 | 1239 | if (TARGET_APCS_REENT && flag_pic) |
f060a027 | 1240 | error ("-fpic and -mapcs-reent are incompatible"); |
9e7454d0 | 1241 | |
64266206 | 1242 | if (TARGET_APCS_REENT) |
c3ceba8e | 1243 | warning (0, "APCS reentrant code not supported. Ignored"); |
9e7454d0 | 1244 | |
cffb2a26 | 1245 | /* If this target is normally configured to use APCS frames, warn if they |
1246 | are turned off and debugging is turned on. */ | |
1247 | if (TARGET_ARM | |
1248 | && write_symbols != NO_DEBUG | |
215b30b3 | 1249 | && !TARGET_APCS_FRAME |
81f6de2f | 1250 | && (TARGET_DEFAULT & MASK_APCS_FRAME)) |
c3ceba8e | 1251 | warning (0, "-g with -mno-apcs-frame may not give sensible debugging"); |
9e7454d0 | 1252 | |
64266206 | 1253 | if (TARGET_APCS_FLOAT) |
c3ceba8e | 1254 | warning (0, "passing floating point arguments in fp regs not yet supported"); |
9e7454d0 | 1255 | |
457275b6 | 1256 | /* Initialize boolean versions of the flags, for use in the arm.md file. */ |
a2cd141b | 1257 | arm_arch3m = (insn_flags & FL_ARCH3M) != 0; |
1258 | arm_arch4 = (insn_flags & FL_ARCH4) != 0; | |
f1039640 | 1259 | arm_arch4t = arm_arch4 & ((insn_flags & FL_THUMB) != 0); |
a2cd141b | 1260 | arm_arch5 = (insn_flags & FL_ARCH5) != 0; |
1261 | arm_arch5e = (insn_flags & FL_ARCH5E) != 0; | |
1262 | arm_arch6 = (insn_flags & FL_ARCH6) != 0; | |
f655717d | 1263 | arm_arch6k = (insn_flags & FL_ARCH6K) != 0; |
25f905c2 | 1264 | arm_arch_notm = (insn_flags & FL_NOTM) != 0; |
1265 | arm_arch_thumb2 = (insn_flags & FL_THUMB2) != 0; | |
a2cd141b | 1266 | arm_arch_xscale = (insn_flags & FL_XSCALE) != 0; |
e65e4284 | 1267 | arm_arch_cirrus = (insn_flags & FL_CIRRUS) != 0; |
a2cd141b | 1268 | |
1269 | arm_ld_sched = (tune_flags & FL_LDSCHED) != 0; | |
74a71f7d | 1270 | arm_tune_strongarm = (tune_flags & FL_STRONG) != 0; |
a2cd141b | 1271 | thumb_code = (TARGET_ARM == 0); |
74a71f7d | 1272 | arm_tune_wbuf = (tune_flags & FL_WBUF) != 0; |
a2cd141b | 1273 | arm_tune_xscale = (tune_flags & FL_XSCALE) != 0; |
1274 | arm_arch_iwmmxt = (insn_flags & FL_IWMMXT) != 0; | |
25f905c2 | 1275 | arm_arch_hwdiv = (insn_flags & FL_DIV) != 0; |
755eb2b4 | 1276 | |
f1039640 | 1277 | /* V5 code we generate is completely interworking capable, so we turn off |
1278 | TARGET_INTERWORK here to avoid many tests later on. */ | |
80b6339a | 1279 | |
1280 | /* XXX However, we must pass the right pre-processor defines to CPP | |
1281 | or GLD can get confused. This is a hack. */ | |
1282 | if (TARGET_INTERWORK) | |
1283 | arm_cpp_interwork = 1; | |
1284 | ||
f1039640 | 1285 | if (arm_arch5) |
81f6de2f | 1286 | target_flags &= ~MASK_INTERWORK; |
f1039640 | 1287 | |
f9273c43 | 1288 | if (target_abi_name) |
1289 | { | |
1290 | for (i = 0; i < ARRAY_SIZE (arm_all_abis); i++) | |
1291 | { | |
1292 | if (streq (arm_all_abis[i].name, target_abi_name)) | |
1293 | { | |
1294 | arm_abi = arm_all_abis[i].abi_type; | |
1295 | break; | |
1296 | } | |
1297 | } | |
1298 | if (i == ARRAY_SIZE (arm_all_abis)) | |
1299 | error ("invalid ABI option: -mabi=%s", target_abi_name); | |
1300 | } | |
1301 | else | |
13a46000 | 1302 | arm_abi = ARM_DEFAULT_ABI; |
f9273c43 | 1303 | |
1304 | if (TARGET_IWMMXT && !ARM_DOUBLEWORD_ALIGN) | |
1305 | error ("iwmmxt requires an AAPCS compatible ABI for proper operation"); | |
1306 | ||
1307 | if (TARGET_IWMMXT_ABI && !TARGET_IWMMXT) | |
1308 | error ("iwmmxt abi requires an iwmmxt capable cpu"); | |
89b30d75 | 1309 | |
a2cd141b | 1310 | arm_fp_model = ARM_FP_MODEL_UNKNOWN; |
1311 | if (target_fpu_name == NULL && target_fpe_name != NULL) | |
7d57ec45 | 1312 | { |
a2cd141b | 1313 | if (streq (target_fpe_name, "2")) |
1314 | target_fpu_name = "fpe2"; | |
1315 | else if (streq (target_fpe_name, "3")) | |
1316 | target_fpu_name = "fpe3"; | |
1317 | else | |
1318 | error ("invalid floating point emulation option: -mfpe=%s", | |
1319 | target_fpe_name); | |
1320 | } | |
1321 | if (target_fpu_name != NULL) | |
1322 | { | |
1323 | /* The user specified a FPU. */ | |
1324 | for (i = 0; i < ARRAY_SIZE (all_fpus); i++) | |
1325 | { | |
1326 | if (streq (all_fpus[i].name, target_fpu_name)) | |
1327 | { | |
1328 | arm_fpu_arch = all_fpus[i].fpu; | |
1329 | arm_fpu_tune = arm_fpu_arch; | |
1330 | arm_fp_model = fp_model_for_fpu[arm_fpu_arch]; | |
1331 | break; | |
1332 | } | |
1333 | } | |
1334 | if (arm_fp_model == ARM_FP_MODEL_UNKNOWN) | |
1335 | error ("invalid floating point option: -mfpu=%s", target_fpu_name); | |
7d57ec45 | 1336 | } |
1337 | else | |
64266206 | 1338 | { |
a2cd141b | 1339 | #ifdef FPUTYPE_DEFAULT |
e65e4284 | 1340 | /* Use the default if it is specified for this platform. */ |
a2cd141b | 1341 | arm_fpu_arch = FPUTYPE_DEFAULT; |
1342 | arm_fpu_tune = FPUTYPE_DEFAULT; | |
1343 | #else | |
1344 | /* Pick one based on CPU type. */ | |
e65e4284 | 1345 | /* ??? Some targets assume FPA is the default. |
a2cd141b | 1346 | if ((insn_flags & FL_VFP) != 0) |
1347 | arm_fpu_arch = FPUTYPE_VFP; | |
e65e4284 | 1348 | else |
1349 | */ | |
1350 | if (arm_arch_cirrus) | |
a2cd141b | 1351 | arm_fpu_arch = FPUTYPE_MAVERICK; |
1352 | else | |
c7f506fd | 1353 | arm_fpu_arch = FPUTYPE_FPA_EMU2; |
a2cd141b | 1354 | #endif |
1355 | if (tune_flags & FL_CO_PROC && arm_fpu_arch == FPUTYPE_FPA_EMU2) | |
1356 | arm_fpu_tune = FPUTYPE_FPA; | |
64266206 | 1357 | else |
a2cd141b | 1358 | arm_fpu_tune = arm_fpu_arch; |
1359 | arm_fp_model = fp_model_for_fpu[arm_fpu_arch]; | |
ed29c566 | 1360 | gcc_assert (arm_fp_model != ARM_FP_MODEL_UNKNOWN); |
a2cd141b | 1361 | } |
1362 | ||
1363 | if (target_float_abi_name != NULL) | |
1364 | { | |
1365 | /* The user specified a FP ABI. */ | |
1366 | for (i = 0; i < ARRAY_SIZE (all_float_abis); i++) | |
1367 | { | |
1368 | if (streq (all_float_abis[i].name, target_float_abi_name)) | |
1369 | { | |
1370 | arm_float_abi = all_float_abis[i].abi_type; | |
1371 | break; | |
1372 | } | |
1373 | } | |
1374 | if (i == ARRAY_SIZE (all_float_abis)) | |
1375 | error ("invalid floating point abi: -mfloat-abi=%s", | |
1376 | target_float_abi_name); | |
64266206 | 1377 | } |
87f8017f | 1378 | else |
1379 | arm_float_abi = TARGET_DEFAULT_FLOAT_ABI; | |
a2cd141b | 1380 | |
eb03dcf8 | 1381 | if (arm_float_abi == ARM_FLOAT_ABI_HARD && TARGET_VFP) |
1382 | sorry ("-mfloat-abi=hard and VFP"); | |
1383 | ||
7f5997af | 1384 | /* FPA and iWMMXt are incompatible because the insn encodings overlap. |
1385 | VFP and iWMMXt can theoretically coexist, but it's unlikely such silicon | |
1386 | will ever exist. GCC makes no attempt to support this combination. */ | |
1387 | if (TARGET_IWMMXT && !TARGET_SOFT_FLOAT) | |
1388 | sorry ("iWMMXt and hardware floating point"); | |
1389 | ||
25f905c2 | 1390 | /* ??? iWMMXt insn patterns need auditing for Thumb-2. */ |
1391 | if (TARGET_THUMB2 && TARGET_IWMMXT) | |
1392 | sorry ("Thumb-2 iWMMXt"); | |
1393 | ||
a2cd141b | 1394 | /* If soft-float is specified then don't use FPU. */ |
1395 | if (TARGET_SOFT_FLOAT) | |
1396 | arm_fpu_arch = FPUTYPE_NONE; | |
9e7454d0 | 1397 | |
9888ad6d | 1398 | /* For arm2/3 there is no need to do any scheduling if there is only |
1399 | a floating point emulator, or we are doing software floating-point. */ | |
a2cd141b | 1400 | if ((TARGET_SOFT_FLOAT |
1401 | || arm_fpu_tune == FPUTYPE_FPA_EMU2 | |
1402 | || arm_fpu_tune == FPUTYPE_FPA_EMU3) | |
55c1e470 | 1403 | && (tune_flags & FL_MODE32) == 0) |
9888ad6d | 1404 | flag_schedule_insns = flag_schedule_insns_after_reload = 0; |
9e7454d0 | 1405 | |
f655717d | 1406 | if (target_thread_switch) |
1407 | { | |
1408 | if (strcmp (target_thread_switch, "soft") == 0) | |
1409 | target_thread_pointer = TP_SOFT; | |
1410 | else if (strcmp (target_thread_switch, "auto") == 0) | |
1411 | target_thread_pointer = TP_AUTO; | |
1412 | else if (strcmp (target_thread_switch, "cp15") == 0) | |
1413 | target_thread_pointer = TP_CP15; | |
1414 | else | |
1415 | error ("invalid thread pointer option: -mtp=%s", target_thread_switch); | |
1416 | } | |
1417 | ||
1418 | /* Use the cp15 method if it is available. */ | |
1419 | if (target_thread_pointer == TP_AUTO) | |
1420 | { | |
1421 | if (arm_arch6k && !TARGET_THUMB) | |
1422 | target_thread_pointer = TP_CP15; | |
1423 | else | |
1424 | target_thread_pointer = TP_SOFT; | |
1425 | } | |
1426 | ||
25f905c2 | 1427 | if (TARGET_HARD_TP && TARGET_THUMB1) |
1428 | error ("can not use -mtp=cp15 with 16-bit Thumb"); | |
f655717d | 1429 | |
f9273c43 | 1430 | /* Override the default structure alignment for AAPCS ABI. */ |
6783d878 | 1431 | if (TARGET_AAPCS_BASED) |
f9273c43 | 1432 | arm_structure_size_boundary = 8; |
1433 | ||
6cd47762 | 1434 | if (structure_size_string != NULL) |
1435 | { | |
1436 | int size = strtol (structure_size_string, NULL, 0); | |
f9273c43 | 1437 | |
1438 | if (size == 8 || size == 32 | |
1439 | || (ARM_DOUBLEWORD_ALIGN && size == 64)) | |
6cd47762 | 1440 | arm_structure_size_boundary = size; |
1441 | else | |
c3ceba8e | 1442 | warning (0, "structure size boundary can only be set to %s", |
f9273c43 | 1443 | ARM_DOUBLEWORD_ALIGN ? "8, 32 or 64": "8 or 32"); |
6cd47762 | 1444 | } |
55c1e470 | 1445 | |
bac7fc85 | 1446 | if (!TARGET_ARM && TARGET_VXWORKS_RTP && flag_pic) |
1447 | { | |
1448 | error ("RTP PIC is incompatible with Thumb"); | |
1449 | flag_pic = 0; | |
1450 | } | |
1451 | ||
1ad38120 | 1452 | /* If stack checking is disabled, we can use r10 as the PIC register, |
1453 | which keeps r9 available. The EABI specifies r9 as the PIC register. */ | |
1454 | if (flag_pic && TARGET_SINGLE_PIC_BASE) | |
bac7fc85 | 1455 | { |
1456 | if (TARGET_VXWORKS_RTP) | |
1457 | warning (0, "RTP PIC is incompatible with -msingle-pic-base"); | |
1458 | arm_pic_register = (TARGET_APCS_STACK || TARGET_AAPCS_BASED) ? 9 : 10; | |
1459 | } | |
1460 | ||
1461 | if (flag_pic && TARGET_VXWORKS_RTP) | |
1462 | arm_pic_register = 9; | |
1ad38120 | 1463 | |
55c1e470 | 1464 | if (arm_pic_register_string != NULL) |
1465 | { | |
df9f5cf8 | 1466 | int pic_register = decode_reg_name (arm_pic_register_string); |
9f468b67 | 1467 | |
215b30b3 | 1468 | if (!flag_pic) |
c3ceba8e | 1469 | warning (0, "-mpic-register= is useless without -fpic"); |
55c1e470 | 1470 | |
55c1e470 | 1471 | /* Prevent the user from choosing an obviously stupid PIC register. */ |
df9f5cf8 | 1472 | else if (pic_register < 0 || call_used_regs[pic_register] |
1473 | || pic_register == HARD_FRAME_POINTER_REGNUM | |
1474 | || pic_register == STACK_POINTER_REGNUM | |
bac7fc85 | 1475 | || pic_register >= PC_REGNUM |
1476 | || (TARGET_VXWORKS_RTP | |
1477 | && (unsigned int) pic_register != arm_pic_register)) | |
68435912 | 1478 | error ("unable to use '%s' for PIC register", arm_pic_register_string); |
55c1e470 | 1479 | else |
1480 | arm_pic_register = pic_register; | |
1481 | } | |
cffb2a26 | 1482 | |
25f905c2 | 1483 | /* ??? We might want scheduling for thumb2. */ |
cffb2a26 | 1484 | if (TARGET_THUMB && flag_schedule_insns) |
1485 | { | |
1486 | /* Don't warn since it's on by default in -O2. */ | |
1487 | flag_schedule_insns = 0; | |
1488 | } | |
1489 | ||
9888ad6d | 1490 | if (optimize_size) |
dde5cb68 | 1491 | { |
7e28d849 | 1492 | arm_constant_limit = 1; |
dde5cb68 | 1493 | |
1494 | /* If optimizing for size, bump the number of instructions that we | |
674a8f0b | 1495 | are prepared to conditionally execute (even on a StrongARM). */ |
dde5cb68 | 1496 | max_insns_skipped = 6; |
1497 | } | |
1498 | else | |
1499 | { | |
1500 | /* For processors with load scheduling, it never costs more than | |
1501 | 2 cycles to load a constant, and the load scheduler may well | |
1502 | reduce that to 1. */ | |
359a6e9f | 1503 | if (arm_ld_sched) |
dde5cb68 | 1504 | arm_constant_limit = 1; |
1505 | ||
1506 | /* On XScale the longer latency of a load makes it more difficult | |
1507 | to achieve a good schedule, so it's faster to synthesize | |
674a8f0b | 1508 | constants that can be done in two insns. */ |
dde5cb68 | 1509 | if (arm_tune_xscale) |
1510 | arm_constant_limit = 2; | |
1511 | ||
1512 | /* StrongARM has early execution of branches, so a sequence | |
1513 | that is worth skipping is shorter. */ | |
74a71f7d | 1514 | if (arm_tune_strongarm) |
dde5cb68 | 1515 | max_insns_skipped = 3; |
1516 | } | |
db80f614 | 1517 | |
1518 | /* Register global variables with the garbage collector. */ | |
1519 | arm_add_gc_roots (); | |
1520 | } | |
1521 | ||
1522 | static void | |
ebd88a36 | 1523 | arm_add_gc_roots (void) |
db80f614 | 1524 | { |
a220ccce | 1525 | gcc_obstack_init(&minipool_obstack); |
1526 | minipool_startobj = (char *) obstack_alloc (&minipool_obstack, 0); | |
64266206 | 1527 | } |
8cd42699 | 1528 | \f |
e27ad2d5 | 1529 | /* A table of known ARM exception types. |
1530 | For use with the interrupt function attribute. */ | |
1531 | ||
1532 | typedef struct | |
1533 | { | |
e99c3a1d | 1534 | const char *const arg; |
1535 | const unsigned long return_value; | |
e27ad2d5 | 1536 | } |
1537 | isr_attribute_arg; | |
1538 | ||
e99c3a1d | 1539 | static const isr_attribute_arg isr_attribute_args [] = |
e27ad2d5 | 1540 | { |
1541 | { "IRQ", ARM_FT_ISR }, | |
1542 | { "irq", ARM_FT_ISR }, | |
1543 | { "FIQ", ARM_FT_FIQ }, | |
1544 | { "fiq", ARM_FT_FIQ }, | |
1545 | { "ABORT", ARM_FT_ISR }, | |
1546 | { "abort", ARM_FT_ISR }, | |
1547 | { "ABORT", ARM_FT_ISR }, | |
1548 | { "abort", ARM_FT_ISR }, | |
1549 | { "UNDEF", ARM_FT_EXCEPTION }, | |
1550 | { "undef", ARM_FT_EXCEPTION }, | |
1551 | { "SWI", ARM_FT_EXCEPTION }, | |
1552 | { "swi", ARM_FT_EXCEPTION }, | |
1553 | { NULL, ARM_FT_NORMAL } | |
1554 | }; | |
1555 | ||
1556 | /* Returns the (interrupt) function type of the current | |
1557 | function, or ARM_FT_UNKNOWN if the type cannot be determined. */ | |
1558 | ||
1559 | static unsigned long | |
ebd88a36 | 1560 | arm_isr_value (tree argument) |
e27ad2d5 | 1561 | { |
e99c3a1d | 1562 | const isr_attribute_arg * ptr; |
a12be13c | 1563 | const char * arg; |
e27ad2d5 | 1564 | |
25f905c2 | 1565 | if (!arm_arch_notm) |
1566 | return ARM_FT_NORMAL | ARM_FT_STACKALIGN; | |
1567 | ||
e27ad2d5 | 1568 | /* No argument - default to IRQ. */ |
1569 | if (argument == NULL_TREE) | |
1570 | return ARM_FT_ISR; | |
1571 | ||
1572 | /* Get the value of the argument. */ | |
1573 | if (TREE_VALUE (argument) == NULL_TREE | |
1574 | || TREE_CODE (TREE_VALUE (argument)) != STRING_CST) | |
1575 | return ARM_FT_UNKNOWN; | |
1576 | ||
1577 | arg = TREE_STRING_POINTER (TREE_VALUE (argument)); | |
1578 | ||
1579 | /* Check it against the list of known arguments. */ | |
755eb2b4 | 1580 | for (ptr = isr_attribute_args; ptr->arg != NULL; ptr++) |
a12be13c | 1581 | if (streq (arg, ptr->arg)) |
1582 | return ptr->return_value; | |
e27ad2d5 | 1583 | |
a2f10574 | 1584 | /* An unrecognized interrupt type. */ |
e27ad2d5 | 1585 | return ARM_FT_UNKNOWN; |
1586 | } | |
1587 | ||
1588 | /* Computes the type of the current function. */ | |
1589 | ||
1590 | static unsigned long | |
ebd88a36 | 1591 | arm_compute_func_type (void) |
e27ad2d5 | 1592 | { |
1593 | unsigned long type = ARM_FT_UNKNOWN; | |
1594 | tree a; | |
1595 | tree attr; | |
9e7454d0 | 1596 | |
ed29c566 | 1597 | gcc_assert (TREE_CODE (current_function_decl) == FUNCTION_DECL); |
e27ad2d5 | 1598 | |
1599 | /* Decide if the current function is volatile. Such functions | |
1600 | never return, and many memory cycles can be saved by not storing | |
1601 | register values that will never be needed again. This optimization | |
1602 | was added to speed up context switching in a kernel application. */ | |
1603 | if (optimize > 0 | |
c4d51de2 | 1604 | && (TREE_NOTHROW (current_function_decl) |
1605 | || !(flag_unwind_tables | |
1606 | || (flag_exceptions && !USING_SJLJ_EXCEPTIONS))) | |
e27ad2d5 | 1607 | && TREE_THIS_VOLATILE (current_function_decl)) |
1608 | type |= ARM_FT_VOLATILE; | |
9e7454d0 | 1609 | |
4ee9c684 | 1610 | if (cfun->static_chain_decl != NULL) |
e27ad2d5 | 1611 | type |= ARM_FT_NESTED; |
1612 | ||
e3c541f0 | 1613 | attr = DECL_ATTRIBUTES (current_function_decl); |
9e7454d0 | 1614 | |
e27ad2d5 | 1615 | a = lookup_attribute ("naked", attr); |
1616 | if (a != NULL_TREE) | |
1617 | type |= ARM_FT_NAKED; | |
1618 | ||
4c44712e | 1619 | a = lookup_attribute ("isr", attr); |
1620 | if (a == NULL_TREE) | |
1621 | a = lookup_attribute ("interrupt", attr); | |
9e7454d0 | 1622 | |
4c44712e | 1623 | if (a == NULL_TREE) |
1624 | type |= TARGET_INTERWORK ? ARM_FT_INTERWORKED : ARM_FT_NORMAL; | |
e27ad2d5 | 1625 | else |
4c44712e | 1626 | type |= arm_isr_value (TREE_VALUE (a)); |
9e7454d0 | 1627 | |
e27ad2d5 | 1628 | return type; |
1629 | } | |
1630 | ||
1631 | /* Returns the type of the current function. */ | |
1632 | ||
1633 | unsigned long | |
ebd88a36 | 1634 | arm_current_func_type (void) |
e27ad2d5 | 1635 | { |
1636 | if (ARM_FUNC_TYPE (cfun->machine->func_type) == ARM_FT_UNKNOWN) | |
1637 | cfun->machine->func_type = arm_compute_func_type (); | |
1638 | ||
1639 | return cfun->machine->func_type; | |
1640 | } | |
1641 | \f | |
9e7454d0 | 1642 | /* Return 1 if it is possible to return using a single instruction. |
ffc9d00c | 1643 | If SIBLING is non-null, this is a test for a return before a sibling |
1644 | call. SIBLING is the call insn, so we can examine its register usage. */ | |
e27ad2d5 | 1645 | |
9c08d1fa | 1646 | int |
ffc9d00c | 1647 | use_return_insn (int iscond, rtx sibling) |
9c08d1fa | 1648 | { |
1649 | int regno; | |
5cf3595a | 1650 | unsigned int func_type; |
50368af8 | 1651 | unsigned long saved_int_regs; |
ffc9d00c | 1652 | unsigned HOST_WIDE_INT stack_adjust; |
f9273c43 | 1653 | arm_stack_offsets *offsets; |
9c08d1fa | 1654 | |
cffb2a26 | 1655 | /* Never use a return instruction before reload has run. */ |
e27ad2d5 | 1656 | if (!reload_completed) |
1657 | return 0; | |
01c5c234 | 1658 | |
5cf3595a | 1659 | func_type = arm_current_func_type (); |
1660 | ||
25f905c2 | 1661 | /* Naked, volatile and stack alignment functions need special |
4a5f62e2 | 1662 | consideration. */ |
25f905c2 | 1663 | if (func_type & (ARM_FT_VOLATILE | ARM_FT_NAKED | ARM_FT_STACKALIGN)) |
e27ad2d5 | 1664 | return 0; |
000c5098 | 1665 | |
8b438b4f | 1666 | /* So do interrupt functions that use the frame pointer and Thumb |
1667 | interrupt functions. */ | |
1668 | if (IS_INTERRUPT (func_type) && (frame_pointer_needed || TARGET_THUMB)) | |
000c5098 | 1669 | return 0; |
ffc9d00c | 1670 | |
f9273c43 | 1671 | offsets = arm_get_frame_offsets (); |
1672 | stack_adjust = offsets->outgoing_args - offsets->saved_regs; | |
ffc9d00c | 1673 | |
e27ad2d5 | 1674 | /* As do variadic functions. */ |
1675 | if (current_function_pretend_args_size | |
71904594 | 1676 | || cfun->machine->uses_anonymous_args |
8d69f2ad | 1677 | /* Or if the function calls __builtin_eh_return () */ |
4c44712e | 1678 | || current_function_calls_eh_return |
8d69f2ad | 1679 | /* Or if the function calls alloca */ |
1680 | || current_function_calls_alloca | |
ffc9d00c | 1681 | /* Or if there is a stack adjustment. However, if the stack pointer |
1682 | is saved on the stack, we can use a pre-incrementing stack load. */ | |
1683 | || !(stack_adjust == 0 || (frame_pointer_needed && stack_adjust == 4))) | |
9c08d1fa | 1684 | return 0; |
1685 | ||
50368af8 | 1686 | saved_int_regs = arm_compute_save_reg_mask (); |
1687 | ||
ffc9d00c | 1688 | /* Unfortunately, the insn |
1689 | ||
1690 | ldmib sp, {..., sp, ...} | |
1691 | ||
1692 | triggers a bug on most SA-110 based devices, such that the stack | |
1693 | pointer won't be correctly restored if the instruction takes a | |
efee20da | 1694 | page fault. We work around this problem by popping r3 along with |
ffc9d00c | 1695 | the other registers, since that is never slower than executing |
9e7454d0 | 1696 | another instruction. |
ffc9d00c | 1697 | |
1698 | We test for !arm_arch5 here, because code for any architecture | |
1699 | less than this could potentially be run on one of the buggy | |
1700 | chips. */ | |
25f905c2 | 1701 | if (stack_adjust == 4 && !arm_arch5 && TARGET_ARM) |
ffc9d00c | 1702 | { |
1703 | /* Validate that r3 is a call-clobbered register (always true in | |
674a8f0b | 1704 | the default abi) ... */ |
ffc9d00c | 1705 | if (!call_used_regs[3]) |
1706 | return 0; | |
1707 | ||
0858d94e | 1708 | /* ... that it isn't being used for a return value ... */ |
1709 | if (arm_size_return_regs () >= (4 * UNITS_PER_WORD)) | |
1710 | return 0; | |
1711 | ||
1712 | /* ... or for a tail-call argument ... */ | |
ffc9d00c | 1713 | if (sibling) |
1714 | { | |
ed29c566 | 1715 | gcc_assert (GET_CODE (sibling) == CALL_INSN); |
ffc9d00c | 1716 | |
1717 | if (find_regno_fusage (sibling, USE, 3)) | |
1718 | return 0; | |
1719 | } | |
1720 | ||
1721 | /* ... and that there are no call-saved registers in r0-r2 | |
1722 | (always true in the default ABI). */ | |
1723 | if (saved_int_regs & 0x7) | |
1724 | return 0; | |
1725 | } | |
1726 | ||
f21249d5 | 1727 | /* Can't be done if interworking with Thumb, and any registers have been |
50368af8 | 1728 | stacked. */ |
8b438b4f | 1729 | if (TARGET_INTERWORK && saved_int_regs != 0 && !IS_INTERRUPT(func_type)) |
129a2fe4 | 1730 | return 0; |
50368af8 | 1731 | |
1732 | /* On StrongARM, conditional returns are expensive if they aren't | |
1733 | taken and multiple registers have been stacked. */ | |
74a71f7d | 1734 | if (iscond && arm_tune_strongarm) |
0eb4b5ff | 1735 | { |
9e7454d0 | 1736 | /* Conditional return when just the LR is stored is a simple |
50368af8 | 1737 | conditional-load instruction, that's not expensive. */ |
1738 | if (saved_int_regs != 0 && saved_int_regs != (1 << LR_REGNUM)) | |
1739 | return 0; | |
0eb4b5ff | 1740 | |
2cb7d577 | 1741 | if (flag_pic |
1742 | && arm_pic_register != INVALID_REGNUM | |
3072d30e | 1743 | && df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM)) |
f21249d5 | 1744 | return 0; |
0eb4b5ff | 1745 | } |
50368af8 | 1746 | |
1747 | /* If there are saved registers but the LR isn't saved, then we need | |
1748 | two instructions for the return. */ | |
1749 | if (saved_int_regs && !(saved_int_regs & (1 << LR_REGNUM))) | |
1750 | return 0; | |
1751 | ||
e8ff69a4 | 1752 | /* Can't be done if any of the FPA regs are pushed, |
e27ad2d5 | 1753 | since this also requires an insn. */ |
a2cd141b | 1754 | if (TARGET_HARD_FLOAT && TARGET_FPA) |
1755 | for (regno = FIRST_FPA_REGNUM; regno <= LAST_FPA_REGNUM; regno++) | |
3072d30e | 1756 | if (df_regs_ever_live_p (regno) && !call_used_regs[regno]) |
a2cd141b | 1757 | return 0; |
1758 | ||
1759 | /* Likewise VFP regs. */ | |
1760 | if (TARGET_HARD_FLOAT && TARGET_VFP) | |
1761 | for (regno = FIRST_VFP_REGNUM; regno <= LAST_VFP_REGNUM; regno++) | |
3072d30e | 1762 | if (df_regs_ever_live_p (regno) && !call_used_regs[regno]) |
cffb2a26 | 1763 | return 0; |
9c08d1fa | 1764 | |
755eb2b4 | 1765 | if (TARGET_REALLY_IWMMXT) |
1766 | for (regno = FIRST_IWMMXT_REGNUM; regno <= LAST_IWMMXT_REGNUM; regno++) | |
3072d30e | 1767 | if (df_regs_ever_live_p (regno) && ! call_used_regs[regno]) |
755eb2b4 | 1768 | return 0; |
1769 | ||
9c08d1fa | 1770 | return 1; |
1771 | } | |
1772 | ||
8cd42699 | 1773 | /* Return TRUE if int I is a valid immediate ARM constant. */ |
1774 | ||
1775 | int | |
ebd88a36 | 1776 | const_ok_for_arm (HOST_WIDE_INT i) |
8cd42699 | 1777 | { |
eaae9d8c | 1778 | int lowbit; |
a60f1840 | 1779 | |
9e7454d0 | 1780 | /* For machines with >32 bit HOST_WIDE_INT, the bits above bit 31 must |
6019008a | 1781 | be all zero, or all one. */ |
35cea10c | 1782 | if ((i & ~(unsigned HOST_WIDE_INT) 0xffffffff) != 0 |
1783 | && ((i & ~(unsigned HOST_WIDE_INT) 0xffffffff) | |
1784 | != ((~(unsigned HOST_WIDE_INT) 0) | |
1785 | & ~(unsigned HOST_WIDE_INT) 0xffffffff))) | |
6019008a | 1786 | return FALSE; |
9e7454d0 | 1787 | |
eaae9d8c | 1788 | i &= (unsigned HOST_WIDE_INT) 0xffffffff; |
a60f1840 | 1789 | |
eaae9d8c | 1790 | /* Fast return for 0 and small values. We must do this for zero, since |
1791 | the code below can't handle that one case. */ | |
1792 | if ((i & ~(unsigned HOST_WIDE_INT) 0xff) == 0) | |
4b6897e9 | 1793 | return TRUE; |
1794 | ||
25f905c2 | 1795 | /* Get the number of trailing zeros. */ |
1796 | lowbit = ffs((int) i) - 1; | |
1797 | ||
1798 | /* Only even shifts are allowed in ARM mode so round down to the | |
1799 | nearest even number. */ | |
1800 | if (TARGET_ARM) | |
1801 | lowbit &= ~1; | |
eaae9d8c | 1802 | |
1803 | if ((i & ~(((unsigned HOST_WIDE_INT) 0xff) << lowbit)) == 0) | |
1804 | return TRUE; | |
25f905c2 | 1805 | |
1806 | if (TARGET_ARM) | |
1807 | { | |
1808 | /* Allow rotated constants in ARM mode. */ | |
1809 | if (lowbit <= 4 | |
eaae9d8c | 1810 | && ((i & ~0xc000003f) == 0 |
1811 | || (i & ~0xf000000f) == 0 | |
1812 | || (i & ~0xfc000003) == 0)) | |
25f905c2 | 1813 | return TRUE; |
1814 | } | |
1815 | else | |
1816 | { | |
1817 | HOST_WIDE_INT v; | |
1818 | ||
1819 | /* Allow repeated pattern. */ | |
1820 | v = i & 0xff; | |
1821 | v |= v << 16; | |
1822 | if (i == v || i == (v | (v << 8))) | |
1823 | return TRUE; | |
1824 | } | |
8cd42699 | 1825 | |
a690ede2 | 1826 | return FALSE; |
1827 | } | |
8cd42699 | 1828 | |
66aacf41 | 1829 | /* Return true if I is a valid constant for the operation CODE. */ |
016696d4 | 1830 | static int |
ebd88a36 | 1831 | const_ok_for_op (HOST_WIDE_INT i, enum rtx_code code) |
4b6897e9 | 1832 | { |
1833 | if (const_ok_for_arm (i)) | |
1834 | return 1; | |
1835 | ||
1836 | switch (code) | |
1837 | { | |
1838 | case PLUS: | |
1839 | return const_ok_for_arm (ARM_SIGN_EXTEND (-i)); | |
1840 | ||
1841 | case MINUS: /* Should only occur with (MINUS I reg) => rsb */ | |
1842 | case XOR: | |
1843 | case IOR: | |
1844 | return 0; | |
1845 | ||
1846 | case AND: | |
1847 | return const_ok_for_arm (ARM_SIGN_EXTEND (~i)); | |
1848 | ||
1849 | default: | |
ed29c566 | 1850 | gcc_unreachable (); |
4b6897e9 | 1851 | } |
1852 | } | |
1853 | ||
1854 | /* Emit a sequence of insns to handle a large constant. | |
1855 | CODE is the code of the operation required, it can be any of SET, PLUS, | |
1856 | IOR, AND, XOR, MINUS; | |
1857 | MODE is the mode in which the operation is being performed; | |
1858 | VAL is the integer to operate on; | |
1859 | SOURCE is the other operand (a register, or a null-pointer for SET); | |
1860 | SUBTARGETS means it is safe to create scratch registers if that will | |
64266206 | 1861 | either produce a simpler sequence, or we will want to cse the values. |
1862 | Return value is the number of insns emitted. */ | |
4b6897e9 | 1863 | |
25f905c2 | 1864 | /* ??? Tweak this for thumb2. */ |
4b6897e9 | 1865 | int |
96f57e36 | 1866 | arm_split_constant (enum rtx_code code, enum machine_mode mode, rtx insn, |
ebd88a36 | 1867 | HOST_WIDE_INT val, rtx target, rtx source, int subtargets) |
64266206 | 1868 | { |
96f57e36 | 1869 | rtx cond; |
1870 | ||
1871 | if (insn && GET_CODE (PATTERN (insn)) == COND_EXEC) | |
1872 | cond = COND_EXEC_TEST (PATTERN (insn)); | |
1873 | else | |
1874 | cond = NULL_RTX; | |
1875 | ||
64266206 | 1876 | if (subtargets || code == SET |
1877 | || (GET_CODE (target) == REG && GET_CODE (source) == REG | |
1878 | && REGNO (target) != REGNO (source))) | |
1879 | { | |
11de4c7f | 1880 | /* After arm_reorg has been called, we can't fix up expensive |
a2f10574 | 1881 | constants by pushing them into memory so we must synthesize |
11de4c7f | 1882 | them in-line, regardless of the cost. This is only likely to |
1883 | be more costly on chips that have load delay slots and we are | |
1884 | compiling without running the scheduler (so no splitting | |
82ec49ae | 1885 | occurred before the final instruction emission). |
1886 | ||
1887 | Ref: gcc -O1 -mcpu=strongarm gcc.c-torture/compile/980506-2.c | |
82ec49ae | 1888 | */ |
215b30b3 | 1889 | if (!after_arm_reorg |
96f57e36 | 1890 | && !cond |
9e7454d0 | 1891 | && (arm_gen_constant (code, mode, NULL_RTX, val, target, source, |
96f57e36 | 1892 | 1, 0) |
11de4c7f | 1893 | > arm_constant_limit + (code != SET))) |
64266206 | 1894 | { |
1895 | if (code == SET) | |
1896 | { | |
1897 | /* Currently SET is the only monadic value for CODE, all | |
1898 | the rest are diadic. */ | |
0083a3b9 | 1899 | emit_set_insn (target, GEN_INT (val)); |
64266206 | 1900 | return 1; |
1901 | } | |
1902 | else | |
1903 | { | |
1904 | rtx temp = subtargets ? gen_reg_rtx (mode) : target; | |
1905 | ||
0083a3b9 | 1906 | emit_set_insn (temp, GEN_INT (val)); |
64266206 | 1907 | /* For MINUS, the value is subtracted from, since we never |
1908 | have subtraction of a constant. */ | |
1909 | if (code == MINUS) | |
0083a3b9 | 1910 | emit_set_insn (target, gen_rtx_MINUS (mode, temp, source)); |
64266206 | 1911 | else |
0083a3b9 | 1912 | emit_set_insn (target, |
1913 | gen_rtx_fmt_ee (code, mode, source, temp)); | |
64266206 | 1914 | return 2; |
1915 | } | |
1916 | } | |
1917 | } | |
1918 | ||
9e7454d0 | 1919 | return arm_gen_constant (code, mode, cond, val, target, source, subtargets, |
96f57e36 | 1920 | 1); |
64266206 | 1921 | } |
1922 | ||
25f905c2 | 1923 | /* Return the number of ARM instructions required to synthesize the given |
1924 | constant. */ | |
cd8df93e | 1925 | static int |
ebd88a36 | 1926 | count_insns_for_constant (HOST_WIDE_INT remainder, int i) |
cd8df93e | 1927 | { |
1928 | HOST_WIDE_INT temp1; | |
1929 | int num_insns = 0; | |
1930 | do | |
1931 | { | |
1932 | int end; | |
9e7454d0 | 1933 | |
cd8df93e | 1934 | if (i <= 0) |
1935 | i += 32; | |
1936 | if (remainder & (3 << (i - 2))) | |
1937 | { | |
1938 | end = i - 8; | |
1939 | if (end < 0) | |
1940 | end += 32; | |
1941 | temp1 = remainder & ((0x0ff << end) | |
1942 | | ((i < end) ? (0xff >> (32 - end)) : 0)); | |
1943 | remainder &= ~temp1; | |
1944 | num_insns++; | |
1945 | i -= 6; | |
1946 | } | |
1947 | i -= 2; | |
1948 | } while (remainder); | |
1949 | return num_insns; | |
1950 | } | |
1951 | ||
96f57e36 | 1952 | /* Emit an instruction with the indicated PATTERN. If COND is |
1953 | non-NULL, conditionalize the execution of the instruction on COND | |
1954 | being true. */ | |
1955 | ||
1956 | static void | |
1957 | emit_constant_insn (rtx cond, rtx pattern) | |
1958 | { | |
1959 | if (cond) | |
1960 | pattern = gen_rtx_COND_EXEC (VOIDmode, copy_rtx (cond), pattern); | |
1961 | emit_insn (pattern); | |
1962 | } | |
1963 | ||
64266206 | 1964 | /* As above, but extra parameter GENERATE which, if clear, suppresses |
1965 | RTL generation. */ | |
25f905c2 | 1966 | /* ??? This needs more work for thumb2. */ |
a12be13c | 1967 | |
cffb2a26 | 1968 | static int |
96f57e36 | 1969 | arm_gen_constant (enum rtx_code code, enum machine_mode mode, rtx cond, |
ebd88a36 | 1970 | HOST_WIDE_INT val, rtx target, rtx source, int subtargets, |
1971 | int generate) | |
4b6897e9 | 1972 | { |
4b6897e9 | 1973 | int can_invert = 0; |
1974 | int can_negate = 0; | |
1975 | int can_negate_initial = 0; | |
1976 | int can_shift = 0; | |
1977 | int i; | |
1978 | int num_bits_set = 0; | |
1979 | int set_sign_bit_copies = 0; | |
1980 | int clear_sign_bit_copies = 0; | |
1981 | int clear_zero_bit_copies = 0; | |
1982 | int set_zero_bit_copies = 0; | |
1983 | int insns = 0; | |
4b6897e9 | 1984 | unsigned HOST_WIDE_INT temp1, temp2; |
35cea10c | 1985 | unsigned HOST_WIDE_INT remainder = val & 0xffffffff; |
4b6897e9 | 1986 | |
cffb2a26 | 1987 | /* Find out which operations are safe for a given CODE. Also do a quick |
4b6897e9 | 1988 | check for degenerate cases; these can occur when DImode operations |
1989 | are split. */ | |
1990 | switch (code) | |
1991 | { | |
1992 | case SET: | |
1993 | can_invert = 1; | |
1994 | can_shift = 1; | |
1995 | can_negate = 1; | |
1996 | break; | |
1997 | ||
1998 | case PLUS: | |
1999 | can_negate = 1; | |
2000 | can_negate_initial = 1; | |
2001 | break; | |
2002 | ||
2003 | case IOR: | |
35cea10c | 2004 | if (remainder == 0xffffffff) |
4b6897e9 | 2005 | { |
64266206 | 2006 | if (generate) |
96f57e36 | 2007 | emit_constant_insn (cond, |
2008 | gen_rtx_SET (VOIDmode, target, | |
2009 | GEN_INT (ARM_SIGN_EXTEND (val)))); | |
4b6897e9 | 2010 | return 1; |
2011 | } | |
2012 | if (remainder == 0) | |
2013 | { | |
2014 | if (reload_completed && rtx_equal_p (target, source)) | |
2015 | return 0; | |
64266206 | 2016 | if (generate) |
96f57e36 | 2017 | emit_constant_insn (cond, |
2018 | gen_rtx_SET (VOIDmode, target, source)); | |
4b6897e9 | 2019 | return 1; |
2020 | } | |
2021 | break; | |
2022 | ||
2023 | case AND: | |
2024 | if (remainder == 0) | |
2025 | { | |
64266206 | 2026 | if (generate) |
96f57e36 | 2027 | emit_constant_insn (cond, |
2028 | gen_rtx_SET (VOIDmode, target, const0_rtx)); | |
4b6897e9 | 2029 | return 1; |
2030 | } | |
35cea10c | 2031 | if (remainder == 0xffffffff) |
4b6897e9 | 2032 | { |
2033 | if (reload_completed && rtx_equal_p (target, source)) | |
2034 | return 0; | |
64266206 | 2035 | if (generate) |
96f57e36 | 2036 | emit_constant_insn (cond, |
2037 | gen_rtx_SET (VOIDmode, target, source)); | |
4b6897e9 | 2038 | return 1; |
2039 | } | |
2040 | can_invert = 1; | |
2041 | break; | |
2042 | ||
2043 | case XOR: | |
2044 | if (remainder == 0) | |
2045 | { | |
2046 | if (reload_completed && rtx_equal_p (target, source)) | |
2047 | return 0; | |
64266206 | 2048 | if (generate) |
96f57e36 | 2049 | emit_constant_insn (cond, |
2050 | gen_rtx_SET (VOIDmode, target, source)); | |
4b6897e9 | 2051 | return 1; |
2052 | } | |
a60f1840 | 2053 | |
ed29c566 | 2054 | /* We don't know how to handle other cases yet. */ |
2055 | gcc_assert (remainder == 0xffffffff); | |
a60f1840 | 2056 | |
ed29c566 | 2057 | if (generate) |
2058 | emit_constant_insn (cond, | |
2059 | gen_rtx_SET (VOIDmode, target, | |
2060 | gen_rtx_NOT (mode, source))); | |
2061 | return 1; | |
4b6897e9 | 2062 | |
2063 | case MINUS: | |
2064 | /* We treat MINUS as (val - source), since (source - val) is always | |
2065 | passed as (source + (-val)). */ | |
2066 | if (remainder == 0) | |
2067 | { | |
64266206 | 2068 | if (generate) |
96f57e36 | 2069 | emit_constant_insn (cond, |
2070 | gen_rtx_SET (VOIDmode, target, | |
2071 | gen_rtx_NEG (mode, source))); | |
4b6897e9 | 2072 | return 1; |
2073 | } | |
2074 | if (const_ok_for_arm (val)) | |
2075 | { | |
64266206 | 2076 | if (generate) |
96f57e36 | 2077 | emit_constant_insn (cond, |
9e7454d0 | 2078 | gen_rtx_SET (VOIDmode, target, |
96f57e36 | 2079 | gen_rtx_MINUS (mode, GEN_INT (val), |
2080 | source))); | |
4b6897e9 | 2081 | return 1; |
2082 | } | |
2083 | can_negate = 1; | |
2084 | ||
2085 | break; | |
2086 | ||
2087 | default: | |
ed29c566 | 2088 | gcc_unreachable (); |
4b6897e9 | 2089 | } |
2090 | ||
66aacf41 | 2091 | /* If we can do it in one insn get out quickly. */ |
4b6897e9 | 2092 | if (const_ok_for_arm (val) |
2093 | || (can_negate_initial && const_ok_for_arm (-val)) | |
2094 | || (can_invert && const_ok_for_arm (~val))) | |
2095 | { | |
64266206 | 2096 | if (generate) |
96f57e36 | 2097 | emit_constant_insn (cond, |
2098 | gen_rtx_SET (VOIDmode, target, | |
9e7454d0 | 2099 | (source |
96f57e36 | 2100 | ? gen_rtx_fmt_ee (code, mode, source, |
2101 | GEN_INT (val)) | |
2102 | : GEN_INT (val)))); | |
4b6897e9 | 2103 | return 1; |
2104 | } | |
2105 | ||
4b6897e9 | 2106 | /* Calculate a few attributes that may be useful for specific |
66aacf41 | 2107 | optimizations. */ |
4b6897e9 | 2108 | for (i = 31; i >= 0; i--) |
2109 | { | |
2110 | if ((remainder & (1 << i)) == 0) | |
2111 | clear_sign_bit_copies++; | |
2112 | else | |
2113 | break; | |
2114 | } | |
2115 | ||
2116 | for (i = 31; i >= 0; i--) | |
2117 | { | |
2118 | if ((remainder & (1 << i)) != 0) | |
2119 | set_sign_bit_copies++; | |
2120 | else | |
2121 | break; | |
2122 | } | |
2123 | ||
2124 | for (i = 0; i <= 31; i++) | |
2125 | { | |
2126 | if ((remainder & (1 << i)) == 0) | |
2127 | clear_zero_bit_copies++; | |
2128 | else | |
2129 | break; | |
2130 | } | |
2131 | ||
2132 | for (i = 0; i <= 31; i++) | |
2133 | { | |
2134 | if ((remainder & (1 << i)) != 0) | |
2135 | set_zero_bit_copies++; | |
2136 | else | |
2137 | break; | |
2138 | } | |
2139 | ||
2140 | switch (code) | |
2141 | { | |
2142 | case SET: | |
25f905c2 | 2143 | /* See if we can use movw. */ |
2144 | if (arm_arch_thumb2 && (remainder & 0xffff0000) == 0) | |
2145 | { | |
2146 | if (generate) | |
2147 | emit_constant_insn (cond, gen_rtx_SET (VOIDmode, target, | |
2148 | GEN_INT (val))); | |
2149 | return 1; | |
2150 | } | |
2151 | ||
4b6897e9 | 2152 | /* See if we can do this by sign_extending a constant that is known |
2153 | to be negative. This is a good, way of doing it, since the shift | |
2154 | may well merge into a subsequent insn. */ | |
2155 | if (set_sign_bit_copies > 1) | |
2156 | { | |
2157 | if (const_ok_for_arm | |
9e7454d0 | 2158 | (temp1 = ARM_SIGN_EXTEND (remainder |
4b6897e9 | 2159 | << (set_sign_bit_copies - 1)))) |
2160 | { | |
64266206 | 2161 | if (generate) |
2162 | { | |
997e4174 | 2163 | rtx new_src = subtargets ? gen_reg_rtx (mode) : target; |
96f57e36 | 2164 | emit_constant_insn (cond, |
9e7454d0 | 2165 | gen_rtx_SET (VOIDmode, new_src, |
96f57e36 | 2166 | GEN_INT (temp1))); |
2167 | emit_constant_insn (cond, | |
9e7454d0 | 2168 | gen_ashrsi3 (target, new_src, |
96f57e36 | 2169 | GEN_INT (set_sign_bit_copies - 1))); |
64266206 | 2170 | } |
4b6897e9 | 2171 | return 2; |
2172 | } | |
2173 | /* For an inverted constant, we will need to set the low bits, | |
2174 | these will be shifted out of harm's way. */ | |
2175 | temp1 |= (1 << (set_sign_bit_copies - 1)) - 1; | |
2176 | if (const_ok_for_arm (~temp1)) | |
2177 | { | |
64266206 | 2178 | if (generate) |
2179 | { | |
997e4174 | 2180 | rtx new_src = subtargets ? gen_reg_rtx (mode) : target; |
96f57e36 | 2181 | emit_constant_insn (cond, |
2182 | gen_rtx_SET (VOIDmode, new_src, | |
2183 | GEN_INT (temp1))); | |
2184 | emit_constant_insn (cond, | |
9e7454d0 | 2185 | gen_ashrsi3 (target, new_src, |
96f57e36 | 2186 | GEN_INT (set_sign_bit_copies - 1))); |
64266206 | 2187 | } |
4b6897e9 | 2188 | return 2; |
2189 | } | |
2190 | } | |
2191 | ||
d430cffb | 2192 | /* See if we can calculate the value as the difference between two |
2193 | valid immediates. */ | |
2194 | if (clear_sign_bit_copies + clear_zero_bit_copies <= 16) | |
2195 | { | |
2196 | int topshift = clear_sign_bit_copies & ~1; | |
2197 | ||
3b85fd4d | 2198 | temp1 = ARM_SIGN_EXTEND ((remainder + (0x00800000 >> topshift)) |
2199 | & (0xff000000 >> topshift)); | |
d430cffb | 2200 | |
2201 | /* If temp1 is zero, then that means the 9 most significant | |
2202 | bits of remainder were 1 and we've caused it to overflow. | |
2203 | When topshift is 0 we don't need to do anything since we | |
2204 | can borrow from 'bit 32'. */ | |
2205 | if (temp1 == 0 && topshift != 0) | |
2206 | temp1 = 0x80000000 >> (topshift - 1); | |
2207 | ||
3b85fd4d | 2208 | temp2 = ARM_SIGN_EXTEND (temp1 - remainder); |
a60f1840 | 2209 | |
d430cffb | 2210 | if (const_ok_for_arm (temp2)) |
2211 | { | |
2212 | if (generate) | |
2213 | { | |
2214 | rtx new_src = subtargets ? gen_reg_rtx (mode) : target; | |
2215 | emit_constant_insn (cond, | |
2216 | gen_rtx_SET (VOIDmode, new_src, | |
2217 | GEN_INT (temp1))); | |
2218 | emit_constant_insn (cond, | |
2219 | gen_addsi3 (target, new_src, | |
2220 | GEN_INT (-temp2))); | |
2221 | } | |
2222 | ||
2223 | return 2; | |
2224 | } | |
2225 | } | |
2226 | ||
4b6897e9 | 2227 | /* See if we can generate this by setting the bottom (or the top) |
2228 | 16 bits, and then shifting these into the other half of the | |
2229 | word. We only look for the simplest cases, to do more would cost | |
2230 | too much. Be careful, however, not to generate this when the | |
2231 | alternative would take fewer insns. */ | |
35cea10c | 2232 | if (val & 0xffff0000) |
4b6897e9 | 2233 | { |
35cea10c | 2234 | temp1 = remainder & 0xffff0000; |
4b6897e9 | 2235 | temp2 = remainder & 0x0000ffff; |
2236 | ||
66aacf41 | 2237 | /* Overlaps outside this range are best done using other methods. */ |
4b6897e9 | 2238 | for (i = 9; i < 24; i++) |
2239 | { | |
35cea10c | 2240 | if ((((temp2 | (temp2 << i)) & 0xffffffff) == remainder) |
215b30b3 | 2241 | && !const_ok_for_arm (temp2)) |
4b6897e9 | 2242 | { |
997e4174 | 2243 | rtx new_src = (subtargets |
2244 | ? (generate ? gen_reg_rtx (mode) : NULL_RTX) | |
2245 | : target); | |
96f57e36 | 2246 | insns = arm_gen_constant (code, mode, cond, temp2, new_src, |
64266206 | 2247 | source, subtargets, generate); |
4b6897e9 | 2248 | source = new_src; |
64266206 | 2249 | if (generate) |
9e7454d0 | 2250 | emit_constant_insn |
96f57e36 | 2251 | (cond, |
2252 | gen_rtx_SET | |
2253 | (VOIDmode, target, | |
2254 | gen_rtx_IOR (mode, | |
2255 | gen_rtx_ASHIFT (mode, source, | |
2256 | GEN_INT (i)), | |
2257 | source))); | |
4b6897e9 | 2258 | return insns + 1; |
2259 | } | |
2260 | } | |
2261 | ||
66aacf41 | 2262 | /* Don't duplicate cases already considered. */ |
4b6897e9 | 2263 | for (i = 17; i < 24; i++) |
2264 | { | |
2265 | if (((temp1 | (temp1 >> i)) == remainder) | |
215b30b3 | 2266 | && !const_ok_for_arm (temp1)) |
4b6897e9 | 2267 | { |
997e4174 | 2268 | rtx new_src = (subtargets |
2269 | ? (generate ? gen_reg_rtx (mode) : NULL_RTX) | |
2270 | : target); | |
96f57e36 | 2271 | insns = arm_gen_constant (code, mode, cond, temp1, new_src, |
64266206 | 2272 | source, subtargets, generate); |
4b6897e9 | 2273 | source = new_src; |
64266206 | 2274 | if (generate) |
96f57e36 | 2275 | emit_constant_insn |
2276 | (cond, | |
2277 | gen_rtx_SET (VOIDmode, target, | |
76676c8e | 2278 | gen_rtx_IOR |
2279 | (mode, | |
2280 | gen_rtx_LSHIFTRT (mode, source, | |
2281 | GEN_INT (i)), | |
2282 | source))); | |
4b6897e9 | 2283 | return insns + 1; |
2284 | } | |
2285 | } | |
2286 | } | |
2287 | break; | |
2288 | ||
2289 | case IOR: | |
2290 | case XOR: | |
57f96f27 | 2291 | /* If we have IOR or XOR, and the constant can be loaded in a |
2292 | single instruction, and we can find a temporary to put it in, | |
4b6897e9 | 2293 | then this can be done in two instructions instead of 3-4. */ |
2294 | if (subtargets | |
997e4174 | 2295 | /* TARGET can't be NULL if SUBTARGETS is 0 */ |
215b30b3 | 2296 | || (reload_completed && !reg_mentioned_p (target, source))) |
4b6897e9 | 2297 | { |
215b30b3 | 2298 | if (const_ok_for_arm (ARM_SIGN_EXTEND (~val))) |
4b6897e9 | 2299 | { |
64266206 | 2300 | if (generate) |
2301 | { | |
2302 | rtx sub = subtargets ? gen_reg_rtx (mode) : target; | |
4b6897e9 | 2303 | |
96f57e36 | 2304 | emit_constant_insn (cond, |
9e7454d0 | 2305 | gen_rtx_SET (VOIDmode, sub, |
96f57e36 | 2306 | GEN_INT (val))); |
2307 | emit_constant_insn (cond, | |
9e7454d0 | 2308 | gen_rtx_SET (VOIDmode, target, |
96f57e36 | 2309 | gen_rtx_fmt_ee (code, mode, |
2310 | source, sub))); | |
64266206 | 2311 | } |
4b6897e9 | 2312 | return 2; |
2313 | } | |
2314 | } | |
2315 | ||
2316 | if (code == XOR) | |
2317 | break; | |
2318 | ||
2319 | if (set_sign_bit_copies > 8 | |
2320 | && (val & (-1 << (32 - set_sign_bit_copies))) == val) | |
2321 | { | |
64266206 | 2322 | if (generate) |
2323 | { | |
2324 | rtx sub = subtargets ? gen_reg_rtx (mode) : target; | |
2325 | rtx shift = GEN_INT (set_sign_bit_copies); | |
2326 | ||
9e7454d0 | 2327 | emit_constant_insn |
96f57e36 | 2328 | (cond, |
2329 | gen_rtx_SET (VOIDmode, sub, | |
9e7454d0 | 2330 | gen_rtx_NOT (mode, |
96f57e36 | 2331 | gen_rtx_ASHIFT (mode, |
9e7454d0 | 2332 | source, |
96f57e36 | 2333 | shift)))); |
9e7454d0 | 2334 | emit_constant_insn |
96f57e36 | 2335 | (cond, |
2336 | gen_rtx_SET (VOIDmode, target, | |
2337 | gen_rtx_NOT (mode, | |
2338 | gen_rtx_LSHIFTRT (mode, sub, | |
2339 | shift)))); | |
64266206 | 2340 | } |
4b6897e9 | 2341 | return 2; |
2342 | } | |
2343 | ||
2344 | if (set_zero_bit_copies > 8 | |
2345 | && (remainder & ((1 << set_zero_bit_copies) - 1)) == remainder) | |
2346 | { | |
64266206 | 2347 | if (generate) |
2348 | { | |
2349 | rtx sub = subtargets ? gen_reg_rtx (mode) : target; | |
2350 | rtx shift = GEN_INT (set_zero_bit_copies); | |
2351 | ||
96f57e36 | 2352 | emit_constant_insn |
2353 | (cond, | |
2354 | gen_rtx_SET (VOIDmode, sub, | |
2355 | gen_rtx_NOT (mode, | |
2356 | gen_rtx_LSHIFTRT (mode, | |
2357 | source, | |
2358 | shift)))); | |
9e7454d0 | 2359 | emit_constant_insn |
96f57e36 | 2360 | (cond, |
2361 | gen_rtx_SET (VOIDmode, target, | |
2362 | gen_rtx_NOT (mode, | |
2363 | gen_rtx_ASHIFT (mode, sub, | |
2364 | shift)))); | |
64266206 | 2365 | } |
4b6897e9 | 2366 | return 2; |
2367 | } | |
2368 | ||
215b30b3 | 2369 | if (const_ok_for_arm (temp1 = ARM_SIGN_EXTEND (~val))) |
4b6897e9 | 2370 | { |
64266206 | 2371 | if (generate) |
2372 | { | |
2373 | rtx sub = subtargets ? gen_reg_rtx (mode) : target; | |
96f57e36 | 2374 | emit_constant_insn (cond, |
2375 | gen_rtx_SET (VOIDmode, sub, | |
2376 | gen_rtx_NOT (mode, source))); | |
64266206 | 2377 | source = sub; |
2378 | if (subtargets) | |
2379 | sub = gen_reg_rtx (mode); | |
96f57e36 | 2380 | emit_constant_insn (cond, |
2381 | gen_rtx_SET (VOIDmode, sub, | |
9e7454d0 | 2382 | gen_rtx_AND (mode, source, |
96f57e36 | 2383 | GEN_INT (temp1)))); |
2384 | emit_constant_insn (cond, | |
2385 | gen_rtx_SET (VOIDmode, target, | |
2386 | gen_rtx_NOT (mode, sub))); | |
64266206 | 2387 | } |
4b6897e9 | 2388 | return 3; |
2389 | } | |
2390 | break; | |
2391 | ||
2392 | case AND: | |
2393 | /* See if two shifts will do 2 or more insn's worth of work. */ | |
2394 | if (clear_sign_bit_copies >= 16 && clear_sign_bit_copies < 24) | |
2395 | { | |
35cea10c | 2396 | HOST_WIDE_INT shift_mask = ((0xffffffff |
4b6897e9 | 2397 | << (32 - clear_sign_bit_copies)) |
35cea10c | 2398 | & 0xffffffff); |
4b6897e9 | 2399 | |
35cea10c | 2400 | if ((remainder | shift_mask) != 0xffffffff) |
4b6897e9 | 2401 | { |
64266206 | 2402 | if (generate) |
2403 | { | |
997e4174 | 2404 | rtx new_src = subtargets ? gen_reg_rtx (mode) : target; |
9e7454d0 | 2405 | insns = arm_gen_constant (AND, mode, cond, |
96f57e36 | 2406 | remainder | shift_mask, |
997e4174 | 2407 | new_src, source, subtargets, 1); |
2408 | source = new_src; | |
64266206 | 2409 | } |
2410 | else | |
997e4174 | 2411 | { |
2412 | rtx targ = subtargets ? NULL_RTX : target; | |
96f57e36 | 2413 | insns = arm_gen_constant (AND, mode, cond, |
2414 | remainder | shift_mask, | |
997e4174 | 2415 | targ, source, subtargets, 0); |
2416 | } | |
64266206 | 2417 | } |
2418 | ||
2419 | if (generate) | |
2420 | { | |
997e4174 | 2421 | rtx new_src = subtargets ? gen_reg_rtx (mode) : target; |
2422 | rtx shift = GEN_INT (clear_sign_bit_copies); | |
2423 | ||
2424 | emit_insn (gen_ashlsi3 (new_src, source, shift)); | |
2425 | emit_insn (gen_lshrsi3 (target, new_src, shift)); | |
4b6897e9 | 2426 | } |
2427 | ||
4b6897e9 | 2428 | return insns + 2; |
2429 | } | |
2430 | ||
2431 | if (clear_zero_bit_copies >= 16 && clear_zero_bit_copies < 24) | |
2432 | { | |
2433 | HOST_WIDE_INT shift_mask = (1 << clear_zero_bit_copies) - 1; | |
9e7454d0 | 2434 | |
35cea10c | 2435 | if ((remainder | shift_mask) != 0xffffffff) |
4b6897e9 | 2436 | { |
64266206 | 2437 | if (generate) |
2438 | { | |
997e4174 | 2439 | rtx new_src = subtargets ? gen_reg_rtx (mode) : target; |
2440 | ||
96f57e36 | 2441 | insns = arm_gen_constant (AND, mode, cond, |
2442 | remainder | shift_mask, | |
997e4174 | 2443 | new_src, source, subtargets, 1); |
2444 | source = new_src; | |
64266206 | 2445 | } |
2446 | else | |
997e4174 | 2447 | { |
2448 | rtx targ = subtargets ? NULL_RTX : target; | |
2449 | ||
96f57e36 | 2450 | insns = arm_gen_constant (AND, mode, cond, |
2451 | remainder | shift_mask, | |
997e4174 | 2452 | targ, source, subtargets, 0); |
2453 | } | |
64266206 | 2454 | } |
2455 | ||
2456 | if (generate) | |
2457 | { | |
997e4174 | 2458 | rtx new_src = subtargets ? gen_reg_rtx (mode) : target; |
2459 | rtx shift = GEN_INT (clear_zero_bit_copies); | |
2460 | ||
2461 | emit_insn (gen_lshrsi3 (new_src, source, shift)); | |
2462 | emit_insn (gen_ashlsi3 (target, new_src, shift)); | |
4b6897e9 | 2463 | } |
2464 | ||
4b6897e9 | 2465 | return insns + 2; |
2466 | } | |
2467 | ||
2468 | break; | |
2469 | ||
2470 | default: | |
2471 | break; | |
2472 | } | |
2473 | ||
2474 | for (i = 0; i < 32; i++) | |
2475 | if (remainder & (1 << i)) | |
2476 | num_bits_set++; | |
2477 | ||
2478 | if (code == AND || (can_invert && num_bits_set > 16)) | |
35cea10c | 2479 | remainder = (~remainder) & 0xffffffff; |
4b6897e9 | 2480 | else if (code == PLUS && num_bits_set > 16) |
35cea10c | 2481 | remainder = (-remainder) & 0xffffffff; |
4b6897e9 | 2482 | else |
2483 | { | |
2484 | can_invert = 0; | |
2485 | can_negate = 0; | |
2486 | } | |
2487 | ||
2488 | /* Now try and find a way of doing the job in either two or three | |
2489 | instructions. | |
2490 | We start by looking for the largest block of zeros that are aligned on | |
2491 | a 2-bit boundary, we then fill up the temps, wrapping around to the | |
2492 | top of the word when we drop off the bottom. | |
25f905c2 | 2493 | In the worst case this code should produce no more than four insns. |
2494 | Thumb-2 constants are shifted, not rotated, so the MSB is always the | |
2495 | best place to start. */ | |
2496 | ||
2497 | /* ??? Use thumb2 replicated constants when the high and low halfwords are | |
2498 | the same. */ | |
4b6897e9 | 2499 | { |
2500 | int best_start = 0; | |
25f905c2 | 2501 | if (!TARGET_THUMB2) |
4b6897e9 | 2502 | { |
25f905c2 | 2503 | int best_consecutive_zeros = 0; |
4b6897e9 | 2504 | |
25f905c2 | 2505 | for (i = 0; i < 32; i += 2) |
4b6897e9 | 2506 | { |
25f905c2 | 2507 | int consecutive_zeros = 0; |
2508 | ||
2509 | if (!(remainder & (3 << i))) | |
4b6897e9 | 2510 | { |
25f905c2 | 2511 | while ((i < 32) && !(remainder & (3 << i))) |
2512 | { | |
2513 | consecutive_zeros += 2; | |
2514 | i += 2; | |
2515 | } | |
2516 | if (consecutive_zeros > best_consecutive_zeros) | |
2517 | { | |
2518 | best_consecutive_zeros = consecutive_zeros; | |
2519 | best_start = i - consecutive_zeros; | |
2520 | } | |
2521 | i -= 2; | |
4b6897e9 | 2522 | } |
4b6897e9 | 2523 | } |
cd8df93e | 2524 | |
25f905c2 | 2525 | /* So long as it won't require any more insns to do so, it's |
2526 | desirable to emit a small constant (in bits 0...9) in the last | |
2527 | insn. This way there is more chance that it can be combined with | |
2528 | a later addressing insn to form a pre-indexed load or store | |
2529 | operation. Consider: | |
2530 | ||
2531 | *((volatile int *)0xe0000100) = 1; | |
2532 | *((volatile int *)0xe0000110) = 2; | |
2533 | ||
2534 | We want this to wind up as: | |
2535 | ||
2536 | mov rA, #0xe0000000 | |
2537 | mov rB, #1 | |
2538 | str rB, [rA, #0x100] | |
2539 | mov rB, #2 | |
2540 | str rB, [rA, #0x110] | |
2541 | ||
2542 | rather than having to synthesize both large constants from scratch. | |
2543 | ||
2544 | Therefore, we calculate how many insns would be required to emit | |
2545 | the constant starting from `best_start', and also starting from | |
2546 | zero (i.e. with bit 31 first to be output). If `best_start' doesn't | |
2547 | yield a shorter sequence, we may as well use zero. */ | |
2548 | if (best_start != 0 | |
2549 | && ((((unsigned HOST_WIDE_INT) 1) << best_start) < remainder) | |
2550 | && (count_insns_for_constant (remainder, 0) <= | |
2551 | count_insns_for_constant (remainder, best_start))) | |
2552 | best_start = 0; | |
2553 | } | |
cd8df93e | 2554 | |
2555 | /* Now start emitting the insns. */ | |
4b6897e9 | 2556 | i = best_start; |
2557 | do | |
2558 | { | |
2559 | int end; | |
2560 | ||
2561 | if (i <= 0) | |
2562 | i += 32; | |
2563 | if (remainder & (3 << (i - 2))) | |
2564 | { | |
2565 | end = i - 8; | |
2566 | if (end < 0) | |
2567 | end += 32; | |
2568 | temp1 = remainder & ((0x0ff << end) | |
2569 | | ((i < end) ? (0xff >> (32 - end)) : 0)); | |
2570 | remainder &= ~temp1; | |
2571 | ||
997e4174 | 2572 | if (generate) |
4b6897e9 | 2573 | { |
6382d316 | 2574 | rtx new_src, temp1_rtx; |
2575 | ||
2576 | if (code == SET || code == MINUS) | |
2577 | { | |
2578 | new_src = (subtargets ? gen_reg_rtx (mode) : target); | |
b6557085 | 2579 | if (can_invert && code != MINUS) |
6382d316 | 2580 | temp1 = ~temp1; |
2581 | } | |
2582 | else | |
2583 | { | |
b6557085 | 2584 | if (remainder && subtargets) |
6382d316 | 2585 | new_src = gen_reg_rtx (mode); |
b6557085 | 2586 | else |
2587 | new_src = target; | |
6382d316 | 2588 | if (can_invert) |
2589 | temp1 = ~temp1; | |
2590 | else if (can_negate) | |
2591 | temp1 = -temp1; | |
2592 | } | |
2593 | ||
2594 | temp1 = trunc_int_for_mode (temp1, mode); | |
2595 | temp1_rtx = GEN_INT (temp1); | |
997e4174 | 2596 | |
2597 | if (code == SET) | |
6382d316 | 2598 | ; |
997e4174 | 2599 | else if (code == MINUS) |
6382d316 | 2600 | temp1_rtx = gen_rtx_MINUS (mode, temp1_rtx, source); |
997e4174 | 2601 | else |
6382d316 | 2602 | temp1_rtx = gen_rtx_fmt_ee (code, mode, source, temp1_rtx); |
2603 | ||
96f57e36 | 2604 | emit_constant_insn (cond, |
9e7454d0 | 2605 | gen_rtx_SET (VOIDmode, new_src, |
96f57e36 | 2606 | temp1_rtx)); |
997e4174 | 2607 | source = new_src; |
4b6897e9 | 2608 | } |
2609 | ||
997e4174 | 2610 | if (code == SET) |
2611 | { | |
2612 | can_invert = 0; | |
2613 | code = PLUS; | |
2614 | } | |
2615 | else if (code == MINUS) | |
2616 | code = PLUS; | |
2617 | ||
4b6897e9 | 2618 | insns++; |
25f905c2 | 2619 | if (TARGET_ARM) |
2620 | i -= 6; | |
2621 | else | |
2622 | i -= 7; | |
4b6897e9 | 2623 | } |
542d5028 | 2624 | /* Arm allows rotates by a multiple of two. Thumb-2 allows arbitrary |
25f905c2 | 2625 | shifts. */ |
2626 | if (TARGET_ARM) | |
2627 | i -= 2; | |
2628 | else | |
2629 | i--; | |
a12be13c | 2630 | } |
2631 | while (remainder); | |
4b6897e9 | 2632 | } |
a12be13c | 2633 | |
4b6897e9 | 2634 | return insns; |
2635 | } | |
2636 | ||
9480e6c7 | 2637 | /* Canonicalize a comparison so that we are more likely to recognize it. |
2638 | This can be done for a few constant compares, where we can make the | |
2639 | immediate value easier to load. */ | |
a12be13c | 2640 | |
9480e6c7 | 2641 | enum rtx_code |
aa3c2d40 | 2642 | arm_canonicalize_comparison (enum rtx_code code, enum machine_mode mode, |
2643 | rtx * op1) | |
9480e6c7 | 2644 | { |
ecd3fcc7 | 2645 | unsigned HOST_WIDE_INT i = INTVAL (*op1); |
aa3c2d40 | 2646 | unsigned HOST_WIDE_INT maxval; |
2647 | maxval = (((unsigned HOST_WIDE_INT) 1) << (GET_MODE_BITSIZE(mode) - 1)) - 1; | |
9480e6c7 | 2648 | |
2649 | switch (code) | |
2650 | { | |
2651 | case EQ: | |
2652 | case NE: | |
2653 | return code; | |
2654 | ||
2655 | case GT: | |
2656 | case LE: | |
aa3c2d40 | 2657 | if (i != maxval |
215b30b3 | 2658 | && (const_ok_for_arm (i + 1) || const_ok_for_arm (-(i + 1)))) |
9480e6c7 | 2659 | { |
215b30b3 | 2660 | *op1 = GEN_INT (i + 1); |
9480e6c7 | 2661 | return code == GT ? GE : LT; |
2662 | } | |
2663 | break; | |
2664 | ||
2665 | case GE: | |
2666 | case LT: | |
aa3c2d40 | 2667 | if (i != ~maxval |
215b30b3 | 2668 | && (const_ok_for_arm (i - 1) || const_ok_for_arm (-(i - 1)))) |
9480e6c7 | 2669 | { |
215b30b3 | 2670 | *op1 = GEN_INT (i - 1); |
9480e6c7 | 2671 | return code == GE ? GT : LE; |
2672 | } | |
2673 | break; | |
2674 | ||
2675 | case GTU: | |
2676 | case LEU: | |
35cea10c | 2677 | if (i != ~((unsigned HOST_WIDE_INT) 0) |
215b30b3 | 2678 | && (const_ok_for_arm (i + 1) || const_ok_for_arm (-(i + 1)))) |
9480e6c7 | 2679 | { |
2680 | *op1 = GEN_INT (i + 1); | |
2681 | return code == GTU ? GEU : LTU; | |
2682 | } | |
2683 | break; | |
2684 | ||
2685 | case GEU: | |
2686 | case LTU: | |
2687 | if (i != 0 | |
215b30b3 | 2688 | && (const_ok_for_arm (i - 1) || const_ok_for_arm (-(i - 1)))) |
9480e6c7 | 2689 | { |
2690 | *op1 = GEN_INT (i - 1); | |
2691 | return code == GEU ? GTU : LEU; | |
2692 | } | |
2693 | break; | |
2694 | ||
2695 | default: | |
ed29c566 | 2696 | gcc_unreachable (); |
9480e6c7 | 2697 | } |
2698 | ||
2699 | return code; | |
2700 | } | |
9480e6c7 | 2701 | |
19347327 | 2702 | |
2703 | /* Define how to find the value returned by a function. */ | |
2704 | ||
ccd90aaa | 2705 | rtx |
2706 | arm_function_value(tree type, tree func ATTRIBUTE_UNUSED) | |
19347327 | 2707 | { |
2708 | enum machine_mode mode; | |
2709 | int unsignedp ATTRIBUTE_UNUSED; | |
2710 | rtx r ATTRIBUTE_UNUSED; | |
2711 | ||
19347327 | 2712 | mode = TYPE_MODE (type); |
2713 | /* Promote integer types. */ | |
2714 | if (INTEGRAL_TYPE_P (type)) | |
2715 | PROMOTE_FUNCTION_MODE (mode, unsignedp, type); | |
46b5d878 | 2716 | |
2717 | /* Promotes small structs returned in a register to full-word size | |
2718 | for big-endian AAPCS. */ | |
2719 | if (arm_return_in_msb (type)) | |
2720 | { | |
2721 | HOST_WIDE_INT size = int_size_in_bytes (type); | |
2722 | if (size % UNITS_PER_WORD != 0) | |
2723 | { | |
2724 | size += UNITS_PER_WORD - size % UNITS_PER_WORD; | |
2725 | mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0); | |
2726 | } | |
2727 | } | |
a60f1840 | 2728 | |
19347327 | 2729 | return LIBCALL_VALUE(mode); |
2730 | } | |
2731 | ||
a60f1840 | 2732 | /* Determine the amount of memory needed to store the possible return |
ccd90aaa | 2733 | registers of an untyped call. */ |
2734 | int | |
2735 | arm_apply_result_size (void) | |
2736 | { | |
2737 | int size = 16; | |
2738 | ||
2739 | if (TARGET_ARM) | |
2740 | { | |
2741 | if (TARGET_HARD_FLOAT_ABI) | |
2742 | { | |
2743 | if (TARGET_FPA) | |
2744 | size += 12; | |
2745 | if (TARGET_MAVERICK) | |
2746 | size += 8; | |
2747 | } | |
2748 | if (TARGET_IWMMXT_ABI) | |
2749 | size += 8; | |
2750 | } | |
2751 | ||
2752 | return size; | |
2753 | } | |
19347327 | 2754 | |
9888ad6d | 2755 | /* Decide whether a type should be returned in memory (true) |
2756 | or in a register (false). This is called by the macro | |
2757 | RETURN_IN_MEMORY. */ | |
64266206 | 2758 | int |
ebd88a36 | 2759 | arm_return_in_memory (tree type) |
64266206 | 2760 | { |
1e00c159 | 2761 | HOST_WIDE_INT size; |
2762 | ||
d98a3884 | 2763 | size = int_size_in_bytes (type); |
2764 | ||
2765 | /* Vector values should be returned using ARM registers, not memory (unless | |
2766 | they're over 16 bytes, which will break since we only have four | |
2767 | call-clobbered registers to play with). */ | |
2768 | if (TREE_CODE (type) == VECTOR_TYPE) | |
2769 | return (size < 0 || size > (4 * UNITS_PER_WORD)); | |
2770 | ||
1a3ba46f | 2771 | if (!AGGREGATE_TYPE_P (type) && |
2772 | !(TARGET_AAPCS_BASED && TREE_CODE (type) == COMPLEX_TYPE)) | |
2773 | /* All simple types are returned in registers. | |
2774 | For AAPCS, complex types are treated the same as aggregates. */ | |
29a6fe51 | 2775 | return 0; |
1e00c159 | 2776 | |
f9273c43 | 2777 | if (arm_abi != ARM_ABI_APCS) |
1e00c159 | 2778 | { |
f9273c43 | 2779 | /* ATPCS and later return aggregate types in memory only if they are |
1e00c159 | 2780 | larger than a word (or are variable size). */ |
2781 | return (size < 0 || size > UNITS_PER_WORD); | |
2782 | } | |
9e7454d0 | 2783 | |
b89db9a7 | 2784 | /* For the arm-wince targets we choose to be compatible with Microsoft's |
cffb2a26 | 2785 | ARM and Thumb compilers, which always return aggregates in memory. */ |
2786 | #ifndef ARM_WINCE | |
b7e2e8e6 | 2787 | /* All structures/unions bigger than one word are returned in memory. |
2788 | Also catch the case where int_size_in_bytes returns -1. In this case | |
b89db9a7 | 2789 | the aggregate is either huge or of variable size, and in either case |
b7e2e8e6 | 2790 | we will want to return it via memory and not in a register. */ |
1e00c159 | 2791 | if (size < 0 || size > UNITS_PER_WORD) |
29a6fe51 | 2792 | return 1; |
9e7454d0 | 2793 | |
29a6fe51 | 2794 | if (TREE_CODE (type) == RECORD_TYPE) |
64266206 | 2795 | { |
2796 | tree field; | |
2797 | ||
3accc59a | 2798 | /* For a struct the APCS says that we only return in a register |
2799 | if the type is 'integer like' and every addressable element | |
2800 | has an offset of zero. For practical purposes this means | |
2801 | that the structure can have at most one non bit-field element | |
2802 | and that this element must be the first one in the structure. */ | |
9e7454d0 | 2803 | |
9888ad6d | 2804 | /* Find the first field, ignoring non FIELD_DECL things which will |
2805 | have been created by C++. */ | |
2806 | for (field = TYPE_FIELDS (type); | |
2807 | field && TREE_CODE (field) != FIELD_DECL; | |
2808 | field = TREE_CHAIN (field)) | |
2809 | continue; | |
9e7454d0 | 2810 | |
9888ad6d | 2811 | if (field == NULL) |
f5daaea4 | 2812 | return 0; /* An empty structure. Allowed by an extension to ANSI C. */ |
9888ad6d | 2813 | |
cffb2a26 | 2814 | /* Check that the first field is valid for returning in a register. */ |
2815 | ||
2816 | /* ... Floats are not allowed */ | |
f5daaea4 | 2817 | if (FLOAT_TYPE_P (TREE_TYPE (field))) |
3accc59a | 2818 | return 1; |
2819 | ||
cffb2a26 | 2820 | /* ... Aggregates that are not themselves valid for returning in |
2821 | a register are not allowed. */ | |
f5daaea4 | 2822 | if (RETURN_IN_MEMORY (TREE_TYPE (field))) |
3accc59a | 2823 | return 1; |
89b30d75 | 2824 | |
3accc59a | 2825 | /* Now check the remaining fields, if any. Only bitfields are allowed, |
2826 | since they are not addressable. */ | |
9888ad6d | 2827 | for (field = TREE_CHAIN (field); |
2828 | field; | |
2829 | field = TREE_CHAIN (field)) | |
2830 | { | |
2831 | if (TREE_CODE (field) != FIELD_DECL) | |
2832 | continue; | |
9e7454d0 | 2833 | |
215b30b3 | 2834 | if (!DECL_BIT_FIELD_TYPE (field)) |
9888ad6d | 2835 | return 1; |
2836 | } | |
64266206 | 2837 | |
2838 | return 0; | |
2839 | } | |
9e7454d0 | 2840 | |
29a6fe51 | 2841 | if (TREE_CODE (type) == UNION_TYPE) |
64266206 | 2842 | { |
2843 | tree field; | |
2844 | ||
2845 | /* Unions can be returned in registers if every element is | |
2846 | integral, or can be returned in an integer register. */ | |
9888ad6d | 2847 | for (field = TYPE_FIELDS (type); |
2848 | field; | |
2849 | field = TREE_CHAIN (field)) | |
64266206 | 2850 | { |
9888ad6d | 2851 | if (TREE_CODE (field) != FIELD_DECL) |
2852 | continue; | |
2853 | ||
7bd8ccc9 | 2854 | if (FLOAT_TYPE_P (TREE_TYPE (field))) |
2855 | return 1; | |
9e7454d0 | 2856 | |
9888ad6d | 2857 | if (RETURN_IN_MEMORY (TREE_TYPE (field))) |
64266206 | 2858 | return 1; |
2859 | } | |
9e7454d0 | 2860 | |
64266206 | 2861 | return 0; |
2862 | } | |
9e7454d0 | 2863 | #endif /* not ARM_WINCE */ |
2864 | ||
cffb2a26 | 2865 | /* Return all other types in memory. */ |
64266206 | 2866 | return 1; |
2867 | } | |
2868 | ||
674a8f0b | 2869 | /* Indicate whether or not words of a double are in big-endian order. */ |
3d94d3fa | 2870 | |
2871 | int | |
ebd88a36 | 2872 | arm_float_words_big_endian (void) |
3d94d3fa | 2873 | { |
a2cd141b | 2874 | if (TARGET_MAVERICK) |
7d57ec45 | 2875 | return 0; |
3d94d3fa | 2876 | |
2877 | /* For FPA, float words are always big-endian. For VFP, floats words | |
2878 | follow the memory system mode. */ | |
2879 | ||
a2cd141b | 2880 | if (TARGET_FPA) |
3d94d3fa | 2881 | { |
3d94d3fa | 2882 | return 1; |
2883 | } | |
2884 | ||
2885 | if (TARGET_VFP) | |
2886 | return (TARGET_BIG_END ? 1 : 0); | |
2887 | ||
2888 | return 1; | |
2889 | } | |
2890 | ||
a52eca60 | 2891 | /* Initialize a variable CUM of type CUMULATIVE_ARGS |
2892 | for a call to a function whose data type is FNTYPE. | |
2893 | For a library call, FNTYPE is NULL. */ | |
2894 | void | |
9e7454d0 | 2895 | arm_init_cumulative_args (CUMULATIVE_ARGS *pcum, tree fntype, |
ebd88a36 | 2896 | rtx libname ATTRIBUTE_UNUSED, |
2897 | tree fndecl ATTRIBUTE_UNUSED) | |
a52eca60 | 2898 | { |
2899 | /* On the ARM, the offset starts at 0. */ | |
f2be8f2b | 2900 | pcum->nregs = 0; |
755eb2b4 | 2901 | pcum->iwmmxt_nregs = 0; |
f9273c43 | 2902 | pcum->can_split = true; |
9e7454d0 | 2903 | |
755eb2b4 | 2904 | /* Varargs vectors are treated the same as long long. |
2905 | named_count avoids having to change the way arm handles 'named' */ | |
2906 | pcum->named_count = 0; | |
2907 | pcum->nargs = 0; | |
2908 | ||
2909 | if (TARGET_REALLY_IWMMXT && fntype) | |
2910 | { | |
2911 | tree fn_arg; | |
2912 | ||
2913 | for (fn_arg = TYPE_ARG_TYPES (fntype); | |
2914 | fn_arg; | |
2915 | fn_arg = TREE_CHAIN (fn_arg)) | |
2916 | pcum->named_count += 1; | |
2917 | ||
2918 | if (! pcum->named_count) | |
2919 | pcum->named_count = INT_MAX; | |
2920 | } | |
a52eca60 | 2921 | } |
2922 | ||
f9273c43 | 2923 | |
2924 | /* Return true if mode/type need doubleword alignment. */ | |
2925 | bool | |
2926 | arm_needs_doubleword_align (enum machine_mode mode, tree type) | |
2927 | { | |
9c291064 | 2928 | return (GET_MODE_ALIGNMENT (mode) > PARM_BOUNDARY |
2929 | || (type && TYPE_ALIGN (type) > PARM_BOUNDARY)); | |
f9273c43 | 2930 | } |
2931 | ||
2932 | ||
a52eca60 | 2933 | /* Determine where to put an argument to a function. |
2934 | Value is zero to push the argument on the stack, | |
2935 | or a hard register in which to store the argument. | |
2936 | ||
2937 | MODE is the argument's machine mode. | |
2938 | TYPE is the data type of the argument (as a tree). | |
2939 | This is null for libcalls where that information may | |
2940 | not be available. | |
2941 | CUM is a variable of type CUMULATIVE_ARGS which gives info about | |
2942 | the preceding args and about the function being called. | |
2943 | NAMED is nonzero if this argument is a named parameter | |
2944 | (otherwise it is an extra parameter matching an ellipsis). */ | |
a12be13c | 2945 | |
a52eca60 | 2946 | rtx |
ebd88a36 | 2947 | arm_function_arg (CUMULATIVE_ARGS *pcum, enum machine_mode mode, |
f9273c43 | 2948 | tree type, int named) |
a52eca60 | 2949 | { |
f9273c43 | 2950 | int nregs; |
2951 | ||
2952 | /* Varargs vectors are treated the same as long long. | |
2953 | named_count avoids having to change the way arm handles 'named' */ | |
2954 | if (TARGET_IWMMXT_ABI | |
9e7454d0 | 2955 | && arm_vector_mode_supported_p (mode) |
f9273c43 | 2956 | && pcum->named_count > pcum->nargs + 1) |
755eb2b4 | 2957 | { |
f9273c43 | 2958 | if (pcum->iwmmxt_nregs <= 9) |
2959 | return gen_rtx_REG (mode, pcum->iwmmxt_nregs + FIRST_IWMMXT_REGNUM); | |
2960 | else | |
755eb2b4 | 2961 | { |
f9273c43 | 2962 | pcum->can_split = false; |
2963 | return NULL_RTX; | |
755eb2b4 | 2964 | } |
755eb2b4 | 2965 | } |
2966 | ||
f9273c43 | 2967 | /* Put doubleword aligned quantities in even register pairs. */ |
2968 | if (pcum->nregs & 1 | |
2969 | && ARM_DOUBLEWORD_ALIGN | |
2970 | && arm_needs_doubleword_align (mode, type)) | |
2971 | pcum->nregs++; | |
2972 | ||
a52eca60 | 2973 | if (mode == VOIDmode) |
de55252a | 2974 | /* Pick an arbitrary value for operand 2 of the call insn. */ |
2975 | return const0_rtx; | |
f9273c43 | 2976 | |
af6ed417 | 2977 | /* Only allow splitting an arg between regs and memory if all preceding |
f9273c43 | 2978 | args were allocated to regs. For args passed by reference we only count |
2979 | the reference pointer. */ | |
2980 | if (pcum->can_split) | |
2981 | nregs = 1; | |
2982 | else | |
2983 | nregs = ARM_NUM_REGS2 (mode, type); | |
2984 | ||
2985 | if (!named || pcum->nregs + nregs > NUM_ARG_REGS) | |
a52eca60 | 2986 | return NULL_RTX; |
9e7454d0 | 2987 | |
a52eca60 | 2988 | return gen_rtx_REG (mode, pcum->nregs); |
2989 | } | |
78aea6a2 | 2990 | |
f054eb3c | 2991 | static int |
2992 | arm_arg_partial_bytes (CUMULATIVE_ARGS *pcum, enum machine_mode mode, | |
2993 | tree type, bool named ATTRIBUTE_UNUSED) | |
2994 | { | |
2995 | int nregs = pcum->nregs; | |
2996 | ||
d98a3884 | 2997 | if (TARGET_IWMMXT_ABI && arm_vector_mode_supported_p (mode)) |
f054eb3c | 2998 | return 0; |
2999 | ||
3000 | if (NUM_ARG_REGS > nregs | |
3001 | && (NUM_ARG_REGS < nregs + ARM_NUM_REGS2 (mode, type)) | |
3002 | && pcum->can_split) | |
3003 | return (NUM_ARG_REGS - nregs) * UNITS_PER_WORD; | |
3004 | ||
3005 | return 0; | |
3006 | } | |
3007 | ||
78aea6a2 | 3008 | /* Variable sized types are passed by reference. This is a GCC |
3009 | extension to the ARM ABI. */ | |
3010 | ||
b981d932 | 3011 | static bool |
3012 | arm_pass_by_reference (CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED, | |
3013 | enum machine_mode mode ATTRIBUTE_UNUSED, | |
3014 | tree type, bool named ATTRIBUTE_UNUSED) | |
78aea6a2 | 3015 | { |
3016 | return type && TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST; | |
3017 | } | |
a52eca60 | 3018 | \f |
78fe751b | 3019 | /* Encode the current state of the #pragma [no_]long_calls. */ |
3020 | typedef enum | |
a52eca60 | 3021 | { |
554f2707 | 3022 | OFF, /* No #pragma [no_]long_calls is in effect. */ |
78fe751b | 3023 | LONG, /* #pragma long_calls is in effect. */ |
3024 | SHORT /* #pragma no_long_calls is in effect. */ | |
3025 | } arm_pragma_enum; | |
a52eca60 | 3026 | |
78fe751b | 3027 | static arm_pragma_enum arm_pragma_long_calls = OFF; |
a52eca60 | 3028 | |
1fcd08b1 | 3029 | void |
ebd88a36 | 3030 | arm_pr_long_calls (struct cpp_reader * pfile ATTRIBUTE_UNUSED) |
a52eca60 | 3031 | { |
1fcd08b1 | 3032 | arm_pragma_long_calls = LONG; |
3033 | } | |
3034 | ||
3035 | void | |
ebd88a36 | 3036 | arm_pr_no_long_calls (struct cpp_reader * pfile ATTRIBUTE_UNUSED) |
1fcd08b1 | 3037 | { |
3038 | arm_pragma_long_calls = SHORT; | |
3039 | } | |
3040 | ||
3041 | void | |
ebd88a36 | 3042 | arm_pr_long_calls_off (struct cpp_reader * pfile ATTRIBUTE_UNUSED) |
1fcd08b1 | 3043 | { |
3044 | arm_pragma_long_calls = OFF; | |
a52eca60 | 3045 | } |
3046 | \f | |
e3c541f0 | 3047 | /* Table of machine attributes. */ |
3048 | const struct attribute_spec arm_attribute_table[] = | |
a52eca60 | 3049 | { |
e3c541f0 | 3050 | /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */ |
a52eca60 | 3051 | /* Function calls made to this symbol must be done indirectly, because |
3052 | it may lie outside of the 26 bit addressing range of a normal function | |
3053 | call. */ | |
e3c541f0 | 3054 | { "long_call", 0, 0, false, true, true, NULL }, |
a52eca60 | 3055 | /* Whereas these functions are always known to reside within the 26 bit |
3056 | addressing range. */ | |
e3c541f0 | 3057 | { "short_call", 0, 0, false, true, true, NULL }, |
9e7454d0 | 3058 | /* Interrupt Service Routines have special prologue and epilogue requirements. */ |
e3c541f0 | 3059 | { "isr", 0, 1, false, false, false, arm_handle_isr_attribute }, |
3060 | { "interrupt", 0, 1, false, false, false, arm_handle_isr_attribute }, | |
3061 | { "naked", 0, 0, true, false, false, arm_handle_fndecl_attribute }, | |
3062 | #ifdef ARM_PE | |
3063 | /* ARM/PE has three new attributes: | |
3064 | interfacearm - ? | |
3065 | dllexport - for exporting a function/variable that will live in a dll | |
3066 | dllimport - for importing a function/variable from a dll | |
3067 | ||
3068 | Microsoft allows multiple declspecs in one __declspec, separating | |
3069 | them with spaces. We do NOT support this. Instead, use __declspec | |
3070 | multiple times. | |
3071 | */ | |
3072 | { "dllimport", 0, 0, true, false, false, NULL }, | |
3073 | { "dllexport", 0, 0, true, false, false, NULL }, | |
3074 | { "interfacearm", 0, 0, true, false, false, arm_handle_fndecl_attribute }, | |
3aa0c315 | 3075 | #elif TARGET_DLLIMPORT_DECL_ATTRIBUTES |
3076 | { "dllimport", 0, 0, false, false, false, handle_dll_attribute }, | |
3077 | { "dllexport", 0, 0, false, false, false, handle_dll_attribute }, | |
a79b1d93 | 3078 | { "notshared", 0, 0, false, true, false, arm_handle_notshared_attribute }, |
e3c541f0 | 3079 | #endif |
3080 | { NULL, 0, 0, false, false, false, NULL } | |
3081 | }; | |
e27ad2d5 | 3082 | |
e3c541f0 | 3083 | /* Handle an attribute requiring a FUNCTION_DECL; |
3084 | arguments as in struct attribute_spec.handler. */ | |
3085 | static tree | |
ebd88a36 | 3086 | arm_handle_fndecl_attribute (tree *node, tree name, tree args ATTRIBUTE_UNUSED, |
3087 | int flags ATTRIBUTE_UNUSED, bool *no_add_attrs) | |
e3c541f0 | 3088 | { |
3089 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
3090 | { | |
9b2d6d13 | 3091 | warning (OPT_Wattributes, "%qs attribute only applies to functions", |
e3c541f0 | 3092 | IDENTIFIER_POINTER (name)); |
3093 | *no_add_attrs = true; | |
3094 | } | |
3095 | ||
3096 | return NULL_TREE; | |
3097 | } | |
3098 | ||
3099 | /* Handle an "interrupt" or "isr" attribute; | |
3100 | arguments as in struct attribute_spec.handler. */ | |
3101 | static tree | |
ebd88a36 | 3102 | arm_handle_isr_attribute (tree *node, tree name, tree args, int flags, |
3103 | bool *no_add_attrs) | |
e3c541f0 | 3104 | { |
3105 | if (DECL_P (*node)) | |
3106 | { | |
3107 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
3108 | { | |
9b2d6d13 | 3109 | warning (OPT_Wattributes, "%qs attribute only applies to functions", |
e3c541f0 | 3110 | IDENTIFIER_POINTER (name)); |
3111 | *no_add_attrs = true; | |
3112 | } | |
3113 | /* FIXME: the argument if any is checked for type attributes; | |
3114 | should it be checked for decl ones? */ | |
3115 | } | |
3116 | else | |
3117 | { | |
3118 | if (TREE_CODE (*node) == FUNCTION_TYPE | |
3119 | || TREE_CODE (*node) == METHOD_TYPE) | |
3120 | { | |
3121 | if (arm_isr_value (args) == ARM_FT_UNKNOWN) | |
3122 | { | |
9b2d6d13 | 3123 | warning (OPT_Wattributes, "%qs attribute ignored", |
3124 | IDENTIFIER_POINTER (name)); | |
e3c541f0 | 3125 | *no_add_attrs = true; |
3126 | } | |
3127 | } | |
3128 | else if (TREE_CODE (*node) == POINTER_TYPE | |
3129 | && (TREE_CODE (TREE_TYPE (*node)) == FUNCTION_TYPE | |
3130 | || TREE_CODE (TREE_TYPE (*node)) == METHOD_TYPE) | |
3131 | && arm_isr_value (args) != ARM_FT_UNKNOWN) | |
3132 | { | |
e086912e | 3133 | *node = build_variant_type_copy (*node); |
a12be13c | 3134 | TREE_TYPE (*node) = build_type_attribute_variant |
3135 | (TREE_TYPE (*node), | |
3136 | tree_cons (name, args, TYPE_ATTRIBUTES (TREE_TYPE (*node)))); | |
e3c541f0 | 3137 | *no_add_attrs = true; |
3138 | } | |
3139 | else | |
3140 | { | |
3141 | /* Possibly pass this attribute on from the type to a decl. */ | |
3142 | if (flags & ((int) ATTR_FLAG_DECL_NEXT | |
3143 | | (int) ATTR_FLAG_FUNCTION_NEXT | |
3144 | | (int) ATTR_FLAG_ARRAY_NEXT)) | |
3145 | { | |
3146 | *no_add_attrs = true; | |
3147 | return tree_cons (name, args, NULL_TREE); | |
3148 | } | |
3149 | else | |
3150 | { | |
9b2d6d13 | 3151 | warning (OPT_Wattributes, "%qs attribute ignored", |
3152 | IDENTIFIER_POINTER (name)); | |
e3c541f0 | 3153 | } |
3154 | } | |
3155 | } | |
3156 | ||
3157 | return NULL_TREE; | |
a52eca60 | 3158 | } |
3159 | ||
56826b58 | 3160 | #if TARGET_DLLIMPORT_DECL_ATTRIBUTES |
a79b1d93 | 3161 | /* Handle the "notshared" attribute. This attribute is another way of |
3162 | requesting hidden visibility. ARM's compiler supports | |
3163 | "__declspec(notshared)"; we support the same thing via an | |
3164 | attribute. */ | |
3165 | ||
3166 | static tree | |
a60f1840 | 3167 | arm_handle_notshared_attribute (tree *node, |
3168 | tree name ATTRIBUTE_UNUSED, | |
3169 | tree args ATTRIBUTE_UNUSED, | |
3170 | int flags ATTRIBUTE_UNUSED, | |
a79b1d93 | 3171 | bool *no_add_attrs) |
3172 | { | |
3173 | tree decl = TYPE_NAME (*node); | |
3174 | ||
3175 | if (decl) | |
3176 | { | |
3177 | DECL_VISIBILITY (decl) = VISIBILITY_HIDDEN; | |
3178 | DECL_VISIBILITY_SPECIFIED (decl) = 1; | |
3179 | *no_add_attrs = false; | |
3180 | } | |
3181 | return NULL_TREE; | |
3182 | } | |
56826b58 | 3183 | #endif |
a79b1d93 | 3184 | |
a52eca60 | 3185 | /* Return 0 if the attributes for two types are incompatible, 1 if they |
3186 | are compatible, and 2 if they are nearly compatible (which causes a | |
3187 | warning to be generated). */ | |
f32321da | 3188 | static int |
ebd88a36 | 3189 | arm_comp_type_attributes (tree type1, tree type2) |
a52eca60 | 3190 | { |
d6c334ee | 3191 | int l1, l2, s1, s2; |
9e7454d0 | 3192 | |
a52eca60 | 3193 | /* Check for mismatch of non-default calling convention. */ |
3194 | if (TREE_CODE (type1) != FUNCTION_TYPE) | |
3195 | return 1; | |
3196 | ||
3197 | /* Check for mismatched call attributes. */ | |
d6c334ee | 3198 | l1 = lookup_attribute ("long_call", TYPE_ATTRIBUTES (type1)) != NULL; |
3199 | l2 = lookup_attribute ("long_call", TYPE_ATTRIBUTES (type2)) != NULL; | |
3200 | s1 = lookup_attribute ("short_call", TYPE_ATTRIBUTES (type1)) != NULL; | |
3201 | s2 = lookup_attribute ("short_call", TYPE_ATTRIBUTES (type2)) != NULL; | |
f7387eb0 | 3202 | |
3203 | /* Only bother to check if an attribute is defined. */ | |
3204 | if (l1 | l2 | s1 | s2) | |
3205 | { | |
3206 | /* If one type has an attribute, the other must have the same attribute. */ | |
d6c334ee | 3207 | if ((l1 != l2) || (s1 != s2)) |
f7387eb0 | 3208 | return 0; |
a52eca60 | 3209 | |
f7387eb0 | 3210 | /* Disallow mixed attributes. */ |
3211 | if ((l1 & s2) || (l2 & s1)) | |
3212 | return 0; | |
3213 | } | |
9e7454d0 | 3214 | |
e27ad2d5 | 3215 | /* Check for mismatched ISR attribute. */ |
3216 | l1 = lookup_attribute ("isr", TYPE_ATTRIBUTES (type1)) != NULL; | |
3217 | if (! l1) | |
3218 | l1 = lookup_attribute ("interrupt", TYPE_ATTRIBUTES (type1)) != NULL; | |
3219 | l2 = lookup_attribute ("isr", TYPE_ATTRIBUTES (type2)) != NULL; | |
3220 | if (! l2) | |
3221 | l1 = lookup_attribute ("interrupt", TYPE_ATTRIBUTES (type2)) != NULL; | |
3222 | if (l1 != l2) | |
3223 | return 0; | |
3224 | ||
f7387eb0 | 3225 | return 1; |
a52eca60 | 3226 | } |
3227 | ||
78fe751b | 3228 | /* Assigns default attributes to newly defined type. This is used to |
3229 | set short_call/long_call attributes for function types of | |
3230 | functions defined inside corresponding #pragma scopes. */ | |
f32321da | 3231 | static void |
ebd88a36 | 3232 | arm_set_default_type_attributes (tree type) |
78fe751b | 3233 | { |
3234 | /* Add __attribute__ ((long_call)) to all functions, when | |
3235 | inside #pragma long_calls or __attribute__ ((short_call)), | |
3236 | when inside #pragma no_long_calls. */ | |
3237 | if (TREE_CODE (type) == FUNCTION_TYPE || TREE_CODE (type) == METHOD_TYPE) | |
3238 | { | |
3239 | tree type_attr_list, attr_name; | |
3240 | type_attr_list = TYPE_ATTRIBUTES (type); | |
3241 | ||
3242 | if (arm_pragma_long_calls == LONG) | |
3243 | attr_name = get_identifier ("long_call"); | |
3244 | else if (arm_pragma_long_calls == SHORT) | |
3245 | attr_name = get_identifier ("short_call"); | |
3246 | else | |
3247 | return; | |
3248 | ||
3249 | type_attr_list = tree_cons (attr_name, NULL_TREE, type_attr_list); | |
3250 | TYPE_ATTRIBUTES (type) = type_attr_list; | |
3251 | } | |
3252 | } | |
3253 | \f | |
de55252a | 3254 | /* Return true if DECL is known to be linked into section SECTION. */ |
3255 | ||
3256 | static bool | |
3257 | arm_function_in_section_p (tree decl, section *section) | |
78fe751b | 3258 | { |
de55252a | 3259 | /* We can only be certain about functions defined in the same |
3260 | compilation unit. */ | |
3261 | if (!TREE_STATIC (decl)) | |
3262 | return false; | |
78fe751b | 3263 | |
de55252a | 3264 | /* Make sure that SYMBOL always binds to the definition in this |
3265 | compilation unit. */ | |
3266 | if (!targetm.binds_local_p (decl)) | |
3267 | return false; | |
78fe751b | 3268 | |
de55252a | 3269 | /* If DECL_SECTION_NAME is set, assume it is trustworthy. */ |
3270 | if (!DECL_SECTION_NAME (decl)) | |
3271 | { | |
3272 | /* Only cater for unit-at-a-time mode, where we know that the user | |
3273 | cannot later specify a section for DECL. */ | |
3274 | if (!flag_unit_at_a_time) | |
3275 | return false; | |
3276 | ||
3277 | /* Make sure that we will not create a unique section for DECL. */ | |
3278 | if (flag_function_sections || DECL_ONE_ONLY (decl)) | |
3279 | return false; | |
3280 | } | |
3281 | ||
3282 | return function_section (decl) == section; | |
78fe751b | 3283 | } |
3284 | ||
85c36fd1 | 3285 | /* Return nonzero if a 32-bit "long_call" should be generated for |
de55252a | 3286 | a call from the current function to DECL. We generate a long_call |
3287 | if the function: | |
78fe751b | 3288 | |
3289 | a. has an __attribute__((long call)) | |
3290 | or b. is within the scope of a #pragma long_calls | |
3291 | or c. the -mlong-calls command line switch has been specified | |
3292 | ||
3293 | However we do not generate a long call if the function: | |
9e7454d0 | 3294 | |
78fe751b | 3295 | d. has an __attribute__ ((short_call)) |
3296 | or e. is inside the scope of a #pragma no_long_calls | |
de55252a | 3297 | or f. is defined in the same section as the current function. */ |
78fe751b | 3298 | |
de55252a | 3299 | bool |
3300 | arm_is_long_call_p (tree decl) | |
3301 | { | |
3302 | tree attrs; | |
78fe751b | 3303 | |
de55252a | 3304 | if (!decl) |
3305 | return TARGET_LONG_CALLS; | |
78fe751b | 3306 | |
de55252a | 3307 | attrs = TYPE_ATTRIBUTES (TREE_TYPE (decl)); |
3308 | if (lookup_attribute ("short_call", attrs)) | |
3309 | return false; | |
78fe751b | 3310 | |
de55252a | 3311 | /* For "f", be conservative, and only cater for cases in which the |
3312 | whole of the current function is placed in the same section. */ | |
3313 | if (!flag_reorder_blocks_and_partition | |
3314 | && arm_function_in_section_p (decl, current_function_section ())) | |
3315 | return false; | |
a2f6cf84 | 3316 | |
de55252a | 3317 | if (lookup_attribute ("long_call", attrs)) |
3318 | return true; | |
9e7454d0 | 3319 | |
de55252a | 3320 | return TARGET_LONG_CALLS; |
78fe751b | 3321 | } |
9e0922c4 | 3322 | |
e3e08e7f | 3323 | /* Return nonzero if it is ok to make a tail-call to DECL. */ |
805e22b2 | 3324 | static bool |
ebd88a36 | 3325 | arm_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED) |
9e0922c4 | 3326 | { |
25f905c2 | 3327 | unsigned long func_type; |
9e0922c4 | 3328 | |
755eb2b4 | 3329 | if (cfun->machine->sibcall_blocked) |
3330 | return false; | |
3331 | ||
9e0922c4 | 3332 | /* Never tailcall something for which we have no decl, or if we |
3333 | are in Thumb mode. */ | |
3334 | if (decl == NULL || TARGET_THUMB) | |
805e22b2 | 3335 | return false; |
9e0922c4 | 3336 | |
bac7fc85 | 3337 | /* The PIC register is live on entry to VxWorks PLT entries, so we |
3338 | must make the call before restoring the PIC register. */ | |
3339 | if (TARGET_VXWORKS_RTP && flag_pic && !targetm.binds_local_p (decl)) | |
3340 | return false; | |
3341 | ||
9e0922c4 | 3342 | /* Cannot tail-call to long calls, since these are out of range of |
de55252a | 3343 | a branch instruction. */ |
3344 | if (arm_is_long_call_p (decl)) | |
805e22b2 | 3345 | return false; |
9e0922c4 | 3346 | |
3347 | /* If we are interworking and the function is not declared static | |
9e7454d0 | 3348 | then we can't tail-call it unless we know that it exists in this |
9e0922c4 | 3349 | compilation unit (since it might be a Thumb routine). */ |
215b30b3 | 3350 | if (TARGET_INTERWORK && TREE_PUBLIC (decl) && !TREE_ASM_WRITTEN (decl)) |
805e22b2 | 3351 | return false; |
9e0922c4 | 3352 | |
25f905c2 | 3353 | func_type = arm_current_func_type (); |
e27ad2d5 | 3354 | /* Never tailcall from an ISR routine - it needs a special exit sequence. */ |
25f905c2 | 3355 | if (IS_INTERRUPT (func_type)) |
3356 | return false; | |
3357 | ||
3358 | /* Never tailcall if function may be called with a misaligned SP. */ | |
3359 | if (IS_STACKALIGN (func_type)) | |
805e22b2 | 3360 | return false; |
e27ad2d5 | 3361 | |
9e0922c4 | 3362 | /* Everything else is ok. */ |
805e22b2 | 3363 | return true; |
9e0922c4 | 3364 | } |
3365 | ||
a52eca60 | 3366 | \f |
d32b3557 | 3367 | /* Addressing mode support functions. */ |
3368 | ||
bf02170c | 3369 | /* Return nonzero if X is a legitimate immediate operand when compiling |
2cb7d577 | 3370 | for PIC. We know that X satisfies CONSTANT_P and flag_pic is true. */ |
69120d26 | 3371 | int |
ebd88a36 | 3372 | legitimate_pic_operand_p (rtx x) |
69120d26 | 3373 | { |
2cb7d577 | 3374 | if (GET_CODE (x) == SYMBOL_REF |
3375 | || (GET_CODE (x) == CONST | |
3376 | && GET_CODE (XEXP (x, 0)) == PLUS | |
3377 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF)) | |
69120d26 | 3378 | return 0; |
3379 | ||
3380 | return 1; | |
3381 | } | |
3382 | ||
bac7fc85 | 3383 | /* Record that the current function needs a PIC register. Initialize |
3384 | cfun->machine->pic_reg if we have not already done so. */ | |
3385 | ||
3386 | static void | |
3387 | require_pic_register (void) | |
3388 | { | |
3389 | /* A lot of the logic here is made obscure by the fact that this | |
3390 | routine gets called as part of the rtx cost estimation process. | |
3391 | We don't want those calls to affect any assumptions about the real | |
3392 | function; and further, we can't call entry_of_function() until we | |
3393 | start the real expansion process. */ | |
3394 | if (!current_function_uses_pic_offset_table) | |
3395 | { | |
e1ba4a27 | 3396 | gcc_assert (can_create_pseudo_p ()); |
bac7fc85 | 3397 | if (arm_pic_register != INVALID_REGNUM) |
3398 | { | |
3399 | cfun->machine->pic_reg = gen_rtx_REG (Pmode, arm_pic_register); | |
3400 | ||
3401 | /* Play games to avoid marking the function as needing pic | |
3402 | if we are being called as part of the cost-estimation | |
3403 | process. */ | |
3404 | if (current_ir_type () != IR_GIMPLE) | |
3405 | current_function_uses_pic_offset_table = 1; | |
3406 | } | |
3407 | else | |
3408 | { | |
3409 | rtx seq; | |
3410 | ||
3411 | cfun->machine->pic_reg = gen_reg_rtx (Pmode); | |
3412 | ||
3413 | /* Play games to avoid marking the function as needing pic | |
3414 | if we are being called as part of the cost-estimation | |
3415 | process. */ | |
3416 | if (current_ir_type () != IR_GIMPLE) | |
3417 | { | |
3418 | current_function_uses_pic_offset_table = 1; | |
3419 | start_sequence (); | |
3420 | ||
3421 | arm_load_pic_register (0UL); | |
3422 | ||
3423 | seq = get_insns (); | |
3424 | end_sequence (); | |
3425 | emit_insn_after (seq, entry_of_function ()); | |
3426 | } | |
3427 | } | |
3428 | } | |
3429 | } | |
3430 | ||
69120d26 | 3431 | rtx |
ebd88a36 | 3432 | legitimize_pic_address (rtx orig, enum machine_mode mode, rtx reg) |
69120d26 | 3433 | { |
1644aab7 | 3434 | if (GET_CODE (orig) == SYMBOL_REF |
3435 | || GET_CODE (orig) == LABEL_REF) | |
69120d26 | 3436 | { |
29122112 | 3437 | #ifndef AOF_ASSEMBLER |
69120d26 | 3438 | rtx pic_ref, address; |
29122112 | 3439 | #endif |
69120d26 | 3440 | rtx insn; |
3441 | int subregs = 0; | |
3442 | ||
bac7fc85 | 3443 | /* If this function doesn't have a pic register, create one now. */ |
3444 | require_pic_register (); | |
2cb7d577 | 3445 | |
69120d26 | 3446 | if (reg == 0) |
3447 | { | |
e1ba4a27 | 3448 | gcc_assert (can_create_pseudo_p ()); |
ed29c566 | 3449 | reg = gen_reg_rtx (Pmode); |
69120d26 | 3450 | |
3451 | subregs = 1; | |
3452 | } | |
3453 | ||
3454 | #ifdef AOF_ASSEMBLER | |
3455 | /* The AOF assembler can generate relocations for these directly, and | |
66aacf41 | 3456 | understands that the PIC register has to be added into the offset. */ |
69120d26 | 3457 | insn = emit_insn (gen_pic_load_addr_based (reg, orig)); |
3458 | #else | |
3459 | if (subregs) | |
3460 | address = gen_reg_rtx (Pmode); | |
3461 | else | |
3462 | address = reg; | |
3463 | ||
8c4d8060 | 3464 | if (TARGET_ARM) |
3465 | emit_insn (gen_pic_load_addr_arm (address, orig)); | |
25f905c2 | 3466 | else if (TARGET_THUMB2) |
3467 | emit_insn (gen_pic_load_addr_thumb2 (address, orig)); | |
3468 | else /* TARGET_THUMB1 */ | |
3469 | emit_insn (gen_pic_load_addr_thumb1 (address, orig)); | |
69120d26 | 3470 | |
bac7fc85 | 3471 | /* VxWorks does not impose a fixed gap between segments; the run-time |
3472 | gap can be different from the object-file gap. We therefore can't | |
3473 | use GOTOFF unless we are absolutely sure that the symbol is in the | |
3474 | same segment as the GOT. Unfortunately, the flexibility of linker | |
3475 | scripts means that we can't be sure of that in general, so assume | |
3476 | that GOTOFF is never valid on VxWorks. */ | |
77744024 | 3477 | if ((GET_CODE (orig) == LABEL_REF |
9e7454d0 | 3478 | || (GET_CODE (orig) == SYMBOL_REF && |
e052c34d | 3479 | SYMBOL_REF_LOCAL_P (orig))) |
bac7fc85 | 3480 | && NEED_GOT_RELOC |
3481 | && !TARGET_VXWORKS_RTP) | |
2cb7d577 | 3482 | pic_ref = gen_rtx_PLUS (Pmode, cfun->machine->pic_reg, address); |
1644aab7 | 3483 | else |
3484 | { | |
e265a6da | 3485 | pic_ref = gen_const_mem (Pmode, |
2cb7d577 | 3486 | gen_rtx_PLUS (Pmode, cfun->machine->pic_reg, |
e265a6da | 3487 | address)); |
1644aab7 | 3488 | } |
3489 | ||
69120d26 | 3490 | insn = emit_move_insn (reg, pic_ref); |
3491 | #endif | |
69120d26 | 3492 | /* Put a REG_EQUAL note on this insn, so that it can be optimized |
3493 | by loop. */ | |
24153880 | 3494 | set_unique_reg_note (insn, REG_EQUAL, orig); |
3495 | ||
69120d26 | 3496 | return reg; |
3497 | } | |
3498 | else if (GET_CODE (orig) == CONST) | |
3499 | { | |
3500 | rtx base, offset; | |
3501 | ||
3502 | if (GET_CODE (XEXP (orig, 0)) == PLUS | |
2cb7d577 | 3503 | && XEXP (XEXP (orig, 0), 0) == cfun->machine->pic_reg) |
69120d26 | 3504 | return orig; |
3505 | ||
f655717d | 3506 | if (GET_CODE (XEXP (orig, 0)) == UNSPEC |
3507 | && XINT (XEXP (orig, 0), 1) == UNSPEC_TLS) | |
3508 | return orig; | |
3509 | ||
69120d26 | 3510 | if (reg == 0) |
3511 | { | |
e1ba4a27 | 3512 | gcc_assert (can_create_pseudo_p ()); |
ed29c566 | 3513 | reg = gen_reg_rtx (Pmode); |
69120d26 | 3514 | } |
3515 | ||
ed29c566 | 3516 | gcc_assert (GET_CODE (XEXP (orig, 0)) == PLUS); |
a60f1840 | 3517 | |
ed29c566 | 3518 | base = legitimize_pic_address (XEXP (XEXP (orig, 0), 0), Pmode, reg); |
3519 | offset = legitimize_pic_address (XEXP (XEXP (orig, 0), 1), Pmode, | |
3520 | base == reg ? 0 : reg); | |
69120d26 | 3521 | |
3522 | if (GET_CODE (offset) == CONST_INT) | |
3523 | { | |
3524 | /* The base register doesn't really matter, we only want to | |
3525 | test the index for the appropriate mode. */ | |
b4e8a300 | 3526 | if (!arm_legitimate_index_p (mode, offset, SET, 0)) |
d32b3557 | 3527 | { |
e1ba4a27 | 3528 | gcc_assert (can_create_pseudo_p ()); |
ed29c566 | 3529 | offset = force_reg (Pmode, offset); |
d32b3557 | 3530 | } |
69120d26 | 3531 | |
69120d26 | 3532 | if (GET_CODE (offset) == CONST_INT) |
b244d4c7 | 3533 | return plus_constant (base, INTVAL (offset)); |
69120d26 | 3534 | } |
3535 | ||
3536 | if (GET_MODE_SIZE (mode) > 4 | |
3537 | && (GET_MODE_CLASS (mode) == MODE_INT | |
3538 | || TARGET_SOFT_FLOAT)) | |
3539 | { | |
3540 | emit_insn (gen_addsi3 (reg, base, offset)); | |
3541 | return reg; | |
3542 | } | |
3543 | ||
76676c8e | 3544 | return gen_rtx_PLUS (Pmode, base, offset); |
69120d26 | 3545 | } |
69120d26 | 3546 | |
3547 | return orig; | |
3548 | } | |
3549 | ||
237533cc | 3550 | |
25f905c2 | 3551 | /* Find a spare register to use during the prolog of a function. */ |
237533cc | 3552 | |
3553 | static int | |
53f179c0 | 3554 | thumb_find_work_register (unsigned long pushed_regs_mask) |
237533cc | 3555 | { |
3556 | int reg; | |
3557 | ||
53f179c0 | 3558 | /* Check the argument registers first as these are call-used. The |
3559 | register allocation order means that sometimes r3 might be used | |
3560 | but earlier argument registers might not, so check them all. */ | |
3561 | for (reg = LAST_ARG_REGNUM; reg >= 0; reg --) | |
3072d30e | 3562 | if (!df_regs_ever_live_p (reg)) |
53f179c0 | 3563 | return reg; |
3564 | ||
3565 | /* Before going on to check the call-saved registers we can try a couple | |
3566 | more ways of deducing that r3 is available. The first is when we are | |
3567 | pushing anonymous arguments onto the stack and we have less than 4 | |
3568 | registers worth of fixed arguments(*). In this case r3 will be part of | |
3569 | the variable argument list and so we can be sure that it will be | |
3570 | pushed right at the start of the function. Hence it will be available | |
3571 | for the rest of the prologue. | |
3572 | (*): ie current_function_pretend_args_size is greater than 0. */ | |
3573 | if (cfun->machine->uses_anonymous_args | |
3574 | && current_function_pretend_args_size > 0) | |
237533cc | 3575 | return LAST_ARG_REGNUM; |
3576 | ||
53f179c0 | 3577 | /* The other case is when we have fixed arguments but less than 4 registers |
3578 | worth. In this case r3 might be used in the body of the function, but | |
3579 | it is not being used to convey an argument into the function. In theory | |
3580 | we could just check current_function_args_size to see how many bytes are | |
3581 | being passed in argument registers, but it seems that it is unreliable. | |
3582 | Sometimes it will have the value 0 when in fact arguments are being | |
3583 | passed. (See testcase execute/20021111-1.c for an example). So we also | |
3584 | check the args_info.nregs field as well. The problem with this field is | |
3585 | that it makes no allowances for arguments that are passed to the | |
3586 | function but which are not used. Hence we could miss an opportunity | |
3587 | when a function has an unused argument in r3. But it is better to be | |
3588 | safe than to be sorry. */ | |
3589 | if (! cfun->machine->uses_anonymous_args | |
3590 | && current_function_args_size >= 0 | |
3591 | && current_function_args_size <= (LAST_ARG_REGNUM * UNITS_PER_WORD) | |
3592 | && cfun->args_info.nregs < 4) | |
3593 | return LAST_ARG_REGNUM; | |
a60f1840 | 3594 | |
53f179c0 | 3595 | /* Otherwise look for a call-saved register that is going to be pushed. */ |
3596 | for (reg = LAST_LO_REGNUM; reg > LAST_ARG_REGNUM; reg --) | |
3597 | if (pushed_regs_mask & (1 << reg)) | |
237533cc | 3598 | return reg; |
3599 | ||
25f905c2 | 3600 | if (TARGET_THUMB2) |
3601 | { | |
3602 | /* Thumb-2 can use high regs. */ | |
3603 | for (reg = FIRST_HI_REGNUM; reg < 15; reg ++) | |
3604 | if (pushed_regs_mask & (1 << reg)) | |
3605 | return reg; | |
3606 | } | |
53f179c0 | 3607 | /* Something went wrong - thumb_compute_save_reg_mask() |
3608 | should have arranged for a suitable register to be pushed. */ | |
ed29c566 | 3609 | gcc_unreachable (); |
237533cc | 3610 | } |
3611 | ||
6cdcb15c | 3612 | static GTY(()) int pic_labelno; |
ea32a06e | 3613 | |
b935b306 | 3614 | /* Generate code to load the PIC register. In thumb mode SCRATCH is a |
3615 | low register. */ | |
ea32a06e | 3616 | |
69120d26 | 3617 | void |
aa6af490 | 3618 | arm_load_pic_register (unsigned long saved_regs ATTRIBUTE_UNUSED) |
69120d26 | 3619 | { |
3620 | #ifndef AOF_ASSEMBLER | |
bac7fc85 | 3621 | rtx l1, labelno, pic_tmp, pic_tmp2, pic_rtx, pic_reg; |
69120d26 | 3622 | rtx global_offset_table; |
3623 | ||
55c1e470 | 3624 | if (current_function_uses_pic_offset_table == 0 || TARGET_SINGLE_PIC_BASE) |
69120d26 | 3625 | return; |
3626 | ||
ed29c566 | 3627 | gcc_assert (flag_pic); |
69120d26 | 3628 | |
bac7fc85 | 3629 | pic_reg = cfun->machine->pic_reg; |
3630 | if (TARGET_VXWORKS_RTP) | |
3631 | { | |
3632 | pic_rtx = gen_rtx_SYMBOL_REF (Pmode, VXWORKS_GOTT_BASE); | |
3633 | pic_rtx = gen_rtx_CONST (Pmode, pic_rtx); | |
3634 | emit_insn (gen_pic_load_addr_arm (pic_reg, pic_rtx)); | |
76676c8e | 3635 | |
bac7fc85 | 3636 | emit_insn (gen_rtx_SET (Pmode, pic_reg, gen_rtx_MEM (Pmode, pic_reg))); |
9e7454d0 | 3637 | |
bac7fc85 | 3638 | pic_tmp = gen_rtx_SYMBOL_REF (Pmode, VXWORKS_GOTT_INDEX); |
3639 | emit_insn (gen_pic_offset_arm (pic_reg, pic_reg, pic_tmp)); | |
8c4d8060 | 3640 | } |
bac7fc85 | 3641 | else |
25f905c2 | 3642 | { |
bac7fc85 | 3643 | /* We use an UNSPEC rather than a LABEL_REF because this label |
3644 | never appears in the code stream. */ | |
3645 | ||
3646 | labelno = GEN_INT (pic_labelno++); | |
3647 | l1 = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, labelno), UNSPEC_PIC_LABEL); | |
3648 | l1 = gen_rtx_CONST (VOIDmode, l1); | |
3649 | ||
3650 | global_offset_table | |
3651 | = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_"); | |
3652 | /* On the ARM the PC register contains 'dot + 8' at the time of the | |
3653 | addition, on the Thumb it is 'dot + 4'. */ | |
3654 | pic_tmp = plus_constant (l1, TARGET_ARM ? 8 : 4); | |
3655 | if (GOT_PCREL) | |
25f905c2 | 3656 | { |
bac7fc85 | 3657 | pic_tmp2 = gen_rtx_PLUS (Pmode, global_offset_table, pc_rtx); |
3658 | pic_tmp2 = gen_rtx_CONST (VOIDmode, pic_tmp2); | |
25f905c2 | 3659 | } |
3660 | else | |
bac7fc85 | 3661 | pic_tmp2 = gen_rtx_CONST (VOIDmode, global_offset_table); |
3662 | ||
3663 | pic_rtx = gen_rtx_MINUS (Pmode, pic_tmp2, pic_tmp); | |
3664 | pic_rtx = gen_rtx_CONST (Pmode, pic_rtx); | |
3665 | ||
3666 | if (TARGET_ARM) | |
25f905c2 | 3667 | { |
bac7fc85 | 3668 | emit_insn (gen_pic_load_addr_arm (pic_reg, pic_rtx)); |
3669 | emit_insn (gen_pic_add_dot_plus_eight (pic_reg, pic_reg, labelno)); | |
25f905c2 | 3670 | } |
bac7fc85 | 3671 | else if (TARGET_THUMB2) |
3672 | { | |
3673 | /* Thumb-2 only allows very limited access to the PC. Calculate the | |
3674 | address in a temporary register. */ | |
3675 | if (arm_pic_register != INVALID_REGNUM) | |
3676 | { | |
3677 | pic_tmp = gen_rtx_REG (SImode, | |
3678 | thumb_find_work_register (saved_regs)); | |
3679 | } | |
3680 | else | |
3681 | { | |
e1ba4a27 | 3682 | gcc_assert (can_create_pseudo_p ()); |
bac7fc85 | 3683 | pic_tmp = gen_reg_rtx (Pmode); |
3684 | } | |
25f905c2 | 3685 | |
bac7fc85 | 3686 | emit_insn (gen_pic_load_addr_thumb2 (pic_reg, pic_rtx)); |
3687 | emit_insn (gen_pic_load_dot_plus_four (pic_tmp, labelno)); | |
3688 | emit_insn (gen_addsi3 (pic_reg, pic_reg, pic_tmp)); | |
3689 | } | |
3690 | else /* TARGET_THUMB1 */ | |
ea32a06e | 3691 | { |
bac7fc85 | 3692 | if (arm_pic_register != INVALID_REGNUM |
3693 | && REGNO (pic_reg) > LAST_LO_REGNUM) | |
3694 | { | |
3695 | /* We will have pushed the pic register, so we should always be | |
3696 | able to find a work register. */ | |
3697 | pic_tmp = gen_rtx_REG (SImode, | |
3698 | thumb_find_work_register (saved_regs)); | |
3699 | emit_insn (gen_pic_load_addr_thumb1 (pic_tmp, pic_rtx)); | |
3700 | emit_insn (gen_movsi (pic_offset_table_rtx, pic_tmp)); | |
3701 | } | |
3702 | else | |
3703 | emit_insn (gen_pic_load_addr_thumb1 (pic_reg, pic_rtx)); | |
3704 | emit_insn (gen_pic_add_dot_plus_four (pic_reg, pic_reg, labelno)); | |
ea32a06e | 3705 | } |
8c4d8060 | 3706 | } |
69120d26 | 3707 | |
69120d26 | 3708 | /* Need to emit this whether or not we obey regdecls, |
3709 | since setjmp/longjmp can cause life info to screw up. */ | |
bac7fc85 | 3710 | emit_insn (gen_rtx_USE (VOIDmode, pic_reg)); |
69120d26 | 3711 | #endif /* AOF_ASSEMBLER */ |
3712 | } | |
3713 | ||
ea32a06e | 3714 | |
d32b3557 | 3715 | /* Return nonzero if X is valid as an ARM state addressing register. */ |
3716 | static int | |
ebd88a36 | 3717 | arm_address_register_rtx_p (rtx x, int strict_p) |
d32b3557 | 3718 | { |
3719 | int regno; | |
3720 | ||
3721 | if (GET_CODE (x) != REG) | |
3722 | return 0; | |
3723 | ||
3724 | regno = REGNO (x); | |
3725 | ||
3726 | if (strict_p) | |
3727 | return ARM_REGNO_OK_FOR_BASE_P (regno); | |
3728 | ||
3729 | return (regno <= LAST_ARM_REGNUM | |
3730 | || regno >= FIRST_PSEUDO_REGISTER | |
3731 | || regno == FRAME_POINTER_REGNUM | |
3732 | || regno == ARG_POINTER_REGNUM); | |
3733 | } | |
3734 | ||
f655717d | 3735 | /* Return TRUE if this rtx is the difference of a symbol and a label, |
3736 | and will reduce to a PC-relative relocation in the object file. | |
3737 | Expressions like this can be left alone when generating PIC, rather | |
3738 | than forced through the GOT. */ | |
3739 | static int | |
3740 | pcrel_constant_p (rtx x) | |
3741 | { | |
3742 | if (GET_CODE (x) == MINUS) | |
3743 | return symbol_mentioned_p (XEXP (x, 0)) && label_mentioned_p (XEXP (x, 1)); | |
3744 | ||
3745 | return FALSE; | |
3746 | } | |
3747 | ||
d32b3557 | 3748 | /* Return nonzero if X is a valid ARM state address operand. */ |
3749 | int | |
b4e8a300 | 3750 | arm_legitimate_address_p (enum machine_mode mode, rtx x, RTX_CODE outer, |
3751 | int strict_p) | |
d32b3557 | 3752 | { |
a8a3b539 | 3753 | bool use_ldrd; |
3754 | enum rtx_code code = GET_CODE (x); | |
9e7454d0 | 3755 | |
d32b3557 | 3756 | if (arm_address_register_rtx_p (x, strict_p)) |
3757 | return 1; | |
3758 | ||
a8a3b539 | 3759 | use_ldrd = (TARGET_LDRD |
3760 | && (mode == DImode | |
3761 | || (mode == DFmode && (TARGET_SOFT_FLOAT || TARGET_VFP)))); | |
3762 | ||
3763 | if (code == POST_INC || code == PRE_DEC | |
3764 | || ((code == PRE_INC || code == POST_DEC) | |
3765 | && (use_ldrd || GET_MODE_SIZE (mode) <= 4))) | |
d32b3557 | 3766 | return arm_address_register_rtx_p (XEXP (x, 0), strict_p); |
3767 | ||
a8a3b539 | 3768 | else if ((code == POST_MODIFY || code == PRE_MODIFY) |
d32b3557 | 3769 | && arm_address_register_rtx_p (XEXP (x, 0), strict_p) |
3770 | && GET_CODE (XEXP (x, 1)) == PLUS | |
0bb91b60 | 3771 | && rtx_equal_p (XEXP (XEXP (x, 1), 0), XEXP (x, 0))) |
a8a3b539 | 3772 | { |
3773 | rtx addend = XEXP (XEXP (x, 1), 1); | |
3774 | ||
a361b456 | 3775 | /* Don't allow ldrd post increment by register because it's hard |
a8a3b539 | 3776 | to fixup invalid register choices. */ |
3777 | if (use_ldrd | |
3778 | && GET_CODE (x) == POST_MODIFY | |
3779 | && GET_CODE (addend) == REG) | |
3780 | return 0; | |
3781 | ||
3782 | return ((use_ldrd || GET_MODE_SIZE (mode) <= 4) | |
3783 | && arm_legitimate_index_p (mode, addend, outer, strict_p)); | |
3784 | } | |
d32b3557 | 3785 | |
3786 | /* After reload constants split into minipools will have addresses | |
3787 | from a LABEL_REF. */ | |
e027ee9c | 3788 | else if (reload_completed |
a8a3b539 | 3789 | && (code == LABEL_REF |
3790 | || (code == CONST | |
d32b3557 | 3791 | && GET_CODE (XEXP (x, 0)) == PLUS |
3792 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == LABEL_REF | |
3793 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT))) | |
3794 | return 1; | |
3795 | ||
d98a3884 | 3796 | else if (mode == TImode || (TARGET_NEON && VALID_NEON_STRUCT_MODE (mode))) |
d32b3557 | 3797 | return 0; |
3798 | ||
a8a3b539 | 3799 | else if (code == PLUS) |
d32b3557 | 3800 | { |
3801 | rtx xop0 = XEXP (x, 0); | |
3802 | rtx xop1 = XEXP (x, 1); | |
3803 | ||
3804 | return ((arm_address_register_rtx_p (xop0, strict_p) | |
b4e8a300 | 3805 | && arm_legitimate_index_p (mode, xop1, outer, strict_p)) |
d32b3557 | 3806 | || (arm_address_register_rtx_p (xop1, strict_p) |
b4e8a300 | 3807 | && arm_legitimate_index_p (mode, xop0, outer, strict_p))); |
d32b3557 | 3808 | } |
3809 | ||
3810 | #if 0 | |
3811 | /* Reload currently can't handle MINUS, so disable this for now */ | |
3812 | else if (GET_CODE (x) == MINUS) | |
3813 | { | |
3814 | rtx xop0 = XEXP (x, 0); | |
3815 | rtx xop1 = XEXP (x, 1); | |
3816 | ||
3817 | return (arm_address_register_rtx_p (xop0, strict_p) | |
b4e8a300 | 3818 | && arm_legitimate_index_p (mode, xop1, outer, strict_p)); |
d32b3557 | 3819 | } |
3820 | #endif | |
3821 | ||
3822 | else if (GET_MODE_CLASS (mode) != MODE_FLOAT | |
a8a3b539 | 3823 | && code == SYMBOL_REF |
d32b3557 | 3824 | && CONSTANT_POOL_ADDRESS_P (x) |
3825 | && ! (flag_pic | |
f655717d | 3826 | && symbol_mentioned_p (get_pool_constant (x)) |
3827 | && ! pcrel_constant_p (get_pool_constant (x)))) | |
d32b3557 | 3828 | return 1; |
3829 | ||
d32b3557 | 3830 | return 0; |
3831 | } | |
3832 | ||
25f905c2 | 3833 | /* Return nonzero if X is a valid Thumb-2 address operand. */ |
3834 | int | |
3835 | thumb2_legitimate_address_p (enum machine_mode mode, rtx x, int strict_p) | |
3836 | { | |
3837 | bool use_ldrd; | |
3838 | enum rtx_code code = GET_CODE (x); | |
3839 | ||
3840 | if (arm_address_register_rtx_p (x, strict_p)) | |
3841 | return 1; | |
3842 | ||
3843 | use_ldrd = (TARGET_LDRD | |
3844 | && (mode == DImode | |
3845 | || (mode == DFmode && (TARGET_SOFT_FLOAT || TARGET_VFP)))); | |
3846 | ||
3847 | if (code == POST_INC || code == PRE_DEC | |
3848 | || ((code == PRE_INC || code == POST_DEC) | |
3849 | && (use_ldrd || GET_MODE_SIZE (mode) <= 4))) | |
3850 | return arm_address_register_rtx_p (XEXP (x, 0), strict_p); | |
3851 | ||
3852 | else if ((code == POST_MODIFY || code == PRE_MODIFY) | |
3853 | && arm_address_register_rtx_p (XEXP (x, 0), strict_p) | |
3854 | && GET_CODE (XEXP (x, 1)) == PLUS | |
3855 | && rtx_equal_p (XEXP (XEXP (x, 1), 0), XEXP (x, 0))) | |
3856 | { | |
3857 | /* Thumb-2 only has autoincrement by constant. */ | |
3858 | rtx addend = XEXP (XEXP (x, 1), 1); | |
3859 | HOST_WIDE_INT offset; | |
3860 | ||
3861 | if (GET_CODE (addend) != CONST_INT) | |
3862 | return 0; | |
3863 | ||
3864 | offset = INTVAL(addend); | |
3865 | if (GET_MODE_SIZE (mode) <= 4) | |
3866 | return (offset > -256 && offset < 256); | |
3867 | ||
3868 | return (use_ldrd && offset > -1024 && offset < 1024 | |
3869 | && (offset & 3) == 0); | |
3870 | } | |
3871 | ||
3872 | /* After reload constants split into minipools will have addresses | |
3873 | from a LABEL_REF. */ | |
3874 | else if (reload_completed | |
3875 | && (code == LABEL_REF | |
3876 | || (code == CONST | |
3877 | && GET_CODE (XEXP (x, 0)) == PLUS | |
3878 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == LABEL_REF | |
3879 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT))) | |
3880 | return 1; | |
3881 | ||
d98a3884 | 3882 | else if (mode == TImode || (TARGET_NEON && VALID_NEON_STRUCT_MODE (mode))) |
25f905c2 | 3883 | return 0; |
3884 | ||
3885 | else if (code == PLUS) | |
3886 | { | |
3887 | rtx xop0 = XEXP (x, 0); | |
3888 | rtx xop1 = XEXP (x, 1); | |
3889 | ||
3890 | return ((arm_address_register_rtx_p (xop0, strict_p) | |
3891 | && thumb2_legitimate_index_p (mode, xop1, strict_p)) | |
3892 | || (arm_address_register_rtx_p (xop1, strict_p) | |
3893 | && thumb2_legitimate_index_p (mode, xop0, strict_p))); | |
3894 | } | |
3895 | ||
3896 | else if (GET_MODE_CLASS (mode) != MODE_FLOAT | |
3897 | && code == SYMBOL_REF | |
3898 | && CONSTANT_POOL_ADDRESS_P (x) | |
3899 | && ! (flag_pic | |
3900 | && symbol_mentioned_p (get_pool_constant (x)) | |
3901 | && ! pcrel_constant_p (get_pool_constant (x)))) | |
3902 | return 1; | |
3903 | ||
3904 | return 0; | |
3905 | } | |
3906 | ||
d32b3557 | 3907 | /* Return nonzero if INDEX is valid for an address index operand in |
3908 | ARM state. */ | |
3909 | static int | |
b4e8a300 | 3910 | arm_legitimate_index_p (enum machine_mode mode, rtx index, RTX_CODE outer, |
3911 | int strict_p) | |
d32b3557 | 3912 | { |
3913 | HOST_WIDE_INT range; | |
3914 | enum rtx_code code = GET_CODE (index); | |
3915 | ||
69130f53 | 3916 | /* Standard coprocessor addressing modes. */ |
3917 | if (TARGET_HARD_FLOAT | |
3918 | && (TARGET_FPA || TARGET_MAVERICK) | |
3919 | && (GET_MODE_CLASS (mode) == MODE_FLOAT | |
3920 | || (TARGET_MAVERICK && mode == DImode))) | |
d32b3557 | 3921 | return (code == CONST_INT && INTVAL (index) < 1024 |
3922 | && INTVAL (index) > -1024 | |
3923 | && (INTVAL (index) & 3) == 0); | |
3924 | ||
d98a3884 | 3925 | if (TARGET_NEON |
3926 | && (VALID_NEON_DREG_MODE (mode) || VALID_NEON_QREG_MODE (mode))) | |
3927 | return (code == CONST_INT | |
3928 | && INTVAL (index) < 1016 | |
3929 | && INTVAL (index) > -1024 | |
3930 | && (INTVAL (index) & 3) == 0); | |
3931 | ||
755eb2b4 | 3932 | if (TARGET_REALLY_IWMMXT && VALID_IWMMXT_REG_MODE (mode)) |
3933 | return (code == CONST_INT | |
7d647a20 | 3934 | && INTVAL (index) < 1024 |
3935 | && INTVAL (index) > -1024 | |
3936 | && (INTVAL (index) & 3) == 0); | |
755eb2b4 | 3937 | |
a8a3b539 | 3938 | if (arm_address_register_rtx_p (index, strict_p) |
3939 | && (GET_MODE_SIZE (mode) <= 4)) | |
3940 | return 1; | |
3941 | ||
3942 | if (mode == DImode || mode == DFmode) | |
3943 | { | |
3944 | if (code == CONST_INT) | |
3945 | { | |
3946 | HOST_WIDE_INT val = INTVAL (index); | |
3947 | ||
3948 | if (TARGET_LDRD) | |
3949 | return val > -256 && val < 256; | |
3950 | else | |
da873b17 | 3951 | return val > -4096 && val < 4092; |
a8a3b539 | 3952 | } |
3953 | ||
3954 | return TARGET_LDRD && arm_address_register_rtx_p (index, strict_p); | |
3955 | } | |
3956 | ||
d32b3557 | 3957 | if (GET_MODE_SIZE (mode) <= 4 |
b4e8a300 | 3958 | && ! (arm_arch4 |
3959 | && (mode == HImode | |
3960 | || (mode == QImode && outer == SIGN_EXTEND)))) | |
d32b3557 | 3961 | { |
b4e8a300 | 3962 | if (code == MULT) |
3963 | { | |
3964 | rtx xiop0 = XEXP (index, 0); | |
3965 | rtx xiop1 = XEXP (index, 1); | |
3966 | ||
3967 | return ((arm_address_register_rtx_p (xiop0, strict_p) | |
3968 | && power_of_two_operand (xiop1, SImode)) | |
3969 | || (arm_address_register_rtx_p (xiop1, strict_p) | |
3970 | && power_of_two_operand (xiop0, SImode))); | |
3971 | } | |
3972 | else if (code == LSHIFTRT || code == ASHIFTRT | |
3973 | || code == ASHIFT || code == ROTATERT) | |
3974 | { | |
3975 | rtx op = XEXP (index, 1); | |
d32b3557 | 3976 | |
b4e8a300 | 3977 | return (arm_address_register_rtx_p (XEXP (index, 0), strict_p) |
3978 | && GET_CODE (op) == CONST_INT | |
3979 | && INTVAL (op) > 0 | |
3980 | && INTVAL (op) <= 31); | |
3981 | } | |
d32b3557 | 3982 | } |
3983 | ||
b4e8a300 | 3984 | /* For ARM v4 we may be doing a sign-extend operation during the |
3985 | load. */ | |
806c15d6 | 3986 | if (arm_arch4) |
b4e8a300 | 3987 | { |
3988 | if (mode == HImode || (outer == SIGN_EXTEND && mode == QImode)) | |
3989 | range = 256; | |
3990 | else | |
3991 | range = 4096; | |
3992 | } | |
806c15d6 | 3993 | else |
3994 | range = (mode == HImode) ? 4095 : 4096; | |
d32b3557 | 3995 | |
3996 | return (code == CONST_INT | |
3997 | && INTVAL (index) < range | |
3998 | && INTVAL (index) > -range); | |
f2989adc | 3999 | } |
4000 | ||
25f905c2 | 4001 | /* Return true if OP is a valid index scaling factor for Thumb-2 address |
4002 | index operand. i.e. 1, 2, 4 or 8. */ | |
4003 | static bool | |
4004 | thumb2_index_mul_operand (rtx op) | |
4005 | { | |
4006 | HOST_WIDE_INT val; | |
4007 | ||
4008 | if (GET_CODE(op) != CONST_INT) | |
4009 | return false; | |
4010 | ||
4011 | val = INTVAL(op); | |
4012 | return (val == 1 || val == 2 || val == 4 || val == 8); | |
4013 | } | |
4014 | ||
4015 | /* Return nonzero if INDEX is a valid Thumb-2 address index operand. */ | |
4016 | static int | |
4017 | thumb2_legitimate_index_p (enum machine_mode mode, rtx index, int strict_p) | |
4018 | { | |
4019 | enum rtx_code code = GET_CODE (index); | |
4020 | ||
4021 | /* ??? Combine arm and thumb2 coprocessor addressing modes. */ | |
4022 | /* Standard coprocessor addressing modes. */ | |
4023 | if (TARGET_HARD_FLOAT | |
4024 | && (TARGET_FPA || TARGET_MAVERICK) | |
4025 | && (GET_MODE_CLASS (mode) == MODE_FLOAT | |
4026 | || (TARGET_MAVERICK && mode == DImode))) | |
4027 | return (code == CONST_INT && INTVAL (index) < 1024 | |
4028 | && INTVAL (index) > -1024 | |
4029 | && (INTVAL (index) & 3) == 0); | |
4030 | ||
4031 | if (TARGET_REALLY_IWMMXT && VALID_IWMMXT_REG_MODE (mode)) | |
f1225f6f | 4032 | { |
4033 | /* For DImode assume values will usually live in core regs | |
4034 | and only allow LDRD addressing modes. */ | |
4035 | if (!TARGET_LDRD || mode != DImode) | |
4036 | return (code == CONST_INT | |
4037 | && INTVAL (index) < 1024 | |
4038 | && INTVAL (index) > -1024 | |
4039 | && (INTVAL (index) & 3) == 0); | |
4040 | } | |
25f905c2 | 4041 | |
d98a3884 | 4042 | if (TARGET_NEON |
4043 | && (VALID_NEON_DREG_MODE (mode) || VALID_NEON_QREG_MODE (mode))) | |
4044 | return (code == CONST_INT | |
4045 | && INTVAL (index) < 1016 | |
4046 | && INTVAL (index) > -1024 | |
4047 | && (INTVAL (index) & 3) == 0); | |
4048 | ||
25f905c2 | 4049 | if (arm_address_register_rtx_p (index, strict_p) |
4050 | && (GET_MODE_SIZE (mode) <= 4)) | |
4051 | return 1; | |
4052 | ||
4053 | if (mode == DImode || mode == DFmode) | |
4054 | { | |
4055 | HOST_WIDE_INT val = INTVAL (index); | |
4056 | /* ??? Can we assume ldrd for thumb2? */ | |
4057 | /* Thumb-2 ldrd only has reg+const addressing modes. */ | |
4058 | if (code != CONST_INT) | |
4059 | return 0; | |
4060 | ||
4061 | /* ldrd supports offsets of +-1020. | |
4062 | However the ldr fallback does not. */ | |
4063 | return val > -256 && val < 256 && (val & 3) == 0; | |
4064 | } | |
4065 | ||
4066 | if (code == MULT) | |
4067 | { | |
4068 | rtx xiop0 = XEXP (index, 0); | |
4069 | rtx xiop1 = XEXP (index, 1); | |
4070 | ||
4071 | return ((arm_address_register_rtx_p (xiop0, strict_p) | |
4072 | && thumb2_index_mul_operand (xiop1)) | |
4073 | || (arm_address_register_rtx_p (xiop1, strict_p) | |
4074 | && thumb2_index_mul_operand (xiop0))); | |
4075 | } | |
4076 | else if (code == ASHIFT) | |
4077 | { | |
4078 | rtx op = XEXP (index, 1); | |
4079 | ||
4080 | return (arm_address_register_rtx_p (XEXP (index, 0), strict_p) | |
4081 | && GET_CODE (op) == CONST_INT | |
4082 | && INTVAL (op) > 0 | |
4083 | && INTVAL (op) <= 3); | |
4084 | } | |
4085 | ||
4086 | return (code == CONST_INT | |
4087 | && INTVAL (index) < 4096 | |
4088 | && INTVAL (index) > -256); | |
4089 | } | |
4090 | ||
4091 | /* Return nonzero if X is valid as a 16-bit Thumb state base register. */ | |
f2989adc | 4092 | static int |
25f905c2 | 4093 | thumb1_base_register_rtx_p (rtx x, enum machine_mode mode, int strict_p) |
f2989adc | 4094 | { |
4095 | int regno; | |
4096 | ||
4097 | if (GET_CODE (x) != REG) | |
4098 | return 0; | |
4099 | ||
4100 | regno = REGNO (x); | |
4101 | ||
4102 | if (strict_p) | |
25f905c2 | 4103 | return THUMB1_REGNO_MODE_OK_FOR_BASE_P (regno, mode); |
f2989adc | 4104 | |
4105 | return (regno <= LAST_LO_REGNUM | |
221038d3 | 4106 | || regno > LAST_VIRTUAL_REGISTER |
f2989adc | 4107 | || regno == FRAME_POINTER_REGNUM |
4108 | || (GET_MODE_SIZE (mode) >= 4 | |
4109 | && (regno == STACK_POINTER_REGNUM | |
11e7c7fd | 4110 | || regno >= FIRST_PSEUDO_REGISTER |
f2989adc | 4111 | || x == hard_frame_pointer_rtx |
4112 | || x == arg_pointer_rtx))); | |
4113 | } | |
4114 | ||
4115 | /* Return nonzero if x is a legitimate index register. This is the case | |
4116 | for any base register that can access a QImode object. */ | |
4117 | inline static int | |
25f905c2 | 4118 | thumb1_index_register_rtx_p (rtx x, int strict_p) |
f2989adc | 4119 | { |
25f905c2 | 4120 | return thumb1_base_register_rtx_p (x, QImode, strict_p); |
f2989adc | 4121 | } |
4122 | ||
25f905c2 | 4123 | /* Return nonzero if x is a legitimate 16-bit Thumb-state address. |
9e7454d0 | 4124 | |
f2989adc | 4125 | The AP may be eliminated to either the SP or the FP, so we use the |
4126 | least common denominator, e.g. SImode, and offsets from 0 to 64. | |
4127 | ||
4128 | ??? Verify whether the above is the right approach. | |
4129 | ||
4130 | ??? Also, the FP may be eliminated to the SP, so perhaps that | |
4131 | needs special handling also. | |
4132 | ||
4133 | ??? Look at how the mips16 port solves this problem. It probably uses | |
4134 | better ways to solve some of these problems. | |
4135 | ||
4136 | Although it is not incorrect, we don't accept QImode and HImode | |
4137 | addresses based on the frame pointer or arg pointer until the | |
4138 | reload pass starts. This is so that eliminating such addresses | |
4139 | into stack based ones won't produce impossible code. */ | |
4140 | int | |
25f905c2 | 4141 | thumb1_legitimate_address_p (enum machine_mode mode, rtx x, int strict_p) |
f2989adc | 4142 | { |
4143 | /* ??? Not clear if this is right. Experiment. */ | |
4144 | if (GET_MODE_SIZE (mode) < 4 | |
4145 | && !(reload_in_progress || reload_completed) | |
4146 | && (reg_mentioned_p (frame_pointer_rtx, x) | |
4147 | || reg_mentioned_p (arg_pointer_rtx, x) | |
4148 | || reg_mentioned_p (virtual_incoming_args_rtx, x) | |
4149 | || reg_mentioned_p (virtual_outgoing_args_rtx, x) | |
4150 | || reg_mentioned_p (virtual_stack_dynamic_rtx, x) | |
4151 | || reg_mentioned_p (virtual_stack_vars_rtx, x))) | |
4152 | return 0; | |
4153 | ||
4154 | /* Accept any base register. SP only in SImode or larger. */ | |
25f905c2 | 4155 | else if (thumb1_base_register_rtx_p (x, mode, strict_p)) |
f2989adc | 4156 | return 1; |
4157 | ||
2efea8c0 | 4158 | /* This is PC relative data before arm_reorg runs. */ |
f2989adc | 4159 | else if (GET_MODE_SIZE (mode) >= 4 && CONSTANT_P (x) |
4160 | && GET_CODE (x) == SYMBOL_REF | |
2cb7d577 | 4161 | && CONSTANT_POOL_ADDRESS_P (x) && !flag_pic) |
f2989adc | 4162 | return 1; |
4163 | ||
2efea8c0 | 4164 | /* This is PC relative data after arm_reorg runs. */ |
f2989adc | 4165 | else if (GET_MODE_SIZE (mode) >= 4 && reload_completed |
4166 | && (GET_CODE (x) == LABEL_REF | |
4167 | || (GET_CODE (x) == CONST | |
4168 | && GET_CODE (XEXP (x, 0)) == PLUS | |
4169 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == LABEL_REF | |
4170 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT))) | |
4171 | return 1; | |
4172 | ||
4173 | /* Post-inc indexing only supported for SImode and larger. */ | |
4174 | else if (GET_CODE (x) == POST_INC && GET_MODE_SIZE (mode) >= 4 | |
25f905c2 | 4175 | && thumb1_index_register_rtx_p (XEXP (x, 0), strict_p)) |
f2989adc | 4176 | return 1; |
4177 | ||
4178 | else if (GET_CODE (x) == PLUS) | |
4179 | { | |
4180 | /* REG+REG address can be any two index registers. */ | |
4181 | /* We disallow FRAME+REG addressing since we know that FRAME | |
4182 | will be replaced with STACK, and SP relative addressing only | |
4183 | permits SP+OFFSET. */ | |
4184 | if (GET_MODE_SIZE (mode) <= 4 | |
4185 | && XEXP (x, 0) != frame_pointer_rtx | |
4186 | && XEXP (x, 1) != frame_pointer_rtx | |
25f905c2 | 4187 | && thumb1_index_register_rtx_p (XEXP (x, 0), strict_p) |
4188 | && thumb1_index_register_rtx_p (XEXP (x, 1), strict_p)) | |
f2989adc | 4189 | return 1; |
4190 | ||
4191 | /* REG+const has 5-7 bit offset for non-SP registers. */ | |
25f905c2 | 4192 | else if ((thumb1_index_register_rtx_p (XEXP (x, 0), strict_p) |
f2989adc | 4193 | || XEXP (x, 0) == arg_pointer_rtx) |
4194 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
4195 | && thumb_legitimate_offset_p (mode, INTVAL (XEXP (x, 1)))) | |
4196 | return 1; | |
4197 | ||
85c36fd1 | 4198 | /* REG+const has 10-bit offset for SP, but only SImode and |
f2989adc | 4199 | larger is supported. */ |
4200 | /* ??? Should probably check for DI/DFmode overflow here | |
4201 | just like GO_IF_LEGITIMATE_OFFSET does. */ | |
4202 | else if (GET_CODE (XEXP (x, 0)) == REG | |
4203 | && REGNO (XEXP (x, 0)) == STACK_POINTER_REGNUM | |
4204 | && GET_MODE_SIZE (mode) >= 4 | |
4205 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
4206 | && INTVAL (XEXP (x, 1)) >= 0 | |
4207 | && INTVAL (XEXP (x, 1)) + GET_MODE_SIZE (mode) <= 1024 | |
4208 | && (INTVAL (XEXP (x, 1)) & 3) == 0) | |
4209 | return 1; | |
4210 | ||
4211 | else if (GET_CODE (XEXP (x, 0)) == REG | |
2e9ff3c0 | 4212 | && (REGNO (XEXP (x, 0)) == FRAME_POINTER_REGNUM |
4213 | || REGNO (XEXP (x, 0)) == ARG_POINTER_REGNUM | |
4214 | || (REGNO (XEXP (x, 0)) >= FIRST_VIRTUAL_REGISTER | |
4215 | && REGNO (XEXP (x, 0)) <= LAST_VIRTUAL_REGISTER)) | |
f2989adc | 4216 | && GET_MODE_SIZE (mode) >= 4 |
4217 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
4218 | && (INTVAL (XEXP (x, 1)) & 3) == 0) | |
4219 | return 1; | |
4220 | } | |
4221 | ||
4222 | else if (GET_MODE_CLASS (mode) != MODE_FLOAT | |
a941568e | 4223 | && GET_MODE_SIZE (mode) == 4 |
f2989adc | 4224 | && GET_CODE (x) == SYMBOL_REF |
4225 | && CONSTANT_POOL_ADDRESS_P (x) | |
f655717d | 4226 | && ! (flag_pic |
4227 | && symbol_mentioned_p (get_pool_constant (x)) | |
4228 | && ! pcrel_constant_p (get_pool_constant (x)))) | |
f2989adc | 4229 | return 1; |
4230 | ||
4231 | return 0; | |
4232 | } | |
4233 | ||
4234 | /* Return nonzero if VAL can be used as an offset in a Thumb-state address | |
4235 | instruction of mode MODE. */ | |
4236 | int | |
ebd88a36 | 4237 | thumb_legitimate_offset_p (enum machine_mode mode, HOST_WIDE_INT val) |
f2989adc | 4238 | { |
4239 | switch (GET_MODE_SIZE (mode)) | |
4240 | { | |
4241 | case 1: | |
4242 | return val >= 0 && val < 32; | |
4243 | ||
4244 | case 2: | |
4245 | return val >= 0 && val < 64 && (val & 1) == 0; | |
4246 | ||
4247 | default: | |
4248 | return (val >= 0 | |
4249 | && (val + GET_MODE_SIZE (mode)) <= 128 | |
4250 | && (val & 3) == 0); | |
4251 | } | |
4252 | } | |
4253 | ||
f655717d | 4254 | /* Build the SYMBOL_REF for __tls_get_addr. */ |
4255 | ||
4256 | static GTY(()) rtx tls_get_addr_libfunc; | |
4257 | ||
4258 | static rtx | |
4259 | get_tls_get_addr (void) | |
4260 | { | |
4261 | if (!tls_get_addr_libfunc) | |
4262 | tls_get_addr_libfunc = init_one_libfunc ("__tls_get_addr"); | |
4263 | return tls_get_addr_libfunc; | |
4264 | } | |
4265 | ||
4266 | static rtx | |
4267 | arm_load_tp (rtx target) | |
4268 | { | |
4269 | if (!target) | |
4270 | target = gen_reg_rtx (SImode); | |
4271 | ||
4272 | if (TARGET_HARD_TP) | |
4273 | { | |
4274 | /* Can return in any reg. */ | |
4275 | emit_insn (gen_load_tp_hard (target)); | |
4276 | } | |
4277 | else | |
4278 | { | |
4279 | /* Always returned in r0. Immediately copy the result into a pseudo, | |
4280 | otherwise other uses of r0 (e.g. setting up function arguments) may | |
4281 | clobber the value. */ | |
4282 | ||
4283 | rtx tmp; | |
4284 | ||
4285 | emit_insn (gen_load_tp_soft ()); | |
4286 | ||
4287 | tmp = gen_rtx_REG (SImode, 0); | |
4288 | emit_move_insn (target, tmp); | |
4289 | } | |
4290 | return target; | |
4291 | } | |
4292 | ||
4293 | static rtx | |
4294 | load_tls_operand (rtx x, rtx reg) | |
4295 | { | |
4296 | rtx tmp; | |
4297 | ||
4298 | if (reg == NULL_RTX) | |
4299 | reg = gen_reg_rtx (SImode); | |
4300 | ||
4301 | tmp = gen_rtx_CONST (SImode, x); | |
4302 | ||
4303 | emit_move_insn (reg, tmp); | |
4304 | ||
4305 | return reg; | |
4306 | } | |
4307 | ||
4308 | static rtx | |
4309 | arm_call_tls_get_addr (rtx x, rtx reg, rtx *valuep, int reloc) | |
4310 | { | |
6cdcb15c | 4311 | rtx insns, label, labelno, sum; |
f655717d | 4312 | |
4313 | start_sequence (); | |
4314 | ||
6cdcb15c | 4315 | labelno = GEN_INT (pic_labelno++); |
4316 | label = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, labelno), UNSPEC_PIC_LABEL); | |
4317 | label = gen_rtx_CONST (VOIDmode, label); | |
4318 | ||
f655717d | 4319 | sum = gen_rtx_UNSPEC (Pmode, |
6cdcb15c | 4320 | gen_rtvec (4, x, GEN_INT (reloc), label, |
f655717d | 4321 | GEN_INT (TARGET_ARM ? 8 : 4)), |
4322 | UNSPEC_TLS); | |
4323 | reg = load_tls_operand (sum, reg); | |
4324 | ||
4325 | if (TARGET_ARM) | |
6cdcb15c | 4326 | emit_insn (gen_pic_add_dot_plus_eight (reg, reg, labelno)); |
25f905c2 | 4327 | else if (TARGET_THUMB2) |
4328 | { | |
4329 | rtx tmp; | |
4330 | /* Thumb-2 only allows very limited access to the PC. Calculate | |
4331 | the address in a temporary register. */ | |
4332 | tmp = gen_reg_rtx (SImode); | |
4333 | emit_insn (gen_pic_load_dot_plus_four (tmp, labelno)); | |
4334 | emit_insn (gen_addsi3(reg, reg, tmp)); | |
4335 | } | |
4336 | else /* TARGET_THUMB1 */ | |
6cdcb15c | 4337 | emit_insn (gen_pic_add_dot_plus_four (reg, reg, labelno)); |
f655717d | 4338 | |
4339 | *valuep = emit_library_call_value (get_tls_get_addr (), NULL_RTX, LCT_PURE, /* LCT_CONST? */ | |
4340 | Pmode, 1, reg, Pmode); | |
4341 | ||
4342 | insns = get_insns (); | |
4343 | end_sequence (); | |
4344 | ||
4345 | return insns; | |
4346 | } | |
4347 | ||
4348 | rtx | |
4349 | legitimize_tls_address (rtx x, rtx reg) | |
4350 | { | |
6cdcb15c | 4351 | rtx dest, tp, label, labelno, sum, insns, ret, eqv, addend; |
f655717d | 4352 | unsigned int model = SYMBOL_REF_TLS_MODEL (x); |
4353 | ||
4354 | switch (model) | |
4355 | { | |
4356 | case TLS_MODEL_GLOBAL_DYNAMIC: | |
4357 | insns = arm_call_tls_get_addr (x, reg, &ret, TLS_GD32); | |
4358 | dest = gen_reg_rtx (Pmode); | |
4359 | emit_libcall_block (insns, dest, ret, x); | |
4360 | return dest; | |
4361 | ||
4362 | case TLS_MODEL_LOCAL_DYNAMIC: | |
4363 | insns = arm_call_tls_get_addr (x, reg, &ret, TLS_LDM32); | |
4364 | ||
4365 | /* Attach a unique REG_EQUIV, to allow the RTL optimizers to | |
4366 | share the LDM result with other LD model accesses. */ | |
4367 | eqv = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const1_rtx), | |
4368 | UNSPEC_TLS); | |
4369 | dest = gen_reg_rtx (Pmode); | |
e307d75d | 4370 | emit_libcall_block (insns, dest, ret, eqv); |
f655717d | 4371 | |
4372 | /* Load the addend. */ | |
4373 | addend = gen_rtx_UNSPEC (Pmode, gen_rtvec (2, x, GEN_INT (TLS_LDO32)), | |
4374 | UNSPEC_TLS); | |
4375 | addend = force_reg (SImode, gen_rtx_CONST (SImode, addend)); | |
4376 | return gen_rtx_PLUS (Pmode, dest, addend); | |
4377 | ||
4378 | case TLS_MODEL_INITIAL_EXEC: | |
6cdcb15c | 4379 | labelno = GEN_INT (pic_labelno++); |
4380 | label = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, labelno), UNSPEC_PIC_LABEL); | |
4381 | label = gen_rtx_CONST (VOIDmode, label); | |
f655717d | 4382 | sum = gen_rtx_UNSPEC (Pmode, |
6cdcb15c | 4383 | gen_rtvec (4, x, GEN_INT (TLS_IE32), label, |
f655717d | 4384 | GEN_INT (TARGET_ARM ? 8 : 4)), |
4385 | UNSPEC_TLS); | |
4386 | reg = load_tls_operand (sum, reg); | |
4387 | ||
4388 | if (TARGET_ARM) | |
6cdcb15c | 4389 | emit_insn (gen_tls_load_dot_plus_eight (reg, reg, labelno)); |
25f905c2 | 4390 | else if (TARGET_THUMB2) |
4391 | { | |
4392 | rtx tmp; | |
4393 | /* Thumb-2 only allows very limited access to the PC. Calculate | |
4394 | the address in a temporary register. */ | |
4395 | tmp = gen_reg_rtx (SImode); | |
4396 | emit_insn (gen_pic_load_dot_plus_four (tmp, labelno)); | |
4397 | emit_insn (gen_addsi3(reg, reg, tmp)); | |
4398 | emit_move_insn (reg, gen_const_mem (SImode, reg)); | |
4399 | } | |
f655717d | 4400 | else |
4401 | { | |
6cdcb15c | 4402 | emit_insn (gen_pic_add_dot_plus_four (reg, reg, labelno)); |
f655717d | 4403 | emit_move_insn (reg, gen_const_mem (SImode, reg)); |
4404 | } | |
4405 | ||
4406 | tp = arm_load_tp (NULL_RTX); | |
4407 | ||
4408 | return gen_rtx_PLUS (Pmode, tp, reg); | |
4409 | ||
4410 | case TLS_MODEL_LOCAL_EXEC: | |
4411 | tp = arm_load_tp (NULL_RTX); | |
4412 | ||
4413 | reg = gen_rtx_UNSPEC (Pmode, | |
4414 | gen_rtvec (2, x, GEN_INT (TLS_LE32)), | |
4415 | UNSPEC_TLS); | |
4416 | reg = force_reg (SImode, gen_rtx_CONST (SImode, reg)); | |
4417 | ||
4418 | return gen_rtx_PLUS (Pmode, tp, reg); | |
4419 | ||
4420 | default: | |
4421 | abort (); | |
4422 | } | |
4423 | } | |
4424 | ||
1ff79d84 | 4425 | /* Try machine-dependent ways of modifying an illegitimate address |
4426 | to be legitimate. If we find one, return the new, valid address. */ | |
1ff79d84 | 4427 | rtx |
ebd88a36 | 4428 | arm_legitimize_address (rtx x, rtx orig_x, enum machine_mode mode) |
1ff79d84 | 4429 | { |
f655717d | 4430 | if (arm_tls_symbol_p (x)) |
4431 | return legitimize_tls_address (x, NULL_RTX); | |
4432 | ||
1ff79d84 | 4433 | if (GET_CODE (x) == PLUS) |
4434 | { | |
4435 | rtx xop0 = XEXP (x, 0); | |
4436 | rtx xop1 = XEXP (x, 1); | |
4437 | ||
4438 | if (CONSTANT_P (xop0) && !symbol_mentioned_p (xop0)) | |
4439 | xop0 = force_reg (SImode, xop0); | |
4440 | ||
4441 | if (CONSTANT_P (xop1) && !symbol_mentioned_p (xop1)) | |
4442 | xop1 = force_reg (SImode, xop1); | |
4443 | ||
4444 | if (ARM_BASE_REGISTER_RTX_P (xop0) | |
4445 | && GET_CODE (xop1) == CONST_INT) | |
4446 | { | |
4447 | HOST_WIDE_INT n, low_n; | |
4448 | rtx base_reg, val; | |
4449 | n = INTVAL (xop1); | |
4450 | ||
a2cd141b | 4451 | /* VFP addressing modes actually allow greater offsets, but for |
4452 | now we just stick with the lowest common denominator. */ | |
4453 | if (mode == DImode | |
4454 | || ((TARGET_SOFT_FLOAT || TARGET_VFP) && mode == DFmode)) | |
1ff79d84 | 4455 | { |
4456 | low_n = n & 0x0f; | |
4457 | n &= ~0x0f; | |
4458 | if (low_n > 4) | |
4459 | { | |
4460 | n += 16; | |
4461 | low_n -= 16; | |
4462 | } | |
4463 | } | |
4464 | else | |
4465 | { | |
4466 | low_n = ((mode) == TImode ? 0 | |
4467 | : n >= 0 ? (n & 0xfff) : -((-n) & 0xfff)); | |
4468 | n -= low_n; | |
4469 | } | |
4470 | ||
4471 | base_reg = gen_reg_rtx (SImode); | |
0083a3b9 | 4472 | val = force_operand (plus_constant (xop0, n), NULL_RTX); |
1ff79d84 | 4473 | emit_move_insn (base_reg, val); |
0083a3b9 | 4474 | x = plus_constant (base_reg, low_n); |
1ff79d84 | 4475 | } |
4476 | else if (xop0 != XEXP (x, 0) || xop1 != XEXP (x, 1)) | |
4477 | x = gen_rtx_PLUS (SImode, xop0, xop1); | |
4478 | } | |
4479 | ||
4480 | /* XXX We don't allow MINUS any more -- see comment in | |
4481 | arm_legitimate_address_p (). */ | |
4482 | else if (GET_CODE (x) == MINUS) | |
4483 | { | |
4484 | rtx xop0 = XEXP (x, 0); | |
4485 | rtx xop1 = XEXP (x, 1); | |
4486 | ||
4487 | if (CONSTANT_P (xop0)) | |
4488 | xop0 = force_reg (SImode, xop0); | |
4489 | ||
4490 | if (CONSTANT_P (xop1) && ! symbol_mentioned_p (xop1)) | |
4491 | xop1 = force_reg (SImode, xop1); | |
4492 | ||
4493 | if (xop0 != XEXP (x, 0) || xop1 != XEXP (x, 1)) | |
4494 | x = gen_rtx_MINUS (SImode, xop0, xop1); | |
4495 | } | |
4496 | ||
61e6ce01 | 4497 | /* Make sure to take full advantage of the pre-indexed addressing mode |
4498 | with absolute addresses which often allows for the base register to | |
4499 | be factorized for multiple adjacent memory references, and it might | |
4500 | even allows for the mini pool to be avoided entirely. */ | |
4501 | else if (GET_CODE (x) == CONST_INT && optimize > 0) | |
4502 | { | |
4503 | unsigned int bits; | |
4504 | HOST_WIDE_INT mask, base, index; | |
4505 | rtx base_reg; | |
4506 | ||
85c36fd1 | 4507 | /* ldr and ldrb can use a 12-bit index, ldrsb and the rest can only |
4508 | use a 8-bit index. So let's use a 12-bit index for SImode only and | |
61e6ce01 | 4509 | hope that arm_gen_constant will enable ldrb to use more bits. */ |
4510 | bits = (mode == SImode) ? 12 : 8; | |
4511 | mask = (1 << bits) - 1; | |
4512 | base = INTVAL (x) & ~mask; | |
4513 | index = INTVAL (x) & mask; | |
4c527b3e | 4514 | if (bit_count (base & 0xffffffff) > (32 - bits)/2) |
61e6ce01 | 4515 | { |
4516 | /* It'll most probably be more efficient to generate the base | |
4517 | with more bits set and use a negative index instead. */ | |
4518 | base |= mask; | |
4519 | index -= mask; | |
4520 | } | |
4521 | base_reg = force_reg (SImode, GEN_INT (base)); | |
0083a3b9 | 4522 | x = plus_constant (base_reg, index); |
61e6ce01 | 4523 | } |
4524 | ||
1ff79d84 | 4525 | if (flag_pic) |
4526 | { | |
4527 | /* We need to find and carefully transform any SYMBOL and LABEL | |
4528 | references; so go back to the original address expression. */ | |
4529 | rtx new_x = legitimize_pic_address (orig_x, mode, NULL_RTX); | |
4530 | ||
4531 | if (new_x != orig_x) | |
4532 | x = new_x; | |
4533 | } | |
4534 | ||
4535 | return x; | |
4536 | } | |
4537 | ||
bca653c5 | 4538 | |
4539 | /* Try machine-dependent ways of modifying an illegitimate Thumb address | |
4540 | to be legitimate. If we find one, return the new, valid address. */ | |
4541 | rtx | |
4542 | thumb_legitimize_address (rtx x, rtx orig_x, enum machine_mode mode) | |
4543 | { | |
f655717d | 4544 | if (arm_tls_symbol_p (x)) |
4545 | return legitimize_tls_address (x, NULL_RTX); | |
4546 | ||
bca653c5 | 4547 | if (GET_CODE (x) == PLUS |
4548 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
4549 | && (INTVAL (XEXP (x, 1)) >= 32 * GET_MODE_SIZE (mode) | |
4550 | || INTVAL (XEXP (x, 1)) < 0)) | |
4551 | { | |
4552 | rtx xop0 = XEXP (x, 0); | |
4553 | rtx xop1 = XEXP (x, 1); | |
4554 | HOST_WIDE_INT offset = INTVAL (xop1); | |
4555 | ||
4556 | /* Try and fold the offset into a biasing of the base register and | |
4557 | then offsetting that. Don't do this when optimizing for space | |
4558 | since it can cause too many CSEs. */ | |
4559 | if (optimize_size && offset >= 0 | |
4560 | && offset < 256 + 31 * GET_MODE_SIZE (mode)) | |
4561 | { | |
4562 | HOST_WIDE_INT delta; | |
4563 | ||
4564 | if (offset >= 256) | |
4565 | delta = offset - (256 - GET_MODE_SIZE (mode)); | |
4566 | else if (offset < 32 * GET_MODE_SIZE (mode) + 8) | |
4567 | delta = 31 * GET_MODE_SIZE (mode); | |
4568 | else | |
4569 | delta = offset & (~31 * GET_MODE_SIZE (mode)); | |
4570 | ||
4571 | xop0 = force_operand (plus_constant (xop0, offset - delta), | |
4572 | NULL_RTX); | |
4573 | x = plus_constant (xop0, delta); | |
4574 | } | |
4575 | else if (offset < 0 && offset > -256) | |
4576 | /* Small negative offsets are best done with a subtract before the | |
4577 | dereference, forcing these into a register normally takes two | |
4578 | instructions. */ | |
4579 | x = force_operand (x, NULL_RTX); | |
4580 | else | |
4581 | { | |
4582 | /* For the remaining cases, force the constant into a register. */ | |
4583 | xop1 = force_reg (SImode, xop1); | |
4584 | x = gen_rtx_PLUS (SImode, xop0, xop1); | |
4585 | } | |
4586 | } | |
4587 | else if (GET_CODE (x) == PLUS | |
4588 | && s_register_operand (XEXP (x, 1), SImode) | |
4589 | && !s_register_operand (XEXP (x, 0), SImode)) | |
4590 | { | |
4591 | rtx xop0 = force_operand (XEXP (x, 0), NULL_RTX); | |
4592 | ||
4593 | x = gen_rtx_PLUS (SImode, xop0, XEXP (x, 1)); | |
4594 | } | |
4595 | ||
4596 | if (flag_pic) | |
4597 | { | |
4598 | /* We need to find and carefully transform any SYMBOL and LABEL | |
4599 | references; so go back to the original address expression. */ | |
4600 | rtx new_x = legitimize_pic_address (orig_x, mode, NULL_RTX); | |
4601 | ||
4602 | if (new_x != orig_x) | |
4603 | x = new_x; | |
4604 | } | |
4605 | ||
4606 | return x; | |
4607 | } | |
4608 | ||
5ede62f6 | 4609 | rtx |
a60f1840 | 4610 | thumb_legitimize_reload_address (rtx *x_p, |
4611 | enum machine_mode mode, | |
4612 | int opnum, int type, | |
4613 | int ind_levels ATTRIBUTE_UNUSED) | |
5ede62f6 | 4614 | { |
4615 | rtx x = *x_p; | |
a60f1840 | 4616 | |
5ede62f6 | 4617 | if (GET_CODE (x) == PLUS |
4618 | && GET_MODE_SIZE (mode) < 4 | |
4619 | && REG_P (XEXP (x, 0)) | |
4620 | && XEXP (x, 0) == stack_pointer_rtx | |
4621 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
4622 | && !thumb_legitimate_offset_p (mode, INTVAL (XEXP (x, 1)))) | |
4623 | { | |
4624 | rtx orig_x = x; | |
4625 | ||
4626 | x = copy_rtx (x); | |
4627 | push_reload (orig_x, NULL_RTX, x_p, NULL, MODE_BASE_REG_CLASS (mode), | |
4628 | Pmode, VOIDmode, 0, 0, opnum, type); | |
4629 | return x; | |
4630 | } | |
4631 | ||
4632 | /* If both registers are hi-regs, then it's better to reload the | |
4633 | entire expression rather than each register individually. That | |
4634 | only requires one reload register rather than two. */ | |
4635 | if (GET_CODE (x) == PLUS | |
4636 | && REG_P (XEXP (x, 0)) | |
4637 | && REG_P (XEXP (x, 1)) | |
4638 | && !REG_MODE_OK_FOR_REG_BASE_P (XEXP (x, 0), mode) | |
4639 | && !REG_MODE_OK_FOR_REG_BASE_P (XEXP (x, 1), mode)) | |
4640 | { | |
4641 | rtx orig_x = x; | |
4642 | ||
4643 | x = copy_rtx (x); | |
4644 | push_reload (orig_x, NULL_RTX, x_p, NULL, MODE_BASE_REG_CLASS (mode), | |
4645 | Pmode, VOIDmode, 0, 0, opnum, type); | |
4646 | return x; | |
4647 | } | |
4648 | ||
4649 | return NULL; | |
4650 | } | |
f655717d | 4651 | |
4652 | /* Test for various thread-local symbols. */ | |
4653 | ||
4654 | /* Return TRUE if X is a thread-local symbol. */ | |
4655 | ||
4656 | static bool | |
4657 | arm_tls_symbol_p (rtx x) | |
4658 | { | |
4659 | if (! TARGET_HAVE_TLS) | |
4660 | return false; | |
4661 | ||
4662 | if (GET_CODE (x) != SYMBOL_REF) | |
4663 | return false; | |
4664 | ||
4665 | return SYMBOL_REF_TLS_MODEL (x) != 0; | |
4666 | } | |
4667 | ||
4668 | /* Helper for arm_tls_referenced_p. */ | |
4669 | ||
4670 | static int | |
4671 | arm_tls_operand_p_1 (rtx *x, void *data ATTRIBUTE_UNUSED) | |
4672 | { | |
4673 | if (GET_CODE (*x) == SYMBOL_REF) | |
4674 | return SYMBOL_REF_TLS_MODEL (*x) != 0; | |
4675 | ||
4676 | /* Don't recurse into UNSPEC_TLS looking for TLS symbols; these are | |
4677 | TLS offsets, not real symbol references. */ | |
4678 | if (GET_CODE (*x) == UNSPEC | |
4679 | && XINT (*x, 1) == UNSPEC_TLS) | |
4680 | return -1; | |
4681 | ||
4682 | return 0; | |
4683 | } | |
4684 | ||
4685 | /* Return TRUE if X contains any TLS symbol references. */ | |
4686 | ||
4687 | bool | |
4688 | arm_tls_referenced_p (rtx x) | |
4689 | { | |
4690 | if (! TARGET_HAVE_TLS) | |
4691 | return false; | |
4692 | ||
4693 | return for_each_rtx (&x, arm_tls_operand_p_1, NULL); | |
4694 | } | |
e348ff3e | 4695 | |
4696 | /* Implement TARGET_CANNOT_FORCE_CONST_MEM. */ | |
4697 | ||
4698 | bool | |
4699 | arm_cannot_force_const_mem (rtx x) | |
4700 | { | |
4701 | rtx base, offset; | |
4702 | ||
4703 | if (ARM_OFFSETS_MUST_BE_WITHIN_SECTIONS_P) | |
4704 | { | |
4705 | split_const (x, &base, &offset); | |
4706 | if (GET_CODE (base) == SYMBOL_REF | |
4707 | && !offset_within_block_p (base, INTVAL (offset))) | |
4708 | return true; | |
4709 | } | |
4710 | return arm_tls_referenced_p (x); | |
4711 | } | |
d32b3557 | 4712 | \f |
4b6897e9 | 4713 | #define REG_OR_SUBREG_REG(X) \ |
4714 | (GET_CODE (X) == REG \ | |
4715 | || (GET_CODE (X) == SUBREG && GET_CODE (SUBREG_REG (X)) == REG)) | |
4716 | ||
4717 | #define REG_OR_SUBREG_RTX(X) \ | |
4718 | (GET_CODE (X) == REG ? (X) : SUBREG_REG (X)) | |
4719 | ||
cffb2a26 | 4720 | #ifndef COSTS_N_INSNS |
4721 | #define COSTS_N_INSNS(N) ((N) * 4 - 2) | |
4722 | #endif | |
fab7adbf | 4723 | static inline int |
25f905c2 | 4724 | thumb1_rtx_costs (rtx x, enum rtx_code code, enum rtx_code outer) |
4b6897e9 | 4725 | { |
4726 | enum machine_mode mode = GET_MODE (x); | |
4b6897e9 | 4727 | |
a2cd141b | 4728 | switch (code) |
cffb2a26 | 4729 | { |
a2cd141b | 4730 | case ASHIFT: |
4731 | case ASHIFTRT: | |
4732 | case LSHIFTRT: | |
9e7454d0 | 4733 | case ROTATERT: |
a2cd141b | 4734 | case PLUS: |
4735 | case MINUS: | |
4736 | case COMPARE: | |
4737 | case NEG: | |
9e7454d0 | 4738 | case NOT: |
a2cd141b | 4739 | return COSTS_N_INSNS (1); |
9e7454d0 | 4740 | |
4741 | case MULT: | |
4742 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
4743 | { | |
4744 | int cycles = 0; | |
a2cd141b | 4745 | unsigned HOST_WIDE_INT i = INTVAL (XEXP (x, 1)); |
9e7454d0 | 4746 | |
4747 | while (i) | |
4748 | { | |
4749 | i >>= 2; | |
4750 | cycles++; | |
4751 | } | |
4752 | return COSTS_N_INSNS (2) + cycles; | |
a2cd141b | 4753 | } |
4754 | return COSTS_N_INSNS (1) + 16; | |
9e7454d0 | 4755 | |
4756 | case SET: | |
4757 | return (COSTS_N_INSNS (1) | |
4758 | + 4 * ((GET_CODE (SET_SRC (x)) == MEM) | |
a2cd141b | 4759 | + GET_CODE (SET_DEST (x)) == MEM)); |
9e7454d0 | 4760 | |
4761 | case CONST_INT: | |
4762 | if (outer == SET) | |
4763 | { | |
4764 | if ((unsigned HOST_WIDE_INT) INTVAL (x) < 256) | |
4765 | return 0; | |
4766 | if (thumb_shiftable_const (INTVAL (x))) | |
4767 | return COSTS_N_INSNS (2); | |
4768 | return COSTS_N_INSNS (3); | |
4769 | } | |
a2cd141b | 4770 | else if ((outer == PLUS || outer == COMPARE) |
9e7454d0 | 4771 | && INTVAL (x) < 256 && INTVAL (x) > -256) |
a2cd141b | 4772 | return 0; |
4773 | else if (outer == AND | |
4774 | && INTVAL (x) < 256 && INTVAL (x) >= -256) | |
4775 | return COSTS_N_INSNS (1); | |
9e7454d0 | 4776 | else if (outer == ASHIFT || outer == ASHIFTRT |
4777 | || outer == LSHIFTRT) | |
4778 | return 0; | |
a2cd141b | 4779 | return COSTS_N_INSNS (2); |
9e7454d0 | 4780 | |
4781 | case CONST: | |
4782 | case CONST_DOUBLE: | |
4783 | case LABEL_REF: | |
4784 | case SYMBOL_REF: | |
a2cd141b | 4785 | return COSTS_N_INSNS (3); |
9e7454d0 | 4786 | |
a2cd141b | 4787 | case UDIV: |
4788 | case UMOD: | |
4789 | case DIV: | |
4790 | case MOD: | |
4791 | return 100; | |
cffb2a26 | 4792 | |
a2cd141b | 4793 | case TRUNCATE: |
4794 | return 99; | |
cffb2a26 | 4795 | |
a2cd141b | 4796 | case AND: |
4797 | case XOR: | |
9e7454d0 | 4798 | case IOR: |
1d60d981 | 4799 | /* XXX guess. */ |
a2cd141b | 4800 | return 8; |
cffb2a26 | 4801 | |
a2cd141b | 4802 | case MEM: |
4803 | /* XXX another guess. */ | |
4804 | /* Memory costs quite a lot for the first word, but subsequent words | |
4805 | load at the equivalent of a single insn each. */ | |
4806 | return (10 + 4 * ((GET_MODE_SIZE (mode) - 1) / UNITS_PER_WORD) | |
4807 | + ((GET_CODE (x) == SYMBOL_REF && CONSTANT_POOL_ADDRESS_P (x)) | |
4808 | ? 4 : 0)); | |
4809 | ||
4810 | case IF_THEN_ELSE: | |
1d60d981 | 4811 | /* XXX a guess. */ |
a2cd141b | 4812 | if (GET_CODE (XEXP (x, 1)) == PC || GET_CODE (XEXP (x, 2)) == PC) |
4813 | return 14; | |
4814 | return 2; | |
4815 | ||
4816 | case ZERO_EXTEND: | |
4817 | /* XXX still guessing. */ | |
4818 | switch (GET_MODE (XEXP (x, 0))) | |
4819 | { | |
4820 | case QImode: | |
4821 | return (1 + (mode == DImode ? 4 : 0) | |
4822 | + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0)); | |
9e7454d0 | 4823 | |
a2cd141b | 4824 | case HImode: |
4825 | return (4 + (mode == DImode ? 4 : 0) | |
4826 | + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0)); | |
9e7454d0 | 4827 | |
a2cd141b | 4828 | case SImode: |
4829 | return (1 + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0)); | |
9e7454d0 | 4830 | |
cffb2a26 | 4831 | default: |
4832 | return 99; | |
cffb2a26 | 4833 | } |
9e7454d0 | 4834 | |
a2cd141b | 4835 | default: |
4836 | return 99; | |
cffb2a26 | 4837 | } |
a2cd141b | 4838 | } |
4839 | ||
4840 | ||
4841 | /* Worker routine for arm_rtx_costs. */ | |
25f905c2 | 4842 | /* ??? This needs updating for thumb2. */ |
a2cd141b | 4843 | static inline int |
4844 | arm_rtx_costs_1 (rtx x, enum rtx_code code, enum rtx_code outer) | |
4845 | { | |
4846 | enum machine_mode mode = GET_MODE (x); | |
4847 | enum rtx_code subcode; | |
4848 | int extra_cost; | |
4849 | ||
4b6897e9 | 4850 | switch (code) |
4851 | { | |
4852 | case MEM: | |
4853 | /* Memory costs quite a lot for the first word, but subsequent words | |
4854 | load at the equivalent of a single insn each. */ | |
4855 | return (10 + 4 * ((GET_MODE_SIZE (mode) - 1) / UNITS_PER_WORD) | |
3ccc58a6 | 4856 | + (GET_CODE (x) == SYMBOL_REF |
4857 | && CONSTANT_POOL_ADDRESS_P (x) ? 4 : 0)); | |
4b6897e9 | 4858 | |
4859 | case DIV: | |
4860 | case MOD: | |
6b784824 | 4861 | case UDIV: |
4862 | case UMOD: | |
4863 | return optimize_size ? COSTS_N_INSNS (2) : 100; | |
4b6897e9 | 4864 | |
4865 | case ROTATE: | |
4866 | if (mode == SImode && GET_CODE (XEXP (x, 1)) == REG) | |
4867 | return 4; | |
4868 | /* Fall through */ | |
4869 | case ROTATERT: | |
4870 | if (mode != SImode) | |
4871 | return 8; | |
4872 | /* Fall through */ | |
4873 | case ASHIFT: case LSHIFTRT: case ASHIFTRT: | |
4874 | if (mode == DImode) | |
4875 | return (8 + (GET_CODE (XEXP (x, 1)) == CONST_INT ? 0 : 8) | |
9e7454d0 | 4876 | + ((GET_CODE (XEXP (x, 0)) == REG |
4b6897e9 | 4877 | || (GET_CODE (XEXP (x, 0)) == SUBREG |
4878 | && GET_CODE (SUBREG_REG (XEXP (x, 0))) == REG)) | |
4879 | ? 0 : 8)); | |
4880 | return (1 + ((GET_CODE (XEXP (x, 0)) == REG | |
4881 | || (GET_CODE (XEXP (x, 0)) == SUBREG | |
4882 | && GET_CODE (SUBREG_REG (XEXP (x, 0))) == REG)) | |
4883 | ? 0 : 4) | |
4884 | + ((GET_CODE (XEXP (x, 1)) == REG | |
4885 | || (GET_CODE (XEXP (x, 1)) == SUBREG | |
4886 | && GET_CODE (SUBREG_REG (XEXP (x, 1))) == REG) | |
4887 | || (GET_CODE (XEXP (x, 1)) == CONST_INT)) | |
4888 | ? 0 : 4)); | |
4889 | ||
4890 | case MINUS: | |
89545238 | 4891 | if (GET_CODE (XEXP (x, 1)) == MULT && mode == SImode && arm_arch_thumb2) |
4892 | { | |
4893 | extra_cost = rtx_cost (XEXP (x, 1), code); | |
4894 | if (!REG_OR_SUBREG_REG (XEXP (x, 0))) | |
4895 | extra_cost += 4 * ARM_NUM_REGS (mode); | |
4896 | return extra_cost; | |
4897 | } | |
4898 | ||
4b6897e9 | 4899 | if (mode == DImode) |
4900 | return (4 + (REG_OR_SUBREG_REG (XEXP (x, 1)) ? 0 : 8) | |
4901 | + ((REG_OR_SUBREG_REG (XEXP (x, 0)) | |
4902 | || (GET_CODE (XEXP (x, 0)) == CONST_INT | |
4903 | && const_ok_for_arm (INTVAL (XEXP (x, 0))))) | |
4904 | ? 0 : 8)); | |
4905 | ||
4906 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
4907 | return (2 + ((REG_OR_SUBREG_REG (XEXP (x, 1)) | |
4908 | || (GET_CODE (XEXP (x, 1)) == CONST_DOUBLE | |
a2cd141b | 4909 | && arm_const_double_rtx (XEXP (x, 1)))) |
4b6897e9 | 4910 | ? 0 : 8) |
4911 | + ((REG_OR_SUBREG_REG (XEXP (x, 0)) | |
4912 | || (GET_CODE (XEXP (x, 0)) == CONST_DOUBLE | |
a2cd141b | 4913 | && arm_const_double_rtx (XEXP (x, 0)))) |
4b6897e9 | 4914 | ? 0 : 8)); |
4915 | ||
4916 | if (((GET_CODE (XEXP (x, 0)) == CONST_INT | |
4917 | && const_ok_for_arm (INTVAL (XEXP (x, 0))) | |
4918 | && REG_OR_SUBREG_REG (XEXP (x, 1)))) | |
4919 | || (((subcode = GET_CODE (XEXP (x, 1))) == ASHIFT | |
4920 | || subcode == ASHIFTRT || subcode == LSHIFTRT | |
4921 | || subcode == ROTATE || subcode == ROTATERT | |
4922 | || (subcode == MULT | |
4923 | && GET_CODE (XEXP (XEXP (x, 1), 1)) == CONST_INT | |
4924 | && ((INTVAL (XEXP (XEXP (x, 1), 1)) & | |
4925 | (INTVAL (XEXP (XEXP (x, 1), 1)) - 1)) == 0))) | |
4926 | && REG_OR_SUBREG_REG (XEXP (XEXP (x, 1), 0)) | |
4927 | && (REG_OR_SUBREG_REG (XEXP (XEXP (x, 1), 1)) | |
4928 | || GET_CODE (XEXP (XEXP (x, 1), 1)) == CONST_INT) | |
4929 | && REG_OR_SUBREG_REG (XEXP (x, 0)))) | |
4930 | return 1; | |
4931 | /* Fall through */ | |
4932 | ||
9e7454d0 | 4933 | case PLUS: |
bd7a45c4 | 4934 | if (GET_CODE (XEXP (x, 0)) == MULT) |
4935 | { | |
4936 | extra_cost = rtx_cost (XEXP (x, 0), code); | |
4937 | if (!REG_OR_SUBREG_REG (XEXP (x, 1))) | |
4938 | extra_cost += 4 * ARM_NUM_REGS (mode); | |
4939 | return extra_cost; | |
4940 | } | |
4941 | ||
4b6897e9 | 4942 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
4943 | return (2 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 8) | |
4944 | + ((REG_OR_SUBREG_REG (XEXP (x, 1)) | |
4945 | || (GET_CODE (XEXP (x, 1)) == CONST_DOUBLE | |
a2cd141b | 4946 | && arm_const_double_rtx (XEXP (x, 1)))) |
4b6897e9 | 4947 | ? 0 : 8)); |
4948 | ||
4949 | /* Fall through */ | |
9e7454d0 | 4950 | case AND: case XOR: case IOR: |
4b6897e9 | 4951 | extra_cost = 0; |
4952 | ||
4953 | /* Normally the frame registers will be spilt into reg+const during | |
4954 | reload, so it is a bad idea to combine them with other instructions, | |
4955 | since then they might not be moved outside of loops. As a compromise | |
4956 | we allow integration with ops that have a constant as their second | |
4957 | operand. */ | |
4958 | if ((REG_OR_SUBREG_REG (XEXP (x, 0)) | |
4959 | && ARM_FRAME_RTX (REG_OR_SUBREG_RTX (XEXP (x, 0))) | |
4960 | && GET_CODE (XEXP (x, 1)) != CONST_INT) | |
4961 | || (REG_OR_SUBREG_REG (XEXP (x, 0)) | |
4962 | && ARM_FRAME_RTX (REG_OR_SUBREG_RTX (XEXP (x, 0))))) | |
4963 | extra_cost = 4; | |
4964 | ||
4965 | if (mode == DImode) | |
4966 | return (4 + extra_cost + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 8) | |
4967 | + ((REG_OR_SUBREG_REG (XEXP (x, 1)) | |
4968 | || (GET_CODE (XEXP (x, 1)) == CONST_INT | |
016696d4 | 4969 | && const_ok_for_op (INTVAL (XEXP (x, 1)), code))) |
4b6897e9 | 4970 | ? 0 : 8)); |
4971 | ||
4972 | if (REG_OR_SUBREG_REG (XEXP (x, 0))) | |
4973 | return (1 + (GET_CODE (XEXP (x, 1)) == CONST_INT ? 0 : extra_cost) | |
4974 | + ((REG_OR_SUBREG_REG (XEXP (x, 1)) | |
4975 | || (GET_CODE (XEXP (x, 1)) == CONST_INT | |
016696d4 | 4976 | && const_ok_for_op (INTVAL (XEXP (x, 1)), code))) |
4b6897e9 | 4977 | ? 0 : 4)); |
4978 | ||
4979 | else if (REG_OR_SUBREG_REG (XEXP (x, 1))) | |
4980 | return (1 + extra_cost | |
4981 | + ((((subcode = GET_CODE (XEXP (x, 0))) == ASHIFT | |
4982 | || subcode == LSHIFTRT || subcode == ASHIFTRT | |
4983 | || subcode == ROTATE || subcode == ROTATERT | |
4984 | || (subcode == MULT | |
4985 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
4986 | && ((INTVAL (XEXP (XEXP (x, 0), 1)) & | |
ecd3fcc7 | 4987 | (INTVAL (XEXP (XEXP (x, 0), 1)) - 1)) == 0))) |
4b6897e9 | 4988 | && (REG_OR_SUBREG_REG (XEXP (XEXP (x, 0), 0))) |
4989 | && ((REG_OR_SUBREG_REG (XEXP (XEXP (x, 0), 1))) | |
ecd3fcc7 | 4990 | || GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)) |
4b6897e9 | 4991 | ? 0 : 4)); |
4992 | ||
4993 | return 8; | |
4994 | ||
4995 | case MULT: | |
a2cd141b | 4996 | /* This should have been handled by the CPU specific routines. */ |
ed29c566 | 4997 | gcc_unreachable (); |
4b6897e9 | 4998 | |
6019008a | 4999 | case TRUNCATE: |
a2cd141b | 5000 | if (arm_arch3m && mode == SImode |
6019008a | 5001 | && GET_CODE (XEXP (x, 0)) == LSHIFTRT |
5002 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT | |
5003 | && (GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) | |
5004 | == GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1))) | |
5005 | && (GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == ZERO_EXTEND | |
5006 | || GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == SIGN_EXTEND)) | |
5007 | return 8; | |
5008 | return 99; | |
5009 | ||
4b6897e9 | 5010 | case NEG: |
5011 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
5012 | return 4 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 6); | |
5013 | /* Fall through */ | |
5014 | case NOT: | |
5015 | if (mode == DImode) | |
5016 | return 4 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 4); | |
5017 | ||
5018 | return 1 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 4); | |
5019 | ||
5020 | case IF_THEN_ELSE: | |
5021 | if (GET_CODE (XEXP (x, 1)) == PC || GET_CODE (XEXP (x, 2)) == PC) | |
5022 | return 14; | |
5023 | return 2; | |
5024 | ||
5025 | case COMPARE: | |
5026 | return 1; | |
5027 | ||
5028 | case ABS: | |
5029 | return 4 + (mode == DImode ? 4 : 0); | |
5030 | ||
5031 | case SIGN_EXTEND: | |
25f905c2 | 5032 | /* ??? value extensions are cheaper on armv6. */ |
4b6897e9 | 5033 | if (GET_MODE (XEXP (x, 0)) == QImode) |
5034 | return (4 + (mode == DImode ? 4 : 0) | |
5035 | + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0)); | |
5036 | /* Fall through */ | |
5037 | case ZERO_EXTEND: | |
5038 | switch (GET_MODE (XEXP (x, 0))) | |
5039 | { | |
5040 | case QImode: | |
5041 | return (1 + (mode == DImode ? 4 : 0) | |
5042 | + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0)); | |
5043 | ||
5044 | case HImode: | |
5045 | return (4 + (mode == DImode ? 4 : 0) | |
5046 | + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0)); | |
5047 | ||
5048 | case SImode: | |
5049 | return (1 + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0)); | |
ecd3fcc7 | 5050 | |
755eb2b4 | 5051 | case V8QImode: |
5052 | case V4HImode: | |
5053 | case V2SImode: | |
5054 | case V4QImode: | |
5055 | case V2HImode: | |
5056 | return 1; | |
5057 | ||
ecd3fcc7 | 5058 | default: |
ed29c566 | 5059 | gcc_unreachable (); |
4b6897e9 | 5060 | } |
ed29c566 | 5061 | gcc_unreachable (); |
4b6897e9 | 5062 | |
9e7454d0 | 5063 | case CONST_INT: |
5064 | if (const_ok_for_arm (INTVAL (x))) | |
5065 | return outer == SET ? 2 : -1; | |
5066 | else if (outer == AND | |
5067 | && const_ok_for_arm (~INTVAL (x))) | |
5068 | return -1; | |
5069 | else if ((outer == COMPARE | |
5070 | || outer == PLUS || outer == MINUS) | |
5071 | && const_ok_for_arm (-INTVAL (x))) | |
5072 | return -1; | |
5073 | else | |
cffb2a26 | 5074 | return 5; |
9e7454d0 | 5075 | |
5076 | case CONST: | |
5077 | case LABEL_REF: | |
5078 | case SYMBOL_REF: | |
cffb2a26 | 5079 | return 6; |
9e7454d0 | 5080 | |
5081 | case CONST_DOUBLE: | |
b5a0636d | 5082 | if (arm_const_double_rtx (x) || vfp3_const_double_rtx (x)) |
9e7454d0 | 5083 | return outer == SET ? 2 : -1; |
5084 | else if ((outer == COMPARE || outer == PLUS) | |
5085 | && neg_const_double_rtx_ok_for_fpa (x)) | |
5086 | return -1; | |
cffb2a26 | 5087 | return 7; |
9e7454d0 | 5088 | |
4b6897e9 | 5089 | default: |
5090 | return 99; | |
5091 | } | |
5092 | } | |
69120d26 | 5093 | |
c1877da8 | 5094 | /* RTX costs when optimizing for size. */ |
5095 | static bool | |
5096 | arm_size_rtx_costs (rtx x, int code, int outer_code, int *total) | |
5097 | { | |
5098 | enum machine_mode mode = GET_MODE (x); | |
5099 | ||
5100 | if (TARGET_THUMB) | |
5101 | { | |
5102 | /* XXX TBD. For now, use the standard costs. */ | |
25f905c2 | 5103 | *total = thumb1_rtx_costs (x, code, outer_code); |
c1877da8 | 5104 | return true; |
5105 | } | |
5106 | ||
5107 | switch (code) | |
5108 | { | |
5109 | case MEM: | |
9e7454d0 | 5110 | /* A memory access costs 1 insn if the mode is small, or the address is |
c1877da8 | 5111 | a single register, otherwise it costs one insn per word. */ |
5112 | if (REG_P (XEXP (x, 0))) | |
5113 | *total = COSTS_N_INSNS (1); | |
5114 | else | |
5115 | *total = COSTS_N_INSNS (ARM_NUM_REGS (mode)); | |
5116 | return true; | |
5117 | ||
5118 | case DIV: | |
5119 | case MOD: | |
5120 | case UDIV: | |
5121 | case UMOD: | |
5122 | /* Needs a libcall, so it costs about this. */ | |
5123 | *total = COSTS_N_INSNS (2); | |
5124 | return false; | |
5125 | ||
5126 | case ROTATE: | |
5127 | if (mode == SImode && GET_CODE (XEXP (x, 1)) == REG) | |
5128 | { | |
5129 | *total = COSTS_N_INSNS (2) + rtx_cost (XEXP (x, 0), code); | |
5130 | return true; | |
5131 | } | |
5132 | /* Fall through */ | |
5133 | case ROTATERT: | |
5134 | case ASHIFT: | |
5135 | case LSHIFTRT: | |
5136 | case ASHIFTRT: | |
5137 | if (mode == DImode && GET_CODE (XEXP (x, 1)) == CONST_INT) | |
5138 | { | |
5139 | *total = COSTS_N_INSNS (3) + rtx_cost (XEXP (x, 0), code); | |
5140 | return true; | |
5141 | } | |
5142 | else if (mode == SImode) | |
5143 | { | |
5144 | *total = COSTS_N_INSNS (1) + rtx_cost (XEXP (x, 0), code); | |
5145 | /* Slightly disparage register shifts, but not by much. */ | |
5146 | if (GET_CODE (XEXP (x, 1)) != CONST_INT) | |
5147 | *total += 1 + rtx_cost (XEXP (x, 1), code); | |
5148 | return true; | |
5149 | } | |
5150 | ||
5151 | /* Needs a libcall. */ | |
5152 | *total = COSTS_N_INSNS (2); | |
5153 | return false; | |
5154 | ||
5155 | case MINUS: | |
5156 | if (TARGET_HARD_FLOAT && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
5157 | { | |
5158 | *total = COSTS_N_INSNS (1); | |
5159 | return false; | |
5160 | } | |
5161 | ||
5162 | if (mode == SImode) | |
5163 | { | |
5164 | enum rtx_code subcode0 = GET_CODE (XEXP (x, 0)); | |
5165 | enum rtx_code subcode1 = GET_CODE (XEXP (x, 1)); | |
5166 | ||
5167 | if (subcode0 == ROTATE || subcode0 == ROTATERT || subcode0 == ASHIFT | |
5168 | || subcode0 == LSHIFTRT || subcode0 == ASHIFTRT | |
5169 | || subcode1 == ROTATE || subcode1 == ROTATERT | |
5170 | || subcode1 == ASHIFT || subcode1 == LSHIFTRT | |
5171 | || subcode1 == ASHIFTRT) | |
5172 | { | |
5173 | /* It's just the cost of the two operands. */ | |
5174 | *total = 0; | |
5175 | return false; | |
5176 | } | |
5177 | ||
5178 | *total = COSTS_N_INSNS (1); | |
5179 | return false; | |
5180 | } | |
5181 | ||
5182 | *total = COSTS_N_INSNS (ARM_NUM_REGS (mode)); | |
5183 | return false; | |
5184 | ||
9e7454d0 | 5185 | case PLUS: |
c1877da8 | 5186 | if (TARGET_HARD_FLOAT && GET_MODE_CLASS (mode) == MODE_FLOAT) |
5187 | { | |
5188 | *total = COSTS_N_INSNS (1); | |
5189 | return false; | |
5190 | } | |
5191 | ||
5192 | /* Fall through */ | |
5193 | case AND: case XOR: case IOR: | |
5194 | if (mode == SImode) | |
5195 | { | |
5196 | enum rtx_code subcode = GET_CODE (XEXP (x, 0)); | |
5197 | ||
5198 | if (subcode == ROTATE || subcode == ROTATERT || subcode == ASHIFT | |
5199 | || subcode == LSHIFTRT || subcode == ASHIFTRT | |
5200 | || (code == AND && subcode == NOT)) | |
5201 | { | |
5202 | /* It's just the cost of the two operands. */ | |
5203 | *total = 0; | |
5204 | return false; | |
5205 | } | |
5206 | } | |
5207 | ||
5208 | *total = COSTS_N_INSNS (ARM_NUM_REGS (mode)); | |
5209 | return false; | |
5210 | ||
5211 | case MULT: | |
5212 | *total = COSTS_N_INSNS (ARM_NUM_REGS (mode)); | |
5213 | return false; | |
5214 | ||
5215 | case NEG: | |
5216 | if (TARGET_HARD_FLOAT && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
5217 | *total = COSTS_N_INSNS (1); | |
5218 | /* Fall through */ | |
5219 | case NOT: | |
5220 | *total = COSTS_N_INSNS (ARM_NUM_REGS (mode)); | |
5221 | ||
5222 | return false; | |
5223 | ||
5224 | case IF_THEN_ELSE: | |
5225 | *total = 0; | |
5226 | return false; | |
5227 | ||
5228 | case COMPARE: | |
5229 | if (cc_register (XEXP (x, 0), VOIDmode)) | |
5230 | * total = 0; | |
5231 | else | |
5232 | *total = COSTS_N_INSNS (1); | |
5233 | return false; | |
5234 | ||
5235 | case ABS: | |
5236 | if (TARGET_HARD_FLOAT && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
5237 | *total = COSTS_N_INSNS (1); | |
5238 | else | |
5239 | *total = COSTS_N_INSNS (1 + ARM_NUM_REGS (mode)); | |
5240 | return false; | |
5241 | ||
5242 | case SIGN_EXTEND: | |
5243 | *total = 0; | |
5244 | if (GET_MODE_SIZE (GET_MODE (XEXP (x, 0))) < 4) | |
5245 | { | |
5246 | if (!(arm_arch4 && MEM_P (XEXP (x, 0)))) | |
5247 | *total += COSTS_N_INSNS (arm_arch6 ? 1 : 2); | |
5248 | } | |
5249 | if (mode == DImode) | |
5250 | *total += COSTS_N_INSNS (1); | |
5251 | return false; | |
5252 | ||
5253 | case ZERO_EXTEND: | |
5254 | *total = 0; | |
5255 | if (!(arm_arch4 && MEM_P (XEXP (x, 0)))) | |
5256 | { | |
5257 | switch (GET_MODE (XEXP (x, 0))) | |
5258 | { | |
5259 | case QImode: | |
5260 | *total += COSTS_N_INSNS (1); | |
5261 | break; | |
5262 | ||
5263 | case HImode: | |
5264 | *total += COSTS_N_INSNS (arm_arch6 ? 1 : 2); | |
9e7454d0 | 5265 | |
c1877da8 | 5266 | case SImode: |
5267 | break; | |
5268 | ||
5269 | default: | |
5270 | *total += COSTS_N_INSNS (2); | |
5271 | } | |
5272 | } | |
5273 | ||
5274 | if (mode == DImode) | |
5275 | *total += COSTS_N_INSNS (1); | |
5276 | ||
5277 | return false; | |
5278 | ||
9e7454d0 | 5279 | case CONST_INT: |
5280 | if (const_ok_for_arm (INTVAL (x))) | |
c1877da8 | 5281 | *total = COSTS_N_INSNS (outer_code == SET ? 1 : 0); |
5282 | else if (const_ok_for_arm (~INTVAL (x))) | |
5283 | *total = COSTS_N_INSNS (outer_code == AND ? 0 : 1); | |
5284 | else if (const_ok_for_arm (-INTVAL (x))) | |
5285 | { | |
5286 | if (outer_code == COMPARE || outer_code == PLUS | |
5287 | || outer_code == MINUS) | |
5288 | *total = 0; | |
5289 | else | |
5290 | *total = COSTS_N_INSNS (1); | |
5291 | } | |
5292 | else | |
5293 | *total = COSTS_N_INSNS (2); | |
5294 | return true; | |
9e7454d0 | 5295 | |
5296 | case CONST: | |
5297 | case LABEL_REF: | |
5298 | case SYMBOL_REF: | |
c1877da8 | 5299 | *total = COSTS_N_INSNS (2); |
5300 | return true; | |
9e7454d0 | 5301 | |
c1877da8 | 5302 | case CONST_DOUBLE: |
5303 | *total = COSTS_N_INSNS (4); | |
5304 | return true; | |
5305 | ||
5306 | default: | |
5307 | if (mode != VOIDmode) | |
5308 | *total = COSTS_N_INSNS (ARM_NUM_REGS (mode)); | |
5309 | else | |
5310 | *total = COSTS_N_INSNS (4); /* How knows? */ | |
5311 | return false; | |
5312 | } | |
5313 | } | |
5314 | ||
25f905c2 | 5315 | /* RTX costs for cores with a slow MUL implementation. Thumb-2 is not |
5316 | supported on any "slowmul" cores, so it can be ignored. */ | |
a2cd141b | 5317 | |
fab7adbf | 5318 | static bool |
a2cd141b | 5319 | arm_slowmul_rtx_costs (rtx x, int code, int outer_code, int *total) |
fab7adbf | 5320 | { |
a2cd141b | 5321 | enum machine_mode mode = GET_MODE (x); |
5322 | ||
5323 | if (TARGET_THUMB) | |
5324 | { | |
25f905c2 | 5325 | *total = thumb1_rtx_costs (x, code, outer_code); |
a2cd141b | 5326 | return true; |
5327 | } | |
9e7454d0 | 5328 | |
a2cd141b | 5329 | switch (code) |
5330 | { | |
5331 | case MULT: | |
5332 | if (GET_MODE_CLASS (mode) == MODE_FLOAT | |
5333 | || mode == DImode) | |
5334 | { | |
5335 | *total = 30; | |
5336 | return true; | |
5337 | } | |
5338 | ||
5339 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
5340 | { | |
5341 | unsigned HOST_WIDE_INT i = (INTVAL (XEXP (x, 1)) | |
5342 | & (unsigned HOST_WIDE_INT) 0xffffffff); | |
5343 | int cost, const_ok = const_ok_for_arm (i); | |
5344 | int j, booth_unit_size; | |
5345 | ||
9e7454d0 | 5346 | /* Tune as appropriate. */ |
a2cd141b | 5347 | cost = const_ok ? 4 : 8; |
5348 | booth_unit_size = 2; | |
5349 | for (j = 0; i && j < 32; j += booth_unit_size) | |
5350 | { | |
5351 | i >>= booth_unit_size; | |
5352 | cost += 2; | |
5353 | } | |
5354 | ||
5355 | *total = cost; | |
5356 | return true; | |
5357 | } | |
5358 | ||
5359 | *total = 30 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 4) | |
5360 | + (REG_OR_SUBREG_REG (XEXP (x, 1)) ? 0 : 4); | |
5361 | return true; | |
9e7454d0 | 5362 | |
a2cd141b | 5363 | default: |
5364 | *total = arm_rtx_costs_1 (x, code, outer_code); | |
5365 | return true; | |
5366 | } | |
fab7adbf | 5367 | } |
5368 | ||
a2cd141b | 5369 | |
5370 | /* RTX cost for cores with a fast multiply unit (M variants). */ | |
5371 | ||
5372 | static bool | |
5373 | arm_fastmul_rtx_costs (rtx x, int code, int outer_code, int *total) | |
5374 | { | |
5375 | enum machine_mode mode = GET_MODE (x); | |
5376 | ||
25f905c2 | 5377 | if (TARGET_THUMB1) |
a2cd141b | 5378 | { |
25f905c2 | 5379 | *total = thumb1_rtx_costs (x, code, outer_code); |
a2cd141b | 5380 | return true; |
5381 | } | |
9e7454d0 | 5382 | |
25f905c2 | 5383 | /* ??? should thumb2 use different costs? */ |
a2cd141b | 5384 | switch (code) |
5385 | { | |
5386 | case MULT: | |
5387 | /* There is no point basing this on the tuning, since it is always the | |
5388 | fast variant if it exists at all. */ | |
5389 | if (mode == DImode | |
5390 | && (GET_CODE (XEXP (x, 0)) == GET_CODE (XEXP (x, 1))) | |
5391 | && (GET_CODE (XEXP (x, 0)) == ZERO_EXTEND | |
5392 | || GET_CODE (XEXP (x, 0)) == SIGN_EXTEND)) | |
5393 | { | |
5394 | *total = 8; | |
5395 | return true; | |
5396 | } | |
9e7454d0 | 5397 | |
a2cd141b | 5398 | |
5399 | if (GET_MODE_CLASS (mode) == MODE_FLOAT | |
5400 | || mode == DImode) | |
5401 | { | |
5402 | *total = 30; | |
5403 | return true; | |
5404 | } | |
5405 | ||
5406 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
5407 | { | |
5408 | unsigned HOST_WIDE_INT i = (INTVAL (XEXP (x, 1)) | |
5409 | & (unsigned HOST_WIDE_INT) 0xffffffff); | |
5410 | int cost, const_ok = const_ok_for_arm (i); | |
5411 | int j, booth_unit_size; | |
5412 | ||
9e7454d0 | 5413 | /* Tune as appropriate. */ |
a2cd141b | 5414 | cost = const_ok ? 4 : 8; |
5415 | booth_unit_size = 8; | |
5416 | for (j = 0; i && j < 32; j += booth_unit_size) | |
5417 | { | |
5418 | i >>= booth_unit_size; | |
5419 | cost += 2; | |
5420 | } | |
5421 | ||
5422 | *total = cost; | |
5423 | return true; | |
5424 | } | |
5425 | ||
5426 | *total = 8 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 4) | |
5427 | + (REG_OR_SUBREG_REG (XEXP (x, 1)) ? 0 : 4); | |
5428 | return true; | |
9e7454d0 | 5429 | |
a2cd141b | 5430 | default: |
5431 | *total = arm_rtx_costs_1 (x, code, outer_code); | |
5432 | return true; | |
5433 | } | |
5434 | } | |
5435 | ||
5436 | ||
25f905c2 | 5437 | /* RTX cost for XScale CPUs. Thumb-2 is not supported on any xscale cores, |
5438 | so it can be ignored. */ | |
a2cd141b | 5439 | |
5440 | static bool | |
5441 | arm_xscale_rtx_costs (rtx x, int code, int outer_code, int *total) | |
5442 | { | |
5443 | enum machine_mode mode = GET_MODE (x); | |
5444 | ||
5445 | if (TARGET_THUMB) | |
5446 | { | |
25f905c2 | 5447 | *total = thumb1_rtx_costs (x, code, outer_code); |
a2cd141b | 5448 | return true; |
5449 | } | |
9e7454d0 | 5450 | |
a2cd141b | 5451 | switch (code) |
5452 | { | |
5453 | case MULT: | |
5454 | /* There is no point basing this on the tuning, since it is always the | |
5455 | fast variant if it exists at all. */ | |
5456 | if (mode == DImode | |
5457 | && (GET_CODE (XEXP (x, 0)) == GET_CODE (XEXP (x, 1))) | |
5458 | && (GET_CODE (XEXP (x, 0)) == ZERO_EXTEND | |
5459 | || GET_CODE (XEXP (x, 0)) == SIGN_EXTEND)) | |
5460 | { | |
5461 | *total = 8; | |
5462 | return true; | |
5463 | } | |
9e7454d0 | 5464 | |
a2cd141b | 5465 | |
5466 | if (GET_MODE_CLASS (mode) == MODE_FLOAT | |
5467 | || mode == DImode) | |
5468 | { | |
5469 | *total = 30; | |
5470 | return true; | |
5471 | } | |
5472 | ||
5473 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
5474 | { | |
5475 | unsigned HOST_WIDE_INT i = (INTVAL (XEXP (x, 1)) | |
5476 | & (unsigned HOST_WIDE_INT) 0xffffffff); | |
5477 | int cost, const_ok = const_ok_for_arm (i); | |
5478 | unsigned HOST_WIDE_INT masked_const; | |
5479 | ||
5480 | /* The cost will be related to two insns. | |
1d60d981 | 5481 | First a load of the constant (MOV or LDR), then a multiply. */ |
a2cd141b | 5482 | cost = 2; |
5483 | if (! const_ok) | |
5484 | cost += 1; /* LDR is probably more expensive because | |
1d60d981 | 5485 | of longer result latency. */ |
a2cd141b | 5486 | masked_const = i & 0xffff8000; |
5487 | if (masked_const != 0 && masked_const != 0xffff8000) | |
5488 | { | |
5489 | masked_const = i & 0xf8000000; | |
5490 | if (masked_const == 0 || masked_const == 0xf8000000) | |
5491 | cost += 1; | |
5492 | else | |
5493 | cost += 2; | |
5494 | } | |
5495 | *total = cost; | |
5496 | return true; | |
5497 | } | |
5498 | ||
5499 | *total = 8 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 4) | |
5500 | + (REG_OR_SUBREG_REG (XEXP (x, 1)) ? 0 : 4); | |
5501 | return true; | |
9e7454d0 | 5502 | |
e8a4aa75 | 5503 | case COMPARE: |
5504 | /* A COMPARE of a MULT is slow on XScale; the muls instruction | |
5505 | will stall until the multiplication is complete. */ | |
5506 | if (GET_CODE (XEXP (x, 0)) == MULT) | |
5507 | *total = 4 + rtx_cost (XEXP (x, 0), code); | |
5508 | else | |
5509 | *total = arm_rtx_costs_1 (x, code, outer_code); | |
5510 | return true; | |
5511 | ||
a2cd141b | 5512 | default: |
5513 | *total = arm_rtx_costs_1 (x, code, outer_code); | |
5514 | return true; | |
5515 | } | |
5516 | } | |
5517 | ||
5518 | ||
5519 | /* RTX costs for 9e (and later) cores. */ | |
5520 | ||
5521 | static bool | |
5522 | arm_9e_rtx_costs (rtx x, int code, int outer_code, int *total) | |
5523 | { | |
5524 | enum machine_mode mode = GET_MODE (x); | |
5525 | int nonreg_cost; | |
5526 | int cost; | |
9e7454d0 | 5527 | |
25f905c2 | 5528 | if (TARGET_THUMB1) |
a2cd141b | 5529 | { |
5530 | switch (code) | |
5531 | { | |
5532 | case MULT: | |
5533 | *total = COSTS_N_INSNS (3); | |
5534 | return true; | |
9e7454d0 | 5535 | |
a2cd141b | 5536 | default: |
25f905c2 | 5537 | *total = thumb1_rtx_costs (x, code, outer_code); |
a2cd141b | 5538 | return true; |
5539 | } | |
5540 | } | |
9e7454d0 | 5541 | |
a2cd141b | 5542 | switch (code) |
5543 | { | |
5544 | case MULT: | |
5545 | /* There is no point basing this on the tuning, since it is always the | |
5546 | fast variant if it exists at all. */ | |
5547 | if (mode == DImode | |
5548 | && (GET_CODE (XEXP (x, 0)) == GET_CODE (XEXP (x, 1))) | |
5549 | && (GET_CODE (XEXP (x, 0)) == ZERO_EXTEND | |
5550 | || GET_CODE (XEXP (x, 0)) == SIGN_EXTEND)) | |
5551 | { | |
5552 | *total = 3; | |
5553 | return true; | |
5554 | } | |
9e7454d0 | 5555 | |
a2cd141b | 5556 | |
5557 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
5558 | { | |
5559 | *total = 30; | |
5560 | return true; | |
5561 | } | |
5562 | if (mode == DImode) | |
5563 | { | |
5564 | cost = 7; | |
5565 | nonreg_cost = 8; | |
5566 | } | |
5567 | else | |
5568 | { | |
5569 | cost = 2; | |
5570 | nonreg_cost = 4; | |
5571 | } | |
5572 | ||
5573 | ||
5574 | *total = cost + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : nonreg_cost) | |
5575 | + (REG_OR_SUBREG_REG (XEXP (x, 1)) ? 0 : nonreg_cost); | |
5576 | return true; | |
9e7454d0 | 5577 | |
a2cd141b | 5578 | default: |
5579 | *total = arm_rtx_costs_1 (x, code, outer_code); | |
5580 | return true; | |
5581 | } | |
5582 | } | |
ec0457a8 | 5583 | /* All address computations that can be done are free, but rtx cost returns |
5584 | the same for practically all of them. So we weight the different types | |
5585 | of address here in the order (most pref first): | |
674a8f0b | 5586 | PRE/POST_INC/DEC, SHIFT or NON-INT sum, INT sum, REG, MEM or LABEL. */ |
d0472904 | 5587 | static inline int |
5588 | arm_arm_address_cost (rtx x) | |
5589 | { | |
5590 | enum rtx_code c = GET_CODE (x); | |
5591 | ||
5592 | if (c == PRE_INC || c == PRE_DEC || c == POST_INC || c == POST_DEC) | |
5593 | return 0; | |
5594 | if (c == MEM || c == LABEL_REF || c == SYMBOL_REF) | |
5595 | return 10; | |
5596 | ||
5597 | if (c == PLUS || c == MINUS) | |
5598 | { | |
d0472904 | 5599 | if (GET_CODE (XEXP (x, 0)) == CONST_INT) |
5600 | return 2; | |
5601 | ||
6720e96c | 5602 | if (ARITHMETIC_P (XEXP (x, 0)) || ARITHMETIC_P (XEXP (x, 1))) |
d0472904 | 5603 | return 3; |
5604 | ||
5605 | return 4; | |
5606 | } | |
5607 | ||
5608 | return 6; | |
5609 | } | |
5610 | ||
5611 | static inline int | |
5612 | arm_thumb_address_cost (rtx x) | |
5613 | { | |
5614 | enum rtx_code c = GET_CODE (x); | |
5615 | ||
5616 | if (c == REG) | |
5617 | return 1; | |
5618 | if (c == PLUS | |
5619 | && GET_CODE (XEXP (x, 0)) == REG | |
5620 | && GET_CODE (XEXP (x, 1)) == CONST_INT) | |
5621 | return 1; | |
5622 | ||
5623 | return 2; | |
5624 | } | |
5625 | ||
ec0457a8 | 5626 | static int |
ebd88a36 | 5627 | arm_address_cost (rtx x) |
ec0457a8 | 5628 | { |
25f905c2 | 5629 | return TARGET_32BIT ? arm_arm_address_cost (x) : arm_thumb_address_cost (x); |
ec0457a8 | 5630 | } |
5631 | ||
747af5e7 | 5632 | static int |
ebd88a36 | 5633 | arm_adjust_cost (rtx insn, rtx link, rtx dep, int cost) |
69120d26 | 5634 | { |
5635 | rtx i_pat, d_pat; | |
5636 | ||
331beb1a | 5637 | /* Some true dependencies can have a higher cost depending |
5638 | on precisely how certain input operands are used. */ | |
f8e17552 | 5639 | if (arm_tune_xscale |
331beb1a | 5640 | && REG_NOTE_KIND (link) == 0 |
86437b62 | 5641 | && recog_memoized (insn) >= 0 |
5642 | && recog_memoized (dep) >= 0) | |
331beb1a | 5643 | { |
5644 | int shift_opnum = get_attr_shift (insn); | |
5645 | enum attr_type attr_type = get_attr_type (dep); | |
5646 | ||
5647 | /* If nonzero, SHIFT_OPNUM contains the operand number of a shifted | |
5648 | operand for INSN. If we have a shifted input operand and the | |
5649 | instruction we depend on is another ALU instruction, then we may | |
5650 | have to account for an additional stall. */ | |
a2cd141b | 5651 | if (shift_opnum != 0 |
5652 | && (attr_type == TYPE_ALU_SHIFT || attr_type == TYPE_ALU_SHIFT_REG)) | |
331beb1a | 5653 | { |
5654 | rtx shifted_operand; | |
5655 | int opno; | |
9e7454d0 | 5656 | |
331beb1a | 5657 | /* Get the shifted operand. */ |
5658 | extract_insn (insn); | |
5659 | shifted_operand = recog_data.operand[shift_opnum]; | |
5660 | ||
5661 | /* Iterate over all the operands in DEP. If we write an operand | |
5662 | that overlaps with SHIFTED_OPERAND, then we have increase the | |
5663 | cost of this dependency. */ | |
5664 | extract_insn (dep); | |
5665 | preprocess_constraints (); | |
5666 | for (opno = 0; opno < recog_data.n_operands; opno++) | |
5667 | { | |
5668 | /* We can ignore strict inputs. */ | |
5669 | if (recog_data.operand_type[opno] == OP_IN) | |
5670 | continue; | |
5671 | ||
5672 | if (reg_overlap_mentioned_p (recog_data.operand[opno], | |
5673 | shifted_operand)) | |
5674 | return 2; | |
5675 | } | |
5676 | } | |
5677 | } | |
5678 | ||
66aacf41 | 5679 | /* XXX This is not strictly true for the FPA. */ |
cffb2a26 | 5680 | if (REG_NOTE_KIND (link) == REG_DEP_ANTI |
5681 | || REG_NOTE_KIND (link) == REG_DEP_OUTPUT) | |
129a2fe4 | 5682 | return 0; |
5683 | ||
cffb2a26 | 5684 | /* Call insns don't incur a stall, even if they follow a load. */ |
5685 | if (REG_NOTE_KIND (link) == 0 | |
5686 | && GET_CODE (insn) == CALL_INSN) | |
5687 | return 1; | |
5688 | ||
69120d26 | 5689 | if ((i_pat = single_set (insn)) != NULL |
5690 | && GET_CODE (SET_SRC (i_pat)) == MEM | |
5691 | && (d_pat = single_set (dep)) != NULL | |
5692 | && GET_CODE (SET_DEST (d_pat)) == MEM) | |
5693 | { | |
3ccc58a6 | 5694 | rtx src_mem = XEXP (SET_SRC (i_pat), 0); |
69120d26 | 5695 | /* This is a load after a store, there is no conflict if the load reads |
5696 | from a cached area. Assume that loads from the stack, and from the | |
9e7454d0 | 5697 | constant pool are cached, and that others will miss. This is a |
66aacf41 | 5698 | hack. */ |
9e7454d0 | 5699 | |
3ccc58a6 | 5700 | if ((GET_CODE (src_mem) == SYMBOL_REF && CONSTANT_POOL_ADDRESS_P (src_mem)) |
5701 | || reg_mentioned_p (stack_pointer_rtx, src_mem) | |
5702 | || reg_mentioned_p (frame_pointer_rtx, src_mem) | |
5703 | || reg_mentioned_p (hard_frame_pointer_rtx, src_mem)) | |
56d27660 | 5704 | return 1; |
69120d26 | 5705 | } |
5706 | ||
5707 | return cost; | |
5708 | } | |
5709 | ||
a2cd141b | 5710 | static int fp_consts_inited = 0; |
9c08d1fa | 5711 | |
a2cd141b | 5712 | /* Only zero is valid for VFP. Other values are also valid for FPA. */ |
5713 | static const char * const strings_fp[8] = | |
bba10fb8 | 5714 | { |
64266206 | 5715 | "0", "1", "2", "3", |
5716 | "4", "5", "0.5", "10" | |
5717 | }; | |
9c08d1fa | 5718 | |
a2cd141b | 5719 | static REAL_VALUE_TYPE values_fp[8]; |
9c08d1fa | 5720 | |
5721 | static void | |
a2cd141b | 5722 | init_fp_table (void) |
9c08d1fa | 5723 | { |
5724 | int i; | |
5725 | REAL_VALUE_TYPE r; | |
5726 | ||
a2cd141b | 5727 | if (TARGET_VFP) |
5728 | fp_consts_inited = 1; | |
5729 | else | |
5730 | fp_consts_inited = 8; | |
5731 | ||
5732 | for (i = 0; i < fp_consts_inited; i++) | |
9c08d1fa | 5733 | { |
a2cd141b | 5734 | r = REAL_VALUE_ATOF (strings_fp[i], DFmode); |
5735 | values_fp[i] = r; | |
9c08d1fa | 5736 | } |
9c08d1fa | 5737 | } |
5738 | ||
a2cd141b | 5739 | /* Return TRUE if rtx X is a valid immediate FP constant. */ |
8cd42699 | 5740 | int |
a2cd141b | 5741 | arm_const_double_rtx (rtx x) |
8cd42699 | 5742 | { |
9c08d1fa | 5743 | REAL_VALUE_TYPE r; |
5744 | int i; | |
9e7454d0 | 5745 | |
a2cd141b | 5746 | if (!fp_consts_inited) |
5747 | init_fp_table (); | |
9e7454d0 | 5748 | |
9c08d1fa | 5749 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); |
5750 | if (REAL_VALUE_MINUS_ZERO (r)) | |
5751 | return 0; | |
a690ede2 | 5752 | |
a2cd141b | 5753 | for (i = 0; i < fp_consts_inited; i++) |
5754 | if (REAL_VALUES_EQUAL (r, values_fp[i])) | |
9c08d1fa | 5755 | return 1; |
a690ede2 | 5756 | |
9c08d1fa | 5757 | return 0; |
a690ede2 | 5758 | } |
9c08d1fa | 5759 | |
e8ff69a4 | 5760 | /* Return TRUE if rtx X is a valid immediate FPA constant. */ |
9c08d1fa | 5761 | int |
ebd88a36 | 5762 | neg_const_double_rtx_ok_for_fpa (rtx x) |
9c08d1fa | 5763 | { |
5764 | REAL_VALUE_TYPE r; | |
5765 | int i; | |
9e7454d0 | 5766 | |
a2cd141b | 5767 | if (!fp_consts_inited) |
5768 | init_fp_table (); | |
9e7454d0 | 5769 | |
9c08d1fa | 5770 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); |
5771 | r = REAL_VALUE_NEGATE (r); | |
5772 | if (REAL_VALUE_MINUS_ZERO (r)) | |
5773 | return 0; | |
a690ede2 | 5774 | |
9c08d1fa | 5775 | for (i = 0; i < 8; i++) |
a2cd141b | 5776 | if (REAL_VALUES_EQUAL (r, values_fp[i])) |
9c08d1fa | 5777 | return 1; |
a690ede2 | 5778 | |
9c08d1fa | 5779 | return 0; |
a690ede2 | 5780 | } |
b5a0636d | 5781 | |
5782 | ||
5783 | /* VFPv3 has a fairly wide range of representable immediates, formed from | |
5784 | "quarter-precision" floating-point values. These can be evaluated using this | |
5785 | formula (with ^ for exponentiation): | |
5786 | ||
5787 | -1^s * n * 2^-r | |
5788 | ||
5789 | Where 's' is a sign bit (0/1), 'n' and 'r' are integers such that | |
5790 | 16 <= n <= 31 and 0 <= r <= 7. | |
5791 | ||
5792 | These values are mapped onto an 8-bit integer ABCDEFGH s.t. | |
5793 | ||
5794 | - A (most-significant) is the sign bit. | |
5795 | - BCD are the exponent (encoded as r XOR 3). | |
5796 | - EFGH are the mantissa (encoded as n - 16). | |
5797 | */ | |
5798 | ||
5799 | /* Return an integer index for a VFPv3 immediate operand X suitable for the | |
5800 | fconst[sd] instruction, or -1 if X isn't suitable. */ | |
5801 | static int | |
5802 | vfp3_const_double_index (rtx x) | |
5803 | { | |
5804 | REAL_VALUE_TYPE r, m; | |
5805 | int sign, exponent; | |
5806 | unsigned HOST_WIDE_INT mantissa, mant_hi; | |
5807 | unsigned HOST_WIDE_INT mask; | |
2e5b424c | 5808 | HOST_WIDE_INT m1, m2; |
b5a0636d | 5809 | int point_pos = 2 * HOST_BITS_PER_WIDE_INT - 1; |
5810 | ||
5811 | if (!TARGET_VFP3 || GET_CODE (x) != CONST_DOUBLE) | |
5812 | return -1; | |
5813 | ||
5814 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); | |
5815 | ||
5816 | /* We can't represent these things, so detect them first. */ | |
5817 | if (REAL_VALUE_ISINF (r) || REAL_VALUE_ISNAN (r) || REAL_VALUE_MINUS_ZERO (r)) | |
5818 | return -1; | |
5819 | ||
5820 | /* Extract sign, exponent and mantissa. */ | |
5821 | sign = REAL_VALUE_NEGATIVE (r) ? 1 : 0; | |
5822 | r = REAL_VALUE_ABS (r); | |
5823 | exponent = REAL_EXP (&r); | |
5824 | /* For the mantissa, we expand into two HOST_WIDE_INTS, apart from the | |
5825 | highest (sign) bit, with a fixed binary point at bit point_pos. | |
5826 | WARNING: If there's ever a VFP version which uses more than 2 * H_W_I - 1 | |
5827 | bits for the mantissa, this may fail (low bits would be lost). */ | |
5828 | real_ldexp (&m, &r, point_pos - exponent); | |
2e5b424c | 5829 | REAL_VALUE_TO_INT (&m1, &m2, m); |
5830 | mantissa = m1; | |
5831 | mant_hi = m2; | |
b5a0636d | 5832 | |
5833 | /* If there are bits set in the low part of the mantissa, we can't | |
5834 | represent this value. */ | |
5835 | if (mantissa != 0) | |
5836 | return -1; | |
5837 | ||
5838 | /* Now make it so that mantissa contains the most-significant bits, and move | |
5839 | the point_pos to indicate that the least-significant bits have been | |
5840 | discarded. */ | |
5841 | point_pos -= HOST_BITS_PER_WIDE_INT; | |
5842 | mantissa = mant_hi; | |
5843 | ||
5844 | /* We can permit four significant bits of mantissa only, plus a high bit | |
5845 | which is always 1. */ | |
5846 | mask = ((unsigned HOST_WIDE_INT)1 << (point_pos - 5)) - 1; | |
5847 | if ((mantissa & mask) != 0) | |
5848 | return -1; | |
5849 | ||
5850 | /* Now we know the mantissa is in range, chop off the unneeded bits. */ | |
5851 | mantissa >>= point_pos - 5; | |
5852 | ||
5853 | /* The mantissa may be zero. Disallow that case. (It's possible to load the | |
5854 | floating-point immediate zero with Neon using an integer-zero load, but | |
5855 | that case is handled elsewhere.) */ | |
5856 | if (mantissa == 0) | |
5857 | return -1; | |
5858 | ||
5859 | gcc_assert (mantissa >= 16 && mantissa <= 31); | |
5860 | ||
5861 | /* The value of 5 here would be 4 if GCC used IEEE754-like encoding (where | |
6dfdc153 | 5862 | normalized significands are in the range [1, 2). (Our mantissa is shifted |
5863 | left 4 places at this point relative to normalized IEEE754 values). GCC | |
b5a0636d | 5864 | internally uses [0.5, 1) (see real.c), so the exponent returned from |
5865 | REAL_EXP must be altered. */ | |
5866 | exponent = 5 - exponent; | |
5867 | ||
5868 | if (exponent < 0 || exponent > 7) | |
5869 | return -1; | |
5870 | ||
5871 | /* Sign, mantissa and exponent are now in the correct form to plug into the | |
5872 | formulae described in the comment above. */ | |
5873 | return (sign << 7) | ((exponent ^ 3) << 4) | (mantissa - 16); | |
5874 | } | |
5875 | ||
5876 | /* Return TRUE if rtx X is a valid immediate VFPv3 constant. */ | |
5877 | int | |
5878 | vfp3_const_double_rtx (rtx x) | |
5879 | { | |
5880 | if (!TARGET_VFP3) | |
5881 | return 0; | |
5882 | ||
5883 | return vfp3_const_double_index (x) != -1; | |
5884 | } | |
5885 | ||
d98a3884 | 5886 | /* Recognize immediates which can be used in various Neon instructions. Legal |
5887 | immediates are described by the following table (for VMVN variants, the | |
5888 | bitwise inverse of the constant shown is recognized. In either case, VMOV | |
5889 | is output and the correct instruction to use for a given constant is chosen | |
5890 | by the assembler). The constant shown is replicated across all elements of | |
5891 | the destination vector. | |
5892 | ||
5893 | insn elems variant constant (binary) | |
5894 | ---- ----- ------- ----------------- | |
5895 | vmov i32 0 00000000 00000000 00000000 abcdefgh | |
5896 | vmov i32 1 00000000 00000000 abcdefgh 00000000 | |
5897 | vmov i32 2 00000000 abcdefgh 00000000 00000000 | |
5898 | vmov i32 3 abcdefgh 00000000 00000000 00000000 | |
5899 | vmov i16 4 00000000 abcdefgh | |
5900 | vmov i16 5 abcdefgh 00000000 | |
5901 | vmvn i32 6 00000000 00000000 00000000 abcdefgh | |
5902 | vmvn i32 7 00000000 00000000 abcdefgh 00000000 | |
5903 | vmvn i32 8 00000000 abcdefgh 00000000 00000000 | |
5904 | vmvn i32 9 abcdefgh 00000000 00000000 00000000 | |
5905 | vmvn i16 10 00000000 abcdefgh | |
5906 | vmvn i16 11 abcdefgh 00000000 | |
5907 | vmov i32 12 00000000 00000000 abcdefgh 11111111 | |
5908 | vmvn i32 13 00000000 00000000 abcdefgh 11111111 | |
5909 | vmov i32 14 00000000 abcdefgh 11111111 11111111 | |
5910 | vmvn i32 15 00000000 abcdefgh 11111111 11111111 | |
5911 | vmov i8 16 abcdefgh | |
5912 | vmov i64 17 aaaaaaaa bbbbbbbb cccccccc dddddddd | |
5913 | eeeeeeee ffffffff gggggggg hhhhhhhh | |
5914 | vmov f32 18 aBbbbbbc defgh000 00000000 00000000 | |
5915 | ||
5916 | For case 18, B = !b. Representable values are exactly those accepted by | |
5917 | vfp3_const_double_index, but are output as floating-point numbers rather | |
5918 | than indices. | |
5919 | ||
5920 | Variants 0-5 (inclusive) may also be used as immediates for the second | |
5921 | operand of VORR/VBIC instructions. | |
5922 | ||
5923 | The INVERSE argument causes the bitwise inverse of the given operand to be | |
5924 | recognized instead (used for recognizing legal immediates for the VAND/VORN | |
5925 | pseudo-instructions). If INVERSE is true, the value placed in *MODCONST is | |
5926 | *not* inverted (i.e. the pseudo-instruction forms vand/vorn should still be | |
5927 | output, rather than the real insns vbic/vorr). | |
5928 | ||
5929 | INVERSE makes no difference to the recognition of float vectors. | |
5930 | ||
5931 | The return value is the variant of immediate as shown in the above table, or | |
5932 | -1 if the given value doesn't match any of the listed patterns. | |
5933 | */ | |
5934 | static int | |
5935 | neon_valid_immediate (rtx op, enum machine_mode mode, int inverse, | |
5936 | rtx *modconst, int *elementwidth) | |
5937 | { | |
5938 | #define CHECK(STRIDE, ELSIZE, CLASS, TEST) \ | |
5939 | matches = 1; \ | |
5940 | for (i = 0; i < idx; i += (STRIDE)) \ | |
5941 | if (!(TEST)) \ | |
5942 | matches = 0; \ | |
5943 | if (matches) \ | |
5944 | { \ | |
5945 | immtype = (CLASS); \ | |
5946 | elsize = (ELSIZE); \ | |
5947 | break; \ | |
5948 | } | |
5949 | ||
5950 | unsigned int i, elsize, idx = 0, n_elts = CONST_VECTOR_NUNITS (op); | |
5951 | unsigned int innersize = GET_MODE_SIZE (GET_MODE_INNER (mode)); | |
5952 | unsigned char bytes[16]; | |
5953 | int immtype = -1, matches; | |
5954 | unsigned int invmask = inverse ? 0xff : 0; | |
5955 | ||
5956 | /* Vectors of float constants. */ | |
5957 | if (GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT) | |
5958 | { | |
5959 | rtx el0 = CONST_VECTOR_ELT (op, 0); | |
5960 | REAL_VALUE_TYPE r0; | |
5961 | ||
5962 | if (!vfp3_const_double_rtx (el0)) | |
5963 | return -1; | |
5964 | ||
5965 | REAL_VALUE_FROM_CONST_DOUBLE (r0, el0); | |
5966 | ||
5967 | for (i = 1; i < n_elts; i++) | |
5968 | { | |
5969 | rtx elt = CONST_VECTOR_ELT (op, i); | |
5970 | REAL_VALUE_TYPE re; | |
5971 | ||
5972 | REAL_VALUE_FROM_CONST_DOUBLE (re, elt); | |
5973 | ||
5974 | if (!REAL_VALUES_EQUAL (r0, re)) | |
5975 | return -1; | |
5976 | } | |
5977 | ||
5978 | if (modconst) | |
5979 | *modconst = CONST_VECTOR_ELT (op, 0); | |
5980 | ||
5981 | if (elementwidth) | |
5982 | *elementwidth = 0; | |
5983 | ||
5984 | return 18; | |
5985 | } | |
5986 | ||
5987 | /* Splat vector constant out into a byte vector. */ | |
5988 | for (i = 0; i < n_elts; i++) | |
5989 | { | |
5990 | rtx el = CONST_VECTOR_ELT (op, i); | |
5991 | unsigned HOST_WIDE_INT elpart; | |
5992 | unsigned int part, parts; | |
5993 | ||
5994 | if (GET_CODE (el) == CONST_INT) | |
5995 | { | |
5996 | elpart = INTVAL (el); | |
5997 | parts = 1; | |
5998 | } | |
5999 | else if (GET_CODE (el) == CONST_DOUBLE) | |
6000 | { | |
6001 | elpart = CONST_DOUBLE_LOW (el); | |
6002 | parts = 2; | |
6003 | } | |
6004 | else | |
6005 | gcc_unreachable (); | |
6006 | ||
6007 | for (part = 0; part < parts; part++) | |
6008 | { | |
6009 | unsigned int byte; | |
6010 | for (byte = 0; byte < innersize; byte++) | |
6011 | { | |
6012 | bytes[idx++] = (elpart & 0xff) ^ invmask; | |
6013 | elpart >>= BITS_PER_UNIT; | |
6014 | } | |
6015 | if (GET_CODE (el) == CONST_DOUBLE) | |
6016 | elpart = CONST_DOUBLE_HIGH (el); | |
6017 | } | |
6018 | } | |
6019 | ||
6020 | /* Sanity check. */ | |
6021 | gcc_assert (idx == GET_MODE_SIZE (mode)); | |
6022 | ||
6023 | do | |
6024 | { | |
6025 | CHECK (4, 32, 0, bytes[i] == bytes[0] && bytes[i + 1] == 0 | |
6026 | && bytes[i + 2] == 0 && bytes[i + 3] == 0); | |
6027 | ||
6028 | CHECK (4, 32, 1, bytes[i] == 0 && bytes[i + 1] == bytes[1] | |
6029 | && bytes[i + 2] == 0 && bytes[i + 3] == 0); | |
6030 | ||
6031 | CHECK (4, 32, 2, bytes[i] == 0 && bytes[i + 1] == 0 | |
6032 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0); | |
6033 | ||
6034 | CHECK (4, 32, 3, bytes[i] == 0 && bytes[i + 1] == 0 | |
6035 | && bytes[i + 2] == 0 && bytes[i + 3] == bytes[3]); | |
6036 | ||
6037 | CHECK (2, 16, 4, bytes[i] == bytes[0] && bytes[i + 1] == 0); | |
6038 | ||
6039 | CHECK (2, 16, 5, bytes[i] == 0 && bytes[i + 1] == bytes[1]); | |
6040 | ||
6041 | CHECK (4, 32, 6, bytes[i] == bytes[0] && bytes[i + 1] == 0xff | |
6042 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff); | |
6043 | ||
6044 | CHECK (4, 32, 7, bytes[i] == 0xff && bytes[i + 1] == bytes[1] | |
6045 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff); | |
6046 | ||
6047 | CHECK (4, 32, 8, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
6048 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0xff); | |
6049 | ||
6050 | CHECK (4, 32, 9, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
6051 | && bytes[i + 2] == 0xff && bytes[i + 3] == bytes[3]); | |
6052 | ||
6053 | CHECK (2, 16, 10, bytes[i] == bytes[0] && bytes[i + 1] == 0xff); | |
6054 | ||
6055 | CHECK (2, 16, 11, bytes[i] == 0xff && bytes[i + 1] == bytes[1]); | |
6056 | ||
6057 | CHECK (4, 32, 12, bytes[i] == 0xff && bytes[i + 1] == bytes[1] | |
6058 | && bytes[i + 2] == 0 && bytes[i + 3] == 0); | |
6059 | ||
6060 | CHECK (4, 32, 13, bytes[i] == 0 && bytes[i + 1] == bytes[1] | |
6061 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff); | |
6062 | ||
6063 | CHECK (4, 32, 14, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
6064 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0); | |
6065 | ||
6066 | CHECK (4, 32, 15, bytes[i] == 0 && bytes[i + 1] == 0 | |
6067 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0xff); | |
6068 | ||
6069 | CHECK (1, 8, 16, bytes[i] == bytes[0]); | |
6070 | ||
6071 | CHECK (1, 64, 17, (bytes[i] == 0 || bytes[i] == 0xff) | |
6072 | && bytes[i] == bytes[(i + 8) % idx]); | |
6073 | } | |
6074 | while (0); | |
6075 | ||
6076 | if (immtype == -1) | |
6077 | return -1; | |
6078 | ||
6079 | if (elementwidth) | |
6080 | *elementwidth = elsize; | |
6081 | ||
6082 | if (modconst) | |
6083 | { | |
6084 | unsigned HOST_WIDE_INT imm = 0; | |
6085 | ||
6086 | /* Un-invert bytes of recognized vector, if neccessary. */ | |
6087 | if (invmask != 0) | |
6088 | for (i = 0; i < idx; i++) | |
6089 | bytes[i] ^= invmask; | |
6090 | ||
6091 | if (immtype == 17) | |
6092 | { | |
6093 | /* FIXME: Broken on 32-bit H_W_I hosts. */ | |
6094 | gcc_assert (sizeof (HOST_WIDE_INT) == 8); | |
6095 | ||
6096 | for (i = 0; i < 8; i++) | |
6097 | imm |= (unsigned HOST_WIDE_INT) (bytes[i] ? 0xff : 0) | |
6098 | << (i * BITS_PER_UNIT); | |
6099 | ||
6100 | *modconst = GEN_INT (imm); | |
6101 | } | |
6102 | else | |
6103 | { | |
6104 | unsigned HOST_WIDE_INT imm = 0; | |
6105 | ||
6106 | for (i = 0; i < elsize / BITS_PER_UNIT; i++) | |
6107 | imm |= (unsigned HOST_WIDE_INT) bytes[i] << (i * BITS_PER_UNIT); | |
6108 | ||
6109 | *modconst = GEN_INT (imm); | |
6110 | } | |
6111 | } | |
6112 | ||
6113 | return immtype; | |
6114 | #undef CHECK | |
6115 | } | |
6116 | ||
6117 | /* Return TRUE if rtx X is legal for use as either a Neon VMOV (or, implicitly, | |
6118 | VMVN) immediate. Write back width per element to *ELEMENTWIDTH (or zero for | |
6119 | float elements), and a modified constant (whatever should be output for a | |
6120 | VMOV) in *MODCONST. */ | |
6121 | ||
6122 | int | |
6123 | neon_immediate_valid_for_move (rtx op, enum machine_mode mode, | |
6124 | rtx *modconst, int *elementwidth) | |
6125 | { | |
6126 | rtx tmpconst; | |
6127 | int tmpwidth; | |
6128 | int retval = neon_valid_immediate (op, mode, 0, &tmpconst, &tmpwidth); | |
6129 | ||
6130 | if (retval == -1) | |
6131 | return 0; | |
6132 | ||
6133 | if (modconst) | |
6134 | *modconst = tmpconst; | |
6135 | ||
6136 | if (elementwidth) | |
6137 | *elementwidth = tmpwidth; | |
6138 | ||
6139 | return 1; | |
6140 | } | |
6141 | ||
6142 | /* Return TRUE if rtx X is legal for use in a VORR or VBIC instruction. If | |
6143 | the immediate is valid, write a constant suitable for using as an operand | |
6144 | to VORR/VBIC/VAND/VORN to *MODCONST and the corresponding element width to | |
6145 | *ELEMENTWIDTH. See neon_valid_immediate for description of INVERSE. */ | |
6146 | ||
6147 | int | |
6148 | neon_immediate_valid_for_logic (rtx op, enum machine_mode mode, int inverse, | |
6149 | rtx *modconst, int *elementwidth) | |
6150 | { | |
6151 | rtx tmpconst; | |
6152 | int tmpwidth; | |
6153 | int retval = neon_valid_immediate (op, mode, inverse, &tmpconst, &tmpwidth); | |
6154 | ||
6155 | if (retval < 0 || retval > 5) | |
6156 | return 0; | |
6157 | ||
6158 | if (modconst) | |
6159 | *modconst = tmpconst; | |
6160 | ||
6161 | if (elementwidth) | |
6162 | *elementwidth = tmpwidth; | |
6163 | ||
6164 | return 1; | |
6165 | } | |
6166 | ||
6167 | /* Return a string suitable for output of Neon immediate logic operation | |
6168 | MNEM. */ | |
6169 | ||
6170 | char * | |
6171 | neon_output_logic_immediate (const char *mnem, rtx *op2, enum machine_mode mode, | |
6172 | int inverse, int quad) | |
6173 | { | |
6174 | int width, is_valid; | |
6175 | static char templ[40]; | |
6176 | ||
6177 | is_valid = neon_immediate_valid_for_logic (*op2, mode, inverse, op2, &width); | |
6178 | ||
6179 | gcc_assert (is_valid != 0); | |
6180 | ||
6181 | if (quad) | |
6182 | sprintf (templ, "%s.i%d\t%%q0, %%2", mnem, width); | |
6183 | else | |
6184 | sprintf (templ, "%s.i%d\t%%P0, %%2", mnem, width); | |
6185 | ||
6186 | return templ; | |
6187 | } | |
6188 | ||
6189 | /* Output a sequence of pairwise operations to implement a reduction. | |
6190 | NOTE: We do "too much work" here, because pairwise operations work on two | |
6191 | registers-worth of operands in one go. Unfortunately we can't exploit those | |
6192 | extra calculations to do the full operation in fewer steps, I don't think. | |
6193 | Although all vector elements of the result but the first are ignored, we | |
6194 | actually calculate the same result in each of the elements. An alternative | |
6195 | such as initially loading a vector with zero to use as each of the second | |
6196 | operands would use up an additional register and take an extra instruction, | |
6197 | for no particular gain. */ | |
6198 | ||
6199 | void | |
6200 | neon_pairwise_reduce (rtx op0, rtx op1, enum machine_mode mode, | |
6201 | rtx (*reduc) (rtx, rtx, rtx)) | |
6202 | { | |
6203 | enum machine_mode inner = GET_MODE_INNER (mode); | |
6204 | unsigned int i, parts = GET_MODE_SIZE (mode) / GET_MODE_SIZE (inner); | |
6205 | rtx tmpsum = op1; | |
6206 | ||
6207 | for (i = parts / 2; i >= 1; i /= 2) | |
6208 | { | |
6209 | rtx dest = (i == 1) ? op0 : gen_reg_rtx (mode); | |
6210 | emit_insn (reduc (dest, tmpsum, tmpsum)); | |
6211 | tmpsum = dest; | |
6212 | } | |
6213 | } | |
6214 | ||
6215 | /* Initialise a vector with non-constant elements. FIXME: We can do better | |
6216 | than the current implementation (building a vector on the stack and then | |
6217 | loading it) in many cases. See rs6000.c. */ | |
6218 | ||
6219 | void | |
6220 | neon_expand_vector_init (rtx target, rtx vals) | |
6221 | { | |
6222 | enum machine_mode mode = GET_MODE (target); | |
6223 | enum machine_mode inner = GET_MODE_INNER (mode); | |
6224 | unsigned int i, n_elts = GET_MODE_NUNITS (mode); | |
6225 | rtx mem; | |
6226 | ||
6227 | gcc_assert (VECTOR_MODE_P (mode)); | |
6228 | ||
6229 | mem = assign_stack_temp (mode, GET_MODE_SIZE (mode), 0); | |
6230 | for (i = 0; i < n_elts; i++) | |
6231 | emit_move_insn (adjust_address_nv (mem, inner, i * GET_MODE_SIZE (inner)), | |
6232 | XVECEXP (vals, 0, i)); | |
6233 | ||
6234 | emit_move_insn (target, mem); | |
6235 | } | |
6236 | ||
8cd42699 | 6237 | \f |
6238 | /* Predicates for `match_operand' and `match_operator'. */ | |
6239 | ||
7d57ec45 | 6240 | /* Return nonzero if OP is a valid Cirrus memory address pattern. */ |
7d57ec45 | 6241 | int |
ebd88a36 | 6242 | cirrus_memory_offset (rtx op) |
7d57ec45 | 6243 | { |
6244 | /* Reject eliminable registers. */ | |
6245 | if (! (reload_in_progress || reload_completed) | |
6246 | && ( reg_mentioned_p (frame_pointer_rtx, op) | |
6247 | || reg_mentioned_p (arg_pointer_rtx, op) | |
6248 | || reg_mentioned_p (virtual_incoming_args_rtx, op) | |
6249 | || reg_mentioned_p (virtual_outgoing_args_rtx, op) | |
6250 | || reg_mentioned_p (virtual_stack_dynamic_rtx, op) | |
6251 | || reg_mentioned_p (virtual_stack_vars_rtx, op))) | |
6252 | return 0; | |
6253 | ||
6254 | if (GET_CODE (op) == MEM) | |
6255 | { | |
6256 | rtx ind; | |
6257 | ||
6258 | ind = XEXP (op, 0); | |
6259 | ||
6260 | /* Match: (mem (reg)). */ | |
6261 | if (GET_CODE (ind) == REG) | |
6262 | return 1; | |
6263 | ||
6264 | /* Match: | |
6265 | (mem (plus (reg) | |
6266 | (const))). */ | |
6267 | if (GET_CODE (ind) == PLUS | |
6268 | && GET_CODE (XEXP (ind, 0)) == REG | |
6269 | && REG_MODE_OK_FOR_BASE_P (XEXP (ind, 0), VOIDmode) | |
6270 | && GET_CODE (XEXP (ind, 1)) == CONST_INT) | |
6271 | return 1; | |
6272 | } | |
6273 | ||
6274 | return 0; | |
6275 | } | |
6276 | ||
eb38fdf8 | 6277 | /* Return TRUE if OP is a valid coprocessor memory address pattern. |
25f905c2 | 6278 | WB is true if full writeback address modes are allowed and is false |
6279 | if limited writeback address modes (POST_INC and PRE_DEC) are | |
6280 | allowed. */ | |
a2cd141b | 6281 | |
6282 | int | |
a8a3b539 | 6283 | arm_coproc_mem_operand (rtx op, bool wb) |
a2cd141b | 6284 | { |
a8a3b539 | 6285 | rtx ind; |
a2cd141b | 6286 | |
a8a3b539 | 6287 | /* Reject eliminable registers. */ |
a2cd141b | 6288 | if (! (reload_in_progress || reload_completed) |
6289 | && ( reg_mentioned_p (frame_pointer_rtx, op) | |
6290 | || reg_mentioned_p (arg_pointer_rtx, op) | |
6291 | || reg_mentioned_p (virtual_incoming_args_rtx, op) | |
6292 | || reg_mentioned_p (virtual_outgoing_args_rtx, op) | |
6293 | || reg_mentioned_p (virtual_stack_dynamic_rtx, op) | |
6294 | || reg_mentioned_p (virtual_stack_vars_rtx, op))) | |
6295 | return FALSE; | |
6296 | ||
58552ed0 | 6297 | /* Constants are converted into offsets from labels. */ |
a8a3b539 | 6298 | if (GET_CODE (op) != MEM) |
6299 | return FALSE; | |
a2cd141b | 6300 | |
a8a3b539 | 6301 | ind = XEXP (op, 0); |
a2cd141b | 6302 | |
a8a3b539 | 6303 | if (reload_completed |
6304 | && (GET_CODE (ind) == LABEL_REF | |
6305 | || (GET_CODE (ind) == CONST | |
6306 | && GET_CODE (XEXP (ind, 0)) == PLUS | |
6307 | && GET_CODE (XEXP (XEXP (ind, 0), 0)) == LABEL_REF | |
6308 | && GET_CODE (XEXP (XEXP (ind, 0), 1)) == CONST_INT))) | |
6309 | return TRUE; | |
a2cd141b | 6310 | |
a8a3b539 | 6311 | /* Match: (mem (reg)). */ |
6312 | if (GET_CODE (ind) == REG) | |
6313 | return arm_address_register_rtx_p (ind, 0); | |
6314 | ||
25f905c2 | 6315 | /* Autoincremment addressing modes. POST_INC and PRE_DEC are |
6316 | acceptable in any case (subject to verification by | |
6317 | arm_address_register_rtx_p). We need WB to be true to accept | |
6318 | PRE_INC and POST_DEC. */ | |
6319 | if (GET_CODE (ind) == POST_INC | |
6320 | || GET_CODE (ind) == PRE_DEC | |
6321 | || (wb | |
6322 | && (GET_CODE (ind) == PRE_INC | |
6323 | || GET_CODE (ind) == POST_DEC))) | |
a8a3b539 | 6324 | return arm_address_register_rtx_p (XEXP (ind, 0), 0); |
6325 | ||
6326 | if (wb | |
6327 | && (GET_CODE (ind) == POST_MODIFY || GET_CODE (ind) == PRE_MODIFY) | |
6328 | && arm_address_register_rtx_p (XEXP (ind, 0), 0) | |
6329 | && GET_CODE (XEXP (ind, 1)) == PLUS | |
6330 | && rtx_equal_p (XEXP (XEXP (ind, 1), 0), XEXP (ind, 0))) | |
6331 | ind = XEXP (ind, 1); | |
6332 | ||
6333 | /* Match: | |
6334 | (plus (reg) | |
6335 | (const)). */ | |
6336 | if (GET_CODE (ind) == PLUS | |
6337 | && GET_CODE (XEXP (ind, 0)) == REG | |
6338 | && REG_MODE_OK_FOR_BASE_P (XEXP (ind, 0), VOIDmode) | |
6339 | && GET_CODE (XEXP (ind, 1)) == CONST_INT | |
6340 | && INTVAL (XEXP (ind, 1)) > -1024 | |
6341 | && INTVAL (XEXP (ind, 1)) < 1024 | |
6342 | && (INTVAL (XEXP (ind, 1)) & 3) == 0) | |
6343 | return TRUE; | |
a2cd141b | 6344 | |
6345 | return FALSE; | |
6346 | } | |
6347 | ||
d98a3884 | 6348 | /* Return TRUE if OP is a memory operand which we can load or store a vector |
6349 | to/from. If CORE is true, we're moving from ARM registers not Neon | |
6350 | registers. */ | |
6351 | int | |
6352 | neon_vector_mem_operand (rtx op, bool core) | |
6353 | { | |
6354 | rtx ind; | |
6355 | ||
6356 | /* Reject eliminable registers. */ | |
6357 | if (! (reload_in_progress || reload_completed) | |
6358 | && ( reg_mentioned_p (frame_pointer_rtx, op) | |
6359 | || reg_mentioned_p (arg_pointer_rtx, op) | |
6360 | || reg_mentioned_p (virtual_incoming_args_rtx, op) | |
6361 | || reg_mentioned_p (virtual_outgoing_args_rtx, op) | |
6362 | || reg_mentioned_p (virtual_stack_dynamic_rtx, op) | |
6363 | || reg_mentioned_p (virtual_stack_vars_rtx, op))) | |
6364 | return FALSE; | |
6365 | ||
6366 | /* Constants are converted into offsets from labels. */ | |
6367 | if (GET_CODE (op) != MEM) | |
6368 | return FALSE; | |
6369 | ||
6370 | ind = XEXP (op, 0); | |
6371 | ||
6372 | if (reload_completed | |
6373 | && (GET_CODE (ind) == LABEL_REF | |
6374 | || (GET_CODE (ind) == CONST | |
6375 | && GET_CODE (XEXP (ind, 0)) == PLUS | |
6376 | && GET_CODE (XEXP (XEXP (ind, 0), 0)) == LABEL_REF | |
6377 | && GET_CODE (XEXP (XEXP (ind, 0), 1)) == CONST_INT))) | |
6378 | return TRUE; | |
6379 | ||
6380 | /* Match: (mem (reg)). */ | |
6381 | if (GET_CODE (ind) == REG) | |
6382 | return arm_address_register_rtx_p (ind, 0); | |
6383 | ||
6384 | /* Allow post-increment with Neon registers. */ | |
6385 | if (!core && GET_CODE (ind) == POST_INC) | |
6386 | return arm_address_register_rtx_p (XEXP (ind, 0), 0); | |
6387 | ||
6388 | #if 0 | |
6389 | /* FIXME: We can support this too if we use VLD1/VST1. */ | |
6390 | if (!core | |
6391 | && GET_CODE (ind) == POST_MODIFY | |
6392 | && arm_address_register_rtx_p (XEXP (ind, 0), 0) | |
6393 | && GET_CODE (XEXP (ind, 1)) == PLUS | |
6394 | && rtx_equal_p (XEXP (XEXP (ind, 1), 0), XEXP (ind, 0))) | |
6395 | ind = XEXP (ind, 1); | |
6396 | #endif | |
6397 | ||
6398 | /* Match: | |
6399 | (plus (reg) | |
6400 | (const)). */ | |
6401 | if (!core | |
6402 | && GET_CODE (ind) == PLUS | |
6403 | && GET_CODE (XEXP (ind, 0)) == REG | |
6404 | && REG_MODE_OK_FOR_BASE_P (XEXP (ind, 0), VOIDmode) | |
6405 | && GET_CODE (XEXP (ind, 1)) == CONST_INT | |
6406 | && INTVAL (XEXP (ind, 1)) > -1024 | |
6407 | && INTVAL (XEXP (ind, 1)) < 1016 | |
6408 | && (INTVAL (XEXP (ind, 1)) & 3) == 0) | |
6409 | return TRUE; | |
6410 | ||
6411 | return FALSE; | |
6412 | } | |
6413 | ||
6414 | /* Return TRUE if OP is a mem suitable for loading/storing a Neon struct | |
6415 | type. */ | |
6416 | int | |
6417 | neon_struct_mem_operand (rtx op) | |
6418 | { | |
6419 | rtx ind; | |
6420 | ||
6421 | /* Reject eliminable registers. */ | |
6422 | if (! (reload_in_progress || reload_completed) | |
6423 | && ( reg_mentioned_p (frame_pointer_rtx, op) | |
6424 | || reg_mentioned_p (arg_pointer_rtx, op) | |
6425 | || reg_mentioned_p (virtual_incoming_args_rtx, op) | |
6426 | || reg_mentioned_p (virtual_outgoing_args_rtx, op) | |
6427 | || reg_mentioned_p (virtual_stack_dynamic_rtx, op) | |
6428 | || reg_mentioned_p (virtual_stack_vars_rtx, op))) | |
6429 | return FALSE; | |
6430 | ||
6431 | /* Constants are converted into offsets from labels. */ | |
6432 | if (GET_CODE (op) != MEM) | |
6433 | return FALSE; | |
6434 | ||
6435 | ind = XEXP (op, 0); | |
6436 | ||
6437 | if (reload_completed | |
6438 | && (GET_CODE (ind) == LABEL_REF | |
6439 | || (GET_CODE (ind) == CONST | |
6440 | && GET_CODE (XEXP (ind, 0)) == PLUS | |
6441 | && GET_CODE (XEXP (XEXP (ind, 0), 0)) == LABEL_REF | |
6442 | && GET_CODE (XEXP (XEXP (ind, 0), 1)) == CONST_INT))) | |
6443 | return TRUE; | |
6444 | ||
6445 | /* Match: (mem (reg)). */ | |
6446 | if (GET_CODE (ind) == REG) | |
6447 | return arm_address_register_rtx_p (ind, 0); | |
6448 | ||
6449 | return FALSE; | |
6450 | } | |
6451 | ||
841b213d | 6452 | /* Return true if X is a register that will be eliminated later on. */ |
6453 | int | |
6454 | arm_eliminable_register (rtx x) | |
6455 | { | |
6456 | return REG_P (x) && (REGNO (x) == FRAME_POINTER_REGNUM | |
6457 | || REGNO (x) == ARG_POINTER_REGNUM | |
6458 | || (REGNO (x) >= FIRST_VIRTUAL_REGISTER | |
6459 | && REGNO (x) <= LAST_VIRTUAL_REGISTER)); | |
6460 | } | |
a2cd141b | 6461 | |
a2cd141b | 6462 | /* Return GENERAL_REGS if a scratch register required to reload x to/from |
f1225f6f | 6463 | coprocessor registers. Otherwise return NO_REGS. */ |
a2cd141b | 6464 | |
6465 | enum reg_class | |
f1225f6f | 6466 | coproc_secondary_reload_class (enum machine_mode mode, rtx x, bool wb) |
a2cd141b | 6467 | { |
d98a3884 | 6468 | if (TARGET_NEON |
6469 | && (GET_MODE_CLASS (mode) == MODE_VECTOR_INT | |
6470 | || GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT) | |
6471 | && neon_vector_mem_operand (x, FALSE)) | |
6472 | return NO_REGS; | |
6473 | ||
f1225f6f | 6474 | if (arm_coproc_mem_operand (x, wb) || s_register_operand (x, mode)) |
a2cd141b | 6475 | return NO_REGS; |
6476 | ||
6477 | return GENERAL_REGS; | |
6478 | } | |
6479 | ||
46b5d878 | 6480 | /* Values which must be returned in the most-significant end of the return |
6481 | register. */ | |
6482 | ||
6483 | static bool | |
6484 | arm_return_in_msb (tree valtype) | |
6485 | { | |
6486 | return (TARGET_AAPCS_BASED | |
6487 | && BYTES_BIG_ENDIAN | |
6488 | && (AGGREGATE_TYPE_P (valtype) | |
6489 | || TREE_CODE (valtype) == COMPLEX_TYPE)); | |
6490 | } | |
a2cd141b | 6491 | |
2c6c7d8b | 6492 | /* Returns TRUE if INSN is an "LDR REG, ADDR" instruction. |
6493 | Use by the Cirrus Maverick code which has to workaround | |
6494 | a hardware bug triggered by such instructions. */ | |
2c6c7d8b | 6495 | static bool |
ebd88a36 | 6496 | arm_memory_load_p (rtx insn) |
7d57ec45 | 6497 | { |
6498 | rtx body, lhs, rhs;; | |
6499 | ||
2c6c7d8b | 6500 | if (insn == NULL_RTX || GET_CODE (insn) != INSN) |
6501 | return false; | |
7d57ec45 | 6502 | |
6503 | body = PATTERN (insn); | |
6504 | ||
6505 | if (GET_CODE (body) != SET) | |
2c6c7d8b | 6506 | return false; |
7d57ec45 | 6507 | |
6508 | lhs = XEXP (body, 0); | |
6509 | rhs = XEXP (body, 1); | |
6510 | ||
2c6c7d8b | 6511 | lhs = REG_OR_SUBREG_RTX (lhs); |
6512 | ||
6513 | /* If the destination is not a general purpose | |
6514 | register we do not have to worry. */ | |
6515 | if (GET_CODE (lhs) != REG | |
6516 | || REGNO_REG_CLASS (REGNO (lhs)) != GENERAL_REGS) | |
6517 | return false; | |
6518 | ||
6519 | /* As well as loads from memory we also have to react | |
6520 | to loads of invalid constants which will be turned | |
6521 | into loads from the minipool. */ | |
6522 | return (GET_CODE (rhs) == MEM | |
6523 | || GET_CODE (rhs) == SYMBOL_REF | |
6524 | || note_invalid_constants (insn, -1, false)); | |
7d57ec45 | 6525 | } |
6526 | ||
2c6c7d8b | 6527 | /* Return TRUE if INSN is a Cirrus instruction. */ |
2c6c7d8b | 6528 | static bool |
ebd88a36 | 6529 | arm_cirrus_insn_p (rtx insn) |
7d57ec45 | 6530 | { |
6531 | enum attr_cirrus attr; | |
6532 | ||
ed29c566 | 6533 | /* get_attr cannot accept USE or CLOBBER. */ |
7d57ec45 | 6534 | if (!insn |
6535 | || GET_CODE (insn) != INSN | |
6536 | || GET_CODE (PATTERN (insn)) == USE | |
6537 | || GET_CODE (PATTERN (insn)) == CLOBBER) | |
6538 | return 0; | |
6539 | ||
6540 | attr = get_attr_cirrus (insn); | |
6541 | ||
2c6c7d8b | 6542 | return attr != CIRRUS_NOT; |
7d57ec45 | 6543 | } |
6544 | ||
6545 | /* Cirrus reorg for invalid instruction combinations. */ | |
7d57ec45 | 6546 | static void |
ebd88a36 | 6547 | cirrus_reorg (rtx first) |
7d57ec45 | 6548 | { |
6549 | enum attr_cirrus attr; | |
6550 | rtx body = PATTERN (first); | |
6551 | rtx t; | |
6552 | int nops; | |
6553 | ||
6554 | /* Any branch must be followed by 2 non Cirrus instructions. */ | |
6555 | if (GET_CODE (first) == JUMP_INSN && GET_CODE (body) != RETURN) | |
6556 | { | |
6557 | nops = 0; | |
6558 | t = next_nonnote_insn (first); | |
6559 | ||
2c6c7d8b | 6560 | if (arm_cirrus_insn_p (t)) |
7d57ec45 | 6561 | ++ nops; |
6562 | ||
2c6c7d8b | 6563 | if (arm_cirrus_insn_p (next_nonnote_insn (t))) |
7d57ec45 | 6564 | ++ nops; |
6565 | ||
6566 | while (nops --) | |
6567 | emit_insn_after (gen_nop (), first); | |
6568 | ||
6569 | return; | |
6570 | } | |
6571 | ||
6572 | /* (float (blah)) is in parallel with a clobber. */ | |
6573 | if (GET_CODE (body) == PARALLEL && XVECLEN (body, 0) > 0) | |
6574 | body = XVECEXP (body, 0, 0); | |
6575 | ||
6576 | if (GET_CODE (body) == SET) | |
6577 | { | |
6578 | rtx lhs = XEXP (body, 0), rhs = XEXP (body, 1); | |
6579 | ||
6580 | /* cfldrd, cfldr64, cfstrd, cfstr64 must | |
6581 | be followed by a non Cirrus insn. */ | |
6582 | if (get_attr_cirrus (first) == CIRRUS_DOUBLE) | |
6583 | { | |
2c6c7d8b | 6584 | if (arm_cirrus_insn_p (next_nonnote_insn (first))) |
7d57ec45 | 6585 | emit_insn_after (gen_nop (), first); |
6586 | ||
6587 | return; | |
6588 | } | |
2c6c7d8b | 6589 | else if (arm_memory_load_p (first)) |
7d57ec45 | 6590 | { |
6591 | unsigned int arm_regno; | |
6592 | ||
6593 | /* Any ldr/cfmvdlr, ldr/cfmvdhr, ldr/cfmvsr, ldr/cfmv64lr, | |
6594 | ldr/cfmv64hr combination where the Rd field is the same | |
6595 | in both instructions must be split with a non Cirrus | |
6596 | insn. Example: | |
6597 | ||
6598 | ldr r0, blah | |
6599 | nop | |
6600 | cfmvsr mvf0, r0. */ | |
6601 | ||
6602 | /* Get Arm register number for ldr insn. */ | |
6603 | if (GET_CODE (lhs) == REG) | |
6604 | arm_regno = REGNO (lhs); | |
7d57ec45 | 6605 | else |
ed29c566 | 6606 | { |
6607 | gcc_assert (GET_CODE (rhs) == REG); | |
6608 | arm_regno = REGNO (rhs); | |
6609 | } | |
7d57ec45 | 6610 | |
6611 | /* Next insn. */ | |
6612 | first = next_nonnote_insn (first); | |
6613 | ||
2c6c7d8b | 6614 | if (! arm_cirrus_insn_p (first)) |
7d57ec45 | 6615 | return; |
6616 | ||
6617 | body = PATTERN (first); | |
6618 | ||
6619 | /* (float (blah)) is in parallel with a clobber. */ | |
6620 | if (GET_CODE (body) == PARALLEL && XVECLEN (body, 0)) | |
6621 | body = XVECEXP (body, 0, 0); | |
6622 | ||
6623 | if (GET_CODE (body) == FLOAT) | |
6624 | body = XEXP (body, 0); | |
6625 | ||
6626 | if (get_attr_cirrus (first) == CIRRUS_MOVE | |
6627 | && GET_CODE (XEXP (body, 1)) == REG | |
6628 | && arm_regno == REGNO (XEXP (body, 1))) | |
6629 | emit_insn_after (gen_nop (), first); | |
6630 | ||
6631 | return; | |
6632 | } | |
6633 | } | |
6634 | ||
ed29c566 | 6635 | /* get_attr cannot accept USE or CLOBBER. */ |
7d57ec45 | 6636 | if (!first |
6637 | || GET_CODE (first) != INSN | |
6638 | || GET_CODE (PATTERN (first)) == USE | |
6639 | || GET_CODE (PATTERN (first)) == CLOBBER) | |
6640 | return; | |
6641 | ||
6642 | attr = get_attr_cirrus (first); | |
6643 | ||
6644 | /* Any coprocessor compare instruction (cfcmps, cfcmpd, ...) | |
6645 | must be followed by a non-coprocessor instruction. */ | |
6646 | if (attr == CIRRUS_COMPARE) | |
6647 | { | |
6648 | nops = 0; | |
6649 | ||
6650 | t = next_nonnote_insn (first); | |
6651 | ||
2c6c7d8b | 6652 | if (arm_cirrus_insn_p (t)) |
7d57ec45 | 6653 | ++ nops; |
6654 | ||
2c6c7d8b | 6655 | if (arm_cirrus_insn_p (next_nonnote_insn (t))) |
7d57ec45 | 6656 | ++ nops; |
6657 | ||
6658 | while (nops --) | |
6659 | emit_insn_after (gen_nop (), first); | |
6660 | ||
6661 | return; | |
6662 | } | |
6663 | } | |
6664 | ||
64266206 | 6665 | /* Return TRUE if X references a SYMBOL_REF. */ |
6666 | int | |
ebd88a36 | 6667 | symbol_mentioned_p (rtx x) |
64266206 | 6668 | { |
a12be13c | 6669 | const char * fmt; |
6670 | int i; | |
64266206 | 6671 | |
6672 | if (GET_CODE (x) == SYMBOL_REF) | |
6673 | return 1; | |
6674 | ||
f655717d | 6675 | /* UNSPEC_TLS entries for a symbol include the SYMBOL_REF, but they |
6676 | are constant offsets, not symbols. */ | |
6677 | if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLS) | |
6678 | return 0; | |
6679 | ||
64266206 | 6680 | fmt = GET_RTX_FORMAT (GET_CODE (x)); |
9e7454d0 | 6681 | |
64266206 | 6682 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) |
6683 | { | |
6684 | if (fmt[i] == 'E') | |
6685 | { | |
a12be13c | 6686 | int j; |
64266206 | 6687 | |
6688 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
6689 | if (symbol_mentioned_p (XVECEXP (x, i, j))) | |
6690 | return 1; | |
6691 | } | |
6692 | else if (fmt[i] == 'e' && symbol_mentioned_p (XEXP (x, i))) | |
6693 | return 1; | |
6694 | } | |
6695 | ||
6696 | return 0; | |
6697 | } | |
6698 | ||
6699 | /* Return TRUE if X references a LABEL_REF. */ | |
6700 | int | |
ebd88a36 | 6701 | label_mentioned_p (rtx x) |
64266206 | 6702 | { |
a12be13c | 6703 | const char * fmt; |
6704 | int i; | |
64266206 | 6705 | |
6706 | if (GET_CODE (x) == LABEL_REF) | |
6707 | return 1; | |
6708 | ||
f655717d | 6709 | /* UNSPEC_TLS entries for a symbol include a LABEL_REF for the referencing |
6710 | instruction, but they are constant offsets, not symbols. */ | |
6711 | if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLS) | |
6712 | return 0; | |
6713 | ||
64266206 | 6714 | fmt = GET_RTX_FORMAT (GET_CODE (x)); |
6715 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
6716 | { | |
6717 | if (fmt[i] == 'E') | |
6718 | { | |
a12be13c | 6719 | int j; |
64266206 | 6720 | |
6721 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
6722 | if (label_mentioned_p (XVECEXP (x, i, j))) | |
6723 | return 1; | |
6724 | } | |
6725 | else if (fmt[i] == 'e' && label_mentioned_p (XEXP (x, i))) | |
6726 | return 1; | |
6727 | } | |
6728 | ||
6729 | return 0; | |
6730 | } | |
6731 | ||
f655717d | 6732 | int |
6733 | tls_mentioned_p (rtx x) | |
6734 | { | |
6735 | switch (GET_CODE (x)) | |
6736 | { | |
6737 | case CONST: | |
6738 | return tls_mentioned_p (XEXP (x, 0)); | |
6739 | ||
6740 | case UNSPEC: | |
6741 | if (XINT (x, 1) == UNSPEC_TLS) | |
6742 | return 1; | |
6743 | ||
6744 | default: | |
6745 | return 0; | |
6746 | } | |
6747 | } | |
6748 | ||
6749 | /* Must not copy a SET whose source operand is PC-relative. */ | |
6750 | ||
6751 | static bool | |
6752 | arm_cannot_copy_insn_p (rtx insn) | |
6753 | { | |
6754 | rtx pat = PATTERN (insn); | |
6755 | ||
beef0fb5 | 6756 | if (GET_CODE (pat) == SET) |
f655717d | 6757 | { |
beef0fb5 | 6758 | rtx rhs = SET_SRC (pat); |
f655717d | 6759 | |
6760 | if (GET_CODE (rhs) == UNSPEC | |
6761 | && XINT (rhs, 1) == UNSPEC_PIC_BASE) | |
6762 | return TRUE; | |
6763 | ||
6764 | if (GET_CODE (rhs) == MEM | |
6765 | && GET_CODE (XEXP (rhs, 0)) == UNSPEC | |
6766 | && XINT (XEXP (rhs, 0), 1) == UNSPEC_PIC_BASE) | |
6767 | return TRUE; | |
6768 | } | |
6769 | ||
6770 | return FALSE; | |
6771 | } | |
6772 | ||
9c08d1fa | 6773 | enum rtx_code |
ebd88a36 | 6774 | minmax_code (rtx x) |
9c08d1fa | 6775 | { |
6776 | enum rtx_code code = GET_CODE (x); | |
6777 | ||
ed29c566 | 6778 | switch (code) |
6779 | { | |
6780 | case SMAX: | |
6781 | return GE; | |
6782 | case SMIN: | |
6783 | return LE; | |
6784 | case UMIN: | |
6785 | return LEU; | |
6786 | case UMAX: | |
6787 | return GEU; | |
6788 | default: | |
6789 | gcc_unreachable (); | |
6790 | } | |
9c08d1fa | 6791 | } |
6792 | ||
66aacf41 | 6793 | /* Return 1 if memory locations are adjacent. */ |
a690ede2 | 6794 | int |
ebd88a36 | 6795 | adjacent_mem_locations (rtx a, rtx b) |
9c08d1fa | 6796 | { |
94dee231 | 6797 | /* We don't guarantee to preserve the order of these memory refs. */ |
6798 | if (volatile_refs_p (a) || volatile_refs_p (b)) | |
6799 | return 0; | |
6800 | ||
9c08d1fa | 6801 | if ((GET_CODE (XEXP (a, 0)) == REG |
6802 | || (GET_CODE (XEXP (a, 0)) == PLUS | |
6803 | && GET_CODE (XEXP (XEXP (a, 0), 1)) == CONST_INT)) | |
6804 | && (GET_CODE (XEXP (b, 0)) == REG | |
6805 | || (GET_CODE (XEXP (b, 0)) == PLUS | |
6806 | && GET_CODE (XEXP (XEXP (b, 0), 1)) == CONST_INT))) | |
6807 | { | |
841b213d | 6808 | HOST_WIDE_INT val0 = 0, val1 = 0; |
6809 | rtx reg0, reg1; | |
6810 | int val_diff; | |
9e7454d0 | 6811 | |
9c08d1fa | 6812 | if (GET_CODE (XEXP (a, 0)) == PLUS) |
6813 | { | |
841b213d | 6814 | reg0 = XEXP (XEXP (a, 0), 0); |
9c08d1fa | 6815 | val0 = INTVAL (XEXP (XEXP (a, 0), 1)); |
6816 | } | |
6817 | else | |
841b213d | 6818 | reg0 = XEXP (a, 0); |
a12be13c | 6819 | |
9c08d1fa | 6820 | if (GET_CODE (XEXP (b, 0)) == PLUS) |
6821 | { | |
841b213d | 6822 | reg1 = XEXP (XEXP (b, 0), 0); |
9c08d1fa | 6823 | val1 = INTVAL (XEXP (XEXP (b, 0), 1)); |
6824 | } | |
6825 | else | |
841b213d | 6826 | reg1 = XEXP (b, 0); |
a12be13c | 6827 | |
ebd88a36 | 6828 | /* Don't accept any offset that will require multiple |
6829 | instructions to handle, since this would cause the | |
6830 | arith_adjacentmem pattern to output an overlong sequence. */ | |
ce8e2d2d | 6831 | if (!const_ok_for_op (PLUS, val0) || !const_ok_for_op (PLUS, val1)) |
6832 | return 0; | |
9e7454d0 | 6833 | |
841b213d | 6834 | /* Don't allow an eliminable register: register elimination can make |
6835 | the offset too large. */ | |
6836 | if (arm_eliminable_register (reg0)) | |
6837 | return 0; | |
6838 | ||
6839 | val_diff = val1 - val0; | |
94dee231 | 6840 | |
6841 | if (arm_ld_sched) | |
6842 | { | |
6843 | /* If the target has load delay slots, then there's no benefit | |
6844 | to using an ldm instruction unless the offset is zero and | |
6845 | we are optimizing for size. */ | |
6846 | return (optimize_size && (REGNO (reg0) == REGNO (reg1)) | |
6847 | && (val0 == 0 || val1 == 0 || val0 == 4 || val1 == 4) | |
6848 | && (val_diff == 4 || val_diff == -4)); | |
6849 | } | |
6850 | ||
841b213d | 6851 | return ((REGNO (reg0) == REGNO (reg1)) |
6852 | && (val_diff == 4 || val_diff == -4)); | |
9c08d1fa | 6853 | } |
841b213d | 6854 | |
9c08d1fa | 6855 | return 0; |
6856 | } | |
6857 | ||
1275b964 | 6858 | int |
ebd88a36 | 6859 | load_multiple_sequence (rtx *operands, int nops, int *regs, int *base, |
6860 | HOST_WIDE_INT *load_offset) | |
1275b964 | 6861 | { |
6862 | int unsorted_regs[4]; | |
6863 | HOST_WIDE_INT unsorted_offsets[4]; | |
6864 | int order[4]; | |
ecd3fcc7 | 6865 | int base_reg = -1; |
1275b964 | 6866 | int i; |
6867 | ||
a12be13c | 6868 | /* Can only handle 2, 3, or 4 insns at present, |
6869 | though could be easily extended if required. */ | |
ed29c566 | 6870 | gcc_assert (nops >= 2 && nops <= 4); |
1275b964 | 6871 | |
6872 | /* Loop over the operands and check that the memory references are | |
a361b456 | 6873 | suitable (i.e. immediate offsets from the same base register). At |
1275b964 | 6874 | the same time, extract the target register, and the memory |
6875 | offsets. */ | |
6876 | for (i = 0; i < nops; i++) | |
6877 | { | |
6878 | rtx reg; | |
6879 | rtx offset; | |
6880 | ||
6019008a | 6881 | /* Convert a subreg of a mem into the mem itself. */ |
6882 | if (GET_CODE (operands[nops + i]) == SUBREG) | |
f2dd364c | 6883 | operands[nops + i] = alter_subreg (operands + (nops + i)); |
6019008a | 6884 | |
ed29c566 | 6885 | gcc_assert (GET_CODE (operands[nops + i]) == MEM); |
1275b964 | 6886 | |
6887 | /* Don't reorder volatile memory references; it doesn't seem worth | |
6888 | looking for the case where the order is ok anyway. */ | |
6889 | if (MEM_VOLATILE_P (operands[nops + i])) | |
6890 | return 0; | |
6891 | ||
6892 | offset = const0_rtx; | |
6893 | ||
6894 | if ((GET_CODE (reg = XEXP (operands[nops + i], 0)) == REG | |
6895 | || (GET_CODE (reg) == SUBREG | |
6896 | && GET_CODE (reg = SUBREG_REG (reg)) == REG)) | |
6897 | || (GET_CODE (XEXP (operands[nops + i], 0)) == PLUS | |
6898 | && ((GET_CODE (reg = XEXP (XEXP (operands[nops + i], 0), 0)) | |
6899 | == REG) | |
6900 | || (GET_CODE (reg) == SUBREG | |
6901 | && GET_CODE (reg = SUBREG_REG (reg)) == REG)) | |
6902 | && (GET_CODE (offset = XEXP (XEXP (operands[nops + i], 0), 1)) | |
6903 | == CONST_INT))) | |
6904 | { | |
6905 | if (i == 0) | |
6906 | { | |
cffb2a26 | 6907 | base_reg = REGNO (reg); |
1275b964 | 6908 | unsorted_regs[0] = (GET_CODE (operands[i]) == REG |
6909 | ? REGNO (operands[i]) | |
6910 | : REGNO (SUBREG_REG (operands[i]))); | |
6911 | order[0] = 0; | |
6912 | } | |
9e7454d0 | 6913 | else |
1275b964 | 6914 | { |
66aacf41 | 6915 | if (base_reg != (int) REGNO (reg)) |
1275b964 | 6916 | /* Not addressed from the same base register. */ |
6917 | return 0; | |
6918 | ||
6919 | unsorted_regs[i] = (GET_CODE (operands[i]) == REG | |
6920 | ? REGNO (operands[i]) | |
6921 | : REGNO (SUBREG_REG (operands[i]))); | |
6922 | if (unsorted_regs[i] < unsorted_regs[order[0]]) | |
6923 | order[0] = i; | |
6924 | } | |
6925 | ||
6926 | /* If it isn't an integer register, or if it overwrites the | |
6927 | base register but isn't the last insn in the list, then | |
6928 | we can't do this. */ | |
6929 | if (unsorted_regs[i] < 0 || unsorted_regs[i] > 14 | |
6930 | || (i != nops - 1 && unsorted_regs[i] == base_reg)) | |
6931 | return 0; | |
6932 | ||
6933 | unsorted_offsets[i] = INTVAL (offset); | |
6934 | } | |
6935 | else | |
6936 | /* Not a suitable memory address. */ | |
6937 | return 0; | |
6938 | } | |
6939 | ||
6940 | /* All the useful information has now been extracted from the | |
6941 | operands into unsorted_regs and unsorted_offsets; additionally, | |
6942 | order[0] has been set to the lowest numbered register in the | |
6943 | list. Sort the registers into order, and check that the memory | |
6944 | offsets are ascending and adjacent. */ | |
6945 | ||
6946 | for (i = 1; i < nops; i++) | |
6947 | { | |
6948 | int j; | |
6949 | ||
6950 | order[i] = order[i - 1]; | |
6951 | for (j = 0; j < nops; j++) | |
6952 | if (unsorted_regs[j] > unsorted_regs[order[i - 1]] | |
6953 | && (order[i] == order[i - 1] | |
6954 | || unsorted_regs[j] < unsorted_regs[order[i]])) | |
6955 | order[i] = j; | |
6956 | ||
6957 | /* Have we found a suitable register? if not, one must be used more | |
6958 | than once. */ | |
6959 | if (order[i] == order[i - 1]) | |
6960 | return 0; | |
6961 | ||
6962 | /* Is the memory address adjacent and ascending? */ | |
6963 | if (unsorted_offsets[order[i]] != unsorted_offsets[order[i - 1]] + 4) | |
6964 | return 0; | |
6965 | } | |
6966 | ||
6967 | if (base) | |
6968 | { | |
6969 | *base = base_reg; | |
6970 | ||
6971 | for (i = 0; i < nops; i++) | |
6972 | regs[i] = unsorted_regs[order[i]]; | |
6973 | ||
6974 | *load_offset = unsorted_offsets[order[0]]; | |
6975 | } | |
6976 | ||
6977 | if (unsorted_offsets[order[0]] == 0) | |
6978 | return 1; /* ldmia */ | |
6979 | ||
25f905c2 | 6980 | if (TARGET_ARM && unsorted_offsets[order[0]] == 4) |
1275b964 | 6981 | return 2; /* ldmib */ |
6982 | ||
25f905c2 | 6983 | if (TARGET_ARM && unsorted_offsets[order[nops - 1]] == 0) |
1275b964 | 6984 | return 3; /* ldmda */ |
6985 | ||
6986 | if (unsorted_offsets[order[nops - 1]] == -4) | |
6987 | return 4; /* ldmdb */ | |
6988 | ||
56d27660 | 6989 | /* For ARM8,9 & StrongARM, 2 ldr instructions are faster than an ldm |
6990 | if the offset isn't small enough. The reason 2 ldrs are faster | |
6991 | is because these ARMs are able to do more than one cache access | |
6992 | in a single cycle. The ARM9 and StrongARM have Harvard caches, | |
6993 | whilst the ARM8 has a double bandwidth cache. This means that | |
6994 | these cores can do both an instruction fetch and a data fetch in | |
6995 | a single cycle, so the trick of calculating the address into a | |
6996 | scratch register (one of the result regs) and then doing a load | |
6997 | multiple actually becomes slower (and no smaller in code size). | |
6998 | That is the transformation | |
9e7454d0 | 6999 | |
7bd8ccc9 | 7000 | ldr rd1, [rbase + offset] |
7001 | ldr rd2, [rbase + offset + 4] | |
9e7454d0 | 7002 | |
7bd8ccc9 | 7003 | to |
9e7454d0 | 7004 | |
7bd8ccc9 | 7005 | add rd1, rbase, offset |
7006 | ldmia rd1, {rd1, rd2} | |
9e7454d0 | 7007 | |
56d27660 | 7008 | produces worse code -- '3 cycles + any stalls on rd2' instead of |
7009 | '2 cycles + any stalls on rd2'. On ARMs with only one cache | |
7010 | access per cycle, the first sequence could never complete in less | |
7011 | than 6 cycles, whereas the ldm sequence would only take 5 and | |
7012 | would make better use of sequential accesses if not hitting the | |
7013 | cache. | |
7014 | ||
7015 | We cheat here and test 'arm_ld_sched' which we currently know to | |
7016 | only be true for the ARM8, ARM9 and StrongARM. If this ever | |
7017 | changes, then the test below needs to be reworked. */ | |
9888ad6d | 7018 | if (nops == 2 && arm_ld_sched) |
129a2fe4 | 7019 | return 0; |
7020 | ||
1275b964 | 7021 | /* Can't do it without setting up the offset, only do this if it takes |
7022 | no more than one insn. */ | |
9e7454d0 | 7023 | return (const_ok_for_arm (unsorted_offsets[order[0]]) |
1275b964 | 7024 | || const_ok_for_arm (-unsorted_offsets[order[0]])) ? 5 : 0; |
7025 | } | |
7026 | ||
0bd59681 | 7027 | const char * |
ebd88a36 | 7028 | emit_ldm_seq (rtx *operands, int nops) |
1275b964 | 7029 | { |
7030 | int regs[4]; | |
7031 | int base_reg; | |
7032 | HOST_WIDE_INT offset; | |
7033 | char buf[100]; | |
7034 | int i; | |
7035 | ||
7036 | switch (load_multiple_sequence (operands, nops, regs, &base_reg, &offset)) | |
7037 | { | |
7038 | case 1: | |
25f905c2 | 7039 | strcpy (buf, "ldm%(ia%)\t"); |
1275b964 | 7040 | break; |
7041 | ||
7042 | case 2: | |
25f905c2 | 7043 | strcpy (buf, "ldm%(ib%)\t"); |
1275b964 | 7044 | break; |
7045 | ||
7046 | case 3: | |
25f905c2 | 7047 | strcpy (buf, "ldm%(da%)\t"); |
1275b964 | 7048 | break; |
7049 | ||
7050 | case 4: | |
25f905c2 | 7051 | strcpy (buf, "ldm%(db%)\t"); |
1275b964 | 7052 | break; |
7053 | ||
7054 | case 5: | |
7055 | if (offset >= 0) | |
7056 | sprintf (buf, "add%%?\t%s%s, %s%s, #%ld", REGISTER_PREFIX, | |
7057 | reg_names[regs[0]], REGISTER_PREFIX, reg_names[base_reg], | |
7058 | (long) offset); | |
7059 | else | |
7060 | sprintf (buf, "sub%%?\t%s%s, %s%s, #%ld", REGISTER_PREFIX, | |
7061 | reg_names[regs[0]], REGISTER_PREFIX, reg_names[base_reg], | |
7062 | (long) -offset); | |
7063 | output_asm_insn (buf, operands); | |
7064 | base_reg = regs[0]; | |
25f905c2 | 7065 | strcpy (buf, "ldm%(ia%)\t"); |
1275b964 | 7066 | break; |
7067 | ||
7068 | default: | |
ed29c566 | 7069 | gcc_unreachable (); |
1275b964 | 7070 | } |
7071 | ||
9e7454d0 | 7072 | sprintf (buf + strlen (buf), "%s%s, {%s%s", REGISTER_PREFIX, |
1275b964 | 7073 | reg_names[base_reg], REGISTER_PREFIX, reg_names[regs[0]]); |
7074 | ||
7075 | for (i = 1; i < nops; i++) | |
7076 | sprintf (buf + strlen (buf), ", %s%s", REGISTER_PREFIX, | |
7077 | reg_names[regs[i]]); | |
7078 | ||
7079 | strcat (buf, "}\t%@ phole ldm"); | |
7080 | ||
7081 | output_asm_insn (buf, operands); | |
7082 | return ""; | |
7083 | } | |
7084 | ||
7085 | int | |
ebd88a36 | 7086 | store_multiple_sequence (rtx *operands, int nops, int *regs, int *base, |
7087 | HOST_WIDE_INT * load_offset) | |
1275b964 | 7088 | { |
7089 | int unsorted_regs[4]; | |
7090 | HOST_WIDE_INT unsorted_offsets[4]; | |
7091 | int order[4]; | |
ecd3fcc7 | 7092 | int base_reg = -1; |
1275b964 | 7093 | int i; |
7094 | ||
7095 | /* Can only handle 2, 3, or 4 insns at present, though could be easily | |
7096 | extended if required. */ | |
ed29c566 | 7097 | gcc_assert (nops >= 2 && nops <= 4); |
1275b964 | 7098 | |
7099 | /* Loop over the operands and check that the memory references are | |
a361b456 | 7100 | suitable (i.e. immediate offsets from the same base register). At |
1275b964 | 7101 | the same time, extract the target register, and the memory |
7102 | offsets. */ | |
7103 | for (i = 0; i < nops; i++) | |
7104 | { | |
7105 | rtx reg; | |
7106 | rtx offset; | |
7107 | ||
6019008a | 7108 | /* Convert a subreg of a mem into the mem itself. */ |
7109 | if (GET_CODE (operands[nops + i]) == SUBREG) | |
f2dd364c | 7110 | operands[nops + i] = alter_subreg (operands + (nops + i)); |
6019008a | 7111 | |
ed29c566 | 7112 | gcc_assert (GET_CODE (operands[nops + i]) == MEM); |
1275b964 | 7113 | |
7114 | /* Don't reorder volatile memory references; it doesn't seem worth | |
7115 | looking for the case where the order is ok anyway. */ | |
7116 | if (MEM_VOLATILE_P (operands[nops + i])) | |
7117 | return 0; | |
7118 | ||
7119 | offset = const0_rtx; | |
7120 | ||
7121 | if ((GET_CODE (reg = XEXP (operands[nops + i], 0)) == REG | |
7122 | || (GET_CODE (reg) == SUBREG | |
7123 | && GET_CODE (reg = SUBREG_REG (reg)) == REG)) | |
7124 | || (GET_CODE (XEXP (operands[nops + i], 0)) == PLUS | |
7125 | && ((GET_CODE (reg = XEXP (XEXP (operands[nops + i], 0), 0)) | |
7126 | == REG) | |
7127 | || (GET_CODE (reg) == SUBREG | |
7128 | && GET_CODE (reg = SUBREG_REG (reg)) == REG)) | |
7129 | && (GET_CODE (offset = XEXP (XEXP (operands[nops + i], 0), 1)) | |
7130 | == CONST_INT))) | |
7131 | { | |
7132 | if (i == 0) | |
7133 | { | |
bba10fb8 | 7134 | base_reg = REGNO (reg); |
1275b964 | 7135 | unsorted_regs[0] = (GET_CODE (operands[i]) == REG |
7136 | ? REGNO (operands[i]) | |
7137 | : REGNO (SUBREG_REG (operands[i]))); | |
7138 | order[0] = 0; | |
7139 | } | |
9e7454d0 | 7140 | else |
1275b964 | 7141 | { |
66aacf41 | 7142 | if (base_reg != (int) REGNO (reg)) |
1275b964 | 7143 | /* Not addressed from the same base register. */ |
7144 | return 0; | |
7145 | ||
7146 | unsorted_regs[i] = (GET_CODE (operands[i]) == REG | |
7147 | ? REGNO (operands[i]) | |
7148 | : REGNO (SUBREG_REG (operands[i]))); | |
7149 | if (unsorted_regs[i] < unsorted_regs[order[0]]) | |
7150 | order[0] = i; | |
7151 | } | |
7152 | ||
7153 | /* If it isn't an integer register, then we can't do this. */ | |
7154 | if (unsorted_regs[i] < 0 || unsorted_regs[i] > 14) | |
7155 | return 0; | |
7156 | ||
7157 | unsorted_offsets[i] = INTVAL (offset); | |
7158 | } | |
7159 | else | |
7160 | /* Not a suitable memory address. */ | |
7161 | return 0; | |
7162 | } | |
7163 | ||
7164 | /* All the useful information has now been extracted from the | |
7165 | operands into unsorted_regs and unsorted_offsets; additionally, | |
7166 | order[0] has been set to the lowest numbered register in the | |
7167 | list. Sort the registers into order, and check that the memory | |
7168 | offsets are ascending and adjacent. */ | |
7169 | ||
7170 | for (i = 1; i < nops; i++) | |
7171 | { | |
7172 | int j; | |
7173 | ||
7174 | order[i] = order[i - 1]; | |
7175 | for (j = 0; j < nops; j++) | |
7176 | if (unsorted_regs[j] > unsorted_regs[order[i - 1]] | |
7177 | && (order[i] == order[i - 1] | |
7178 | || unsorted_regs[j] < unsorted_regs[order[i]])) | |
7179 | order[i] = j; | |
7180 | ||
7181 | /* Have we found a suitable register? if not, one must be used more | |
7182 | than once. */ | |
7183 | if (order[i] == order[i - 1]) | |
7184 | return 0; | |
7185 | ||
7186 | /* Is the memory address adjacent and ascending? */ | |
7187 | if (unsorted_offsets[order[i]] != unsorted_offsets[order[i - 1]] + 4) | |
7188 | return 0; | |
7189 | } | |
7190 | ||
7191 | if (base) | |
7192 | { | |
7193 | *base = base_reg; | |
7194 | ||
7195 | for (i = 0; i < nops; i++) | |
7196 | regs[i] = unsorted_regs[order[i]]; | |
7197 | ||
7198 | *load_offset = unsorted_offsets[order[0]]; | |
7199 | } | |
7200 | ||
7201 | if (unsorted_offsets[order[0]] == 0) | |
7202 | return 1; /* stmia */ | |
7203 | ||
7204 | if (unsorted_offsets[order[0]] == 4) | |
7205 | return 2; /* stmib */ | |
7206 | ||
7207 | if (unsorted_offsets[order[nops - 1]] == 0) | |
7208 | return 3; /* stmda */ | |
7209 | ||
7210 | if (unsorted_offsets[order[nops - 1]] == -4) | |
7211 | return 4; /* stmdb */ | |
7212 | ||
7213 | return 0; | |
7214 | } | |
7215 | ||
0bd59681 | 7216 | const char * |
ebd88a36 | 7217 | emit_stm_seq (rtx *operands, int nops) |
1275b964 | 7218 | { |
7219 | int regs[4]; | |
7220 | int base_reg; | |
7221 | HOST_WIDE_INT offset; | |
7222 | char buf[100]; | |
7223 | int i; | |
7224 | ||
7225 | switch (store_multiple_sequence (operands, nops, regs, &base_reg, &offset)) | |
7226 | { | |
7227 | case 1: | |
25f905c2 | 7228 | strcpy (buf, "stm%(ia%)\t"); |
1275b964 | 7229 | break; |
7230 | ||
7231 | case 2: | |
25f905c2 | 7232 | strcpy (buf, "stm%(ib%)\t"); |
1275b964 | 7233 | break; |
7234 | ||
7235 | case 3: | |
25f905c2 | 7236 | strcpy (buf, "stm%(da%)\t"); |
1275b964 | 7237 | break; |
7238 | ||
7239 | case 4: | |
25f905c2 | 7240 | strcpy (buf, "stm%(db%)\t"); |
1275b964 | 7241 | break; |
7242 | ||
7243 | default: | |
ed29c566 | 7244 | gcc_unreachable (); |
1275b964 | 7245 | } |
7246 | ||
9e7454d0 | 7247 | sprintf (buf + strlen (buf), "%s%s, {%s%s", REGISTER_PREFIX, |
1275b964 | 7248 | reg_names[base_reg], REGISTER_PREFIX, reg_names[regs[0]]); |
7249 | ||
7250 | for (i = 1; i < nops; i++) | |
7251 | sprintf (buf + strlen (buf), ", %s%s", REGISTER_PREFIX, | |
7252 | reg_names[regs[i]]); | |
7253 | ||
7254 | strcat (buf, "}\t%@ phole stm"); | |
7255 | ||
7256 | output_asm_insn (buf, operands); | |
7257 | return ""; | |
7258 | } | |
9c08d1fa | 7259 | \f |
66aacf41 | 7260 | /* Routines for use in generating RTL. */ |
a12be13c | 7261 | |
a690ede2 | 7262 | rtx |
ebd88a36 | 7263 | arm_gen_load_multiple (int base_regno, int count, rtx from, int up, |
9580c25f | 7264 | int write_back, rtx basemem, HOST_WIDE_INT *offsetp) |
9c08d1fa | 7265 | { |
9580c25f | 7266 | HOST_WIDE_INT offset = *offsetp; |
9c08d1fa | 7267 | int i = 0, j; |
7268 | rtx result; | |
7269 | int sign = up ? 1 : -1; | |
9580c25f | 7270 | rtx mem, addr; |
9c08d1fa | 7271 | |
331beb1a | 7272 | /* XScale has load-store double instructions, but they have stricter |
33f88b1c | 7273 | alignment requirements than load-store multiple, so we cannot |
331beb1a | 7274 | use them. |
7275 | ||
7276 | For XScale ldm requires 2 + NREGS cycles to complete and blocks | |
7277 | the pipeline until completion. | |
7278 | ||
7279 | NREGS CYCLES | |
7280 | 1 3 | |
7281 | 2 4 | |
7282 | 3 5 | |
7283 | 4 6 | |
7284 | ||
7285 | An ldr instruction takes 1-3 cycles, but does not block the | |
7286 | pipeline. | |
7287 | ||
7288 | NREGS CYCLES | |
7289 | 1 1-3 | |
7290 | 2 2-6 | |
7291 | 3 3-9 | |
7292 | 4 4-12 | |
7293 | ||
7294 | Best case ldr will always win. However, the more ldr instructions | |
7295 | we issue, the less likely we are to be able to schedule them well. | |
7296 | Using ldr instructions also increases code size. | |
7297 | ||
7298 | As a compromise, we use ldr for counts of 1 or 2 regs, and ldm | |
7299 | for counts of 3 or 4 regs. */ | |
f8e17552 | 7300 | if (arm_tune_xscale && count <= 2 && ! optimize_size) |
331beb1a | 7301 | { |
7302 | rtx seq; | |
9e7454d0 | 7303 | |
331beb1a | 7304 | start_sequence (); |
9e7454d0 | 7305 | |
331beb1a | 7306 | for (i = 0; i < count; i++) |
7307 | { | |
9580c25f | 7308 | addr = plus_constant (from, i * 4 * sign); |
7309 | mem = adjust_automodify_address (basemem, SImode, addr, offset); | |
331beb1a | 7310 | emit_move_insn (gen_rtx_REG (SImode, base_regno + i), mem); |
9580c25f | 7311 | offset += 4 * sign; |
331beb1a | 7312 | } |
7313 | ||
7314 | if (write_back) | |
9580c25f | 7315 | { |
7316 | emit_move_insn (from, plus_constant (from, count * 4 * sign)); | |
7317 | *offsetp = offset; | |
7318 | } | |
331beb1a | 7319 | |
31d3e01c | 7320 | seq = get_insns (); |
331beb1a | 7321 | end_sequence (); |
9e7454d0 | 7322 | |
331beb1a | 7323 | return seq; |
7324 | } | |
7325 | ||
76676c8e | 7326 | result = gen_rtx_PARALLEL (VOIDmode, |
dd6d7504 | 7327 | rtvec_alloc (count + (write_back ? 1 : 0))); |
9c08d1fa | 7328 | if (write_back) |
a690ede2 | 7329 | { |
9c08d1fa | 7330 | XVECEXP (result, 0, 0) |
0083a3b9 | 7331 | = gen_rtx_SET (VOIDmode, from, plus_constant (from, count * 4 * sign)); |
9c08d1fa | 7332 | i = 1; |
7333 | count++; | |
a690ede2 | 7334 | } |
7335 | ||
9c08d1fa | 7336 | for (j = 0; i < count; i++, j++) |
a690ede2 | 7337 | { |
9580c25f | 7338 | addr = plus_constant (from, j * 4 * sign); |
7339 | mem = adjust_automodify_address_nv (basemem, SImode, addr, offset); | |
76676c8e | 7340 | XVECEXP (result, 0, i) |
7341 | = gen_rtx_SET (VOIDmode, gen_rtx_REG (SImode, base_regno + j), mem); | |
9580c25f | 7342 | offset += 4 * sign; |
a690ede2 | 7343 | } |
7344 | ||
9580c25f | 7345 | if (write_back) |
7346 | *offsetp = offset; | |
7347 | ||
9c08d1fa | 7348 | return result; |
7349 | } | |
7350 | ||
a690ede2 | 7351 | rtx |
ebd88a36 | 7352 | arm_gen_store_multiple (int base_regno, int count, rtx to, int up, |
9580c25f | 7353 | int write_back, rtx basemem, HOST_WIDE_INT *offsetp) |
9c08d1fa | 7354 | { |
9580c25f | 7355 | HOST_WIDE_INT offset = *offsetp; |
9c08d1fa | 7356 | int i = 0, j; |
7357 | rtx result; | |
7358 | int sign = up ? 1 : -1; | |
9580c25f | 7359 | rtx mem, addr; |
9c08d1fa | 7360 | |
331beb1a | 7361 | /* See arm_gen_load_multiple for discussion of |
7362 | the pros/cons of ldm/stm usage for XScale. */ | |
f8e17552 | 7363 | if (arm_tune_xscale && count <= 2 && ! optimize_size) |
331beb1a | 7364 | { |
7365 | rtx seq; | |
9e7454d0 | 7366 | |
331beb1a | 7367 | start_sequence (); |
9e7454d0 | 7368 | |
331beb1a | 7369 | for (i = 0; i < count; i++) |
7370 | { | |
9580c25f | 7371 | addr = plus_constant (to, i * 4 * sign); |
7372 | mem = adjust_automodify_address (basemem, SImode, addr, offset); | |
331beb1a | 7373 | emit_move_insn (mem, gen_rtx_REG (SImode, base_regno + i)); |
9580c25f | 7374 | offset += 4 * sign; |
331beb1a | 7375 | } |
7376 | ||
7377 | if (write_back) | |
9580c25f | 7378 | { |
7379 | emit_move_insn (to, plus_constant (to, count * 4 * sign)); | |
7380 | *offsetp = offset; | |
7381 | } | |
331beb1a | 7382 | |
31d3e01c | 7383 | seq = get_insns (); |
331beb1a | 7384 | end_sequence (); |
9e7454d0 | 7385 | |
331beb1a | 7386 | return seq; |
7387 | } | |
7388 | ||
76676c8e | 7389 | result = gen_rtx_PARALLEL (VOIDmode, |
dd6d7504 | 7390 | rtvec_alloc (count + (write_back ? 1 : 0))); |
9c08d1fa | 7391 | if (write_back) |
a690ede2 | 7392 | { |
9c08d1fa | 7393 | XVECEXP (result, 0, 0) |
0083a3b9 | 7394 | = gen_rtx_SET (VOIDmode, to, |
76676c8e | 7395 | plus_constant (to, count * 4 * sign)); |
9c08d1fa | 7396 | i = 1; |
7397 | count++; | |
a690ede2 | 7398 | } |
7399 | ||
9c08d1fa | 7400 | for (j = 0; i < count; i++, j++) |
a690ede2 | 7401 | { |
9580c25f | 7402 | addr = plus_constant (to, j * 4 * sign); |
7403 | mem = adjust_automodify_address_nv (basemem, SImode, addr, offset); | |
76676c8e | 7404 | XVECEXP (result, 0, i) |
7405 | = gen_rtx_SET (VOIDmode, mem, gen_rtx_REG (SImode, base_regno + j)); | |
9580c25f | 7406 | offset += 4 * sign; |
a690ede2 | 7407 | } |
7408 | ||
9580c25f | 7409 | if (write_back) |
7410 | *offsetp = offset; | |
7411 | ||
9c08d1fa | 7412 | return result; |
7413 | } | |
7414 | ||
c29feb68 | 7415 | int |
008c057d | 7416 | arm_gen_movmemqi (rtx *operands) |
c29feb68 | 7417 | { |
7418 | HOST_WIDE_INT in_words_to_go, out_words_to_go, last_bytes; | |
9580c25f | 7419 | HOST_WIDE_INT srcoffset, dstoffset; |
ecd3fcc7 | 7420 | int i; |
9580c25f | 7421 | rtx src, dst, srcbase, dstbase; |
c29feb68 | 7422 | rtx part_bytes_reg = NULL; |
6019008a | 7423 | rtx mem; |
c29feb68 | 7424 | |
7425 | if (GET_CODE (operands[2]) != CONST_INT | |
7426 | || GET_CODE (operands[3]) != CONST_INT | |
7427 | || INTVAL (operands[2]) > 64 | |
7428 | || INTVAL (operands[3]) & 3) | |
7429 | return 0; | |
7430 | ||
9580c25f | 7431 | dstbase = operands[0]; |
7432 | srcbase = operands[1]; | |
6019008a | 7433 | |
9580c25f | 7434 | dst = copy_to_mode_reg (SImode, XEXP (dstbase, 0)); |
7435 | src = copy_to_mode_reg (SImode, XEXP (srcbase, 0)); | |
c29feb68 | 7436 | |
36837fde | 7437 | in_words_to_go = ARM_NUM_INTS (INTVAL (operands[2])); |
c29feb68 | 7438 | out_words_to_go = INTVAL (operands[2]) / 4; |
7439 | last_bytes = INTVAL (operands[2]) & 3; | |
9580c25f | 7440 | dstoffset = srcoffset = 0; |
c29feb68 | 7441 | |
7442 | if (out_words_to_go != in_words_to_go && ((in_words_to_go - 1) & 3) != 0) | |
76676c8e | 7443 | part_bytes_reg = gen_rtx_REG (SImode, (in_words_to_go - 1) & 3); |
c29feb68 | 7444 | |
7445 | for (i = 0; in_words_to_go >= 2; i+=4) | |
7446 | { | |
9480e6c7 | 7447 | if (in_words_to_go > 4) |
6019008a | 7448 | emit_insn (arm_gen_load_multiple (0, 4, src, TRUE, TRUE, |
9580c25f | 7449 | srcbase, &srcoffset)); |
9480e6c7 | 7450 | else |
9e7454d0 | 7451 | emit_insn (arm_gen_load_multiple (0, in_words_to_go, src, TRUE, |
9580c25f | 7452 | FALSE, srcbase, &srcoffset)); |
9480e6c7 | 7453 | |
c29feb68 | 7454 | if (out_words_to_go) |
7455 | { | |
9480e6c7 | 7456 | if (out_words_to_go > 4) |
6019008a | 7457 | emit_insn (arm_gen_store_multiple (0, 4, dst, TRUE, TRUE, |
9580c25f | 7458 | dstbase, &dstoffset)); |
9480e6c7 | 7459 | else if (out_words_to_go != 1) |
7460 | emit_insn (arm_gen_store_multiple (0, out_words_to_go, | |
9e7454d0 | 7461 | dst, TRUE, |
9480e6c7 | 7462 | (last_bytes == 0 |
6019008a | 7463 | ? FALSE : TRUE), |
9580c25f | 7464 | dstbase, &dstoffset)); |
c29feb68 | 7465 | else |
7466 | { | |
9580c25f | 7467 | mem = adjust_automodify_address (dstbase, SImode, dst, dstoffset); |
76676c8e | 7468 | emit_move_insn (mem, gen_rtx_REG (SImode, 0)); |
9480e6c7 | 7469 | if (last_bytes != 0) |
9580c25f | 7470 | { |
7471 | emit_insn (gen_addsi3 (dst, dst, GEN_INT (4))); | |
7472 | dstoffset += 4; | |
7473 | } | |
c29feb68 | 7474 | } |
7475 | } | |
7476 | ||
7477 | in_words_to_go -= in_words_to_go < 4 ? in_words_to_go : 4; | |
7478 | out_words_to_go -= out_words_to_go < 4 ? out_words_to_go : 4; | |
7479 | } | |
7480 | ||
7481 | /* OUT_WORDS_TO_GO will be zero here if there are byte stores to do. */ | |
7482 | if (out_words_to_go) | |
bba10fb8 | 7483 | { |
7484 | rtx sreg; | |
9e7454d0 | 7485 | |
9580c25f | 7486 | mem = adjust_automodify_address (srcbase, SImode, src, srcoffset); |
7487 | sreg = copy_to_reg (mem); | |
7488 | ||
7489 | mem = adjust_automodify_address (dstbase, SImode, dst, dstoffset); | |
bba10fb8 | 7490 | emit_move_insn (mem, sreg); |
bba10fb8 | 7491 | in_words_to_go--; |
9e7454d0 | 7492 | |
ed29c566 | 7493 | gcc_assert (!in_words_to_go); /* Sanity check */ |
bba10fb8 | 7494 | } |
c29feb68 | 7495 | |
7496 | if (in_words_to_go) | |
7497 | { | |
ed29c566 | 7498 | gcc_assert (in_words_to_go > 0); |
c29feb68 | 7499 | |
9580c25f | 7500 | mem = adjust_automodify_address (srcbase, SImode, src, srcoffset); |
6019008a | 7501 | part_bytes_reg = copy_to_mode_reg (SImode, mem); |
c29feb68 | 7502 | } |
7503 | ||
ed29c566 | 7504 | gcc_assert (!last_bytes || part_bytes_reg); |
cffb2a26 | 7505 | |
c29feb68 | 7506 | if (BYTES_BIG_ENDIAN && last_bytes) |
7507 | { | |
7508 | rtx tmp = gen_reg_rtx (SImode); | |
7509 | ||
66aacf41 | 7510 | /* The bytes we want are in the top end of the word. */ |
411b0ddb | 7511 | emit_insn (gen_lshrsi3 (tmp, part_bytes_reg, |
7512 | GEN_INT (8 * (4 - last_bytes)))); | |
c29feb68 | 7513 | part_bytes_reg = tmp; |
9e7454d0 | 7514 | |
c29feb68 | 7515 | while (last_bytes) |
7516 | { | |
9580c25f | 7517 | mem = adjust_automodify_address (dstbase, QImode, |
7518 | plus_constant (dst, last_bytes - 1), | |
7519 | dstoffset + last_bytes - 1); | |
9e8503e6 | 7520 | emit_move_insn (mem, gen_lowpart (QImode, part_bytes_reg)); |
7521 | ||
c29feb68 | 7522 | if (--last_bytes) |
7523 | { | |
7524 | tmp = gen_reg_rtx (SImode); | |
7525 | emit_insn (gen_lshrsi3 (tmp, part_bytes_reg, GEN_INT (8))); | |
7526 | part_bytes_reg = tmp; | |
7527 | } | |
7528 | } | |
9e7454d0 | 7529 | |
c29feb68 | 7530 | } |
7531 | else | |
7532 | { | |
cffb2a26 | 7533 | if (last_bytes > 1) |
c29feb68 | 7534 | { |
9580c25f | 7535 | mem = adjust_automodify_address (dstbase, HImode, dst, dstoffset); |
9e8503e6 | 7536 | emit_move_insn (mem, gen_lowpart (HImode, part_bytes_reg)); |
cffb2a26 | 7537 | last_bytes -= 2; |
7538 | if (last_bytes) | |
c29feb68 | 7539 | { |
7540 | rtx tmp = gen_reg_rtx (SImode); | |
bcd9bd66 | 7541 | emit_insn (gen_addsi3 (dst, dst, const2_rtx)); |
cffb2a26 | 7542 | emit_insn (gen_lshrsi3 (tmp, part_bytes_reg, GEN_INT (16))); |
c29feb68 | 7543 | part_bytes_reg = tmp; |
9580c25f | 7544 | dstoffset += 2; |
c29feb68 | 7545 | } |
7546 | } | |
9e7454d0 | 7547 | |
cffb2a26 | 7548 | if (last_bytes) |
7549 | { | |
9580c25f | 7550 | mem = adjust_automodify_address (dstbase, QImode, dst, dstoffset); |
9e8503e6 | 7551 | emit_move_insn (mem, gen_lowpart (QImode, part_bytes_reg)); |
cffb2a26 | 7552 | } |
c29feb68 | 7553 | } |
7554 | ||
7555 | return 1; | |
7556 | } | |
7557 | ||
3c5afce6 | 7558 | /* Select a dominance comparison mode if possible for a test of the general |
7559 | form (OP (COND_OR (X) (Y)) (const_int 0)). We support three forms. | |
9e7454d0 | 7560 | COND_OR == DOM_CC_X_AND_Y => (X && Y) |
3c5afce6 | 7561 | COND_OR == DOM_CC_NX_OR_Y => ((! X) || Y) |
9e7454d0 | 7562 | COND_OR == DOM_CC_X_OR_Y => (X || Y) |
3c5afce6 | 7563 | In all cases OP will be either EQ or NE, but we don't need to know which |
9e7454d0 | 7564 | here. If we are unable to support a dominance comparison we return |
3c5afce6 | 7565 | CC mode. This will then fail to match for the RTL expressions that |
7566 | generate this call. */ | |
3c5afce6 | 7567 | enum machine_mode |
ebd88a36 | 7568 | arm_select_dominance_cc_mode (rtx x, rtx y, HOST_WIDE_INT cond_or) |
1275b964 | 7569 | { |
7570 | enum rtx_code cond1, cond2; | |
7571 | int swapped = 0; | |
7572 | ||
7573 | /* Currently we will probably get the wrong result if the individual | |
7574 | comparisons are not simple. This also ensures that it is safe to | |
ad87de1e | 7575 | reverse a comparison if necessary. */ |
1275b964 | 7576 | if ((arm_select_cc_mode (cond1 = GET_CODE (x), XEXP (x, 0), XEXP (x, 1)) |
7577 | != CCmode) | |
7578 | || (arm_select_cc_mode (cond2 = GET_CODE (y), XEXP (y, 0), XEXP (y, 1)) | |
7579 | != CCmode)) | |
7580 | return CCmode; | |
7581 | ||
f6c53574 | 7582 | /* The if_then_else variant of this tests the second condition if the |
7583 | first passes, but is true if the first fails. Reverse the first | |
7584 | condition to get a true "inclusive-or" expression. */ | |
3c5afce6 | 7585 | if (cond_or == DOM_CC_NX_OR_Y) |
1275b964 | 7586 | cond1 = reverse_condition (cond1); |
7587 | ||
7588 | /* If the comparisons are not equal, and one doesn't dominate the other, | |
7589 | then we can't do this. */ | |
9e7454d0 | 7590 | if (cond1 != cond2 |
215b30b3 | 7591 | && !comparison_dominates_p (cond1, cond2) |
7592 | && (swapped = 1, !comparison_dominates_p (cond2, cond1))) | |
1275b964 | 7593 | return CCmode; |
7594 | ||
7595 | if (swapped) | |
7596 | { | |
7597 | enum rtx_code temp = cond1; | |
7598 | cond1 = cond2; | |
7599 | cond2 = temp; | |
7600 | } | |
7601 | ||
7602 | switch (cond1) | |
7603 | { | |
7604 | case EQ: | |
ed29c566 | 7605 | if (cond_or == DOM_CC_X_AND_Y) |
1275b964 | 7606 | return CC_DEQmode; |
7607 | ||
7608 | switch (cond2) | |
7609 | { | |
ed29c566 | 7610 | case EQ: return CC_DEQmode; |
1275b964 | 7611 | case LE: return CC_DLEmode; |
7612 | case LEU: return CC_DLEUmode; | |
7613 | case GE: return CC_DGEmode; | |
7614 | case GEU: return CC_DGEUmode; | |
ed29c566 | 7615 | default: gcc_unreachable (); |
1275b964 | 7616 | } |
7617 | ||
1275b964 | 7618 | case LT: |
ed29c566 | 7619 | if (cond_or == DOM_CC_X_AND_Y) |
1275b964 | 7620 | return CC_DLTmode; |
a60f1840 | 7621 | |
ed29c566 | 7622 | switch (cond2) |
7623 | { | |
7624 | case LT: | |
7625 | return CC_DLTmode; | |
7626 | case LE: | |
7627 | return CC_DLEmode; | |
7628 | case NE: | |
7629 | return CC_DNEmode; | |
7630 | default: | |
7631 | gcc_unreachable (); | |
7632 | } | |
1275b964 | 7633 | |
7634 | case GT: | |
ed29c566 | 7635 | if (cond_or == DOM_CC_X_AND_Y) |
1275b964 | 7636 | return CC_DGTmode; |
ed29c566 | 7637 | |
7638 | switch (cond2) | |
7639 | { | |
7640 | case GT: | |
7641 | return CC_DGTmode; | |
7642 | case GE: | |
7643 | return CC_DGEmode; | |
7644 | case NE: | |
7645 | return CC_DNEmode; | |
7646 | default: | |
7647 | gcc_unreachable (); | |
7648 | } | |
9e7454d0 | 7649 | |
1275b964 | 7650 | case LTU: |
ed29c566 | 7651 | if (cond_or == DOM_CC_X_AND_Y) |
1275b964 | 7652 | return CC_DLTUmode; |
ed29c566 | 7653 | |
7654 | switch (cond2) | |
7655 | { | |
7656 | case LTU: | |
7657 | return CC_DLTUmode; | |
7658 | case LEU: | |
7659 | return CC_DLEUmode; | |
7660 | case NE: | |
7661 | return CC_DNEmode; | |
7662 | default: | |
7663 | gcc_unreachable (); | |
7664 | } | |
1275b964 | 7665 | |
7666 | case GTU: | |
ed29c566 | 7667 | if (cond_or == DOM_CC_X_AND_Y) |
1275b964 | 7668 | return CC_DGTUmode; |
a60f1840 | 7669 | |
ed29c566 | 7670 | switch (cond2) |
7671 | { | |
7672 | case GTU: | |
7673 | return CC_DGTUmode; | |
7674 | case GEU: | |
7675 | return CC_DGEUmode; | |
7676 | case NE: | |
7677 | return CC_DNEmode; | |
7678 | default: | |
7679 | gcc_unreachable (); | |
7680 | } | |
1275b964 | 7681 | |
7682 | /* The remaining cases only occur when both comparisons are the | |
7683 | same. */ | |
7684 | case NE: | |
ed29c566 | 7685 | gcc_assert (cond1 == cond2); |
1275b964 | 7686 | return CC_DNEmode; |
7687 | ||
7688 | case LE: | |
ed29c566 | 7689 | gcc_assert (cond1 == cond2); |
1275b964 | 7690 | return CC_DLEmode; |
7691 | ||
7692 | case GE: | |
ed29c566 | 7693 | gcc_assert (cond1 == cond2); |
1275b964 | 7694 | return CC_DGEmode; |
7695 | ||
7696 | case LEU: | |
ed29c566 | 7697 | gcc_assert (cond1 == cond2); |
1275b964 | 7698 | return CC_DLEUmode; |
7699 | ||
7700 | case GEU: | |
ed29c566 | 7701 | gcc_assert (cond1 == cond2); |
1275b964 | 7702 | return CC_DGEUmode; |
ecd3fcc7 | 7703 | |
7704 | default: | |
ed29c566 | 7705 | gcc_unreachable (); |
1275b964 | 7706 | } |
1275b964 | 7707 | } |
7708 | ||
7709 | enum machine_mode | |
ebd88a36 | 7710 | arm_select_cc_mode (enum rtx_code op, rtx x, rtx y) |
1275b964 | 7711 | { |
7712 | /* All floating point compares return CCFP if it is an equality | |
7713 | comparison, and CCFPE otherwise. */ | |
7714 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
f3d87a93 | 7715 | { |
7716 | switch (op) | |
7717 | { | |
7718 | case EQ: | |
7719 | case NE: | |
7720 | case UNORDERED: | |
7721 | case ORDERED: | |
7722 | case UNLT: | |
7723 | case UNLE: | |
7724 | case UNGT: | |
7725 | case UNGE: | |
7726 | case UNEQ: | |
7727 | case LTGT: | |
7728 | return CCFPmode; | |
7729 | ||
7730 | case LT: | |
7731 | case LE: | |
7732 | case GT: | |
7733 | case GE: | |
a2cd141b | 7734 | if (TARGET_HARD_FLOAT && TARGET_MAVERICK) |
7d57ec45 | 7735 | return CCFPmode; |
f3d87a93 | 7736 | return CCFPEmode; |
7737 | ||
7738 | default: | |
ed29c566 | 7739 | gcc_unreachable (); |
f3d87a93 | 7740 | } |
7741 | } | |
9e7454d0 | 7742 | |
1275b964 | 7743 | /* A compare with a shifted operand. Because of canonicalization, the |
7744 | comparison will have to be swapped when we emit the assembler. */ | |
7745 | if (GET_MODE (y) == SImode && GET_CODE (y) == REG | |
7746 | && (GET_CODE (x) == ASHIFT || GET_CODE (x) == ASHIFTRT | |
7747 | || GET_CODE (x) == LSHIFTRT || GET_CODE (x) == ROTATE | |
7748 | || GET_CODE (x) == ROTATERT)) | |
7749 | return CC_SWPmode; | |
7750 | ||
4c9d8377 | 7751 | /* This operation is performed swapped, but since we only rely on the Z |
7752 | flag we don't need an additional mode. */ | |
7753 | if (GET_MODE (y) == SImode && REG_P (y) | |
7754 | && GET_CODE (x) == NEG | |
7755 | && (op == EQ || op == NE)) | |
7756 | return CC_Zmode; | |
7757 | ||
9e7454d0 | 7758 | /* This is a special case that is used by combine to allow a |
ad87de1e | 7759 | comparison of a shifted byte load to be split into a zero-extend |
1275b964 | 7760 | followed by a comparison of the shifted integer (only valid for |
ad87de1e | 7761 | equalities and unsigned inequalities). */ |
1275b964 | 7762 | if (GET_MODE (x) == SImode |
7763 | && GET_CODE (x) == ASHIFT | |
7764 | && GET_CODE (XEXP (x, 1)) == CONST_INT && INTVAL (XEXP (x, 1)) == 24 | |
7765 | && GET_CODE (XEXP (x, 0)) == SUBREG | |
7766 | && GET_CODE (SUBREG_REG (XEXP (x, 0))) == MEM | |
7767 | && GET_MODE (SUBREG_REG (XEXP (x, 0))) == QImode | |
7768 | && (op == EQ || op == NE | |
7769 | || op == GEU || op == GTU || op == LTU || op == LEU) | |
7770 | && GET_CODE (y) == CONST_INT) | |
7771 | return CC_Zmode; | |
7772 | ||
f6c53574 | 7773 | /* A construct for a conditional compare, if the false arm contains |
7774 | 0, then both conditions must be true, otherwise either condition | |
7775 | must be true. Not all conditions are possible, so CCmode is | |
7776 | returned if it can't be done. */ | |
7777 | if (GET_CODE (x) == IF_THEN_ELSE | |
7778 | && (XEXP (x, 2) == const0_rtx | |
7779 | || XEXP (x, 2) == const1_rtx) | |
6720e96c | 7780 | && COMPARISON_P (XEXP (x, 0)) |
7781 | && COMPARISON_P (XEXP (x, 1))) | |
9e7454d0 | 7782 | return arm_select_dominance_cc_mode (XEXP (x, 0), XEXP (x, 1), |
3c5afce6 | 7783 | INTVAL (XEXP (x, 2))); |
f6c53574 | 7784 | |
7785 | /* Alternate canonicalizations of the above. These are somewhat cleaner. */ | |
7786 | if (GET_CODE (x) == AND | |
6720e96c | 7787 | && COMPARISON_P (XEXP (x, 0)) |
7788 | && COMPARISON_P (XEXP (x, 1))) | |
3c5afce6 | 7789 | return arm_select_dominance_cc_mode (XEXP (x, 0), XEXP (x, 1), |
7790 | DOM_CC_X_AND_Y); | |
f6c53574 | 7791 | |
7792 | if (GET_CODE (x) == IOR | |
6720e96c | 7793 | && COMPARISON_P (XEXP (x, 0)) |
7794 | && COMPARISON_P (XEXP (x, 1))) | |
3c5afce6 | 7795 | return arm_select_dominance_cc_mode (XEXP (x, 0), XEXP (x, 1), |
7796 | DOM_CC_X_OR_Y); | |
f6c53574 | 7797 | |
58d6528b | 7798 | /* An operation (on Thumb) where we want to test for a single bit. |
7799 | This is done by shifting that bit up into the top bit of a | |
7800 | scratch register; we can then branch on the sign bit. */ | |
25f905c2 | 7801 | if (TARGET_THUMB1 |
58d6528b | 7802 | && GET_MODE (x) == SImode |
7803 | && (op == EQ || op == NE) | |
86efa74d | 7804 | && GET_CODE (x) == ZERO_EXTRACT |
7805 | && XEXP (x, 1) == const1_rtx) | |
58d6528b | 7806 | return CC_Nmode; |
7807 | ||
1275b964 | 7808 | /* An operation that sets the condition codes as a side-effect, the |
7809 | V flag is not set correctly, so we can only use comparisons where | |
7810 | this doesn't matter. (For LT and GE we can use "mi" and "pl" | |
58d6528b | 7811 | instead.) */ |
25f905c2 | 7812 | /* ??? Does the ZERO_EXTRACT case really apply to thumb2? */ |
1275b964 | 7813 | if (GET_MODE (x) == SImode |
7814 | && y == const0_rtx | |
7815 | && (op == EQ || op == NE || op == LT || op == GE) | |
7816 | && (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS | |
7817 | || GET_CODE (x) == AND || GET_CODE (x) == IOR | |
7818 | || GET_CODE (x) == XOR || GET_CODE (x) == MULT | |
7819 | || GET_CODE (x) == NOT || GET_CODE (x) == NEG | |
7820 | || GET_CODE (x) == LSHIFTRT | |
7821 | || GET_CODE (x) == ASHIFT || GET_CODE (x) == ASHIFTRT | |
58d6528b | 7822 | || GET_CODE (x) == ROTATERT |
25f905c2 | 7823 | || (TARGET_32BIT && GET_CODE (x) == ZERO_EXTRACT))) |
1275b964 | 7824 | return CC_NOOVmode; |
7825 | ||
1275b964 | 7826 | if (GET_MODE (x) == QImode && (op == EQ || op == NE)) |
7827 | return CC_Zmode; | |
7828 | ||
9480e6c7 | 7829 | if (GET_MODE (x) == SImode && (op == LTU || op == GEU) |
7830 | && GET_CODE (x) == PLUS | |
7831 | && (rtx_equal_p (XEXP (x, 0), y) || rtx_equal_p (XEXP (x, 1), y))) | |
7832 | return CC_Cmode; | |
7833 | ||
1275b964 | 7834 | return CCmode; |
7835 | } | |
7836 | ||
9c08d1fa | 7837 | /* X and Y are two things to compare using CODE. Emit the compare insn and |
7838 | return the rtx for register 0 in the proper mode. FP means this is a | |
7839 | floating point compare: I don't think that it is needed on the arm. */ | |
9c08d1fa | 7840 | rtx |
ebd88a36 | 7841 | arm_gen_compare_reg (enum rtx_code code, rtx x, rtx y) |
9c08d1fa | 7842 | { |
7843 | enum machine_mode mode = SELECT_CC_MODE (code, x, y); | |
cffb2a26 | 7844 | rtx cc_reg = gen_rtx_REG (mode, CC_REGNUM); |
9c08d1fa | 7845 | |
0083a3b9 | 7846 | emit_set_insn (cc_reg, gen_rtx_COMPARE (mode, x, y)); |
9c08d1fa | 7847 | |
7848 | return cc_reg; | |
7849 | } | |
7850 | ||
68121397 | 7851 | /* Generate a sequence of insns that will generate the correct return |
7852 | address mask depending on the physical architecture that the program | |
7853 | is running on. */ | |
68121397 | 7854 | rtx |
ebd88a36 | 7855 | arm_gen_return_addr_mask (void) |
68121397 | 7856 | { |
7857 | rtx reg = gen_reg_rtx (Pmode); | |
7858 | ||
7859 | emit_insn (gen_return_addr_mask (reg)); | |
7860 | return reg; | |
7861 | } | |
7862 | ||
91fd5a8d | 7863 | void |
ebd88a36 | 7864 | arm_reload_in_hi (rtx *operands) |
91fd5a8d | 7865 | { |
bc5c7e08 | 7866 | rtx ref = operands[1]; |
7867 | rtx base, scratch; | |
7868 | HOST_WIDE_INT offset = 0; | |
7869 | ||
7870 | if (GET_CODE (ref) == SUBREG) | |
7871 | { | |
701e46d0 | 7872 | offset = SUBREG_BYTE (ref); |
bc5c7e08 | 7873 | ref = SUBREG_REG (ref); |
7874 | } | |
7875 | ||
7876 | if (GET_CODE (ref) == REG) | |
7877 | { | |
7878 | /* We have a pseudo which has been spilt onto the stack; there | |
7879 | are two cases here: the first where there is a simple | |
7880 | stack-slot replacement and a second where the stack-slot is | |
7881 | out of range, or is used as a subreg. */ | |
7882 | if (reg_equiv_mem[REGNO (ref)]) | |
7883 | { | |
7884 | ref = reg_equiv_mem[REGNO (ref)]; | |
7885 | base = find_replacement (&XEXP (ref, 0)); | |
7886 | } | |
7887 | else | |
66aacf41 | 7888 | /* The slot is out of range, or was dressed up in a SUBREG. */ |
bc5c7e08 | 7889 | base = reg_equiv_address[REGNO (ref)]; |
7890 | } | |
7891 | else | |
7892 | base = find_replacement (&XEXP (ref, 0)); | |
91fd5a8d | 7893 | |
997d68fe | 7894 | /* Handle the case where the address is too complex to be offset by 1. */ |
7895 | if (GET_CODE (base) == MINUS | |
7896 | || (GET_CODE (base) == PLUS && GET_CODE (XEXP (base, 1)) != CONST_INT)) | |
7897 | { | |
bc5c7e08 | 7898 | rtx base_plus = gen_rtx_REG (SImode, REGNO (operands[2]) + 1); |
997d68fe | 7899 | |
0083a3b9 | 7900 | emit_set_insn (base_plus, base); |
997d68fe | 7901 | base = base_plus; |
7902 | } | |
bc5c7e08 | 7903 | else if (GET_CODE (base) == PLUS) |
7904 | { | |
66aacf41 | 7905 | /* The addend must be CONST_INT, or we would have dealt with it above. */ |
bc5c7e08 | 7906 | HOST_WIDE_INT hi, lo; |
7907 | ||
7908 | offset += INTVAL (XEXP (base, 1)); | |
7909 | base = XEXP (base, 0); | |
7910 | ||
66aacf41 | 7911 | /* Rework the address into a legal sequence of insns. */ |
bc5c7e08 | 7912 | /* Valid range for lo is -4095 -> 4095 */ |
7913 | lo = (offset >= 0 | |
7914 | ? (offset & 0xfff) | |
7915 | : -((-offset) & 0xfff)); | |
7916 | ||
7917 | /* Corner case, if lo is the max offset then we would be out of range | |
7918 | once we have added the additional 1 below, so bump the msb into the | |
7919 | pre-loading insn(s). */ | |
7920 | if (lo == 4095) | |
7921 | lo &= 0x7ff; | |
7922 | ||
35cea10c | 7923 | hi = ((((offset - lo) & (HOST_WIDE_INT) 0xffffffff) |
7924 | ^ (HOST_WIDE_INT) 0x80000000) | |
7925 | - (HOST_WIDE_INT) 0x80000000); | |
bc5c7e08 | 7926 | |
ed29c566 | 7927 | gcc_assert (hi + lo == offset); |
bc5c7e08 | 7928 | |
7929 | if (hi != 0) | |
7930 | { | |
7931 | rtx base_plus = gen_rtx_REG (SImode, REGNO (operands[2]) + 1); | |
7932 | ||
7933 | /* Get the base address; addsi3 knows how to handle constants | |
66aacf41 | 7934 | that require more than one insn. */ |
bc5c7e08 | 7935 | emit_insn (gen_addsi3 (base_plus, base, GEN_INT (hi))); |
7936 | base = base_plus; | |
7937 | offset = lo; | |
7938 | } | |
7939 | } | |
997d68fe | 7940 | |
dfb47981 | 7941 | /* Operands[2] may overlap operands[0] (though it won't overlap |
7942 | operands[1]), that's why we asked for a DImode reg -- so we can | |
7943 | use the bit that does not overlap. */ | |
7944 | if (REGNO (operands[2]) == REGNO (operands[0])) | |
7945 | scratch = gen_rtx_REG (SImode, REGNO (operands[2]) + 1); | |
7946 | else | |
7947 | scratch = gen_rtx_REG (SImode, REGNO (operands[2])); | |
7948 | ||
bc5c7e08 | 7949 | emit_insn (gen_zero_extendqisi2 (scratch, |
7950 | gen_rtx_MEM (QImode, | |
7951 | plus_constant (base, | |
7952 | offset)))); | |
76676c8e | 7953 | emit_insn (gen_zero_extendqisi2 (gen_rtx_SUBREG (SImode, operands[0], 0), |
9e7454d0 | 7954 | gen_rtx_MEM (QImode, |
bc5c7e08 | 7955 | plus_constant (base, |
7956 | offset + 1)))); | |
215b30b3 | 7957 | if (!BYTES_BIG_ENDIAN) |
0083a3b9 | 7958 | emit_set_insn (gen_rtx_SUBREG (SImode, operands[0], 0), |
7959 | gen_rtx_IOR (SImode, | |
7960 | gen_rtx_ASHIFT | |
7961 | (SImode, | |
7962 | gen_rtx_SUBREG (SImode, operands[0], 0), | |
7963 | GEN_INT (8)), | |
7964 | scratch)); | |
91fd5a8d | 7965 | else |
0083a3b9 | 7966 | emit_set_insn (gen_rtx_SUBREG (SImode, operands[0], 0), |
7967 | gen_rtx_IOR (SImode, | |
7968 | gen_rtx_ASHIFT (SImode, scratch, | |
7969 | GEN_INT (8)), | |
7970 | gen_rtx_SUBREG (SImode, operands[0], 0))); | |
91fd5a8d | 7971 | } |
7972 | ||
5bd751ff | 7973 | /* Handle storing a half-word to memory during reload by synthesizing as two |
bc5c7e08 | 7974 | byte stores. Take care not to clobber the input values until after we |
7975 | have moved them somewhere safe. This code assumes that if the DImode | |
7976 | scratch in operands[2] overlaps either the input value or output address | |
7977 | in some way, then that value must die in this insn (we absolutely need | |
7978 | two scratch registers for some corner cases). */ | |
a690ede2 | 7979 | void |
ebd88a36 | 7980 | arm_reload_out_hi (rtx *operands) |
9c1b36a9 | 7981 | { |
bc5c7e08 | 7982 | rtx ref = operands[0]; |
7983 | rtx outval = operands[1]; | |
7984 | rtx base, scratch; | |
7985 | HOST_WIDE_INT offset = 0; | |
7986 | ||
7987 | if (GET_CODE (ref) == SUBREG) | |
7988 | { | |
701e46d0 | 7989 | offset = SUBREG_BYTE (ref); |
bc5c7e08 | 7990 | ref = SUBREG_REG (ref); |
7991 | } | |
7992 | ||
bc5c7e08 | 7993 | if (GET_CODE (ref) == REG) |
7994 | { | |
7995 | /* We have a pseudo which has been spilt onto the stack; there | |
7996 | are two cases here: the first where there is a simple | |
7997 | stack-slot replacement and a second where the stack-slot is | |
7998 | out of range, or is used as a subreg. */ | |
7999 | if (reg_equiv_mem[REGNO (ref)]) | |
8000 | { | |
8001 | ref = reg_equiv_mem[REGNO (ref)]; | |
8002 | base = find_replacement (&XEXP (ref, 0)); | |
8003 | } | |
8004 | else | |
66aacf41 | 8005 | /* The slot is out of range, or was dressed up in a SUBREG. */ |
bc5c7e08 | 8006 | base = reg_equiv_address[REGNO (ref)]; |
8007 | } | |
8008 | else | |
8009 | base = find_replacement (&XEXP (ref, 0)); | |
8010 | ||
8011 | scratch = gen_rtx_REG (SImode, REGNO (operands[2])); | |
8012 | ||
8013 | /* Handle the case where the address is too complex to be offset by 1. */ | |
8014 | if (GET_CODE (base) == MINUS | |
8015 | || (GET_CODE (base) == PLUS && GET_CODE (XEXP (base, 1)) != CONST_INT)) | |
8016 | { | |
8017 | rtx base_plus = gen_rtx_REG (SImode, REGNO (operands[2]) + 1); | |
8018 | ||
8019 | /* Be careful not to destroy OUTVAL. */ | |
8020 | if (reg_overlap_mentioned_p (base_plus, outval)) | |
8021 | { | |
8022 | /* Updating base_plus might destroy outval, see if we can | |
8023 | swap the scratch and base_plus. */ | |
215b30b3 | 8024 | if (!reg_overlap_mentioned_p (scratch, outval)) |
bc5c7e08 | 8025 | { |
8026 | rtx tmp = scratch; | |
8027 | scratch = base_plus; | |
8028 | base_plus = tmp; | |
8029 | } | |
8030 | else | |
8031 | { | |
8032 | rtx scratch_hi = gen_rtx_REG (HImode, REGNO (operands[2])); | |
8033 | ||
8034 | /* Be conservative and copy OUTVAL into the scratch now, | |
8035 | this should only be necessary if outval is a subreg | |
8036 | of something larger than a word. */ | |
8037 | /* XXX Might this clobber base? I can't see how it can, | |
8038 | since scratch is known to overlap with OUTVAL, and | |
8039 | must be wider than a word. */ | |
8040 | emit_insn (gen_movhi (scratch_hi, outval)); | |
8041 | outval = scratch_hi; | |
8042 | } | |
8043 | } | |
8044 | ||
0083a3b9 | 8045 | emit_set_insn (base_plus, base); |
bc5c7e08 | 8046 | base = base_plus; |
8047 | } | |
8048 | else if (GET_CODE (base) == PLUS) | |
8049 | { | |
66aacf41 | 8050 | /* The addend must be CONST_INT, or we would have dealt with it above. */ |
bc5c7e08 | 8051 | HOST_WIDE_INT hi, lo; |
8052 | ||
8053 | offset += INTVAL (XEXP (base, 1)); | |
8054 | base = XEXP (base, 0); | |
8055 | ||
66aacf41 | 8056 | /* Rework the address into a legal sequence of insns. */ |
bc5c7e08 | 8057 | /* Valid range for lo is -4095 -> 4095 */ |
8058 | lo = (offset >= 0 | |
8059 | ? (offset & 0xfff) | |
8060 | : -((-offset) & 0xfff)); | |
8061 | ||
8062 | /* Corner case, if lo is the max offset then we would be out of range | |
8063 | once we have added the additional 1 below, so bump the msb into the | |
8064 | pre-loading insn(s). */ | |
8065 | if (lo == 4095) | |
8066 | lo &= 0x7ff; | |
8067 | ||
35cea10c | 8068 | hi = ((((offset - lo) & (HOST_WIDE_INT) 0xffffffff) |
8069 | ^ (HOST_WIDE_INT) 0x80000000) | |
8070 | - (HOST_WIDE_INT) 0x80000000); | |
bc5c7e08 | 8071 | |
ed29c566 | 8072 | gcc_assert (hi + lo == offset); |
bc5c7e08 | 8073 | |
8074 | if (hi != 0) | |
8075 | { | |
8076 | rtx base_plus = gen_rtx_REG (SImode, REGNO (operands[2]) + 1); | |
8077 | ||
8078 | /* Be careful not to destroy OUTVAL. */ | |
8079 | if (reg_overlap_mentioned_p (base_plus, outval)) | |
8080 | { | |
8081 | /* Updating base_plus might destroy outval, see if we | |
8082 | can swap the scratch and base_plus. */ | |
215b30b3 | 8083 | if (!reg_overlap_mentioned_p (scratch, outval)) |
bc5c7e08 | 8084 | { |
8085 | rtx tmp = scratch; | |
8086 | scratch = base_plus; | |
8087 | base_plus = tmp; | |
8088 | } | |
8089 | else | |
8090 | { | |
8091 | rtx scratch_hi = gen_rtx_REG (HImode, REGNO (operands[2])); | |
8092 | ||
8093 | /* Be conservative and copy outval into scratch now, | |
8094 | this should only be necessary if outval is a | |
8095 | subreg of something larger than a word. */ | |
8096 | /* XXX Might this clobber base? I can't see how it | |
8097 | can, since scratch is known to overlap with | |
8098 | outval. */ | |
8099 | emit_insn (gen_movhi (scratch_hi, outval)); | |
8100 | outval = scratch_hi; | |
8101 | } | |
8102 | } | |
8103 | ||
8104 | /* Get the base address; addsi3 knows how to handle constants | |
66aacf41 | 8105 | that require more than one insn. */ |
bc5c7e08 | 8106 | emit_insn (gen_addsi3 (base_plus, base, GEN_INT (hi))); |
8107 | base = base_plus; | |
8108 | offset = lo; | |
8109 | } | |
8110 | } | |
9c1b36a9 | 8111 | |
51d172bd | 8112 | if (BYTES_BIG_ENDIAN) |
8113 | { | |
9e7454d0 | 8114 | emit_insn (gen_movqi (gen_rtx_MEM (QImode, |
bc5c7e08 | 8115 | plus_constant (base, offset + 1)), |
9e8503e6 | 8116 | gen_lowpart (QImode, outval))); |
bc5c7e08 | 8117 | emit_insn (gen_lshrsi3 (scratch, |
8118 | gen_rtx_SUBREG (SImode, outval, 0), | |
51d172bd | 8119 | GEN_INT (8))); |
bc5c7e08 | 8120 | emit_insn (gen_movqi (gen_rtx_MEM (QImode, plus_constant (base, offset)), |
9e8503e6 | 8121 | gen_lowpart (QImode, scratch))); |
51d172bd | 8122 | } |
8123 | else | |
8124 | { | |
bc5c7e08 | 8125 | emit_insn (gen_movqi (gen_rtx_MEM (QImode, plus_constant (base, offset)), |
9e8503e6 | 8126 | gen_lowpart (QImode, outval))); |
bc5c7e08 | 8127 | emit_insn (gen_lshrsi3 (scratch, |
8128 | gen_rtx_SUBREG (SImode, outval, 0), | |
51d172bd | 8129 | GEN_INT (8))); |
bc5c7e08 | 8130 | emit_insn (gen_movqi (gen_rtx_MEM (QImode, |
8131 | plus_constant (base, offset + 1)), | |
9e8503e6 | 8132 | gen_lowpart (QImode, scratch))); |
51d172bd | 8133 | } |
9c1b36a9 | 8134 | } |
46b5d878 | 8135 | |
8136 | /* Return true if a type must be passed in memory. For AAPCS, small aggregates | |
8137 | (padded to the size of a word) should be passed in a register. */ | |
8138 | ||
8139 | static bool | |
8140 | arm_must_pass_in_stack (enum machine_mode mode, tree type) | |
8141 | { | |
8142 | if (TARGET_AAPCS_BASED) | |
8143 | return must_pass_in_stack_var_size (mode, type); | |
8144 | else | |
8145 | return must_pass_in_stack_var_size_or_pad (mode, type); | |
8146 | } | |
8147 | ||
8148 | ||
8149 | /* For use by FUNCTION_ARG_PADDING (MODE, TYPE). | |
8150 | Return true if an argument passed on the stack should be padded upwards, | |
a5558ca7 | 8151 | i.e. if the least-significant byte has useful data. |
8152 | For legacy APCS ABIs we use the default. For AAPCS based ABIs small | |
8153 | aggregate types are placed in the lowest memory address. */ | |
46b5d878 | 8154 | |
8155 | bool | |
8156 | arm_pad_arg_upward (enum machine_mode mode, tree type) | |
8157 | { | |
8158 | if (!TARGET_AAPCS_BASED) | |
a5558ca7 | 8159 | return DEFAULT_FUNCTION_ARG_PADDING(mode, type) == upward; |
46b5d878 | 8160 | |
8161 | if (type && BYTES_BIG_ENDIAN && INTEGRAL_TYPE_P (type)) | |
8162 | return false; | |
8163 | ||
8164 | return true; | |
8165 | } | |
8166 | ||
8167 | ||
8168 | /* Similarly, for use by BLOCK_REG_PADDING (MODE, TYPE, FIRST). | |
8169 | For non-AAPCS, return !BYTES_BIG_ENDIAN if the least significant | |
8170 | byte of the register has useful data, and return the opposite if the | |
8171 | most significant byte does. | |
8172 | For AAPCS, small aggregates and small complex types are always padded | |
8173 | upwards. */ | |
8174 | ||
8175 | bool | |
8176 | arm_pad_reg_upward (enum machine_mode mode ATTRIBUTE_UNUSED, | |
8177 | tree type, int first ATTRIBUTE_UNUSED) | |
8178 | { | |
8179 | if (TARGET_AAPCS_BASED | |
8180 | && BYTES_BIG_ENDIAN | |
8181 | && (AGGREGATE_TYPE_P (type) || TREE_CODE (type) == COMPLEX_TYPE) | |
8182 | && int_size_in_bytes (type) <= 4) | |
8183 | return true; | |
8184 | ||
8185 | /* Otherwise, use default padding. */ | |
8186 | return !BYTES_BIG_ENDIAN; | |
8187 | } | |
8188 | ||
64266206 | 8189 | \f |
cffb2a26 | 8190 | /* Print a symbolic form of X to the debug file, F. */ |
8191 | static void | |
ebd88a36 | 8192 | arm_print_value (FILE *f, rtx x) |
cffb2a26 | 8193 | { |
8194 | switch (GET_CODE (x)) | |
8195 | { | |
8196 | case CONST_INT: | |
8197 | fprintf (f, HOST_WIDE_INT_PRINT_HEX, INTVAL (x)); | |
8198 | return; | |
8199 | ||
8200 | case CONST_DOUBLE: | |
8201 | fprintf (f, "<0x%lx,0x%lx>", (long)XWINT (x, 2), (long)XWINT (x, 3)); | |
8202 | return; | |
8203 | ||
755eb2b4 | 8204 | case CONST_VECTOR: |
8205 | { | |
8206 | int i; | |
8207 | ||
8208 | fprintf (f, "<"); | |
8209 | for (i = 0; i < CONST_VECTOR_NUNITS (x); i++) | |
8210 | { | |
8211 | fprintf (f, HOST_WIDE_INT_PRINT_HEX, INTVAL (CONST_VECTOR_ELT (x, i))); | |
8212 | if (i < (CONST_VECTOR_NUNITS (x) - 1)) | |
8213 | fputc (',', f); | |
8214 | } | |
8215 | fprintf (f, ">"); | |
8216 | } | |
8217 | return; | |
8218 | ||
cffb2a26 | 8219 | case CONST_STRING: |
8220 | fprintf (f, "\"%s\"", XSTR (x, 0)); | |
8221 | return; | |
8222 | ||
8223 | case SYMBOL_REF: | |
8224 | fprintf (f, "`%s'", XSTR (x, 0)); | |
8225 | return; | |
8226 | ||
8227 | case LABEL_REF: | |
8228 | fprintf (f, "L%d", INSN_UID (XEXP (x, 0))); | |
8229 | return; | |
8230 | ||
8231 | case CONST: | |
8232 | arm_print_value (f, XEXP (x, 0)); | |
8233 | return; | |
8234 | ||
8235 | case PLUS: | |
8236 | arm_print_value (f, XEXP (x, 0)); | |
8237 | fprintf (f, "+"); | |
8238 | arm_print_value (f, XEXP (x, 1)); | |
8239 | return; | |
8240 | ||
8241 | case PC: | |
8242 | fprintf (f, "pc"); | |
8243 | return; | |
8244 | ||
8245 | default: | |
8246 | fprintf (f, "????"); | |
8247 | return; | |
8248 | } | |
8249 | } | |
8250 | \f | |
64266206 | 8251 | /* Routines for manipulation of the constant pool. */ |
64266206 | 8252 | |
56d27660 | 8253 | /* Arm instructions cannot load a large constant directly into a |
8254 | register; they have to come from a pc relative load. The constant | |
8255 | must therefore be placed in the addressable range of the pc | |
8256 | relative load. Depending on the precise pc relative load | |
8257 | instruction the range is somewhere between 256 bytes and 4k. This | |
8258 | means that we often have to dump a constant inside a function, and | |
64266206 | 8259 | generate code to branch around it. |
8260 | ||
56d27660 | 8261 | It is important to minimize this, since the branches will slow |
8262 | things down and make the code larger. | |
64266206 | 8263 | |
56d27660 | 8264 | Normally we can hide the table after an existing unconditional |
8265 | branch so that there is no interruption of the flow, but in the | |
8266 | worst case the code looks like this: | |
64266206 | 8267 | |
8268 | ldr rn, L1 | |
56d27660 | 8269 | ... |
64266206 | 8270 | b L2 |
8271 | align | |
8272 | L1: .long value | |
8273 | L2: | |
56d27660 | 8274 | ... |
64266206 | 8275 | |
64266206 | 8276 | ldr rn, L3 |
56d27660 | 8277 | ... |
64266206 | 8278 | b L4 |
8279 | align | |
64266206 | 8280 | L3: .long value |
8281 | L4: | |
56d27660 | 8282 | ... |
8283 | ||
8284 | We fix this by performing a scan after scheduling, which notices | |
8285 | which instructions need to have their operands fetched from the | |
8286 | constant table and builds the table. | |
8287 | ||
8288 | The algorithm starts by building a table of all the constants that | |
8289 | need fixing up and all the natural barriers in the function (places | |
8290 | where a constant table can be dropped without breaking the flow). | |
8291 | For each fixup we note how far the pc-relative replacement will be | |
8292 | able to reach and the offset of the instruction into the function. | |
8293 | ||
8294 | Having built the table we then group the fixes together to form | |
8295 | tables that are as large as possible (subject to addressing | |
8296 | constraints) and emit each table of constants after the last | |
8297 | barrier that is within range of all the instructions in the group. | |
8298 | If a group does not contain a barrier, then we forcibly create one | |
8299 | by inserting a jump instruction into the flow. Once the table has | |
8300 | been inserted, the insns are then modified to reference the | |
8301 | relevant entry in the pool. | |
8302 | ||
66aacf41 | 8303 | Possible enhancements to the algorithm (not implemented) are: |
56d27660 | 8304 | |
cffb2a26 | 8305 | 1) For some processors and object formats, there may be benefit in |
56d27660 | 8306 | aligning the pools to the start of cache lines; this alignment |
8307 | would need to be taken into account when calculating addressability | |
66aacf41 | 8308 | of a pool. */ |
64266206 | 8309 | |
cffb2a26 | 8310 | /* These typedefs are located at the start of this file, so that |
8311 | they can be used in the prototypes there. This comment is to | |
8312 | remind readers of that fact so that the following structures | |
8313 | can be understood more easily. | |
8314 | ||
8315 | typedef struct minipool_node Mnode; | |
8316 | typedef struct minipool_fixup Mfix; */ | |
8317 | ||
8318 | struct minipool_node | |
8319 | { | |
8320 | /* Doubly linked chain of entries. */ | |
8321 | Mnode * next; | |
8322 | Mnode * prev; | |
8323 | /* The maximum offset into the code that this entry can be placed. While | |
8324 | pushing fixes for forward references, all entries are sorted in order | |
8325 | of increasing max_address. */ | |
8326 | HOST_WIDE_INT max_address; | |
45981c0a | 8327 | /* Similarly for an entry inserted for a backwards ref. */ |
cffb2a26 | 8328 | HOST_WIDE_INT min_address; |
8329 | /* The number of fixes referencing this entry. This can become zero | |
8330 | if we "unpush" an entry. In this case we ignore the entry when we | |
8331 | come to emit the code. */ | |
8332 | int refcount; | |
8333 | /* The offset from the start of the minipool. */ | |
8334 | HOST_WIDE_INT offset; | |
8335 | /* The value in table. */ | |
8336 | rtx value; | |
8337 | /* The mode of value. */ | |
8338 | enum machine_mode mode; | |
755eb2b4 | 8339 | /* The size of the value. With iWMMXt enabled |
8340 | sizes > 4 also imply an alignment of 8-bytes. */ | |
cffb2a26 | 8341 | int fix_size; |
8342 | }; | |
8343 | ||
8344 | struct minipool_fixup | |
64266206 | 8345 | { |
cffb2a26 | 8346 | Mfix * next; |
8347 | rtx insn; | |
8348 | HOST_WIDE_INT address; | |
8349 | rtx * loc; | |
8350 | enum machine_mode mode; | |
8351 | int fix_size; | |
8352 | rtx value; | |
8353 | Mnode * minipool; | |
8354 | HOST_WIDE_INT forwards; | |
8355 | HOST_WIDE_INT backwards; | |
8356 | }; | |
64266206 | 8357 | |
cffb2a26 | 8358 | /* Fixes less than a word need padding out to a word boundary. */ |
8359 | #define MINIPOOL_FIX_SIZE(mode) \ | |
8360 | (GET_MODE_SIZE ((mode)) >= 4 ? GET_MODE_SIZE ((mode)) : 4) | |
64266206 | 8361 | |
cffb2a26 | 8362 | static Mnode * minipool_vector_head; |
8363 | static Mnode * minipool_vector_tail; | |
8364 | static rtx minipool_vector_label; | |
acc677e9 | 8365 | static int minipool_pad; |
61592d0d | 8366 | |
cffb2a26 | 8367 | /* The linked list of all minipool fixes required for this function. */ |
8368 | Mfix * minipool_fix_head; | |
8369 | Mfix * minipool_fix_tail; | |
8370 | /* The fix entry for the current minipool, once it has been placed. */ | |
8371 | Mfix * minipool_barrier; | |
8372 | ||
8373 | /* Determines if INSN is the start of a jump table. Returns the end | |
8374 | of the TABLE or NULL_RTX. */ | |
8375 | static rtx | |
ebd88a36 | 8376 | is_jump_table (rtx insn) |
64266206 | 8377 | { |
cffb2a26 | 8378 | rtx table; |
9e7454d0 | 8379 | |
cffb2a26 | 8380 | if (GET_CODE (insn) == JUMP_INSN |
8381 | && JUMP_LABEL (insn) != NULL | |
8382 | && ((table = next_real_insn (JUMP_LABEL (insn))) | |
8383 | == next_real_insn (insn)) | |
8384 | && table != NULL | |
8385 | && GET_CODE (table) == JUMP_INSN | |
8386 | && (GET_CODE (PATTERN (table)) == ADDR_VEC | |
8387 | || GET_CODE (PATTERN (table)) == ADDR_DIFF_VEC)) | |
8388 | return table; | |
8389 | ||
8390 | return NULL_RTX; | |
64266206 | 8391 | } |
8392 | ||
fd957ef3 | 8393 | #ifndef JUMP_TABLES_IN_TEXT_SECTION |
8394 | #define JUMP_TABLES_IN_TEXT_SECTION 0 | |
8395 | #endif | |
8396 | ||
cffb2a26 | 8397 | static HOST_WIDE_INT |
ebd88a36 | 8398 | get_jump_table_size (rtx insn) |
64266206 | 8399 | { |
fd957ef3 | 8400 | /* ADDR_VECs only take room if read-only data does into the text |
8401 | section. */ | |
2f14b1f9 | 8402 | if (JUMP_TABLES_IN_TEXT_SECTION || readonly_data_section == text_section) |
fd957ef3 | 8403 | { |
8404 | rtx body = PATTERN (insn); | |
8405 | int elt = GET_CODE (body) == ADDR_DIFF_VEC ? 1 : 0; | |
25f905c2 | 8406 | HOST_WIDE_INT size; |
8407 | HOST_WIDE_INT modesize; | |
64266206 | 8408 | |
25f905c2 | 8409 | modesize = GET_MODE_SIZE (GET_MODE (body)); |
8410 | size = modesize * XVECLEN (body, elt); | |
8411 | switch (modesize) | |
8412 | { | |
8413 | case 1: | |
0d424440 | 8414 | /* Round up size of TBB table to a halfword boundary. */ |
25f905c2 | 8415 | size = (size + 1) & ~(HOST_WIDE_INT)1; |
8416 | break; | |
8417 | case 2: | |
542d5028 | 8418 | /* No padding necessary for TBH. */ |
25f905c2 | 8419 | break; |
8420 | case 4: | |
8421 | /* Add two bytes for alignment on Thumb. */ | |
8422 | if (TARGET_THUMB) | |
8423 | size += 2; | |
8424 | break; | |
8425 | default: | |
8426 | gcc_unreachable (); | |
8427 | } | |
8428 | return size; | |
fd957ef3 | 8429 | } |
8430 | ||
8431 | return 0; | |
cffb2a26 | 8432 | } |
64266206 | 8433 | |
cffb2a26 | 8434 | /* Move a minipool fix MP from its current location to before MAX_MP. |
8435 | If MAX_MP is NULL, then MP doesn't need moving, but the addressing | |
5910bb95 | 8436 | constraints may need updating. */ |
cffb2a26 | 8437 | static Mnode * |
ebd88a36 | 8438 | move_minipool_fix_forward_ref (Mnode *mp, Mnode *max_mp, |
8439 | HOST_WIDE_INT max_address) | |
cffb2a26 | 8440 | { |
ed29c566 | 8441 | /* The code below assumes these are different. */ |
8442 | gcc_assert (mp != max_mp); | |
cffb2a26 | 8443 | |
8444 | if (max_mp == NULL) | |
8445 | { | |
8446 | if (max_address < mp->max_address) | |
8447 | mp->max_address = max_address; | |
8448 | } | |
8449 | else | |
64266206 | 8450 | { |
cffb2a26 | 8451 | if (max_address > max_mp->max_address - mp->fix_size) |
8452 | mp->max_address = max_mp->max_address - mp->fix_size; | |
8453 | else | |
8454 | mp->max_address = max_address; | |
64266206 | 8455 | |
cffb2a26 | 8456 | /* Unlink MP from its current position. Since max_mp is non-null, |
8457 | mp->prev must be non-null. */ | |
8458 | mp->prev->next = mp->next; | |
8459 | if (mp->next != NULL) | |
8460 | mp->next->prev = mp->prev; | |
8461 | else | |
8462 | minipool_vector_tail = mp->prev; | |
64266206 | 8463 | |
cffb2a26 | 8464 | /* Re-insert it before MAX_MP. */ |
8465 | mp->next = max_mp; | |
8466 | mp->prev = max_mp->prev; | |
8467 | max_mp->prev = mp; | |
9e7454d0 | 8468 | |
cffb2a26 | 8469 | if (mp->prev != NULL) |
8470 | mp->prev->next = mp; | |
8471 | else | |
8472 | minipool_vector_head = mp; | |
8473 | } | |
64266206 | 8474 | |
cffb2a26 | 8475 | /* Save the new entry. */ |
8476 | max_mp = mp; | |
8477 | ||
cb0ccc1e | 8478 | /* Scan over the preceding entries and adjust their addresses as |
cffb2a26 | 8479 | required. */ |
8480 | while (mp->prev != NULL | |
8481 | && mp->prev->max_address > mp->max_address - mp->prev->fix_size) | |
8482 | { | |
8483 | mp->prev->max_address = mp->max_address - mp->prev->fix_size; | |
8484 | mp = mp->prev; | |
64266206 | 8485 | } |
8486 | ||
cffb2a26 | 8487 | return max_mp; |
64266206 | 8488 | } |
8489 | ||
cffb2a26 | 8490 | /* Add a constant to the minipool for a forward reference. Returns the |
8491 | node added or NULL if the constant will not fit in this pool. */ | |
8492 | static Mnode * | |
ebd88a36 | 8493 | add_minipool_forward_ref (Mfix *fix) |
cffb2a26 | 8494 | { |
8495 | /* If set, max_mp is the first pool_entry that has a lower | |
8496 | constraint than the one we are trying to add. */ | |
8497 | Mnode * max_mp = NULL; | |
acc677e9 | 8498 | HOST_WIDE_INT max_address = fix->address + fix->forwards - minipool_pad; |
cffb2a26 | 8499 | Mnode * mp; |
9e7454d0 | 8500 | |
8f2b909a | 8501 | /* If the minipool starts before the end of FIX->INSN then this FIX |
8502 | can not be placed into the current pool. Furthermore, adding the | |
8503 | new constant pool entry may cause the pool to start FIX_SIZE bytes | |
8504 | earlier. */ | |
cffb2a26 | 8505 | if (minipool_vector_head && |
8f2b909a | 8506 | (fix->address + get_attr_length (fix->insn) |
8507 | >= minipool_vector_head->max_address - fix->fix_size)) | |
cffb2a26 | 8508 | return NULL; |
64266206 | 8509 | |
cffb2a26 | 8510 | /* Scan the pool to see if a constant with the same value has |
8511 | already been added. While we are doing this, also note the | |
8512 | location where we must insert the constant if it doesn't already | |
8513 | exist. */ | |
8514 | for (mp = minipool_vector_head; mp != NULL; mp = mp->next) | |
8515 | { | |
8516 | if (GET_CODE (fix->value) == GET_CODE (mp->value) | |
8517 | && fix->mode == mp->mode | |
8518 | && (GET_CODE (fix->value) != CODE_LABEL | |
8519 | || (CODE_LABEL_NUMBER (fix->value) | |
8520 | == CODE_LABEL_NUMBER (mp->value))) | |
8521 | && rtx_equal_p (fix->value, mp->value)) | |
8522 | { | |
8523 | /* More than one fix references this entry. */ | |
8524 | mp->refcount++; | |
8525 | return move_minipool_fix_forward_ref (mp, max_mp, max_address); | |
8526 | } | |
8527 | ||
8528 | /* Note the insertion point if necessary. */ | |
8529 | if (max_mp == NULL | |
8530 | && mp->max_address > max_address) | |
8531 | max_mp = mp; | |
755eb2b4 | 8532 | |
8533 | /* If we are inserting an 8-bytes aligned quantity and | |
8534 | we have not already found an insertion point, then | |
8535 | make sure that all such 8-byte aligned quantities are | |
8536 | placed at the start of the pool. */ | |
f9273c43 | 8537 | if (ARM_DOUBLEWORD_ALIGN |
755eb2b4 | 8538 | && max_mp == NULL |
d98a3884 | 8539 | && fix->fix_size >= 8 |
8540 | && mp->fix_size < 8) | |
755eb2b4 | 8541 | { |
8542 | max_mp = mp; | |
8543 | max_address = mp->max_address; | |
8544 | } | |
cffb2a26 | 8545 | } |
8546 | ||
8547 | /* The value is not currently in the minipool, so we need to create | |
8548 | a new entry for it. If MAX_MP is NULL, the entry will be put on | |
8549 | the end of the list since the placement is less constrained than | |
8550 | any existing entry. Otherwise, we insert the new fix before | |
b89db9a7 | 8551 | MAX_MP and, if necessary, adjust the constraints on the other |
cffb2a26 | 8552 | entries. */ |
4c36ffe6 | 8553 | mp = XNEW (Mnode); |
cffb2a26 | 8554 | mp->fix_size = fix->fix_size; |
8555 | mp->mode = fix->mode; | |
8556 | mp->value = fix->value; | |
8557 | mp->refcount = 1; | |
8558 | /* Not yet required for a backwards ref. */ | |
8559 | mp->min_address = -65536; | |
8560 | ||
8561 | if (max_mp == NULL) | |
8562 | { | |
8563 | mp->max_address = max_address; | |
8564 | mp->next = NULL; | |
8565 | mp->prev = minipool_vector_tail; | |
8566 | ||
8567 | if (mp->prev == NULL) | |
8568 | { | |
8569 | minipool_vector_head = mp; | |
8570 | minipool_vector_label = gen_label_rtx (); | |
0da499d4 | 8571 | } |
64266206 | 8572 | else |
cffb2a26 | 8573 | mp->prev->next = mp; |
64266206 | 8574 | |
cffb2a26 | 8575 | minipool_vector_tail = mp; |
8576 | } | |
8577 | else | |
8578 | { | |
8579 | if (max_address > max_mp->max_address - mp->fix_size) | |
8580 | mp->max_address = max_mp->max_address - mp->fix_size; | |
8581 | else | |
8582 | mp->max_address = max_address; | |
8583 | ||
8584 | mp->next = max_mp; | |
8585 | mp->prev = max_mp->prev; | |
8586 | max_mp->prev = mp; | |
8587 | if (mp->prev != NULL) | |
8588 | mp->prev->next = mp; | |
8589 | else | |
8590 | minipool_vector_head = mp; | |
8591 | } | |
8592 | ||
8593 | /* Save the new entry. */ | |
8594 | max_mp = mp; | |
8595 | ||
cb0ccc1e | 8596 | /* Scan over the preceding entries and adjust their addresses as |
cffb2a26 | 8597 | required. */ |
8598 | while (mp->prev != NULL | |
8599 | && mp->prev->max_address > mp->max_address - mp->prev->fix_size) | |
8600 | { | |
8601 | mp->prev->max_address = mp->max_address - mp->prev->fix_size; | |
8602 | mp = mp->prev; | |
64266206 | 8603 | } |
8604 | ||
cffb2a26 | 8605 | return max_mp; |
8606 | } | |
8607 | ||
8608 | static Mnode * | |
ebd88a36 | 8609 | move_minipool_fix_backward_ref (Mnode *mp, Mnode *min_mp, |
8610 | HOST_WIDE_INT min_address) | |
cffb2a26 | 8611 | { |
8612 | HOST_WIDE_INT offset; | |
8613 | ||
ed29c566 | 8614 | /* The code below assumes these are different. */ |
8615 | gcc_assert (mp != min_mp); | |
cffb2a26 | 8616 | |
8617 | if (min_mp == NULL) | |
64266206 | 8618 | { |
cffb2a26 | 8619 | if (min_address > mp->min_address) |
8620 | mp->min_address = min_address; | |
8621 | } | |
8622 | else | |
8623 | { | |
8624 | /* We will adjust this below if it is too loose. */ | |
8625 | mp->min_address = min_address; | |
8626 | ||
8627 | /* Unlink MP from its current position. Since min_mp is non-null, | |
8628 | mp->next must be non-null. */ | |
8629 | mp->next->prev = mp->prev; | |
8630 | if (mp->prev != NULL) | |
8631 | mp->prev->next = mp->next; | |
8632 | else | |
8633 | minipool_vector_head = mp->next; | |
8634 | ||
8635 | /* Reinsert it after MIN_MP. */ | |
8636 | mp->prev = min_mp; | |
8637 | mp->next = min_mp->next; | |
8638 | min_mp->next = mp; | |
8639 | if (mp->next != NULL) | |
8640 | mp->next->prev = mp; | |
64266206 | 8641 | else |
cffb2a26 | 8642 | minipool_vector_tail = mp; |
8643 | } | |
8644 | ||
8645 | min_mp = mp; | |
8646 | ||
8647 | offset = 0; | |
8648 | for (mp = minipool_vector_head; mp != NULL; mp = mp->next) | |
8649 | { | |
8650 | mp->offset = offset; | |
8651 | if (mp->refcount > 0) | |
8652 | offset += mp->fix_size; | |
8653 | ||
8654 | if (mp->next && mp->next->min_address < mp->min_address + mp->fix_size) | |
8655 | mp->next->min_address = mp->min_address + mp->fix_size; | |
8656 | } | |
8657 | ||
8658 | return min_mp; | |
9e7454d0 | 8659 | } |
cffb2a26 | 8660 | |
8661 | /* Add a constant to the minipool for a backward reference. Returns the | |
9e7454d0 | 8662 | node added or NULL if the constant will not fit in this pool. |
cffb2a26 | 8663 | |
8664 | Note that the code for insertion for a backwards reference can be | |
8665 | somewhat confusing because the calculated offsets for each fix do | |
8666 | not take into account the size of the pool (which is still under | |
8667 | construction. */ | |
8668 | static Mnode * | |
ebd88a36 | 8669 | add_minipool_backward_ref (Mfix *fix) |
cffb2a26 | 8670 | { |
8671 | /* If set, min_mp is the last pool_entry that has a lower constraint | |
8672 | than the one we are trying to add. */ | |
ebd88a36 | 8673 | Mnode *min_mp = NULL; |
cffb2a26 | 8674 | /* This can be negative, since it is only a constraint. */ |
8675 | HOST_WIDE_INT min_address = fix->address - fix->backwards; | |
ebd88a36 | 8676 | Mnode *mp; |
cffb2a26 | 8677 | |
8678 | /* If we can't reach the current pool from this insn, or if we can't | |
8679 | insert this entry at the end of the pool without pushing other | |
8680 | fixes out of range, then we don't try. This ensures that we | |
8681 | can't fail later on. */ | |
8682 | if (min_address >= minipool_barrier->address | |
8683 | || (minipool_vector_tail->min_address + fix->fix_size | |
8684 | >= minipool_barrier->address)) | |
8685 | return NULL; | |
8686 | ||
8687 | /* Scan the pool to see if a constant with the same value has | |
8688 | already been added. While we are doing this, also note the | |
8689 | location where we must insert the constant if it doesn't already | |
8690 | exist. */ | |
8691 | for (mp = minipool_vector_tail; mp != NULL; mp = mp->prev) | |
8692 | { | |
8693 | if (GET_CODE (fix->value) == GET_CODE (mp->value) | |
8694 | && fix->mode == mp->mode | |
8695 | && (GET_CODE (fix->value) != CODE_LABEL | |
8696 | || (CODE_LABEL_NUMBER (fix->value) | |
8697 | == CODE_LABEL_NUMBER (mp->value))) | |
8698 | && rtx_equal_p (fix->value, mp->value) | |
8699 | /* Check that there is enough slack to move this entry to the | |
8700 | end of the table (this is conservative). */ | |
9e7454d0 | 8701 | && (mp->max_address |
8702 | > (minipool_barrier->address | |
cffb2a26 | 8703 | + minipool_vector_tail->offset |
8704 | + minipool_vector_tail->fix_size))) | |
8705 | { | |
8706 | mp->refcount++; | |
8707 | return move_minipool_fix_backward_ref (mp, min_mp, min_address); | |
8708 | } | |
8709 | ||
8710 | if (min_mp != NULL) | |
8711 | mp->min_address += fix->fix_size; | |
8712 | else | |
8713 | { | |
8714 | /* Note the insertion point if necessary. */ | |
8715 | if (mp->min_address < min_address) | |
755eb2b4 | 8716 | { |
8717 | /* For now, we do not allow the insertion of 8-byte alignment | |
8718 | requiring nodes anywhere but at the start of the pool. */ | |
f9273c43 | 8719 | if (ARM_DOUBLEWORD_ALIGN |
d98a3884 | 8720 | && fix->fix_size >= 8 && mp->fix_size < 8) |
755eb2b4 | 8721 | return NULL; |
8722 | else | |
8723 | min_mp = mp; | |
8724 | } | |
cffb2a26 | 8725 | else if (mp->max_address |
8726 | < minipool_barrier->address + mp->offset + fix->fix_size) | |
8727 | { | |
8728 | /* Inserting before this entry would push the fix beyond | |
8729 | its maximum address (which can happen if we have | |
8730 | re-located a forwards fix); force the new fix to come | |
8731 | after it. */ | |
8732 | min_mp = mp; | |
8733 | min_address = mp->min_address + fix->fix_size; | |
8734 | } | |
755eb2b4 | 8735 | /* If we are inserting an 8-bytes aligned quantity and |
8736 | we have not already found an insertion point, then | |
8737 | make sure that all such 8-byte aligned quantities are | |
8738 | placed at the start of the pool. */ | |
f9273c43 | 8739 | else if (ARM_DOUBLEWORD_ALIGN |
755eb2b4 | 8740 | && min_mp == NULL |
d98a3884 | 8741 | && fix->fix_size >= 8 |
755eb2b4 | 8742 | && mp->fix_size < 8) |
8743 | { | |
8744 | min_mp = mp; | |
8745 | min_address = mp->min_address + fix->fix_size; | |
8746 | } | |
cffb2a26 | 8747 | } |
8748 | } | |
8749 | ||
8750 | /* We need to create a new entry. */ | |
4c36ffe6 | 8751 | mp = XNEW (Mnode); |
cffb2a26 | 8752 | mp->fix_size = fix->fix_size; |
8753 | mp->mode = fix->mode; | |
8754 | mp->value = fix->value; | |
8755 | mp->refcount = 1; | |
8756 | mp->max_address = minipool_barrier->address + 65536; | |
8757 | ||
8758 | mp->min_address = min_address; | |
8759 | ||
8760 | if (min_mp == NULL) | |
8761 | { | |
8762 | mp->prev = NULL; | |
8763 | mp->next = minipool_vector_head; | |
8764 | ||
8765 | if (mp->next == NULL) | |
8766 | { | |
8767 | minipool_vector_tail = mp; | |
8768 | minipool_vector_label = gen_label_rtx (); | |
8769 | } | |
8770 | else | |
8771 | mp->next->prev = mp; | |
8772 | ||
8773 | minipool_vector_head = mp; | |
8774 | } | |
8775 | else | |
8776 | { | |
8777 | mp->next = min_mp->next; | |
8778 | mp->prev = min_mp; | |
8779 | min_mp->next = mp; | |
9e7454d0 | 8780 | |
cffb2a26 | 8781 | if (mp->next != NULL) |
8782 | mp->next->prev = mp; | |
8783 | else | |
8784 | minipool_vector_tail = mp; | |
8785 | } | |
8786 | ||
8787 | /* Save the new entry. */ | |
8788 | min_mp = mp; | |
8789 | ||
8790 | if (mp->prev) | |
8791 | mp = mp->prev; | |
8792 | else | |
8793 | mp->offset = 0; | |
8794 | ||
8795 | /* Scan over the following entries and adjust their offsets. */ | |
8796 | while (mp->next != NULL) | |
8797 | { | |
8798 | if (mp->next->min_address < mp->min_address + mp->fix_size) | |
8799 | mp->next->min_address = mp->min_address + mp->fix_size; | |
8800 | ||
8801 | if (mp->refcount) | |
8802 | mp->next->offset = mp->offset + mp->fix_size; | |
8803 | else | |
8804 | mp->next->offset = mp->offset; | |
8805 | ||
8806 | mp = mp->next; | |
8807 | } | |
8808 | ||
8809 | return min_mp; | |
8810 | } | |
8811 | ||
8812 | static void | |
ebd88a36 | 8813 | assign_minipool_offsets (Mfix *barrier) |
cffb2a26 | 8814 | { |
8815 | HOST_WIDE_INT offset = 0; | |
ebd88a36 | 8816 | Mnode *mp; |
cffb2a26 | 8817 | |
8818 | minipool_barrier = barrier; | |
8819 | ||
8820 | for (mp = minipool_vector_head; mp != NULL; mp = mp->next) | |
8821 | { | |
8822 | mp->offset = offset; | |
9e7454d0 | 8823 | |
cffb2a26 | 8824 | if (mp->refcount > 0) |
8825 | offset += mp->fix_size; | |
8826 | } | |
8827 | } | |
8828 | ||
8829 | /* Output the literal table */ | |
8830 | static void | |
ebd88a36 | 8831 | dump_minipool (rtx scan) |
cffb2a26 | 8832 | { |
755eb2b4 | 8833 | Mnode * mp; |
8834 | Mnode * nmp; | |
8835 | int align64 = 0; | |
8836 | ||
f9273c43 | 8837 | if (ARM_DOUBLEWORD_ALIGN) |
755eb2b4 | 8838 | for (mp = minipool_vector_head; mp != NULL; mp = mp->next) |
d98a3884 | 8839 | if (mp->refcount > 0 && mp->fix_size >= 8) |
755eb2b4 | 8840 | { |
8841 | align64 = 1; | |
8842 | break; | |
8843 | } | |
cffb2a26 | 8844 | |
450d042a | 8845 | if (dump_file) |
8846 | fprintf (dump_file, | |
755eb2b4 | 8847 | ";; Emitting minipool after insn %u; address %ld; align %d (bytes)\n", |
8848 | INSN_UID (scan), (unsigned long) minipool_barrier->address, align64 ? 8 : 4); | |
cffb2a26 | 8849 | |
8850 | scan = emit_label_after (gen_label_rtx (), scan); | |
755eb2b4 | 8851 | scan = emit_insn_after (align64 ? gen_align_8 () : gen_align_4 (), scan); |
cffb2a26 | 8852 | scan = emit_label_after (minipool_vector_label, scan); |
8853 | ||
8854 | for (mp = minipool_vector_head; mp != NULL; mp = nmp) | |
8855 | { | |
8856 | if (mp->refcount > 0) | |
8857 | { | |
450d042a | 8858 | if (dump_file) |
cffb2a26 | 8859 | { |
9e7454d0 | 8860 | fprintf (dump_file, |
cffb2a26 | 8861 | ";; Offset %u, min %ld, max %ld ", |
8862 | (unsigned) mp->offset, (unsigned long) mp->min_address, | |
8863 | (unsigned long) mp->max_address); | |
450d042a | 8864 | arm_print_value (dump_file, mp->value); |
8865 | fputc ('\n', dump_file); | |
cffb2a26 | 8866 | } |
8867 | ||
8868 | switch (mp->fix_size) | |
8869 | { | |
8870 | #ifdef HAVE_consttable_1 | |
8871 | case 1: | |
8872 | scan = emit_insn_after (gen_consttable_1 (mp->value), scan); | |
8873 | break; | |
8874 | ||
8875 | #endif | |
8876 | #ifdef HAVE_consttable_2 | |
8877 | case 2: | |
8878 | scan = emit_insn_after (gen_consttable_2 (mp->value), scan); | |
8879 | break; | |
8880 | ||
8881 | #endif | |
8882 | #ifdef HAVE_consttable_4 | |
8883 | case 4: | |
8884 | scan = emit_insn_after (gen_consttable_4 (mp->value), scan); | |
8885 | break; | |
8886 | ||
8887 | #endif | |
8888 | #ifdef HAVE_consttable_8 | |
8889 | case 8: | |
8890 | scan = emit_insn_after (gen_consttable_8 (mp->value), scan); | |
8891 | break; | |
8892 | ||
d98a3884 | 8893 | #endif |
8894 | #ifdef HAVE_consttable_16 | |
8895 | case 16: | |
8896 | scan = emit_insn_after (gen_consttable_16 (mp->value), scan); | |
8897 | break; | |
8898 | ||
cffb2a26 | 8899 | #endif |
8900 | default: | |
ed29c566 | 8901 | gcc_unreachable (); |
cffb2a26 | 8902 | } |
8903 | } | |
8904 | ||
8905 | nmp = mp->next; | |
8906 | free (mp); | |
64266206 | 8907 | } |
8908 | ||
cffb2a26 | 8909 | minipool_vector_head = minipool_vector_tail = NULL; |
8910 | scan = emit_insn_after (gen_consttable_end (), scan); | |
8911 | scan = emit_barrier_after (scan); | |
64266206 | 8912 | } |
8913 | ||
cffb2a26 | 8914 | /* Return the cost of forcibly inserting a barrier after INSN. */ |
8915 | static int | |
ebd88a36 | 8916 | arm_barrier_cost (rtx insn) |
56d27660 | 8917 | { |
cffb2a26 | 8918 | /* Basing the location of the pool on the loop depth is preferable, |
8919 | but at the moment, the basic block information seems to be | |
8920 | corrupt by this stage of the compilation. */ | |
8921 | int base_cost = 50; | |
8922 | rtx next = next_nonnote_insn (insn); | |
8923 | ||
8924 | if (next != NULL && GET_CODE (next) == CODE_LABEL) | |
8925 | base_cost -= 20; | |
8926 | ||
8927 | switch (GET_CODE (insn)) | |
8928 | { | |
8929 | case CODE_LABEL: | |
8930 | /* It will always be better to place the table before the label, rather | |
8931 | than after it. */ | |
9e7454d0 | 8932 | return 50; |
56d27660 | 8933 | |
cffb2a26 | 8934 | case INSN: |
8935 | case CALL_INSN: | |
8936 | return base_cost; | |
8937 | ||
8938 | case JUMP_INSN: | |
8939 | return base_cost - 10; | |
8940 | ||
8941 | default: | |
8942 | return base_cost + 10; | |
8943 | } | |
8944 | } | |
8945 | ||
8946 | /* Find the best place in the insn stream in the range | |
8947 | (FIX->address,MAX_ADDRESS) to forcibly insert a minipool barrier. | |
8948 | Create the barrier by inserting a jump and add a new fix entry for | |
8949 | it. */ | |
8950 | static Mfix * | |
ebd88a36 | 8951 | create_fix_barrier (Mfix *fix, HOST_WIDE_INT max_address) |
cffb2a26 | 8952 | { |
8953 | HOST_WIDE_INT count = 0; | |
8954 | rtx barrier; | |
8955 | rtx from = fix->insn; | |
8f2b909a | 8956 | /* The instruction after which we will insert the jump. */ |
8957 | rtx selected = NULL; | |
cffb2a26 | 8958 | int selected_cost; |
8f2b909a | 8959 | /* The address at which the jump instruction will be placed. */ |
cffb2a26 | 8960 | HOST_WIDE_INT selected_address; |
8961 | Mfix * new_fix; | |
8962 | HOST_WIDE_INT max_count = max_address - fix->address; | |
8963 | rtx label = gen_label_rtx (); | |
8964 | ||
8965 | selected_cost = arm_barrier_cost (from); | |
8966 | selected_address = fix->address; | |
8967 | ||
8968 | while (from && count < max_count) | |
8969 | { | |
8970 | rtx tmp; | |
8971 | int new_cost; | |
8972 | ||
8973 | /* This code shouldn't have been called if there was a natural barrier | |
8974 | within range. */ | |
ed29c566 | 8975 | gcc_assert (GET_CODE (from) != BARRIER); |
cffb2a26 | 8976 | |
8977 | /* Count the length of this insn. */ | |
8978 | count += get_attr_length (from); | |
8979 | ||
8980 | /* If there is a jump table, add its length. */ | |
8981 | tmp = is_jump_table (from); | |
8982 | if (tmp != NULL) | |
8983 | { | |
8984 | count += get_jump_table_size (tmp); | |
8985 | ||
8986 | /* Jump tables aren't in a basic block, so base the cost on | |
8987 | the dispatch insn. If we select this location, we will | |
8988 | still put the pool after the table. */ | |
8989 | new_cost = arm_barrier_cost (from); | |
8990 | ||
8f2b909a | 8991 | if (count < max_count |
8992 | && (!selected || new_cost <= selected_cost)) | |
cffb2a26 | 8993 | { |
8994 | selected = tmp; | |
8995 | selected_cost = new_cost; | |
8996 | selected_address = fix->address + count; | |
8997 | } | |
8998 | ||
8999 | /* Continue after the dispatch table. */ | |
9000 | from = NEXT_INSN (tmp); | |
9001 | continue; | |
9002 | } | |
9003 | ||
9004 | new_cost = arm_barrier_cost (from); | |
9e7454d0 | 9005 | |
8f2b909a | 9006 | if (count < max_count |
9007 | && (!selected || new_cost <= selected_cost)) | |
cffb2a26 | 9008 | { |
9009 | selected = from; | |
9010 | selected_cost = new_cost; | |
9011 | selected_address = fix->address + count; | |
9012 | } | |
9013 | ||
9014 | from = NEXT_INSN (from); | |
9015 | } | |
9016 | ||
8f2b909a | 9017 | /* Make sure that we found a place to insert the jump. */ |
9018 | gcc_assert (selected); | |
9019 | ||
cffb2a26 | 9020 | /* Create a new JUMP_INSN that branches around a barrier. */ |
9021 | from = emit_jump_insn_after (gen_jump (label), selected); | |
9022 | JUMP_LABEL (from) = label; | |
9023 | barrier = emit_barrier_after (from); | |
9024 | emit_label_after (label, barrier); | |
9025 | ||
9026 | /* Create a minipool barrier entry for the new barrier. */ | |
a220ccce | 9027 | new_fix = (Mfix *) obstack_alloc (&minipool_obstack, sizeof (* new_fix)); |
cffb2a26 | 9028 | new_fix->insn = barrier; |
9029 | new_fix->address = selected_address; | |
9030 | new_fix->next = fix->next; | |
9031 | fix->next = new_fix; | |
9032 | ||
9033 | return new_fix; | |
9034 | } | |
9035 | ||
9036 | /* Record that there is a natural barrier in the insn stream at | |
9037 | ADDRESS. */ | |
56d27660 | 9038 | static void |
ebd88a36 | 9039 | push_minipool_barrier (rtx insn, HOST_WIDE_INT address) |
64266206 | 9040 | { |
a220ccce | 9041 | Mfix * fix = (Mfix *) obstack_alloc (&minipool_obstack, sizeof (* fix)); |
ecd3fcc7 | 9042 | |
56d27660 | 9043 | fix->insn = insn; |
9044 | fix->address = address; | |
64266206 | 9045 | |
56d27660 | 9046 | fix->next = NULL; |
9047 | if (minipool_fix_head != NULL) | |
9048 | minipool_fix_tail->next = fix; | |
9049 | else | |
9050 | minipool_fix_head = fix; | |
9051 | ||
9052 | minipool_fix_tail = fix; | |
9053 | } | |
64266206 | 9054 | |
cffb2a26 | 9055 | /* Record INSN, which will need fixing up to load a value from the |
9056 | minipool. ADDRESS is the offset of the insn since the start of the | |
9057 | function; LOC is a pointer to the part of the insn which requires | |
9058 | fixing; VALUE is the constant that must be loaded, which is of type | |
9059 | MODE. */ | |
56d27660 | 9060 | static void |
ebd88a36 | 9061 | push_minipool_fix (rtx insn, HOST_WIDE_INT address, rtx *loc, |
9062 | enum machine_mode mode, rtx value) | |
56d27660 | 9063 | { |
a220ccce | 9064 | Mfix * fix = (Mfix *) obstack_alloc (&minipool_obstack, sizeof (* fix)); |
56d27660 | 9065 | |
9066 | #ifdef AOF_ASSEMBLER | |
5910bb95 | 9067 | /* PIC symbol references need to be converted into offsets into the |
56d27660 | 9068 | based area. */ |
cffb2a26 | 9069 | /* XXX This shouldn't be done here. */ |
9070 | if (flag_pic && GET_CODE (value) == SYMBOL_REF) | |
56d27660 | 9071 | value = aof_pic_entry (value); |
9072 | #endif /* AOF_ASSEMBLER */ | |
9073 | ||
9074 | fix->insn = insn; | |
9075 | fix->address = address; | |
9076 | fix->loc = loc; | |
9077 | fix->mode = mode; | |
cffb2a26 | 9078 | fix->fix_size = MINIPOOL_FIX_SIZE (mode); |
56d27660 | 9079 | fix->value = value; |
cffb2a26 | 9080 | fix->forwards = get_attr_pool_range (insn); |
9081 | fix->backwards = get_attr_neg_pool_range (insn); | |
9082 | fix->minipool = NULL; | |
56d27660 | 9083 | |
9084 | /* If an insn doesn't have a range defined for it, then it isn't | |
ed29c566 | 9085 | expecting to be reworked by this code. Better to stop now than |
56d27660 | 9086 | to generate duff assembly code. */ |
ed29c566 | 9087 | gcc_assert (fix->forwards || fix->backwards); |
56d27660 | 9088 | |
acc677e9 | 9089 | /* If an entry requires 8-byte alignment then assume all constant pools |
9090 | require 4 bytes of padding. Trying to do this later on a per-pool | |
9ca2c29a | 9091 | basis is awkward because existing pool entries have to be modified. */ |
d98a3884 | 9092 | if (ARM_DOUBLEWORD_ALIGN && fix->fix_size >= 8) |
acc677e9 | 9093 | minipool_pad = 4; |
755eb2b4 | 9094 | |
450d042a | 9095 | if (dump_file) |
cffb2a26 | 9096 | { |
450d042a | 9097 | fprintf (dump_file, |
cffb2a26 | 9098 | ";; %smode fixup for i%d; addr %lu, range (%ld,%ld): ", |
9099 | GET_MODE_NAME (mode), | |
9e7454d0 | 9100 | INSN_UID (insn), (unsigned long) address, |
cffb2a26 | 9101 | -1 * (long)fix->backwards, (long)fix->forwards); |
450d042a | 9102 | arm_print_value (dump_file, fix->value); |
9103 | fprintf (dump_file, "\n"); | |
cffb2a26 | 9104 | } |
9105 | ||
66aacf41 | 9106 | /* Add it to the chain of fixes. */ |
56d27660 | 9107 | fix->next = NULL; |
9e7454d0 | 9108 | |
56d27660 | 9109 | if (minipool_fix_head != NULL) |
9110 | minipool_fix_tail->next = fix; | |
9111 | else | |
9112 | minipool_fix_head = fix; | |
9113 | ||
9114 | minipool_fix_tail = fix; | |
9115 | } | |
9116 | ||
d51f92df | 9117 | /* Return the cost of synthesizing a 64-bit constant VAL inline. |
359a6e9f | 9118 | Returns the number of insns needed, or 99 if we don't know how to |
9119 | do it. */ | |
9120 | int | |
9121 | arm_const_double_inline_cost (rtx val) | |
3194bc82 | 9122 | { |
d51f92df | 9123 | rtx lowpart, highpart; |
9124 | enum machine_mode mode; | |
a60f1840 | 9125 | |
d51f92df | 9126 | mode = GET_MODE (val); |
3194bc82 | 9127 | |
d51f92df | 9128 | if (mode == VOIDmode) |
9129 | mode = DImode; | |
9130 | ||
9131 | gcc_assert (GET_MODE_SIZE (mode) == 8); | |
a60f1840 | 9132 | |
d51f92df | 9133 | lowpart = gen_lowpart (SImode, val); |
9134 | highpart = gen_highpart_mode (SImode, mode, val); | |
a60f1840 | 9135 | |
d51f92df | 9136 | gcc_assert (GET_CODE (lowpart) == CONST_INT); |
9137 | gcc_assert (GET_CODE (highpart) == CONST_INT); | |
9138 | ||
9139 | return (arm_gen_constant (SET, SImode, NULL_RTX, INTVAL (lowpart), | |
359a6e9f | 9140 | NULL_RTX, NULL_RTX, 0, 0) |
d51f92df | 9141 | + arm_gen_constant (SET, SImode, NULL_RTX, INTVAL (highpart), |
359a6e9f | 9142 | NULL_RTX, NULL_RTX, 0, 0)); |
9143 | } | |
9144 | ||
20833d12 | 9145 | /* Return true if it is worthwhile to split a 64-bit constant into two |
e5ba9289 | 9146 | 32-bit operations. This is the case if optimizing for size, or |
9147 | if we have load delay slots, or if one 32-bit part can be done with | |
9148 | a single data operation. */ | |
9149 | bool | |
9150 | arm_const_double_by_parts (rtx val) | |
9151 | { | |
9152 | enum machine_mode mode = GET_MODE (val); | |
9153 | rtx part; | |
9154 | ||
9155 | if (optimize_size || arm_ld_sched) | |
9156 | return true; | |
9157 | ||
9158 | if (mode == VOIDmode) | |
9159 | mode = DImode; | |
a60f1840 | 9160 | |
e5ba9289 | 9161 | part = gen_highpart_mode (SImode, mode, val); |
a60f1840 | 9162 | |
e5ba9289 | 9163 | gcc_assert (GET_CODE (part) == CONST_INT); |
a60f1840 | 9164 | |
e5ba9289 | 9165 | if (const_ok_for_arm (INTVAL (part)) |
9166 | || const_ok_for_arm (~INTVAL (part))) | |
9167 | return true; | |
a60f1840 | 9168 | |
e5ba9289 | 9169 | part = gen_lowpart (SImode, val); |
a60f1840 | 9170 | |
e5ba9289 | 9171 | gcc_assert (GET_CODE (part) == CONST_INT); |
a60f1840 | 9172 | |
e5ba9289 | 9173 | if (const_ok_for_arm (INTVAL (part)) |
9174 | || const_ok_for_arm (~INTVAL (part))) | |
9175 | return true; | |
a60f1840 | 9176 | |
e5ba9289 | 9177 | return false; |
9178 | } | |
9179 | ||
2c6c7d8b | 9180 | /* Scan INSN and note any of its operands that need fixing. |
9181 | If DO_PUSHES is false we do not actually push any of the fixups | |
d51f92df | 9182 | needed. The function returns TRUE if any fixups were needed/pushed. |
2c6c7d8b | 9183 | This is used by arm_memory_load_p() which needs to know about loads |
9184 | of constants that will be converted into minipool loads. */ | |
2c6c7d8b | 9185 | static bool |
ebd88a36 | 9186 | note_invalid_constants (rtx insn, HOST_WIDE_INT address, int do_pushes) |
56d27660 | 9187 | { |
2c6c7d8b | 9188 | bool result = false; |
56d27660 | 9189 | int opno; |
9190 | ||
cffb2a26 | 9191 | extract_insn (insn); |
56d27660 | 9192 | |
215b30b3 | 9193 | if (!constrain_operands (1)) |
56d27660 | 9194 | fatal_insn_not_found (insn); |
9195 | ||
97088907 | 9196 | if (recog_data.n_alternatives == 0) |
9197 | return false; | |
9198 | ||
d51f92df | 9199 | /* Fill in recog_op_alt with information about the constraints of |
9200 | this insn. */ | |
56d27660 | 9201 | preprocess_constraints (); |
9202 | ||
ed420a25 | 9203 | for (opno = 0; opno < recog_data.n_operands; opno++) |
56d27660 | 9204 | { |
66aacf41 | 9205 | /* Things we need to fix can only occur in inputs. */ |
31146060 | 9206 | if (recog_data.operand_type[opno] != OP_IN) |
56d27660 | 9207 | continue; |
9208 | ||
9209 | /* If this alternative is a memory reference, then any mention | |
9210 | of constants in this alternative is really to fool reload | |
9211 | into allowing us to accept one there. We need to fix them up | |
9212 | now so that we output the right code. */ | |
9213 | if (recog_op_alt[opno][which_alternative].memory_ok) | |
9214 | { | |
ed420a25 | 9215 | rtx op = recog_data.operand[opno]; |
56d27660 | 9216 | |
d51f92df | 9217 | if (CONSTANT_P (op)) |
2c6c7d8b | 9218 | { |
9219 | if (do_pushes) | |
9220 | push_minipool_fix (insn, address, recog_data.operand_loc[opno], | |
9221 | recog_data.operand_mode[opno], op); | |
9222 | result = true; | |
9223 | } | |
cffb2a26 | 9224 | else if (GET_CODE (op) == MEM |
56d27660 | 9225 | && GET_CODE (XEXP (op, 0)) == SYMBOL_REF |
9226 | && CONSTANT_POOL_ADDRESS_P (XEXP (op, 0))) | |
2c6c7d8b | 9227 | { |
9228 | if (do_pushes) | |
28fe824a | 9229 | { |
9230 | rtx cop = avoid_constant_pool_reference (op); | |
9231 | ||
9232 | /* Casting the address of something to a mode narrower | |
9233 | than a word can cause avoid_constant_pool_reference() | |
9234 | to return the pool reference itself. That's no good to | |
9e7454d0 | 9235 | us here. Lets just hope that we can use the |
28fe824a | 9236 | constant pool value directly. */ |
9237 | if (op == cop) | |
aeac46d4 | 9238 | cop = get_pool_constant (XEXP (op, 0)); |
28fe824a | 9239 | |
9240 | push_minipool_fix (insn, address, | |
9241 | recog_data.operand_loc[opno], | |
aeac46d4 | 9242 | recog_data.operand_mode[opno], cop); |
28fe824a | 9243 | } |
2c6c7d8b | 9244 | |
9245 | result = true; | |
9246 | } | |
56d27660 | 9247 | } |
64266206 | 9248 | } |
2c6c7d8b | 9249 | |
9250 | return result; | |
64266206 | 9251 | } |
9252 | ||
2efea8c0 | 9253 | /* Gcc puts the pool in the wrong place for ARM, since we can only |
9254 | load addresses a limited distance around the pc. We do some | |
9255 | special munging to move the constant pool values to the correct | |
9256 | point in the code. */ | |
2efea8c0 | 9257 | static void |
ebd88a36 | 9258 | arm_reorg (void) |
64266206 | 9259 | { |
9260 | rtx insn; | |
cffb2a26 | 9261 | HOST_WIDE_INT address = 0; |
9262 | Mfix * fix; | |
ecd3fcc7 | 9263 | |
56d27660 | 9264 | minipool_fix_head = minipool_fix_tail = NULL; |
64266206 | 9265 | |
56d27660 | 9266 | /* The first insn must always be a note, or the code below won't |
9267 | scan it properly. */ | |
2efea8c0 | 9268 | insn = get_insns (); |
ed29c566 | 9269 | gcc_assert (GET_CODE (insn) == NOTE); |
acc677e9 | 9270 | minipool_pad = 0; |
56d27660 | 9271 | |
9272 | /* Scan all the insns and record the operands that will need fixing. */ | |
2efea8c0 | 9273 | for (insn = next_nonnote_insn (insn); insn; insn = next_nonnote_insn (insn)) |
64266206 | 9274 | { |
7d57ec45 | 9275 | if (TARGET_CIRRUS_FIX_INVALID_INSNS |
2c6c7d8b | 9276 | && (arm_cirrus_insn_p (insn) |
7d57ec45 | 9277 | || GET_CODE (insn) == JUMP_INSN |
2c6c7d8b | 9278 | || arm_memory_load_p (insn))) |
7d57ec45 | 9279 | cirrus_reorg (insn); |
9280 | ||
56d27660 | 9281 | if (GET_CODE (insn) == BARRIER) |
cffb2a26 | 9282 | push_minipool_barrier (insn, address); |
2c6c7d8b | 9283 | else if (INSN_P (insn)) |
56d27660 | 9284 | { |
9285 | rtx table; | |
9286 | ||
2c6c7d8b | 9287 | note_invalid_constants (insn, address, true); |
56d27660 | 9288 | address += get_attr_length (insn); |
cffb2a26 | 9289 | |
56d27660 | 9290 | /* If the insn is a vector jump, add the size of the table |
9291 | and skip the table. */ | |
cffb2a26 | 9292 | if ((table = is_jump_table (insn)) != NULL) |
64266206 | 9293 | { |
cffb2a26 | 9294 | address += get_jump_table_size (table); |
56d27660 | 9295 | insn = table; |
9296 | } | |
9297 | } | |
9298 | } | |
61592d0d | 9299 | |
cffb2a26 | 9300 | fix = minipool_fix_head; |
9e7454d0 | 9301 | |
56d27660 | 9302 | /* Now scan the fixups and perform the required changes. */ |
cffb2a26 | 9303 | while (fix) |
56d27660 | 9304 | { |
cffb2a26 | 9305 | Mfix * ftmp; |
9306 | Mfix * fdel; | |
9307 | Mfix * last_added_fix; | |
9308 | Mfix * last_barrier = NULL; | |
9309 | Mfix * this_fix; | |
56d27660 | 9310 | |
9311 | /* Skip any further barriers before the next fix. */ | |
9312 | while (fix && GET_CODE (fix->insn) == BARRIER) | |
9313 | fix = fix->next; | |
9314 | ||
cffb2a26 | 9315 | /* No more fixes. */ |
56d27660 | 9316 | if (fix == NULL) |
9317 | break; | |
61592d0d | 9318 | |
cffb2a26 | 9319 | last_added_fix = NULL; |
64266206 | 9320 | |
cffb2a26 | 9321 | for (ftmp = fix; ftmp; ftmp = ftmp->next) |
56d27660 | 9322 | { |
56d27660 | 9323 | if (GET_CODE (ftmp->insn) == BARRIER) |
56d27660 | 9324 | { |
cffb2a26 | 9325 | if (ftmp->address >= minipool_vector_head->max_address) |
9326 | break; | |
64266206 | 9327 | |
cffb2a26 | 9328 | last_barrier = ftmp; |
64266206 | 9329 | } |
cffb2a26 | 9330 | else if ((ftmp->minipool = add_minipool_forward_ref (ftmp)) == NULL) |
9331 | break; | |
9332 | ||
9333 | last_added_fix = ftmp; /* Keep track of the last fix added. */ | |
64266206 | 9334 | } |
56d27660 | 9335 | |
cffb2a26 | 9336 | /* If we found a barrier, drop back to that; any fixes that we |
9337 | could have reached but come after the barrier will now go in | |
9338 | the next mini-pool. */ | |
56d27660 | 9339 | if (last_barrier != NULL) |
9340 | { | |
9e7454d0 | 9341 | /* Reduce the refcount for those fixes that won't go into this |
cffb2a26 | 9342 | pool after all. */ |
9343 | for (fdel = last_barrier->next; | |
9344 | fdel && fdel != ftmp; | |
9345 | fdel = fdel->next) | |
9346 | { | |
9347 | fdel->minipool->refcount--; | |
9348 | fdel->minipool = NULL; | |
9349 | } | |
9350 | ||
56d27660 | 9351 | ftmp = last_barrier; |
9352 | } | |
9353 | else | |
6524d6b5 | 9354 | { |
cffb2a26 | 9355 | /* ftmp is first fix that we can't fit into this pool and |
9356 | there no natural barriers that we could use. Insert a | |
9357 | new barrier in the code somewhere between the previous | |
9358 | fix and this one, and arrange to jump around it. */ | |
9359 | HOST_WIDE_INT max_address; | |
9360 | ||
9361 | /* The last item on the list of fixes must be a barrier, so | |
9362 | we can never run off the end of the list of fixes without | |
9363 | last_barrier being set. */ | |
ed29c566 | 9364 | gcc_assert (ftmp); |
cffb2a26 | 9365 | |
9366 | max_address = minipool_vector_head->max_address; | |
6524d6b5 | 9367 | /* Check that there isn't another fix that is in range that |
9368 | we couldn't fit into this pool because the pool was | |
9369 | already too large: we need to put the pool before such an | |
8f2b909a | 9370 | instruction. The pool itself may come just after the |
9371 | fix because create_fix_barrier also allows space for a | |
9372 | jump instruction. */ | |
cffb2a26 | 9373 | if (ftmp->address < max_address) |
8f2b909a | 9374 | max_address = ftmp->address + 1; |
cffb2a26 | 9375 | |
9376 | last_barrier = create_fix_barrier (last_added_fix, max_address); | |
9377 | } | |
9378 | ||
9379 | assign_minipool_offsets (last_barrier); | |
9380 | ||
9381 | while (ftmp) | |
9382 | { | |
9383 | if (GET_CODE (ftmp->insn) != BARRIER | |
9384 | && ((ftmp->minipool = add_minipool_backward_ref (ftmp)) | |
9385 | == NULL)) | |
9386 | break; | |
6524d6b5 | 9387 | |
cffb2a26 | 9388 | ftmp = ftmp->next; |
6524d6b5 | 9389 | } |
56d27660 | 9390 | |
9391 | /* Scan over the fixes we have identified for this pool, fixing them | |
9392 | up and adding the constants to the pool itself. */ | |
cffb2a26 | 9393 | for (this_fix = fix; this_fix && ftmp != this_fix; |
56d27660 | 9394 | this_fix = this_fix->next) |
9395 | if (GET_CODE (this_fix->insn) != BARRIER) | |
9396 | { | |
56d27660 | 9397 | rtx addr |
9e7454d0 | 9398 | = plus_constant (gen_rtx_LABEL_REF (VOIDmode, |
56d27660 | 9399 | minipool_vector_label), |
cffb2a26 | 9400 | this_fix->minipool->offset); |
56d27660 | 9401 | *this_fix->loc = gen_rtx_MEM (this_fix->mode, addr); |
9402 | } | |
9403 | ||
cffb2a26 | 9404 | dump_minipool (last_barrier->insn); |
56d27660 | 9405 | fix = ftmp; |
64266206 | 9406 | } |
11de4c7f | 9407 | |
56d27660 | 9408 | /* From now on we must synthesize any constants that we can't handle |
9409 | directly. This can happen if the RTL gets split during final | |
9410 | instruction generation. */ | |
11de4c7f | 9411 | after_arm_reorg = 1; |
a220ccce | 9412 | |
9413 | /* Free the minipool memory. */ | |
9414 | obstack_free (&minipool_obstack, minipool_startobj); | |
64266206 | 9415 | } |
8cd42699 | 9416 | \f |
9417 | /* Routines to output assembly language. */ | |
9418 | ||
a690ede2 | 9419 | /* If the rtx is the correct value then return the string of the number. |
9c08d1fa | 9420 | In this way we can ensure that valid double constants are generated even |
66aacf41 | 9421 | when cross compiling. */ |
0bd59681 | 9422 | const char * |
ebd88a36 | 9423 | fp_immediate_constant (rtx x) |
9c08d1fa | 9424 | { |
9425 | REAL_VALUE_TYPE r; | |
9426 | int i; | |
9e7454d0 | 9427 | |
a2cd141b | 9428 | if (!fp_consts_inited) |
9429 | init_fp_table (); | |
9e7454d0 | 9430 | |
9c08d1fa | 9431 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); |
9432 | for (i = 0; i < 8; i++) | |
a2cd141b | 9433 | if (REAL_VALUES_EQUAL (r, values_fp[i])) |
9434 | return strings_fp[i]; | |
a690ede2 | 9435 | |
ed29c566 | 9436 | gcc_unreachable (); |
9c08d1fa | 9437 | } |
9438 | ||
f6b01912 | 9439 | /* As for fp_immediate_constant, but value is passed directly, not in rtx. */ |
0bd59681 | 9440 | static const char * |
ebd88a36 | 9441 | fp_const_from_val (REAL_VALUE_TYPE *r) |
f6b01912 | 9442 | { |
9443 | int i; | |
9444 | ||
a2cd141b | 9445 | if (!fp_consts_inited) |
9446 | init_fp_table (); | |
f6b01912 | 9447 | |
9448 | for (i = 0; i < 8; i++) | |
a2cd141b | 9449 | if (REAL_VALUES_EQUAL (*r, values_fp[i])) |
9450 | return strings_fp[i]; | |
f6b01912 | 9451 | |
ed29c566 | 9452 | gcc_unreachable (); |
f6b01912 | 9453 | } |
9c08d1fa | 9454 | |
8cd42699 | 9455 | /* Output the operands of a LDM/STM instruction to STREAM. |
9456 | MASK is the ARM register set mask of which only bits 0-15 are important. | |
e27ad2d5 | 9457 | REG is the base register, either the frame pointer or the stack pointer, |
8b438b4f | 9458 | INSTR is the possibly suffixed load or store instruction. |
9459 | RFE is nonzero if the instruction should also copy spsr to cpsr. */ | |
53f179c0 | 9460 | |
cffb2a26 | 9461 | static void |
53f179c0 | 9462 | print_multi_reg (FILE *stream, const char *instr, unsigned reg, |
8b438b4f | 9463 | unsigned long mask, int rfe) |
8cd42699 | 9464 | { |
53f179c0 | 9465 | unsigned i; |
9466 | bool not_first = FALSE; | |
8cd42699 | 9467 | |
8b438b4f | 9468 | gcc_assert (!rfe || (mask & (1 << PC_REGNUM))); |
39d5d2b6 | 9469 | fputc ('\t', stream); |
caceeb43 | 9470 | asm_fprintf (stream, instr, reg); |
25f905c2 | 9471 | fputc ('{', stream); |
9e7454d0 | 9472 | |
cffb2a26 | 9473 | for (i = 0; i <= LAST_ARM_REGNUM; i++) |
8cd42699 | 9474 | if (mask & (1 << i)) |
9475 | { | |
9476 | if (not_first) | |
9477 | fprintf (stream, ", "); | |
9e7454d0 | 9478 | |
caceeb43 | 9479 | asm_fprintf (stream, "%r", i); |
8cd42699 | 9480 | not_first = TRUE; |
9481 | } | |
a690ede2 | 9482 | |
8b438b4f | 9483 | if (rfe) |
9484 | fprintf (stream, "}^\n"); | |
9485 | else | |
9486 | fprintf (stream, "}\n"); | |
a690ede2 | 9487 | } |
8cd42699 | 9488 | |
a2cd141b | 9489 | |
382a2a57 | 9490 | /* Output a FLDMD instruction to STREAM. |
7b78e739 | 9491 | BASE if the register containing the address. |
9492 | REG and COUNT specify the register range. | |
382a2a57 | 9493 | Extra registers may be added to avoid hardware bugs. |
9494 | ||
9495 | We output FLDMD even for ARMv5 VFP implementations. Although | |
9496 | FLDMD is technically not supported until ARMv6, it is believed | |
9497 | that all VFP implementations support its use in this context. */ | |
a2cd141b | 9498 | |
9499 | static void | |
382a2a57 | 9500 | vfp_output_fldmd (FILE * stream, unsigned int base, int reg, int count) |
a2cd141b | 9501 | { |
9502 | int i; | |
9503 | ||
7b78e739 | 9504 | /* Workaround ARM10 VFPr1 bug. */ |
9505 | if (count == 2 && !arm_arch6) | |
9506 | { | |
9507 | if (reg == 15) | |
9508 | reg--; | |
9509 | count++; | |
9510 | } | |
9511 | ||
b5a0636d | 9512 | /* FLDMD may not load more than 16 doubleword registers at a time. Split the |
9513 | load into multiple parts if we have to handle more than 16 registers. */ | |
9514 | if (count > 16) | |
9515 | { | |
9516 | vfp_output_fldmd (stream, base, reg, 16); | |
9517 | vfp_output_fldmd (stream, base, reg + 16, count - 16); | |
9518 | return; | |
9519 | } | |
9520 | ||
a2cd141b | 9521 | fputc ('\t', stream); |
382a2a57 | 9522 | asm_fprintf (stream, "fldmfdd\t%r!, {", base); |
a2cd141b | 9523 | |
7b78e739 | 9524 | for (i = reg; i < reg + count; i++) |
a2cd141b | 9525 | { |
7b78e739 | 9526 | if (i > reg) |
a2cd141b | 9527 | fputs (", ", stream); |
7b78e739 | 9528 | asm_fprintf (stream, "d%d", i); |
a2cd141b | 9529 | } |
9530 | fputs ("}\n", stream); | |
7b78e739 | 9531 | |
a2cd141b | 9532 | } |
9533 | ||
9534 | ||
9535 | /* Output the assembly for a store multiple. */ | |
9536 | ||
9537 | const char * | |
382a2a57 | 9538 | vfp_output_fstmd (rtx * operands) |
a2cd141b | 9539 | { |
9540 | char pattern[100]; | |
9541 | int p; | |
9542 | int base; | |
9543 | int i; | |
9544 | ||
382a2a57 | 9545 | strcpy (pattern, "fstmfdd\t%m0!, {%P1"); |
a2cd141b | 9546 | p = strlen (pattern); |
9547 | ||
ed29c566 | 9548 | gcc_assert (GET_CODE (operands[1]) == REG); |
a2cd141b | 9549 | |
9550 | base = (REGNO (operands[1]) - FIRST_VFP_REGNUM) / 2; | |
9551 | for (i = 1; i < XVECLEN (operands[2], 0); i++) | |
9552 | { | |
9553 | p += sprintf (&pattern[p], ", d%d", base + i); | |
9554 | } | |
9555 | strcpy (&pattern[p], "}"); | |
9556 | ||
9557 | output_asm_insn (pattern, operands); | |
9558 | return ""; | |
9559 | } | |
9560 | ||
9561 | ||
7b78e739 | 9562 | /* Emit RTL to save block of VFP register pairs to the stack. Returns the |
9563 | number of bytes pushed. */ | |
a2cd141b | 9564 | |
7b78e739 | 9565 | static int |
382a2a57 | 9566 | vfp_emit_fstmd (int base_reg, int count) |
a2cd141b | 9567 | { |
9568 | rtx par; | |
9569 | rtx dwarf; | |
9570 | rtx tmp, reg; | |
9571 | int i; | |
9572 | ||
7b78e739 | 9573 | /* Workaround ARM10 VFPr1 bug. Data corruption can occur when exactly two |
9574 | register pairs are stored by a store multiple insn. We avoid this | |
9575 | by pushing an extra pair. */ | |
9576 | if (count == 2 && !arm_arch6) | |
9577 | { | |
9578 | if (base_reg == LAST_VFP_REGNUM - 3) | |
9579 | base_reg -= 2; | |
9580 | count++; | |
9581 | } | |
9582 | ||
b5a0636d | 9583 | /* FSTMD may not store more than 16 doubleword registers at once. Split |
9584 | larger stores into multiple parts (up to a maximum of two, in | |
9585 | practice). */ | |
9586 | if (count > 16) | |
9587 | { | |
9588 | int saved; | |
9589 | /* NOTE: base_reg is an internal register number, so each D register | |
9590 | counts as 2. */ | |
9591 | saved = vfp_emit_fstmd (base_reg + 32, count - 16); | |
9592 | saved += vfp_emit_fstmd (base_reg, 16); | |
9593 | return saved; | |
9594 | } | |
9595 | ||
a2cd141b | 9596 | par = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (count)); |
9597 | dwarf = gen_rtx_SEQUENCE (VOIDmode, rtvec_alloc (count + 1)); | |
9598 | ||
9599 | reg = gen_rtx_REG (DFmode, base_reg); | |
9600 | base_reg += 2; | |
9601 | ||
9602 | XVECEXP (par, 0, 0) | |
9603 | = gen_rtx_SET (VOIDmode, | |
788fcce0 | 9604 | gen_frame_mem (BLKmode, |
9605 | gen_rtx_PRE_DEC (BLKmode, | |
9606 | stack_pointer_rtx)), | |
a2cd141b | 9607 | gen_rtx_UNSPEC (BLKmode, |
9608 | gen_rtvec (1, reg), | |
9609 | UNSPEC_PUSH_MULT)); | |
9610 | ||
9611 | tmp = gen_rtx_SET (VOIDmode, stack_pointer_rtx, | |
382a2a57 | 9612 | plus_constant (stack_pointer_rtx, -(count * 8))); |
a2cd141b | 9613 | RTX_FRAME_RELATED_P (tmp) = 1; |
9614 | XVECEXP (dwarf, 0, 0) = tmp; | |
9615 | ||
9616 | tmp = gen_rtx_SET (VOIDmode, | |
788fcce0 | 9617 | gen_frame_mem (DFmode, stack_pointer_rtx), |
a2cd141b | 9618 | reg); |
9619 | RTX_FRAME_RELATED_P (tmp) = 1; | |
9620 | XVECEXP (dwarf, 0, 1) = tmp; | |
9621 | ||
9622 | for (i = 1; i < count; i++) | |
9623 | { | |
9624 | reg = gen_rtx_REG (DFmode, base_reg); | |
9625 | base_reg += 2; | |
9626 | XVECEXP (par, 0, i) = gen_rtx_USE (VOIDmode, reg); | |
9627 | ||
9628 | tmp = gen_rtx_SET (VOIDmode, | |
788fcce0 | 9629 | gen_frame_mem (DFmode, |
0083a3b9 | 9630 | plus_constant (stack_pointer_rtx, |
9631 | i * 8)), | |
a2cd141b | 9632 | reg); |
9633 | RTX_FRAME_RELATED_P (tmp) = 1; | |
9634 | XVECEXP (dwarf, 0, i + 1) = tmp; | |
9635 | } | |
9636 | ||
9637 | par = emit_insn (par); | |
9638 | REG_NOTES (par) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, dwarf, | |
9639 | REG_NOTES (par)); | |
7b78e739 | 9640 | RTX_FRAME_RELATED_P (par) = 1; |
9641 | ||
382a2a57 | 9642 | return count * 8; |
a2cd141b | 9643 | } |
9644 | ||
bac7fc85 | 9645 | /* Emit a call instruction with pattern PAT. ADDR is the address of |
9646 | the call target. */ | |
9647 | ||
9648 | void | |
9649 | arm_emit_call_insn (rtx pat, rtx addr) | |
9650 | { | |
9651 | rtx insn; | |
9652 | ||
9653 | insn = emit_call_insn (pat); | |
9654 | ||
9655 | /* The PIC register is live on entry to VxWorks PIC PLT entries. | |
9656 | If the call might use such an entry, add a use of the PIC register | |
9657 | to the instruction's CALL_INSN_FUNCTION_USAGE. */ | |
9658 | if (TARGET_VXWORKS_RTP | |
9659 | && flag_pic | |
9660 | && GET_CODE (addr) == SYMBOL_REF | |
9661 | && (SYMBOL_REF_DECL (addr) | |
9662 | ? !targetm.binds_local_p (SYMBOL_REF_DECL (addr)) | |
9663 | : !SYMBOL_REF_LOCAL_P (addr))) | |
9664 | { | |
9665 | require_pic_register (); | |
9666 | use_reg (&CALL_INSN_FUNCTION_USAGE (insn), cfun->machine->pic_reg); | |
9667 | } | |
9668 | } | |
a2cd141b | 9669 | |
66aacf41 | 9670 | /* Output a 'call' insn. */ |
0bd59681 | 9671 | const char * |
ebd88a36 | 9672 | output_call (rtx *operands) |
8cd42699 | 9673 | { |
ed29c566 | 9674 | gcc_assert (!arm_arch5); /* Patterns should call blx <reg> directly. */ |
8cd42699 | 9675 | |
f1039640 | 9676 | /* Handle calls to lr using ip (which may be clobbered in subr anyway). */ |
bba10fb8 | 9677 | if (REGNO (operands[0]) == LR_REGNUM) |
8cd42699 | 9678 | { |
bba10fb8 | 9679 | operands[0] = gen_rtx_REG (SImode, IP_REGNUM); |
39d5d2b6 | 9680 | output_asm_insn ("mov%?\t%0, %|lr", operands); |
8cd42699 | 9681 | } |
9e7454d0 | 9682 | |
39d5d2b6 | 9683 | output_asm_insn ("mov%?\t%|lr, %|pc", operands); |
9e7454d0 | 9684 | |
f1039640 | 9685 | if (TARGET_INTERWORK || arm_arch4t) |
99ecc60d | 9686 | output_asm_insn ("bx%?\t%0", operands); |
9687 | else | |
9688 | output_asm_insn ("mov%?\t%|pc, %0", operands); | |
9e7454d0 | 9689 | |
a690ede2 | 9690 | return ""; |
9691 | } | |
8cd42699 | 9692 | |
66aacf41 | 9693 | /* Output a 'call' insn that is a reference in memory. */ |
0bd59681 | 9694 | const char * |
ebd88a36 | 9695 | output_call_mem (rtx *operands) |
9c08d1fa | 9696 | { |
f1039640 | 9697 | if (TARGET_INTERWORK && !arm_arch5) |
99ecc60d | 9698 | { |
9699 | output_asm_insn ("ldr%?\t%|ip, %0", operands); | |
9700 | output_asm_insn ("mov%?\t%|lr, %|pc", operands); | |
9701 | output_asm_insn ("bx%?\t%|ip", operands); | |
9702 | } | |
136396e8 | 9703 | else if (regno_use_in (LR_REGNUM, operands[0])) |
9704 | { | |
9705 | /* LR is used in the memory address. We load the address in the | |
9706 | first instruction. It's safe to use IP as the target of the | |
9707 | load since the call will kill it anyway. */ | |
9708 | output_asm_insn ("ldr%?\t%|ip, %0", operands); | |
f1039640 | 9709 | if (arm_arch5) |
9a821292 | 9710 | output_asm_insn ("blx%?\t%|ip", operands); |
f1039640 | 9711 | else |
9712 | { | |
9713 | output_asm_insn ("mov%?\t%|lr, %|pc", operands); | |
9714 | if (arm_arch4t) | |
9715 | output_asm_insn ("bx%?\t%|ip", operands); | |
9716 | else | |
9717 | output_asm_insn ("mov%?\t%|pc, %|ip", operands); | |
9718 | } | |
136396e8 | 9719 | } |
99ecc60d | 9720 | else |
9721 | { | |
9722 | output_asm_insn ("mov%?\t%|lr, %|pc", operands); | |
9723 | output_asm_insn ("ldr%?\t%|pc, %0", operands); | |
9724 | } | |
9725 | ||
a690ede2 | 9726 | return ""; |
9727 | } | |
9c08d1fa | 9728 | |
9729 | ||
e8ff69a4 | 9730 | /* Output a move from arm registers to an fpa registers. |
9731 | OPERANDS[0] is an fpa register. | |
9c08d1fa | 9732 | OPERANDS[1] is the first registers of an arm register pair. */ |
0bd59681 | 9733 | const char * |
ebd88a36 | 9734 | output_mov_long_double_fpa_from_arm (rtx *operands) |
9c08d1fa | 9735 | { |
9736 | int arm_reg0 = REGNO (operands[1]); | |
9737 | rtx ops[3]; | |
9738 | ||
ed29c566 | 9739 | gcc_assert (arm_reg0 != IP_REGNUM); |
a690ede2 | 9740 | |
76676c8e | 9741 | ops[0] = gen_rtx_REG (SImode, arm_reg0); |
9742 | ops[1] = gen_rtx_REG (SImode, 1 + arm_reg0); | |
9743 | ops[2] = gen_rtx_REG (SImode, 2 + arm_reg0); | |
9e7454d0 | 9744 | |
25f905c2 | 9745 | output_asm_insn ("stm%(fd%)\t%|sp!, {%0, %1, %2}", ops); |
39d5d2b6 | 9746 | output_asm_insn ("ldf%?e\t%0, [%|sp], #12", operands); |
9e7454d0 | 9747 | |
a690ede2 | 9748 | return ""; |
9749 | } | |
9c08d1fa | 9750 | |
e8ff69a4 | 9751 | /* Output a move from an fpa register to arm registers. |
9c08d1fa | 9752 | OPERANDS[0] is the first registers of an arm register pair. |
e8ff69a4 | 9753 | OPERANDS[1] is an fpa register. */ |
0bd59681 | 9754 | const char * |
ebd88a36 | 9755 | output_mov_long_double_arm_from_fpa (rtx *operands) |
9c08d1fa | 9756 | { |
9757 | int arm_reg0 = REGNO (operands[0]); | |
9758 | rtx ops[3]; | |
9759 | ||
ed29c566 | 9760 | gcc_assert (arm_reg0 != IP_REGNUM); |
a690ede2 | 9761 | |
76676c8e | 9762 | ops[0] = gen_rtx_REG (SImode, arm_reg0); |
9763 | ops[1] = gen_rtx_REG (SImode, 1 + arm_reg0); | |
9764 | ops[2] = gen_rtx_REG (SImode, 2 + arm_reg0); | |
9c08d1fa | 9765 | |
39d5d2b6 | 9766 | output_asm_insn ("stf%?e\t%1, [%|sp, #-12]!", operands); |
25f905c2 | 9767 | output_asm_insn ("ldm%(fd%)\t%|sp!, {%0, %1, %2}", ops); |
a690ede2 | 9768 | return ""; |
9769 | } | |
9c08d1fa | 9770 | |
9771 | /* Output a move from arm registers to arm registers of a long double | |
9772 | OPERANDS[0] is the destination. | |
9773 | OPERANDS[1] is the source. */ | |
0bd59681 | 9774 | const char * |
ebd88a36 | 9775 | output_mov_long_double_arm_from_arm (rtx *operands) |
9c08d1fa | 9776 | { |
66aacf41 | 9777 | /* We have to be careful here because the two might overlap. */ |
9c08d1fa | 9778 | int dest_start = REGNO (operands[0]); |
9779 | int src_start = REGNO (operands[1]); | |
9780 | rtx ops[2]; | |
9781 | int i; | |
9782 | ||
9783 | if (dest_start < src_start) | |
9784 | { | |
9785 | for (i = 0; i < 3; i++) | |
9786 | { | |
76676c8e | 9787 | ops[0] = gen_rtx_REG (SImode, dest_start + i); |
9788 | ops[1] = gen_rtx_REG (SImode, src_start + i); | |
f6b01912 | 9789 | output_asm_insn ("mov%?\t%0, %1", ops); |
9c08d1fa | 9790 | } |
9791 | } | |
9792 | else | |
9793 | { | |
9794 | for (i = 2; i >= 0; i--) | |
9795 | { | |
76676c8e | 9796 | ops[0] = gen_rtx_REG (SImode, dest_start + i); |
9797 | ops[1] = gen_rtx_REG (SImode, src_start + i); | |
f6b01912 | 9798 | output_asm_insn ("mov%?\t%0, %1", ops); |
9c08d1fa | 9799 | } |
9800 | } | |
a690ede2 | 9801 | |
9c08d1fa | 9802 | return ""; |
9803 | } | |
9804 | ||
9805 | ||
e8ff69a4 | 9806 | /* Output a move from arm registers to an fpa registers. |
9807 | OPERANDS[0] is an fpa register. | |
8cd42699 | 9808 | OPERANDS[1] is the first registers of an arm register pair. */ |
0bd59681 | 9809 | const char * |
ebd88a36 | 9810 | output_mov_double_fpa_from_arm (rtx *operands) |
8cd42699 | 9811 | { |
9812 | int arm_reg0 = REGNO (operands[1]); | |
9813 | rtx ops[2]; | |
9814 | ||
ed29c566 | 9815 | gcc_assert (arm_reg0 != IP_REGNUM); |
9e7454d0 | 9816 | |
76676c8e | 9817 | ops[0] = gen_rtx_REG (SImode, arm_reg0); |
9818 | ops[1] = gen_rtx_REG (SImode, 1 + arm_reg0); | |
25f905c2 | 9819 | output_asm_insn ("stm%(fd%)\t%|sp!, {%0, %1}", ops); |
39d5d2b6 | 9820 | output_asm_insn ("ldf%?d\t%0, [%|sp], #8", operands); |
a690ede2 | 9821 | return ""; |
9822 | } | |
8cd42699 | 9823 | |
e8ff69a4 | 9824 | /* Output a move from an fpa register to arm registers. |
8cd42699 | 9825 | OPERANDS[0] is the first registers of an arm register pair. |
e8ff69a4 | 9826 | OPERANDS[1] is an fpa register. */ |
0bd59681 | 9827 | const char * |
ebd88a36 | 9828 | output_mov_double_arm_from_fpa (rtx *operands) |
8cd42699 | 9829 | { |
9830 | int arm_reg0 = REGNO (operands[0]); | |
9831 | rtx ops[2]; | |
9832 | ||
ed29c566 | 9833 | gcc_assert (arm_reg0 != IP_REGNUM); |
a690ede2 | 9834 | |
76676c8e | 9835 | ops[0] = gen_rtx_REG (SImode, arm_reg0); |
9836 | ops[1] = gen_rtx_REG (SImode, 1 + arm_reg0); | |
39d5d2b6 | 9837 | output_asm_insn ("stf%?d\t%1, [%|sp, #-8]!", operands); |
25f905c2 | 9838 | output_asm_insn ("ldm%(fd%)\t%|sp!, {%0, %1}", ops); |
a690ede2 | 9839 | return ""; |
9840 | } | |
8cd42699 | 9841 | |
9842 | /* Output a move between double words. | |
9843 | It must be REG<-REG, REG<-CONST_DOUBLE, REG<-CONST_INT, REG<-MEM | |
9844 | or MEM<-REG and all MEMs must be offsettable addresses. */ | |
0bd59681 | 9845 | const char * |
ebd88a36 | 9846 | output_move_double (rtx *operands) |
8cd42699 | 9847 | { |
9848 | enum rtx_code code0 = GET_CODE (operands[0]); | |
9849 | enum rtx_code code1 = GET_CODE (operands[1]); | |
6019008a | 9850 | rtx otherops[3]; |
8cd42699 | 9851 | |
9852 | if (code0 == REG) | |
9853 | { | |
9854 | int reg0 = REGNO (operands[0]); | |
9855 | ||
76676c8e | 9856 | otherops[0] = gen_rtx_REG (SImode, 1 + reg0); |
9e7454d0 | 9857 | |
ed29c566 | 9858 | gcc_assert (code1 == MEM); /* Constraints should ensure this. */ |
9859 | ||
9860 | switch (GET_CODE (XEXP (operands[1], 0))) | |
8cd42699 | 9861 | { |
ed29c566 | 9862 | case REG: |
25f905c2 | 9863 | output_asm_insn ("ldm%(ia%)\t%m1, %M0", operands); |
ed29c566 | 9864 | break; |
a60f1840 | 9865 | |
ed29c566 | 9866 | case PRE_INC: |
9867 | gcc_assert (TARGET_LDRD); | |
25f905c2 | 9868 | output_asm_insn ("ldr%(d%)\t%0, [%m1, #8]!", operands); |
ed29c566 | 9869 | break; |
a60f1840 | 9870 | |
ed29c566 | 9871 | case PRE_DEC: |
25f905c2 | 9872 | if (TARGET_LDRD) |
9873 | output_asm_insn ("ldr%(d%)\t%0, [%m1, #-8]!", operands); | |
9874 | else | |
9875 | output_asm_insn ("ldm%(db%)\t%m1!, %M0", operands); | |
ed29c566 | 9876 | break; |
a60f1840 | 9877 | |
ed29c566 | 9878 | case POST_INC: |
25f905c2 | 9879 | output_asm_insn ("ldm%(ia%)\t%m1!, %M0", operands); |
ed29c566 | 9880 | break; |
a60f1840 | 9881 | |
ed29c566 | 9882 | case POST_DEC: |
9883 | gcc_assert (TARGET_LDRD); | |
25f905c2 | 9884 | output_asm_insn ("ldr%(d%)\t%0, [%m1], #-8", operands); |
ed29c566 | 9885 | break; |
a60f1840 | 9886 | |
ed29c566 | 9887 | case PRE_MODIFY: |
9888 | case POST_MODIFY: | |
9889 | otherops[0] = operands[0]; | |
9890 | otherops[1] = XEXP (XEXP (XEXP (operands[1], 0), 1), 0); | |
9891 | otherops[2] = XEXP (XEXP (XEXP (operands[1], 0), 1), 1); | |
a60f1840 | 9892 | |
ed29c566 | 9893 | if (GET_CODE (XEXP (operands[1], 0)) == PRE_MODIFY) |
8cd42699 | 9894 | { |
ed29c566 | 9895 | if (reg_overlap_mentioned_p (otherops[0], otherops[2])) |
a8a3b539 | 9896 | { |
ed29c566 | 9897 | /* Registers overlap so split out the increment. */ |
9898 | output_asm_insn ("add%?\t%1, %1, %2", otherops); | |
25f905c2 | 9899 | output_asm_insn ("ldr%(d%)\t%0, [%1] @split", otherops); |
a8a3b539 | 9900 | } |
9901 | else | |
f1225f6f | 9902 | { |
9903 | /* IWMMXT allows offsets larger than ldrd can handle, | |
9904 | fix these up with a pair of ldr. */ | |
9905 | if (GET_CODE (otherops[2]) == CONST_INT | |
9906 | && (INTVAL(otherops[2]) <= -256 | |
9907 | || INTVAL(otherops[2]) >= 256)) | |
9908 | { | |
9909 | output_asm_insn ("ldr%?\t%0, [%1, %2]!", otherops); | |
9910 | otherops[0] = gen_rtx_REG (SImode, 1 + reg0); | |
9911 | output_asm_insn ("ldr%?\t%0, [%1, #4]", otherops); | |
9912 | } | |
9913 | else | |
9914 | output_asm_insn ("ldr%(d%)\t%0, [%1, %2]!", otherops); | |
9915 | } | |
ed29c566 | 9916 | } |
9917 | else | |
9918 | { | |
f1225f6f | 9919 | /* IWMMXT allows offsets larger than ldrd can handle, |
9920 | fix these up with a pair of ldr. */ | |
9921 | if (GET_CODE (otherops[2]) == CONST_INT | |
9922 | && (INTVAL(otherops[2]) <= -256 | |
9923 | || INTVAL(otherops[2]) >= 256)) | |
9924 | { | |
9925 | otherops[0] = gen_rtx_REG (SImode, 1 + reg0); | |
9926 | output_asm_insn ("ldr%?\t%0, [%1, #4]", otherops); | |
9927 | otherops[0] = operands[0]; | |
9928 | output_asm_insn ("ldr%?\t%0, [%1], %2", otherops); | |
9929 | } | |
9930 | else | |
9931 | /* We only allow constant increments, so this is safe. */ | |
9932 | output_asm_insn ("ldr%(d%)\t%0, [%1], %2", otherops); | |
ed29c566 | 9933 | } |
9934 | break; | |
a60f1840 | 9935 | |
ed29c566 | 9936 | case LABEL_REF: |
9937 | case CONST: | |
9938 | output_asm_insn ("adr%?\t%0, %1", operands); | |
25f905c2 | 9939 | output_asm_insn ("ldm%(ia%)\t%0, %M0", operands); |
ed29c566 | 9940 | break; |
a60f1840 | 9941 | |
25f905c2 | 9942 | /* ??? This needs checking for thumb2. */ |
ed29c566 | 9943 | default: |
9944 | if (arm_add_operand (XEXP (XEXP (operands[1], 0), 1), | |
9945 | GET_MODE (XEXP (XEXP (operands[1], 0), 1)))) | |
9946 | { | |
9947 | otherops[0] = operands[0]; | |
9948 | otherops[1] = XEXP (XEXP (operands[1], 0), 0); | |
9949 | otherops[2] = XEXP (XEXP (operands[1], 0), 1); | |
a60f1840 | 9950 | |
ed29c566 | 9951 | if (GET_CODE (XEXP (operands[1], 0)) == PLUS) |
a8a3b539 | 9952 | { |
ed29c566 | 9953 | if (GET_CODE (otherops[2]) == CONST_INT) |
64266206 | 9954 | { |
ed29c566 | 9955 | switch ((int) INTVAL (otherops[2])) |
64266206 | 9956 | { |
ed29c566 | 9957 | case -8: |
25f905c2 | 9958 | output_asm_insn ("ldm%(db%)\t%1, %M0", otherops); |
ed29c566 | 9959 | return ""; |
9960 | case -4: | |
25f905c2 | 9961 | if (TARGET_THUMB2) |
9962 | break; | |
9963 | output_asm_insn ("ldm%(da%)\t%1, %M0", otherops); | |
ed29c566 | 9964 | return ""; |
9965 | case 4: | |
25f905c2 | 9966 | if (TARGET_THUMB2) |
9967 | break; | |
9968 | output_asm_insn ("ldm%(ib%)\t%1, %M0", otherops); | |
ed29c566 | 9969 | return ""; |
a8a3b539 | 9970 | } |
ed29c566 | 9971 | } |
9972 | if (TARGET_LDRD | |
9973 | && (GET_CODE (otherops[2]) == REG | |
9974 | || (GET_CODE (otherops[2]) == CONST_INT | |
9975 | && INTVAL (otherops[2]) > -256 | |
9976 | && INTVAL (otherops[2]) < 256))) | |
9977 | { | |
9978 | if (reg_overlap_mentioned_p (otherops[0], | |
9979 | otherops[2])) | |
a8a3b539 | 9980 | { |
ed29c566 | 9981 | /* Swap base and index registers over to |
9982 | avoid a conflict. */ | |
9983 | otherops[1] = XEXP (XEXP (operands[1], 0), 1); | |
9984 | otherops[2] = XEXP (XEXP (operands[1], 0), 0); | |
a8a3b539 | 9985 | } |
ed29c566 | 9986 | /* If both registers conflict, it will usually |
9987 | have been fixed by a splitter. */ | |
9988 | if (reg_overlap_mentioned_p (otherops[0], otherops[2])) | |
a8a3b539 | 9989 | { |
ed29c566 | 9990 | output_asm_insn ("add%?\t%1, %1, %2", otherops); |
25f905c2 | 9991 | output_asm_insn ("ldr%(d%)\t%0, [%1]", |
ed29c566 | 9992 | otherops); |
64266206 | 9993 | } |
9994 | else | |
25f905c2 | 9995 | output_asm_insn ("ldr%(d%)\t%0, [%1, %2]", otherops); |
ed29c566 | 9996 | return ""; |
64266206 | 9997 | } |
a60f1840 | 9998 | |
ed29c566 | 9999 | if (GET_CODE (otherops[2]) == CONST_INT) |
64266206 | 10000 | { |
ed29c566 | 10001 | if (!(const_ok_for_arm (INTVAL (otherops[2])))) |
10002 | output_asm_insn ("sub%?\t%0, %1, #%n2", otherops); | |
10003 | else | |
10004 | output_asm_insn ("add%?\t%0, %1, %2", otherops); | |
64266206 | 10005 | } |
10006 | else | |
ed29c566 | 10007 | output_asm_insn ("add%?\t%0, %1, %2", otherops); |
10008 | } | |
10009 | else | |
10010 | output_asm_insn ("sub%?\t%0, %1, %2", otherops); | |
10011 | ||
25f905c2 | 10012 | return "ldm%(ia%)\t%0, %M0"; |
ed29c566 | 10013 | } |
10014 | else | |
10015 | { | |
10016 | otherops[1] = adjust_address (operands[1], SImode, 4); | |
10017 | /* Take care of overlapping base/data reg. */ | |
10018 | if (reg_mentioned_p (operands[0], operands[1])) | |
10019 | { | |
10020 | output_asm_insn ("ldr%?\t%0, %1", otherops); | |
10021 | output_asm_insn ("ldr%?\t%0, %1", operands); | |
10022 | } | |
10023 | else | |
10024 | { | |
10025 | output_asm_insn ("ldr%?\t%0, %1", operands); | |
10026 | output_asm_insn ("ldr%?\t%0, %1", otherops); | |
8cd42699 | 10027 | } |
10028 | } | |
10029 | } | |
8cd42699 | 10030 | } |
ed29c566 | 10031 | else |
8cd42699 | 10032 | { |
ed29c566 | 10033 | /* Constraints should ensure this. */ |
10034 | gcc_assert (code0 == MEM && code1 == REG); | |
10035 | gcc_assert (REGNO (operands[1]) != IP_REGNUM); | |
64266206 | 10036 | |
9c08d1fa | 10037 | switch (GET_CODE (XEXP (operands[0], 0))) |
10038 | { | |
10039 | case REG: | |
25f905c2 | 10040 | output_asm_insn ("stm%(ia%)\t%m0, %M1", operands); |
9c08d1fa | 10041 | break; |
64266206 | 10042 | |
9c08d1fa | 10043 | case PRE_INC: |
ed29c566 | 10044 | gcc_assert (TARGET_LDRD); |
25f905c2 | 10045 | output_asm_insn ("str%(d%)\t%1, [%m0, #8]!", operands); |
9c08d1fa | 10046 | break; |
64266206 | 10047 | |
9c08d1fa | 10048 | case PRE_DEC: |
25f905c2 | 10049 | if (TARGET_LDRD) |
10050 | output_asm_insn ("str%(d%)\t%1, [%m0, #-8]!", operands); | |
10051 | else | |
10052 | output_asm_insn ("stm%(db%)\t%m0!, %M1", operands); | |
9c08d1fa | 10053 | break; |
64266206 | 10054 | |
9c08d1fa | 10055 | case POST_INC: |
25f905c2 | 10056 | output_asm_insn ("stm%(ia%)\t%m0!, %M1", operands); |
9c08d1fa | 10057 | break; |
64266206 | 10058 | |
9c08d1fa | 10059 | case POST_DEC: |
ed29c566 | 10060 | gcc_assert (TARGET_LDRD); |
25f905c2 | 10061 | output_asm_insn ("str%(d%)\t%1, [%m0], #-8", operands); |
a8a3b539 | 10062 | break; |
10063 | ||
10064 | case PRE_MODIFY: | |
10065 | case POST_MODIFY: | |
10066 | otherops[0] = operands[1]; | |
10067 | otherops[1] = XEXP (XEXP (XEXP (operands[0], 0), 1), 0); | |
10068 | otherops[2] = XEXP (XEXP (XEXP (operands[0], 0), 1), 1); | |
10069 | ||
f1225f6f | 10070 | /* IWMMXT allows offsets larger than ldrd can handle, |
10071 | fix these up with a pair of ldr. */ | |
10072 | if (GET_CODE (otherops[2]) == CONST_INT | |
10073 | && (INTVAL(otherops[2]) <= -256 | |
10074 | || INTVAL(otherops[2]) >= 256)) | |
10075 | { | |
10076 | rtx reg1; | |
10077 | reg1 = gen_rtx_REG (SImode, 1 + REGNO (operands[1])); | |
10078 | if (GET_CODE (XEXP (operands[0], 0)) == PRE_MODIFY) | |
10079 | { | |
10080 | output_asm_insn ("ldr%?\t%0, [%1, %2]!", otherops); | |
10081 | otherops[0] = reg1; | |
10082 | output_asm_insn ("ldr%?\t%0, [%1, #4]", otherops); | |
10083 | } | |
10084 | else | |
10085 | { | |
10086 | otherops[0] = reg1; | |
10087 | output_asm_insn ("ldr%?\t%0, [%1, #4]", otherops); | |
10088 | otherops[0] = operands[1]; | |
10089 | output_asm_insn ("ldr%?\t%0, [%1], %2", otherops); | |
10090 | } | |
10091 | } | |
10092 | else if (GET_CODE (XEXP (operands[0], 0)) == PRE_MODIFY) | |
25f905c2 | 10093 | output_asm_insn ("str%(d%)\t%0, [%1, %2]!", otherops); |
a8a3b539 | 10094 | else |
25f905c2 | 10095 | output_asm_insn ("str%(d%)\t%0, [%1], %2", otherops); |
9c08d1fa | 10096 | break; |
64266206 | 10097 | |
10098 | case PLUS: | |
a8a3b539 | 10099 | otherops[2] = XEXP (XEXP (operands[0], 0), 1); |
10100 | if (GET_CODE (otherops[2]) == CONST_INT) | |
64266206 | 10101 | { |
000c5098 | 10102 | switch ((int) INTVAL (XEXP (XEXP (operands[0], 0), 1))) |
64266206 | 10103 | { |
10104 | case -8: | |
25f905c2 | 10105 | output_asm_insn ("stm%(db%)\t%m0, %M1", operands); |
64266206 | 10106 | return ""; |
10107 | ||
10108 | case -4: | |
25f905c2 | 10109 | if (TARGET_THUMB2) |
10110 | break; | |
10111 | output_asm_insn ("stm%(da%)\t%m0, %M1", operands); | |
64266206 | 10112 | return ""; |
10113 | ||
10114 | case 4: | |
25f905c2 | 10115 | if (TARGET_THUMB2) |
10116 | break; | |
10117 | output_asm_insn ("stm%(ib%)\t%m0, %M1", operands); | |
64266206 | 10118 | return ""; |
10119 | } | |
10120 | } | |
a8a3b539 | 10121 | if (TARGET_LDRD |
10122 | && (GET_CODE (otherops[2]) == REG | |
10123 | || (GET_CODE (otherops[2]) == CONST_INT | |
10124 | && INTVAL (otherops[2]) > -256 | |
10125 | && INTVAL (otherops[2]) < 256))) | |
10126 | { | |
10127 | otherops[0] = operands[1]; | |
10128 | otherops[1] = XEXP (XEXP (operands[0], 0), 0); | |
25f905c2 | 10129 | output_asm_insn ("str%(d%)\t%0, [%1, %2]", otherops); |
a8a3b539 | 10130 | return ""; |
10131 | } | |
64266206 | 10132 | /* Fall through */ |
10133 | ||
9c08d1fa | 10134 | default: |
d35c2d18 | 10135 | otherops[0] = adjust_address (operands[0], SImode, 4); |
76676c8e | 10136 | otherops[1] = gen_rtx_REG (SImode, 1 + REGNO (operands[1])); |
f6b01912 | 10137 | output_asm_insn ("str%?\t%1, %0", operands); |
10138 | output_asm_insn ("str%?\t%1, %0", otherops); | |
8cd42699 | 10139 | } |
10140 | } | |
8cd42699 | 10141 | |
f6b01912 | 10142 | return ""; |
10143 | } | |
8cd42699 | 10144 | |
d98a3884 | 10145 | /* Output a move, load or store for quad-word vectors in ARM registers. Only |
10146 | handles MEMs accepted by neon_vector_mem_operand with CORE=true. */ | |
25f905c2 | 10147 | |
10148 | const char * | |
d98a3884 | 10149 | output_move_quad (rtx *operands) |
25f905c2 | 10150 | { |
d98a3884 | 10151 | if (REG_P (operands[0])) |
10152 | { | |
10153 | /* Load, or reg->reg move. */ | |
25f905c2 | 10154 | |
d98a3884 | 10155 | if (MEM_P (operands[1])) |
10156 | { | |
10157 | switch (GET_CODE (XEXP (operands[1], 0))) | |
10158 | { | |
10159 | case REG: | |
10160 | output_asm_insn ("ldm%(ia%)\t%m1, %M0", operands); | |
10161 | break; | |
10162 | ||
10163 | case LABEL_REF: | |
10164 | case CONST: | |
10165 | output_asm_insn ("adr%?\t%0, %1", operands); | |
10166 | output_asm_insn ("ldm%(ia%)\t%0, %M0", operands); | |
10167 | break; | |
10168 | ||
10169 | default: | |
10170 | gcc_unreachable (); | |
10171 | } | |
10172 | } | |
10173 | else | |
10174 | { | |
10175 | rtx ops[2]; | |
10176 | int dest, src, i; | |
25f905c2 | 10177 | |
d98a3884 | 10178 | gcc_assert (REG_P (operands[1])); |
25f905c2 | 10179 | |
d98a3884 | 10180 | dest = REGNO (operands[0]); |
10181 | src = REGNO (operands[1]); | |
25f905c2 | 10182 | |
d98a3884 | 10183 | /* This seems pretty dumb, but hopefully GCC won't try to do it |
10184 | very often. */ | |
10185 | if (dest < src) | |
10186 | for (i = 0; i < 4; i++) | |
10187 | { | |
10188 | ops[0] = gen_rtx_REG (SImode, dest + i); | |
10189 | ops[1] = gen_rtx_REG (SImode, src + i); | |
10190 | output_asm_insn ("mov%?\t%0, %1", ops); | |
10191 | } | |
10192 | else | |
10193 | for (i = 3; i >= 0; i--) | |
10194 | { | |
10195 | ops[0] = gen_rtx_REG (SImode, dest + i); | |
10196 | ops[1] = gen_rtx_REG (SImode, src + i); | |
10197 | output_asm_insn ("mov%?\t%0, %1", ops); | |
10198 | } | |
10199 | } | |
10200 | } | |
10201 | else | |
10202 | { | |
10203 | gcc_assert (MEM_P (operands[0])); | |
10204 | gcc_assert (REG_P (operands[1])); | |
10205 | gcc_assert (!reg_overlap_mentioned_p (operands[1], operands[0])); | |
10206 | ||
10207 | switch (GET_CODE (XEXP (operands[0], 0))) | |
10208 | { | |
10209 | case REG: | |
10210 | output_asm_insn ("stm%(ia%)\t%m0, %M1", operands); | |
10211 | break; | |
10212 | ||
10213 | default: | |
10214 | gcc_unreachable (); | |
10215 | } | |
10216 | } | |
10217 | ||
10218 | return ""; | |
10219 | } | |
10220 | ||
10221 | /* Output a VFP load or store instruction. */ | |
10222 | ||
10223 | const char * | |
10224 | output_move_vfp (rtx *operands) | |
10225 | { | |
10226 | rtx reg, mem, addr, ops[2]; | |
10227 | int load = REG_P (operands[0]); | |
10228 | int dp = GET_MODE_SIZE (GET_MODE (operands[0])) == 8; | |
10229 | int integer_p = GET_MODE_CLASS (GET_MODE (operands[0])) == MODE_INT; | |
10230 | const char *template; | |
10231 | char buff[50]; | |
10232 | enum machine_mode mode; | |
10233 | ||
10234 | reg = operands[!load]; | |
10235 | mem = operands[load]; | |
10236 | ||
10237 | mode = GET_MODE (reg); | |
10238 | ||
10239 | gcc_assert (REG_P (reg)); | |
10240 | gcc_assert (IS_VFP_REGNUM (REGNO (reg))); | |
10241 | gcc_assert (mode == SFmode | |
10242 | || mode == DFmode | |
10243 | || mode == SImode | |
10244 | || mode == DImode | |
10245 | || (TARGET_NEON && VALID_NEON_DREG_MODE (mode))); | |
10246 | gcc_assert (MEM_P (mem)); | |
10247 | ||
10248 | addr = XEXP (mem, 0); | |
10249 | ||
10250 | switch (GET_CODE (addr)) | |
10251 | { | |
10252 | case PRE_DEC: | |
10253 | template = "f%smdb%c%%?\t%%0!, {%%%s1}%s"; | |
10254 | ops[0] = XEXP (addr, 0); | |
10255 | ops[1] = reg; | |
25f905c2 | 10256 | break; |
10257 | ||
10258 | case POST_INC: | |
10259 | template = "f%smia%c%%?\t%%0!, {%%%s1}%s"; | |
10260 | ops[0] = XEXP (addr, 0); | |
10261 | ops[1] = reg; | |
10262 | break; | |
10263 | ||
10264 | default: | |
10265 | template = "f%s%c%%?\t%%%s0, %%1%s"; | |
10266 | ops[0] = reg; | |
10267 | ops[1] = mem; | |
10268 | break; | |
10269 | } | |
10270 | ||
10271 | sprintf (buff, template, | |
10272 | load ? "ld" : "st", | |
10273 | dp ? 'd' : 's', | |
10274 | dp ? "P" : "", | |
10275 | integer_p ? "\t%@ int" : ""); | |
10276 | output_asm_insn (buff, ops); | |
10277 | ||
10278 | return ""; | |
10279 | } | |
10280 | ||
d98a3884 | 10281 | /* Output a Neon quad-word load or store, or a load or store for |
10282 | larger structure modes. We could also support post-modify forms using | |
10283 | VLD1/VST1 (for the vectorizer, and perhaps otherwise), but we don't do that | |
10284 | yet. | |
10285 | WARNING: The ordering of elements in memory is weird in big-endian mode, | |
10286 | because we use VSTM instead of VST1, to make it easy to make vector stores | |
10287 | via ARM registers write values in the same order as stores direct from Neon | |
10288 | registers. For example, the byte ordering of a quadword vector with 16-byte | |
10289 | elements like this: | |
10290 | ||
10291 | [e7:e6:e5:e4:e3:e2:e1:e0] (highest-numbered element first) | |
10292 | ||
10293 | will be (with lowest address first, h = most-significant byte, | |
10294 | l = least-significant byte of element): | |
10295 | ||
10296 | [e3h, e3l, e2h, e2l, e1h, e1l, e0h, e0l, | |
10297 | e7h, e7l, e6h, e6l, e5h, e5l, e4h, e4l] | |
10298 | ||
10299 | When necessary, quadword registers (dN, dN+1) are moved to ARM registers from | |
10300 | rN in the order: | |
10301 | ||
10302 | dN -> (rN+1, rN), dN+1 -> (rN+3, rN+2) | |
10303 | ||
10304 | So that STM/LDM can be used on vectors in ARM registers, and the same memory | |
10305 | layout will result as if VSTM/VLDM were used. */ | |
10306 | ||
10307 | const char * | |
10308 | output_move_neon (rtx *operands) | |
10309 | { | |
10310 | rtx reg, mem, addr, ops[2]; | |
10311 | int regno, load = REG_P (operands[0]); | |
10312 | const char *template; | |
10313 | char buff[50]; | |
10314 | enum machine_mode mode; | |
10315 | ||
10316 | reg = operands[!load]; | |
10317 | mem = operands[load]; | |
10318 | ||
10319 | mode = GET_MODE (reg); | |
10320 | ||
10321 | gcc_assert (REG_P (reg)); | |
10322 | regno = REGNO (reg); | |
10323 | gcc_assert (VFP_REGNO_OK_FOR_DOUBLE (regno) | |
10324 | || NEON_REGNO_OK_FOR_QUAD (regno)); | |
10325 | gcc_assert (VALID_NEON_DREG_MODE (mode) | |
10326 | || VALID_NEON_QREG_MODE (mode) | |
10327 | || VALID_NEON_STRUCT_MODE (mode)); | |
10328 | gcc_assert (MEM_P (mem)); | |
10329 | ||
10330 | addr = XEXP (mem, 0); | |
10331 | ||
10332 | /* Strip off const from addresses like (const (plus (...))). */ | |
10333 | if (GET_CODE (addr) == CONST && GET_CODE (XEXP (addr, 0)) == PLUS) | |
10334 | addr = XEXP (addr, 0); | |
10335 | ||
10336 | switch (GET_CODE (addr)) | |
10337 | { | |
10338 | case POST_INC: | |
10339 | template = "v%smia%%?\t%%0!, %%h1"; | |
10340 | ops[0] = XEXP (addr, 0); | |
10341 | ops[1] = reg; | |
10342 | break; | |
10343 | ||
10344 | case POST_MODIFY: | |
10345 | /* FIXME: Not currently enabled in neon_vector_mem_operand. */ | |
10346 | gcc_unreachable (); | |
10347 | ||
10348 | case LABEL_REF: | |
10349 | case PLUS: | |
10350 | { | |
10351 | int nregs = HARD_REGNO_NREGS (REGNO (reg), mode) / 2; | |
10352 | int i; | |
10353 | int overlap = -1; | |
10354 | for (i = 0; i < nregs; i++) | |
10355 | { | |
10356 | /* We're only using DImode here because it's a convenient size. */ | |
10357 | ops[0] = gen_rtx_REG (DImode, REGNO (reg) + 2 * i); | |
10358 | ops[1] = adjust_address (mem, SImode, 8 * i); | |
10359 | if (reg_overlap_mentioned_p (ops[0], mem)) | |
10360 | { | |
10361 | gcc_assert (overlap == -1); | |
10362 | overlap = i; | |
10363 | } | |
10364 | else | |
10365 | { | |
10366 | sprintf (buff, "v%sr%%?\t%%P0, %%1", load ? "ld" : "st"); | |
10367 | output_asm_insn (buff, ops); | |
10368 | } | |
10369 | } | |
10370 | if (overlap != -1) | |
10371 | { | |
10372 | ops[0] = gen_rtx_REG (DImode, REGNO (reg) + 2 * overlap); | |
10373 | ops[1] = adjust_address (mem, SImode, 8 * overlap); | |
10374 | sprintf (buff, "v%sr%%?\t%%P0, %%1", load ? "ld" : "st"); | |
10375 | output_asm_insn (buff, ops); | |
10376 | } | |
10377 | ||
10378 | return ""; | |
10379 | } | |
10380 | ||
10381 | default: | |
10382 | template = "v%smia%%?\t%%m0, %%h1"; | |
10383 | ops[0] = mem; | |
10384 | ops[1] = reg; | |
10385 | } | |
10386 | ||
10387 | sprintf (buff, template, load ? "ld" : "st"); | |
10388 | output_asm_insn (buff, ops); | |
10389 | ||
10390 | return ""; | |
10391 | } | |
10392 | ||
a12be13c | 10393 | /* Output an ADD r, s, #n where n may be too big for one instruction. |
10394 | If adding zero to one register, output nothing. */ | |
0bd59681 | 10395 | const char * |
ebd88a36 | 10396 | output_add_immediate (rtx *operands) |
8cd42699 | 10397 | { |
a690ede2 | 10398 | HOST_WIDE_INT n = INTVAL (operands[2]); |
8cd42699 | 10399 | |
10400 | if (n != 0 || REGNO (operands[0]) != REGNO (operands[1])) | |
10401 | { | |
10402 | if (n < 0) | |
10403 | output_multi_immediate (operands, | |
f6b01912 | 10404 | "sub%?\t%0, %1, %2", "sub%?\t%0, %0, %2", 2, |
10405 | -n); | |
8cd42699 | 10406 | else |
10407 | output_multi_immediate (operands, | |
f6b01912 | 10408 | "add%?\t%0, %1, %2", "add%?\t%0, %0, %2", 2, |
10409 | n); | |
8cd42699 | 10410 | } |
a690ede2 | 10411 | |
10412 | return ""; | |
10413 | } | |
8cd42699 | 10414 | |
8cd42699 | 10415 | /* Output a multiple immediate operation. |
10416 | OPERANDS is the vector of operands referred to in the output patterns. | |
10417 | INSTR1 is the output pattern to use for the first constant. | |
10418 | INSTR2 is the output pattern to use for subsequent constants. | |
10419 | IMMED_OP is the index of the constant slot in OPERANDS. | |
10420 | N is the constant value. */ | |
0bd59681 | 10421 | static const char * |
ebd88a36 | 10422 | output_multi_immediate (rtx *operands, const char *instr1, const char *instr2, |
10423 | int immed_op, HOST_WIDE_INT n) | |
8cd42699 | 10424 | { |
a690ede2 | 10425 | #if HOST_BITS_PER_WIDE_INT > 32 |
35cea10c | 10426 | n &= 0xffffffff; |
a690ede2 | 10427 | #endif |
10428 | ||
8cd42699 | 10429 | if (n == 0) |
10430 | { | |
a12be13c | 10431 | /* Quick and easy output. */ |
8cd42699 | 10432 | operands[immed_op] = const0_rtx; |
a12be13c | 10433 | output_asm_insn (instr1, operands); |
8cd42699 | 10434 | } |
10435 | else | |
10436 | { | |
10437 | int i; | |
0bd59681 | 10438 | const char * instr = instr1; |
8cd42699 | 10439 | |
66aacf41 | 10440 | /* Note that n is never zero here (which would give no output). */ |
8cd42699 | 10441 | for (i = 0; i < 32; i += 2) |
10442 | { | |
10443 | if (n & (3 << i)) | |
10444 | { | |
a690ede2 | 10445 | operands[immed_op] = GEN_INT (n & (255 << i)); |
10446 | output_asm_insn (instr, operands); | |
8cd42699 | 10447 | instr = instr2; |
10448 | i += 6; | |
10449 | } | |
10450 | } | |
10451 | } | |
9e7454d0 | 10452 | |
a690ede2 | 10453 | return ""; |
f6b01912 | 10454 | } |
8cd42699 | 10455 | |
25f905c2 | 10456 | /* Return the name of a shifter operation. */ |
10457 | static const char * | |
10458 | arm_shift_nmem(enum rtx_code code) | |
10459 | { | |
10460 | switch (code) | |
10461 | { | |
10462 | case ASHIFT: | |
10463 | return ARM_LSL_NAME; | |
10464 | ||
10465 | case ASHIFTRT: | |
10466 | return "asr"; | |
10467 | ||
10468 | case LSHIFTRT: | |
10469 | return "lsr"; | |
10470 | ||
10471 | case ROTATERT: | |
10472 | return "ror"; | |
10473 | ||
10474 | default: | |
10475 | abort(); | |
10476 | } | |
10477 | } | |
10478 | ||
8cd42699 | 10479 | /* Return the appropriate ARM instruction for the operation code. |
10480 | The returned result should not be overwritten. OP is the rtx of the | |
10481 | operation. SHIFT_FIRST_ARG is TRUE if the first argument of the operator | |
10482 | was shifted. */ | |
0bd59681 | 10483 | const char * |
ebd88a36 | 10484 | arithmetic_instr (rtx op, int shift_first_arg) |
8cd42699 | 10485 | { |
f6b01912 | 10486 | switch (GET_CODE (op)) |
8cd42699 | 10487 | { |
10488 | case PLUS: | |
a690ede2 | 10489 | return "add"; |
10490 | ||
8cd42699 | 10491 | case MINUS: |
a690ede2 | 10492 | return shift_first_arg ? "rsb" : "sub"; |
10493 | ||
8cd42699 | 10494 | case IOR: |
a690ede2 | 10495 | return "orr"; |
10496 | ||
8cd42699 | 10497 | case XOR: |
a690ede2 | 10498 | return "eor"; |
10499 | ||
8cd42699 | 10500 | case AND: |
a690ede2 | 10501 | return "and"; |
10502 | ||
25f905c2 | 10503 | case ASHIFT: |
10504 | case ASHIFTRT: | |
10505 | case LSHIFTRT: | |
10506 | case ROTATERT: | |
10507 | return arm_shift_nmem(GET_CODE(op)); | |
10508 | ||
8cd42699 | 10509 | default: |
ed29c566 | 10510 | gcc_unreachable (); |
8cd42699 | 10511 | } |
a690ede2 | 10512 | } |
8cd42699 | 10513 | |
8cd42699 | 10514 | /* Ensure valid constant shifts and return the appropriate shift mnemonic |
10515 | for the operation code. The returned result should not be overwritten. | |
10516 | OP is the rtx code of the shift. | |
f6b01912 | 10517 | On exit, *AMOUNTP will be -1 if the shift is by a register, or a constant |
66aacf41 | 10518 | shift. */ |
0bd59681 | 10519 | static const char * |
ebd88a36 | 10520 | shift_op (rtx op, HOST_WIDE_INT *amountp) |
8cd42699 | 10521 | { |
0bd59681 | 10522 | const char * mnem; |
4b6897e9 | 10523 | enum rtx_code code = GET_CODE (op); |
8cd42699 | 10524 | |
ed29c566 | 10525 | switch (GET_CODE (XEXP (op, 1))) |
10526 | { | |
10527 | case REG: | |
10528 | case SUBREG: | |
10529 | *amountp = -1; | |
10530 | break; | |
10531 | ||
10532 | case CONST_INT: | |
10533 | *amountp = INTVAL (XEXP (op, 1)); | |
10534 | break; | |
10535 | ||
10536 | default: | |
10537 | gcc_unreachable (); | |
10538 | } | |
f6b01912 | 10539 | |
4b6897e9 | 10540 | switch (code) |
8cd42699 | 10541 | { |
9c9db025 | 10542 | case ROTATE: |
ed29c566 | 10543 | gcc_assert (*amountp != -1); |
9c9db025 | 10544 | *amountp = 32 - *amountp; |
25f905c2 | 10545 | code = ROTATERT; |
9c9db025 | 10546 | |
10547 | /* Fall through. */ | |
10548 | ||
25f905c2 | 10549 | case ASHIFT: |
10550 | case ASHIFTRT: | |
10551 | case LSHIFTRT: | |
f6b01912 | 10552 | case ROTATERT: |
25f905c2 | 10553 | mnem = arm_shift_nmem(code); |
f6b01912 | 10554 | break; |
10555 | ||
9c08d1fa | 10556 | case MULT: |
4b6897e9 | 10557 | /* We never have to worry about the amount being other than a |
10558 | power of 2, since this case can never be reloaded from a reg. */ | |
ed29c566 | 10559 | gcc_assert (*amountp != -1); |
10560 | *amountp = int_log2 (*amountp); | |
25f905c2 | 10561 | return ARM_LSL_NAME; |
a690ede2 | 10562 | |
8cd42699 | 10563 | default: |
ed29c566 | 10564 | gcc_unreachable (); |
8cd42699 | 10565 | } |
10566 | ||
4b6897e9 | 10567 | if (*amountp != -1) |
10568 | { | |
10569 | /* This is not 100% correct, but follows from the desire to merge | |
10570 | multiplication by a power of 2 with the recognizer for a | |
25f905c2 | 10571 | shift. >=32 is not a valid shift for "lsl", so we must try and |
4b6897e9 | 10572 | output a shift that produces the correct arithmetical result. |
01cc3b75 | 10573 | Using lsr #32 is identical except for the fact that the carry bit |
9e7454d0 | 10574 | is not set correctly if we set the flags; but we never use the |
4b6897e9 | 10575 | carry bit from such an operation, so we can ignore that. */ |
10576 | if (code == ROTATERT) | |
a12be13c | 10577 | /* Rotate is just modulo 32. */ |
10578 | *amountp &= 31; | |
4b6897e9 | 10579 | else if (*amountp != (*amountp & 31)) |
10580 | { | |
10581 | if (code == ASHIFT) | |
10582 | mnem = "lsr"; | |
10583 | *amountp = 32; | |
10584 | } | |
10585 | ||
10586 | /* Shifts of 0 are no-ops. */ | |
10587 | if (*amountp == 0) | |
10588 | return NULL; | |
9e7454d0 | 10589 | } |
4b6897e9 | 10590 | |
f6b01912 | 10591 | return mnem; |
10592 | } | |
8cd42699 | 10593 | |
66aacf41 | 10594 | /* Obtain the shift from the POWER of two. */ |
a12be13c | 10595 | |
dca34a8d | 10596 | static HOST_WIDE_INT |
ebd88a36 | 10597 | int_log2 (HOST_WIDE_INT power) |
8cd42699 | 10598 | { |
a690ede2 | 10599 | HOST_WIDE_INT shift = 0; |
8cd42699 | 10600 | |
35cea10c | 10601 | while ((((HOST_WIDE_INT) 1 << shift) & power) == 0) |
8cd42699 | 10602 | { |
ed29c566 | 10603 | gcc_assert (shift <= 31); |
ebd88a36 | 10604 | shift++; |
8cd42699 | 10605 | } |
a690ede2 | 10606 | |
10607 | return shift; | |
10608 | } | |
8cd42699 | 10609 | |
624d37a6 | 10610 | /* Output a .ascii pseudo-op, keeping track of lengths. This is |
10611 | because /bin/as is horribly restrictive. The judgement about | |
10612 | whether or not each character is 'printable' (and can be output as | |
10613 | is) or not (and must be printed with an octal escape) must be made | |
10614 | with reference to the *host* character set -- the situation is | |
10615 | similar to that discussed in the comments above pp_c_char in | |
10616 | c-pretty-print.c. */ | |
10617 | ||
96576951 | 10618 | #define MAX_ASCII_LEN 51 |
8cd42699 | 10619 | |
10620 | void | |
ebd88a36 | 10621 | output_ascii_pseudo_op (FILE *stream, const unsigned char *p, int len) |
8cd42699 | 10622 | { |
10623 | int i; | |
96576951 | 10624 | int len_so_far = 0; |
8cd42699 | 10625 | |
96576951 | 10626 | fputs ("\t.ascii\t\"", stream); |
9e7454d0 | 10627 | |
8cd42699 | 10628 | for (i = 0; i < len; i++) |
10629 | { | |
a12be13c | 10630 | int c = p[i]; |
8cd42699 | 10631 | |
96576951 | 10632 | if (len_so_far >= MAX_ASCII_LEN) |
8cd42699 | 10633 | { |
96576951 | 10634 | fputs ("\"\n\t.ascii\t\"", stream); |
8cd42699 | 10635 | len_so_far = 0; |
8cd42699 | 10636 | } |
10637 | ||
624d37a6 | 10638 | if (ISPRINT (c)) |
8cd42699 | 10639 | { |
624d37a6 | 10640 | if (c == '\\' || c == '\"') |
96576951 | 10641 | { |
624d37a6 | 10642 | putc ('\\', stream); |
215b30b3 | 10643 | len_so_far++; |
96576951 | 10644 | } |
624d37a6 | 10645 | putc (c, stream); |
10646 | len_so_far++; | |
10647 | } | |
10648 | else | |
10649 | { | |
10650 | fprintf (stream, "\\%03o", c); | |
10651 | len_so_far += 4; | |
8cd42699 | 10652 | } |
8cd42699 | 10653 | } |
a690ede2 | 10654 | |
8cd42699 | 10655 | fputs ("\"\n", stream); |
a690ede2 | 10656 | } |
8cd42699 | 10657 | \f |
4c44712e | 10658 | /* Compute the register save mask for registers 0 through 12 |
f9273c43 | 10659 | inclusive. This code is used by arm_compute_save_reg_mask. */ |
53f179c0 | 10660 | |
e27ad2d5 | 10661 | static unsigned long |
ebd88a36 | 10662 | arm_compute_save_reg0_reg12_mask (void) |
e27ad2d5 | 10663 | { |
aa1ae31c | 10664 | unsigned long func_type = arm_current_func_type (); |
53f179c0 | 10665 | unsigned long save_reg_mask = 0; |
e27ad2d5 | 10666 | unsigned int reg; |
e27ad2d5 | 10667 | |
2d9065d3 | 10668 | if (IS_INTERRUPT (func_type)) |
e27ad2d5 | 10669 | { |
2d9065d3 | 10670 | unsigned int max_reg; |
2d9065d3 | 10671 | /* Interrupt functions must not corrupt any registers, |
10672 | even call clobbered ones. If this is a leaf function | |
10673 | we can just examine the registers used by the RTL, but | |
10674 | otherwise we have to assume that whatever function is | |
10675 | called might clobber anything, and so we have to save | |
10676 | all the call-clobbered registers as well. */ | |
10677 | if (ARM_FUNC_TYPE (func_type) == ARM_FT_FIQ) | |
10678 | /* FIQ handlers have registers r8 - r12 banked, so | |
10679 | we only need to check r0 - r7, Normal ISRs only | |
aa1ae31c | 10680 | bank r14 and r15, so we must check up to r12. |
2d9065d3 | 10681 | r13 is the stack pointer which is always preserved, |
10682 | so we do not need to consider it here. */ | |
10683 | max_reg = 7; | |
10684 | else | |
10685 | max_reg = 12; | |
9e7454d0 | 10686 | |
2d9065d3 | 10687 | for (reg = 0; reg <= max_reg; reg++) |
3072d30e | 10688 | if (df_regs_ever_live_p (reg) |
10689 | || (! current_function_is_leaf && call_used_regs[reg])) | |
e27ad2d5 | 10690 | save_reg_mask |= (1 << reg); |
37d5afa3 | 10691 | |
f933d49b | 10692 | /* Also save the pic base register if necessary. */ |
37d5afa3 | 10693 | if (flag_pic |
10694 | && !TARGET_SINGLE_PIC_BASE | |
2cb7d577 | 10695 | && arm_pic_register != INVALID_REGNUM |
37d5afa3 | 10696 | && current_function_uses_pic_offset_table) |
10697 | save_reg_mask |= 1 << PIC_OFFSET_TABLE_REGNUM; | |
e27ad2d5 | 10698 | } |
10699 | else | |
10700 | { | |
25f905c2 | 10701 | /* In arm mode we handle r11 (FP) as a special case. */ |
10702 | unsigned last_reg = TARGET_ARM ? 10 : 11; | |
10703 | ||
e27ad2d5 | 10704 | /* In the normal case we only need to save those registers |
10705 | which are call saved and which are used by this function. */ | |
25f905c2 | 10706 | for (reg = 0; reg <= last_reg; reg++) |
3072d30e | 10707 | if (df_regs_ever_live_p (reg) && ! call_used_regs[reg]) |
e27ad2d5 | 10708 | save_reg_mask |= (1 << reg); |
10709 | ||
10710 | /* Handle the frame pointer as a special case. */ | |
3072d30e | 10711 | if (! TARGET_APCS_FRAME |
10712 | && ! frame_pointer_needed | |
10713 | && df_regs_ever_live_p (HARD_FRAME_POINTER_REGNUM) | |
10714 | && ! call_used_regs[HARD_FRAME_POINTER_REGNUM]) | |
25f905c2 | 10715 | save_reg_mask |= 1 << HARD_FRAME_POINTER_REGNUM; |
10716 | else if (! TARGET_APCS_FRAME | |
10717 | && ! frame_pointer_needed | |
3072d30e | 10718 | && df_regs_ever_live_p (HARD_FRAME_POINTER_REGNUM) |
25f905c2 | 10719 | && ! call_used_regs[HARD_FRAME_POINTER_REGNUM]) |
e27ad2d5 | 10720 | save_reg_mask |= 1 << HARD_FRAME_POINTER_REGNUM; |
10721 | ||
10722 | /* If we aren't loading the PIC register, | |
10723 | don't stack it even though it may be live. */ | |
10724 | if (flag_pic | |
a60f1840 | 10725 | && !TARGET_SINGLE_PIC_BASE |
2cb7d577 | 10726 | && arm_pic_register != INVALID_REGNUM |
3072d30e | 10727 | && (df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM) |
37d5afa3 | 10728 | || current_function_uses_pic_offset_table)) |
e27ad2d5 | 10729 | save_reg_mask |= 1 << PIC_OFFSET_TABLE_REGNUM; |
25f905c2 | 10730 | |
10731 | /* The prologue will copy SP into R0, so save it. */ | |
10732 | if (IS_STACKALIGN (func_type)) | |
10733 | save_reg_mask |= 1; | |
e27ad2d5 | 10734 | } |
10735 | ||
4c44712e | 10736 | /* Save registers so the exception handler can modify them. */ |
10737 | if (current_function_calls_eh_return) | |
10738 | { | |
10739 | unsigned int i; | |
9e7454d0 | 10740 | |
4c44712e | 10741 | for (i = 0; ; i++) |
10742 | { | |
10743 | reg = EH_RETURN_DATA_REGNO (i); | |
10744 | if (reg == INVALID_REGNUM) | |
10745 | break; | |
10746 | save_reg_mask |= 1 << reg; | |
10747 | } | |
10748 | } | |
10749 | ||
aa1ae31c | 10750 | return save_reg_mask; |
10751 | } | |
10752 | ||
25f905c2 | 10753 | |
aa1ae31c | 10754 | /* Compute a bit mask of which registers need to be |
10755 | saved on the stack for the current function. */ | |
10756 | ||
10757 | static unsigned long | |
ebd88a36 | 10758 | arm_compute_save_reg_mask (void) |
aa1ae31c | 10759 | { |
10760 | unsigned int save_reg_mask = 0; | |
10761 | unsigned long func_type = arm_current_func_type (); | |
25f905c2 | 10762 | unsigned int reg; |
aa1ae31c | 10763 | |
10764 | if (IS_NAKED (func_type)) | |
10765 | /* This should never really happen. */ | |
10766 | return 0; | |
10767 | ||
10768 | /* If we are creating a stack frame, then we must save the frame pointer, | |
10769 | IP (which will hold the old stack pointer), LR and the PC. */ | |
25f905c2 | 10770 | if (frame_pointer_needed && TARGET_ARM) |
aa1ae31c | 10771 | save_reg_mask |= |
10772 | (1 << ARM_HARD_FRAME_POINTER_REGNUM) | |
10773 | | (1 << IP_REGNUM) | |
10774 | | (1 << LR_REGNUM) | |
10775 | | (1 << PC_REGNUM); | |
10776 | ||
10777 | /* Volatile functions do not return, so there | |
10778 | is no need to save any other registers. */ | |
10779 | if (IS_VOLATILE (func_type)) | |
10780 | return save_reg_mask; | |
10781 | ||
10782 | save_reg_mask |= arm_compute_save_reg0_reg12_mask (); | |
10783 | ||
e27ad2d5 | 10784 | /* Decide if we need to save the link register. |
10785 | Interrupt routines have their own banked link register, | |
10786 | so they never need to save it. | |
71d934e9 | 10787 | Otherwise if we do not use the link register we do not need to save |
e27ad2d5 | 10788 | it. If we are pushing other registers onto the stack however, we |
10789 | can save an instruction in the epilogue by pushing the link register | |
10790 | now and then popping it back into the PC. This incurs extra memory | |
5bd751ff | 10791 | accesses though, so we only do it when optimizing for size, and only |
e27ad2d5 | 10792 | if we know that we will not need a fancy return sequence. */ |
3072d30e | 10793 | if (df_regs_ever_live_p (LR_REGNUM) |
10794 | || (save_reg_mask | |
10795 | && optimize_size | |
10796 | && ARM_FUNC_TYPE (func_type) == ARM_FT_NORMAL | |
10797 | && !current_function_calls_eh_return)) | |
e27ad2d5 | 10798 | save_reg_mask |= 1 << LR_REGNUM; |
10799 | ||
89b30d75 | 10800 | if (cfun->machine->lr_save_eliminated) |
10801 | save_reg_mask &= ~ (1 << LR_REGNUM); | |
10802 | ||
755eb2b4 | 10803 | if (TARGET_REALLY_IWMMXT |
10804 | && ((bit_count (save_reg_mask) | |
10805 | + ARM_NUM_INTS (current_function_pretend_args_size)) % 2) != 0) | |
10806 | { | |
755eb2b4 | 10807 | /* The total number of registers that are going to be pushed |
10808 | onto the stack is odd. We need to ensure that the stack | |
10809 | is 64-bit aligned before we start to save iWMMXt registers, | |
10810 | and also before we start to create locals. (A local variable | |
10811 | might be a double or long long which we will load/store using | |
10812 | an iWMMXt instruction). Therefore we need to push another | |
10813 | ARM register, so that the stack will be 64-bit aligned. We | |
10814 | try to avoid using the arg registers (r0 -r3) as they might be | |
10815 | used to pass values in a tail call. */ | |
10816 | for (reg = 4; reg <= 12; reg++) | |
10817 | if ((save_reg_mask & (1 << reg)) == 0) | |
10818 | break; | |
10819 | ||
10820 | if (reg <= 12) | |
10821 | save_reg_mask |= (1 << reg); | |
10822 | else | |
10823 | { | |
10824 | cfun->machine->sibcall_blocked = 1; | |
10825 | save_reg_mask |= (1 << 3); | |
10826 | } | |
10827 | } | |
10828 | ||
25f905c2 | 10829 | /* We may need to push an additional register for use initializing the |
10830 | PIC base register. */ | |
10831 | if (TARGET_THUMB2 && IS_NESTED (func_type) && flag_pic | |
10832 | && (save_reg_mask & THUMB2_WORK_REGS) == 0) | |
10833 | { | |
10834 | reg = thumb_find_work_register (1 << 4); | |
10835 | if (!call_used_regs[reg]) | |
10836 | save_reg_mask |= (1 << reg); | |
10837 | } | |
10838 | ||
e27ad2d5 | 10839 | return save_reg_mask; |
10840 | } | |
10841 | ||
7b78e739 | 10842 | |
237533cc | 10843 | /* Compute a bit mask of which registers need to be |
10844 | saved on the stack for the current function. */ | |
10845 | static unsigned long | |
25f905c2 | 10846 | thumb1_compute_save_reg_mask (void) |
237533cc | 10847 | { |
10848 | unsigned long mask; | |
53f179c0 | 10849 | unsigned reg; |
237533cc | 10850 | |
10851 | mask = 0; | |
10852 | for (reg = 0; reg < 12; reg ++) | |
3072d30e | 10853 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
53f179c0 | 10854 | mask |= 1 << reg; |
237533cc | 10855 | |
bd0b3b57 | 10856 | if (flag_pic |
10857 | && !TARGET_SINGLE_PIC_BASE | |
2cb7d577 | 10858 | && arm_pic_register != INVALID_REGNUM |
bd0b3b57 | 10859 | && current_function_uses_pic_offset_table) |
10860 | mask |= 1 << PIC_OFFSET_TABLE_REGNUM; | |
53f179c0 | 10861 | |
150502c9 | 10862 | /* See if we might need r11 for calls to _interwork_r11_call_via_rN(). */ |
10863 | if (!frame_pointer_needed && CALLER_INTERWORKING_SLOT_SIZE > 0) | |
10864 | mask |= 1 << ARM_HARD_FRAME_POINTER_REGNUM; | |
237533cc | 10865 | |
53f179c0 | 10866 | /* LR will also be pushed if any lo regs are pushed. */ |
237533cc | 10867 | if (mask & 0xff || thumb_force_lr_save ()) |
10868 | mask |= (1 << LR_REGNUM); | |
10869 | ||
53f179c0 | 10870 | /* Make sure we have a low work register if we need one. |
10871 | We will need one if we are going to push a high register, | |
10872 | but we are not currently intending to push a low register. */ | |
10873 | if ((mask & 0xff) == 0 | |
237533cc | 10874 | && ((mask & 0x0f00) || TARGET_BACKTRACE)) |
53f179c0 | 10875 | { |
10876 | /* Use thumb_find_work_register to choose which register | |
10877 | we will use. If the register is live then we will | |
10878 | have to push it. Use LAST_LO_REGNUM as our fallback | |
10879 | choice for the register to select. */ | |
10880 | reg = thumb_find_work_register (1 << LAST_LO_REGNUM); | |
c550a00b | 10881 | /* Make sure the register returned by thumb_find_work_register is |
10882 | not part of the return value. */ | |
10883 | if (reg * UNITS_PER_WORD <= arm_size_return_regs ()) | |
10884 | reg = LAST_LO_REGNUM; | |
53f179c0 | 10885 | |
10886 | if (! call_used_regs[reg]) | |
10887 | mask |= 1 << reg; | |
10888 | } | |
237533cc | 10889 | |
10890 | return mask; | |
10891 | } | |
10892 | ||
10893 | ||
7b78e739 | 10894 | /* Return the number of bytes required to save VFP registers. */ |
10895 | static int | |
10896 | arm_get_vfp_saved_size (void) | |
10897 | { | |
10898 | unsigned int regno; | |
10899 | int count; | |
10900 | int saved; | |
10901 | ||
10902 | saved = 0; | |
10903 | /* Space for saved VFP registers. */ | |
10904 | if (TARGET_HARD_FLOAT && TARGET_VFP) | |
10905 | { | |
10906 | count = 0; | |
10907 | for (regno = FIRST_VFP_REGNUM; | |
10908 | regno < LAST_VFP_REGNUM; | |
10909 | regno += 2) | |
10910 | { | |
3072d30e | 10911 | if ((!df_regs_ever_live_p (regno) || call_used_regs[regno]) |
10912 | && (!df_regs_ever_live_p (regno + 1) || call_used_regs[regno + 1])) | |
7b78e739 | 10913 | { |
10914 | if (count > 0) | |
10915 | { | |
10916 | /* Workaround ARM10 VFPr1 bug. */ | |
10917 | if (count == 2 && !arm_arch6) | |
10918 | count++; | |
382a2a57 | 10919 | saved += count * 8; |
7b78e739 | 10920 | } |
10921 | count = 0; | |
10922 | } | |
10923 | else | |
10924 | count++; | |
10925 | } | |
10926 | if (count > 0) | |
10927 | { | |
10928 | if (count == 2 && !arm_arch6) | |
10929 | count++; | |
382a2a57 | 10930 | saved += count * 8; |
7b78e739 | 10931 | } |
10932 | } | |
10933 | return saved; | |
10934 | } | |
10935 | ||
10936 | ||
8d69f2ad | 10937 | /* Generate a function exit sequence. If REALLY_RETURN is false, then do |
e27ad2d5 | 10938 | everything bar the final return instruction. */ |
0bd59681 | 10939 | const char * |
ebd88a36 | 10940 | output_return_instruction (rtx operand, int really_return, int reverse) |
9c08d1fa | 10941 | { |
e27ad2d5 | 10942 | char conditional[10]; |
9c08d1fa | 10943 | char instr[100]; |
53f179c0 | 10944 | unsigned reg; |
e27ad2d5 | 10945 | unsigned long live_regs_mask; |
10946 | unsigned long func_type; | |
f9273c43 | 10947 | arm_stack_offsets *offsets; |
9f468b67 | 10948 | |
e27ad2d5 | 10949 | func_type = arm_current_func_type (); |
4b6897e9 | 10950 | |
e27ad2d5 | 10951 | if (IS_NAKED (func_type)) |
cffb2a26 | 10952 | return ""; |
e27ad2d5 | 10953 | |
10954 | if (IS_VOLATILE (func_type) && TARGET_ABORT_NORETURN) | |
4b6897e9 | 10955 | { |
8d69f2ad | 10956 | /* If this function was declared non-returning, and we have |
10957 | found a tail call, then we have to trust that the called | |
10958 | function won't return. */ | |
ed593f11 | 10959 | if (really_return) |
10960 | { | |
10961 | rtx ops[2]; | |
9e7454d0 | 10962 | |
ed593f11 | 10963 | /* Otherwise, trap an attempted return by aborting. */ |
10964 | ops[0] = operand; | |
9e7454d0 | 10965 | ops[1] = gen_rtx_SYMBOL_REF (Pmode, NEED_PLT_RELOC ? "abort(PLT)" |
ed593f11 | 10966 | : "abort"); |
10967 | assemble_external_libcall (ops[1]); | |
10968 | output_asm_insn (reverse ? "bl%D0\t%a1" : "bl%d0\t%a1", ops); | |
10969 | } | |
9e7454d0 | 10970 | |
4b6897e9 | 10971 | return ""; |
10972 | } | |
e27ad2d5 | 10973 | |
ed29c566 | 10974 | gcc_assert (!current_function_calls_alloca || really_return); |
9c08d1fa | 10975 | |
11b07ceb | 10976 | sprintf (conditional, "%%?%%%c0", reverse ? 'D' : 'd'); |
cffb2a26 | 10977 | |
e27ad2d5 | 10978 | return_used_this_function = 1; |
9c08d1fa | 10979 | |
e27ad2d5 | 10980 | live_regs_mask = arm_compute_save_reg_mask (); |
9c08d1fa | 10981 | |
71d934e9 | 10982 | if (live_regs_mask) |
e27ad2d5 | 10983 | { |
71d934e9 | 10984 | const char * return_reg; |
10985 | ||
9e7454d0 | 10986 | /* If we do not have any special requirements for function exit |
8b438b4f | 10987 | (e.g. interworking) then we can load the return address |
71d934e9 | 10988 | directly into the PC. Otherwise we must load it into LR. */ |
10989 | if (really_return | |
8b438b4f | 10990 | && (IS_INTERRUPT (func_type) || !TARGET_INTERWORK)) |
71d934e9 | 10991 | return_reg = reg_names[PC_REGNUM]; |
e27ad2d5 | 10992 | else |
71d934e9 | 10993 | return_reg = reg_names[LR_REGNUM]; |
10994 | ||
e27ad2d5 | 10995 | if ((live_regs_mask & (1 << IP_REGNUM)) == (1 << IP_REGNUM)) |
fa3ae0b9 | 10996 | { |
10997 | /* There are three possible reasons for the IP register | |
10998 | being saved. 1) a stack frame was created, in which case | |
10999 | IP contains the old stack pointer, or 2) an ISR routine | |
11000 | corrupted it, or 3) it was saved to align the stack on | |
11001 | iWMMXt. In case 1, restore IP into SP, otherwise just | |
11002 | restore IP. */ | |
11003 | if (frame_pointer_needed) | |
11004 | { | |
11005 | live_regs_mask &= ~ (1 << IP_REGNUM); | |
11006 | live_regs_mask |= (1 << SP_REGNUM); | |
11007 | } | |
11008 | else | |
ed29c566 | 11009 | gcc_assert (IS_INTERRUPT (func_type) || TARGET_REALLY_IWMMXT); |
fa3ae0b9 | 11010 | } |
a690ede2 | 11011 | |
4a5f62e2 | 11012 | /* On some ARM architectures it is faster to use LDR rather than |
11013 | LDM to load a single register. On other architectures, the | |
11014 | cost is the same. In 26 bit mode, or for exception handlers, | |
11015 | we have to use LDM to load the PC so that the CPSR is also | |
11016 | restored. */ | |
11017 | for (reg = 0; reg <= LAST_ARM_REGNUM; reg++) | |
53f179c0 | 11018 | if (live_regs_mask == (1U << reg)) |
11019 | break; | |
11020 | ||
4a5f62e2 | 11021 | if (reg <= LAST_ARM_REGNUM |
11022 | && (reg != LR_REGNUM | |
9e7454d0 | 11023 | || ! really_return |
c1a66faf | 11024 | || ! IS_INTERRUPT (func_type))) |
4a5f62e2 | 11025 | { |
9e7454d0 | 11026 | sprintf (instr, "ldr%s\t%%|%s, [%%|sp], #4", conditional, |
4a5f62e2 | 11027 | (reg == LR_REGNUM) ? return_reg : reg_names[reg]); |
e27ad2d5 | 11028 | } |
9c08d1fa | 11029 | else |
39d5d2b6 | 11030 | { |
71d934e9 | 11031 | char *p; |
11032 | int first = 1; | |
e27ad2d5 | 11033 | |
8d69f2ad | 11034 | /* Generate the load multiple instruction to restore the |
11035 | registers. Note we can get here, even if | |
11036 | frame_pointer_needed is true, but only if sp already | |
11037 | points to the base of the saved core registers. */ | |
11038 | if (live_regs_mask & (1 << SP_REGNUM)) | |
ffc9d00c | 11039 | { |
f9273c43 | 11040 | unsigned HOST_WIDE_INT stack_adjust; |
11041 | ||
11042 | offsets = arm_get_frame_offsets (); | |
11043 | stack_adjust = offsets->outgoing_args - offsets->saved_regs; | |
ed29c566 | 11044 | gcc_assert (stack_adjust == 0 || stack_adjust == 4); |
ffc9d00c | 11045 | |
25f905c2 | 11046 | if (stack_adjust && arm_arch5 && TARGET_ARM) |
ffc9d00c | 11047 | sprintf (instr, "ldm%sib\t%%|sp, {", conditional); |
11048 | else | |
11049 | { | |
53f179c0 | 11050 | /* If we can't use ldmib (SA110 bug), |
11051 | then try to pop r3 instead. */ | |
ffc9d00c | 11052 | if (stack_adjust) |
11053 | live_regs_mask |= 1 << 3; | |
11054 | sprintf (instr, "ldm%sfd\t%%|sp, {", conditional); | |
11055 | } | |
11056 | } | |
99ecc60d | 11057 | else |
71d934e9 | 11058 | sprintf (instr, "ldm%sfd\t%%|sp!, {", conditional); |
11059 | ||
11060 | p = instr + strlen (instr); | |
e27ad2d5 | 11061 | |
71d934e9 | 11062 | for (reg = 0; reg <= SP_REGNUM; reg++) |
11063 | if (live_regs_mask & (1 << reg)) | |
11064 | { | |
11065 | int l = strlen (reg_names[reg]); | |
11066 | ||
11067 | if (first) | |
11068 | first = 0; | |
11069 | else | |
11070 | { | |
11071 | memcpy (p, ", ", 2); | |
11072 | p += 2; | |
11073 | } | |
11074 | ||
11075 | memcpy (p, "%|", 2); | |
11076 | memcpy (p + 2, reg_names[reg], l); | |
11077 | p += l + 2; | |
11078 | } | |
9e7454d0 | 11079 | |
71d934e9 | 11080 | if (live_regs_mask & (1 << LR_REGNUM)) |
11081 | { | |
f29bf0d0 | 11082 | sprintf (p, "%s%%|%s}", first ? "" : ", ", return_reg); |
c1a66faf | 11083 | /* If returning from an interrupt, restore the CPSR. */ |
11084 | if (IS_INTERRUPT (func_type)) | |
f29bf0d0 | 11085 | strcat (p, "^"); |
71d934e9 | 11086 | } |
11087 | else | |
11088 | strcpy (p, "}"); | |
39d5d2b6 | 11089 | } |
99ecc60d | 11090 | |
71d934e9 | 11091 | output_asm_insn (instr, & operand); |
11092 | ||
4a5f62e2 | 11093 | /* See if we need to generate an extra instruction to |
11094 | perform the actual function return. */ | |
11095 | if (really_return | |
11096 | && func_type != ARM_FT_INTERWORKED | |
11097 | && (live_regs_mask & (1 << LR_REGNUM)) != 0) | |
99ecc60d | 11098 | { |
4a5f62e2 | 11099 | /* The return has already been handled |
11100 | by loading the LR into the PC. */ | |
11101 | really_return = 0; | |
99ecc60d | 11102 | } |
9c08d1fa | 11103 | } |
9f468b67 | 11104 | |
71d934e9 | 11105 | if (really_return) |
9c08d1fa | 11106 | { |
e27ad2d5 | 11107 | switch ((int) ARM_FUNC_TYPE (func_type)) |
11108 | { | |
11109 | case ARM_FT_ISR: | |
11110 | case ARM_FT_FIQ: | |
25f905c2 | 11111 | /* ??? This is wrong for unified assembly syntax. */ |
e27ad2d5 | 11112 | sprintf (instr, "sub%ss\t%%|pc, %%|lr, #4", conditional); |
11113 | break; | |
11114 | ||
11115 | case ARM_FT_INTERWORKED: | |
11116 | sprintf (instr, "bx%s\t%%|lr", conditional); | |
11117 | break; | |
11118 | ||
11119 | case ARM_FT_EXCEPTION: | |
25f905c2 | 11120 | /* ??? This is wrong for unified assembly syntax. */ |
e27ad2d5 | 11121 | sprintf (instr, "mov%ss\t%%|pc, %%|lr", conditional); |
11122 | break; | |
11123 | ||
11124 | default: | |
f1039640 | 11125 | /* Use bx if it's available. */ |
11126 | if (arm_arch5 || arm_arch4t) | |
9e7454d0 | 11127 | sprintf (instr, "bx%s\t%%|lr", conditional); |
71d934e9 | 11128 | else |
c1a66faf | 11129 | sprintf (instr, "mov%s\t%%|pc, %%|lr", conditional); |
e27ad2d5 | 11130 | break; |
11131 | } | |
71d934e9 | 11132 | |
11133 | output_asm_insn (instr, & operand); | |
9c08d1fa | 11134 | } |
a690ede2 | 11135 | |
9c08d1fa | 11136 | return ""; |
11137 | } | |
11138 | ||
c00597d1 | 11139 | /* Write the function name into the code section, directly preceding |
11140 | the function prologue. | |
11141 | ||
11142 | Code will be output similar to this: | |
11143 | t0 | |
11144 | .ascii "arm_poke_function_name", 0 | |
11145 | .align | |
11146 | t1 | |
11147 | .word 0xff000000 + (t1 - t0) | |
11148 | arm_poke_function_name | |
11149 | mov ip, sp | |
11150 | stmfd sp!, {fp, ip, lr, pc} | |
11151 | sub fp, ip, #4 | |
11152 | ||
11153 | When performing a stack backtrace, code can inspect the value | |
11154 | of 'pc' stored at 'fp' + 0. If the trace function then looks | |
11155 | at location pc - 12 and the top 8 bits are set, then we know | |
11156 | that there is a function name embedded immediately preceding this | |
11157 | location and has length ((pc[-3]) & 0xff000000). | |
11158 | ||
11159 | We assume that pc is declared as a pointer to an unsigned long. | |
11160 | ||
11161 | It is of no benefit to output the function name if we are assembling | |
11162 | a leaf function. These function types will not contain a stack | |
11163 | backtrace structure, therefore it is not possible to determine the | |
11164 | function name. */ | |
c00597d1 | 11165 | void |
ebd88a36 | 11166 | arm_poke_function_name (FILE *stream, const char *name) |
c00597d1 | 11167 | { |
11168 | unsigned long alignlength; | |
11169 | unsigned long length; | |
11170 | rtx x; | |
11171 | ||
cffb2a26 | 11172 | length = strlen (name) + 1; |
068e6db6 | 11173 | alignlength = ROUND_UP_WORD (length); |
9e7454d0 | 11174 | |
56d27660 | 11175 | ASM_OUTPUT_ASCII (stream, name, length); |
c00597d1 | 11176 | ASM_OUTPUT_ALIGN (stream, 2); |
35cea10c | 11177 | x = GEN_INT ((unsigned HOST_WIDE_INT) 0xff000000 + alignlength); |
58356836 | 11178 | assemble_aligned_integer (UNITS_PER_WORD, x); |
c00597d1 | 11179 | } |
11180 | ||
e27ad2d5 | 11181 | /* Place some comments into the assembler stream |
11182 | describing the current function. */ | |
17d9b0c3 | 11183 | static void |
ebd88a36 | 11184 | arm_output_function_prologue (FILE *f, HOST_WIDE_INT frame_size) |
8cd42699 | 11185 | { |
e27ad2d5 | 11186 | unsigned long func_type; |
17d9b0c3 | 11187 | |
25f905c2 | 11188 | if (TARGET_THUMB1) |
17d9b0c3 | 11189 | { |
25f905c2 | 11190 | thumb1_output_function_prologue (f, frame_size); |
17d9b0c3 | 11191 | return; |
11192 | } | |
9e7454d0 | 11193 | |
e27ad2d5 | 11194 | /* Sanity check. */ |
ed29c566 | 11195 | gcc_assert (!arm_ccfsm_state && !arm_target_insn); |
9011416b | 11196 | |
e27ad2d5 | 11197 | func_type = arm_current_func_type (); |
9e7454d0 | 11198 | |
e27ad2d5 | 11199 | switch ((int) ARM_FUNC_TYPE (func_type)) |
11200 | { | |
11201 | default: | |
11202 | case ARM_FT_NORMAL: | |
11203 | break; | |
11204 | case ARM_FT_INTERWORKED: | |
11205 | asm_fprintf (f, "\t%@ Function supports interworking.\n"); | |
11206 | break; | |
e27ad2d5 | 11207 | case ARM_FT_ISR: |
11208 | asm_fprintf (f, "\t%@ Interrupt Service Routine.\n"); | |
11209 | break; | |
11210 | case ARM_FT_FIQ: | |
11211 | asm_fprintf (f, "\t%@ Fast Interrupt Service Routine.\n"); | |
11212 | break; | |
11213 | case ARM_FT_EXCEPTION: | |
11214 | asm_fprintf (f, "\t%@ ARM Exception Handler.\n"); | |
11215 | break; | |
11216 | } | |
9e7454d0 | 11217 | |
e27ad2d5 | 11218 | if (IS_NAKED (func_type)) |
11219 | asm_fprintf (f, "\t%@ Naked Function: prologue and epilogue provided by programmer.\n"); | |
11220 | ||
11221 | if (IS_VOLATILE (func_type)) | |
11222 | asm_fprintf (f, "\t%@ Volatile: function does not return.\n"); | |
11223 | ||
11224 | if (IS_NESTED (func_type)) | |
11225 | asm_fprintf (f, "\t%@ Nested: function declared inside another function.\n"); | |
25f905c2 | 11226 | if (IS_STACKALIGN (func_type)) |
11227 | asm_fprintf (f, "\t%@ Stack Align: May be called with mis-aligned SP.\n"); | |
9e7454d0 | 11228 | |
4d1bb299 | 11229 | asm_fprintf (f, "\t%@ args = %d, pretend = %d, frame = %wd\n", |
caceeb43 | 11230 | current_function_args_size, |
11231 | current_function_pretend_args_size, frame_size); | |
e27ad2d5 | 11232 | |
71904594 | 11233 | asm_fprintf (f, "\t%@ frame_needed = %d, uses_anonymous_args = %d\n", |
caceeb43 | 11234 | frame_pointer_needed, |
71904594 | 11235 | cfun->machine->uses_anonymous_args); |
8cd42699 | 11236 | |
89b30d75 | 11237 | if (cfun->machine->lr_save_eliminated) |
11238 | asm_fprintf (f, "\t%@ link register save eliminated.\n"); | |
11239 | ||
4c44712e | 11240 | if (current_function_calls_eh_return) |
11241 | asm_fprintf (f, "\t@ Calls __builtin_eh_return.\n"); | |
11242 | ||
69120d26 | 11243 | #ifdef AOF_ASSEMBLER |
11244 | if (flag_pic) | |
caceeb43 | 11245 | asm_fprintf (f, "\tmov\t%r, %r\n", IP_REGNUM, PIC_OFFSET_TABLE_REGNUM); |
69120d26 | 11246 | #endif |
e27ad2d5 | 11247 | |
9e7454d0 | 11248 | return_used_this_function = 0; |
a690ede2 | 11249 | } |
8cd42699 | 11250 | |
0bd59681 | 11251 | const char * |
ffc9d00c | 11252 | arm_output_epilogue (rtx sibling) |
8cd42699 | 11253 | { |
56d27660 | 11254 | int reg; |
89b30d75 | 11255 | unsigned long saved_regs_mask; |
e27ad2d5 | 11256 | unsigned long func_type; |
9e7454d0 | 11257 | /* Floats_offset is the offset from the "virtual" frame. In an APCS |
73dd1c00 | 11258 | frame that is $fp + 4 for a non-variadic function. */ |
11259 | int floats_offset = 0; | |
8cd42699 | 11260 | rtx operands[3]; |
cffb2a26 | 11261 | FILE * f = asm_out_file; |
755eb2b4 | 11262 | unsigned int lrm_count = 0; |
ffc9d00c | 11263 | int really_return = (sibling == NULL); |
a2cd141b | 11264 | int start_reg; |
f9273c43 | 11265 | arm_stack_offsets *offsets; |
8cd42699 | 11266 | |
e27ad2d5 | 11267 | /* If we have already generated the return instruction |
11268 | then it is futile to generate anything else. */ | |
ffc9d00c | 11269 | if (use_return_insn (FALSE, sibling) && return_used_this_function) |
56d27660 | 11270 | return ""; |
8cd42699 | 11271 | |
e27ad2d5 | 11272 | func_type = arm_current_func_type (); |
cffb2a26 | 11273 | |
e27ad2d5 | 11274 | if (IS_NAKED (func_type)) |
11275 | /* Naked functions don't have epilogues. */ | |
11276 | return ""; | |
1c494086 | 11277 | |
e27ad2d5 | 11278 | if (IS_VOLATILE (func_type) && TARGET_ABORT_NORETURN) |
4b6897e9 | 11279 | { |
6ebaa29d | 11280 | rtx op; |
9e7454d0 | 11281 | |
e27ad2d5 | 11282 | /* A volatile function should never return. Call abort. */ |
55c1e470 | 11283 | op = gen_rtx_SYMBOL_REF (Pmode, NEED_PLT_RELOC ? "abort(PLT)" : "abort"); |
64266206 | 11284 | assemble_external_libcall (op); |
4b6897e9 | 11285 | output_asm_insn ("bl\t%a0", &op); |
9e7454d0 | 11286 | |
56d27660 | 11287 | return ""; |
4b6897e9 | 11288 | } |
11289 | ||
ed29c566 | 11290 | /* If we are throwing an exception, then we really must be doing a |
11291 | return, so we can't tail-call. */ | |
11292 | gcc_assert (!current_function_calls_eh_return || really_return); | |
9e7454d0 | 11293 | |
f9273c43 | 11294 | offsets = arm_get_frame_offsets (); |
89b30d75 | 11295 | saved_regs_mask = arm_compute_save_reg_mask (); |
755eb2b4 | 11296 | |
11297 | if (TARGET_IWMMXT) | |
11298 | lrm_count = bit_count (saved_regs_mask); | |
11299 | ||
f9273c43 | 11300 | floats_offset = offsets->saved_args; |
e27ad2d5 | 11301 | /* Compute how far away the floats will be. */ |
755eb2b4 | 11302 | for (reg = 0; reg <= LAST_ARM_REGNUM; reg++) |
89b30d75 | 11303 | if (saved_regs_mask & (1 << reg)) |
0eb4b5ff | 11304 | floats_offset += 4; |
9e7454d0 | 11305 | |
25f905c2 | 11306 | if (frame_pointer_needed && TARGET_ARM) |
8cd42699 | 11307 | { |
a2cd141b | 11308 | /* This variable is for the Virtual Frame Pointer, not VFP regs. */ |
f9273c43 | 11309 | int vfp_offset = offsets->frame; |
73dd1c00 | 11310 | |
c7f506fd | 11311 | if (arm_fpu_arch == FPUTYPE_FPA_EMU2) |
f21249d5 | 11312 | { |
a2cd141b | 11313 | for (reg = LAST_FPA_REGNUM; reg >= FIRST_FPA_REGNUM; reg--) |
3072d30e | 11314 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
f21249d5 | 11315 | { |
11316 | floats_offset += 12; | |
9e7454d0 | 11317 | asm_fprintf (f, "\tldfe\t%r, [%r, #-%d]\n", |
73dd1c00 | 11318 | reg, FP_REGNUM, floats_offset - vfp_offset); |
f21249d5 | 11319 | } |
11320 | } | |
11321 | else | |
11322 | { | |
a2cd141b | 11323 | start_reg = LAST_FPA_REGNUM; |
f21249d5 | 11324 | |
a2cd141b | 11325 | for (reg = LAST_FPA_REGNUM; reg >= FIRST_FPA_REGNUM; reg--) |
f21249d5 | 11326 | { |
3072d30e | 11327 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
f21249d5 | 11328 | { |
11329 | floats_offset += 12; | |
9e7454d0 | 11330 | |
66aacf41 | 11331 | /* We can't unstack more than four registers at once. */ |
f21249d5 | 11332 | if (start_reg - reg == 3) |
11333 | { | |
caceeb43 | 11334 | asm_fprintf (f, "\tlfm\t%r, 4, [%r, #-%d]\n", |
73dd1c00 | 11335 | reg, FP_REGNUM, floats_offset - vfp_offset); |
f21249d5 | 11336 | start_reg = reg - 1; |
11337 | } | |
11338 | } | |
11339 | else | |
11340 | { | |
11341 | if (reg != start_reg) | |
caceeb43 | 11342 | asm_fprintf (f, "\tlfm\t%r, %d, [%r, #-%d]\n", |
11343 | reg + 1, start_reg - reg, | |
73dd1c00 | 11344 | FP_REGNUM, floats_offset - vfp_offset); |
f21249d5 | 11345 | start_reg = reg - 1; |
11346 | } | |
11347 | } | |
11348 | ||
11349 | /* Just in case the last register checked also needs unstacking. */ | |
11350 | if (reg != start_reg) | |
caceeb43 | 11351 | asm_fprintf (f, "\tlfm\t%r, %d, [%r, #-%d]\n", |
11352 | reg + 1, start_reg - reg, | |
73dd1c00 | 11353 | FP_REGNUM, floats_offset - vfp_offset); |
f21249d5 | 11354 | } |
e27ad2d5 | 11355 | |
a2cd141b | 11356 | if (TARGET_HARD_FLOAT && TARGET_VFP) |
11357 | { | |
7b78e739 | 11358 | int saved_size; |
a2cd141b | 11359 | |
382a2a57 | 11360 | /* The fldmd insns do not have base+offset addressing |
11361 | modes, so we use IP to hold the address. */ | |
7b78e739 | 11362 | saved_size = arm_get_vfp_saved_size (); |
a2cd141b | 11363 | |
7b78e739 | 11364 | if (saved_size > 0) |
a2cd141b | 11365 | { |
7b78e739 | 11366 | floats_offset += saved_size; |
a2cd141b | 11367 | asm_fprintf (f, "\tsub\t%r, %r, #%d\n", IP_REGNUM, |
11368 | FP_REGNUM, floats_offset - vfp_offset); | |
11369 | } | |
11370 | start_reg = FIRST_VFP_REGNUM; | |
11371 | for (reg = FIRST_VFP_REGNUM; reg < LAST_VFP_REGNUM; reg += 2) | |
11372 | { | |
3072d30e | 11373 | if ((!df_regs_ever_live_p (reg) || call_used_regs[reg]) |
11374 | && (!df_regs_ever_live_p (reg + 1) || call_used_regs[reg + 1])) | |
a2cd141b | 11375 | { |
11376 | if (start_reg != reg) | |
382a2a57 | 11377 | vfp_output_fldmd (f, IP_REGNUM, |
7b78e739 | 11378 | (start_reg - FIRST_VFP_REGNUM) / 2, |
11379 | (reg - start_reg) / 2); | |
a2cd141b | 11380 | start_reg = reg + 2; |
11381 | } | |
11382 | } | |
11383 | if (start_reg != reg) | |
382a2a57 | 11384 | vfp_output_fldmd (f, IP_REGNUM, |
7b78e739 | 11385 | (start_reg - FIRST_VFP_REGNUM) / 2, |
11386 | (reg - start_reg) / 2); | |
a2cd141b | 11387 | } |
11388 | ||
755eb2b4 | 11389 | if (TARGET_IWMMXT) |
11390 | { | |
11391 | /* The frame pointer is guaranteed to be non-double-word aligned. | |
11392 | This is because it is set to (old_stack_pointer - 4) and the | |
11393 | old_stack_pointer was double word aligned. Thus the offset to | |
11394 | the iWMMXt registers to be loaded must also be non-double-word | |
11395 | sized, so that the resultant address *is* double-word aligned. | |
11396 | We can ignore floats_offset since that was already included in | |
11397 | the live_regs_mask. */ | |
11398 | lrm_count += (lrm_count % 2 ? 2 : 1); | |
9e7454d0 | 11399 | |
9d63835d | 11400 | for (reg = LAST_IWMMXT_REGNUM; reg >= FIRST_IWMMXT_REGNUM; reg--) |
3072d30e | 11401 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
755eb2b4 | 11402 | { |
9e7454d0 | 11403 | asm_fprintf (f, "\twldrd\t%r, [%r, #-%d]\n", |
755eb2b4 | 11404 | reg, FP_REGNUM, lrm_count * 4); |
9e7454d0 | 11405 | lrm_count += 2; |
755eb2b4 | 11406 | } |
11407 | } | |
11408 | ||
89b30d75 | 11409 | /* saved_regs_mask should contain the IP, which at the time of stack |
e27ad2d5 | 11410 | frame generation actually contains the old stack pointer. So a |
11411 | quick way to unwind the stack is just pop the IP register directly | |
11412 | into the stack pointer. */ | |
ed29c566 | 11413 | gcc_assert (saved_regs_mask & (1 << IP_REGNUM)); |
89b30d75 | 11414 | saved_regs_mask &= ~ (1 << IP_REGNUM); |
11415 | saved_regs_mask |= (1 << SP_REGNUM); | |
e27ad2d5 | 11416 | |
89b30d75 | 11417 | /* There are two registers left in saved_regs_mask - LR and PC. We |
e27ad2d5 | 11418 | only need to restore the LR register (the return address), but to |
11419 | save time we can load it directly into the PC, unless we need a | |
11420 | special function exit sequence, or we are not really returning. */ | |
4c44712e | 11421 | if (really_return |
11422 | && ARM_FUNC_TYPE (func_type) == ARM_FT_NORMAL | |
11423 | && !current_function_calls_eh_return) | |
e27ad2d5 | 11424 | /* Delete the LR from the register mask, so that the LR on |
11425 | the stack is loaded into the PC in the register mask. */ | |
89b30d75 | 11426 | saved_regs_mask &= ~ (1 << LR_REGNUM); |
f21249d5 | 11427 | else |
89b30d75 | 11428 | saved_regs_mask &= ~ (1 << PC_REGNUM); |
01c5c234 | 11429 | |
11430 | /* We must use SP as the base register, because SP is one of the | |
11431 | registers being restored. If an interrupt or page fault | |
11432 | happens in the ldm instruction, the SP might or might not | |
11433 | have been restored. That would be bad, as then SP will no | |
11434 | longer indicate the safe area of stack, and we can get stack | |
11435 | corruption. Using SP as the base register means that it will | |
11436 | be reset correctly to the original value, should an interrupt | |
8d69f2ad | 11437 | occur. If the stack pointer already points at the right |
11438 | place, then omit the subtraction. */ | |
f9273c43 | 11439 | if (offsets->outgoing_args != (1 + (int) bit_count (saved_regs_mask)) |
8d69f2ad | 11440 | || current_function_calls_alloca) |
11441 | asm_fprintf (f, "\tsub\t%r, %r, #%d\n", SP_REGNUM, FP_REGNUM, | |
11442 | 4 * bit_count (saved_regs_mask)); | |
8b438b4f | 11443 | print_multi_reg (f, "ldmfd\t%r, ", SP_REGNUM, saved_regs_mask, 0); |
2d9065d3 | 11444 | |
11445 | if (IS_INTERRUPT (func_type)) | |
11446 | /* Interrupt handlers will have pushed the | |
11447 | IP onto the stack, so restore it now. */ | |
8b438b4f | 11448 | print_multi_reg (f, "ldmfd\t%r!, ", SP_REGNUM, 1 << IP_REGNUM, 0); |
8cd42699 | 11449 | } |
11450 | else | |
11451 | { | |
25f905c2 | 11452 | HOST_WIDE_INT amount; |
8b438b4f | 11453 | int rfe; |
093305b1 | 11454 | /* Restore stack pointer if necessary. */ |
25f905c2 | 11455 | if (frame_pointer_needed) |
093305b1 | 11456 | { |
25f905c2 | 11457 | /* For Thumb-2 restore sp from the frame pointer. |
542d5028 | 11458 | Operand restrictions mean we have to increment FP, then copy |
25f905c2 | 11459 | to SP. */ |
11460 | amount = offsets->locals_base - offsets->saved_regs; | |
11461 | operands[0] = hard_frame_pointer_rtx; | |
11462 | } | |
11463 | else | |
11464 | { | |
11465 | operands[0] = stack_pointer_rtx; | |
11466 | amount = offsets->outgoing_args - offsets->saved_regs; | |
11467 | } | |
11468 | ||
11469 | if (amount) | |
11470 | { | |
11471 | operands[1] = operands[0]; | |
11472 | operands[2] = GEN_INT (amount); | |
093305b1 | 11473 | output_add_immediate (operands); |
11474 | } | |
25f905c2 | 11475 | if (frame_pointer_needed) |
11476 | asm_fprintf (f, "\tmov\t%r, %r\n", | |
11477 | SP_REGNUM, HARD_FRAME_POINTER_REGNUM); | |
093305b1 | 11478 | |
c7f506fd | 11479 | if (arm_fpu_arch == FPUTYPE_FPA_EMU2) |
f21249d5 | 11480 | { |
a2cd141b | 11481 | for (reg = FIRST_FPA_REGNUM; reg <= LAST_FPA_REGNUM; reg++) |
3072d30e | 11482 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
caceeb43 | 11483 | asm_fprintf (f, "\tldfe\t%r, [%r], #12\n", |
11484 | reg, SP_REGNUM); | |
f21249d5 | 11485 | } |
11486 | else | |
11487 | { | |
a2cd141b | 11488 | start_reg = FIRST_FPA_REGNUM; |
f21249d5 | 11489 | |
a2cd141b | 11490 | for (reg = FIRST_FPA_REGNUM; reg <= LAST_FPA_REGNUM; reg++) |
f21249d5 | 11491 | { |
3072d30e | 11492 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
f21249d5 | 11493 | { |
11494 | if (reg - start_reg == 3) | |
11495 | { | |
caceeb43 | 11496 | asm_fprintf (f, "\tlfmfd\t%r, 4, [%r]!\n", |
11497 | start_reg, SP_REGNUM); | |
f21249d5 | 11498 | start_reg = reg + 1; |
11499 | } | |
11500 | } | |
11501 | else | |
11502 | { | |
11503 | if (reg != start_reg) | |
caceeb43 | 11504 | asm_fprintf (f, "\tlfmfd\t%r, %d, [%r]!\n", |
11505 | start_reg, reg - start_reg, | |
11506 | SP_REGNUM); | |
9e7454d0 | 11507 | |
f21249d5 | 11508 | start_reg = reg + 1; |
11509 | } | |
11510 | } | |
11511 | ||
11512 | /* Just in case the last register checked also needs unstacking. */ | |
11513 | if (reg != start_reg) | |
caceeb43 | 11514 | asm_fprintf (f, "\tlfmfd\t%r, %d, [%r]!\n", |
11515 | start_reg, reg - start_reg, SP_REGNUM); | |
f21249d5 | 11516 | } |
11517 | ||
a2cd141b | 11518 | if (TARGET_HARD_FLOAT && TARGET_VFP) |
11519 | { | |
11520 | start_reg = FIRST_VFP_REGNUM; | |
11521 | for (reg = FIRST_VFP_REGNUM; reg < LAST_VFP_REGNUM; reg += 2) | |
11522 | { | |
3072d30e | 11523 | if ((!df_regs_ever_live_p (reg) || call_used_regs[reg]) |
11524 | && (!df_regs_ever_live_p (reg + 1) || call_used_regs[reg + 1])) | |
a2cd141b | 11525 | { |
11526 | if (start_reg != reg) | |
382a2a57 | 11527 | vfp_output_fldmd (f, SP_REGNUM, |
7b78e739 | 11528 | (start_reg - FIRST_VFP_REGNUM) / 2, |
11529 | (reg - start_reg) / 2); | |
a2cd141b | 11530 | start_reg = reg + 2; |
11531 | } | |
11532 | } | |
11533 | if (start_reg != reg) | |
382a2a57 | 11534 | vfp_output_fldmd (f, SP_REGNUM, |
7b78e739 | 11535 | (start_reg - FIRST_VFP_REGNUM) / 2, |
11536 | (reg - start_reg) / 2); | |
a2cd141b | 11537 | } |
755eb2b4 | 11538 | if (TARGET_IWMMXT) |
11539 | for (reg = FIRST_IWMMXT_REGNUM; reg <= LAST_IWMMXT_REGNUM; reg++) | |
3072d30e | 11540 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
9d63835d | 11541 | asm_fprintf (f, "\twldrd\t%r, [%r], #8\n", reg, SP_REGNUM); |
755eb2b4 | 11542 | |
e27ad2d5 | 11543 | /* If we can, restore the LR into the PC. */ |
8b438b4f | 11544 | if (ARM_FUNC_TYPE (func_type) != ARM_FT_INTERWORKED |
11545 | && (TARGET_ARM || ARM_FUNC_TYPE (func_type) == ARM_FT_NORMAL) | |
25f905c2 | 11546 | && !IS_STACKALIGN (func_type) |
e27ad2d5 | 11547 | && really_return |
11548 | && current_function_pretend_args_size == 0 | |
4c44712e | 11549 | && saved_regs_mask & (1 << LR_REGNUM) |
11550 | && !current_function_calls_eh_return) | |
8cd42699 | 11551 | { |
89b30d75 | 11552 | saved_regs_mask &= ~ (1 << LR_REGNUM); |
11553 | saved_regs_mask |= (1 << PC_REGNUM); | |
8b438b4f | 11554 | rfe = IS_INTERRUPT (func_type); |
e27ad2d5 | 11555 | } |
8b438b4f | 11556 | else |
11557 | rfe = 0; | |
cffb2a26 | 11558 | |
e27ad2d5 | 11559 | /* Load the registers off the stack. If we only have one register |
25f905c2 | 11560 | to load use the LDR instruction - it is faster. For Thumb-2 |
11561 | always use pop and the assembler will pick the best instruction.*/ | |
8b438b4f | 11562 | if (TARGET_ARM && saved_regs_mask == (1 << LR_REGNUM) |
11563 | && !IS_INTERRUPT(func_type)) | |
e27ad2d5 | 11564 | { |
4c44712e | 11565 | asm_fprintf (f, "\tldr\t%r, [%r], #4\n", LR_REGNUM, SP_REGNUM); |
8cd42699 | 11566 | } |
89b30d75 | 11567 | else if (saved_regs_mask) |
8cc5c6da | 11568 | { |
11569 | if (saved_regs_mask & (1 << SP_REGNUM)) | |
11570 | /* Note - write back to the stack register is not enabled | |
a361b456 | 11571 | (i.e. "ldmfd sp!..."). We know that the stack pointer is |
8cc5c6da | 11572 | in the list of registers and if we add writeback the |
11573 | instruction becomes UNPREDICTABLE. */ | |
8b438b4f | 11574 | print_multi_reg (f, "ldmfd\t%r, ", SP_REGNUM, saved_regs_mask, |
11575 | rfe); | |
25f905c2 | 11576 | else if (TARGET_ARM) |
8b438b4f | 11577 | print_multi_reg (f, "ldmfd\t%r!, ", SP_REGNUM, saved_regs_mask, |
11578 | rfe); | |
8cc5c6da | 11579 | else |
8b438b4f | 11580 | print_multi_reg (f, "pop\t", SP_REGNUM, saved_regs_mask, 0); |
8cc5c6da | 11581 | } |
e27ad2d5 | 11582 | |
11583 | if (current_function_pretend_args_size) | |
8cd42699 | 11584 | { |
e27ad2d5 | 11585 | /* Unwind the pre-pushed regs. */ |
11586 | operands[0] = operands[1] = stack_pointer_rtx; | |
11587 | operands[2] = GEN_INT (current_function_pretend_args_size); | |
11588 | output_add_immediate (operands); | |
11589 | } | |
11590 | } | |
69120d26 | 11591 | |
cdf460bd | 11592 | /* We may have already restored PC directly from the stack. */ |
025e633f | 11593 | if (!really_return || saved_regs_mask & (1 << PC_REGNUM)) |
e27ad2d5 | 11594 | return ""; |
cffb2a26 | 11595 | |
4c44712e | 11596 | /* Stack adjustment for exception handler. */ |
11597 | if (current_function_calls_eh_return) | |
9e7454d0 | 11598 | asm_fprintf (f, "\tadd\t%r, %r, %r\n", SP_REGNUM, SP_REGNUM, |
4c44712e | 11599 | ARM_EH_STACKADJ_REGNUM); |
11600 | ||
e27ad2d5 | 11601 | /* Generate the return instruction. */ |
11602 | switch ((int) ARM_FUNC_TYPE (func_type)) | |
11603 | { | |
e27ad2d5 | 11604 | case ARM_FT_ISR: |
11605 | case ARM_FT_FIQ: | |
11606 | asm_fprintf (f, "\tsubs\t%r, %r, #4\n", PC_REGNUM, LR_REGNUM); | |
11607 | break; | |
11608 | ||
11609 | case ARM_FT_EXCEPTION: | |
11610 | asm_fprintf (f, "\tmovs\t%r, %r\n", PC_REGNUM, LR_REGNUM); | |
11611 | break; | |
11612 | ||
11613 | case ARM_FT_INTERWORKED: | |
11614 | asm_fprintf (f, "\tbx\t%r\n", LR_REGNUM); | |
11615 | break; | |
11616 | ||
11617 | default: | |
25f905c2 | 11618 | if (IS_STACKALIGN (func_type)) |
11619 | { | |
11620 | /* See comment in arm_expand_prologue. */ | |
11621 | asm_fprintf (f, "\tmov\t%r, %r\n", SP_REGNUM, 0); | |
11622 | } | |
f1039640 | 11623 | if (arm_arch5 || arm_arch4t) |
11624 | asm_fprintf (f, "\tbx\t%r\n", LR_REGNUM); | |
11625 | else | |
11626 | asm_fprintf (f, "\tmov\t%r, %r\n", PC_REGNUM, LR_REGNUM); | |
e27ad2d5 | 11627 | break; |
8cd42699 | 11628 | } |
a690ede2 | 11629 | |
56d27660 | 11630 | return ""; |
11631 | } | |
11632 | ||
17d9b0c3 | 11633 | static void |
ebd88a36 | 11634 | arm_output_function_epilogue (FILE *file ATTRIBUTE_UNUSED, |
f9273c43 | 11635 | HOST_WIDE_INT frame_size ATTRIBUTE_UNUSED) |
56d27660 | 11636 | { |
f9273c43 | 11637 | arm_stack_offsets *offsets; |
11638 | ||
25f905c2 | 11639 | if (TARGET_THUMB1) |
cffb2a26 | 11640 | { |
afe27f3b | 11641 | int regno; |
11642 | ||
11643 | /* Emit any call-via-reg trampolines that are needed for v4t support | |
11644 | of call_reg and call_value_reg type insns. */ | |
a52972b8 | 11645 | for (regno = 0; regno < LR_REGNUM; regno++) |
afe27f3b | 11646 | { |
11647 | rtx label = cfun->machine->call_via[regno]; | |
11648 | ||
11649 | if (label != NULL) | |
11650 | { | |
2f14b1f9 | 11651 | switch_to_section (function_section (current_function_decl)); |
afe27f3b | 11652 | targetm.asm_out.internal_label (asm_out_file, "L", |
11653 | CODE_LABEL_NUMBER (label)); | |
11654 | asm_fprintf (asm_out_file, "\tbx\t%r\n", regno); | |
11655 | } | |
11656 | } | |
11657 | ||
cffb2a26 | 11658 | /* ??? Probably not safe to set this here, since it assumes that a |
11659 | function will be emitted as assembly immediately after we generate | |
11660 | RTL for it. This does not happen for inline functions. */ | |
11661 | return_used_this_function = 0; | |
11662 | } | |
25f905c2 | 11663 | else /* TARGET_32BIT */ |
cffb2a26 | 11664 | { |
a7712927 | 11665 | /* We need to take into account any stack-frame rounding. */ |
f9273c43 | 11666 | offsets = arm_get_frame_offsets (); |
a7712927 | 11667 | |
ed29c566 | 11668 | gcc_assert (!use_return_insn (FALSE, NULL) |
11669 | || !return_used_this_function | |
11670 | || offsets->saved_regs == offsets->outgoing_args | |
11671 | || frame_pointer_needed); | |
a690ede2 | 11672 | |
cffb2a26 | 11673 | /* Reset the ARM-specific per-function variables. */ |
cffb2a26 | 11674 | after_arm_reorg = 0; |
11675 | } | |
a690ede2 | 11676 | } |
4b6897e9 | 11677 | |
220d204b | 11678 | /* Generate and emit an insn that we will recognize as a push_multi. |
11679 | Unfortunately, since this insn does not reflect very well the actual | |
11680 | semantics of the operation, we need to annotate the insn for the benefit | |
11681 | of DWARF2 frame unwind information. */ | |
220d204b | 11682 | static rtx |
53f179c0 | 11683 | emit_multi_reg_push (unsigned long mask) |
4b6897e9 | 11684 | { |
11685 | int num_regs = 0; | |
5cf3595a | 11686 | int num_dwarf_regs; |
4b6897e9 | 11687 | int i, j; |
11688 | rtx par; | |
220d204b | 11689 | rtx dwarf; |
c17534c2 | 11690 | int dwarf_par_index; |
220d204b | 11691 | rtx tmp, reg; |
4b6897e9 | 11692 | |
cffb2a26 | 11693 | for (i = 0; i <= LAST_ARM_REGNUM; i++) |
4b6897e9 | 11694 | if (mask & (1 << i)) |
215b30b3 | 11695 | num_regs++; |
4b6897e9 | 11696 | |
ed29c566 | 11697 | gcc_assert (num_regs && num_regs <= 16); |
4b6897e9 | 11698 | |
5cf3595a | 11699 | /* We don't record the PC in the dwarf frame information. */ |
11700 | num_dwarf_regs = num_regs; | |
11701 | if (mask & (1 << PC_REGNUM)) | |
11702 | num_dwarf_regs--; | |
11703 | ||
c17534c2 | 11704 | /* For the body of the insn we are going to generate an UNSPEC in |
a2f10574 | 11705 | parallel with several USEs. This allows the insn to be recognized |
c17534c2 | 11706 | by the push_multi pattern in the arm.md file. The insn looks |
11707 | something like this: | |
11708 | ||
9e7454d0 | 11709 | (parallel [ |
e1159bbe | 11710 | (set (mem:BLK (pre_dec:BLK (reg:SI sp))) |
11711 | (unspec:BLK [(reg:SI r4)] UNSPEC_PUSH_MULT)) | |
c17534c2 | 11712 | (use (reg:SI 11 fp)) |
11713 | (use (reg:SI 12 ip)) | |
11714 | (use (reg:SI 14 lr)) | |
11715 | (use (reg:SI 15 pc)) | |
11716 | ]) | |
11717 | ||
11718 | For the frame note however, we try to be more explicit and actually | |
11719 | show each register being stored into the stack frame, plus a (single) | |
11720 | decrement of the stack pointer. We do it this way in order to be | |
11721 | friendly to the stack unwinding code, which only wants to see a single | |
11722 | stack decrement per instruction. The RTL we generate for the note looks | |
11723 | something like this: | |
11724 | ||
9e7454d0 | 11725 | (sequence [ |
c17534c2 | 11726 | (set (reg:SI sp) (plus:SI (reg:SI sp) (const_int -20))) |
11727 | (set (mem:SI (reg:SI sp)) (reg:SI r4)) | |
11728 | (set (mem:SI (plus:SI (reg:SI sp) (const_int 4))) (reg:SI fp)) | |
11729 | (set (mem:SI (plus:SI (reg:SI sp) (const_int 8))) (reg:SI ip)) | |
11730 | (set (mem:SI (plus:SI (reg:SI sp) (const_int 12))) (reg:SI lr)) | |
c17534c2 | 11731 | ]) |
11732 | ||
11733 | This sequence is used both by the code to support stack unwinding for | |
11734 | exceptions handlers and the code to generate dwarf2 frame debugging. */ | |
9e7454d0 | 11735 | |
76676c8e | 11736 | par = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (num_regs)); |
5cf3595a | 11737 | dwarf = gen_rtx_SEQUENCE (VOIDmode, rtvec_alloc (num_dwarf_regs + 1)); |
c17534c2 | 11738 | dwarf_par_index = 1; |
4b6897e9 | 11739 | |
cffb2a26 | 11740 | for (i = 0; i <= LAST_ARM_REGNUM; i++) |
4b6897e9 | 11741 | { |
11742 | if (mask & (1 << i)) | |
11743 | { | |
220d204b | 11744 | reg = gen_rtx_REG (SImode, i); |
11745 | ||
4b6897e9 | 11746 | XVECEXP (par, 0, 0) |
76676c8e | 11747 | = gen_rtx_SET (VOIDmode, |
788fcce0 | 11748 | gen_frame_mem (BLKmode, |
11749 | gen_rtx_PRE_DEC (BLKmode, | |
11750 | stack_pointer_rtx)), | |
76676c8e | 11751 | gen_rtx_UNSPEC (BLKmode, |
220d204b | 11752 | gen_rtvec (1, reg), |
5cf3595a | 11753 | UNSPEC_PUSH_MULT)); |
220d204b | 11754 | |
5cf3595a | 11755 | if (i != PC_REGNUM) |
11756 | { | |
11757 | tmp = gen_rtx_SET (VOIDmode, | |
788fcce0 | 11758 | gen_frame_mem (SImode, stack_pointer_rtx), |
5cf3595a | 11759 | reg); |
11760 | RTX_FRAME_RELATED_P (tmp) = 1; | |
11761 | XVECEXP (dwarf, 0, dwarf_par_index) = tmp; | |
11762 | dwarf_par_index++; | |
11763 | } | |
220d204b | 11764 | |
4b6897e9 | 11765 | break; |
11766 | } | |
11767 | } | |
11768 | ||
11769 | for (j = 1, i++; j < num_regs; i++) | |
11770 | { | |
11771 | if (mask & (1 << i)) | |
11772 | { | |
220d204b | 11773 | reg = gen_rtx_REG (SImode, i); |
11774 | ||
11775 | XVECEXP (par, 0, j) = gen_rtx_USE (VOIDmode, reg); | |
11776 | ||
5cf3595a | 11777 | if (i != PC_REGNUM) |
11778 | { | |
788fcce0 | 11779 | tmp |
11780 | = gen_rtx_SET (VOIDmode, | |
11781 | gen_frame_mem (SImode, | |
5cf3595a | 11782 | plus_constant (stack_pointer_rtx, |
11783 | 4 * j)), | |
788fcce0 | 11784 | reg); |
5cf3595a | 11785 | RTX_FRAME_RELATED_P (tmp) = 1; |
11786 | XVECEXP (dwarf, 0, dwarf_par_index++) = tmp; | |
11787 | } | |
11788 | ||
4b6897e9 | 11789 | j++; |
11790 | } | |
11791 | } | |
f21249d5 | 11792 | |
220d204b | 11793 | par = emit_insn (par); |
9e7454d0 | 11794 | |
0083a3b9 | 11795 | tmp = gen_rtx_SET (VOIDmode, |
c17534c2 | 11796 | stack_pointer_rtx, |
0083a3b9 | 11797 | plus_constant (stack_pointer_rtx, -4 * num_regs)); |
c17534c2 | 11798 | RTX_FRAME_RELATED_P (tmp) = 1; |
11799 | XVECEXP (dwarf, 0, 0) = tmp; | |
9e7454d0 | 11800 | |
220d204b | 11801 | REG_NOTES (par) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, dwarf, |
11802 | REG_NOTES (par)); | |
11803 | return par; | |
f21249d5 | 11804 | } |
11805 | ||
0858d94e | 11806 | /* Calculate the size of the return value that is passed in registers. */ |
11807 | static int | |
11808 | arm_size_return_regs (void) | |
11809 | { | |
11810 | enum machine_mode mode; | |
11811 | ||
11812 | if (current_function_return_rtx != 0) | |
11813 | mode = GET_MODE (current_function_return_rtx); | |
11814 | else | |
11815 | mode = DECL_MODE (DECL_RESULT (current_function_decl)); | |
11816 | ||
11817 | return GET_MODE_SIZE (mode); | |
11818 | } | |
11819 | ||
220d204b | 11820 | static rtx |
ebd88a36 | 11821 | emit_sfm (int base_reg, int count) |
f21249d5 | 11822 | { |
11823 | rtx par; | |
220d204b | 11824 | rtx dwarf; |
11825 | rtx tmp, reg; | |
f21249d5 | 11826 | int i; |
11827 | ||
76676c8e | 11828 | par = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (count)); |
4cde8e57 | 11829 | dwarf = gen_rtx_SEQUENCE (VOIDmode, rtvec_alloc (count + 1)); |
220d204b | 11830 | |
11831 | reg = gen_rtx_REG (XFmode, base_reg++); | |
76676c8e | 11832 | |
11833 | XVECEXP (par, 0, 0) | |
9e7454d0 | 11834 | = gen_rtx_SET (VOIDmode, |
788fcce0 | 11835 | gen_frame_mem (BLKmode, |
11836 | gen_rtx_PRE_DEC (BLKmode, | |
11837 | stack_pointer_rtx)), | |
76676c8e | 11838 | gen_rtx_UNSPEC (BLKmode, |
220d204b | 11839 | gen_rtvec (1, reg), |
e1159bbe | 11840 | UNSPEC_PUSH_MULT)); |
9e7454d0 | 11841 | tmp = gen_rtx_SET (VOIDmode, |
788fcce0 | 11842 | gen_frame_mem (XFmode, stack_pointer_rtx), reg); |
220d204b | 11843 | RTX_FRAME_RELATED_P (tmp) = 1; |
9e7454d0 | 11844 | XVECEXP (dwarf, 0, 1) = tmp; |
11845 | ||
f21249d5 | 11846 | for (i = 1; i < count; i++) |
220d204b | 11847 | { |
11848 | reg = gen_rtx_REG (XFmode, base_reg++); | |
11849 | XVECEXP (par, 0, i) = gen_rtx_USE (VOIDmode, reg); | |
11850 | ||
9e7454d0 | 11851 | tmp = gen_rtx_SET (VOIDmode, |
788fcce0 | 11852 | gen_frame_mem (XFmode, |
11853 | plus_constant (stack_pointer_rtx, | |
11854 | i * 12)), | |
220d204b | 11855 | reg); |
11856 | RTX_FRAME_RELATED_P (tmp) = 1; | |
9e7454d0 | 11857 | XVECEXP (dwarf, 0, i + 1) = tmp; |
220d204b | 11858 | } |
f21249d5 | 11859 | |
4cde8e57 | 11860 | tmp = gen_rtx_SET (VOIDmode, |
11861 | stack_pointer_rtx, | |
0083a3b9 | 11862 | plus_constant (stack_pointer_rtx, -12 * count)); |
11863 | ||
4cde8e57 | 11864 | RTX_FRAME_RELATED_P (tmp) = 1; |
11865 | XVECEXP (dwarf, 0, 0) = tmp; | |
11866 | ||
220d204b | 11867 | par = emit_insn (par); |
11868 | REG_NOTES (par) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, dwarf, | |
11869 | REG_NOTES (par)); | |
11870 | return par; | |
4b6897e9 | 11871 | } |
11872 | ||
a2cd141b | 11873 | |
ef5bbe67 | 11874 | /* Return true if the current function needs to save/restore LR. */ |
11875 | ||
11876 | static bool | |
11877 | thumb_force_lr_save (void) | |
11878 | { | |
11879 | return !cfun->machine->lr_save_eliminated | |
11880 | && (!leaf_function_p () | |
11881 | || thumb_far_jump_used_p () | |
3072d30e | 11882 | || df_regs_ever_live_p (LR_REGNUM)); |
ef5bbe67 | 11883 | } |
11884 | ||
11885 | ||
9966e02f | 11886 | /* Compute the distance from register FROM to register TO. |
11887 | These can be the arg pointer (26), the soft frame pointer (25), | |
11888 | the stack pointer (13) or the hard frame pointer (11). | |
4c44712e | 11889 | In thumb mode r7 is used as the soft frame pointer, if needed. |
9966e02f | 11890 | Typical stack layout looks like this: |
11891 | ||
11892 | old stack pointer -> | | | |
11893 | ---- | |
11894 | | | \ | |
11895 | | | saved arguments for | |
11896 | | | vararg functions | |
11897 | | | / | |
11898 | -- | |
11899 | hard FP & arg pointer -> | | \ | |
11900 | | | stack | |
11901 | | | frame | |
11902 | | | / | |
11903 | -- | |
11904 | | | \ | |
11905 | | | call saved | |
11906 | | | registers | |
11907 | soft frame pointer -> | | / | |
11908 | -- | |
11909 | | | \ | |
11910 | | | local | |
11911 | | | variables | |
f90b51f1 | 11912 | locals base pointer -> | | / |
9966e02f | 11913 | -- |
11914 | | | \ | |
11915 | | | outgoing | |
11916 | | | arguments | |
11917 | current stack pointer -> | | / | |
11918 | -- | |
11919 | ||
28c2d844 | 11920 | For a given function some or all of these stack components |
9966e02f | 11921 | may not be needed, giving rise to the possibility of |
11922 | eliminating some of the registers. | |
11923 | ||
e3e08e7f | 11924 | The values returned by this function must reflect the behavior |
9966e02f | 11925 | of arm_expand_prologue() and arm_compute_save_reg_mask(). |
11926 | ||
11927 | The sign of the number returned reflects the direction of stack | |
11928 | growth, so the values are positive for all eliminations except | |
f9273c43 | 11929 | from the soft frame pointer to the hard frame pointer. |
11930 | ||
11931 | SFP may point just inside the local variables block to ensure correct | |
11932 | alignment. */ | |
11933 | ||
11934 | ||
11935 | /* Calculate stack offsets. These are used to calculate register elimination | |
11936 | offsets and in prologue/epilogue code. */ | |
11937 | ||
11938 | static arm_stack_offsets * | |
11939 | arm_get_frame_offsets (void) | |
9966e02f | 11940 | { |
f9273c43 | 11941 | struct arm_stack_offsets *offsets; |
9966e02f | 11942 | unsigned long func_type; |
f9273c43 | 11943 | int leaf; |
f9273c43 | 11944 | int saved; |
11945 | HOST_WIDE_INT frame_size; | |
11946 | ||
11947 | offsets = &cfun->machine->stack_offsets; | |
9e7454d0 | 11948 | |
f9273c43 | 11949 | /* We need to know if we are a leaf function. Unfortunately, it |
11950 | is possible to be called after start_sequence has been called, | |
11951 | which causes get_insns to return the insns for the sequence, | |
11952 | not the function, which will cause leaf_function_p to return | |
11953 | the incorrect result. | |
9966e02f | 11954 | |
f9273c43 | 11955 | to know about leaf functions once reload has completed, and the |
11956 | frame size cannot be changed after that time, so we can safely | |
11957 | use the cached value. */ | |
11958 | ||
11959 | if (reload_completed) | |
11960 | return offsets; | |
11961 | ||
af6ed417 | 11962 | /* Initially this is the size of the local variables. It will translated |
11963 | into an offset once we have determined the size of preceding data. */ | |
f9273c43 | 11964 | frame_size = ROUND_UP_WORD (get_frame_size ()); |
11965 | ||
11966 | leaf = leaf_function_p (); | |
11967 | ||
11968 | /* Space for variadic functions. */ | |
11969 | offsets->saved_args = current_function_pretend_args_size; | |
11970 | ||
25f905c2 | 11971 | /* In Thumb mode this is incorrect, but never used. */ |
f9273c43 | 11972 | offsets->frame = offsets->saved_args + (frame_pointer_needed ? 4 : 0); |
11973 | ||
25f905c2 | 11974 | if (TARGET_32BIT) |
9966e02f | 11975 | { |
f9273c43 | 11976 | unsigned int regno; |
ba3d677f | 11977 | |
f9273c43 | 11978 | saved = bit_count (arm_compute_save_reg_mask ()) * 4; |
755eb2b4 | 11979 | |
f9273c43 | 11980 | /* We know that SP will be doubleword aligned on entry, and we must |
11981 | preserve that condition at any subroutine call. We also require the | |
11982 | soft frame pointer to be doubleword aligned. */ | |
11983 | ||
11984 | if (TARGET_REALLY_IWMMXT) | |
a2cd141b | 11985 | { |
f9273c43 | 11986 | /* Check for the call-saved iWMMXt registers. */ |
11987 | for (regno = FIRST_IWMMXT_REGNUM; | |
11988 | regno <= LAST_IWMMXT_REGNUM; | |
11989 | regno++) | |
3072d30e | 11990 | if (df_regs_ever_live_p (regno) && ! call_used_regs[regno]) |
f9273c43 | 11991 | saved += 8; |
11992 | } | |
11993 | ||
11994 | func_type = arm_current_func_type (); | |
11995 | if (! IS_VOLATILE (func_type)) | |
11996 | { | |
11997 | /* Space for saved FPA registers. */ | |
11998 | for (regno = FIRST_FPA_REGNUM; regno <= LAST_FPA_REGNUM; regno++) | |
3072d30e | 11999 | if (df_regs_ever_live_p (regno) && ! call_used_regs[regno]) |
f9273c43 | 12000 | saved += 12; |
12001 | ||
12002 | /* Space for saved VFP registers. */ | |
12003 | if (TARGET_HARD_FLOAT && TARGET_VFP) | |
7b78e739 | 12004 | saved += arm_get_vfp_saved_size (); |
a2cd141b | 12005 | } |
f9273c43 | 12006 | } |
25f905c2 | 12007 | else /* TARGET_THUMB1 */ |
f9273c43 | 12008 | { |
25f905c2 | 12009 | saved = bit_count (thumb1_compute_save_reg_mask ()) * 4; |
f9273c43 | 12010 | if (TARGET_BACKTRACE) |
237533cc | 12011 | saved += 16; |
f9273c43 | 12012 | } |
a2cd141b | 12013 | |
f9273c43 | 12014 | /* Saved registers include the stack frame. */ |
12015 | offsets->saved_regs = offsets->saved_args + saved; | |
150502c9 | 12016 | offsets->soft_frame = offsets->saved_regs + CALLER_INTERWORKING_SLOT_SIZE; |
f9273c43 | 12017 | /* A leaf function does not need any stack alignment if it has nothing |
12018 | on the stack. */ | |
12019 | if (leaf && frame_size == 0) | |
12020 | { | |
12021 | offsets->outgoing_args = offsets->soft_frame; | |
a650636a | 12022 | offsets->locals_base = offsets->soft_frame; |
f9273c43 | 12023 | return offsets; |
12024 | } | |
12025 | ||
12026 | /* Ensure SFP has the correct alignment. */ | |
12027 | if (ARM_DOUBLEWORD_ALIGN | |
12028 | && (offsets->soft_frame & 7)) | |
12029 | offsets->soft_frame += 4; | |
12030 | ||
f90b51f1 | 12031 | offsets->locals_base = offsets->soft_frame + frame_size; |
12032 | offsets->outgoing_args = (offsets->locals_base | |
12033 | + current_function_outgoing_args_size); | |
f9273c43 | 12034 | |
12035 | if (ARM_DOUBLEWORD_ALIGN) | |
12036 | { | |
12037 | /* Ensure SP remains doubleword aligned. */ | |
12038 | if (offsets->outgoing_args & 7) | |
12039 | offsets->outgoing_args += 4; | |
ed29c566 | 12040 | gcc_assert (!(offsets->outgoing_args & 7)); |
9966e02f | 12041 | } |
12042 | ||
f9273c43 | 12043 | return offsets; |
12044 | } | |
12045 | ||
12046 | ||
af6ed417 | 12047 | /* Calculate the relative offsets for the different stack pointers. Positive |
f9273c43 | 12048 | offsets are in the direction of stack growth. */ |
12049 | ||
30e9913f | 12050 | HOST_WIDE_INT |
f9273c43 | 12051 | arm_compute_initial_elimination_offset (unsigned int from, unsigned int to) |
12052 | { | |
12053 | arm_stack_offsets *offsets; | |
12054 | ||
12055 | offsets = arm_get_frame_offsets (); | |
9966e02f | 12056 | |
9966e02f | 12057 | /* OK, now we have enough information to compute the distances. |
12058 | There must be an entry in these switch tables for each pair | |
12059 | of registers in ELIMINABLE_REGS, even if some of the entries | |
12060 | seem to be redundant or useless. */ | |
12061 | switch (from) | |
12062 | { | |
12063 | case ARG_POINTER_REGNUM: | |
12064 | switch (to) | |
12065 | { | |
12066 | case THUMB_HARD_FRAME_POINTER_REGNUM: | |
12067 | return 0; | |
12068 | ||
12069 | case FRAME_POINTER_REGNUM: | |
12070 | /* This is the reverse of the soft frame pointer | |
12071 | to hard frame pointer elimination below. */ | |
f9273c43 | 12072 | return offsets->soft_frame - offsets->saved_args; |
9966e02f | 12073 | |
12074 | case ARM_HARD_FRAME_POINTER_REGNUM: | |
12075 | /* If there is no stack frame then the hard | |
12076 | frame pointer and the arg pointer coincide. */ | |
f9273c43 | 12077 | if (offsets->frame == offsets->saved_regs) |
9966e02f | 12078 | return 0; |
4ee9c684 | 12079 | /* FIXME: Not sure about this. Maybe we should always return 0 ? */ |
12080 | return (frame_pointer_needed | |
12081 | && cfun->static_chain_decl != NULL | |
12082 | && ! cfun->machine->uses_anonymous_args) ? 4 : 0; | |
9966e02f | 12083 | |
12084 | case STACK_POINTER_REGNUM: | |
12085 | /* If nothing has been pushed on the stack at all | |
12086 | then this will return -4. This *is* correct! */ | |
f9273c43 | 12087 | return offsets->outgoing_args - (offsets->saved_args + 4); |
9966e02f | 12088 | |
12089 | default: | |
ed29c566 | 12090 | gcc_unreachable (); |
9966e02f | 12091 | } |
ed29c566 | 12092 | gcc_unreachable (); |
9966e02f | 12093 | |
12094 | case FRAME_POINTER_REGNUM: | |
12095 | switch (to) | |
12096 | { | |
12097 | case THUMB_HARD_FRAME_POINTER_REGNUM: | |
12098 | return 0; | |
12099 | ||
12100 | case ARM_HARD_FRAME_POINTER_REGNUM: | |
12101 | /* The hard frame pointer points to the top entry in the | |
12102 | stack frame. The soft frame pointer to the bottom entry | |
12103 | in the stack frame. If there is no stack frame at all, | |
12104 | then they are identical. */ | |
f9273c43 | 12105 | |
12106 | return offsets->frame - offsets->soft_frame; | |
9966e02f | 12107 | |
12108 | case STACK_POINTER_REGNUM: | |
f9273c43 | 12109 | return offsets->outgoing_args - offsets->soft_frame; |
9966e02f | 12110 | |
12111 | default: | |
ed29c566 | 12112 | gcc_unreachable (); |
9966e02f | 12113 | } |
ed29c566 | 12114 | gcc_unreachable (); |
9966e02f | 12115 | |
12116 | default: | |
12117 | /* You cannot eliminate from the stack pointer. | |
12118 | In theory you could eliminate from the hard frame | |
12119 | pointer to the stack pointer, but this will never | |
12120 | happen, since if a stack frame is not needed the | |
12121 | hard frame pointer will never be used. */ | |
ed29c566 | 12122 | gcc_unreachable (); |
9966e02f | 12123 | } |
12124 | } | |
12125 | ||
a7712927 | 12126 | |
542d5028 | 12127 | /* Emit RTL to save coprocessor registers on function entry. Returns the |
25f905c2 | 12128 | number of bytes pushed. */ |
12129 | ||
12130 | static int | |
12131 | arm_save_coproc_regs(void) | |
12132 | { | |
12133 | int saved_size = 0; | |
12134 | unsigned reg; | |
12135 | unsigned start_reg; | |
12136 | rtx insn; | |
12137 | ||
12138 | for (reg = LAST_IWMMXT_REGNUM; reg >= FIRST_IWMMXT_REGNUM; reg--) | |
3072d30e | 12139 | if (df_regs_ever_live_p (reg) && ! call_used_regs[reg]) |
25f905c2 | 12140 | { |
12141 | insn = gen_rtx_PRE_DEC (V2SImode, stack_pointer_rtx); | |
12142 | insn = gen_rtx_MEM (V2SImode, insn); | |
12143 | insn = emit_set_insn (insn, gen_rtx_REG (V2SImode, reg)); | |
12144 | RTX_FRAME_RELATED_P (insn) = 1; | |
12145 | saved_size += 8; | |
12146 | } | |
12147 | ||
12148 | /* Save any floating point call-saved registers used by this | |
12149 | function. */ | |
12150 | if (arm_fpu_arch == FPUTYPE_FPA_EMU2) | |
12151 | { | |
12152 | for (reg = LAST_FPA_REGNUM; reg >= FIRST_FPA_REGNUM; reg--) | |
3072d30e | 12153 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
25f905c2 | 12154 | { |
12155 | insn = gen_rtx_PRE_DEC (XFmode, stack_pointer_rtx); | |
12156 | insn = gen_rtx_MEM (XFmode, insn); | |
12157 | insn = emit_set_insn (insn, gen_rtx_REG (XFmode, reg)); | |
12158 | RTX_FRAME_RELATED_P (insn) = 1; | |
12159 | saved_size += 12; | |
12160 | } | |
12161 | } | |
12162 | else | |
12163 | { | |
12164 | start_reg = LAST_FPA_REGNUM; | |
12165 | ||
12166 | for (reg = LAST_FPA_REGNUM; reg >= FIRST_FPA_REGNUM; reg--) | |
12167 | { | |
3072d30e | 12168 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
25f905c2 | 12169 | { |
12170 | if (start_reg - reg == 3) | |
12171 | { | |
12172 | insn = emit_sfm (reg, 4); | |
12173 | RTX_FRAME_RELATED_P (insn) = 1; | |
12174 | saved_size += 48; | |
12175 | start_reg = reg - 1; | |
12176 | } | |
12177 | } | |
12178 | else | |
12179 | { | |
12180 | if (start_reg != reg) | |
12181 | { | |
12182 | insn = emit_sfm (reg + 1, start_reg - reg); | |
12183 | RTX_FRAME_RELATED_P (insn) = 1; | |
12184 | saved_size += (start_reg - reg) * 12; | |
12185 | } | |
12186 | start_reg = reg - 1; | |
12187 | } | |
12188 | } | |
12189 | ||
12190 | if (start_reg != reg) | |
12191 | { | |
12192 | insn = emit_sfm (reg + 1, start_reg - reg); | |
12193 | saved_size += (start_reg - reg) * 12; | |
12194 | RTX_FRAME_RELATED_P (insn) = 1; | |
12195 | } | |
12196 | } | |
12197 | if (TARGET_HARD_FLOAT && TARGET_VFP) | |
12198 | { | |
12199 | start_reg = FIRST_VFP_REGNUM; | |
12200 | ||
12201 | for (reg = FIRST_VFP_REGNUM; reg < LAST_VFP_REGNUM; reg += 2) | |
12202 | { | |
3072d30e | 12203 | if ((!df_regs_ever_live_p (reg) || call_used_regs[reg]) |
12204 | && (!df_regs_ever_live_p (reg + 1) || call_used_regs[reg + 1])) | |
25f905c2 | 12205 | { |
12206 | if (start_reg != reg) | |
12207 | saved_size += vfp_emit_fstmd (start_reg, | |
12208 | (reg - start_reg) / 2); | |
12209 | start_reg = reg + 2; | |
12210 | } | |
12211 | } | |
12212 | if (start_reg != reg) | |
12213 | saved_size += vfp_emit_fstmd (start_reg, | |
12214 | (reg - start_reg) / 2); | |
12215 | } | |
12216 | return saved_size; | |
12217 | } | |
12218 | ||
12219 | ||
12220 | /* Set the Thumb frame pointer from the stack pointer. */ | |
12221 | ||
12222 | static void | |
12223 | thumb_set_frame_pointer (arm_stack_offsets *offsets) | |
12224 | { | |
12225 | HOST_WIDE_INT amount; | |
12226 | rtx insn, dwarf; | |
12227 | ||
12228 | amount = offsets->outgoing_args - offsets->locals_base; | |
12229 | if (amount < 1024) | |
12230 | insn = emit_insn (gen_addsi3 (hard_frame_pointer_rtx, | |
12231 | stack_pointer_rtx, GEN_INT (amount))); | |
12232 | else | |
12233 | { | |
12234 | emit_insn (gen_movsi (hard_frame_pointer_rtx, GEN_INT (amount))); | |
12235 | insn = emit_insn (gen_addsi3 (hard_frame_pointer_rtx, | |
12236 | hard_frame_pointer_rtx, | |
12237 | stack_pointer_rtx)); | |
12238 | dwarf = gen_rtx_SET (VOIDmode, hard_frame_pointer_rtx, | |
12239 | plus_constant (stack_pointer_rtx, amount)); | |
12240 | RTX_FRAME_RELATED_P (dwarf) = 1; | |
12241 | REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, dwarf, | |
12242 | REG_NOTES (insn)); | |
12243 | } | |
12244 | ||
12245 | RTX_FRAME_RELATED_P (insn) = 1; | |
12246 | } | |
12247 | ||
12248 | /* Generate the prologue instructions for entry into an ARM or Thumb-2 | |
12249 | function. */ | |
4b6897e9 | 12250 | void |
ebd88a36 | 12251 | arm_expand_prologue (void) |
4b6897e9 | 12252 | { |
e27ad2d5 | 12253 | rtx amount; |
220d204b | 12254 | rtx insn; |
209368fd | 12255 | rtx ip_rtx; |
e27ad2d5 | 12256 | unsigned long live_regs_mask; |
12257 | unsigned long func_type; | |
209368fd | 12258 | int fp_offset = 0; |
9966e02f | 12259 | int saved_pretend_args = 0; |
f9273c43 | 12260 | int saved_regs = 0; |
816c161a | 12261 | unsigned HOST_WIDE_INT args_to_push; |
f9273c43 | 12262 | arm_stack_offsets *offsets; |
dd3c06e0 | 12263 | |
e27ad2d5 | 12264 | func_type = arm_current_func_type (); |
4b6897e9 | 12265 | |
9011416b | 12266 | /* Naked functions don't have prologues. */ |
e27ad2d5 | 12267 | if (IS_NAKED (func_type)) |
9011416b | 12268 | return; |
12269 | ||
9966e02f | 12270 | /* Make a copy of c_f_p_a_s as we may need to modify it locally. */ |
12271 | args_to_push = current_function_pretend_args_size; | |
9e7454d0 | 12272 | |
e27ad2d5 | 12273 | /* Compute which register we will have to save onto the stack. */ |
12274 | live_regs_mask = arm_compute_save_reg_mask (); | |
4b6897e9 | 12275 | |
209368fd | 12276 | ip_rtx = gen_rtx_REG (SImode, IP_REGNUM); |
dd3c06e0 | 12277 | |
25f905c2 | 12278 | if (IS_STACKALIGN (func_type)) |
12279 | { | |
12280 | rtx dwarf; | |
12281 | rtx r0; | |
12282 | rtx r1; | |
12283 | /* Handle a word-aligned stack pointer. We generate the following: | |
12284 | ||
12285 | mov r0, sp | |
12286 | bic r1, r0, #7 | |
12287 | mov sp, r1 | |
12288 | <save and restore r0 in normal prologue/epilogue> | |
12289 | mov sp, r0 | |
12290 | bx lr | |
12291 | ||
12292 | The unwinder doesn't need to know about the stack realignment. | |
12293 | Just tell it we saved SP in r0. */ | |
12294 | gcc_assert (TARGET_THUMB2 && !arm_arch_notm && args_to_push == 0); | |
12295 | ||
12296 | r0 = gen_rtx_REG (SImode, 0); | |
12297 | r1 = gen_rtx_REG (SImode, 1); | |
12298 | dwarf = gen_rtx_UNSPEC (SImode, NULL_RTVEC, UNSPEC_STACK_ALIGN); | |
12299 | dwarf = gen_rtx_SET (VOIDmode, r0, dwarf); | |
12300 | insn = gen_movsi (r0, stack_pointer_rtx); | |
12301 | RTX_FRAME_RELATED_P (insn) = 1; | |
12302 | REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, | |
12303 | dwarf, REG_NOTES (insn)); | |
12304 | emit_insn (insn); | |
12305 | emit_insn (gen_andsi3 (r1, r0, GEN_INT (~(HOST_WIDE_INT)7))); | |
12306 | emit_insn (gen_movsi (stack_pointer_rtx, r1)); | |
12307 | } | |
12308 | ||
12309 | if (frame_pointer_needed && TARGET_ARM) | |
4b6897e9 | 12310 | { |
2d9065d3 | 12311 | if (IS_INTERRUPT (func_type)) |
12312 | { | |
12313 | /* Interrupt functions must not corrupt any registers. | |
12314 | Creating a frame pointer however, corrupts the IP | |
12315 | register, so we must push it first. */ | |
12316 | insn = emit_multi_reg_push (1 << IP_REGNUM); | |
aa1ae31c | 12317 | |
12318 | /* Do not set RTX_FRAME_RELATED_P on this insn. | |
12319 | The dwarf stack unwinding code only wants to see one | |
12320 | stack decrement per function, and this is not it. If | |
12321 | this instruction is labeled as being part of the frame | |
12322 | creation sequence then dwarf2out_frame_debug_expr will | |
ed29c566 | 12323 | die when it encounters the assignment of IP to FP |
aa1ae31c | 12324 | later on, since the use of SP here establishes SP as |
12325 | the CFA register and not IP. | |
12326 | ||
12327 | Anyway this instruction is not really part of the stack | |
12328 | frame creation although it is part of the prologue. */ | |
2d9065d3 | 12329 | } |
12330 | else if (IS_NESTED (func_type)) | |
209368fd | 12331 | { |
12332 | /* The Static chain register is the same as the IP register | |
12333 | used as a scratch register during stack frame creation. | |
12334 | To get around this need to find somewhere to store IP | |
12335 | whilst the frame is being created. We try the following | |
12336 | places in order: | |
9e7454d0 | 12337 | |
e27ad2d5 | 12338 | 1. The last argument register. |
209368fd | 12339 | 2. A slot on the stack above the frame. (This only |
12340 | works if the function is not a varargs function). | |
9966e02f | 12341 | 3. Register r3, after pushing the argument registers |
12342 | onto the stack. | |
e27ad2d5 | 12343 | |
9b536fa6 | 12344 | Note - we only need to tell the dwarf2 backend about the SP |
12345 | adjustment in the second variant; the static chain register | |
12346 | doesn't need to be unwound, as it doesn't contain a value | |
12347 | inherited from the caller. */ | |
dd3c06e0 | 12348 | |
3072d30e | 12349 | if (df_regs_ever_live_p (3) == false) |
0083a3b9 | 12350 | insn = emit_set_insn (gen_rtx_REG (SImode, 3), ip_rtx); |
9966e02f | 12351 | else if (args_to_push == 0) |
209368fd | 12352 | { |
9b536fa6 | 12353 | rtx dwarf; |
9b536fa6 | 12354 | |
0083a3b9 | 12355 | insn = gen_rtx_PRE_DEC (SImode, stack_pointer_rtx); |
12356 | insn = emit_set_insn (gen_frame_mem (SImode, insn), ip_rtx); | |
209368fd | 12357 | fp_offset = 4; |
9b536fa6 | 12358 | |
12359 | /* Just tell the dwarf backend that we adjusted SP. */ | |
12360 | dwarf = gen_rtx_SET (VOIDmode, stack_pointer_rtx, | |
0083a3b9 | 12361 | plus_constant (stack_pointer_rtx, |
12362 | -fp_offset)); | |
9b536fa6 | 12363 | RTX_FRAME_RELATED_P (insn) = 1; |
12364 | REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, | |
12365 | dwarf, REG_NOTES (insn)); | |
209368fd | 12366 | } |
12367 | else | |
9966e02f | 12368 | { |
12369 | /* Store the args on the stack. */ | |
71904594 | 12370 | if (cfun->machine->uses_anonymous_args) |
9966e02f | 12371 | insn = emit_multi_reg_push |
12372 | ((0xf0 >> (args_to_push / 4)) & 0xf); | |
12373 | else | |
12374 | insn = emit_insn | |
9e7454d0 | 12375 | (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, |
9966e02f | 12376 | GEN_INT (- args_to_push))); |
12377 | ||
12378 | RTX_FRAME_RELATED_P (insn) = 1; | |
12379 | ||
12380 | saved_pretend_args = 1; | |
12381 | fp_offset = args_to_push; | |
12382 | args_to_push = 0; | |
12383 | ||
12384 | /* Now reuse r3 to preserve IP. */ | |
0083a3b9 | 12385 | emit_set_insn (gen_rtx_REG (SImode, 3), ip_rtx); |
9966e02f | 12386 | } |
209368fd | 12387 | } |
12388 | ||
0083a3b9 | 12389 | insn = emit_set_insn (ip_rtx, |
12390 | plus_constant (stack_pointer_rtx, fp_offset)); | |
5209dc22 | 12391 | RTX_FRAME_RELATED_P (insn) = 1; |
4b6897e9 | 12392 | } |
12393 | ||
9966e02f | 12394 | if (args_to_push) |
4b6897e9 | 12395 | { |
e27ad2d5 | 12396 | /* Push the argument registers, or reserve space for them. */ |
71904594 | 12397 | if (cfun->machine->uses_anonymous_args) |
220d204b | 12398 | insn = emit_multi_reg_push |
9966e02f | 12399 | ((0xf0 >> (args_to_push / 4)) & 0xf); |
4b6897e9 | 12400 | else |
220d204b | 12401 | insn = emit_insn |
9e7454d0 | 12402 | (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, |
9966e02f | 12403 | GEN_INT (- args_to_push))); |
220d204b | 12404 | RTX_FRAME_RELATED_P (insn) = 1; |
4b6897e9 | 12405 | } |
12406 | ||
000c5098 | 12407 | /* If this is an interrupt service routine, and the link register |
12408 | is going to be pushed, and we are not creating a stack frame, | |
12409 | (which would involve an extra push of IP and a pop in the epilogue) | |
12410 | subtracting four from LR now will mean that the function return | |
12411 | can be done with a single instruction. */ | |
4a5f62e2 | 12412 | if ((func_type == ARM_FT_ISR || func_type == ARM_FT_FIQ) |
000c5098 | 12413 | && (live_regs_mask & (1 << LR_REGNUM)) != 0 |
8b438b4f | 12414 | && ! frame_pointer_needed |
12415 | && TARGET_ARM) | |
0083a3b9 | 12416 | { |
12417 | rtx lr = gen_rtx_REG (SImode, LR_REGNUM); | |
12418 | ||
12419 | emit_set_insn (lr, plus_constant (lr, -4)); | |
12420 | } | |
4a5f62e2 | 12421 | |
4b6897e9 | 12422 | if (live_regs_mask) |
12423 | { | |
220d204b | 12424 | insn = emit_multi_reg_push (live_regs_mask); |
f9273c43 | 12425 | saved_regs += bit_count (live_regs_mask) * 4; |
220d204b | 12426 | RTX_FRAME_RELATED_P (insn) = 1; |
4b6897e9 | 12427 | } |
cffb2a26 | 12428 | |
e27ad2d5 | 12429 | if (! IS_VOLATILE (func_type)) |
25f905c2 | 12430 | saved_regs += arm_save_coproc_regs (); |
f21249d5 | 12431 | |
25f905c2 | 12432 | if (frame_pointer_needed && TARGET_ARM) |
12433 | { | |
12434 | /* Create the new frame pointer. */ | |
a2cd141b | 12435 | { |
25f905c2 | 12436 | insn = GEN_INT (-(4 + args_to_push + fp_offset)); |
12437 | insn = emit_insn (gen_addsi3 (hard_frame_pointer_rtx, ip_rtx, insn)); | |
12438 | RTX_FRAME_RELATED_P (insn) = 1; | |
a2cd141b | 12439 | |
25f905c2 | 12440 | if (IS_NESTED (func_type)) |
a2cd141b | 12441 | { |
25f905c2 | 12442 | /* Recover the static chain register. */ |
3072d30e | 12443 | if (!df_regs_ever_live_p (3) |
25f905c2 | 12444 | || saved_pretend_args) |
12445 | insn = gen_rtx_REG (SImode, 3); | |
12446 | else /* if (current_function_pretend_args_size == 0) */ | |
a2cd141b | 12447 | { |
25f905c2 | 12448 | insn = plus_constant (hard_frame_pointer_rtx, 4); |
12449 | insn = gen_frame_mem (SImode, insn); | |
a2cd141b | 12450 | } |
25f905c2 | 12451 | emit_set_insn (ip_rtx, insn); |
12452 | /* Add a USE to stop propagate_one_insn() from barfing. */ | |
12453 | emit_insn (gen_prologue_use (ip_rtx)); | |
a2cd141b | 12454 | } |
209368fd | 12455 | } |
220d204b | 12456 | } |
4b6897e9 | 12457 | |
f9273c43 | 12458 | offsets = arm_get_frame_offsets (); |
12459 | if (offsets->outgoing_args != offsets->saved_args + saved_regs) | |
4b6897e9 | 12460 | { |
4f8f8a17 | 12461 | /* This add can produce multiple insns for a large constant, so we |
12462 | need to get tricky. */ | |
12463 | rtx last = get_last_insn (); | |
f9273c43 | 12464 | |
12465 | amount = GEN_INT (offsets->saved_args + saved_regs | |
12466 | - offsets->outgoing_args); | |
12467 | ||
220d204b | 12468 | insn = emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, |
12469 | amount)); | |
4f8f8a17 | 12470 | do |
12471 | { | |
12472 | last = last ? NEXT_INSN (last) : get_insns (); | |
12473 | RTX_FRAME_RELATED_P (last) = 1; | |
12474 | } | |
12475 | while (last != insn); | |
3060ab98 | 12476 | |
12477 | /* If the frame pointer is needed, emit a special barrier that | |
12478 | will prevent the scheduler from moving stores to the frame | |
12479 | before the stack adjustment. */ | |
12480 | if (frame_pointer_needed) | |
4c58c898 | 12481 | insn = emit_insn (gen_stack_tie (stack_pointer_rtx, |
12482 | hard_frame_pointer_rtx)); | |
4b6897e9 | 12483 | } |
12484 | ||
ea32a06e | 12485 | |
25f905c2 | 12486 | if (frame_pointer_needed && TARGET_THUMB2) |
12487 | thumb_set_frame_pointer (offsets); | |
12488 | ||
2cb7d577 | 12489 | if (flag_pic && arm_pic_register != INVALID_REGNUM) |
25f905c2 | 12490 | { |
12491 | unsigned long mask; | |
12492 | ||
12493 | mask = live_regs_mask; | |
12494 | mask &= THUMB2_WORK_REGS; | |
12495 | if (!IS_NESTED (func_type)) | |
12496 | mask |= (1 << IP_REGNUM); | |
12497 | arm_load_pic_register (mask); | |
12498 | } | |
ea32a06e | 12499 | |
4b6897e9 | 12500 | /* If we are profiling, make sure no instructions are scheduled before |
9888ad6d | 12501 | the call to mcount. Similarly if the user has requested no |
363d487e | 12502 | scheduling in the prolog. Similarly if we want non-call exceptions |
12503 | using the EABI unwinder, to prevent faulting instructions from being | |
12504 | swapped with a stack adjustment. */ | |
12505 | if (current_function_profile || !TARGET_SCHED_PROLOG | |
12506 | || (ARM_EABI_UNWIND_TABLES && flag_non_call_exceptions)) | |
4b6897e9 | 12507 | emit_insn (gen_blockage ()); |
89b30d75 | 12508 | |
12509 | /* If the link register is being kept alive, with the return address in it, | |
12510 | then make sure that it does not get reused by the ce2 pass. */ | |
12511 | if ((live_regs_mask & (1 << LR_REGNUM)) == 0) | |
3072d30e | 12512 | cfun->machine->lr_save_eliminated = 1; |
4b6897e9 | 12513 | } |
8cd42699 | 12514 | \f |
25f905c2 | 12515 | /* Print condition code to STREAM. Helper function for arm_print_operand. */ |
12516 | static void | |
12517 | arm_print_condition (FILE *stream) | |
12518 | { | |
12519 | if (arm_ccfsm_state == 3 || arm_ccfsm_state == 4) | |
12520 | { | |
12521 | /* Branch conversion is not implemented for Thumb-2. */ | |
12522 | if (TARGET_THUMB) | |
12523 | { | |
12524 | output_operand_lossage ("predicated Thumb instruction"); | |
12525 | return; | |
12526 | } | |
12527 | if (current_insn_predicate != NULL) | |
12528 | { | |
12529 | output_operand_lossage | |
12530 | ("predicated instruction in conditional sequence"); | |
12531 | return; | |
12532 | } | |
12533 | ||
12534 | fputs (arm_condition_codes[arm_current_cc], stream); | |
12535 | } | |
12536 | else if (current_insn_predicate) | |
12537 | { | |
12538 | enum arm_cond_code code; | |
12539 | ||
12540 | if (TARGET_THUMB1) | |
12541 | { | |
12542 | output_operand_lossage ("predicated Thumb instruction"); | |
12543 | return; | |
12544 | } | |
12545 | ||
12546 | code = get_arm_condition_code (current_insn_predicate); | |
12547 | fputs (arm_condition_codes[code], stream); | |
12548 | } | |
12549 | } | |
12550 | ||
12551 | ||
f6b01912 | 12552 | /* If CODE is 'd', then the X is a condition operand and the instruction |
12553 | should only be executed if the condition is true. | |
01cc3b75 | 12554 | if CODE is 'D', then the X is a condition operand and the instruction |
f6b01912 | 12555 | should only be executed if the condition is false: however, if the mode |
12556 | of the comparison is CCFPEmode, then always execute the instruction -- we | |
12557 | do this because in these circumstances !GE does not necessarily imply LT; | |
12558 | in these cases the instruction pattern will take care to make sure that | |
12559 | an instruction containing %d will follow, thereby undoing the effects of | |
01cc3b75 | 12560 | doing this instruction unconditionally. |
f6b01912 | 12561 | If CODE is 'N' then X is a floating point operand that must be negated |
12562 | before output. | |
12563 | If CODE is 'B' then output a bitwise inverted value of X (a const int). | |
12564 | If X is a REG and CODE is `M', output a ldm/stm style multi-reg. */ | |
f6b01912 | 12565 | void |
ebd88a36 | 12566 | arm_print_operand (FILE *stream, rtx x, int code) |
f6b01912 | 12567 | { |
12568 | switch (code) | |
12569 | { | |
12570 | case '@': | |
fdb4a2ea | 12571 | fputs (ASM_COMMENT_START, stream); |
f6b01912 | 12572 | return; |
12573 | ||
cffb2a26 | 12574 | case '_': |
12575 | fputs (user_label_prefix, stream); | |
12576 | return; | |
9e7454d0 | 12577 | |
f6b01912 | 12578 | case '|': |
fdb4a2ea | 12579 | fputs (REGISTER_PREFIX, stream); |
f6b01912 | 12580 | return; |
12581 | ||
12582 | case '?': | |
25f905c2 | 12583 | arm_print_condition (stream); |
12584 | return; | |
0d66636f | 12585 | |
25f905c2 | 12586 | case '(': |
12587 | /* Nothing in unified syntax, otherwise the current condition code. */ | |
12588 | if (!TARGET_UNIFIED_ASM) | |
12589 | arm_print_condition (stream); | |
12590 | break; | |
12591 | ||
12592 | case ')': | |
12593 | /* The current condition code in unified syntax, otherwise nothing. */ | |
12594 | if (TARGET_UNIFIED_ASM) | |
12595 | arm_print_condition (stream); | |
12596 | break; | |
12597 | ||
12598 | case '.': | |
12599 | /* The current condition code for a condition code setting instruction. | |
542d5028 | 12600 | Preceded by 's' in unified syntax, otherwise followed by 's'. */ |
25f905c2 | 12601 | if (TARGET_UNIFIED_ASM) |
12602 | { | |
12603 | fputc('s', stream); | |
12604 | arm_print_condition (stream); | |
0d66636f | 12605 | } |
25f905c2 | 12606 | else |
0d66636f | 12607 | { |
25f905c2 | 12608 | arm_print_condition (stream); |
12609 | fputc('s', stream); | |
0d66636f | 12610 | } |
f6b01912 | 12611 | return; |
12612 | ||
25f905c2 | 12613 | case '!': |
12614 | /* If the instruction is conditionally executed then print | |
12615 | the current condition code, otherwise print 's'. */ | |
12616 | gcc_assert (TARGET_THUMB2 && TARGET_UNIFIED_ASM); | |
12617 | if (current_insn_predicate) | |
12618 | arm_print_condition (stream); | |
12619 | else | |
12620 | fputc('s', stream); | |
12621 | break; | |
12622 | ||
d98a3884 | 12623 | /* %# is a "break" sequence. It doesn't output anything, but is used to |
12624 | seperate e.g. operand numbers from following text, if that text consists | |
12625 | of further digits which we don't want to be part of the operand | |
12626 | number. */ | |
12627 | case '#': | |
12628 | return; | |
12629 | ||
f6b01912 | 12630 | case 'N': |
12631 | { | |
12632 | REAL_VALUE_TYPE r; | |
12633 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); | |
12634 | r = REAL_VALUE_NEGATE (r); | |
12635 | fprintf (stream, "%s", fp_const_from_val (&r)); | |
12636 | } | |
12637 | return; | |
12638 | ||
d98a3884 | 12639 | /* An integer without a preceding # sign. */ |
12640 | case 'c': | |
12641 | gcc_assert (GET_CODE (x) == CONST_INT); | |
12642 | fprintf (stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (x)); | |
12643 | return; | |
12644 | ||
f6b01912 | 12645 | case 'B': |
12646 | if (GET_CODE (x) == CONST_INT) | |
f13e56bb | 12647 | { |
12648 | HOST_WIDE_INT val; | |
215b30b3 | 12649 | val = ARM_SIGN_EXTEND (~INTVAL (x)); |
b7ecc820 | 12650 | fprintf (stream, HOST_WIDE_INT_PRINT_DEC, val); |
f13e56bb | 12651 | } |
f6b01912 | 12652 | else |
12653 | { | |
12654 | putc ('~', stream); | |
12655 | output_addr_const (stream, x); | |
12656 | } | |
12657 | return; | |
12658 | ||
25f905c2 | 12659 | case 'L': |
12660 | /* The low 16 bits of an immediate constant. */ | |
12661 | fprintf (stream, HOST_WIDE_INT_PRINT_DEC, INTVAL(x) & 0xffff); | |
12662 | return; | |
12663 | ||
f6b01912 | 12664 | case 'i': |
12665 | fprintf (stream, "%s", arithmetic_instr (x, 1)); | |
12666 | return; | |
12667 | ||
7d57ec45 | 12668 | /* Truncate Cirrus shift counts. */ |
12669 | case 's': | |
12670 | if (GET_CODE (x) == CONST_INT) | |
12671 | { | |
12672 | fprintf (stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (x) & 0x3f); | |
12673 | return; | |
12674 | } | |
12675 | arm_print_operand (stream, x, 0); | |
12676 | return; | |
12677 | ||
f6b01912 | 12678 | case 'I': |
12679 | fprintf (stream, "%s", arithmetic_instr (x, 0)); | |
12680 | return; | |
12681 | ||
12682 | case 'S': | |
12683 | { | |
12684 | HOST_WIDE_INT val; | |
26dff611 | 12685 | const char *shift; |
12686 | ||
12687 | if (!shift_operator (x, SImode)) | |
12688 | { | |
12689 | output_operand_lossage ("invalid shift operand"); | |
12690 | break; | |
12691 | } | |
12692 | ||
12693 | shift = shift_op (x, &val); | |
f6b01912 | 12694 | |
4b6897e9 | 12695 | if (shift) |
12696 | { | |
26dff611 | 12697 | fprintf (stream, ", %s ", shift); |
4b6897e9 | 12698 | if (val == -1) |
12699 | arm_print_operand (stream, XEXP (x, 1), 0); | |
12700 | else | |
4840a03a | 12701 | fprintf (stream, "#" HOST_WIDE_INT_PRINT_DEC, val); |
4b6897e9 | 12702 | } |
f6b01912 | 12703 | } |
12704 | return; | |
12705 | ||
cffb2a26 | 12706 | /* An explanation of the 'Q', 'R' and 'H' register operands: |
9e7454d0 | 12707 | |
cffb2a26 | 12708 | In a pair of registers containing a DI or DF value the 'Q' |
12709 | operand returns the register number of the register containing | |
5910bb95 | 12710 | the least significant part of the value. The 'R' operand returns |
cffb2a26 | 12711 | the register number of the register containing the most |
12712 | significant part of the value. | |
9e7454d0 | 12713 | |
cffb2a26 | 12714 | The 'H' operand returns the higher of the two register numbers. |
12715 | On a run where WORDS_BIG_ENDIAN is true the 'H' operand is the | |
5910bb95 | 12716 | same as the 'Q' operand, since the most significant part of the |
cffb2a26 | 12717 | value is held in the lower number register. The reverse is true |
12718 | on systems where WORDS_BIG_ENDIAN is false. | |
9e7454d0 | 12719 | |
cffb2a26 | 12720 | The purpose of these operands is to distinguish between cases |
12721 | where the endian-ness of the values is important (for example | |
12722 | when they are added together), and cases where the endian-ness | |
12723 | is irrelevant, but the order of register operations is important. | |
12724 | For example when loading a value from memory into a register | |
12725 | pair, the endian-ness does not matter. Provided that the value | |
12726 | from the lower memory address is put into the lower numbered | |
12727 | register, and the value from the higher address is put into the | |
12728 | higher numbered register, the load will work regardless of whether | |
12729 | the value being loaded is big-wordian or little-wordian. The | |
12730 | order of the two register loads can matter however, if the address | |
12731 | of the memory location is actually held in one of the registers | |
12732 | being overwritten by the load. */ | |
cb88963f | 12733 | case 'Q': |
47378eff | 12734 | if (GET_CODE (x) != REG || REGNO (x) > LAST_ARM_REGNUM) |
12735 | { | |
12736 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12737 | return; | |
12738 | } | |
12739 | ||
cffb2a26 | 12740 | asm_fprintf (stream, "%r", REGNO (x) + (WORDS_BIG_ENDIAN ? 1 : 0)); |
cb88963f | 12741 | return; |
12742 | ||
f6b01912 | 12743 | case 'R': |
47378eff | 12744 | if (GET_CODE (x) != REG || REGNO (x) > LAST_ARM_REGNUM) |
12745 | { | |
12746 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12747 | return; | |
12748 | } | |
12749 | ||
cffb2a26 | 12750 | asm_fprintf (stream, "%r", REGNO (x) + (WORDS_BIG_ENDIAN ? 0 : 1)); |
12751 | return; | |
12752 | ||
12753 | case 'H': | |
47378eff | 12754 | if (GET_CODE (x) != REG || REGNO (x) > LAST_ARM_REGNUM) |
12755 | { | |
12756 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12757 | return; | |
12758 | } | |
12759 | ||
cffb2a26 | 12760 | asm_fprintf (stream, "%r", REGNO (x) + 1); |
f6b01912 | 12761 | return; |
12762 | ||
d98a3884 | 12763 | case 'J': |
12764 | if (GET_CODE (x) != REG || REGNO (x) > LAST_ARM_REGNUM) | |
12765 | { | |
12766 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12767 | return; | |
12768 | } | |
12769 | ||
12770 | asm_fprintf (stream, "%r", REGNO (x) + (WORDS_BIG_ENDIAN ? 3 : 2)); | |
12771 | return; | |
12772 | ||
12773 | case 'K': | |
12774 | if (GET_CODE (x) != REG || REGNO (x) > LAST_ARM_REGNUM) | |
12775 | { | |
12776 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12777 | return; | |
12778 | } | |
12779 | ||
12780 | asm_fprintf (stream, "%r", REGNO (x) + (WORDS_BIG_ENDIAN ? 2 : 3)); | |
12781 | return; | |
12782 | ||
f6b01912 | 12783 | case 'm': |
9e7454d0 | 12784 | asm_fprintf (stream, "%r", |
cffb2a26 | 12785 | GET_CODE (XEXP (x, 0)) == REG |
12786 | ? REGNO (XEXP (x, 0)) : REGNO (XEXP (XEXP (x, 0), 0))); | |
f6b01912 | 12787 | return; |
12788 | ||
12789 | case 'M': | |
caceeb43 | 12790 | asm_fprintf (stream, "{%r-%r}", |
cffb2a26 | 12791 | REGNO (x), |
36837fde | 12792 | REGNO (x) + ARM_NUM_REGS (GET_MODE (x)) - 1); |
f6b01912 | 12793 | return; |
12794 | ||
d98a3884 | 12795 | /* Like 'M', but writing doubleword vector registers, for use by Neon |
12796 | insns. */ | |
12797 | case 'h': | |
12798 | { | |
12799 | int regno = (REGNO (x) - FIRST_VFP_REGNUM) / 2; | |
12800 | int numregs = ARM_NUM_REGS (GET_MODE (x)) / 2; | |
12801 | if (numregs == 1) | |
12802 | asm_fprintf (stream, "{d%d}", regno); | |
12803 | else | |
12804 | asm_fprintf (stream, "{d%d-d%d}", regno, regno + numregs - 1); | |
12805 | } | |
12806 | return; | |
12807 | ||
f6b01912 | 12808 | case 'd': |
5db468b7 | 12809 | /* CONST_TRUE_RTX means always -- that's the default. */ |
12810 | if (x == const_true_rtx) | |
cffb2a26 | 12811 | return; |
9e7454d0 | 12812 | |
47378eff | 12813 | if (!COMPARISON_P (x)) |
12814 | { | |
12815 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12816 | return; | |
12817 | } | |
12818 | ||
58d6528b | 12819 | fputs (arm_condition_codes[get_arm_condition_code (x)], |
12820 | stream); | |
f6b01912 | 12821 | return; |
12822 | ||
12823 | case 'D': | |
a361b456 | 12824 | /* CONST_TRUE_RTX means not always -- i.e. never. We shouldn't ever |
5db468b7 | 12825 | want to do that. */ |
12826 | if (x == const_true_rtx) | |
47378eff | 12827 | { |
666137bc | 12828 | output_operand_lossage ("instruction never executed"); |
47378eff | 12829 | return; |
12830 | } | |
12831 | if (!COMPARISON_P (x)) | |
12832 | { | |
12833 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12834 | return; | |
12835 | } | |
cffb2a26 | 12836 | |
58d6528b | 12837 | fputs (arm_condition_codes[ARM_INVERSE_CONDITION_CODE |
12838 | (get_arm_condition_code (x))], | |
12839 | stream); | |
f6b01912 | 12840 | return; |
12841 | ||
7d57ec45 | 12842 | /* Cirrus registers can be accessed in a variety of ways: |
12843 | single floating point (f) | |
12844 | double floating point (d) | |
12845 | 32bit integer (fx) | |
12846 | 64bit integer (dx). */ | |
12847 | case 'W': /* Cirrus register in F mode. */ | |
12848 | case 'X': /* Cirrus register in D mode. */ | |
12849 | case 'Y': /* Cirrus register in FX mode. */ | |
12850 | case 'Z': /* Cirrus register in DX mode. */ | |
ed29c566 | 12851 | gcc_assert (GET_CODE (x) == REG |
12852 | && REGNO_REG_CLASS (REGNO (x)) == CIRRUS_REGS); | |
7d57ec45 | 12853 | |
12854 | fprintf (stream, "mv%s%s", | |
12855 | code == 'W' ? "f" | |
12856 | : code == 'X' ? "d" | |
12857 | : code == 'Y' ? "fx" : "dx", reg_names[REGNO (x)] + 2); | |
12858 | ||
12859 | return; | |
12860 | ||
12861 | /* Print cirrus register in the mode specified by the register's mode. */ | |
12862 | case 'V': | |
12863 | { | |
12864 | int mode = GET_MODE (x); | |
12865 | ||
12866 | if (GET_CODE (x) != REG || REGNO_REG_CLASS (REGNO (x)) != CIRRUS_REGS) | |
47378eff | 12867 | { |
12868 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12869 | return; | |
12870 | } | |
7d57ec45 | 12871 | |
12872 | fprintf (stream, "mv%s%s", | |
12873 | mode == DFmode ? "d" | |
12874 | : mode == SImode ? "fx" | |
12875 | : mode == DImode ? "dx" | |
12876 | : "f", reg_names[REGNO (x)] + 2); | |
12877 | ||
12878 | return; | |
12879 | } | |
12880 | ||
755eb2b4 | 12881 | case 'U': |
12882 | if (GET_CODE (x) != REG | |
12883 | || REGNO (x) < FIRST_IWMMXT_GR_REGNUM | |
12884 | || REGNO (x) > LAST_IWMMXT_GR_REGNUM) | |
12885 | /* Bad value for wCG register number. */ | |
47378eff | 12886 | { |
12887 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12888 | return; | |
12889 | } | |
12890 | ||
755eb2b4 | 12891 | else |
12892 | fprintf (stream, "%d", REGNO (x) - FIRST_IWMMXT_GR_REGNUM); | |
12893 | return; | |
12894 | ||
12895 | /* Print an iWMMXt control register name. */ | |
12896 | case 'w': | |
12897 | if (GET_CODE (x) != CONST_INT | |
12898 | || INTVAL (x) < 0 | |
12899 | || INTVAL (x) >= 16) | |
12900 | /* Bad value for wC register number. */ | |
47378eff | 12901 | { |
12902 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12903 | return; | |
12904 | } | |
12905 | ||
755eb2b4 | 12906 | else |
12907 | { | |
12908 | static const char * wc_reg_names [16] = | |
12909 | { | |
12910 | "wCID", "wCon", "wCSSF", "wCASF", | |
12911 | "wC4", "wC5", "wC6", "wC7", | |
12912 | "wCGR0", "wCGR1", "wCGR2", "wCGR3", | |
12913 | "wC12", "wC13", "wC14", "wC15" | |
12914 | }; | |
9e7454d0 | 12915 | |
755eb2b4 | 12916 | fprintf (stream, wc_reg_names [INTVAL (x)]); |
12917 | } | |
12918 | return; | |
12919 | ||
d98a3884 | 12920 | /* Print a VFP/Neon double precision or quad precision register name. */ |
a2cd141b | 12921 | case 'P': |
d98a3884 | 12922 | case 'q': |
a2cd141b | 12923 | { |
12924 | int mode = GET_MODE (x); | |
d98a3884 | 12925 | int is_quad = (code == 'q'); |
12926 | int regno; | |
a2cd141b | 12927 | |
d98a3884 | 12928 | if (GET_MODE_SIZE (mode) != (is_quad ? 16 : 8)) |
47378eff | 12929 | { |
12930 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12931 | return; | |
12932 | } | |
a2cd141b | 12933 | |
12934 | if (GET_CODE (x) != REG | |
12935 | || !IS_VFP_REGNUM (REGNO (x))) | |
47378eff | 12936 | { |
12937 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12938 | return; | |
12939 | } | |
a2cd141b | 12940 | |
d98a3884 | 12941 | regno = REGNO (x); |
12942 | if ((is_quad && !NEON_REGNO_OK_FOR_QUAD (regno)) | |
12943 | || (!is_quad && !VFP_REGNO_OK_FOR_DOUBLE (regno))) | |
47378eff | 12944 | { |
12945 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12946 | return; | |
12947 | } | |
a2cd141b | 12948 | |
d98a3884 | 12949 | fprintf (stream, "%c%d", is_quad ? 'q' : 'd', |
12950 | (regno - FIRST_VFP_REGNUM) >> (is_quad ? 2 : 1)); | |
12951 | } | |
12952 | return; | |
12953 | ||
12954 | /* These two codes print the low/high doubleword register of a Neon quad | |
12955 | register, respectively. For pair-structure types, can also print | |
12956 | low/high quadword registers. */ | |
12957 | case 'e': | |
12958 | case 'f': | |
12959 | { | |
12960 | int mode = GET_MODE (x); | |
12961 | int regno; | |
12962 | ||
12963 | if ((GET_MODE_SIZE (mode) != 16 | |
12964 | && GET_MODE_SIZE (mode) != 32) || GET_CODE (x) != REG) | |
12965 | { | |
12966 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12967 | return; | |
12968 | } | |
12969 | ||
12970 | regno = REGNO (x); | |
12971 | if (!NEON_REGNO_OK_FOR_QUAD (regno)) | |
12972 | { | |
12973 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12974 | return; | |
12975 | } | |
12976 | ||
12977 | if (GET_MODE_SIZE (mode) == 16) | |
12978 | fprintf (stream, "d%d", ((regno - FIRST_VFP_REGNUM) >> 1) | |
12979 | + (code == 'f' ? 1 : 0)); | |
12980 | else | |
12981 | fprintf (stream, "q%d", ((regno - FIRST_VFP_REGNUM) >> 2) | |
12982 | + (code == 'f' ? 1 : 0)); | |
a2cd141b | 12983 | } |
12984 | return; | |
12985 | ||
b5a0636d | 12986 | /* Print a VFPv3 floating-point constant, represented as an integer |
12987 | index. */ | |
12988 | case 'G': | |
12989 | { | |
12990 | int index = vfp3_const_double_index (x); | |
12991 | gcc_assert (index != -1); | |
12992 | fprintf (stream, "%d", index); | |
12993 | } | |
12994 | return; | |
12995 | ||
d98a3884 | 12996 | /* Print bits representing opcode features for Neon. |
12997 | ||
12998 | Bit 0 is 1 for signed, 0 for unsigned. Floats count as signed | |
12999 | and polynomials as unsigned. | |
13000 | ||
13001 | Bit 1 is 1 for floats and polynomials, 0 for ordinary integers. | |
13002 | ||
13003 | Bit 2 is 1 for rounding functions, 0 otherwise. */ | |
13004 | ||
13005 | /* Identify the type as 's', 'u', 'p' or 'f'. */ | |
13006 | case 'T': | |
13007 | { | |
13008 | HOST_WIDE_INT bits = INTVAL (x); | |
13009 | fputc ("uspf"[bits & 3], stream); | |
13010 | } | |
13011 | return; | |
13012 | ||
13013 | /* Likewise, but signed and unsigned integers are both 'i'. */ | |
13014 | case 'F': | |
13015 | { | |
13016 | HOST_WIDE_INT bits = INTVAL (x); | |
13017 | fputc ("iipf"[bits & 3], stream); | |
13018 | } | |
13019 | return; | |
13020 | ||
13021 | /* As for 'T', but emit 'u' instead of 'p'. */ | |
13022 | case 't': | |
13023 | { | |
13024 | HOST_WIDE_INT bits = INTVAL (x); | |
13025 | fputc ("usuf"[bits & 3], stream); | |
13026 | } | |
13027 | return; | |
13028 | ||
13029 | /* Bit 2: rounding (vs none). */ | |
13030 | case 'O': | |
13031 | { | |
13032 | HOST_WIDE_INT bits = INTVAL (x); | |
13033 | fputs ((bits & 4) != 0 ? "r" : "", stream); | |
13034 | } | |
13035 | return; | |
13036 | ||
f6b01912 | 13037 | default: |
13038 | if (x == 0) | |
47378eff | 13039 | { |
13040 | output_operand_lossage ("missing operand"); | |
13041 | return; | |
13042 | } | |
f6b01912 | 13043 | |
ed29c566 | 13044 | switch (GET_CODE (x)) |
f6b01912 | 13045 | { |
ed29c566 | 13046 | case REG: |
13047 | asm_fprintf (stream, "%r", REGNO (x)); | |
13048 | break; | |
13049 | ||
13050 | case MEM: | |
f6b01912 | 13051 | output_memory_reference_mode = GET_MODE (x); |
13052 | output_address (XEXP (x, 0)); | |
ed29c566 | 13053 | break; |
13054 | ||
13055 | case CONST_DOUBLE: | |
d98a3884 | 13056 | if (TARGET_NEON) |
13057 | { | |
13058 | char fpstr[20]; | |
13059 | real_to_decimal (fpstr, CONST_DOUBLE_REAL_VALUE (x), | |
13060 | sizeof (fpstr), 0, 1); | |
13061 | fprintf (stream, "#%s", fpstr); | |
13062 | } | |
13063 | else | |
13064 | fprintf (stream, "#%s", fp_immediate_constant (x)); | |
ed29c566 | 13065 | break; |
13066 | ||
13067 | default: | |
13068 | gcc_assert (GET_CODE (x) != NEG); | |
f6b01912 | 13069 | fputc ('#', stream); |
13070 | output_addr_const (stream, x); | |
ed29c566 | 13071 | break; |
f6b01912 | 13072 | } |
13073 | } | |
13074 | } | |
8cd42699 | 13075 | \f |
58356836 | 13076 | #ifndef AOF_ASSEMBLER |
13077 | /* Target hook for assembling integer objects. The ARM version needs to | |
13078 | handle word-sized values specially. */ | |
58356836 | 13079 | static bool |
ebd88a36 | 13080 | arm_assemble_integer (rtx x, unsigned int size, int aligned_p) |
58356836 | 13081 | { |
d98a3884 | 13082 | enum machine_mode mode; |
13083 | ||
58356836 | 13084 | if (size == UNITS_PER_WORD && aligned_p) |
13085 | { | |
13086 | fputs ("\t.word\t", asm_out_file); | |
13087 | output_addr_const (asm_out_file, x); | |
13088 | ||
13089 | /* Mark symbols as position independent. We only do this in the | |
674a8f0b | 13090 | .text segment, not in the .data segment. */ |
58356836 | 13091 | if (NEED_GOT_RELOC && flag_pic && making_const_table && |
13092 | (GET_CODE (x) == SYMBOL_REF || GET_CODE (x) == LABEL_REF)) | |
13093 | { | |
bac7fc85 | 13094 | /* See legitimize_pic_address for an explanation of the |
13095 | TARGET_VXWORKS_RTP check. */ | |
13096 | if (TARGET_VXWORKS_RTP | |
13097 | || (GET_CODE (x) == SYMBOL_REF && !SYMBOL_REF_LOCAL_P (x))) | |
58356836 | 13098 | fputs ("(GOT)", asm_out_file); |
bac7fc85 | 13099 | else |
13100 | fputs ("(GOTOFF)", asm_out_file); | |
58356836 | 13101 | } |
13102 | fputc ('\n', asm_out_file); | |
13103 | return true; | |
13104 | } | |
a12be13c | 13105 | |
d98a3884 | 13106 | mode = GET_MODE (x); |
13107 | ||
13108 | if (arm_vector_mode_supported_p (mode)) | |
755eb2b4 | 13109 | { |
13110 | int i, units; | |
d98a3884 | 13111 | unsigned int invmask = 0, parts_per_word; |
755eb2b4 | 13112 | |
ed29c566 | 13113 | gcc_assert (GET_CODE (x) == CONST_VECTOR); |
755eb2b4 | 13114 | |
13115 | units = CONST_VECTOR_NUNITS (x); | |
d98a3884 | 13116 | size = GET_MODE_SIZE (GET_MODE_INNER (mode)); |
755eb2b4 | 13117 | |
d98a3884 | 13118 | /* For big-endian Neon vectors, we must permute the vector to the form |
13119 | which, when loaded by a VLDR or VLDM instruction, will give a vector | |
13120 | with the elements in the right order. */ | |
13121 | if (TARGET_NEON && WORDS_BIG_ENDIAN) | |
13122 | { | |
13123 | parts_per_word = UNITS_PER_WORD / size; | |
13124 | /* FIXME: This might be wrong for 64-bit vector elements, but we don't | |
13125 | support those anywhere yet. */ | |
13126 | invmask = (parts_per_word == 0) ? 0 : (1 << (parts_per_word - 1)) - 1; | |
13127 | } | |
755eb2b4 | 13128 | |
d98a3884 | 13129 | if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT) |
13130 | for (i = 0; i < units; i++) | |
13131 | { | |
13132 | rtx elt = CONST_VECTOR_ELT (x, i ^ invmask); | |
13133 | assemble_integer | |
13134 | (elt, size, i == 0 ? BIGGEST_ALIGNMENT : size * BITS_PER_UNIT, 1); | |
13135 | } | |
13136 | else | |
13137 | for (i = 0; i < units; i++) | |
13138 | { | |
13139 | rtx elt = CONST_VECTOR_ELT (x, i); | |
13140 | REAL_VALUE_TYPE rval; | |
755eb2b4 | 13141 | |
d98a3884 | 13142 | REAL_VALUE_FROM_CONST_DOUBLE (rval, elt); |
13143 | ||
13144 | assemble_real | |
13145 | (rval, GET_MODE_INNER (mode), | |
13146 | i == 0 ? BIGGEST_ALIGNMENT : size * BITS_PER_UNIT); | |
13147 | } | |
755eb2b4 | 13148 | |
13149 | return true; | |
13150 | } | |
13151 | ||
58356836 | 13152 | return default_assemble_integer (x, size, aligned_p); |
13153 | } | |
d24bc145 | 13154 | |
d24bc145 | 13155 | static void |
a457f9f8 | 13156 | arm_elf_asm_cdtor (rtx symbol, int priority, bool is_ctor) |
d24bc145 | 13157 | { |
df816ae6 | 13158 | section *s; |
13159 | ||
d24bc145 | 13160 | if (!TARGET_AAPCS_BASED) |
13161 | { | |
a457f9f8 | 13162 | (is_ctor ? |
13163 | default_named_section_asm_out_constructor | |
13164 | : default_named_section_asm_out_destructor) (symbol, priority); | |
d24bc145 | 13165 | return; |
13166 | } | |
13167 | ||
13168 | /* Put these in the .init_array section, using a special relocation. */ | |
df816ae6 | 13169 | if (priority != DEFAULT_INIT_PRIORITY) |
13170 | { | |
13171 | char buf[18]; | |
a457f9f8 | 13172 | sprintf (buf, "%s.%.5u", |
13173 | is_ctor ? ".init_array" : ".fini_array", | |
13174 | priority); | |
df816ae6 | 13175 | s = get_section (buf, SECTION_WRITE, NULL_TREE); |
13176 | } | |
a457f9f8 | 13177 | else if (is_ctor) |
df816ae6 | 13178 | s = ctors_section; |
a457f9f8 | 13179 | else |
13180 | s = dtors_section; | |
df816ae6 | 13181 | |
13182 | switch_to_section (s); | |
d24bc145 | 13183 | assemble_align (POINTER_SIZE); |
13184 | fputs ("\t.word\t", asm_out_file); | |
13185 | output_addr_const (asm_out_file, symbol); | |
13186 | fputs ("(target1)\n", asm_out_file); | |
13187 | } | |
a457f9f8 | 13188 | |
13189 | /* Add a function to the list of static constructors. */ | |
13190 | ||
13191 | static void | |
13192 | arm_elf_asm_constructor (rtx symbol, int priority) | |
13193 | { | |
13194 | arm_elf_asm_cdtor (symbol, priority, /*is_ctor=*/true); | |
13195 | } | |
13196 | ||
13197 | /* Add a function to the list of static destructors. */ | |
13198 | ||
13199 | static void | |
13200 | arm_elf_asm_destructor (rtx symbol, int priority) | |
13201 | { | |
13202 | arm_elf_asm_cdtor (symbol, priority, /*is_ctor=*/false); | |
13203 | } | |
58356836 | 13204 | #endif |
13205 | \f | |
8cd42699 | 13206 | /* A finite state machine takes care of noticing whether or not instructions |
13207 | can be conditionally executed, and thus decrease execution time and code | |
13208 | size by deleting branch instructions. The fsm is controlled by | |
13209 | final_prescan_insn, and controls the actions of ASM_OUTPUT_OPCODE. */ | |
13210 | ||
13211 | /* The state of the fsm controlling condition codes are: | |
13212 | 0: normal, do nothing special | |
13213 | 1: make ASM_OUTPUT_OPCODE not output this instruction | |
13214 | 2: make ASM_OUTPUT_OPCODE not output this instruction | |
13215 | 3: make instructions conditional | |
13216 | 4: make instructions conditional | |
13217 | ||
13218 | State transitions (state->state by whom under condition): | |
13219 | 0 -> 1 final_prescan_insn if the `target' is a label | |
13220 | 0 -> 2 final_prescan_insn if the `target' is an unconditional branch | |
13221 | 1 -> 3 ASM_OUTPUT_OPCODE after not having output the conditional branch | |
13222 | 2 -> 4 ASM_OUTPUT_OPCODE after not having output the conditional branch | |
805e22b2 | 13223 | 3 -> 0 (*targetm.asm_out.internal_label) if the `target' label is reached |
8cd42699 | 13224 | (the target label has CODE_LABEL_NUMBER equal to arm_target_label). |
13225 | 4 -> 0 final_prescan_insn if the `target' unconditional branch is reached | |
13226 | (the target insn is arm_target_insn). | |
13227 | ||
9c08d1fa | 13228 | If the jump clobbers the conditions then we use states 2 and 4. |
13229 | ||
13230 | A similar thing can be done with conditional return insns. | |
13231 | ||
8cd42699 | 13232 | XXX In case the `target' is an unconditional branch, this conditionalising |
13233 | of the instructions always reduces code size, but not always execution | |
13234 | time. But then, I want to reduce the code size to somewhere near what | |
13235 | /bin/cc produces. */ | |
13236 | ||
25f905c2 | 13237 | /* In addition to this, state is maintained for Thumb-2 COND_EXEC |
13238 | instructions. When a COND_EXEC instruction is seen the subsequent | |
13239 | instructions are scanned so that multiple conditional instructions can be | |
13240 | combined into a single IT block. arm_condexec_count and arm_condexec_mask | |
13241 | specify the length and true/false mask for the IT block. These will be | |
13242 | decremented/zeroed by arm_asm_output_opcode as the insns are output. */ | |
13243 | ||
8cd42699 | 13244 | /* Returns the index of the ARM condition code string in |
13245 | `arm_condition_codes'. COMPARISON should be an rtx like | |
13246 | `(eq (...) (...))'. */ | |
1275b964 | 13247 | static enum arm_cond_code |
ebd88a36 | 13248 | get_arm_condition_code (rtx comparison) |
8cd42699 | 13249 | { |
ac577012 | 13250 | enum machine_mode mode = GET_MODE (XEXP (comparison, 0)); |
a12be13c | 13251 | int code; |
13252 | enum rtx_code comp_code = GET_CODE (comparison); | |
ac577012 | 13253 | |
13254 | if (GET_MODE_CLASS (mode) != MODE_CC) | |
1275b964 | 13255 | mode = SELECT_CC_MODE (comp_code, XEXP (comparison, 0), |
ac577012 | 13256 | XEXP (comparison, 1)); |
13257 | ||
13258 | switch (mode) | |
8cd42699 | 13259 | { |
1275b964 | 13260 | case CC_DNEmode: code = ARM_NE; goto dominance; |
13261 | case CC_DEQmode: code = ARM_EQ; goto dominance; | |
13262 | case CC_DGEmode: code = ARM_GE; goto dominance; | |
13263 | case CC_DGTmode: code = ARM_GT; goto dominance; | |
13264 | case CC_DLEmode: code = ARM_LE; goto dominance; | |
13265 | case CC_DLTmode: code = ARM_LT; goto dominance; | |
13266 | case CC_DGEUmode: code = ARM_CS; goto dominance; | |
13267 | case CC_DGTUmode: code = ARM_HI; goto dominance; | |
13268 | case CC_DLEUmode: code = ARM_LS; goto dominance; | |
13269 | case CC_DLTUmode: code = ARM_CC; | |
13270 | ||
13271 | dominance: | |
ed29c566 | 13272 | gcc_assert (comp_code == EQ || comp_code == NE); |
1275b964 | 13273 | |
13274 | if (comp_code == EQ) | |
13275 | return ARM_INVERSE_CONDITION_CODE (code); | |
13276 | return code; | |
13277 | ||
ac577012 | 13278 | case CC_NOOVmode: |
1275b964 | 13279 | switch (comp_code) |
ac577012 | 13280 | { |
1275b964 | 13281 | case NE: return ARM_NE; |
13282 | case EQ: return ARM_EQ; | |
13283 | case GE: return ARM_PL; | |
13284 | case LT: return ARM_MI; | |
ed29c566 | 13285 | default: gcc_unreachable (); |
ac577012 | 13286 | } |
13287 | ||
13288 | case CC_Zmode: | |
1275b964 | 13289 | switch (comp_code) |
ac577012 | 13290 | { |
1275b964 | 13291 | case NE: return ARM_NE; |
13292 | case EQ: return ARM_EQ; | |
ed29c566 | 13293 | default: gcc_unreachable (); |
ac577012 | 13294 | } |
13295 | ||
58d6528b | 13296 | case CC_Nmode: |
13297 | switch (comp_code) | |
13298 | { | |
13299 | case NE: return ARM_MI; | |
13300 | case EQ: return ARM_PL; | |
ed29c566 | 13301 | default: gcc_unreachable (); |
58d6528b | 13302 | } |
13303 | ||
ac577012 | 13304 | case CCFPEmode: |
f3d87a93 | 13305 | case CCFPmode: |
13306 | /* These encodings assume that AC=1 in the FPA system control | |
13307 | byte. This allows us to handle all cases except UNEQ and | |
13308 | LTGT. */ | |
1275b964 | 13309 | switch (comp_code) |
13310 | { | |
13311 | case GE: return ARM_GE; | |
13312 | case GT: return ARM_GT; | |
13313 | case LE: return ARM_LS; | |
13314 | case LT: return ARM_MI; | |
f3d87a93 | 13315 | case NE: return ARM_NE; |
13316 | case EQ: return ARM_EQ; | |
13317 | case ORDERED: return ARM_VC; | |
13318 | case UNORDERED: return ARM_VS; | |
13319 | case UNLT: return ARM_LT; | |
13320 | case UNLE: return ARM_LE; | |
13321 | case UNGT: return ARM_HI; | |
13322 | case UNGE: return ARM_PL; | |
13323 | /* UNEQ and LTGT do not have a representation. */ | |
13324 | case UNEQ: /* Fall through. */ | |
13325 | case LTGT: /* Fall through. */ | |
ed29c566 | 13326 | default: gcc_unreachable (); |
1275b964 | 13327 | } |
13328 | ||
13329 | case CC_SWPmode: | |
13330 | switch (comp_code) | |
13331 | { | |
13332 | case NE: return ARM_NE; | |
13333 | case EQ: return ARM_EQ; | |
13334 | case GE: return ARM_LE; | |
13335 | case GT: return ARM_LT; | |
13336 | case LE: return ARM_GE; | |
13337 | case LT: return ARM_GT; | |
13338 | case GEU: return ARM_LS; | |
13339 | case GTU: return ARM_CC; | |
13340 | case LEU: return ARM_CS; | |
13341 | case LTU: return ARM_HI; | |
ed29c566 | 13342 | default: gcc_unreachable (); |
1275b964 | 13343 | } |
13344 | ||
9480e6c7 | 13345 | case CC_Cmode: |
13346 | switch (comp_code) | |
13347 | { | |
13348 | case LTU: return ARM_CS; | |
13349 | case GEU: return ARM_CC; | |
ed29c566 | 13350 | default: gcc_unreachable (); |
9480e6c7 | 13351 | } |
9e7454d0 | 13352 | |
ac577012 | 13353 | case CCmode: |
1275b964 | 13354 | switch (comp_code) |
ac577012 | 13355 | { |
1275b964 | 13356 | case NE: return ARM_NE; |
13357 | case EQ: return ARM_EQ; | |
13358 | case GE: return ARM_GE; | |
13359 | case GT: return ARM_GT; | |
13360 | case LE: return ARM_LE; | |
13361 | case LT: return ARM_LT; | |
13362 | case GEU: return ARM_CS; | |
13363 | case GTU: return ARM_HI; | |
13364 | case LEU: return ARM_LS; | |
13365 | case LTU: return ARM_CC; | |
ed29c566 | 13366 | default: gcc_unreachable (); |
ac577012 | 13367 | } |
13368 | ||
ed29c566 | 13369 | default: gcc_unreachable (); |
8cd42699 | 13370 | } |
a690ede2 | 13371 | } |
8cd42699 | 13372 | |
f6a0d06f | 13373 | /* Tell arm_asm_output_opcode to output IT blocks for conditionally executed |
25f905c2 | 13374 | instructions. */ |
13375 | void | |
13376 | thumb2_final_prescan_insn (rtx insn) | |
13377 | { | |
13378 | rtx first_insn = insn; | |
13379 | rtx body = PATTERN (insn); | |
13380 | rtx predicate; | |
13381 | enum arm_cond_code code; | |
13382 | int n; | |
13383 | int mask; | |
13384 | ||
13385 | /* Remove the previous insn from the count of insns to be output. */ | |
13386 | if (arm_condexec_count) | |
13387 | arm_condexec_count--; | |
13388 | ||
13389 | /* Nothing to do if we are already inside a conditional block. */ | |
13390 | if (arm_condexec_count) | |
13391 | return; | |
13392 | ||
13393 | if (GET_CODE (body) != COND_EXEC) | |
13394 | return; | |
13395 | ||
13396 | /* Conditional jumps are implemented directly. */ | |
13397 | if (GET_CODE (insn) == JUMP_INSN) | |
13398 | return; | |
13399 | ||
13400 | predicate = COND_EXEC_TEST (body); | |
13401 | arm_current_cc = get_arm_condition_code (predicate); | |
13402 | ||
13403 | n = get_attr_ce_count (insn); | |
13404 | arm_condexec_count = 1; | |
13405 | arm_condexec_mask = (1 << n) - 1; | |
13406 | arm_condexec_masklen = n; | |
13407 | /* See if subsequent instructions can be combined into the same block. */ | |
13408 | for (;;) | |
13409 | { | |
13410 | insn = next_nonnote_insn (insn); | |
13411 | ||
13412 | /* Jumping into the middle of an IT block is illegal, so a label or | |
13413 | barrier terminates the block. */ | |
13414 | if (GET_CODE (insn) != INSN && GET_CODE(insn) != JUMP_INSN) | |
13415 | break; | |
13416 | ||
13417 | body = PATTERN (insn); | |
13418 | /* USE and CLOBBER aren't really insns, so just skip them. */ | |
13419 | if (GET_CODE (body) == USE | |
13420 | || GET_CODE (body) == CLOBBER) | |
7dd28afd | 13421 | continue; |
25f905c2 | 13422 | |
542d5028 | 13423 | /* ??? Recognize conditional jumps, and combine them with IT blocks. */ |
25f905c2 | 13424 | if (GET_CODE (body) != COND_EXEC) |
13425 | break; | |
13426 | /* Allow up to 4 conditionally executed instructions in a block. */ | |
13427 | n = get_attr_ce_count (insn); | |
13428 | if (arm_condexec_masklen + n > 4) | |
13429 | break; | |
13430 | ||
13431 | predicate = COND_EXEC_TEST (body); | |
13432 | code = get_arm_condition_code (predicate); | |
13433 | mask = (1 << n) - 1; | |
13434 | if (arm_current_cc == code) | |
13435 | arm_condexec_mask |= (mask << arm_condexec_masklen); | |
13436 | else if (arm_current_cc != ARM_INVERSE_CONDITION_CODE(code)) | |
13437 | break; | |
13438 | ||
13439 | arm_condexec_count++; | |
13440 | arm_condexec_masklen += n; | |
13441 | ||
13442 | /* A jump must be the last instruction in a conditional block. */ | |
13443 | if (GET_CODE(insn) == JUMP_INSN) | |
13444 | break; | |
13445 | } | |
13446 | /* Restore recog_data (getting the attributes of other insns can | |
13447 | destroy this array, but final.c assumes that it remains intact | |
13448 | across this call). */ | |
13449 | extract_constrain_insn_cached (first_insn); | |
13450 | } | |
13451 | ||
8cd42699 | 13452 | void |
ebd88a36 | 13453 | arm_final_prescan_insn (rtx insn) |
8cd42699 | 13454 | { |
13455 | /* BODY will hold the body of INSN. */ | |
a12be13c | 13456 | rtx body = PATTERN (insn); |
8cd42699 | 13457 | |
13458 | /* This will be 1 if trying to repeat the trick, and things need to be | |
13459 | reversed if it appears to fail. */ | |
13460 | int reverse = 0; | |
13461 | ||
9c08d1fa | 13462 | /* JUMP_CLOBBERS will be one implies that the conditions if a branch is |
13463 | taken are clobbered, even if the rtl suggests otherwise. It also | |
13464 | means that we have to grub around within the jump expression to find | |
13465 | out what the conditions are when the jump isn't taken. */ | |
13466 | int jump_clobbers = 0; | |
9e7454d0 | 13467 | |
66aacf41 | 13468 | /* If we start with a return insn, we only succeed if we find another one. */ |
9c08d1fa | 13469 | int seeking_return = 0; |
9e7454d0 | 13470 | |
8cd42699 | 13471 | /* START_INSN will hold the insn from where we start looking. This is the |
13472 | first insn after the following code_label if REVERSE is true. */ | |
13473 | rtx start_insn = insn; | |
13474 | ||
13475 | /* If in state 4, check if the target branch is reached, in order to | |
13476 | change back to state 0. */ | |
13477 | if (arm_ccfsm_state == 4) | |
13478 | { | |
13479 | if (insn == arm_target_insn) | |
9888ad6d | 13480 | { |
13481 | arm_target_insn = NULL; | |
13482 | arm_ccfsm_state = 0; | |
13483 | } | |
8cd42699 | 13484 | return; |
13485 | } | |
13486 | ||
13487 | /* If in state 3, it is possible to repeat the trick, if this insn is an | |
13488 | unconditional branch to a label, and immediately following this branch | |
13489 | is the previous target label which is only used once, and the label this | |
13490 | branch jumps to is not too far off. */ | |
13491 | if (arm_ccfsm_state == 3) | |
13492 | { | |
13493 | if (simplejump_p (insn)) | |
13494 | { | |
13495 | start_insn = next_nonnote_insn (start_insn); | |
13496 | if (GET_CODE (start_insn) == BARRIER) | |
13497 | { | |
13498 | /* XXX Isn't this always a barrier? */ | |
13499 | start_insn = next_nonnote_insn (start_insn); | |
13500 | } | |
13501 | if (GET_CODE (start_insn) == CODE_LABEL | |
13502 | && CODE_LABEL_NUMBER (start_insn) == arm_target_label | |
13503 | && LABEL_NUSES (start_insn) == 1) | |
13504 | reverse = TRUE; | |
13505 | else | |
13506 | return; | |
13507 | } | |
9c08d1fa | 13508 | else if (GET_CODE (body) == RETURN) |
13509 | { | |
13510 | start_insn = next_nonnote_insn (start_insn); | |
13511 | if (GET_CODE (start_insn) == BARRIER) | |
13512 | start_insn = next_nonnote_insn (start_insn); | |
13513 | if (GET_CODE (start_insn) == CODE_LABEL | |
13514 | && CODE_LABEL_NUMBER (start_insn) == arm_target_label | |
13515 | && LABEL_NUSES (start_insn) == 1) | |
13516 | { | |
13517 | reverse = TRUE; | |
13518 | seeking_return = 1; | |
13519 | } | |
13520 | else | |
13521 | return; | |
13522 | } | |
8cd42699 | 13523 | else |
13524 | return; | |
13525 | } | |
13526 | ||
ed29c566 | 13527 | gcc_assert (!arm_ccfsm_state || reverse); |
8cd42699 | 13528 | if (GET_CODE (insn) != JUMP_INSN) |
13529 | return; | |
13530 | ||
9e7454d0 | 13531 | /* This jump might be paralleled with a clobber of the condition codes |
9c08d1fa | 13532 | the jump should always come first */ |
13533 | if (GET_CODE (body) == PARALLEL && XVECLEN (body, 0) > 0) | |
13534 | body = XVECEXP (body, 0, 0); | |
13535 | ||
8cd42699 | 13536 | if (reverse |
13537 | || (GET_CODE (body) == SET && GET_CODE (SET_DEST (body)) == PC | |
13538 | && GET_CODE (SET_SRC (body)) == IF_THEN_ELSE)) | |
13539 | { | |
9480e6c7 | 13540 | int insns_skipped; |
13541 | int fail = FALSE, succeed = FALSE; | |
8cd42699 | 13542 | /* Flag which part of the IF_THEN_ELSE is the LABEL_REF. */ |
13543 | int then_not_else = TRUE; | |
9c08d1fa | 13544 | rtx this_insn = start_insn, label = 0; |
8cd42699 | 13545 | |
9e7454d0 | 13546 | /* If the jump cannot be done with one instruction, we cannot |
f3d87a93 | 13547 | conditionally execute the instruction in the inverse case. */ |
9c08d1fa | 13548 | if (get_attr_conds (insn) == CONDS_JUMP_CLOB) |
81380c80 | 13549 | { |
81380c80 | 13550 | jump_clobbers = 1; |
13551 | return; | |
13552 | } | |
9e7454d0 | 13553 | |
8cd42699 | 13554 | /* Register the insn jumped to. */ |
13555 | if (reverse) | |
9c08d1fa | 13556 | { |
13557 | if (!seeking_return) | |
13558 | label = XEXP (SET_SRC (body), 0); | |
13559 | } | |
8cd42699 | 13560 | else if (GET_CODE (XEXP (SET_SRC (body), 1)) == LABEL_REF) |
13561 | label = XEXP (XEXP (SET_SRC (body), 1), 0); | |
13562 | else if (GET_CODE (XEXP (SET_SRC (body), 2)) == LABEL_REF) | |
13563 | { | |
13564 | label = XEXP (XEXP (SET_SRC (body), 2), 0); | |
13565 | then_not_else = FALSE; | |
13566 | } | |
9c08d1fa | 13567 | else if (GET_CODE (XEXP (SET_SRC (body), 1)) == RETURN) |
13568 | seeking_return = 1; | |
13569 | else if (GET_CODE (XEXP (SET_SRC (body), 2)) == RETURN) | |
13570 | { | |
13571 | seeking_return = 1; | |
13572 | then_not_else = FALSE; | |
13573 | } | |
8cd42699 | 13574 | else |
ed29c566 | 13575 | gcc_unreachable (); |
8cd42699 | 13576 | |
13577 | /* See how many insns this branch skips, and what kind of insns. If all | |
13578 | insns are okay, and the label or unconditional branch to the same | |
13579 | label is not too far away, succeed. */ | |
13580 | for (insns_skipped = 0; | |
129a2fe4 | 13581 | !fail && !succeed && insns_skipped++ < max_insns_skipped;) |
8cd42699 | 13582 | { |
13583 | rtx scanbody; | |
13584 | ||
13585 | this_insn = next_nonnote_insn (this_insn); | |
13586 | if (!this_insn) | |
13587 | break; | |
13588 | ||
8cd42699 | 13589 | switch (GET_CODE (this_insn)) |
13590 | { | |
13591 | case CODE_LABEL: | |
13592 | /* Succeed if it is the target label, otherwise fail since | |
13593 | control falls in from somewhere else. */ | |
13594 | if (this_insn == label) | |
13595 | { | |
9c08d1fa | 13596 | if (jump_clobbers) |
13597 | { | |
13598 | arm_ccfsm_state = 2; | |
13599 | this_insn = next_nonnote_insn (this_insn); | |
13600 | } | |
13601 | else | |
13602 | arm_ccfsm_state = 1; | |
8cd42699 | 13603 | succeed = TRUE; |
13604 | } | |
13605 | else | |
13606 | fail = TRUE; | |
13607 | break; | |
13608 | ||
9c08d1fa | 13609 | case BARRIER: |
8cd42699 | 13610 | /* Succeed if the following insn is the target label. |
9e7454d0 | 13611 | Otherwise fail. |
13612 | If return insns are used then the last insn in a function | |
66aacf41 | 13613 | will be a barrier. */ |
8cd42699 | 13614 | this_insn = next_nonnote_insn (this_insn); |
9c08d1fa | 13615 | if (this_insn && this_insn == label) |
8cd42699 | 13616 | { |
9c08d1fa | 13617 | if (jump_clobbers) |
13618 | { | |
13619 | arm_ccfsm_state = 2; | |
13620 | this_insn = next_nonnote_insn (this_insn); | |
13621 | } | |
13622 | else | |
13623 | arm_ccfsm_state = 1; | |
8cd42699 | 13624 | succeed = TRUE; |
13625 | } | |
13626 | else | |
13627 | fail = TRUE; | |
13628 | break; | |
13629 | ||
9c08d1fa | 13630 | case CALL_INSN: |
f1039640 | 13631 | /* The AAPCS says that conditional calls should not be |
13632 | used since they make interworking inefficient (the | |
13633 | linker can't transform BL<cond> into BLX). That's | |
13634 | only a problem if the machine has BLX. */ | |
13635 | if (arm_arch5) | |
13636 | { | |
13637 | fail = TRUE; | |
13638 | break; | |
13639 | } | |
13640 | ||
c1a66faf | 13641 | /* Succeed if the following insn is the target label, or |
13642 | if the following two insns are a barrier and the | |
13643 | target label. */ | |
13644 | this_insn = next_nonnote_insn (this_insn); | |
13645 | if (this_insn && GET_CODE (this_insn) == BARRIER) | |
13646 | this_insn = next_nonnote_insn (this_insn); | |
9480e6c7 | 13647 | |
c1a66faf | 13648 | if (this_insn && this_insn == label |
13649 | && insns_skipped < max_insns_skipped) | |
13650 | { | |
13651 | if (jump_clobbers) | |
9480e6c7 | 13652 | { |
c1a66faf | 13653 | arm_ccfsm_state = 2; |
13654 | this_insn = next_nonnote_insn (this_insn); | |
9480e6c7 | 13655 | } |
13656 | else | |
c1a66faf | 13657 | arm_ccfsm_state = 1; |
13658 | succeed = TRUE; | |
9480e6c7 | 13659 | } |
c1a66faf | 13660 | else |
13661 | fail = TRUE; | |
9c08d1fa | 13662 | break; |
64266206 | 13663 | |
8cd42699 | 13664 | case JUMP_INSN: |
13665 | /* If this is an unconditional branch to the same label, succeed. | |
13666 | If it is to another label, do nothing. If it is conditional, | |
13667 | fail. */ | |
ebd88a36 | 13668 | /* XXX Probably, the tests for SET and the PC are |
13669 | unnecessary. */ | |
8cd42699 | 13670 | |
ae175e8f | 13671 | scanbody = PATTERN (this_insn); |
9c08d1fa | 13672 | if (GET_CODE (scanbody) == SET |
13673 | && GET_CODE (SET_DEST (scanbody)) == PC) | |
8cd42699 | 13674 | { |
13675 | if (GET_CODE (SET_SRC (scanbody)) == LABEL_REF | |
13676 | && XEXP (SET_SRC (scanbody), 0) == label && !reverse) | |
13677 | { | |
13678 | arm_ccfsm_state = 2; | |
13679 | succeed = TRUE; | |
13680 | } | |
13681 | else if (GET_CODE (SET_SRC (scanbody)) == IF_THEN_ELSE) | |
13682 | fail = TRUE; | |
13683 | } | |
a361b456 | 13684 | /* Fail if a conditional return is undesirable (e.g. on a |
129a2fe4 | 13685 | StrongARM), but still allow this if optimizing for size. */ |
13686 | else if (GET_CODE (scanbody) == RETURN | |
ffc9d00c | 13687 | && !use_return_insn (TRUE, NULL) |
215b30b3 | 13688 | && !optimize_size) |
129a2fe4 | 13689 | fail = TRUE; |
9c08d1fa | 13690 | else if (GET_CODE (scanbody) == RETURN |
13691 | && seeking_return) | |
13692 | { | |
13693 | arm_ccfsm_state = 2; | |
13694 | succeed = TRUE; | |
13695 | } | |
13696 | else if (GET_CODE (scanbody) == PARALLEL) | |
13697 | { | |
13698 | switch (get_attr_conds (this_insn)) | |
13699 | { | |
13700 | case CONDS_NOCOND: | |
13701 | break; | |
13702 | default: | |
13703 | fail = TRUE; | |
13704 | break; | |
13705 | } | |
13706 | } | |
5e3840e1 | 13707 | else |
a361b456 | 13708 | fail = TRUE; /* Unrecognized jump (e.g. epilogue). */ |
5e3840e1 | 13709 | |
8cd42699 | 13710 | break; |
13711 | ||
13712 | case INSN: | |
9c08d1fa | 13713 | /* Instructions using or affecting the condition codes make it |
13714 | fail. */ | |
ae175e8f | 13715 | scanbody = PATTERN (this_insn); |
215b30b3 | 13716 | if (!(GET_CODE (scanbody) == SET |
13717 | || GET_CODE (scanbody) == PARALLEL) | |
3acc3606 | 13718 | || get_attr_conds (this_insn) != CONDS_NOCOND) |
8cd42699 | 13719 | fail = TRUE; |
7d57ec45 | 13720 | |
13721 | /* A conditional cirrus instruction must be followed by | |
13722 | a non Cirrus instruction. However, since we | |
13723 | conditionalize instructions in this function and by | |
13724 | the time we get here we can't add instructions | |
13725 | (nops), because shorten_branches() has already been | |
13726 | called, we will disable conditionalizing Cirrus | |
13727 | instructions to be safe. */ | |
13728 | if (GET_CODE (scanbody) != USE | |
13729 | && GET_CODE (scanbody) != CLOBBER | |
2c6c7d8b | 13730 | && get_attr_cirrus (this_insn) != CIRRUS_NOT) |
7d57ec45 | 13731 | fail = TRUE; |
8cd42699 | 13732 | break; |
13733 | ||
13734 | default: | |
13735 | break; | |
13736 | } | |
13737 | } | |
13738 | if (succeed) | |
13739 | { | |
9c08d1fa | 13740 | if ((!seeking_return) && (arm_ccfsm_state == 1 || reverse)) |
8cd42699 | 13741 | arm_target_label = CODE_LABEL_NUMBER (label); |
ed29c566 | 13742 | else |
9c08d1fa | 13743 | { |
ed29c566 | 13744 | gcc_assert (seeking_return || arm_ccfsm_state == 2); |
a60f1840 | 13745 | |
9c08d1fa | 13746 | while (this_insn && GET_CODE (PATTERN (this_insn)) == USE) |
13747 | { | |
13748 | this_insn = next_nonnote_insn (this_insn); | |
ed29c566 | 13749 | gcc_assert (!this_insn |
13750 | || (GET_CODE (this_insn) != BARRIER | |
13751 | && GET_CODE (this_insn) != CODE_LABEL)); | |
9c08d1fa | 13752 | } |
13753 | if (!this_insn) | |
13754 | { | |
674a8f0b | 13755 | /* Oh, dear! we ran off the end.. give up. */ |
25f905c2 | 13756 | extract_constrain_insn_cached (insn); |
9c08d1fa | 13757 | arm_ccfsm_state = 0; |
48af475c | 13758 | arm_target_insn = NULL; |
9c08d1fa | 13759 | return; |
13760 | } | |
13761 | arm_target_insn = this_insn; | |
13762 | } | |
9c08d1fa | 13763 | if (jump_clobbers) |
13764 | { | |
ed29c566 | 13765 | gcc_assert (!reverse); |
9e7454d0 | 13766 | arm_current_cc = |
9c08d1fa | 13767 | get_arm_condition_code (XEXP (XEXP (XEXP (SET_SRC (body), |
13768 | 0), 0), 1)); | |
13769 | if (GET_CODE (XEXP (XEXP (SET_SRC (body), 0), 0)) == AND) | |
13770 | arm_current_cc = ARM_INVERSE_CONDITION_CODE (arm_current_cc); | |
13771 | if (GET_CODE (XEXP (SET_SRC (body), 0)) == NE) | |
13772 | arm_current_cc = ARM_INVERSE_CONDITION_CODE (arm_current_cc); | |
13773 | } | |
13774 | else | |
13775 | { | |
13776 | /* If REVERSE is true, ARM_CURRENT_CC needs to be inverted from | |
13777 | what it was. */ | |
13778 | if (!reverse) | |
13779 | arm_current_cc = get_arm_condition_code (XEXP (SET_SRC (body), | |
13780 | 0)); | |
13781 | } | |
8cd42699 | 13782 | |
8cd42699 | 13783 | if (reverse || then_not_else) |
13784 | arm_current_cc = ARM_INVERSE_CONDITION_CODE (arm_current_cc); | |
13785 | } | |
9e7454d0 | 13786 | |
ed420a25 | 13787 | /* Restore recog_data (getting the attributes of other insns can |
9c08d1fa | 13788 | destroy this array, but final.c assumes that it remains intact |
25f905c2 | 13789 | across this call. */ |
13790 | extract_constrain_insn_cached (insn); | |
13791 | } | |
13792 | } | |
13793 | ||
13794 | /* Output IT instructions. */ | |
13795 | void | |
13796 | thumb2_asm_output_opcode (FILE * stream) | |
13797 | { | |
13798 | char buff[5]; | |
13799 | int n; | |
13800 | ||
13801 | if (arm_condexec_mask) | |
13802 | { | |
13803 | for (n = 0; n < arm_condexec_masklen; n++) | |
13804 | buff[n] = (arm_condexec_mask & (1 << n)) ? 't' : 'e'; | |
13805 | buff[n] = 0; | |
13806 | asm_fprintf(stream, "i%s\t%s\n\t", buff, | |
13807 | arm_condition_codes[arm_current_cc]); | |
13808 | arm_condexec_mask = 0; | |
8cd42699 | 13809 | } |
a690ede2 | 13810 | } |
8cd42699 | 13811 | |
d19ef66f | 13812 | /* Returns true if REGNO is a valid register |
c1877da8 | 13813 | for holding a quantity of type MODE. */ |
d19ef66f | 13814 | int |
ebd88a36 | 13815 | arm_hard_regno_mode_ok (unsigned int regno, enum machine_mode mode) |
d19ef66f | 13816 | { |
13817 | if (GET_MODE_CLASS (mode) == MODE_CC) | |
8901ae22 | 13818 | return (regno == CC_REGNUM |
13819 | || (TARGET_HARD_FLOAT && TARGET_VFP | |
13820 | && regno == VFPCC_REGNUM)); | |
9e7454d0 | 13821 | |
25f905c2 | 13822 | if (TARGET_THUMB1) |
d19ef66f | 13823 | /* For the Thumb we only allow values bigger than SImode in |
13824 | registers 0 - 6, so that there is always a second low | |
13825 | register available to hold the upper part of the value. | |
13826 | We probably we ought to ensure that the register is the | |
13827 | start of an even numbered register pair. */ | |
36837fde | 13828 | return (ARM_NUM_REGS (mode) < 2) || (regno < LAST_LO_REGNUM); |
d19ef66f | 13829 | |
8901ae22 | 13830 | if (TARGET_HARD_FLOAT && TARGET_MAVERICK |
13831 | && IS_CIRRUS_REGNUM (regno)) | |
7d57ec45 | 13832 | /* We have outlawed SI values in Cirrus registers because they |
13833 | reside in the lower 32 bits, but SF values reside in the | |
13834 | upper 32 bits. This causes gcc all sorts of grief. We can't | |
13835 | even split the registers into pairs because Cirrus SI values | |
13836 | get sign extended to 64bits-- aldyh. */ | |
13837 | return (GET_MODE_CLASS (mode) == MODE_FLOAT) || (mode == DImode); | |
13838 | ||
8901ae22 | 13839 | if (TARGET_HARD_FLOAT && TARGET_VFP |
13840 | && IS_VFP_REGNUM (regno)) | |
a2cd141b | 13841 | { |
13842 | if (mode == SFmode || mode == SImode) | |
b5a0636d | 13843 | return VFP_REGNO_OK_FOR_SINGLE (regno); |
a2cd141b | 13844 | |
a2cd141b | 13845 | if (mode == DFmode) |
b5a0636d | 13846 | return VFP_REGNO_OK_FOR_DOUBLE (regno); |
d98a3884 | 13847 | |
13848 | if (TARGET_NEON) | |
13849 | return (VALID_NEON_DREG_MODE (mode) && VFP_REGNO_OK_FOR_DOUBLE (regno)) | |
13850 | || (VALID_NEON_QREG_MODE (mode) | |
13851 | && NEON_REGNO_OK_FOR_QUAD (regno)) | |
13852 | || (mode == TImode && NEON_REGNO_OK_FOR_NREGS (regno, 2)) | |
13853 | || (mode == EImode && NEON_REGNO_OK_FOR_NREGS (regno, 3)) | |
13854 | || (mode == OImode && NEON_REGNO_OK_FOR_NREGS (regno, 4)) | |
13855 | || (mode == CImode && NEON_REGNO_OK_FOR_NREGS (regno, 6)) | |
13856 | || (mode == XImode && NEON_REGNO_OK_FOR_NREGS (regno, 8)); | |
13857 | ||
a2cd141b | 13858 | return FALSE; |
13859 | } | |
13860 | ||
8901ae22 | 13861 | if (TARGET_REALLY_IWMMXT) |
13862 | { | |
13863 | if (IS_IWMMXT_GR_REGNUM (regno)) | |
13864 | return mode == SImode; | |
755eb2b4 | 13865 | |
8901ae22 | 13866 | if (IS_IWMMXT_REGNUM (regno)) |
13867 | return VALID_IWMMXT_REG_MODE (mode); | |
13868 | } | |
13869 | ||
a8a3b539 | 13870 | /* We allow any value to be stored in the general registers. |
13871 | Restrict doubleword quantities to even register pairs so that we can | |
d98a3884 | 13872 | use ldrd. Do not allow Neon structure opaque modes in general registers; |
13873 | they would use too many. */ | |
d19ef66f | 13874 | if (regno <= LAST_ARM_REGNUM) |
d98a3884 | 13875 | return !(TARGET_LDRD && GET_MODE_SIZE (mode) > 4 && (regno & 1) != 0) |
13876 | && !VALID_NEON_STRUCT_MODE (mode); | |
d19ef66f | 13877 | |
8901ae22 | 13878 | if (regno == FRAME_POINTER_REGNUM |
d19ef66f | 13879 | || regno == ARG_POINTER_REGNUM) |
13880 | /* We only allow integers in the fake hard registers. */ | |
13881 | return GET_MODE_CLASS (mode) == MODE_INT; | |
13882 | ||
e8ff69a4 | 13883 | /* The only registers left are the FPA registers |
d19ef66f | 13884 | which we only allow to hold FP values. */ |
8901ae22 | 13885 | return (TARGET_HARD_FLOAT && TARGET_FPA |
13886 | && GET_MODE_CLASS (mode) == MODE_FLOAT | |
13887 | && regno >= FIRST_FPA_REGNUM | |
13888 | && regno <= LAST_FPA_REGNUM); | |
d19ef66f | 13889 | } |
13890 | ||
25f905c2 | 13891 | /* For efficiency and historical reasons LO_REGS, HI_REGS and CC_REGS are |
13892 | not used in arm mode. */ | |
cffb2a26 | 13893 | int |
ebd88a36 | 13894 | arm_regno_class (int regno) |
cffb2a26 | 13895 | { |
25f905c2 | 13896 | if (TARGET_THUMB1) |
cffb2a26 | 13897 | { |
13898 | if (regno == STACK_POINTER_REGNUM) | |
13899 | return STACK_REG; | |
13900 | if (regno == CC_REGNUM) | |
13901 | return CC_REG; | |
13902 | if (regno < 8) | |
13903 | return LO_REGS; | |
13904 | return HI_REGS; | |
13905 | } | |
13906 | ||
25f905c2 | 13907 | if (TARGET_THUMB2 && regno < 8) |
13908 | return LO_REGS; | |
13909 | ||
cffb2a26 | 13910 | if ( regno <= LAST_ARM_REGNUM |
13911 | || regno == FRAME_POINTER_REGNUM | |
13912 | || regno == ARG_POINTER_REGNUM) | |
25f905c2 | 13913 | return TARGET_THUMB2 ? HI_REGS : GENERAL_REGS; |
9e7454d0 | 13914 | |
a2cd141b | 13915 | if (regno == CC_REGNUM || regno == VFPCC_REGNUM) |
25f905c2 | 13916 | return TARGET_THUMB2 ? CC_REG : NO_REGS; |
cffb2a26 | 13917 | |
7d57ec45 | 13918 | if (IS_CIRRUS_REGNUM (regno)) |
13919 | return CIRRUS_REGS; | |
13920 | ||
a2cd141b | 13921 | if (IS_VFP_REGNUM (regno)) |
b5a0636d | 13922 | { |
13923 | if (regno <= D7_VFP_REGNUM) | |
13924 | return VFP_D0_D7_REGS; | |
13925 | else if (regno <= LAST_LO_VFP_REGNUM) | |
13926 | return VFP_LO_REGS; | |
13927 | else | |
13928 | return VFP_HI_REGS; | |
13929 | } | |
a2cd141b | 13930 | |
755eb2b4 | 13931 | if (IS_IWMMXT_REGNUM (regno)) |
13932 | return IWMMXT_REGS; | |
13933 | ||
6ee861f1 | 13934 | if (IS_IWMMXT_GR_REGNUM (regno)) |
13935 | return IWMMXT_GR_REGS; | |
13936 | ||
e8ff69a4 | 13937 | return FPA_REGS; |
cffb2a26 | 13938 | } |
13939 | ||
13940 | /* Handle a special case when computing the offset | |
13941 | of an argument from the frame pointer. */ | |
13942 | int | |
ebd88a36 | 13943 | arm_debugger_arg_offset (int value, rtx addr) |
cffb2a26 | 13944 | { |
13945 | rtx insn; | |
13946 | ||
13947 | /* We are only interested if dbxout_parms() failed to compute the offset. */ | |
13948 | if (value != 0) | |
13949 | return 0; | |
13950 | ||
13951 | /* We can only cope with the case where the address is held in a register. */ | |
13952 | if (GET_CODE (addr) != REG) | |
13953 | return 0; | |
13954 | ||
13955 | /* If we are using the frame pointer to point at the argument, then | |
13956 | an offset of 0 is correct. */ | |
0bd59681 | 13957 | if (REGNO (addr) == (unsigned) HARD_FRAME_POINTER_REGNUM) |
cffb2a26 | 13958 | return 0; |
9e7454d0 | 13959 | |
cffb2a26 | 13960 | /* If we are using the stack pointer to point at the |
13961 | argument, then an offset of 0 is correct. */ | |
25f905c2 | 13962 | /* ??? Check this is consistent with thumb2 frame layout. */ |
215b30b3 | 13963 | if ((TARGET_THUMB || !frame_pointer_needed) |
cffb2a26 | 13964 | && REGNO (addr) == SP_REGNUM) |
13965 | return 0; | |
9e7454d0 | 13966 | |
cffb2a26 | 13967 | /* Oh dear. The argument is pointed to by a register rather |
13968 | than being held in a register, or being stored at a known | |
13969 | offset from the frame pointer. Since GDB only understands | |
13970 | those two kinds of argument we must translate the address | |
13971 | held in the register into an offset from the frame pointer. | |
13972 | We do this by searching through the insns for the function | |
13973 | looking to see where this register gets its value. If the | |
457275b6 | 13974 | register is initialized from the frame pointer plus an offset |
cffb2a26 | 13975 | then we are in luck and we can continue, otherwise we give up. |
9e7454d0 | 13976 | |
cffb2a26 | 13977 | This code is exercised by producing debugging information |
13978 | for a function with arguments like this: | |
9e7454d0 | 13979 | |
cffb2a26 | 13980 | double func (double a, double b, int c, double d) {return d;} |
9e7454d0 | 13981 | |
cffb2a26 | 13982 | Without this code the stab for parameter 'd' will be set to |
13983 | an offset of 0 from the frame pointer, rather than 8. */ | |
13984 | ||
13985 | /* The if() statement says: | |
13986 | ||
13987 | If the insn is a normal instruction | |
13988 | and if the insn is setting the value in a register | |
13989 | and if the register being set is the register holding the address of the argument | |
13990 | and if the address is computing by an addition | |
13991 | that involves adding to a register | |
13992 | which is the frame pointer | |
13993 | a constant integer | |
13994 | ||
674a8f0b | 13995 | then... */ |
9e7454d0 | 13996 | |
cffb2a26 | 13997 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) |
13998 | { | |
9e7454d0 | 13999 | if ( GET_CODE (insn) == INSN |
cffb2a26 | 14000 | && GET_CODE (PATTERN (insn)) == SET |
14001 | && REGNO (XEXP (PATTERN (insn), 0)) == REGNO (addr) | |
14002 | && GET_CODE (XEXP (PATTERN (insn), 1)) == PLUS | |
14003 | && GET_CODE (XEXP (XEXP (PATTERN (insn), 1), 0)) == REG | |
0bd59681 | 14004 | && REGNO (XEXP (XEXP (PATTERN (insn), 1), 0)) == (unsigned) HARD_FRAME_POINTER_REGNUM |
cffb2a26 | 14005 | && GET_CODE (XEXP (XEXP (PATTERN (insn), 1), 1)) == CONST_INT |
14006 | ) | |
14007 | { | |
14008 | value = INTVAL (XEXP (XEXP (PATTERN (insn), 1), 1)); | |
9e7454d0 | 14009 | |
cffb2a26 | 14010 | break; |
14011 | } | |
14012 | } | |
9e7454d0 | 14013 | |
cffb2a26 | 14014 | if (value == 0) |
14015 | { | |
14016 | debug_rtx (addr); | |
c3ceba8e | 14017 | warning (0, "unable to compute real location of stacked parameter"); |
cffb2a26 | 14018 | value = 8; /* XXX magic hack */ |
14019 | } | |
14020 | ||
14021 | return value; | |
14022 | } | |
cffb2a26 | 14023 | \f |
755eb2b4 | 14024 | #define def_mbuiltin(MASK, NAME, TYPE, CODE) \ |
14025 | do \ | |
14026 | { \ | |
14027 | if ((MASK) & insn_flags) \ | |
54be5d7e | 14028 | add_builtin_function ((NAME), (TYPE), (CODE), \ |
14029 | BUILT_IN_MD, NULL, NULL_TREE); \ | |
755eb2b4 | 14030 | } \ |
14031 | while (0) | |
14032 | ||
14033 | struct builtin_description | |
14034 | { | |
14035 | const unsigned int mask; | |
14036 | const enum insn_code icode; | |
14037 | const char * const name; | |
14038 | const enum arm_builtins code; | |
14039 | const enum rtx_code comparison; | |
14040 | const unsigned int flag; | |
14041 | }; | |
14042 | ||
14043 | static const struct builtin_description bdesc_2arg[] = | |
14044 | { | |
14045 | #define IWMMXT_BUILTIN(code, string, builtin) \ | |
14046 | { FL_IWMMXT, CODE_FOR_##code, "__builtin_arm_" string, \ | |
14047 | ARM_BUILTIN_##builtin, 0, 0 }, | |
14048 | ||
14049 | IWMMXT_BUILTIN (addv8qi3, "waddb", WADDB) | |
14050 | IWMMXT_BUILTIN (addv4hi3, "waddh", WADDH) | |
14051 | IWMMXT_BUILTIN (addv2si3, "waddw", WADDW) | |
14052 | IWMMXT_BUILTIN (subv8qi3, "wsubb", WSUBB) | |
14053 | IWMMXT_BUILTIN (subv4hi3, "wsubh", WSUBH) | |
14054 | IWMMXT_BUILTIN (subv2si3, "wsubw", WSUBW) | |
14055 | IWMMXT_BUILTIN (ssaddv8qi3, "waddbss", WADDSSB) | |
14056 | IWMMXT_BUILTIN (ssaddv4hi3, "waddhss", WADDSSH) | |
14057 | IWMMXT_BUILTIN (ssaddv2si3, "waddwss", WADDSSW) | |
14058 | IWMMXT_BUILTIN (sssubv8qi3, "wsubbss", WSUBSSB) | |
14059 | IWMMXT_BUILTIN (sssubv4hi3, "wsubhss", WSUBSSH) | |
14060 | IWMMXT_BUILTIN (sssubv2si3, "wsubwss", WSUBSSW) | |
14061 | IWMMXT_BUILTIN (usaddv8qi3, "waddbus", WADDUSB) | |
14062 | IWMMXT_BUILTIN (usaddv4hi3, "waddhus", WADDUSH) | |
14063 | IWMMXT_BUILTIN (usaddv2si3, "waddwus", WADDUSW) | |
14064 | IWMMXT_BUILTIN (ussubv8qi3, "wsubbus", WSUBUSB) | |
14065 | IWMMXT_BUILTIN (ussubv4hi3, "wsubhus", WSUBUSH) | |
14066 | IWMMXT_BUILTIN (ussubv2si3, "wsubwus", WSUBUSW) | |
14067 | IWMMXT_BUILTIN (mulv4hi3, "wmulul", WMULUL) | |
08476a09 | 14068 | IWMMXT_BUILTIN (smulv4hi3_highpart, "wmulsm", WMULSM) |
14069 | IWMMXT_BUILTIN (umulv4hi3_highpart, "wmulum", WMULUM) | |
755eb2b4 | 14070 | IWMMXT_BUILTIN (eqv8qi3, "wcmpeqb", WCMPEQB) |
14071 | IWMMXT_BUILTIN (eqv4hi3, "wcmpeqh", WCMPEQH) | |
14072 | IWMMXT_BUILTIN (eqv2si3, "wcmpeqw", WCMPEQW) | |
14073 | IWMMXT_BUILTIN (gtuv8qi3, "wcmpgtub", WCMPGTUB) | |
14074 | IWMMXT_BUILTIN (gtuv4hi3, "wcmpgtuh", WCMPGTUH) | |
14075 | IWMMXT_BUILTIN (gtuv2si3, "wcmpgtuw", WCMPGTUW) | |
14076 | IWMMXT_BUILTIN (gtv8qi3, "wcmpgtsb", WCMPGTSB) | |
14077 | IWMMXT_BUILTIN (gtv4hi3, "wcmpgtsh", WCMPGTSH) | |
14078 | IWMMXT_BUILTIN (gtv2si3, "wcmpgtsw", WCMPGTSW) | |
14079 | IWMMXT_BUILTIN (umaxv8qi3, "wmaxub", WMAXUB) | |
14080 | IWMMXT_BUILTIN (smaxv8qi3, "wmaxsb", WMAXSB) | |
14081 | IWMMXT_BUILTIN (umaxv4hi3, "wmaxuh", WMAXUH) | |
14082 | IWMMXT_BUILTIN (smaxv4hi3, "wmaxsh", WMAXSH) | |
14083 | IWMMXT_BUILTIN (umaxv2si3, "wmaxuw", WMAXUW) | |
14084 | IWMMXT_BUILTIN (smaxv2si3, "wmaxsw", WMAXSW) | |
14085 | IWMMXT_BUILTIN (uminv8qi3, "wminub", WMINUB) | |
14086 | IWMMXT_BUILTIN (sminv8qi3, "wminsb", WMINSB) | |
14087 | IWMMXT_BUILTIN (uminv4hi3, "wminuh", WMINUH) | |
14088 | IWMMXT_BUILTIN (sminv4hi3, "wminsh", WMINSH) | |
14089 | IWMMXT_BUILTIN (uminv2si3, "wminuw", WMINUW) | |
14090 | IWMMXT_BUILTIN (sminv2si3, "wminsw", WMINSW) | |
14091 | IWMMXT_BUILTIN (iwmmxt_anddi3, "wand", WAND) | |
14092 | IWMMXT_BUILTIN (iwmmxt_nanddi3, "wandn", WANDN) | |
14093 | IWMMXT_BUILTIN (iwmmxt_iordi3, "wor", WOR) | |
14094 | IWMMXT_BUILTIN (iwmmxt_xordi3, "wxor", WXOR) | |
14095 | IWMMXT_BUILTIN (iwmmxt_uavgv8qi3, "wavg2b", WAVG2B) | |
14096 | IWMMXT_BUILTIN (iwmmxt_uavgv4hi3, "wavg2h", WAVG2H) | |
14097 | IWMMXT_BUILTIN (iwmmxt_uavgrndv8qi3, "wavg2br", WAVG2BR) | |
14098 | IWMMXT_BUILTIN (iwmmxt_uavgrndv4hi3, "wavg2hr", WAVG2HR) | |
14099 | IWMMXT_BUILTIN (iwmmxt_wunpckilb, "wunpckilb", WUNPCKILB) | |
14100 | IWMMXT_BUILTIN (iwmmxt_wunpckilh, "wunpckilh", WUNPCKILH) | |
14101 | IWMMXT_BUILTIN (iwmmxt_wunpckilw, "wunpckilw", WUNPCKILW) | |
14102 | IWMMXT_BUILTIN (iwmmxt_wunpckihb, "wunpckihb", WUNPCKIHB) | |
14103 | IWMMXT_BUILTIN (iwmmxt_wunpckihh, "wunpckihh", WUNPCKIHH) | |
14104 | IWMMXT_BUILTIN (iwmmxt_wunpckihw, "wunpckihw", WUNPCKIHW) | |
14105 | IWMMXT_BUILTIN (iwmmxt_wmadds, "wmadds", WMADDS) | |
14106 | IWMMXT_BUILTIN (iwmmxt_wmaddu, "wmaddu", WMADDU) | |
14107 | ||
14108 | #define IWMMXT_BUILTIN2(code, builtin) \ | |
14109 | { FL_IWMMXT, CODE_FOR_##code, NULL, ARM_BUILTIN_##builtin, 0, 0 }, | |
9e7454d0 | 14110 | |
755eb2b4 | 14111 | IWMMXT_BUILTIN2 (iwmmxt_wpackhss, WPACKHSS) |
14112 | IWMMXT_BUILTIN2 (iwmmxt_wpackwss, WPACKWSS) | |
14113 | IWMMXT_BUILTIN2 (iwmmxt_wpackdss, WPACKDSS) | |
14114 | IWMMXT_BUILTIN2 (iwmmxt_wpackhus, WPACKHUS) | |
14115 | IWMMXT_BUILTIN2 (iwmmxt_wpackwus, WPACKWUS) | |
14116 | IWMMXT_BUILTIN2 (iwmmxt_wpackdus, WPACKDUS) | |
14117 | IWMMXT_BUILTIN2 (ashlv4hi3_di, WSLLH) | |
d98a3884 | 14118 | IWMMXT_BUILTIN2 (ashlv4hi3_iwmmxt, WSLLHI) |
755eb2b4 | 14119 | IWMMXT_BUILTIN2 (ashlv2si3_di, WSLLW) |
d98a3884 | 14120 | IWMMXT_BUILTIN2 (ashlv2si3_iwmmxt, WSLLWI) |
755eb2b4 | 14121 | IWMMXT_BUILTIN2 (ashldi3_di, WSLLD) |
14122 | IWMMXT_BUILTIN2 (ashldi3_iwmmxt, WSLLDI) | |
14123 | IWMMXT_BUILTIN2 (lshrv4hi3_di, WSRLH) | |
d98a3884 | 14124 | IWMMXT_BUILTIN2 (lshrv4hi3_iwmmxt, WSRLHI) |
755eb2b4 | 14125 | IWMMXT_BUILTIN2 (lshrv2si3_di, WSRLW) |
d98a3884 | 14126 | IWMMXT_BUILTIN2 (lshrv2si3_iwmmxt, WSRLWI) |
755eb2b4 | 14127 | IWMMXT_BUILTIN2 (lshrdi3_di, WSRLD) |
a2cd141b | 14128 | IWMMXT_BUILTIN2 (lshrdi3_iwmmxt, WSRLDI) |
755eb2b4 | 14129 | IWMMXT_BUILTIN2 (ashrv4hi3_di, WSRAH) |
d98a3884 | 14130 | IWMMXT_BUILTIN2 (ashrv4hi3_iwmmxt, WSRAHI) |
755eb2b4 | 14131 | IWMMXT_BUILTIN2 (ashrv2si3_di, WSRAW) |
d98a3884 | 14132 | IWMMXT_BUILTIN2 (ashrv2si3_iwmmxt, WSRAWI) |
755eb2b4 | 14133 | IWMMXT_BUILTIN2 (ashrdi3_di, WSRAD) |
a2cd141b | 14134 | IWMMXT_BUILTIN2 (ashrdi3_iwmmxt, WSRADI) |
755eb2b4 | 14135 | IWMMXT_BUILTIN2 (rorv4hi3_di, WRORH) |
14136 | IWMMXT_BUILTIN2 (rorv4hi3, WRORHI) | |
14137 | IWMMXT_BUILTIN2 (rorv2si3_di, WRORW) | |
14138 | IWMMXT_BUILTIN2 (rorv2si3, WRORWI) | |
14139 | IWMMXT_BUILTIN2 (rordi3_di, WRORD) | |
14140 | IWMMXT_BUILTIN2 (rordi3, WRORDI) | |
14141 | IWMMXT_BUILTIN2 (iwmmxt_wmacuz, WMACUZ) | |
14142 | IWMMXT_BUILTIN2 (iwmmxt_wmacsz, WMACSZ) | |
14143 | }; | |
14144 | ||
14145 | static const struct builtin_description bdesc_1arg[] = | |
14146 | { | |
14147 | IWMMXT_BUILTIN (iwmmxt_tmovmskb, "tmovmskb", TMOVMSKB) | |
14148 | IWMMXT_BUILTIN (iwmmxt_tmovmskh, "tmovmskh", TMOVMSKH) | |
14149 | IWMMXT_BUILTIN (iwmmxt_tmovmskw, "tmovmskw", TMOVMSKW) | |
14150 | IWMMXT_BUILTIN (iwmmxt_waccb, "waccb", WACCB) | |
14151 | IWMMXT_BUILTIN (iwmmxt_wacch, "wacch", WACCH) | |
14152 | IWMMXT_BUILTIN (iwmmxt_waccw, "waccw", WACCW) | |
14153 | IWMMXT_BUILTIN (iwmmxt_wunpckehub, "wunpckehub", WUNPCKEHUB) | |
14154 | IWMMXT_BUILTIN (iwmmxt_wunpckehuh, "wunpckehuh", WUNPCKEHUH) | |
14155 | IWMMXT_BUILTIN (iwmmxt_wunpckehuw, "wunpckehuw", WUNPCKEHUW) | |
14156 | IWMMXT_BUILTIN (iwmmxt_wunpckehsb, "wunpckehsb", WUNPCKEHSB) | |
14157 | IWMMXT_BUILTIN (iwmmxt_wunpckehsh, "wunpckehsh", WUNPCKEHSH) | |
14158 | IWMMXT_BUILTIN (iwmmxt_wunpckehsw, "wunpckehsw", WUNPCKEHSW) | |
14159 | IWMMXT_BUILTIN (iwmmxt_wunpckelub, "wunpckelub", WUNPCKELUB) | |
14160 | IWMMXT_BUILTIN (iwmmxt_wunpckeluh, "wunpckeluh", WUNPCKELUH) | |
14161 | IWMMXT_BUILTIN (iwmmxt_wunpckeluw, "wunpckeluw", WUNPCKELUW) | |
14162 | IWMMXT_BUILTIN (iwmmxt_wunpckelsb, "wunpckelsb", WUNPCKELSB) | |
14163 | IWMMXT_BUILTIN (iwmmxt_wunpckelsh, "wunpckelsh", WUNPCKELSH) | |
14164 | IWMMXT_BUILTIN (iwmmxt_wunpckelsw, "wunpckelsw", WUNPCKELSW) | |
14165 | }; | |
14166 | ||
14167 | /* Set up all the iWMMXt builtins. This is | |
14168 | not called if TARGET_IWMMXT is zero. */ | |
14169 | ||
14170 | static void | |
14171 | arm_init_iwmmxt_builtins (void) | |
14172 | { | |
14173 | const struct builtin_description * d; | |
14174 | size_t i; | |
14175 | tree endlink = void_list_node; | |
14176 | ||
4917c376 | 14177 | tree V2SI_type_node = build_vector_type_for_mode (intSI_type_node, V2SImode); |
14178 | tree V4HI_type_node = build_vector_type_for_mode (intHI_type_node, V4HImode); | |
14179 | tree V8QI_type_node = build_vector_type_for_mode (intQI_type_node, V8QImode); | |
14180 | ||
755eb2b4 | 14181 | tree int_ftype_int |
14182 | = build_function_type (integer_type_node, | |
14183 | tree_cons (NULL_TREE, integer_type_node, endlink)); | |
14184 | tree v8qi_ftype_v8qi_v8qi_int | |
14185 | = build_function_type (V8QI_type_node, | |
14186 | tree_cons (NULL_TREE, V8QI_type_node, | |
14187 | tree_cons (NULL_TREE, V8QI_type_node, | |
14188 | tree_cons (NULL_TREE, | |
14189 | integer_type_node, | |
14190 | endlink)))); | |
14191 | tree v4hi_ftype_v4hi_int | |
14192 | = build_function_type (V4HI_type_node, | |
14193 | tree_cons (NULL_TREE, V4HI_type_node, | |
14194 | tree_cons (NULL_TREE, integer_type_node, | |
14195 | endlink))); | |
14196 | tree v2si_ftype_v2si_int | |
14197 | = build_function_type (V2SI_type_node, | |
14198 | tree_cons (NULL_TREE, V2SI_type_node, | |
14199 | tree_cons (NULL_TREE, integer_type_node, | |
14200 | endlink))); | |
14201 | tree v2si_ftype_di_di | |
14202 | = build_function_type (V2SI_type_node, | |
14203 | tree_cons (NULL_TREE, long_long_integer_type_node, | |
14204 | tree_cons (NULL_TREE, long_long_integer_type_node, | |
14205 | endlink))); | |
14206 | tree di_ftype_di_int | |
14207 | = build_function_type (long_long_integer_type_node, | |
14208 | tree_cons (NULL_TREE, long_long_integer_type_node, | |
14209 | tree_cons (NULL_TREE, integer_type_node, | |
14210 | endlink))); | |
14211 | tree di_ftype_di_int_int | |
14212 | = build_function_type (long_long_integer_type_node, | |
14213 | tree_cons (NULL_TREE, long_long_integer_type_node, | |
14214 | tree_cons (NULL_TREE, integer_type_node, | |
14215 | tree_cons (NULL_TREE, | |
14216 | integer_type_node, | |
14217 | endlink)))); | |
14218 | tree int_ftype_v8qi | |
14219 | = build_function_type (integer_type_node, | |
14220 | tree_cons (NULL_TREE, V8QI_type_node, | |
14221 | endlink)); | |
14222 | tree int_ftype_v4hi | |
14223 | = build_function_type (integer_type_node, | |
14224 | tree_cons (NULL_TREE, V4HI_type_node, | |
14225 | endlink)); | |
14226 | tree int_ftype_v2si | |
14227 | = build_function_type (integer_type_node, | |
14228 | tree_cons (NULL_TREE, V2SI_type_node, | |
14229 | endlink)); | |
14230 | tree int_ftype_v8qi_int | |
14231 | = build_function_type (integer_type_node, | |
14232 | tree_cons (NULL_TREE, V8QI_type_node, | |
14233 | tree_cons (NULL_TREE, integer_type_node, | |
14234 | endlink))); | |
14235 | tree int_ftype_v4hi_int | |
14236 | = build_function_type (integer_type_node, | |
14237 | tree_cons (NULL_TREE, V4HI_type_node, | |
14238 | tree_cons (NULL_TREE, integer_type_node, | |
14239 | endlink))); | |
14240 | tree int_ftype_v2si_int | |
14241 | = build_function_type (integer_type_node, | |
14242 | tree_cons (NULL_TREE, V2SI_type_node, | |
14243 | tree_cons (NULL_TREE, integer_type_node, | |
14244 | endlink))); | |
14245 | tree v8qi_ftype_v8qi_int_int | |
14246 | = build_function_type (V8QI_type_node, | |
14247 | tree_cons (NULL_TREE, V8QI_type_node, | |
14248 | tree_cons (NULL_TREE, integer_type_node, | |
14249 | tree_cons (NULL_TREE, | |
14250 | integer_type_node, | |
14251 | endlink)))); | |
14252 | tree v4hi_ftype_v4hi_int_int | |
14253 | = build_function_type (V4HI_type_node, | |
14254 | tree_cons (NULL_TREE, V4HI_type_node, | |
14255 | tree_cons (NULL_TREE, integer_type_node, | |
14256 | tree_cons (NULL_TREE, | |
14257 | integer_type_node, | |
14258 | endlink)))); | |
14259 | tree v2si_ftype_v2si_int_int | |
14260 | = build_function_type (V2SI_type_node, | |
14261 | tree_cons (NULL_TREE, V2SI_type_node, | |
14262 | tree_cons (NULL_TREE, integer_type_node, | |
14263 | tree_cons (NULL_TREE, | |
14264 | integer_type_node, | |
14265 | endlink)))); | |
14266 | /* Miscellaneous. */ | |
14267 | tree v8qi_ftype_v4hi_v4hi | |
14268 | = build_function_type (V8QI_type_node, | |
14269 | tree_cons (NULL_TREE, V4HI_type_node, | |
14270 | tree_cons (NULL_TREE, V4HI_type_node, | |
14271 | endlink))); | |
14272 | tree v4hi_ftype_v2si_v2si | |
14273 | = build_function_type (V4HI_type_node, | |
14274 | tree_cons (NULL_TREE, V2SI_type_node, | |
14275 | tree_cons (NULL_TREE, V2SI_type_node, | |
14276 | endlink))); | |
14277 | tree v2si_ftype_v4hi_v4hi | |
14278 | = build_function_type (V2SI_type_node, | |
14279 | tree_cons (NULL_TREE, V4HI_type_node, | |
14280 | tree_cons (NULL_TREE, V4HI_type_node, | |
14281 | endlink))); | |
14282 | tree v2si_ftype_v8qi_v8qi | |
14283 | = build_function_type (V2SI_type_node, | |
14284 | tree_cons (NULL_TREE, V8QI_type_node, | |
14285 | tree_cons (NULL_TREE, V8QI_type_node, | |
14286 | endlink))); | |
14287 | tree v4hi_ftype_v4hi_di | |
14288 | = build_function_type (V4HI_type_node, | |
14289 | tree_cons (NULL_TREE, V4HI_type_node, | |
14290 | tree_cons (NULL_TREE, | |
14291 | long_long_integer_type_node, | |
14292 | endlink))); | |
14293 | tree v2si_ftype_v2si_di | |
14294 | = build_function_type (V2SI_type_node, | |
14295 | tree_cons (NULL_TREE, V2SI_type_node, | |
14296 | tree_cons (NULL_TREE, | |
14297 | long_long_integer_type_node, | |
14298 | endlink))); | |
14299 | tree void_ftype_int_int | |
14300 | = build_function_type (void_type_node, | |
14301 | tree_cons (NULL_TREE, integer_type_node, | |
14302 | tree_cons (NULL_TREE, integer_type_node, | |
14303 | endlink))); | |
14304 | tree di_ftype_void | |
14305 | = build_function_type (long_long_unsigned_type_node, endlink); | |
14306 | tree di_ftype_v8qi | |
14307 | = build_function_type (long_long_integer_type_node, | |
14308 | tree_cons (NULL_TREE, V8QI_type_node, | |
14309 | endlink)); | |
14310 | tree di_ftype_v4hi | |
14311 | = build_function_type (long_long_integer_type_node, | |
14312 | tree_cons (NULL_TREE, V4HI_type_node, | |
14313 | endlink)); | |
14314 | tree di_ftype_v2si | |
14315 | = build_function_type (long_long_integer_type_node, | |
14316 | tree_cons (NULL_TREE, V2SI_type_node, | |
14317 | endlink)); | |
14318 | tree v2si_ftype_v4hi | |
14319 | = build_function_type (V2SI_type_node, | |
14320 | tree_cons (NULL_TREE, V4HI_type_node, | |
14321 | endlink)); | |
14322 | tree v4hi_ftype_v8qi | |
14323 | = build_function_type (V4HI_type_node, | |
14324 | tree_cons (NULL_TREE, V8QI_type_node, | |
14325 | endlink)); | |
14326 | ||
14327 | tree di_ftype_di_v4hi_v4hi | |
14328 | = build_function_type (long_long_unsigned_type_node, | |
14329 | tree_cons (NULL_TREE, | |
14330 | long_long_unsigned_type_node, | |
14331 | tree_cons (NULL_TREE, V4HI_type_node, | |
14332 | tree_cons (NULL_TREE, | |
14333 | V4HI_type_node, | |
14334 | endlink)))); | |
14335 | ||
14336 | tree di_ftype_v4hi_v4hi | |
14337 | = build_function_type (long_long_unsigned_type_node, | |
14338 | tree_cons (NULL_TREE, V4HI_type_node, | |
14339 | tree_cons (NULL_TREE, V4HI_type_node, | |
14340 | endlink))); | |
14341 | ||
14342 | /* Normal vector binops. */ | |
14343 | tree v8qi_ftype_v8qi_v8qi | |
14344 | = build_function_type (V8QI_type_node, | |
14345 | tree_cons (NULL_TREE, V8QI_type_node, | |
14346 | tree_cons (NULL_TREE, V8QI_type_node, | |
14347 | endlink))); | |
14348 | tree v4hi_ftype_v4hi_v4hi | |
14349 | = build_function_type (V4HI_type_node, | |
14350 | tree_cons (NULL_TREE, V4HI_type_node, | |
14351 | tree_cons (NULL_TREE, V4HI_type_node, | |
14352 | endlink))); | |
14353 | tree v2si_ftype_v2si_v2si | |
14354 | = build_function_type (V2SI_type_node, | |
14355 | tree_cons (NULL_TREE, V2SI_type_node, | |
14356 | tree_cons (NULL_TREE, V2SI_type_node, | |
14357 | endlink))); | |
14358 | tree di_ftype_di_di | |
14359 | = build_function_type (long_long_unsigned_type_node, | |
14360 | tree_cons (NULL_TREE, long_long_unsigned_type_node, | |
14361 | tree_cons (NULL_TREE, | |
14362 | long_long_unsigned_type_node, | |
14363 | endlink))); | |
14364 | ||
14365 | /* Add all builtins that are more or less simple operations on two | |
14366 | operands. */ | |
3b9d9482 | 14367 | for (i = 0, d = bdesc_2arg; i < ARRAY_SIZE (bdesc_2arg); i++, d++) |
755eb2b4 | 14368 | { |
14369 | /* Use one of the operands; the target can have a different mode for | |
14370 | mask-generating compares. */ | |
14371 | enum machine_mode mode; | |
14372 | tree type; | |
14373 | ||
14374 | if (d->name == 0) | |
14375 | continue; | |
14376 | ||
14377 | mode = insn_data[d->icode].operand[1].mode; | |
14378 | ||
14379 | switch (mode) | |
14380 | { | |
14381 | case V8QImode: | |
14382 | type = v8qi_ftype_v8qi_v8qi; | |
14383 | break; | |
14384 | case V4HImode: | |
14385 | type = v4hi_ftype_v4hi_v4hi; | |
14386 | break; | |
14387 | case V2SImode: | |
14388 | type = v2si_ftype_v2si_v2si; | |
14389 | break; | |
14390 | case DImode: | |
14391 | type = di_ftype_di_di; | |
14392 | break; | |
14393 | ||
14394 | default: | |
ed29c566 | 14395 | gcc_unreachable (); |
755eb2b4 | 14396 | } |
14397 | ||
14398 | def_mbuiltin (d->mask, d->name, type, d->code); | |
14399 | } | |
14400 | ||
14401 | /* Add the remaining MMX insns with somewhat more complicated types. */ | |
14402 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wzero", di_ftype_void, ARM_BUILTIN_WZERO); | |
14403 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_setwcx", void_ftype_int_int, ARM_BUILTIN_SETWCX); | |
14404 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_getwcx", int_ftype_int, ARM_BUILTIN_GETWCX); | |
14405 | ||
14406 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsllh", v4hi_ftype_v4hi_di, ARM_BUILTIN_WSLLH); | |
14407 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsllw", v2si_ftype_v2si_di, ARM_BUILTIN_WSLLW); | |
14408 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wslld", di_ftype_di_di, ARM_BUILTIN_WSLLD); | |
14409 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsllhi", v4hi_ftype_v4hi_int, ARM_BUILTIN_WSLLHI); | |
14410 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsllwi", v2si_ftype_v2si_int, ARM_BUILTIN_WSLLWI); | |
14411 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wslldi", di_ftype_di_int, ARM_BUILTIN_WSLLDI); | |
14412 | ||
14413 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrlh", v4hi_ftype_v4hi_di, ARM_BUILTIN_WSRLH); | |
14414 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrlw", v2si_ftype_v2si_di, ARM_BUILTIN_WSRLW); | |
14415 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrld", di_ftype_di_di, ARM_BUILTIN_WSRLD); | |
14416 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrlhi", v4hi_ftype_v4hi_int, ARM_BUILTIN_WSRLHI); | |
14417 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrlwi", v2si_ftype_v2si_int, ARM_BUILTIN_WSRLWI); | |
14418 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrldi", di_ftype_di_int, ARM_BUILTIN_WSRLDI); | |
14419 | ||
14420 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrah", v4hi_ftype_v4hi_di, ARM_BUILTIN_WSRAH); | |
14421 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsraw", v2si_ftype_v2si_di, ARM_BUILTIN_WSRAW); | |
14422 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrad", di_ftype_di_di, ARM_BUILTIN_WSRAD); | |
14423 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrahi", v4hi_ftype_v4hi_int, ARM_BUILTIN_WSRAHI); | |
14424 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrawi", v2si_ftype_v2si_int, ARM_BUILTIN_WSRAWI); | |
14425 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsradi", di_ftype_di_int, ARM_BUILTIN_WSRADI); | |
14426 | ||
14427 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wrorh", v4hi_ftype_v4hi_di, ARM_BUILTIN_WRORH); | |
14428 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wrorw", v2si_ftype_v2si_di, ARM_BUILTIN_WRORW); | |
14429 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wrord", di_ftype_di_di, ARM_BUILTIN_WRORD); | |
14430 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wrorhi", v4hi_ftype_v4hi_int, ARM_BUILTIN_WRORHI); | |
14431 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wrorwi", v2si_ftype_v2si_int, ARM_BUILTIN_WRORWI); | |
14432 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wrordi", di_ftype_di_int, ARM_BUILTIN_WRORDI); | |
14433 | ||
14434 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wshufh", v4hi_ftype_v4hi_int, ARM_BUILTIN_WSHUFH); | |
14435 | ||
14436 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsadb", v2si_ftype_v8qi_v8qi, ARM_BUILTIN_WSADB); | |
14437 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsadh", v2si_ftype_v4hi_v4hi, ARM_BUILTIN_WSADH); | |
14438 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsadbz", v2si_ftype_v8qi_v8qi, ARM_BUILTIN_WSADBZ); | |
14439 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsadhz", v2si_ftype_v4hi_v4hi, ARM_BUILTIN_WSADHZ); | |
14440 | ||
14441 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_textrmsb", int_ftype_v8qi_int, ARM_BUILTIN_TEXTRMSB); | |
14442 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_textrmsh", int_ftype_v4hi_int, ARM_BUILTIN_TEXTRMSH); | |
14443 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_textrmsw", int_ftype_v2si_int, ARM_BUILTIN_TEXTRMSW); | |
14444 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_textrmub", int_ftype_v8qi_int, ARM_BUILTIN_TEXTRMUB); | |
14445 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_textrmuh", int_ftype_v4hi_int, ARM_BUILTIN_TEXTRMUH); | |
14446 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_textrmuw", int_ftype_v2si_int, ARM_BUILTIN_TEXTRMUW); | |
14447 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tinsrb", v8qi_ftype_v8qi_int_int, ARM_BUILTIN_TINSRB); | |
14448 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tinsrh", v4hi_ftype_v4hi_int_int, ARM_BUILTIN_TINSRH); | |
14449 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tinsrw", v2si_ftype_v2si_int_int, ARM_BUILTIN_TINSRW); | |
14450 | ||
14451 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_waccb", di_ftype_v8qi, ARM_BUILTIN_WACCB); | |
14452 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wacch", di_ftype_v4hi, ARM_BUILTIN_WACCH); | |
14453 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_waccw", di_ftype_v2si, ARM_BUILTIN_WACCW); | |
14454 | ||
14455 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmovmskb", int_ftype_v8qi, ARM_BUILTIN_TMOVMSKB); | |
14456 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmovmskh", int_ftype_v4hi, ARM_BUILTIN_TMOVMSKH); | |
14457 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmovmskw", int_ftype_v2si, ARM_BUILTIN_TMOVMSKW); | |
14458 | ||
14459 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wpackhss", v8qi_ftype_v4hi_v4hi, ARM_BUILTIN_WPACKHSS); | |
14460 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wpackhus", v8qi_ftype_v4hi_v4hi, ARM_BUILTIN_WPACKHUS); | |
14461 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wpackwus", v4hi_ftype_v2si_v2si, ARM_BUILTIN_WPACKWUS); | |
14462 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wpackwss", v4hi_ftype_v2si_v2si, ARM_BUILTIN_WPACKWSS); | |
14463 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wpackdus", v2si_ftype_di_di, ARM_BUILTIN_WPACKDUS); | |
14464 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wpackdss", v2si_ftype_di_di, ARM_BUILTIN_WPACKDSS); | |
14465 | ||
14466 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckehub", v4hi_ftype_v8qi, ARM_BUILTIN_WUNPCKEHUB); | |
14467 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckehuh", v2si_ftype_v4hi, ARM_BUILTIN_WUNPCKEHUH); | |
14468 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckehuw", di_ftype_v2si, ARM_BUILTIN_WUNPCKEHUW); | |
14469 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckehsb", v4hi_ftype_v8qi, ARM_BUILTIN_WUNPCKEHSB); | |
14470 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckehsh", v2si_ftype_v4hi, ARM_BUILTIN_WUNPCKEHSH); | |
14471 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckehsw", di_ftype_v2si, ARM_BUILTIN_WUNPCKEHSW); | |
14472 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckelub", v4hi_ftype_v8qi, ARM_BUILTIN_WUNPCKELUB); | |
14473 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckeluh", v2si_ftype_v4hi, ARM_BUILTIN_WUNPCKELUH); | |
14474 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckeluw", di_ftype_v2si, ARM_BUILTIN_WUNPCKELUW); | |
14475 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckelsb", v4hi_ftype_v8qi, ARM_BUILTIN_WUNPCKELSB); | |
14476 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckelsh", v2si_ftype_v4hi, ARM_BUILTIN_WUNPCKELSH); | |
14477 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckelsw", di_ftype_v2si, ARM_BUILTIN_WUNPCKELSW); | |
14478 | ||
14479 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wmacs", di_ftype_di_v4hi_v4hi, ARM_BUILTIN_WMACS); | |
14480 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wmacsz", di_ftype_v4hi_v4hi, ARM_BUILTIN_WMACSZ); | |
14481 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wmacu", di_ftype_di_v4hi_v4hi, ARM_BUILTIN_WMACU); | |
14482 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wmacuz", di_ftype_v4hi_v4hi, ARM_BUILTIN_WMACUZ); | |
14483 | ||
14484 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_walign", v8qi_ftype_v8qi_v8qi_int, ARM_BUILTIN_WALIGN); | |
14485 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmia", di_ftype_di_int_int, ARM_BUILTIN_TMIA); | |
14486 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmiaph", di_ftype_di_int_int, ARM_BUILTIN_TMIAPH); | |
14487 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmiabb", di_ftype_di_int_int, ARM_BUILTIN_TMIABB); | |
14488 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmiabt", di_ftype_di_int_int, ARM_BUILTIN_TMIABT); | |
14489 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmiatb", di_ftype_di_int_int, ARM_BUILTIN_TMIATB); | |
14490 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmiatt", di_ftype_di_int_int, ARM_BUILTIN_TMIATT); | |
14491 | } | |
14492 | ||
f655717d | 14493 | static void |
14494 | arm_init_tls_builtins (void) | |
14495 | { | |
14496 | tree ftype; | |
14497 | tree nothrow = tree_cons (get_identifier ("nothrow"), NULL, NULL); | |
14498 | tree const_nothrow = tree_cons (get_identifier ("const"), NULL, nothrow); | |
14499 | ||
14500 | ftype = build_function_type (ptr_type_node, void_list_node); | |
54be5d7e | 14501 | add_builtin_function ("__builtin_thread_pointer", ftype, |
14502 | ARM_BUILTIN_THREAD_POINTER, BUILT_IN_MD, | |
14503 | NULL, const_nothrow); | |
f655717d | 14504 | } |
14505 | ||
d98a3884 | 14506 | typedef enum { |
14507 | T_V8QI = 0x0001, | |
14508 | T_V4HI = 0x0002, | |
14509 | T_V2SI = 0x0004, | |
14510 | T_V2SF = 0x0008, | |
14511 | T_DI = 0x0010, | |
14512 | T_V16QI = 0x0020, | |
14513 | T_V8HI = 0x0040, | |
14514 | T_V4SI = 0x0080, | |
14515 | T_V4SF = 0x0100, | |
14516 | T_V2DI = 0x0200, | |
14517 | T_TI = 0x0400, | |
14518 | T_EI = 0x0800, | |
14519 | T_OI = 0x1000 | |
14520 | } neon_builtin_type_bits; | |
14521 | ||
14522 | #define v8qi_UP T_V8QI | |
14523 | #define v4hi_UP T_V4HI | |
14524 | #define v2si_UP T_V2SI | |
14525 | #define v2sf_UP T_V2SF | |
14526 | #define di_UP T_DI | |
14527 | #define v16qi_UP T_V16QI | |
14528 | #define v8hi_UP T_V8HI | |
14529 | #define v4si_UP T_V4SI | |
14530 | #define v4sf_UP T_V4SF | |
14531 | #define v2di_UP T_V2DI | |
14532 | #define ti_UP T_TI | |
14533 | #define ei_UP T_EI | |
14534 | #define oi_UP T_OI | |
14535 | ||
14536 | #define UP(X) X##_UP | |
14537 | ||
14538 | #define T_MAX 13 | |
14539 | ||
14540 | typedef enum { | |
14541 | NEON_BINOP, | |
14542 | NEON_TERNOP, | |
14543 | NEON_UNOP, | |
14544 | NEON_GETLANE, | |
14545 | NEON_SETLANE, | |
14546 | NEON_CREATE, | |
14547 | NEON_DUP, | |
14548 | NEON_DUPLANE, | |
14549 | NEON_COMBINE, | |
14550 | NEON_SPLIT, | |
14551 | NEON_LANEMUL, | |
14552 | NEON_LANEMULL, | |
14553 | NEON_LANEMULH, | |
14554 | NEON_LANEMAC, | |
14555 | NEON_SCALARMUL, | |
14556 | NEON_SCALARMULL, | |
14557 | NEON_SCALARMULH, | |
14558 | NEON_SCALARMAC, | |
14559 | NEON_CONVERT, | |
14560 | NEON_FIXCONV, | |
14561 | NEON_SELECT, | |
14562 | NEON_RESULTPAIR, | |
14563 | NEON_REINTERP, | |
14564 | NEON_VTBL, | |
14565 | NEON_VTBX, | |
14566 | NEON_LOAD1, | |
14567 | NEON_LOAD1LANE, | |
14568 | NEON_STORE1, | |
14569 | NEON_STORE1LANE, | |
14570 | NEON_LOADSTRUCT, | |
14571 | NEON_LOADSTRUCTLANE, | |
14572 | NEON_STORESTRUCT, | |
14573 | NEON_STORESTRUCTLANE, | |
14574 | NEON_LOGICBINOP, | |
14575 | NEON_SHIFTINSERT, | |
14576 | NEON_SHIFTIMM, | |
14577 | NEON_SHIFTACC | |
14578 | } neon_itype; | |
14579 | ||
14580 | typedef struct { | |
14581 | const char *name; | |
14582 | const neon_itype itype; | |
14583 | const neon_builtin_type_bits bits; | |
14584 | const enum insn_code codes[T_MAX]; | |
14585 | const unsigned int num_vars; | |
14586 | unsigned int base_fcode; | |
14587 | } neon_builtin_datum; | |
14588 | ||
14589 | #define CF(N,X) CODE_FOR_neon_##N##X | |
14590 | ||
14591 | #define VAR1(T, N, A) \ | |
14592 | #N, NEON_##T, UP (A), { CF (N, A) }, 1, 0 | |
14593 | #define VAR2(T, N, A, B) \ | |
14594 | #N, NEON_##T, UP (A) | UP (B), { CF (N, A), CF (N, B) }, 2, 0 | |
14595 | #define VAR3(T, N, A, B, C) \ | |
14596 | #N, NEON_##T, UP (A) | UP (B) | UP (C), \ | |
14597 | { CF (N, A), CF (N, B), CF (N, C) }, 3, 0 | |
14598 | #define VAR4(T, N, A, B, C, D) \ | |
14599 | #N, NEON_##T, UP (A) | UP (B) | UP (C) | UP (D), \ | |
14600 | { CF (N, A), CF (N, B), CF (N, C), CF (N, D) }, 4, 0 | |
14601 | #define VAR5(T, N, A, B, C, D, E) \ | |
14602 | #N, NEON_##T, UP (A) | UP (B) | UP (C) | UP (D) | UP (E), \ | |
14603 | { CF (N, A), CF (N, B), CF (N, C), CF (N, D), CF (N, E) }, 5, 0 | |
14604 | #define VAR6(T, N, A, B, C, D, E, F) \ | |
14605 | #N, NEON_##T, UP (A) | UP (B) | UP (C) | UP (D) | UP (E) | UP (F), \ | |
14606 | { CF (N, A), CF (N, B), CF (N, C), CF (N, D), CF (N, E), CF (N, F) }, 6, 0 | |
14607 | #define VAR7(T, N, A, B, C, D, E, F, G) \ | |
14608 | #N, NEON_##T, UP (A) | UP (B) | UP (C) | UP (D) | UP (E) | UP (F) | UP (G), \ | |
14609 | { CF (N, A), CF (N, B), CF (N, C), CF (N, D), CF (N, E), CF (N, F), \ | |
14610 | CF (N, G) }, 7, 0 | |
14611 | #define VAR8(T, N, A, B, C, D, E, F, G, H) \ | |
14612 | #N, NEON_##T, UP (A) | UP (B) | UP (C) | UP (D) | UP (E) | UP (F) | UP (G) \ | |
14613 | | UP (H), \ | |
14614 | { CF (N, A), CF (N, B), CF (N, C), CF (N, D), CF (N, E), CF (N, F), \ | |
14615 | CF (N, G), CF (N, H) }, 8, 0 | |
14616 | #define VAR9(T, N, A, B, C, D, E, F, G, H, I) \ | |
14617 | #N, NEON_##T, UP (A) | UP (B) | UP (C) | UP (D) | UP (E) | UP (F) | UP (G) \ | |
14618 | | UP (H) | UP (I), \ | |
14619 | { CF (N, A), CF (N, B), CF (N, C), CF (N, D), CF (N, E), CF (N, F), \ | |
14620 | CF (N, G), CF (N, H), CF (N, I) }, 9, 0 | |
14621 | #define VAR10(T, N, A, B, C, D, E, F, G, H, I, J) \ | |
14622 | #N, NEON_##T, UP (A) | UP (B) | UP (C) | UP (D) | UP (E) | UP (F) | UP (G) \ | |
14623 | | UP (H) | UP (I) | UP (J), \ | |
14624 | { CF (N, A), CF (N, B), CF (N, C), CF (N, D), CF (N, E), CF (N, F), \ | |
14625 | CF (N, G), CF (N, H), CF (N, I), CF (N, J) }, 10, 0 | |
14626 | ||
14627 | /* The mode entries in the following table correspond to the "key" type of the | |
14628 | instruction variant, i.e. equivalent to that which would be specified after | |
14629 | the assembler mnemonic, which usually refers to the last vector operand. | |
14630 | (Signed/unsigned/polynomial types are not differentiated between though, and | |
14631 | are all mapped onto the same mode for a given element size.) The modes | |
14632 | listed per instruction should be the same as those defined for that | |
14633 | instruction's pattern in neon.md. | |
14634 | WARNING: Variants should be listed in the same increasing order as | |
14635 | neon_builtin_type_bits. */ | |
14636 | ||
14637 | static neon_builtin_datum neon_builtin_data[] = | |
14638 | { | |
14639 | { VAR10 (BINOP, vadd, | |
14640 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14641 | { VAR3 (BINOP, vaddl, v8qi, v4hi, v2si) }, | |
14642 | { VAR3 (BINOP, vaddw, v8qi, v4hi, v2si) }, | |
14643 | { VAR6 (BINOP, vhadd, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14644 | { VAR8 (BINOP, vqadd, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14645 | { VAR3 (BINOP, vaddhn, v8hi, v4si, v2di) }, | |
14646 | { VAR8 (BINOP, vmul, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14647 | { VAR8 (TERNOP, vmla, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14648 | { VAR3 (TERNOP, vmlal, v8qi, v4hi, v2si) }, | |
14649 | { VAR8 (TERNOP, vmls, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14650 | { VAR3 (TERNOP, vmlsl, v8qi, v4hi, v2si) }, | |
14651 | { VAR4 (BINOP, vqdmulh, v4hi, v2si, v8hi, v4si) }, | |
14652 | { VAR2 (TERNOP, vqdmlal, v4hi, v2si) }, | |
14653 | { VAR2 (TERNOP, vqdmlsl, v4hi, v2si) }, | |
14654 | { VAR3 (BINOP, vmull, v8qi, v4hi, v2si) }, | |
14655 | { VAR2 (SCALARMULL, vmull_n, v4hi, v2si) }, | |
14656 | { VAR2 (LANEMULL, vmull_lane, v4hi, v2si) }, | |
14657 | { VAR2 (SCALARMULL, vqdmull_n, v4hi, v2si) }, | |
14658 | { VAR2 (LANEMULL, vqdmull_lane, v4hi, v2si) }, | |
14659 | { VAR4 (SCALARMULH, vqdmulh_n, v4hi, v2si, v8hi, v4si) }, | |
14660 | { VAR4 (LANEMULH, vqdmulh_lane, v4hi, v2si, v8hi, v4si) }, | |
14661 | { VAR2 (BINOP, vqdmull, v4hi, v2si) }, | |
14662 | { VAR8 (BINOP, vshl, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14663 | { VAR8 (BINOP, vqshl, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14664 | { VAR8 (SHIFTIMM, vshr_n, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14665 | { VAR3 (SHIFTIMM, vshrn_n, v8hi, v4si, v2di) }, | |
14666 | { VAR3 (SHIFTIMM, vqshrn_n, v8hi, v4si, v2di) }, | |
14667 | { VAR3 (SHIFTIMM, vqshrun_n, v8hi, v4si, v2di) }, | |
14668 | { VAR8 (SHIFTIMM, vshl_n, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14669 | { VAR8 (SHIFTIMM, vqshl_n, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14670 | { VAR8 (SHIFTIMM, vqshlu_n, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14671 | { VAR3 (SHIFTIMM, vshll_n, v8qi, v4hi, v2si) }, | |
14672 | { VAR8 (SHIFTACC, vsra_n, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14673 | { VAR10 (BINOP, vsub, | |
14674 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14675 | { VAR3 (BINOP, vsubl, v8qi, v4hi, v2si) }, | |
14676 | { VAR3 (BINOP, vsubw, v8qi, v4hi, v2si) }, | |
14677 | { VAR8 (BINOP, vqsub, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14678 | { VAR6 (BINOP, vhsub, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14679 | { VAR3 (BINOP, vsubhn, v8hi, v4si, v2di) }, | |
14680 | { VAR8 (BINOP, vceq, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14681 | { VAR8 (BINOP, vcge, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14682 | { VAR8 (BINOP, vcgt, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14683 | { VAR2 (BINOP, vcage, v2sf, v4sf) }, | |
14684 | { VAR2 (BINOP, vcagt, v2sf, v4sf) }, | |
14685 | { VAR6 (BINOP, vtst, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14686 | { VAR8 (BINOP, vabd, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14687 | { VAR3 (BINOP, vabdl, v8qi, v4hi, v2si) }, | |
14688 | { VAR6 (TERNOP, vaba, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14689 | { VAR3 (TERNOP, vabal, v8qi, v4hi, v2si) }, | |
14690 | { VAR8 (BINOP, vmax, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14691 | { VAR8 (BINOP, vmin, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14692 | { VAR4 (BINOP, vpadd, v8qi, v4hi, v2si, v2sf) }, | |
14693 | { VAR6 (UNOP, vpaddl, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14694 | { VAR6 (BINOP, vpadal, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14695 | { VAR4 (BINOP, vpmax, v8qi, v4hi, v2si, v2sf) }, | |
14696 | { VAR4 (BINOP, vpmin, v8qi, v4hi, v2si, v2sf) }, | |
14697 | { VAR2 (BINOP, vrecps, v2sf, v4sf) }, | |
14698 | { VAR2 (BINOP, vrsqrts, v2sf, v4sf) }, | |
14699 | { VAR8 (SHIFTINSERT, vsri_n, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14700 | { VAR8 (SHIFTINSERT, vsli_n, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14701 | { VAR8 (UNOP, vabs, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14702 | { VAR6 (UNOP, vqabs, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14703 | { VAR8 (UNOP, vneg, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14704 | { VAR6 (UNOP, vqneg, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14705 | { VAR6 (UNOP, vcls, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14706 | { VAR6 (UNOP, vclz, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14707 | { VAR2 (UNOP, vcnt, v8qi, v16qi) }, | |
14708 | { VAR4 (UNOP, vrecpe, v2si, v2sf, v4si, v4sf) }, | |
14709 | { VAR4 (UNOP, vrsqrte, v2si, v2sf, v4si, v4sf) }, | |
14710 | { VAR6 (UNOP, vmvn, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14711 | /* FIXME: vget_lane supports more variants than this! */ | |
14712 | { VAR10 (GETLANE, vget_lane, | |
14713 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14714 | { VAR10 (SETLANE, vset_lane, | |
14715 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14716 | { VAR5 (CREATE, vcreate, v8qi, v4hi, v2si, v2sf, di) }, | |
14717 | { VAR10 (DUP, vdup_n, | |
14718 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14719 | { VAR10 (DUPLANE, vdup_lane, | |
14720 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14721 | { VAR5 (COMBINE, vcombine, v8qi, v4hi, v2si, v2sf, di) }, | |
14722 | { VAR5 (SPLIT, vget_high, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14723 | { VAR5 (SPLIT, vget_low, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14724 | { VAR3 (UNOP, vmovn, v8hi, v4si, v2di) }, | |
14725 | { VAR3 (UNOP, vqmovn, v8hi, v4si, v2di) }, | |
14726 | { VAR3 (UNOP, vqmovun, v8hi, v4si, v2di) }, | |
14727 | { VAR3 (UNOP, vmovl, v8qi, v4hi, v2si) }, | |
14728 | { VAR6 (LANEMUL, vmul_lane, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14729 | { VAR6 (LANEMAC, vmla_lane, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14730 | { VAR2 (LANEMAC, vmlal_lane, v4hi, v2si) }, | |
14731 | { VAR2 (LANEMAC, vqdmlal_lane, v4hi, v2si) }, | |
14732 | { VAR6 (LANEMAC, vmls_lane, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14733 | { VAR2 (LANEMAC, vmlsl_lane, v4hi, v2si) }, | |
14734 | { VAR2 (LANEMAC, vqdmlsl_lane, v4hi, v2si) }, | |
14735 | { VAR6 (SCALARMUL, vmul_n, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14736 | { VAR6 (SCALARMAC, vmla_n, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14737 | { VAR2 (SCALARMAC, vmlal_n, v4hi, v2si) }, | |
14738 | { VAR2 (SCALARMAC, vqdmlal_n, v4hi, v2si) }, | |
14739 | { VAR6 (SCALARMAC, vmls_n, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14740 | { VAR2 (SCALARMAC, vmlsl_n, v4hi, v2si) }, | |
14741 | { VAR2 (SCALARMAC, vqdmlsl_n, v4hi, v2si) }, | |
14742 | { VAR10 (BINOP, vext, | |
14743 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14744 | { VAR8 (UNOP, vrev64, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14745 | { VAR4 (UNOP, vrev32, v8qi, v4hi, v16qi, v8hi) }, | |
14746 | { VAR2 (UNOP, vrev16, v8qi, v16qi) }, | |
14747 | { VAR4 (CONVERT, vcvt, v2si, v2sf, v4si, v4sf) }, | |
14748 | { VAR4 (FIXCONV, vcvt_n, v2si, v2sf, v4si, v4sf) }, | |
14749 | { VAR10 (SELECT, vbsl, | |
14750 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14751 | { VAR1 (VTBL, vtbl1, v8qi) }, | |
14752 | { VAR1 (VTBL, vtbl2, v8qi) }, | |
14753 | { VAR1 (VTBL, vtbl3, v8qi) }, | |
14754 | { VAR1 (VTBL, vtbl4, v8qi) }, | |
14755 | { VAR1 (VTBX, vtbx1, v8qi) }, | |
14756 | { VAR1 (VTBX, vtbx2, v8qi) }, | |
14757 | { VAR1 (VTBX, vtbx3, v8qi) }, | |
14758 | { VAR1 (VTBX, vtbx4, v8qi) }, | |
14759 | { VAR8 (RESULTPAIR, vtrn, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14760 | { VAR8 (RESULTPAIR, vzip, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14761 | { VAR8 (RESULTPAIR, vuzp, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14762 | { VAR5 (REINTERP, vreinterpretv8qi, v8qi, v4hi, v2si, v2sf, di) }, | |
14763 | { VAR5 (REINTERP, vreinterpretv4hi, v8qi, v4hi, v2si, v2sf, di) }, | |
14764 | { VAR5 (REINTERP, vreinterpretv2si, v8qi, v4hi, v2si, v2sf, di) }, | |
14765 | { VAR5 (REINTERP, vreinterpretv2sf, v8qi, v4hi, v2si, v2sf, di) }, | |
14766 | { VAR5 (REINTERP, vreinterpretdi, v8qi, v4hi, v2si, v2sf, di) }, | |
14767 | { VAR5 (REINTERP, vreinterpretv16qi, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14768 | { VAR5 (REINTERP, vreinterpretv8hi, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14769 | { VAR5 (REINTERP, vreinterpretv4si, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14770 | { VAR5 (REINTERP, vreinterpretv4sf, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14771 | { VAR5 (REINTERP, vreinterpretv2di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14772 | { VAR10 (LOAD1, vld1, | |
14773 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14774 | { VAR10 (LOAD1LANE, vld1_lane, | |
14775 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14776 | { VAR10 (LOAD1, vld1_dup, | |
14777 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14778 | { VAR10 (STORE1, vst1, | |
14779 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14780 | { VAR10 (STORE1LANE, vst1_lane, | |
14781 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14782 | { VAR9 (LOADSTRUCT, | |
14783 | vld2, v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf) }, | |
14784 | { VAR7 (LOADSTRUCTLANE, vld2_lane, | |
14785 | v8qi, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14786 | { VAR5 (LOADSTRUCT, vld2_dup, v8qi, v4hi, v2si, v2sf, di) }, | |
14787 | { VAR9 (STORESTRUCT, vst2, | |
14788 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf) }, | |
14789 | { VAR7 (STORESTRUCTLANE, vst2_lane, | |
14790 | v8qi, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14791 | { VAR9 (LOADSTRUCT, | |
14792 | vld3, v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf) }, | |
14793 | { VAR7 (LOADSTRUCTLANE, vld3_lane, | |
14794 | v8qi, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14795 | { VAR5 (LOADSTRUCT, vld3_dup, v8qi, v4hi, v2si, v2sf, di) }, | |
14796 | { VAR9 (STORESTRUCT, vst3, | |
14797 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf) }, | |
14798 | { VAR7 (STORESTRUCTLANE, vst3_lane, | |
14799 | v8qi, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14800 | { VAR9 (LOADSTRUCT, vld4, | |
14801 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf) }, | |
14802 | { VAR7 (LOADSTRUCTLANE, vld4_lane, | |
14803 | v8qi, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14804 | { VAR5 (LOADSTRUCT, vld4_dup, v8qi, v4hi, v2si, v2sf, di) }, | |
14805 | { VAR9 (STORESTRUCT, vst4, | |
14806 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf) }, | |
14807 | { VAR7 (STORESTRUCTLANE, vst4_lane, | |
14808 | v8qi, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14809 | { VAR10 (LOGICBINOP, vand, | |
14810 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14811 | { VAR10 (LOGICBINOP, vorr, | |
14812 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14813 | { VAR10 (BINOP, veor, | |
14814 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14815 | { VAR10 (LOGICBINOP, vbic, | |
14816 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14817 | { VAR10 (LOGICBINOP, vorn, | |
14818 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) } | |
14819 | }; | |
14820 | ||
14821 | #undef CF | |
14822 | #undef VAR1 | |
14823 | #undef VAR2 | |
14824 | #undef VAR3 | |
14825 | #undef VAR4 | |
14826 | #undef VAR5 | |
14827 | #undef VAR6 | |
14828 | #undef VAR7 | |
14829 | #undef VAR8 | |
14830 | #undef VAR9 | |
14831 | #undef VAR10 | |
14832 | ||
14833 | static void | |
14834 | arm_init_neon_builtins (void) | |
14835 | { | |
14836 | unsigned int i, fcode = ARM_BUILTIN_NEON_BASE; | |
14837 | ||
14838 | /* Create distinguished type nodes for NEON vector element types, | |
14839 | and pointers to values of such types, so we can detect them later. */ | |
14840 | tree neon_intQI_type_node = make_signed_type (GET_MODE_PRECISION (QImode)); | |
14841 | tree neon_intHI_type_node = make_signed_type (GET_MODE_PRECISION (HImode)); | |
14842 | tree neon_polyQI_type_node = make_signed_type (GET_MODE_PRECISION (QImode)); | |
14843 | tree neon_polyHI_type_node = make_signed_type (GET_MODE_PRECISION (HImode)); | |
14844 | tree neon_intSI_type_node = make_signed_type (GET_MODE_PRECISION (SImode)); | |
14845 | tree neon_intDI_type_node = make_signed_type (GET_MODE_PRECISION (DImode)); | |
14846 | tree neon_float_type_node = make_node (REAL_TYPE); | |
14847 | TYPE_PRECISION (neon_float_type_node) = FLOAT_TYPE_SIZE; | |
14848 | layout_type (neon_float_type_node); | |
14849 | ||
14850 | /* Define typedefs which exactly correspond to the modes we are basing vector | |
14851 | types on. If you change these names you'll need to change | |
eddcdde1 | 14852 | the table used by arm_mangle_type too. */ |
d98a3884 | 14853 | (*lang_hooks.types.register_builtin_type) (neon_intQI_type_node, |
14854 | "__builtin_neon_qi"); | |
14855 | (*lang_hooks.types.register_builtin_type) (neon_intHI_type_node, | |
14856 | "__builtin_neon_hi"); | |
14857 | (*lang_hooks.types.register_builtin_type) (neon_intSI_type_node, | |
14858 | "__builtin_neon_si"); | |
14859 | (*lang_hooks.types.register_builtin_type) (neon_float_type_node, | |
14860 | "__builtin_neon_sf"); | |
14861 | (*lang_hooks.types.register_builtin_type) (neon_intDI_type_node, | |
14862 | "__builtin_neon_di"); | |
14863 | ||
14864 | (*lang_hooks.types.register_builtin_type) (neon_polyQI_type_node, | |
14865 | "__builtin_neon_poly8"); | |
14866 | (*lang_hooks.types.register_builtin_type) (neon_polyHI_type_node, | |
14867 | "__builtin_neon_poly16"); | |
14868 | ||
14869 | tree intQI_pointer_node = build_pointer_type (neon_intQI_type_node); | |
14870 | tree intHI_pointer_node = build_pointer_type (neon_intHI_type_node); | |
14871 | tree intSI_pointer_node = build_pointer_type (neon_intSI_type_node); | |
14872 | tree intDI_pointer_node = build_pointer_type (neon_intDI_type_node); | |
14873 | tree float_pointer_node = build_pointer_type (neon_float_type_node); | |
14874 | ||
14875 | /* Next create constant-qualified versions of the above types. */ | |
14876 | tree const_intQI_node = build_qualified_type (neon_intQI_type_node, | |
14877 | TYPE_QUAL_CONST); | |
14878 | tree const_intHI_node = build_qualified_type (neon_intHI_type_node, | |
14879 | TYPE_QUAL_CONST); | |
14880 | tree const_intSI_node = build_qualified_type (neon_intSI_type_node, | |
14881 | TYPE_QUAL_CONST); | |
14882 | tree const_intDI_node = build_qualified_type (neon_intDI_type_node, | |
14883 | TYPE_QUAL_CONST); | |
14884 | tree const_float_node = build_qualified_type (neon_float_type_node, | |
14885 | TYPE_QUAL_CONST); | |
14886 | ||
14887 | tree const_intQI_pointer_node = build_pointer_type (const_intQI_node); | |
14888 | tree const_intHI_pointer_node = build_pointer_type (const_intHI_node); | |
14889 | tree const_intSI_pointer_node = build_pointer_type (const_intSI_node); | |
14890 | tree const_intDI_pointer_node = build_pointer_type (const_intDI_node); | |
14891 | tree const_float_pointer_node = build_pointer_type (const_float_node); | |
14892 | ||
14893 | /* Now create vector types based on our NEON element types. */ | |
14894 | /* 64-bit vectors. */ | |
14895 | tree V8QI_type_node = | |
14896 | build_vector_type_for_mode (neon_intQI_type_node, V8QImode); | |
14897 | tree V4HI_type_node = | |
14898 | build_vector_type_for_mode (neon_intHI_type_node, V4HImode); | |
14899 | tree V2SI_type_node = | |
14900 | build_vector_type_for_mode (neon_intSI_type_node, V2SImode); | |
14901 | tree V2SF_type_node = | |
14902 | build_vector_type_for_mode (neon_float_type_node, V2SFmode); | |
14903 | /* 128-bit vectors. */ | |
14904 | tree V16QI_type_node = | |
14905 | build_vector_type_for_mode (neon_intQI_type_node, V16QImode); | |
14906 | tree V8HI_type_node = | |
14907 | build_vector_type_for_mode (neon_intHI_type_node, V8HImode); | |
14908 | tree V4SI_type_node = | |
14909 | build_vector_type_for_mode (neon_intSI_type_node, V4SImode); | |
14910 | tree V4SF_type_node = | |
14911 | build_vector_type_for_mode (neon_float_type_node, V4SFmode); | |
14912 | tree V2DI_type_node = | |
14913 | build_vector_type_for_mode (neon_intDI_type_node, V2DImode); | |
14914 | ||
14915 | /* Unsigned integer types for various mode sizes. */ | |
14916 | tree intUQI_type_node = make_unsigned_type (GET_MODE_PRECISION (QImode)); | |
14917 | tree intUHI_type_node = make_unsigned_type (GET_MODE_PRECISION (HImode)); | |
14918 | tree intUSI_type_node = make_unsigned_type (GET_MODE_PRECISION (SImode)); | |
14919 | tree intUDI_type_node = make_unsigned_type (GET_MODE_PRECISION (DImode)); | |
14920 | ||
14921 | (*lang_hooks.types.register_builtin_type) (intUQI_type_node, | |
14922 | "__builtin_neon_uqi"); | |
14923 | (*lang_hooks.types.register_builtin_type) (intUHI_type_node, | |
14924 | "__builtin_neon_uhi"); | |
14925 | (*lang_hooks.types.register_builtin_type) (intUSI_type_node, | |
14926 | "__builtin_neon_usi"); | |
14927 | (*lang_hooks.types.register_builtin_type) (intUDI_type_node, | |
14928 | "__builtin_neon_udi"); | |
14929 | ||
14930 | /* Opaque integer types for structures of vectors. */ | |
14931 | tree intEI_type_node = make_signed_type (GET_MODE_PRECISION (EImode)); | |
14932 | tree intOI_type_node = make_signed_type (GET_MODE_PRECISION (OImode)); | |
14933 | tree intCI_type_node = make_signed_type (GET_MODE_PRECISION (CImode)); | |
14934 | tree intXI_type_node = make_signed_type (GET_MODE_PRECISION (XImode)); | |
14935 | ||
14936 | (*lang_hooks.types.register_builtin_type) (intTI_type_node, | |
14937 | "__builtin_neon_ti"); | |
14938 | (*lang_hooks.types.register_builtin_type) (intEI_type_node, | |
14939 | "__builtin_neon_ei"); | |
14940 | (*lang_hooks.types.register_builtin_type) (intOI_type_node, | |
14941 | "__builtin_neon_oi"); | |
14942 | (*lang_hooks.types.register_builtin_type) (intCI_type_node, | |
14943 | "__builtin_neon_ci"); | |
14944 | (*lang_hooks.types.register_builtin_type) (intXI_type_node, | |
14945 | "__builtin_neon_xi"); | |
14946 | ||
14947 | /* Pointers to vector types. */ | |
14948 | tree V8QI_pointer_node = build_pointer_type (V8QI_type_node); | |
14949 | tree V4HI_pointer_node = build_pointer_type (V4HI_type_node); | |
14950 | tree V2SI_pointer_node = build_pointer_type (V2SI_type_node); | |
14951 | tree V2SF_pointer_node = build_pointer_type (V2SF_type_node); | |
14952 | tree V16QI_pointer_node = build_pointer_type (V16QI_type_node); | |
14953 | tree V8HI_pointer_node = build_pointer_type (V8HI_type_node); | |
14954 | tree V4SI_pointer_node = build_pointer_type (V4SI_type_node); | |
14955 | tree V4SF_pointer_node = build_pointer_type (V4SF_type_node); | |
14956 | tree V2DI_pointer_node = build_pointer_type (V2DI_type_node); | |
14957 | ||
14958 | /* Operations which return results as pairs. */ | |
14959 | tree void_ftype_pv8qi_v8qi_v8qi = | |
14960 | build_function_type_list (void_type_node, V8QI_pointer_node, V8QI_type_node, | |
14961 | V8QI_type_node, NULL); | |
14962 | tree void_ftype_pv4hi_v4hi_v4hi = | |
14963 | build_function_type_list (void_type_node, V4HI_pointer_node, V4HI_type_node, | |
14964 | V4HI_type_node, NULL); | |
14965 | tree void_ftype_pv2si_v2si_v2si = | |
14966 | build_function_type_list (void_type_node, V2SI_pointer_node, V2SI_type_node, | |
14967 | V2SI_type_node, NULL); | |
14968 | tree void_ftype_pv2sf_v2sf_v2sf = | |
14969 | build_function_type_list (void_type_node, V2SF_pointer_node, V2SF_type_node, | |
14970 | V2SF_type_node, NULL); | |
14971 | tree void_ftype_pdi_di_di = | |
14972 | build_function_type_list (void_type_node, intDI_pointer_node, | |
14973 | neon_intDI_type_node, neon_intDI_type_node, NULL); | |
14974 | tree void_ftype_pv16qi_v16qi_v16qi = | |
14975 | build_function_type_list (void_type_node, V16QI_pointer_node, | |
14976 | V16QI_type_node, V16QI_type_node, NULL); | |
14977 | tree void_ftype_pv8hi_v8hi_v8hi = | |
14978 | build_function_type_list (void_type_node, V8HI_pointer_node, V8HI_type_node, | |
14979 | V8HI_type_node, NULL); | |
14980 | tree void_ftype_pv4si_v4si_v4si = | |
14981 | build_function_type_list (void_type_node, V4SI_pointer_node, V4SI_type_node, | |
14982 | V4SI_type_node, NULL); | |
14983 | tree void_ftype_pv4sf_v4sf_v4sf = | |
14984 | build_function_type_list (void_type_node, V4SF_pointer_node, V4SF_type_node, | |
14985 | V4SF_type_node, NULL); | |
14986 | tree void_ftype_pv2di_v2di_v2di = | |
14987 | build_function_type_list (void_type_node, V2DI_pointer_node, V2DI_type_node, | |
14988 | V2DI_type_node, NULL); | |
14989 | ||
14990 | tree reinterp_ftype_dreg[5][5]; | |
14991 | tree reinterp_ftype_qreg[5][5]; | |
14992 | tree dreg_types[5], qreg_types[5]; | |
14993 | ||
14994 | dreg_types[0] = V8QI_type_node; | |
14995 | dreg_types[1] = V4HI_type_node; | |
14996 | dreg_types[2] = V2SI_type_node; | |
14997 | dreg_types[3] = V2SF_type_node; | |
14998 | dreg_types[4] = neon_intDI_type_node; | |
14999 | ||
15000 | qreg_types[0] = V16QI_type_node; | |
15001 | qreg_types[1] = V8HI_type_node; | |
15002 | qreg_types[2] = V4SI_type_node; | |
15003 | qreg_types[3] = V4SF_type_node; | |
15004 | qreg_types[4] = V2DI_type_node; | |
15005 | ||
15006 | for (i = 0; i < 5; i++) | |
15007 | { | |
15008 | int j; | |
15009 | for (j = 0; j < 5; j++) | |
15010 | { | |
15011 | reinterp_ftype_dreg[i][j] | |
15012 | = build_function_type_list (dreg_types[i], dreg_types[j], NULL); | |
15013 | reinterp_ftype_qreg[i][j] | |
15014 | = build_function_type_list (qreg_types[i], qreg_types[j], NULL); | |
15015 | } | |
15016 | } | |
15017 | ||
15018 | for (i = 0; i < ARRAY_SIZE (neon_builtin_data); i++) | |
15019 | { | |
15020 | neon_builtin_datum *d = &neon_builtin_data[i]; | |
15021 | unsigned int j, codeidx = 0; | |
15022 | ||
15023 | d->base_fcode = fcode; | |
15024 | ||
15025 | for (j = 0; j < T_MAX; j++) | |
15026 | { | |
15027 | const char* const modenames[] = { | |
15028 | "v8qi", "v4hi", "v2si", "v2sf", "di", | |
15029 | "v16qi", "v8hi", "v4si", "v4sf", "v2di" | |
15030 | }; | |
15031 | char namebuf[60]; | |
15032 | tree ftype = NULL; | |
15033 | enum insn_code icode; | |
15034 | int is_load = 0, is_store = 0; | |
15035 | ||
15036 | if ((d->bits & (1 << j)) == 0) | |
15037 | continue; | |
15038 | ||
15039 | icode = d->codes[codeidx++]; | |
15040 | ||
15041 | switch (d->itype) | |
15042 | { | |
15043 | case NEON_LOAD1: | |
15044 | case NEON_LOAD1LANE: | |
15045 | case NEON_LOADSTRUCT: | |
15046 | case NEON_LOADSTRUCTLANE: | |
15047 | is_load = 1; | |
15048 | /* Fall through. */ | |
15049 | case NEON_STORE1: | |
15050 | case NEON_STORE1LANE: | |
15051 | case NEON_STORESTRUCT: | |
15052 | case NEON_STORESTRUCTLANE: | |
15053 | if (!is_load) | |
15054 | is_store = 1; | |
15055 | /* Fall through. */ | |
15056 | case NEON_UNOP: | |
15057 | case NEON_BINOP: | |
15058 | case NEON_LOGICBINOP: | |
15059 | case NEON_SHIFTINSERT: | |
15060 | case NEON_TERNOP: | |
15061 | case NEON_GETLANE: | |
15062 | case NEON_SETLANE: | |
15063 | case NEON_CREATE: | |
15064 | case NEON_DUP: | |
15065 | case NEON_DUPLANE: | |
15066 | case NEON_SHIFTIMM: | |
15067 | case NEON_SHIFTACC: | |
15068 | case NEON_COMBINE: | |
15069 | case NEON_SPLIT: | |
15070 | case NEON_CONVERT: | |
15071 | case NEON_FIXCONV: | |
15072 | case NEON_LANEMUL: | |
15073 | case NEON_LANEMULL: | |
15074 | case NEON_LANEMULH: | |
15075 | case NEON_LANEMAC: | |
15076 | case NEON_SCALARMUL: | |
15077 | case NEON_SCALARMULL: | |
15078 | case NEON_SCALARMULH: | |
15079 | case NEON_SCALARMAC: | |
15080 | case NEON_SELECT: | |
15081 | case NEON_VTBL: | |
15082 | case NEON_VTBX: | |
15083 | { | |
15084 | int k; | |
15085 | tree return_type = void_type_node, args = void_list_node; | |
15086 | ||
15087 | /* Build a function type directly from the insn_data for this | |
15088 | builtin. The build_function_type() function takes care of | |
15089 | removing duplicates for us. */ | |
15090 | for (k = insn_data[icode].n_operands - 1; k >= 0; k--) | |
15091 | { | |
15092 | tree eltype; | |
15093 | ||
15094 | if (is_load && k == 1) | |
15095 | { | |
15096 | /* Neon load patterns always have the memory operand | |
15097 | (a SImode pointer) in the operand 1 position. We | |
15098 | want a const pointer to the element type in that | |
15099 | position. */ | |
15100 | gcc_assert (insn_data[icode].operand[k].mode == SImode); | |
15101 | ||
15102 | switch (1 << j) | |
15103 | { | |
15104 | case T_V8QI: | |
15105 | case T_V16QI: | |
15106 | eltype = const_intQI_pointer_node; | |
15107 | break; | |
15108 | ||
15109 | case T_V4HI: | |
15110 | case T_V8HI: | |
15111 | eltype = const_intHI_pointer_node; | |
15112 | break; | |
15113 | ||
15114 | case T_V2SI: | |
15115 | case T_V4SI: | |
15116 | eltype = const_intSI_pointer_node; | |
15117 | break; | |
15118 | ||
15119 | case T_V2SF: | |
15120 | case T_V4SF: | |
15121 | eltype = const_float_pointer_node; | |
15122 | break; | |
15123 | ||
15124 | case T_DI: | |
15125 | case T_V2DI: | |
15126 | eltype = const_intDI_pointer_node; | |
15127 | break; | |
15128 | ||
15129 | default: gcc_unreachable (); | |
15130 | } | |
15131 | } | |
15132 | else if (is_store && k == 0) | |
15133 | { | |
15134 | /* Similarly, Neon store patterns use operand 0 as | |
15135 | the memory location to store to (a SImode pointer). | |
15136 | Use a pointer to the element type of the store in | |
15137 | that position. */ | |
15138 | gcc_assert (insn_data[icode].operand[k].mode == SImode); | |
15139 | ||
15140 | switch (1 << j) | |
15141 | { | |
15142 | case T_V8QI: | |
15143 | case T_V16QI: | |
15144 | eltype = intQI_pointer_node; | |
15145 | break; | |
15146 | ||
15147 | case T_V4HI: | |
15148 | case T_V8HI: | |
15149 | eltype = intHI_pointer_node; | |
15150 | break; | |
15151 | ||
15152 | case T_V2SI: | |
15153 | case T_V4SI: | |
15154 | eltype = intSI_pointer_node; | |
15155 | break; | |
15156 | ||
15157 | case T_V2SF: | |
15158 | case T_V4SF: | |
15159 | eltype = float_pointer_node; | |
15160 | break; | |
15161 | ||
15162 | case T_DI: | |
15163 | case T_V2DI: | |
15164 | eltype = intDI_pointer_node; | |
15165 | break; | |
15166 | ||
15167 | default: gcc_unreachable (); | |
15168 | } | |
15169 | } | |
15170 | else | |
15171 | { | |
15172 | switch (insn_data[icode].operand[k].mode) | |
15173 | { | |
15174 | case VOIDmode: eltype = void_type_node; break; | |
15175 | /* Scalars. */ | |
15176 | case QImode: eltype = neon_intQI_type_node; break; | |
15177 | case HImode: eltype = neon_intHI_type_node; break; | |
15178 | case SImode: eltype = neon_intSI_type_node; break; | |
15179 | case SFmode: eltype = neon_float_type_node; break; | |
15180 | case DImode: eltype = neon_intDI_type_node; break; | |
15181 | case TImode: eltype = intTI_type_node; break; | |
15182 | case EImode: eltype = intEI_type_node; break; | |
15183 | case OImode: eltype = intOI_type_node; break; | |
15184 | case CImode: eltype = intCI_type_node; break; | |
15185 | case XImode: eltype = intXI_type_node; break; | |
15186 | /* 64-bit vectors. */ | |
15187 | case V8QImode: eltype = V8QI_type_node; break; | |
15188 | case V4HImode: eltype = V4HI_type_node; break; | |
15189 | case V2SImode: eltype = V2SI_type_node; break; | |
15190 | case V2SFmode: eltype = V2SF_type_node; break; | |
15191 | /* 128-bit vectors. */ | |
15192 | case V16QImode: eltype = V16QI_type_node; break; | |
15193 | case V8HImode: eltype = V8HI_type_node; break; | |
15194 | case V4SImode: eltype = V4SI_type_node; break; | |
15195 | case V4SFmode: eltype = V4SF_type_node; break; | |
15196 | case V2DImode: eltype = V2DI_type_node; break; | |
15197 | default: gcc_unreachable (); | |
15198 | } | |
15199 | } | |
15200 | ||
15201 | if (k == 0 && !is_store) | |
15202 | return_type = eltype; | |
15203 | else | |
15204 | args = tree_cons (NULL_TREE, eltype, args); | |
15205 | } | |
15206 | ||
15207 | ftype = build_function_type (return_type, args); | |
15208 | } | |
15209 | break; | |
15210 | ||
15211 | case NEON_RESULTPAIR: | |
15212 | { | |
15213 | switch (insn_data[icode].operand[1].mode) | |
15214 | { | |
15215 | case V8QImode: ftype = void_ftype_pv8qi_v8qi_v8qi; break; | |
15216 | case V4HImode: ftype = void_ftype_pv4hi_v4hi_v4hi; break; | |
15217 | case V2SImode: ftype = void_ftype_pv2si_v2si_v2si; break; | |
15218 | case V2SFmode: ftype = void_ftype_pv2sf_v2sf_v2sf; break; | |
15219 | case DImode: ftype = void_ftype_pdi_di_di; break; | |
15220 | case V16QImode: ftype = void_ftype_pv16qi_v16qi_v16qi; break; | |
15221 | case V8HImode: ftype = void_ftype_pv8hi_v8hi_v8hi; break; | |
15222 | case V4SImode: ftype = void_ftype_pv4si_v4si_v4si; break; | |
15223 | case V4SFmode: ftype = void_ftype_pv4sf_v4sf_v4sf; break; | |
15224 | case V2DImode: ftype = void_ftype_pv2di_v2di_v2di; break; | |
15225 | default: gcc_unreachable (); | |
15226 | } | |
15227 | } | |
15228 | break; | |
15229 | ||
15230 | case NEON_REINTERP: | |
15231 | { | |
15232 | /* We iterate over 5 doubleword types, then 5 quadword | |
15233 | types. */ | |
15234 | int rhs = j % 5; | |
15235 | switch (insn_data[icode].operand[0].mode) | |
15236 | { | |
15237 | case V8QImode: ftype = reinterp_ftype_dreg[0][rhs]; break; | |
15238 | case V4HImode: ftype = reinterp_ftype_dreg[1][rhs]; break; | |
15239 | case V2SImode: ftype = reinterp_ftype_dreg[2][rhs]; break; | |
15240 | case V2SFmode: ftype = reinterp_ftype_dreg[3][rhs]; break; | |
15241 | case DImode: ftype = reinterp_ftype_dreg[4][rhs]; break; | |
15242 | case V16QImode: ftype = reinterp_ftype_qreg[0][rhs]; break; | |
15243 | case V8HImode: ftype = reinterp_ftype_qreg[1][rhs]; break; | |
15244 | case V4SImode: ftype = reinterp_ftype_qreg[2][rhs]; break; | |
15245 | case V4SFmode: ftype = reinterp_ftype_qreg[3][rhs]; break; | |
15246 | case V2DImode: ftype = reinterp_ftype_qreg[4][rhs]; break; | |
15247 | default: gcc_unreachable (); | |
15248 | } | |
15249 | } | |
15250 | break; | |
15251 | ||
15252 | default: | |
15253 | gcc_unreachable (); | |
15254 | } | |
15255 | ||
15256 | gcc_assert (ftype != NULL); | |
15257 | ||
15258 | sprintf (namebuf, "__builtin_neon_%s%s", d->name, modenames[j]); | |
15259 | ||
15260 | add_builtin_function (namebuf, ftype, fcode++, BUILT_IN_MD, NULL, | |
15261 | NULL_TREE); | |
15262 | } | |
15263 | } | |
15264 | } | |
15265 | ||
755eb2b4 | 15266 | static void |
15267 | arm_init_builtins (void) | |
15268 | { | |
f655717d | 15269 | arm_init_tls_builtins (); |
15270 | ||
755eb2b4 | 15271 | if (TARGET_REALLY_IWMMXT) |
15272 | arm_init_iwmmxt_builtins (); | |
d98a3884 | 15273 | |
15274 | if (TARGET_NEON) | |
15275 | arm_init_neon_builtins (); | |
755eb2b4 | 15276 | } |
15277 | ||
15278 | /* Errors in the source file can cause expand_expr to return const0_rtx | |
15279 | where we expect a vector. To avoid crashing, use one of the vector | |
15280 | clear instructions. */ | |
15281 | ||
15282 | static rtx | |
15283 | safe_vector_operand (rtx x, enum machine_mode mode) | |
15284 | { | |
15285 | if (x != const0_rtx) | |
15286 | return x; | |
15287 | x = gen_reg_rtx (mode); | |
15288 | ||
15289 | emit_insn (gen_iwmmxt_clrdi (mode == DImode ? x | |
15290 | : gen_rtx_SUBREG (DImode, x, 0))); | |
15291 | return x; | |
15292 | } | |
15293 | ||
15294 | /* Subroutine of arm_expand_builtin to take care of binop insns. */ | |
15295 | ||
15296 | static rtx | |
15297 | arm_expand_binop_builtin (enum insn_code icode, | |
c2f47e15 | 15298 | tree exp, rtx target) |
755eb2b4 | 15299 | { |
15300 | rtx pat; | |
c2f47e15 | 15301 | tree arg0 = CALL_EXPR_ARG (exp, 0); |
15302 | tree arg1 = CALL_EXPR_ARG (exp, 1); | |
8ec3c5c2 | 15303 | rtx op0 = expand_normal (arg0); |
15304 | rtx op1 = expand_normal (arg1); | |
755eb2b4 | 15305 | enum machine_mode tmode = insn_data[icode].operand[0].mode; |
15306 | enum machine_mode mode0 = insn_data[icode].operand[1].mode; | |
15307 | enum machine_mode mode1 = insn_data[icode].operand[2].mode; | |
15308 | ||
15309 | if (VECTOR_MODE_P (mode0)) | |
15310 | op0 = safe_vector_operand (op0, mode0); | |
15311 | if (VECTOR_MODE_P (mode1)) | |
15312 | op1 = safe_vector_operand (op1, mode1); | |
15313 | ||
15314 | if (! target | |
15315 | || GET_MODE (target) != tmode | |
15316 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
15317 | target = gen_reg_rtx (tmode); | |
15318 | ||
ed29c566 | 15319 | gcc_assert (GET_MODE (op0) == mode0 && GET_MODE (op1) == mode1); |
755eb2b4 | 15320 | |
15321 | if (! (*insn_data[icode].operand[1].predicate) (op0, mode0)) | |
15322 | op0 = copy_to_mode_reg (mode0, op0); | |
15323 | if (! (*insn_data[icode].operand[2].predicate) (op1, mode1)) | |
15324 | op1 = copy_to_mode_reg (mode1, op1); | |
15325 | ||
15326 | pat = GEN_FCN (icode) (target, op0, op1); | |
15327 | if (! pat) | |
15328 | return 0; | |
15329 | emit_insn (pat); | |
15330 | return target; | |
15331 | } | |
15332 | ||
15333 | /* Subroutine of arm_expand_builtin to take care of unop insns. */ | |
15334 | ||
15335 | static rtx | |
15336 | arm_expand_unop_builtin (enum insn_code icode, | |
c2f47e15 | 15337 | tree exp, rtx target, int do_load) |
755eb2b4 | 15338 | { |
15339 | rtx pat; | |
c2f47e15 | 15340 | tree arg0 = CALL_EXPR_ARG (exp, 0); |
8ec3c5c2 | 15341 | rtx op0 = expand_normal (arg0); |
755eb2b4 | 15342 | enum machine_mode tmode = insn_data[icode].operand[0].mode; |
15343 | enum machine_mode mode0 = insn_data[icode].operand[1].mode; | |
15344 | ||
15345 | if (! target | |
15346 | || GET_MODE (target) != tmode | |
15347 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
15348 | target = gen_reg_rtx (tmode); | |
15349 | if (do_load) | |
15350 | op0 = gen_rtx_MEM (mode0, copy_to_mode_reg (Pmode, op0)); | |
15351 | else | |
15352 | { | |
15353 | if (VECTOR_MODE_P (mode0)) | |
15354 | op0 = safe_vector_operand (op0, mode0); | |
15355 | ||
15356 | if (! (*insn_data[icode].operand[1].predicate) (op0, mode0)) | |
15357 | op0 = copy_to_mode_reg (mode0, op0); | |
15358 | } | |
15359 | ||
15360 | pat = GEN_FCN (icode) (target, op0); | |
15361 | if (! pat) | |
15362 | return 0; | |
15363 | emit_insn (pat); | |
15364 | return target; | |
15365 | } | |
15366 | ||
d98a3884 | 15367 | static int |
15368 | neon_builtin_compare (const void *a, const void *b) | |
15369 | { | |
15370 | const neon_builtin_datum *key = a; | |
15371 | const neon_builtin_datum *memb = b; | |
15372 | unsigned int soughtcode = key->base_fcode; | |
15373 | ||
15374 | if (soughtcode >= memb->base_fcode | |
15375 | && soughtcode < memb->base_fcode + memb->num_vars) | |
15376 | return 0; | |
15377 | else if (soughtcode < memb->base_fcode) | |
15378 | return -1; | |
15379 | else | |
15380 | return 1; | |
15381 | } | |
15382 | ||
15383 | static enum insn_code | |
15384 | locate_neon_builtin_icode (int fcode, neon_itype *itype) | |
15385 | { | |
15386 | neon_builtin_datum key, *found; | |
15387 | int idx; | |
15388 | ||
15389 | key.base_fcode = fcode; | |
15390 | found = bsearch (&key, &neon_builtin_data[0], ARRAY_SIZE (neon_builtin_data), | |
15391 | sizeof (neon_builtin_data[0]), neon_builtin_compare); | |
15392 | gcc_assert (found); | |
15393 | idx = fcode - (int) found->base_fcode; | |
15394 | gcc_assert (idx >= 0 && idx < T_MAX && idx < (int)found->num_vars); | |
15395 | ||
15396 | if (itype) | |
15397 | *itype = found->itype; | |
15398 | ||
15399 | return found->codes[idx]; | |
15400 | } | |
15401 | ||
15402 | typedef enum { | |
15403 | NEON_ARG_COPY_TO_REG, | |
15404 | NEON_ARG_CONSTANT, | |
15405 | NEON_ARG_STOP | |
15406 | } builtin_arg; | |
15407 | ||
15408 | #define NEON_MAX_BUILTIN_ARGS 5 | |
15409 | ||
15410 | /* Expand a Neon builtin. */ | |
15411 | static rtx | |
15412 | arm_expand_neon_args (rtx target, int icode, int have_retval, | |
15413 | tree exp, ...) | |
15414 | { | |
15415 | va_list ap; | |
15416 | rtx pat; | |
15417 | tree arg[NEON_MAX_BUILTIN_ARGS]; | |
15418 | rtx op[NEON_MAX_BUILTIN_ARGS]; | |
15419 | enum machine_mode tmode = insn_data[icode].operand[0].mode; | |
15420 | enum machine_mode mode[NEON_MAX_BUILTIN_ARGS]; | |
15421 | int argc = 0; | |
15422 | ||
15423 | if (have_retval | |
15424 | && (!target | |
15425 | || GET_MODE (target) != tmode | |
15426 | || !(*insn_data[icode].operand[0].predicate) (target, tmode))) | |
15427 | target = gen_reg_rtx (tmode); | |
15428 | ||
15429 | va_start (ap, exp); | |
15430 | ||
15431 | for (;;) | |
15432 | { | |
15433 | builtin_arg thisarg = va_arg (ap, int); | |
15434 | ||
15435 | if (thisarg == NEON_ARG_STOP) | |
15436 | break; | |
15437 | else | |
15438 | { | |
15439 | arg[argc] = CALL_EXPR_ARG (exp, argc); | |
15440 | op[argc] = expand_normal (arg[argc]); | |
15441 | mode[argc] = insn_data[icode].operand[argc + have_retval].mode; | |
15442 | ||
15443 | switch (thisarg) | |
15444 | { | |
15445 | case NEON_ARG_COPY_TO_REG: | |
15446 | /*gcc_assert (GET_MODE (op[argc]) == mode[argc]);*/ | |
15447 | if (!(*insn_data[icode].operand[argc + have_retval].predicate) | |
15448 | (op[argc], mode[argc])) | |
15449 | op[argc] = copy_to_mode_reg (mode[argc], op[argc]); | |
15450 | break; | |
15451 | ||
15452 | case NEON_ARG_CONSTANT: | |
15453 | /* FIXME: This error message is somewhat unhelpful. */ | |
15454 | if (!(*insn_data[icode].operand[argc + have_retval].predicate) | |
15455 | (op[argc], mode[argc])) | |
15456 | error ("argument must be a constant"); | |
15457 | break; | |
15458 | ||
15459 | case NEON_ARG_STOP: | |
15460 | gcc_unreachable (); | |
15461 | } | |
15462 | ||
15463 | argc++; | |
15464 | } | |
15465 | } | |
15466 | ||
15467 | va_end (ap); | |
15468 | ||
15469 | if (have_retval) | |
15470 | switch (argc) | |
15471 | { | |
15472 | case 1: | |
15473 | pat = GEN_FCN (icode) (target, op[0]); | |
15474 | break; | |
15475 | ||
15476 | case 2: | |
15477 | pat = GEN_FCN (icode) (target, op[0], op[1]); | |
15478 | break; | |
15479 | ||
15480 | case 3: | |
15481 | pat = GEN_FCN (icode) (target, op[0], op[1], op[2]); | |
15482 | break; | |
15483 | ||
15484 | case 4: | |
15485 | pat = GEN_FCN (icode) (target, op[0], op[1], op[2], op[3]); | |
15486 | break; | |
15487 | ||
15488 | case 5: | |
15489 | pat = GEN_FCN (icode) (target, op[0], op[1], op[2], op[3], op[4]); | |
15490 | break; | |
15491 | ||
15492 | default: | |
15493 | gcc_unreachable (); | |
15494 | } | |
15495 | else | |
15496 | switch (argc) | |
15497 | { | |
15498 | case 1: | |
15499 | pat = GEN_FCN (icode) (op[0]); | |
15500 | break; | |
15501 | ||
15502 | case 2: | |
15503 | pat = GEN_FCN (icode) (op[0], op[1]); | |
15504 | break; | |
15505 | ||
15506 | case 3: | |
15507 | pat = GEN_FCN (icode) (op[0], op[1], op[2]); | |
15508 | break; | |
15509 | ||
15510 | case 4: | |
15511 | pat = GEN_FCN (icode) (op[0], op[1], op[2], op[3]); | |
15512 | break; | |
15513 | ||
15514 | case 5: | |
15515 | pat = GEN_FCN (icode) (op[0], op[1], op[2], op[3], op[4]); | |
15516 | break; | |
15517 | ||
15518 | default: | |
15519 | gcc_unreachable (); | |
15520 | } | |
15521 | ||
15522 | if (!pat) | |
15523 | return 0; | |
15524 | ||
15525 | emit_insn (pat); | |
15526 | ||
15527 | return target; | |
15528 | } | |
15529 | ||
15530 | /* Expand a Neon builtin. These are "special" because they don't have symbolic | |
15531 | constants defined per-instruction or per instruction-variant. Instead, the | |
15532 | required info is looked up in the table neon_builtin_data. */ | |
15533 | static rtx | |
15534 | arm_expand_neon_builtin (int fcode, tree exp, rtx target) | |
15535 | { | |
15536 | neon_itype itype; | |
15537 | enum insn_code icode = locate_neon_builtin_icode (fcode, &itype); | |
15538 | ||
15539 | switch (itype) | |
15540 | { | |
15541 | case NEON_UNOP: | |
15542 | case NEON_CONVERT: | |
15543 | case NEON_DUPLANE: | |
15544 | return arm_expand_neon_args (target, icode, 1, exp, | |
15545 | NEON_ARG_COPY_TO_REG, NEON_ARG_CONSTANT, NEON_ARG_STOP); | |
15546 | ||
15547 | case NEON_BINOP: | |
15548 | case NEON_SETLANE: | |
15549 | case NEON_SCALARMUL: | |
15550 | case NEON_SCALARMULL: | |
15551 | case NEON_SCALARMULH: | |
15552 | case NEON_SHIFTINSERT: | |
15553 | case NEON_LOGICBINOP: | |
15554 | return arm_expand_neon_args (target, icode, 1, exp, | |
15555 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_CONSTANT, | |
15556 | NEON_ARG_STOP); | |
15557 | ||
15558 | case NEON_TERNOP: | |
15559 | return arm_expand_neon_args (target, icode, 1, exp, | |
15560 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, | |
15561 | NEON_ARG_CONSTANT, NEON_ARG_STOP); | |
15562 | ||
15563 | case NEON_GETLANE: | |
15564 | case NEON_FIXCONV: | |
15565 | case NEON_SHIFTIMM: | |
15566 | return arm_expand_neon_args (target, icode, 1, exp, | |
15567 | NEON_ARG_COPY_TO_REG, NEON_ARG_CONSTANT, NEON_ARG_CONSTANT, | |
15568 | NEON_ARG_STOP); | |
15569 | ||
15570 | case NEON_CREATE: | |
15571 | return arm_expand_neon_args (target, icode, 1, exp, | |
15572 | NEON_ARG_COPY_TO_REG, NEON_ARG_STOP); | |
15573 | ||
15574 | case NEON_DUP: | |
15575 | case NEON_SPLIT: | |
15576 | case NEON_REINTERP: | |
15577 | return arm_expand_neon_args (target, icode, 1, exp, | |
15578 | NEON_ARG_COPY_TO_REG, NEON_ARG_STOP); | |
15579 | ||
15580 | case NEON_COMBINE: | |
15581 | case NEON_VTBL: | |
15582 | return arm_expand_neon_args (target, icode, 1, exp, | |
15583 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_STOP); | |
15584 | ||
15585 | case NEON_RESULTPAIR: | |
15586 | return arm_expand_neon_args (target, icode, 0, exp, | |
15587 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, | |
15588 | NEON_ARG_STOP); | |
15589 | ||
15590 | case NEON_LANEMUL: | |
15591 | case NEON_LANEMULL: | |
15592 | case NEON_LANEMULH: | |
15593 | return arm_expand_neon_args (target, icode, 1, exp, | |
15594 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_CONSTANT, | |
15595 | NEON_ARG_CONSTANT, NEON_ARG_STOP); | |
15596 | ||
15597 | case NEON_LANEMAC: | |
15598 | return arm_expand_neon_args (target, icode, 1, exp, | |
15599 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, | |
15600 | NEON_ARG_CONSTANT, NEON_ARG_CONSTANT, NEON_ARG_STOP); | |
15601 | ||
15602 | case NEON_SHIFTACC: | |
15603 | return arm_expand_neon_args (target, icode, 1, exp, | |
15604 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_CONSTANT, | |
15605 | NEON_ARG_CONSTANT, NEON_ARG_STOP); | |
15606 | ||
15607 | case NEON_SCALARMAC: | |
15608 | return arm_expand_neon_args (target, icode, 1, exp, | |
15609 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, | |
15610 | NEON_ARG_CONSTANT, NEON_ARG_STOP); | |
15611 | ||
15612 | case NEON_SELECT: | |
15613 | case NEON_VTBX: | |
15614 | return arm_expand_neon_args (target, icode, 1, exp, | |
15615 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, | |
15616 | NEON_ARG_STOP); | |
15617 | ||
15618 | case NEON_LOAD1: | |
15619 | case NEON_LOADSTRUCT: | |
15620 | return arm_expand_neon_args (target, icode, 1, exp, | |
15621 | NEON_ARG_COPY_TO_REG, NEON_ARG_STOP); | |
15622 | ||
15623 | case NEON_LOAD1LANE: | |
15624 | case NEON_LOADSTRUCTLANE: | |
15625 | return arm_expand_neon_args (target, icode, 1, exp, | |
15626 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_CONSTANT, | |
15627 | NEON_ARG_STOP); | |
15628 | ||
15629 | case NEON_STORE1: | |
15630 | case NEON_STORESTRUCT: | |
15631 | return arm_expand_neon_args (target, icode, 0, exp, | |
15632 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_STOP); | |
15633 | ||
15634 | case NEON_STORE1LANE: | |
15635 | case NEON_STORESTRUCTLANE: | |
15636 | return arm_expand_neon_args (target, icode, 0, exp, | |
15637 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_CONSTANT, | |
15638 | NEON_ARG_STOP); | |
15639 | } | |
15640 | ||
15641 | gcc_unreachable (); | |
15642 | } | |
15643 | ||
15644 | /* Emit code to reinterpret one Neon type as another, without altering bits. */ | |
15645 | void | |
15646 | neon_reinterpret (rtx dest, rtx src) | |
15647 | { | |
15648 | emit_move_insn (dest, gen_lowpart (GET_MODE (dest), src)); | |
15649 | } | |
15650 | ||
15651 | /* Emit code to place a Neon pair result in memory locations (with equal | |
15652 | registers). */ | |
15653 | void | |
15654 | neon_emit_pair_result_insn (enum machine_mode mode, | |
15655 | rtx (*intfn) (rtx, rtx, rtx, rtx), rtx destaddr, | |
15656 | rtx op1, rtx op2) | |
15657 | { | |
15658 | rtx mem = gen_rtx_MEM (mode, destaddr); | |
15659 | rtx tmp1 = gen_reg_rtx (mode); | |
15660 | rtx tmp2 = gen_reg_rtx (mode); | |
15661 | ||
15662 | emit_insn (intfn (tmp1, op1, tmp2, op2)); | |
15663 | ||
15664 | emit_move_insn (mem, tmp1); | |
15665 | mem = adjust_address (mem, mode, GET_MODE_SIZE (mode)); | |
15666 | emit_move_insn (mem, tmp2); | |
15667 | } | |
15668 | ||
15669 | /* Set up operands for a register copy from src to dest, taking care not to | |
15670 | clobber registers in the process. | |
15671 | FIXME: This has rather high polynomial complexity (O(n^3)?) but shouldn't | |
15672 | be called with a large N, so that should be OK. */ | |
15673 | ||
15674 | void | |
15675 | neon_disambiguate_copy (rtx *operands, rtx *dest, rtx *src, unsigned int count) | |
15676 | { | |
15677 | unsigned int copied = 0, opctr = 0; | |
15678 | unsigned int done = (1 << count) - 1; | |
15679 | unsigned int i, j; | |
15680 | ||
15681 | while (copied != done) | |
15682 | { | |
15683 | for (i = 0; i < count; i++) | |
15684 | { | |
15685 | int good = 1; | |
15686 | ||
15687 | for (j = 0; good && j < count; j++) | |
15688 | if (i != j && (copied & (1 << j)) == 0 | |
15689 | && reg_overlap_mentioned_p (src[j], dest[i])) | |
15690 | good = 0; | |
15691 | ||
15692 | if (good) | |
15693 | { | |
15694 | operands[opctr++] = dest[i]; | |
15695 | operands[opctr++] = src[i]; | |
15696 | copied |= 1 << i; | |
15697 | } | |
15698 | } | |
15699 | } | |
15700 | ||
15701 | gcc_assert (opctr == count * 2); | |
15702 | } | |
15703 | ||
755eb2b4 | 15704 | /* Expand an expression EXP that calls a built-in function, |
15705 | with result going to TARGET if that's convenient | |
15706 | (and in mode MODE if that's convenient). | |
15707 | SUBTARGET may be used as the target for computing one of EXP's operands. | |
15708 | IGNORE is nonzero if the value is to be ignored. */ | |
15709 | ||
15710 | static rtx | |
15711 | arm_expand_builtin (tree exp, | |
15712 | rtx target, | |
15713 | rtx subtarget ATTRIBUTE_UNUSED, | |
15714 | enum machine_mode mode ATTRIBUTE_UNUSED, | |
15715 | int ignore ATTRIBUTE_UNUSED) | |
15716 | { | |
15717 | const struct builtin_description * d; | |
15718 | enum insn_code icode; | |
c2f47e15 | 15719 | tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); |
755eb2b4 | 15720 | tree arg0; |
15721 | tree arg1; | |
15722 | tree arg2; | |
15723 | rtx op0; | |
15724 | rtx op1; | |
15725 | rtx op2; | |
15726 | rtx pat; | |
15727 | int fcode = DECL_FUNCTION_CODE (fndecl); | |
15728 | size_t i; | |
15729 | enum machine_mode tmode; | |
15730 | enum machine_mode mode0; | |
15731 | enum machine_mode mode1; | |
15732 | enum machine_mode mode2; | |
15733 | ||
d98a3884 | 15734 | if (fcode >= ARM_BUILTIN_NEON_BASE) |
15735 | return arm_expand_neon_builtin (fcode, exp, target); | |
15736 | ||
755eb2b4 | 15737 | switch (fcode) |
15738 | { | |
15739 | case ARM_BUILTIN_TEXTRMSB: | |
15740 | case ARM_BUILTIN_TEXTRMUB: | |
15741 | case ARM_BUILTIN_TEXTRMSH: | |
15742 | case ARM_BUILTIN_TEXTRMUH: | |
15743 | case ARM_BUILTIN_TEXTRMSW: | |
15744 | case ARM_BUILTIN_TEXTRMUW: | |
15745 | icode = (fcode == ARM_BUILTIN_TEXTRMSB ? CODE_FOR_iwmmxt_textrmsb | |
15746 | : fcode == ARM_BUILTIN_TEXTRMUB ? CODE_FOR_iwmmxt_textrmub | |
15747 | : fcode == ARM_BUILTIN_TEXTRMSH ? CODE_FOR_iwmmxt_textrmsh | |
15748 | : fcode == ARM_BUILTIN_TEXTRMUH ? CODE_FOR_iwmmxt_textrmuh | |
15749 | : CODE_FOR_iwmmxt_textrmw); | |
15750 | ||
c2f47e15 | 15751 | arg0 = CALL_EXPR_ARG (exp, 0); |
15752 | arg1 = CALL_EXPR_ARG (exp, 1); | |
8ec3c5c2 | 15753 | op0 = expand_normal (arg0); |
15754 | op1 = expand_normal (arg1); | |
755eb2b4 | 15755 | tmode = insn_data[icode].operand[0].mode; |
15756 | mode0 = insn_data[icode].operand[1].mode; | |
15757 | mode1 = insn_data[icode].operand[2].mode; | |
15758 | ||
15759 | if (! (*insn_data[icode].operand[1].predicate) (op0, mode0)) | |
15760 | op0 = copy_to_mode_reg (mode0, op0); | |
15761 | if (! (*insn_data[icode].operand[2].predicate) (op1, mode1)) | |
15762 | { | |
15763 | /* @@@ better error message */ | |
15764 | error ("selector must be an immediate"); | |
15765 | return gen_reg_rtx (tmode); | |
15766 | } | |
15767 | if (target == 0 | |
15768 | || GET_MODE (target) != tmode | |
15769 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
15770 | target = gen_reg_rtx (tmode); | |
15771 | pat = GEN_FCN (icode) (target, op0, op1); | |
15772 | if (! pat) | |
15773 | return 0; | |
15774 | emit_insn (pat); | |
15775 | return target; | |
15776 | ||
15777 | case ARM_BUILTIN_TINSRB: | |
15778 | case ARM_BUILTIN_TINSRH: | |
15779 | case ARM_BUILTIN_TINSRW: | |
15780 | icode = (fcode == ARM_BUILTIN_TINSRB ? CODE_FOR_iwmmxt_tinsrb | |
15781 | : fcode == ARM_BUILTIN_TINSRH ? CODE_FOR_iwmmxt_tinsrh | |
15782 | : CODE_FOR_iwmmxt_tinsrw); | |
c2f47e15 | 15783 | arg0 = CALL_EXPR_ARG (exp, 0); |
15784 | arg1 = CALL_EXPR_ARG (exp, 1); | |
15785 | arg2 = CALL_EXPR_ARG (exp, 2); | |
8ec3c5c2 | 15786 | op0 = expand_normal (arg0); |
15787 | op1 = expand_normal (arg1); | |
15788 | op2 = expand_normal (arg2); | |
755eb2b4 | 15789 | tmode = insn_data[icode].operand[0].mode; |
15790 | mode0 = insn_data[icode].operand[1].mode; | |
15791 | mode1 = insn_data[icode].operand[2].mode; | |
15792 | mode2 = insn_data[icode].operand[3].mode; | |
15793 | ||
15794 | if (! (*insn_data[icode].operand[1].predicate) (op0, mode0)) | |
15795 | op0 = copy_to_mode_reg (mode0, op0); | |
15796 | if (! (*insn_data[icode].operand[2].predicate) (op1, mode1)) | |
15797 | op1 = copy_to_mode_reg (mode1, op1); | |
15798 | if (! (*insn_data[icode].operand[3].predicate) (op2, mode2)) | |
15799 | { | |
15800 | /* @@@ better error message */ | |
15801 | error ("selector must be an immediate"); | |
15802 | return const0_rtx; | |
15803 | } | |
15804 | if (target == 0 | |
15805 | || GET_MODE (target) != tmode | |
15806 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
15807 | target = gen_reg_rtx (tmode); | |
15808 | pat = GEN_FCN (icode) (target, op0, op1, op2); | |
15809 | if (! pat) | |
15810 | return 0; | |
15811 | emit_insn (pat); | |
15812 | return target; | |
15813 | ||
15814 | case ARM_BUILTIN_SETWCX: | |
c2f47e15 | 15815 | arg0 = CALL_EXPR_ARG (exp, 0); |
15816 | arg1 = CALL_EXPR_ARG (exp, 1); | |
8ec3c5c2 | 15817 | op0 = force_reg (SImode, expand_normal (arg0)); |
15818 | op1 = expand_normal (arg1); | |
08476a09 | 15819 | emit_insn (gen_iwmmxt_tmcr (op1, op0)); |
755eb2b4 | 15820 | return 0; |
15821 | ||
15822 | case ARM_BUILTIN_GETWCX: | |
c2f47e15 | 15823 | arg0 = CALL_EXPR_ARG (exp, 0); |
8ec3c5c2 | 15824 | op0 = expand_normal (arg0); |
755eb2b4 | 15825 | target = gen_reg_rtx (SImode); |
15826 | emit_insn (gen_iwmmxt_tmrc (target, op0)); | |
15827 | return target; | |
15828 | ||
15829 | case ARM_BUILTIN_WSHUFH: | |
15830 | icode = CODE_FOR_iwmmxt_wshufh; | |
c2f47e15 | 15831 | arg0 = CALL_EXPR_ARG (exp, 0); |
15832 | arg1 = CALL_EXPR_ARG (exp, 1); | |
8ec3c5c2 | 15833 | op0 = expand_normal (arg0); |
15834 | op1 = expand_normal (arg1); | |
755eb2b4 | 15835 | tmode = insn_data[icode].operand[0].mode; |
15836 | mode1 = insn_data[icode].operand[1].mode; | |
15837 | mode2 = insn_data[icode].operand[2].mode; | |
15838 | ||
15839 | if (! (*insn_data[icode].operand[1].predicate) (op0, mode1)) | |
15840 | op0 = copy_to_mode_reg (mode1, op0); | |
15841 | if (! (*insn_data[icode].operand[2].predicate) (op1, mode2)) | |
15842 | { | |
15843 | /* @@@ better error message */ | |
15844 | error ("mask must be an immediate"); | |
15845 | return const0_rtx; | |
15846 | } | |
15847 | if (target == 0 | |
15848 | || GET_MODE (target) != tmode | |
15849 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
15850 | target = gen_reg_rtx (tmode); | |
15851 | pat = GEN_FCN (icode) (target, op0, op1); | |
15852 | if (! pat) | |
15853 | return 0; | |
15854 | emit_insn (pat); | |
15855 | return target; | |
15856 | ||
15857 | case ARM_BUILTIN_WSADB: | |
c2f47e15 | 15858 | return arm_expand_binop_builtin (CODE_FOR_iwmmxt_wsadb, exp, target); |
755eb2b4 | 15859 | case ARM_BUILTIN_WSADH: |
c2f47e15 | 15860 | return arm_expand_binop_builtin (CODE_FOR_iwmmxt_wsadh, exp, target); |
755eb2b4 | 15861 | case ARM_BUILTIN_WSADBZ: |
c2f47e15 | 15862 | return arm_expand_binop_builtin (CODE_FOR_iwmmxt_wsadbz, exp, target); |
755eb2b4 | 15863 | case ARM_BUILTIN_WSADHZ: |
c2f47e15 | 15864 | return arm_expand_binop_builtin (CODE_FOR_iwmmxt_wsadhz, exp, target); |
755eb2b4 | 15865 | |
15866 | /* Several three-argument builtins. */ | |
15867 | case ARM_BUILTIN_WMACS: | |
15868 | case ARM_BUILTIN_WMACU: | |
15869 | case ARM_BUILTIN_WALIGN: | |
15870 | case ARM_BUILTIN_TMIA: | |
15871 | case ARM_BUILTIN_TMIAPH: | |
15872 | case ARM_BUILTIN_TMIATT: | |
15873 | case ARM_BUILTIN_TMIATB: | |
15874 | case ARM_BUILTIN_TMIABT: | |
15875 | case ARM_BUILTIN_TMIABB: | |
15876 | icode = (fcode == ARM_BUILTIN_WMACS ? CODE_FOR_iwmmxt_wmacs | |
15877 | : fcode == ARM_BUILTIN_WMACU ? CODE_FOR_iwmmxt_wmacu | |
15878 | : fcode == ARM_BUILTIN_TMIA ? CODE_FOR_iwmmxt_tmia | |
15879 | : fcode == ARM_BUILTIN_TMIAPH ? CODE_FOR_iwmmxt_tmiaph | |
15880 | : fcode == ARM_BUILTIN_TMIABB ? CODE_FOR_iwmmxt_tmiabb | |
15881 | : fcode == ARM_BUILTIN_TMIABT ? CODE_FOR_iwmmxt_tmiabt | |
15882 | : fcode == ARM_BUILTIN_TMIATB ? CODE_FOR_iwmmxt_tmiatb | |
15883 | : fcode == ARM_BUILTIN_TMIATT ? CODE_FOR_iwmmxt_tmiatt | |
15884 | : CODE_FOR_iwmmxt_walign); | |
c2f47e15 | 15885 | arg0 = CALL_EXPR_ARG (exp, 0); |
15886 | arg1 = CALL_EXPR_ARG (exp, 1); | |
15887 | arg2 = CALL_EXPR_ARG (exp, 2); | |
8ec3c5c2 | 15888 | op0 = expand_normal (arg0); |
15889 | op1 = expand_normal (arg1); | |
15890 | op2 = expand_normal (arg2); | |
755eb2b4 | 15891 | tmode = insn_data[icode].operand[0].mode; |
15892 | mode0 = insn_data[icode].operand[1].mode; | |
15893 | mode1 = insn_data[icode].operand[2].mode; | |
15894 | mode2 = insn_data[icode].operand[3].mode; | |
15895 | ||
15896 | if (! (*insn_data[icode].operand[1].predicate) (op0, mode0)) | |
15897 | op0 = copy_to_mode_reg (mode0, op0); | |
15898 | if (! (*insn_data[icode].operand[2].predicate) (op1, mode1)) | |
15899 | op1 = copy_to_mode_reg (mode1, op1); | |
15900 | if (! (*insn_data[icode].operand[3].predicate) (op2, mode2)) | |
15901 | op2 = copy_to_mode_reg (mode2, op2); | |
15902 | if (target == 0 | |
15903 | || GET_MODE (target) != tmode | |
15904 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
15905 | target = gen_reg_rtx (tmode); | |
15906 | pat = GEN_FCN (icode) (target, op0, op1, op2); | |
15907 | if (! pat) | |
15908 | return 0; | |
15909 | emit_insn (pat); | |
15910 | return target; | |
9e7454d0 | 15911 | |
755eb2b4 | 15912 | case ARM_BUILTIN_WZERO: |
15913 | target = gen_reg_rtx (DImode); | |
15914 | emit_insn (gen_iwmmxt_clrdi (target)); | |
15915 | return target; | |
15916 | ||
f655717d | 15917 | case ARM_BUILTIN_THREAD_POINTER: |
15918 | return arm_load_tp (target); | |
15919 | ||
755eb2b4 | 15920 | default: |
15921 | break; | |
15922 | } | |
15923 | ||
3b9d9482 | 15924 | for (i = 0, d = bdesc_2arg; i < ARRAY_SIZE (bdesc_2arg); i++, d++) |
755eb2b4 | 15925 | if (d->code == (const enum arm_builtins) fcode) |
c2f47e15 | 15926 | return arm_expand_binop_builtin (d->icode, exp, target); |
755eb2b4 | 15927 | |
3b9d9482 | 15928 | for (i = 0, d = bdesc_1arg; i < ARRAY_SIZE (bdesc_1arg); i++, d++) |
755eb2b4 | 15929 | if (d->code == (const enum arm_builtins) fcode) |
c2f47e15 | 15930 | return arm_expand_unop_builtin (d->icode, exp, target, 0); |
755eb2b4 | 15931 | |
15932 | /* @@@ Should really do something sensible here. */ | |
15933 | return NULL_RTX; | |
15934 | } | |
15935 | \f | |
a12be13c | 15936 | /* Return the number (counting from 0) of |
15937 | the least significant set bit in MASK. */ | |
15938 | ||
ebd88a36 | 15939 | inline static int |
53f179c0 | 15940 | number_of_first_bit_set (unsigned mask) |
cffb2a26 | 15941 | { |
15942 | int bit; | |
15943 | ||
15944 | for (bit = 0; | |
15945 | (mask & (1 << bit)) == 0; | |
215b30b3 | 15946 | ++bit) |
cffb2a26 | 15947 | continue; |
15948 | ||
15949 | return bit; | |
15950 | } | |
15951 | ||
53f179c0 | 15952 | /* Emit code to push or pop registers to or from the stack. F is the |
15953 | assembly file. MASK is the registers to push or pop. PUSH is | |
15954 | nonzero if we should push, and zero if we should pop. For debugging | |
15955 | output, if pushing, adjust CFA_OFFSET by the amount of space added | |
15956 | to the stack. REAL_REGS should have the same number of bits set as | |
15957 | MASK, and will be used instead (in the same order) to describe which | |
15958 | registers were saved - this is used to mark the save slots when we | |
15959 | push high registers after moving them to low registers. */ | |
15960 | static void | |
15961 | thumb_pushpop (FILE *f, unsigned long mask, int push, int *cfa_offset, | |
15962 | unsigned long real_regs) | |
15963 | { | |
15964 | int regno; | |
15965 | int lo_mask = mask & 0xFF; | |
15966 | int pushed_words = 0; | |
15967 | ||
ed29c566 | 15968 | gcc_assert (mask); |
53f179c0 | 15969 | |
15970 | if (lo_mask == 0 && !push && (mask & (1 << PC_REGNUM))) | |
15971 | { | |
15972 | /* Special case. Do not generate a POP PC statement here, do it in | |
15973 | thumb_exit() */ | |
15974 | thumb_exit (f, -1); | |
15975 | return; | |
15976 | } | |
15977 | ||
1774763d | 15978 | if (ARM_EABI_UNWIND_TABLES && push) |
15979 | { | |
15980 | fprintf (f, "\t.save\t{"); | |
15981 | for (regno = 0; regno < 15; regno++) | |
15982 | { | |
15983 | if (real_regs & (1 << regno)) | |
15984 | { | |
15985 | if (real_regs & ((1 << regno) -1)) | |
15986 | fprintf (f, ", "); | |
15987 | asm_fprintf (f, "%r", regno); | |
15988 | } | |
15989 | } | |
15990 | fprintf (f, "}\n"); | |
15991 | } | |
15992 | ||
53f179c0 | 15993 | fprintf (f, "\t%s\t{", push ? "push" : "pop"); |
15994 | ||
15995 | /* Look at the low registers first. */ | |
15996 | for (regno = 0; regno <= LAST_LO_REGNUM; regno++, lo_mask >>= 1) | |
15997 | { | |
15998 | if (lo_mask & 1) | |
15999 | { | |
16000 | asm_fprintf (f, "%r", regno); | |
16001 | ||
16002 | if ((lo_mask & ~1) != 0) | |
16003 | fprintf (f, ", "); | |
16004 | ||
16005 | pushed_words++; | |
16006 | } | |
16007 | } | |
16008 | ||
16009 | if (push && (mask & (1 << LR_REGNUM))) | |
16010 | { | |
16011 | /* Catch pushing the LR. */ | |
16012 | if (mask & 0xFF) | |
16013 | fprintf (f, ", "); | |
16014 | ||
16015 | asm_fprintf (f, "%r", LR_REGNUM); | |
16016 | ||
16017 | pushed_words++; | |
16018 | } | |
16019 | else if (!push && (mask & (1 << PC_REGNUM))) | |
16020 | { | |
16021 | /* Catch popping the PC. */ | |
16022 | if (TARGET_INTERWORK || TARGET_BACKTRACE | |
16023 | || current_function_calls_eh_return) | |
16024 | { | |
16025 | /* The PC is never poped directly, instead | |
16026 | it is popped into r3 and then BX is used. */ | |
16027 | fprintf (f, "}\n"); | |
16028 | ||
16029 | thumb_exit (f, -1); | |
16030 | ||
16031 | return; | |
16032 | } | |
16033 | else | |
16034 | { | |
16035 | if (mask & 0xFF) | |
16036 | fprintf (f, ", "); | |
16037 | ||
16038 | asm_fprintf (f, "%r", PC_REGNUM); | |
16039 | } | |
16040 | } | |
16041 | ||
16042 | fprintf (f, "}\n"); | |
16043 | ||
16044 | if (push && pushed_words && dwarf2out_do_frame ()) | |
16045 | { | |
16046 | char *l = dwarf2out_cfi_label (); | |
16047 | int pushed_mask = real_regs; | |
16048 | ||
16049 | *cfa_offset += pushed_words * 4; | |
16050 | dwarf2out_def_cfa (l, SP_REGNUM, *cfa_offset); | |
16051 | ||
16052 | pushed_words = 0; | |
16053 | pushed_mask = real_regs; | |
16054 | for (regno = 0; regno <= 14; regno++, pushed_mask >>= 1) | |
16055 | { | |
16056 | if (pushed_mask & 1) | |
16057 | dwarf2out_reg_save (l, regno, 4 * pushed_words++ - *cfa_offset); | |
16058 | } | |
16059 | } | |
16060 | } | |
16061 | ||
cffb2a26 | 16062 | /* Generate code to return from a thumb function. |
16063 | If 'reg_containing_return_addr' is -1, then the return address is | |
16064 | actually on the stack, at the stack pointer. */ | |
16065 | static void | |
4c44712e | 16066 | thumb_exit (FILE *f, int reg_containing_return_addr) |
cffb2a26 | 16067 | { |
16068 | unsigned regs_available_for_popping; | |
16069 | unsigned regs_to_pop; | |
16070 | int pops_needed; | |
16071 | unsigned available; | |
16072 | unsigned required; | |
16073 | int mode; | |
16074 | int size; | |
16075 | int restore_a4 = FALSE; | |
16076 | ||
16077 | /* Compute the registers we need to pop. */ | |
16078 | regs_to_pop = 0; | |
16079 | pops_needed = 0; | |
16080 | ||
4c44712e | 16081 | if (reg_containing_return_addr == -1) |
cffb2a26 | 16082 | { |
cffb2a26 | 16083 | regs_to_pop |= 1 << LR_REGNUM; |
215b30b3 | 16084 | ++pops_needed; |
cffb2a26 | 16085 | } |
16086 | ||
16087 | if (TARGET_BACKTRACE) | |
16088 | { | |
16089 | /* Restore the (ARM) frame pointer and stack pointer. */ | |
16090 | regs_to_pop |= (1 << ARM_HARD_FRAME_POINTER_REGNUM) | (1 << SP_REGNUM); | |
16091 | pops_needed += 2; | |
16092 | } | |
16093 | ||
16094 | /* If there is nothing to pop then just emit the BX instruction and | |
16095 | return. */ | |
16096 | if (pops_needed == 0) | |
16097 | { | |
4c44712e | 16098 | if (current_function_calls_eh_return) |
16099 | asm_fprintf (f, "\tadd\t%r, %r\n", SP_REGNUM, ARM_EH_STACKADJ_REGNUM); | |
cffb2a26 | 16100 | |
16101 | asm_fprintf (f, "\tbx\t%r\n", reg_containing_return_addr); | |
16102 | return; | |
16103 | } | |
16104 | /* Otherwise if we are not supporting interworking and we have not created | |
16105 | a backtrace structure and the function was not entered in ARM mode then | |
16106 | just pop the return address straight into the PC. */ | |
215b30b3 | 16107 | else if (!TARGET_INTERWORK |
16108 | && !TARGET_BACKTRACE | |
4c44712e | 16109 | && !is_called_in_ARM_mode (current_function_decl) |
16110 | && !current_function_calls_eh_return) | |
cffb2a26 | 16111 | { |
4c44712e | 16112 | asm_fprintf (f, "\tpop\t{%r}\n", PC_REGNUM); |
cffb2a26 | 16113 | return; |
16114 | } | |
16115 | ||
16116 | /* Find out how many of the (return) argument registers we can corrupt. */ | |
16117 | regs_available_for_popping = 0; | |
16118 | ||
16119 | /* If returning via __builtin_eh_return, the bottom three registers | |
16120 | all contain information needed for the return. */ | |
4c44712e | 16121 | if (current_function_calls_eh_return) |
cffb2a26 | 16122 | size = 12; |
16123 | else | |
16124 | { | |
cffb2a26 | 16125 | /* If we can deduce the registers used from the function's |
16126 | return value. This is more reliable that examining | |
3072d30e | 16127 | df_regs_ever_live_p () because that will be set if the register is |
cffb2a26 | 16128 | ever used in the function, not just if the register is used |
16129 | to hold a return value. */ | |
16130 | ||
16131 | if (current_function_return_rtx != 0) | |
16132 | mode = GET_MODE (current_function_return_rtx); | |
16133 | else | |
cffb2a26 | 16134 | mode = DECL_MODE (DECL_RESULT (current_function_decl)); |
16135 | ||
16136 | size = GET_MODE_SIZE (mode); | |
16137 | ||
16138 | if (size == 0) | |
16139 | { | |
16140 | /* In a void function we can use any argument register. | |
16141 | In a function that returns a structure on the stack | |
16142 | we can use the second and third argument registers. */ | |
16143 | if (mode == VOIDmode) | |
16144 | regs_available_for_popping = | |
16145 | (1 << ARG_REGISTER (1)) | |
16146 | | (1 << ARG_REGISTER (2)) | |
16147 | | (1 << ARG_REGISTER (3)); | |
16148 | else | |
16149 | regs_available_for_popping = | |
16150 | (1 << ARG_REGISTER (2)) | |
16151 | | (1 << ARG_REGISTER (3)); | |
16152 | } | |
16153 | else if (size <= 4) | |
16154 | regs_available_for_popping = | |
16155 | (1 << ARG_REGISTER (2)) | |
16156 | | (1 << ARG_REGISTER (3)); | |
16157 | else if (size <= 8) | |
16158 | regs_available_for_popping = | |
16159 | (1 << ARG_REGISTER (3)); | |
16160 | } | |
16161 | ||
16162 | /* Match registers to be popped with registers into which we pop them. */ | |
16163 | for (available = regs_available_for_popping, | |
16164 | required = regs_to_pop; | |
16165 | required != 0 && available != 0; | |
16166 | available &= ~(available & - available), | |
16167 | required &= ~(required & - required)) | |
16168 | -- pops_needed; | |
16169 | ||
16170 | /* If we have any popping registers left over, remove them. */ | |
16171 | if (available > 0) | |
215b30b3 | 16172 | regs_available_for_popping &= ~available; |
9e7454d0 | 16173 | |
cffb2a26 | 16174 | /* Otherwise if we need another popping register we can use |
16175 | the fourth argument register. */ | |
16176 | else if (pops_needed) | |
16177 | { | |
16178 | /* If we have not found any free argument registers and | |
16179 | reg a4 contains the return address, we must move it. */ | |
16180 | if (regs_available_for_popping == 0 | |
16181 | && reg_containing_return_addr == LAST_ARG_REGNUM) | |
16182 | { | |
16183 | asm_fprintf (f, "\tmov\t%r, %r\n", LR_REGNUM, LAST_ARG_REGNUM); | |
16184 | reg_containing_return_addr = LR_REGNUM; | |
16185 | } | |
16186 | else if (size > 12) | |
16187 | { | |
16188 | /* Register a4 is being used to hold part of the return value, | |
16189 | but we have dire need of a free, low register. */ | |
16190 | restore_a4 = TRUE; | |
9e7454d0 | 16191 | |
cffb2a26 | 16192 | asm_fprintf (f, "\tmov\t%r, %r\n",IP_REGNUM, LAST_ARG_REGNUM); |
16193 | } | |
9e7454d0 | 16194 | |
cffb2a26 | 16195 | if (reg_containing_return_addr != LAST_ARG_REGNUM) |
16196 | { | |
16197 | /* The fourth argument register is available. */ | |
16198 | regs_available_for_popping |= 1 << LAST_ARG_REGNUM; | |
9e7454d0 | 16199 | |
215b30b3 | 16200 | --pops_needed; |
cffb2a26 | 16201 | } |
16202 | } | |
16203 | ||
16204 | /* Pop as many registers as we can. */ | |
37826745 | 16205 | thumb_pushpop (f, regs_available_for_popping, FALSE, NULL, |
16206 | regs_available_for_popping); | |
cffb2a26 | 16207 | |
16208 | /* Process the registers we popped. */ | |
16209 | if (reg_containing_return_addr == -1) | |
16210 | { | |
16211 | /* The return address was popped into the lowest numbered register. */ | |
215b30b3 | 16212 | regs_to_pop &= ~(1 << LR_REGNUM); |
9e7454d0 | 16213 | |
cffb2a26 | 16214 | reg_containing_return_addr = |
16215 | number_of_first_bit_set (regs_available_for_popping); | |
16216 | ||
16217 | /* Remove this register for the mask of available registers, so that | |
b89db9a7 | 16218 | the return address will not be corrupted by further pops. */ |
215b30b3 | 16219 | regs_available_for_popping &= ~(1 << reg_containing_return_addr); |
cffb2a26 | 16220 | } |
16221 | ||
16222 | /* If we popped other registers then handle them here. */ | |
16223 | if (regs_available_for_popping) | |
16224 | { | |
16225 | int frame_pointer; | |
9e7454d0 | 16226 | |
cffb2a26 | 16227 | /* Work out which register currently contains the frame pointer. */ |
16228 | frame_pointer = number_of_first_bit_set (regs_available_for_popping); | |
16229 | ||
16230 | /* Move it into the correct place. */ | |
16231 | asm_fprintf (f, "\tmov\t%r, %r\n", | |
16232 | ARM_HARD_FRAME_POINTER_REGNUM, frame_pointer); | |
16233 | ||
16234 | /* (Temporarily) remove it from the mask of popped registers. */ | |
215b30b3 | 16235 | regs_available_for_popping &= ~(1 << frame_pointer); |
16236 | regs_to_pop &= ~(1 << ARM_HARD_FRAME_POINTER_REGNUM); | |
9e7454d0 | 16237 | |
cffb2a26 | 16238 | if (regs_available_for_popping) |
16239 | { | |
16240 | int stack_pointer; | |
9e7454d0 | 16241 | |
cffb2a26 | 16242 | /* We popped the stack pointer as well, |
16243 | find the register that contains it. */ | |
16244 | stack_pointer = number_of_first_bit_set (regs_available_for_popping); | |
16245 | ||
16246 | /* Move it into the stack register. */ | |
16247 | asm_fprintf (f, "\tmov\t%r, %r\n", SP_REGNUM, stack_pointer); | |
9e7454d0 | 16248 | |
cffb2a26 | 16249 | /* At this point we have popped all necessary registers, so |
16250 | do not worry about restoring regs_available_for_popping | |
16251 | to its correct value: | |
16252 | ||
16253 | assert (pops_needed == 0) | |
16254 | assert (regs_available_for_popping == (1 << frame_pointer)) | |
16255 | assert (regs_to_pop == (1 << STACK_POINTER)) */ | |
16256 | } | |
16257 | else | |
16258 | { | |
16259 | /* Since we have just move the popped value into the frame | |
16260 | pointer, the popping register is available for reuse, and | |
16261 | we know that we still have the stack pointer left to pop. */ | |
16262 | regs_available_for_popping |= (1 << frame_pointer); | |
16263 | } | |
16264 | } | |
9e7454d0 | 16265 | |
cffb2a26 | 16266 | /* If we still have registers left on the stack, but we no longer have |
16267 | any registers into which we can pop them, then we must move the return | |
16268 | address into the link register and make available the register that | |
16269 | contained it. */ | |
16270 | if (regs_available_for_popping == 0 && pops_needed > 0) | |
16271 | { | |
16272 | regs_available_for_popping |= 1 << reg_containing_return_addr; | |
9e7454d0 | 16273 | |
cffb2a26 | 16274 | asm_fprintf (f, "\tmov\t%r, %r\n", LR_REGNUM, |
16275 | reg_containing_return_addr); | |
9e7454d0 | 16276 | |
cffb2a26 | 16277 | reg_containing_return_addr = LR_REGNUM; |
16278 | } | |
16279 | ||
16280 | /* If we have registers left on the stack then pop some more. | |
16281 | We know that at most we will want to pop FP and SP. */ | |
16282 | if (pops_needed > 0) | |
16283 | { | |
16284 | int popped_into; | |
16285 | int move_to; | |
9e7454d0 | 16286 | |
37826745 | 16287 | thumb_pushpop (f, regs_available_for_popping, FALSE, NULL, |
16288 | regs_available_for_popping); | |
cffb2a26 | 16289 | |
16290 | /* We have popped either FP or SP. | |
16291 | Move whichever one it is into the correct register. */ | |
16292 | popped_into = number_of_first_bit_set (regs_available_for_popping); | |
16293 | move_to = number_of_first_bit_set (regs_to_pop); | |
16294 | ||
16295 | asm_fprintf (f, "\tmov\t%r, %r\n", move_to, popped_into); | |
16296 | ||
215b30b3 | 16297 | regs_to_pop &= ~(1 << move_to); |
cffb2a26 | 16298 | |
215b30b3 | 16299 | --pops_needed; |
cffb2a26 | 16300 | } |
9e7454d0 | 16301 | |
cffb2a26 | 16302 | /* If we still have not popped everything then we must have only |
16303 | had one register available to us and we are now popping the SP. */ | |
16304 | if (pops_needed > 0) | |
16305 | { | |
16306 | int popped_into; | |
9e7454d0 | 16307 | |
37826745 | 16308 | thumb_pushpop (f, regs_available_for_popping, FALSE, NULL, |
16309 | regs_available_for_popping); | |
cffb2a26 | 16310 | |
16311 | popped_into = number_of_first_bit_set (regs_available_for_popping); | |
16312 | ||
16313 | asm_fprintf (f, "\tmov\t%r, %r\n", SP_REGNUM, popped_into); | |
16314 | /* | |
16315 | assert (regs_to_pop == (1 << STACK_POINTER)) | |
16316 | assert (pops_needed == 1) | |
16317 | */ | |
16318 | } | |
16319 | ||
16320 | /* If necessary restore the a4 register. */ | |
16321 | if (restore_a4) | |
16322 | { | |
16323 | if (reg_containing_return_addr != LR_REGNUM) | |
16324 | { | |
16325 | asm_fprintf (f, "\tmov\t%r, %r\n", LR_REGNUM, LAST_ARG_REGNUM); | |
16326 | reg_containing_return_addr = LR_REGNUM; | |
16327 | } | |
9e7454d0 | 16328 | |
cffb2a26 | 16329 | asm_fprintf (f, "\tmov\t%r, %r\n", LAST_ARG_REGNUM, IP_REGNUM); |
16330 | } | |
16331 | ||
4c44712e | 16332 | if (current_function_calls_eh_return) |
16333 | asm_fprintf (f, "\tadd\t%r, %r\n", SP_REGNUM, ARM_EH_STACKADJ_REGNUM); | |
cffb2a26 | 16334 | |
16335 | /* Return to caller. */ | |
16336 | asm_fprintf (f, "\tbx\t%r\n", reg_containing_return_addr); | |
16337 | } | |
16338 | ||
cffb2a26 | 16339 | \f |
16340 | void | |
25f905c2 | 16341 | thumb1_final_prescan_insn (rtx insn) |
cffb2a26 | 16342 | { |
cffb2a26 | 16343 | if (flag_print_asm_name) |
47fc0706 | 16344 | asm_fprintf (asm_out_file, "%@ 0x%04x\n", |
16345 | INSN_ADDRESSES (INSN_UID (insn))); | |
cffb2a26 | 16346 | } |
16347 | ||
16348 | int | |
ebd88a36 | 16349 | thumb_shiftable_const (unsigned HOST_WIDE_INT val) |
cffb2a26 | 16350 | { |
16351 | unsigned HOST_WIDE_INT mask = 0xff; | |
16352 | int i; | |
16353 | ||
16354 | if (val == 0) /* XXX */ | |
16355 | return 0; | |
9e7454d0 | 16356 | |
cffb2a26 | 16357 | for (i = 0; i < 25; i++) |
16358 | if ((val & (mask << i)) == val) | |
16359 | return 1; | |
16360 | ||
16361 | return 0; | |
16362 | } | |
16363 | ||
e3e08e7f | 16364 | /* Returns nonzero if the current function contains, |
cffb2a26 | 16365 | or might contain a far jump. */ |
f9273c43 | 16366 | static int |
16367 | thumb_far_jump_used_p (void) | |
cffb2a26 | 16368 | { |
16369 | rtx insn; | |
16370 | ||
16371 | /* This test is only important for leaf functions. */ | |
215b30b3 | 16372 | /* assert (!leaf_function_p ()); */ |
9e7454d0 | 16373 | |
cffb2a26 | 16374 | /* If we have already decided that far jumps may be used, |
16375 | do not bother checking again, and always return true even if | |
16376 | it turns out that they are not being used. Once we have made | |
16377 | the decision that far jumps are present (and that hence the link | |
16378 | register will be pushed onto the stack) we cannot go back on it. */ | |
16379 | if (cfun->machine->far_jump_used) | |
16380 | return 1; | |
16381 | ||
16382 | /* If this function is not being called from the prologue/epilogue | |
16383 | generation code then it must be being called from the | |
16384 | INITIAL_ELIMINATION_OFFSET macro. */ | |
f9273c43 | 16385 | if (!(ARM_DOUBLEWORD_ALIGN || reload_completed)) |
cffb2a26 | 16386 | { |
16387 | /* In this case we know that we are being asked about the elimination | |
16388 | of the arg pointer register. If that register is not being used, | |
16389 | then there are no arguments on the stack, and we do not have to | |
16390 | worry that a far jump might force the prologue to push the link | |
16391 | register, changing the stack offsets. In this case we can just | |
16392 | return false, since the presence of far jumps in the function will | |
16393 | not affect stack offsets. | |
16394 | ||
16395 | If the arg pointer is live (or if it was live, but has now been | |
16396 | eliminated and so set to dead) then we do have to test to see if | |
16397 | the function might contain a far jump. This test can lead to some | |
16398 | false negatives, since before reload is completed, then length of | |
16399 | branch instructions is not known, so gcc defaults to returning their | |
16400 | longest length, which in turn sets the far jump attribute to true. | |
16401 | ||
16402 | A false negative will not result in bad code being generated, but it | |
16403 | will result in a needless push and pop of the link register. We | |
f9273c43 | 16404 | hope that this does not occur too often. |
16405 | ||
16406 | If we need doubleword stack alignment this could affect the other | |
16407 | elimination offsets so we can't risk getting it wrong. */ | |
3072d30e | 16408 | if (df_regs_ever_live_p (ARG_POINTER_REGNUM)) |
cffb2a26 | 16409 | cfun->machine->arg_pointer_live = 1; |
215b30b3 | 16410 | else if (!cfun->machine->arg_pointer_live) |
cffb2a26 | 16411 | return 0; |
16412 | } | |
16413 | ||
16414 | /* Check to see if the function contains a branch | |
16415 | insn with the far jump attribute set. */ | |
16416 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
16417 | { | |
16418 | if (GET_CODE (insn) == JUMP_INSN | |
16419 | /* Ignore tablejump patterns. */ | |
16420 | && GET_CODE (PATTERN (insn)) != ADDR_VEC | |
16421 | && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC | |
16422 | && get_attr_far_jump (insn) == FAR_JUMP_YES | |
16423 | ) | |
16424 | { | |
8d232dc7 | 16425 | /* Record the fact that we have decided that |
cffb2a26 | 16426 | the function does use far jumps. */ |
16427 | cfun->machine->far_jump_used = 1; | |
16428 | return 1; | |
16429 | } | |
16430 | } | |
9e7454d0 | 16431 | |
cffb2a26 | 16432 | return 0; |
16433 | } | |
16434 | ||
e3e08e7f | 16435 | /* Return nonzero if FUNC must be entered in ARM mode. */ |
cffb2a26 | 16436 | int |
ebd88a36 | 16437 | is_called_in_ARM_mode (tree func) |
cffb2a26 | 16438 | { |
ed29c566 | 16439 | gcc_assert (TREE_CODE (func) == FUNCTION_DECL); |
cffb2a26 | 16440 | |
fccee353 | 16441 | /* Ignore the problem about functions whose address is taken. */ |
cffb2a26 | 16442 | if (TARGET_CALLEE_INTERWORKING && TREE_PUBLIC (func)) |
16443 | return TRUE; | |
16444 | ||
9e7454d0 | 16445 | #ifdef ARM_PE |
e3c541f0 | 16446 | return lookup_attribute ("interfacearm", DECL_ATTRIBUTES (func)) != NULL_TREE; |
cffb2a26 | 16447 | #else |
16448 | return FALSE; | |
16449 | #endif | |
16450 | } | |
16451 | ||
674a8f0b | 16452 | /* The bits which aren't usefully expanded as rtl. */ |
0bd59681 | 16453 | const char * |
ebd88a36 | 16454 | thumb_unexpanded_epilogue (void) |
cffb2a26 | 16455 | { |
16456 | int regno; | |
53f179c0 | 16457 | unsigned long live_regs_mask = 0; |
cffb2a26 | 16458 | int high_regs_pushed = 0; |
cffb2a26 | 16459 | int had_to_push_lr; |
237533cc | 16460 | int size; |
cffb2a26 | 16461 | |
16462 | if (return_used_this_function) | |
16463 | return ""; | |
16464 | ||
4d0f24bc | 16465 | if (IS_NAKED (arm_current_func_type ())) |
16466 | return ""; | |
16467 | ||
25f905c2 | 16468 | live_regs_mask = thumb1_compute_save_reg_mask (); |
237533cc | 16469 | high_regs_pushed = bit_count (live_regs_mask & 0x0f00); |
16470 | ||
16471 | /* If we can deduce the registers used from the function's return value. | |
3072d30e | 16472 | This is more reliable that examining df_regs_ever_live_p () because that |
237533cc | 16473 | will be set if the register is ever used in the function, not just if |
16474 | the register is used to hold a return value. */ | |
0858d94e | 16475 | size = arm_size_return_regs (); |
cffb2a26 | 16476 | |
16477 | /* The prolog may have pushed some high registers to use as | |
a361b456 | 16478 | work registers. e.g. the testsuite file: |
cffb2a26 | 16479 | gcc/testsuite/gcc/gcc.c-torture/execute/complex-2.c |
16480 | compiles to produce: | |
16481 | push {r4, r5, r6, r7, lr} | |
16482 | mov r7, r9 | |
16483 | mov r6, r8 | |
16484 | push {r6, r7} | |
16485 | as part of the prolog. We have to undo that pushing here. */ | |
9e7454d0 | 16486 | |
cffb2a26 | 16487 | if (high_regs_pushed) |
16488 | { | |
53f179c0 | 16489 | unsigned long mask = live_regs_mask & 0xff; |
cffb2a26 | 16490 | int next_hi_reg; |
cffb2a26 | 16491 | |
237533cc | 16492 | /* The available low registers depend on the size of the value we are |
16493 | returning. */ | |
16494 | if (size <= 12) | |
cffb2a26 | 16495 | mask |= 1 << 3; |
237533cc | 16496 | if (size <= 8) |
16497 | mask |= 1 << 2; | |
cffb2a26 | 16498 | |
16499 | if (mask == 0) | |
16500 | /* Oh dear! We have no low registers into which we can pop | |
16501 | high registers! */ | |
f060a027 | 16502 | internal_error |
16503 | ("no low registers available for popping high registers"); | |
9e7454d0 | 16504 | |
cffb2a26 | 16505 | for (next_hi_reg = 8; next_hi_reg < 13; next_hi_reg++) |
237533cc | 16506 | if (live_regs_mask & (1 << next_hi_reg)) |
cffb2a26 | 16507 | break; |
16508 | ||
16509 | while (high_regs_pushed) | |
16510 | { | |
16511 | /* Find lo register(s) into which the high register(s) can | |
16512 | be popped. */ | |
16513 | for (regno = 0; regno <= LAST_LO_REGNUM; regno++) | |
16514 | { | |
16515 | if (mask & (1 << regno)) | |
16516 | high_regs_pushed--; | |
16517 | if (high_regs_pushed == 0) | |
16518 | break; | |
16519 | } | |
16520 | ||
16521 | mask &= (2 << regno) - 1; /* A noop if regno == 8 */ | |
16522 | ||
674a8f0b | 16523 | /* Pop the values into the low register(s). */ |
37826745 | 16524 | thumb_pushpop (asm_out_file, mask, 0, NULL, mask); |
cffb2a26 | 16525 | |
16526 | /* Move the value(s) into the high registers. */ | |
16527 | for (regno = 0; regno <= LAST_LO_REGNUM; regno++) | |
16528 | { | |
16529 | if (mask & (1 << regno)) | |
16530 | { | |
16531 | asm_fprintf (asm_out_file, "\tmov\t%r, %r\n", next_hi_reg, | |
16532 | regno); | |
9e7454d0 | 16533 | |
cffb2a26 | 16534 | for (next_hi_reg++; next_hi_reg < 13; next_hi_reg++) |
237533cc | 16535 | if (live_regs_mask & (1 << next_hi_reg)) |
cffb2a26 | 16536 | break; |
16537 | } | |
16538 | } | |
16539 | } | |
237533cc | 16540 | live_regs_mask &= ~0x0f00; |
cffb2a26 | 16541 | } |
16542 | ||
237533cc | 16543 | had_to_push_lr = (live_regs_mask & (1 << LR_REGNUM)) != 0; |
16544 | live_regs_mask &= 0xff; | |
16545 | ||
cffb2a26 | 16546 | if (current_function_pretend_args_size == 0 || TARGET_BACKTRACE) |
16547 | { | |
9e7454d0 | 16548 | /* Pop the return address into the PC. */ |
237533cc | 16549 | if (had_to_push_lr) |
cffb2a26 | 16550 | live_regs_mask |= 1 << PC_REGNUM; |
16551 | ||
16552 | /* Either no argument registers were pushed or a backtrace | |
16553 | structure was created which includes an adjusted stack | |
16554 | pointer, so just pop everything. */ | |
16555 | if (live_regs_mask) | |
37826745 | 16556 | thumb_pushpop (asm_out_file, live_regs_mask, FALSE, NULL, |
16557 | live_regs_mask); | |
237533cc | 16558 | |
cffb2a26 | 16559 | /* We have either just popped the return address into the |
237533cc | 16560 | PC or it is was kept in LR for the entire function. */ |
16561 | if (!had_to_push_lr) | |
16562 | thumb_exit (asm_out_file, LR_REGNUM); | |
cffb2a26 | 16563 | } |
16564 | else | |
16565 | { | |
16566 | /* Pop everything but the return address. */ | |
cffb2a26 | 16567 | if (live_regs_mask) |
37826745 | 16568 | thumb_pushpop (asm_out_file, live_regs_mask, FALSE, NULL, |
16569 | live_regs_mask); | |
cffb2a26 | 16570 | |
16571 | if (had_to_push_lr) | |
237533cc | 16572 | { |
16573 | if (size > 12) | |
16574 | { | |
16575 | /* We have no free low regs, so save one. */ | |
16576 | asm_fprintf (asm_out_file, "\tmov\t%r, %r\n", IP_REGNUM, | |
16577 | LAST_ARG_REGNUM); | |
16578 | } | |
16579 | ||
16580 | /* Get the return address into a temporary register. */ | |
16581 | thumb_pushpop (asm_out_file, 1 << LAST_ARG_REGNUM, 0, NULL, | |
16582 | 1 << LAST_ARG_REGNUM); | |
16583 | ||
16584 | if (size > 12) | |
16585 | { | |
16586 | /* Move the return address to lr. */ | |
16587 | asm_fprintf (asm_out_file, "\tmov\t%r, %r\n", LR_REGNUM, | |
16588 | LAST_ARG_REGNUM); | |
16589 | /* Restore the low register. */ | |
16590 | asm_fprintf (asm_out_file, "\tmov\t%r, %r\n", LAST_ARG_REGNUM, | |
16591 | IP_REGNUM); | |
16592 | regno = LR_REGNUM; | |
16593 | } | |
16594 | else | |
16595 | regno = LAST_ARG_REGNUM; | |
16596 | } | |
16597 | else | |
16598 | regno = LR_REGNUM; | |
9e7454d0 | 16599 | |
cffb2a26 | 16600 | /* Remove the argument registers that were pushed onto the stack. */ |
16601 | asm_fprintf (asm_out_file, "\tadd\t%r, %r, #%d\n", | |
16602 | SP_REGNUM, SP_REGNUM, | |
16603 | current_function_pretend_args_size); | |
9e7454d0 | 16604 | |
237533cc | 16605 | thumb_exit (asm_out_file, regno); |
cffb2a26 | 16606 | } |
16607 | ||
16608 | return ""; | |
16609 | } | |
16610 | ||
16611 | /* Functions to save and restore machine-specific function data. */ | |
1f3233d1 | 16612 | static struct machine_function * |
ebd88a36 | 16613 | arm_init_machine_status (void) |
cffb2a26 | 16614 | { |
1f3233d1 | 16615 | struct machine_function *machine; |
16616 | machine = (machine_function *) ggc_alloc_cleared (sizeof (machine_function)); | |
e27ad2d5 | 16617 | |
9e7454d0 | 16618 | #if ARM_FT_UNKNOWN != 0 |
1f3233d1 | 16619 | machine->func_type = ARM_FT_UNKNOWN; |
e27ad2d5 | 16620 | #endif |
1f3233d1 | 16621 | return machine; |
2997ed4d | 16622 | } |
16623 | ||
cffb2a26 | 16624 | /* Return an RTX indicating where the return address to the |
16625 | calling function can be found. */ | |
16626 | rtx | |
ebd88a36 | 16627 | arm_return_addr (int count, rtx frame ATTRIBUTE_UNUSED) |
cffb2a26 | 16628 | { |
cffb2a26 | 16629 | if (count != 0) |
16630 | return NULL_RTX; | |
16631 | ||
c1a66faf | 16632 | return get_hard_reg_initial_val (Pmode, LR_REGNUM); |
cffb2a26 | 16633 | } |
16634 | ||
16635 | /* Do anything needed before RTL is emitted for each function. */ | |
16636 | void | |
ebd88a36 | 16637 | arm_init_expanders (void) |
cffb2a26 | 16638 | { |
16639 | /* Arrange to initialize and mark the machine per-function status. */ | |
16640 | init_machine_status = arm_init_machine_status; | |
0e2db0a2 | 16641 | |
16642 | /* This is to stop the combine pass optimizing away the alignment | |
16643 | adjustment of va_arg. */ | |
16644 | /* ??? It is claimed that this should not be necessary. */ | |
16645 | if (cfun) | |
16646 | mark_reg_pointer (arg_pointer_rtx, PARM_BOUNDARY); | |
cffb2a26 | 16647 | } |
16648 | ||
a7712927 | 16649 | |
f90b51f1 | 16650 | /* Like arm_compute_initial_elimination offset. Simpler because there |
16651 | isn't an ABI specified frame pointer for Thumb. Instead, we set it | |
16652 | to point at the base of the local variables after static stack | |
16653 | space for a function has been allocated. */ | |
a7712927 | 16654 | |
f9273c43 | 16655 | HOST_WIDE_INT |
16656 | thumb_compute_initial_elimination_offset (unsigned int from, unsigned int to) | |
16657 | { | |
16658 | arm_stack_offsets *offsets; | |
a7712927 | 16659 | |
f9273c43 | 16660 | offsets = arm_get_frame_offsets (); |
a7712927 | 16661 | |
f9273c43 | 16662 | switch (from) |
a7712927 | 16663 | { |
f9273c43 | 16664 | case ARG_POINTER_REGNUM: |
16665 | switch (to) | |
16666 | { | |
16667 | case STACK_POINTER_REGNUM: | |
16668 | return offsets->outgoing_args - offsets->saved_args; | |
a7712927 | 16669 | |
f9273c43 | 16670 | case FRAME_POINTER_REGNUM: |
16671 | return offsets->soft_frame - offsets->saved_args; | |
a7712927 | 16672 | |
f9273c43 | 16673 | case ARM_HARD_FRAME_POINTER_REGNUM: |
16674 | return offsets->saved_regs - offsets->saved_args; | |
a7712927 | 16675 | |
f90b51f1 | 16676 | case THUMB_HARD_FRAME_POINTER_REGNUM: |
16677 | return offsets->locals_base - offsets->saved_args; | |
16678 | ||
f9273c43 | 16679 | default: |
ed29c566 | 16680 | gcc_unreachable (); |
f9273c43 | 16681 | } |
16682 | break; | |
a7712927 | 16683 | |
f9273c43 | 16684 | case FRAME_POINTER_REGNUM: |
16685 | switch (to) | |
16686 | { | |
16687 | case STACK_POINTER_REGNUM: | |
16688 | return offsets->outgoing_args - offsets->soft_frame; | |
a7712927 | 16689 | |
f9273c43 | 16690 | case ARM_HARD_FRAME_POINTER_REGNUM: |
16691 | return offsets->saved_regs - offsets->soft_frame; | |
a7712927 | 16692 | |
f90b51f1 | 16693 | case THUMB_HARD_FRAME_POINTER_REGNUM: |
16694 | return offsets->locals_base - offsets->soft_frame; | |
16695 | ||
f9273c43 | 16696 | default: |
ed29c566 | 16697 | gcc_unreachable (); |
f9273c43 | 16698 | } |
16699 | break; | |
a7712927 | 16700 | |
f9273c43 | 16701 | default: |
ed29c566 | 16702 | gcc_unreachable (); |
f9273c43 | 16703 | } |
a7712927 | 16704 | } |
16705 | ||
cffb2a26 | 16706 | /* Generate the rest of a function's prologue. */ |
16707 | void | |
25f905c2 | 16708 | thumb1_expand_prologue (void) |
cffb2a26 | 16709 | { |
37826745 | 16710 | rtx insn, dwarf; |
16711 | ||
f9273c43 | 16712 | HOST_WIDE_INT amount; |
16713 | arm_stack_offsets *offsets; | |
e27ad2d5 | 16714 | unsigned long func_type; |
ef5bbe67 | 16715 | int regno; |
237533cc | 16716 | unsigned long live_regs_mask; |
e27ad2d5 | 16717 | |
16718 | func_type = arm_current_func_type (); | |
9e7454d0 | 16719 | |
cffb2a26 | 16720 | /* Naked functions don't have prologues. */ |
e27ad2d5 | 16721 | if (IS_NAKED (func_type)) |
cffb2a26 | 16722 | return; |
16723 | ||
e27ad2d5 | 16724 | if (IS_INTERRUPT (func_type)) |
16725 | { | |
68435912 | 16726 | error ("interrupt Service Routines cannot be coded in Thumb mode"); |
e27ad2d5 | 16727 | return; |
16728 | } | |
16729 | ||
25f905c2 | 16730 | live_regs_mask = thumb1_compute_save_reg_mask (); |
53f179c0 | 16731 | /* Load the pic register before setting the frame pointer, |
16732 | so we can use r7 as a temporary work register. */ | |
2cb7d577 | 16733 | if (flag_pic && arm_pic_register != INVALID_REGNUM) |
aa6af490 | 16734 | arm_load_pic_register (live_regs_mask); |
ea32a06e | 16735 | |
f90b51f1 | 16736 | if (!frame_pointer_needed && CALLER_INTERWORKING_SLOT_SIZE > 0) |
150502c9 | 16737 | emit_move_insn (gen_rtx_REG (Pmode, ARM_HARD_FRAME_POINTER_REGNUM), |
16738 | stack_pointer_rtx); | |
cffb2a26 | 16739 | |
f90b51f1 | 16740 | offsets = arm_get_frame_offsets (); |
f9273c43 | 16741 | amount = offsets->outgoing_args - offsets->saved_regs; |
cffb2a26 | 16742 | if (amount) |
16743 | { | |
cffb2a26 | 16744 | if (amount < 512) |
37826745 | 16745 | { |
16746 | insn = emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, | |
16747 | GEN_INT (- amount))); | |
16748 | RTX_FRAME_RELATED_P (insn) = 1; | |
16749 | } | |
cffb2a26 | 16750 | else |
16751 | { | |
cffb2a26 | 16752 | rtx reg; |
16753 | ||
16754 | /* The stack decrement is too big for an immediate value in a single | |
16755 | insn. In theory we could issue multiple subtracts, but after | |
16756 | three of them it becomes more space efficient to place the full | |
16757 | value in the constant pool and load into a register. (Also the | |
16758 | ARM debugger really likes to see only one stack decrement per | |
16759 | function). So instead we look for a scratch register into which | |
16760 | we can load the decrement, and then we subtract this from the | |
16761 | stack pointer. Unfortunately on the thumb the only available | |
16762 | scratch registers are the argument registers, and we cannot use | |
16763 | these as they may hold arguments to the function. Instead we | |
16764 | attempt to locate a call preserved register which is used by this | |
16765 | function. If we can find one, then we know that it will have | |
16766 | been pushed at the start of the prologue and so we can corrupt | |
16767 | it now. */ | |
16768 | for (regno = LAST_ARG_REGNUM + 1; regno <= LAST_LO_REGNUM; regno++) | |
237533cc | 16769 | if (live_regs_mask & (1 << regno) |
215b30b3 | 16770 | && !(frame_pointer_needed |
16771 | && (regno == THUMB_HARD_FRAME_POINTER_REGNUM))) | |
cffb2a26 | 16772 | break; |
16773 | ||
9ac46427 | 16774 | if (regno > LAST_LO_REGNUM) /* Very unlikely. */ |
cffb2a26 | 16775 | { |
1a83b3ff | 16776 | rtx spare = gen_rtx_REG (SImode, IP_REGNUM); |
cffb2a26 | 16777 | |
b89db9a7 | 16778 | /* Choose an arbitrary, non-argument low register. */ |
1a83b3ff | 16779 | reg = gen_rtx_REG (SImode, LAST_LO_REGNUM); |
cffb2a26 | 16780 | |
16781 | /* Save it by copying it into a high, scratch register. */ | |
bf82d677 | 16782 | emit_insn (gen_movsi (spare, reg)); |
16783 | /* Add a USE to stop propagate_one_insn() from barfing. */ | |
063a05c7 | 16784 | emit_insn (gen_prologue_use (spare)); |
cffb2a26 | 16785 | |
16786 | /* Decrement the stack. */ | |
a12be13c | 16787 | emit_insn (gen_movsi (reg, GEN_INT (- amount))); |
37826745 | 16788 | insn = emit_insn (gen_addsi3 (stack_pointer_rtx, |
16789 | stack_pointer_rtx, reg)); | |
16790 | RTX_FRAME_RELATED_P (insn) = 1; | |
0083a3b9 | 16791 | dwarf = gen_rtx_SET (VOIDmode, stack_pointer_rtx, |
37826745 | 16792 | plus_constant (stack_pointer_rtx, |
af204f9f | 16793 | -amount)); |
37826745 | 16794 | RTX_FRAME_RELATED_P (dwarf) = 1; |
16795 | REG_NOTES (insn) | |
16796 | = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, dwarf, | |
16797 | REG_NOTES (insn)); | |
cffb2a26 | 16798 | |
16799 | /* Restore the low register's original value. */ | |
16800 | emit_insn (gen_movsi (reg, spare)); | |
9e7454d0 | 16801 | |
cffb2a26 | 16802 | /* Emit a USE of the restored scratch register, so that flow |
16803 | analysis will not consider the restore redundant. The | |
16804 | register won't be used again in this function and isn't | |
16805 | restored by the epilogue. */ | |
063a05c7 | 16806 | emit_insn (gen_prologue_use (reg)); |
cffb2a26 | 16807 | } |
16808 | else | |
16809 | { | |
1a83b3ff | 16810 | reg = gen_rtx_REG (SImode, regno); |
cffb2a26 | 16811 | |
a12be13c | 16812 | emit_insn (gen_movsi (reg, GEN_INT (- amount))); |
37826745 | 16813 | |
16814 | insn = emit_insn (gen_addsi3 (stack_pointer_rtx, | |
16815 | stack_pointer_rtx, reg)); | |
16816 | RTX_FRAME_RELATED_P (insn) = 1; | |
0083a3b9 | 16817 | dwarf = gen_rtx_SET (VOIDmode, stack_pointer_rtx, |
37826745 | 16818 | plus_constant (stack_pointer_rtx, |
af204f9f | 16819 | -amount)); |
37826745 | 16820 | RTX_FRAME_RELATED_P (dwarf) = 1; |
16821 | REG_NOTES (insn) | |
16822 | = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, dwarf, | |
16823 | REG_NOTES (insn)); | |
cffb2a26 | 16824 | } |
16825 | } | |
f90b51f1 | 16826 | } |
16827 | ||
16828 | if (frame_pointer_needed) | |
25f905c2 | 16829 | thumb_set_frame_pointer (offsets); |
9e7454d0 | 16830 | |
363d487e | 16831 | /* If we are profiling, make sure no instructions are scheduled before |
16832 | the call to mcount. Similarly if the user has requested no | |
16833 | scheduling in the prolog. Similarly if we want non-call exceptions | |
16834 | using the EABI unwinder, to prevent faulting instructions from being | |
16835 | swapped with a stack adjustment. */ | |
16836 | if (current_function_profile || !TARGET_SCHED_PROLOG | |
16837 | || (ARM_EABI_UNWIND_TABLES && flag_non_call_exceptions)) | |
cffb2a26 | 16838 | emit_insn (gen_blockage ()); |
ef5bbe67 | 16839 | |
16840 | cfun->machine->lr_save_eliminated = !thumb_force_lr_save (); | |
237533cc | 16841 | if (live_regs_mask & 0xff) |
16842 | cfun->machine->lr_save_eliminated = 0; | |
cffb2a26 | 16843 | } |
16844 | ||
237533cc | 16845 | |
cffb2a26 | 16846 | void |
25f905c2 | 16847 | thumb1_expand_epilogue (void) |
cffb2a26 | 16848 | { |
f9273c43 | 16849 | HOST_WIDE_INT amount; |
16850 | arm_stack_offsets *offsets; | |
58d6528b | 16851 | int regno; |
16852 | ||
e27ad2d5 | 16853 | /* Naked functions don't have prologues. */ |
16854 | if (IS_NAKED (arm_current_func_type ())) | |
cffb2a26 | 16855 | return; |
16856 | ||
f9273c43 | 16857 | offsets = arm_get_frame_offsets (); |
16858 | amount = offsets->outgoing_args - offsets->saved_regs; | |
16859 | ||
cffb2a26 | 16860 | if (frame_pointer_needed) |
f90b51f1 | 16861 | { |
16862 | emit_insn (gen_movsi (stack_pointer_rtx, hard_frame_pointer_rtx)); | |
16863 | amount = offsets->locals_base - offsets->saved_regs; | |
16864 | } | |
a60f1840 | 16865 | |
a650636a | 16866 | gcc_assert (amount >= 0); |
f90b51f1 | 16867 | if (amount) |
cffb2a26 | 16868 | { |
cffb2a26 | 16869 | if (amount < 512) |
16870 | emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, | |
16871 | GEN_INT (amount))); | |
16872 | else | |
16873 | { | |
16874 | /* r3 is always free in the epilogue. */ | |
1a83b3ff | 16875 | rtx reg = gen_rtx_REG (SImode, LAST_ARG_REGNUM); |
cffb2a26 | 16876 | |
16877 | emit_insn (gen_movsi (reg, GEN_INT (amount))); | |
16878 | emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, reg)); | |
16879 | } | |
16880 | } | |
9e7454d0 | 16881 | |
cffb2a26 | 16882 | /* Emit a USE (stack_pointer_rtx), so that |
16883 | the stack adjustment will not be deleted. */ | |
063a05c7 | 16884 | emit_insn (gen_prologue_use (stack_pointer_rtx)); |
cffb2a26 | 16885 | |
81f6de2f | 16886 | if (current_function_profile || !TARGET_SCHED_PROLOG) |
cffb2a26 | 16887 | emit_insn (gen_blockage ()); |
58d6528b | 16888 | |
16889 | /* Emit a clobber for each insn that will be restored in the epilogue, | |
16890 | so that flow2 will get register lifetimes correct. */ | |
16891 | for (regno = 0; regno < 13; regno++) | |
3072d30e | 16892 | if (df_regs_ever_live_p (regno) && !call_used_regs[regno]) |
58d6528b | 16893 | emit_insn (gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (SImode, regno))); |
16894 | ||
3072d30e | 16895 | if (! df_regs_ever_live_p (LR_REGNUM)) |
58d6528b | 16896 | emit_insn (gen_rtx_USE (VOIDmode, gen_rtx_REG (SImode, LR_REGNUM))); |
cffb2a26 | 16897 | } |
16898 | ||
17d9b0c3 | 16899 | static void |
25f905c2 | 16900 | thumb1_output_function_prologue (FILE *f, HOST_WIDE_INT size ATTRIBUTE_UNUSED) |
cffb2a26 | 16901 | { |
53f179c0 | 16902 | unsigned long live_regs_mask = 0; |
16903 | unsigned long l_mask; | |
16904 | unsigned high_regs_pushed = 0; | |
37826745 | 16905 | int cfa_offset = 0; |
cffb2a26 | 16906 | int regno; |
16907 | ||
e27ad2d5 | 16908 | if (IS_NAKED (arm_current_func_type ())) |
cffb2a26 | 16909 | return; |
16910 | ||
16911 | if (is_called_in_ARM_mode (current_function_decl)) | |
16912 | { | |
16913 | const char * name; | |
16914 | ||
ed29c566 | 16915 | gcc_assert (GET_CODE (DECL_RTL (current_function_decl)) == MEM); |
16916 | gcc_assert (GET_CODE (XEXP (DECL_RTL (current_function_decl), 0)) | |
16917 | == SYMBOL_REF); | |
cffb2a26 | 16918 | name = XSTR (XEXP (DECL_RTL (current_function_decl), 0), 0); |
9e7454d0 | 16919 | |
cffb2a26 | 16920 | /* Generate code sequence to switch us into Thumb mode. */ |
16921 | /* The .code 32 directive has already been emitted by | |
f5443994 | 16922 | ASM_DECLARE_FUNCTION_NAME. */ |
cffb2a26 | 16923 | asm_fprintf (f, "\torr\t%r, %r, #1\n", IP_REGNUM, PC_REGNUM); |
16924 | asm_fprintf (f, "\tbx\t%r\n", IP_REGNUM); | |
16925 | ||
16926 | /* Generate a label, so that the debugger will notice the | |
16927 | change in instruction sets. This label is also used by | |
16928 | the assembler to bypass the ARM code when this function | |
16929 | is called from a Thumb encoded function elsewhere in the | |
16930 | same file. Hence the definition of STUB_NAME here must | |
674a8f0b | 16931 | agree with the definition in gas/config/tc-arm.c. */ |
9e7454d0 | 16932 | |
cffb2a26 | 16933 | #define STUB_NAME ".real_start_of" |
9e7454d0 | 16934 | |
7fe1d31c | 16935 | fprintf (f, "\t.code\t16\n"); |
cffb2a26 | 16936 | #ifdef ARM_PE |
16937 | if (arm_dllexport_name_p (name)) | |
08a98e4f | 16938 | name = arm_strip_name_encoding (name); |
9e7454d0 | 16939 | #endif |
cffb2a26 | 16940 | asm_fprintf (f, "\t.globl %s%U%s\n", STUB_NAME, name); |
7fe1d31c | 16941 | fprintf (f, "\t.thumb_func\n"); |
cffb2a26 | 16942 | asm_fprintf (f, "%s%U%s:\n", STUB_NAME, name); |
16943 | } | |
9e7454d0 | 16944 | |
cffb2a26 | 16945 | if (current_function_pretend_args_size) |
16946 | { | |
1774763d | 16947 | /* Output unwind directive for the stack adjustment. */ |
16948 | if (ARM_EABI_UNWIND_TABLES) | |
16949 | fprintf (f, "\t.pad #%d\n", | |
16950 | current_function_pretend_args_size); | |
16951 | ||
71904594 | 16952 | if (cfun->machine->uses_anonymous_args) |
cffb2a26 | 16953 | { |
16954 | int num_pushes; | |
9e7454d0 | 16955 | |
7fe1d31c | 16956 | fprintf (f, "\tpush\t{"); |
cffb2a26 | 16957 | |
36837fde | 16958 | num_pushes = ARM_NUM_INTS (current_function_pretend_args_size); |
9e7454d0 | 16959 | |
cffb2a26 | 16960 | for (regno = LAST_ARG_REGNUM + 1 - num_pushes; |
16961 | regno <= LAST_ARG_REGNUM; | |
215b30b3 | 16962 | regno++) |
cffb2a26 | 16963 | asm_fprintf (f, "%r%s", regno, |
16964 | regno == LAST_ARG_REGNUM ? "" : ", "); | |
16965 | ||
7fe1d31c | 16966 | fprintf (f, "}\n"); |
cffb2a26 | 16967 | } |
16968 | else | |
9e7454d0 | 16969 | asm_fprintf (f, "\tsub\t%r, %r, #%d\n", |
cffb2a26 | 16970 | SP_REGNUM, SP_REGNUM, |
16971 | current_function_pretend_args_size); | |
37826745 | 16972 | |
16973 | /* We don't need to record the stores for unwinding (would it | |
16974 | help the debugger any if we did?), but record the change in | |
16975 | the stack pointer. */ | |
16976 | if (dwarf2out_do_frame ()) | |
16977 | { | |
16978 | char *l = dwarf2out_cfi_label (); | |
53f179c0 | 16979 | |
37826745 | 16980 | cfa_offset = cfa_offset + current_function_pretend_args_size; |
16981 | dwarf2out_def_cfa (l, SP_REGNUM, cfa_offset); | |
16982 | } | |
cffb2a26 | 16983 | } |
16984 | ||
53f179c0 | 16985 | /* Get the registers we are going to push. */ |
25f905c2 | 16986 | live_regs_mask = thumb1_compute_save_reg_mask (); |
53f179c0 | 16987 | /* Extract a mask of the ones we can give to the Thumb's push instruction. */ |
237533cc | 16988 | l_mask = live_regs_mask & 0x40ff; |
53f179c0 | 16989 | /* Then count how many other high registers will need to be pushed. */ |
16990 | high_regs_pushed = bit_count (live_regs_mask & 0x0f00); | |
cffb2a26 | 16991 | |
16992 | if (TARGET_BACKTRACE) | |
16993 | { | |
53f179c0 | 16994 | unsigned offset; |
16995 | unsigned work_register; | |
9e7454d0 | 16996 | |
cffb2a26 | 16997 | /* We have been asked to create a stack backtrace structure. |
16998 | The code looks like this: | |
9e7454d0 | 16999 | |
cffb2a26 | 17000 | 0 .align 2 |
17001 | 0 func: | |
17002 | 0 sub SP, #16 Reserve space for 4 registers. | |
237533cc | 17003 | 2 push {R7} Push low registers. |
cffb2a26 | 17004 | 4 add R7, SP, #20 Get the stack pointer before the push. |
17005 | 6 str R7, [SP, #8] Store the stack pointer (before reserving the space). | |
17006 | 8 mov R7, PC Get hold of the start of this code plus 12. | |
17007 | 10 str R7, [SP, #16] Store it. | |
17008 | 12 mov R7, FP Get hold of the current frame pointer. | |
17009 | 14 str R7, [SP, #4] Store it. | |
17010 | 16 mov R7, LR Get hold of the current return address. | |
17011 | 18 str R7, [SP, #12] Store it. | |
17012 | 20 add R7, SP, #16 Point at the start of the backtrace structure. | |
17013 | 22 mov FP, R7 Put this value into the frame pointer. */ | |
17014 | ||
237533cc | 17015 | work_register = thumb_find_work_register (live_regs_mask); |
9e7454d0 | 17016 | |
1774763d | 17017 | if (ARM_EABI_UNWIND_TABLES) |
17018 | asm_fprintf (f, "\t.pad #16\n"); | |
17019 | ||
cffb2a26 | 17020 | asm_fprintf |
17021 | (f, "\tsub\t%r, %r, #16\t%@ Create stack backtrace structure\n", | |
17022 | SP_REGNUM, SP_REGNUM); | |
37826745 | 17023 | |
17024 | if (dwarf2out_do_frame ()) | |
17025 | { | |
17026 | char *l = dwarf2out_cfi_label (); | |
53f179c0 | 17027 | |
37826745 | 17028 | cfa_offset = cfa_offset + 16; |
17029 | dwarf2out_def_cfa (l, SP_REGNUM, cfa_offset); | |
17030 | } | |
17031 | ||
237533cc | 17032 | if (l_mask) |
17033 | { | |
17034 | thumb_pushpop (f, l_mask, 1, &cfa_offset, l_mask); | |
e54325a1 | 17035 | offset = bit_count (l_mask) * UNITS_PER_WORD; |
237533cc | 17036 | } |
17037 | else | |
17038 | offset = 0; | |
9e7454d0 | 17039 | |
cffb2a26 | 17040 | asm_fprintf (f, "\tadd\t%r, %r, #%d\n", work_register, SP_REGNUM, |
17041 | offset + 16 + current_function_pretend_args_size); | |
9e7454d0 | 17042 | |
cffb2a26 | 17043 | asm_fprintf (f, "\tstr\t%r, [%r, #%d]\n", work_register, SP_REGNUM, |
17044 | offset + 4); | |
17045 | ||
17046 | /* Make sure that the instruction fetching the PC is in the right place | |
17047 | to calculate "start of backtrace creation code + 12". */ | |
237533cc | 17048 | if (l_mask) |
cffb2a26 | 17049 | { |
17050 | asm_fprintf (f, "\tmov\t%r, %r\n", work_register, PC_REGNUM); | |
17051 | asm_fprintf (f, "\tstr\t%r, [%r, #%d]\n", work_register, SP_REGNUM, | |
17052 | offset + 12); | |
17053 | asm_fprintf (f, "\tmov\t%r, %r\n", work_register, | |
17054 | ARM_HARD_FRAME_POINTER_REGNUM); | |
17055 | asm_fprintf (f, "\tstr\t%r, [%r, #%d]\n", work_register, SP_REGNUM, | |
17056 | offset); | |
17057 | } | |
17058 | else | |
17059 | { | |
17060 | asm_fprintf (f, "\tmov\t%r, %r\n", work_register, | |
17061 | ARM_HARD_FRAME_POINTER_REGNUM); | |
17062 | asm_fprintf (f, "\tstr\t%r, [%r, #%d]\n", work_register, SP_REGNUM, | |
17063 | offset); | |
17064 | asm_fprintf (f, "\tmov\t%r, %r\n", work_register, PC_REGNUM); | |
17065 | asm_fprintf (f, "\tstr\t%r, [%r, #%d]\n", work_register, SP_REGNUM, | |
17066 | offset + 12); | |
17067 | } | |
9e7454d0 | 17068 | |
cffb2a26 | 17069 | asm_fprintf (f, "\tmov\t%r, %r\n", work_register, LR_REGNUM); |
17070 | asm_fprintf (f, "\tstr\t%r, [%r, #%d]\n", work_register, SP_REGNUM, | |
17071 | offset + 8); | |
17072 | asm_fprintf (f, "\tadd\t%r, %r, #%d\n", work_register, SP_REGNUM, | |
17073 | offset + 12); | |
17074 | asm_fprintf (f, "\tmov\t%r, %r\t\t%@ Backtrace structure created\n", | |
17075 | ARM_HARD_FRAME_POINTER_REGNUM, work_register); | |
17076 | } | |
442e3cb9 | 17077 | /* Optimization: If we are not pushing any low registers but we are going |
53f179c0 | 17078 | to push some high registers then delay our first push. This will just |
17079 | be a push of LR and we can combine it with the push of the first high | |
17080 | register. */ | |
17081 | else if ((l_mask & 0xff) != 0 | |
17082 | || (high_regs_pushed == 0 && l_mask)) | |
237533cc | 17083 | thumb_pushpop (f, l_mask, 1, &cfa_offset, l_mask); |
cffb2a26 | 17084 | |
cffb2a26 | 17085 | if (high_regs_pushed) |
17086 | { | |
53f179c0 | 17087 | unsigned pushable_regs; |
17088 | unsigned next_hi_reg; | |
cffb2a26 | 17089 | |
17090 | for (next_hi_reg = 12; next_hi_reg > LAST_LO_REGNUM; next_hi_reg--) | |
237533cc | 17091 | if (live_regs_mask & (1 << next_hi_reg)) |
9f468b67 | 17092 | break; |
cffb2a26 | 17093 | |
237533cc | 17094 | pushable_regs = l_mask & 0xff; |
cffb2a26 | 17095 | |
17096 | if (pushable_regs == 0) | |
237533cc | 17097 | pushable_regs = 1 << thumb_find_work_register (live_regs_mask); |
cffb2a26 | 17098 | |
17099 | while (high_regs_pushed > 0) | |
17100 | { | |
53f179c0 | 17101 | unsigned long real_regs_mask = 0; |
37826745 | 17102 | |
53f179c0 | 17103 | for (regno = LAST_LO_REGNUM; regno >= 0; regno --) |
cffb2a26 | 17104 | { |
237533cc | 17105 | if (pushable_regs & (1 << regno)) |
cffb2a26 | 17106 | { |
17107 | asm_fprintf (f, "\tmov\t%r, %r\n", regno, next_hi_reg); | |
9e7454d0 | 17108 | |
53f179c0 | 17109 | high_regs_pushed --; |
37826745 | 17110 | real_regs_mask |= (1 << next_hi_reg); |
9e7454d0 | 17111 | |
cffb2a26 | 17112 | if (high_regs_pushed) |
9ac46427 | 17113 | { |
53f179c0 | 17114 | for (next_hi_reg --; next_hi_reg > LAST_LO_REGNUM; |
17115 | next_hi_reg --) | |
237533cc | 17116 | if (live_regs_mask & (1 << next_hi_reg)) |
cffb2a26 | 17117 | break; |
9ac46427 | 17118 | } |
cffb2a26 | 17119 | else |
17120 | { | |
237533cc | 17121 | pushable_regs &= ~((1 << regno) - 1); |
cffb2a26 | 17122 | break; |
17123 | } | |
17124 | } | |
17125 | } | |
37826745 | 17126 | |
53f179c0 | 17127 | /* If we had to find a work register and we have not yet |
17128 | saved the LR then add it to the list of regs to push. */ | |
17129 | if (l_mask == (1 << LR_REGNUM)) | |
17130 | { | |
17131 | thumb_pushpop (f, pushable_regs | (1 << LR_REGNUM), | |
17132 | 1, &cfa_offset, | |
17133 | real_regs_mask | (1 << LR_REGNUM)); | |
17134 | l_mask = 0; | |
17135 | } | |
17136 | else | |
17137 | thumb_pushpop (f, pushable_regs, 1, &cfa_offset, real_regs_mask); | |
cffb2a26 | 17138 | } |
cffb2a26 | 17139 | } |
17140 | } | |
17141 | ||
17142 | /* Handle the case of a double word load into a low register from | |
17143 | a computed memory address. The computed address may involve a | |
17144 | register which is overwritten by the load. */ | |
0bd59681 | 17145 | const char * |
ebd88a36 | 17146 | thumb_load_double_from_address (rtx *operands) |
cffb2a26 | 17147 | { |
17148 | rtx addr; | |
17149 | rtx base; | |
17150 | rtx offset; | |
17151 | rtx arg1; | |
17152 | rtx arg2; | |
9e7454d0 | 17153 | |
ed29c566 | 17154 | gcc_assert (GET_CODE (operands[0]) == REG); |
17155 | gcc_assert (GET_CODE (operands[1]) == MEM); | |
cffb2a26 | 17156 | |
17157 | /* Get the memory address. */ | |
17158 | addr = XEXP (operands[1], 0); | |
9e7454d0 | 17159 | |
cffb2a26 | 17160 | /* Work out how the memory address is computed. */ |
17161 | switch (GET_CODE (addr)) | |
17162 | { | |
17163 | case REG: | |
788fcce0 | 17164 | operands[2] = adjust_address (operands[1], SImode, 4); |
a60f1840 | 17165 | |
cffb2a26 | 17166 | if (REGNO (operands[0]) == REGNO (addr)) |
17167 | { | |
17168 | output_asm_insn ("ldr\t%H0, %2", operands); | |
17169 | output_asm_insn ("ldr\t%0, %1", operands); | |
17170 | } | |
17171 | else | |
17172 | { | |
17173 | output_asm_insn ("ldr\t%0, %1", operands); | |
17174 | output_asm_insn ("ldr\t%H0, %2", operands); | |
17175 | } | |
17176 | break; | |
9e7454d0 | 17177 | |
cffb2a26 | 17178 | case CONST: |
17179 | /* Compute <address> + 4 for the high order load. */ | |
788fcce0 | 17180 | operands[2] = adjust_address (operands[1], SImode, 4); |
a60f1840 | 17181 | |
cffb2a26 | 17182 | output_asm_insn ("ldr\t%0, %1", operands); |
17183 | output_asm_insn ("ldr\t%H0, %2", operands); | |
17184 | break; | |
9e7454d0 | 17185 | |
cffb2a26 | 17186 | case PLUS: |
17187 | arg1 = XEXP (addr, 0); | |
17188 | arg2 = XEXP (addr, 1); | |
9e7454d0 | 17189 | |
cffb2a26 | 17190 | if (CONSTANT_P (arg1)) |
17191 | base = arg2, offset = arg1; | |
17192 | else | |
17193 | base = arg1, offset = arg2; | |
9e7454d0 | 17194 | |
ed29c566 | 17195 | gcc_assert (GET_CODE (base) == REG); |
cffb2a26 | 17196 | |
17197 | /* Catch the case of <address> = <reg> + <reg> */ | |
17198 | if (GET_CODE (offset) == REG) | |
17199 | { | |
17200 | int reg_offset = REGNO (offset); | |
17201 | int reg_base = REGNO (base); | |
17202 | int reg_dest = REGNO (operands[0]); | |
9e7454d0 | 17203 | |
cffb2a26 | 17204 | /* Add the base and offset registers together into the |
17205 | higher destination register. */ | |
17206 | asm_fprintf (asm_out_file, "\tadd\t%r, %r, %r", | |
17207 | reg_dest + 1, reg_base, reg_offset); | |
9e7454d0 | 17208 | |
cffb2a26 | 17209 | /* Load the lower destination register from the address in |
17210 | the higher destination register. */ | |
17211 | asm_fprintf (asm_out_file, "\tldr\t%r, [%r, #0]", | |
17212 | reg_dest, reg_dest + 1); | |
9e7454d0 | 17213 | |
cffb2a26 | 17214 | /* Load the higher destination register from its own address |
17215 | plus 4. */ | |
17216 | asm_fprintf (asm_out_file, "\tldr\t%r, [%r, #4]", | |
17217 | reg_dest + 1, reg_dest + 1); | |
17218 | } | |
17219 | else | |
17220 | { | |
17221 | /* Compute <address> + 4 for the high order load. */ | |
788fcce0 | 17222 | operands[2] = adjust_address (operands[1], SImode, 4); |
9e7454d0 | 17223 | |
cffb2a26 | 17224 | /* If the computed address is held in the low order register |
17225 | then load the high order register first, otherwise always | |
17226 | load the low order register first. */ | |
17227 | if (REGNO (operands[0]) == REGNO (base)) | |
17228 | { | |
17229 | output_asm_insn ("ldr\t%H0, %2", operands); | |
17230 | output_asm_insn ("ldr\t%0, %1", operands); | |
17231 | } | |
17232 | else | |
17233 | { | |
17234 | output_asm_insn ("ldr\t%0, %1", operands); | |
17235 | output_asm_insn ("ldr\t%H0, %2", operands); | |
17236 | } | |
17237 | } | |
17238 | break; | |
17239 | ||
17240 | case LABEL_REF: | |
17241 | /* With no registers to worry about we can just load the value | |
17242 | directly. */ | |
788fcce0 | 17243 | operands[2] = adjust_address (operands[1], SImode, 4); |
9e7454d0 | 17244 | |
cffb2a26 | 17245 | output_asm_insn ("ldr\t%H0, %2", operands); |
17246 | output_asm_insn ("ldr\t%0, %1", operands); | |
17247 | break; | |
9e7454d0 | 17248 | |
cffb2a26 | 17249 | default: |
ed29c566 | 17250 | gcc_unreachable (); |
cffb2a26 | 17251 | } |
9e7454d0 | 17252 | |
cffb2a26 | 17253 | return ""; |
17254 | } | |
17255 | ||
0bd59681 | 17256 | const char * |
ebd88a36 | 17257 | thumb_output_move_mem_multiple (int n, rtx *operands) |
cffb2a26 | 17258 | { |
17259 | rtx tmp; | |
17260 | ||
17261 | switch (n) | |
17262 | { | |
17263 | case 2: | |
960f3acf | 17264 | if (REGNO (operands[4]) > REGNO (operands[5])) |
cffb2a26 | 17265 | { |
960f3acf | 17266 | tmp = operands[4]; |
17267 | operands[4] = operands[5]; | |
17268 | operands[5] = tmp; | |
cffb2a26 | 17269 | } |
960f3acf | 17270 | output_asm_insn ("ldmia\t%1!, {%4, %5}", operands); |
17271 | output_asm_insn ("stmia\t%0!, {%4, %5}", operands); | |
cffb2a26 | 17272 | break; |
17273 | ||
17274 | case 3: | |
960f3acf | 17275 | if (REGNO (operands[4]) > REGNO (operands[5])) |
cffb2a26 | 17276 | { |
960f3acf | 17277 | tmp = operands[4]; |
17278 | operands[4] = operands[5]; | |
17279 | operands[5] = tmp; | |
cffb2a26 | 17280 | } |
960f3acf | 17281 | if (REGNO (operands[5]) > REGNO (operands[6])) |
cffb2a26 | 17282 | { |
960f3acf | 17283 | tmp = operands[5]; |
17284 | operands[5] = operands[6]; | |
17285 | operands[6] = tmp; | |
cffb2a26 | 17286 | } |
960f3acf | 17287 | if (REGNO (operands[4]) > REGNO (operands[5])) |
cffb2a26 | 17288 | { |
960f3acf | 17289 | tmp = operands[4]; |
17290 | operands[4] = operands[5]; | |
17291 | operands[5] = tmp; | |
cffb2a26 | 17292 | } |
9e7454d0 | 17293 | |
960f3acf | 17294 | output_asm_insn ("ldmia\t%1!, {%4, %5, %6}", operands); |
17295 | output_asm_insn ("stmia\t%0!, {%4, %5, %6}", operands); | |
cffb2a26 | 17296 | break; |
17297 | ||
17298 | default: | |
ed29c566 | 17299 | gcc_unreachable (); |
cffb2a26 | 17300 | } |
17301 | ||
17302 | return ""; | |
17303 | } | |
17304 | ||
afe27f3b | 17305 | /* Output a call-via instruction for thumb state. */ |
17306 | const char * | |
17307 | thumb_call_via_reg (rtx reg) | |
17308 | { | |
17309 | int regno = REGNO (reg); | |
17310 | rtx *labelp; | |
17311 | ||
a52972b8 | 17312 | gcc_assert (regno < LR_REGNUM); |
afe27f3b | 17313 | |
17314 | /* If we are in the normal text section we can use a single instance | |
17315 | per compilation unit. If we are doing function sections, then we need | |
17316 | an entry per section, since we can't rely on reachability. */ | |
2f14b1f9 | 17317 | if (in_section == text_section) |
afe27f3b | 17318 | { |
17319 | thumb_call_reg_needed = 1; | |
17320 | ||
17321 | if (thumb_call_via_label[regno] == NULL) | |
17322 | thumb_call_via_label[regno] = gen_label_rtx (); | |
17323 | labelp = thumb_call_via_label + regno; | |
17324 | } | |
17325 | else | |
17326 | { | |
17327 | if (cfun->machine->call_via[regno] == NULL) | |
17328 | cfun->machine->call_via[regno] = gen_label_rtx (); | |
17329 | labelp = cfun->machine->call_via + regno; | |
17330 | } | |
17331 | ||
17332 | output_asm_insn ("bl\t%a0", labelp); | |
17333 | return ""; | |
17334 | } | |
17335 | ||
a12be13c | 17336 | /* Routines for generating rtl. */ |
cffb2a26 | 17337 | void |
008c057d | 17338 | thumb_expand_movmemqi (rtx *operands) |
cffb2a26 | 17339 | { |
17340 | rtx out = copy_to_mode_reg (SImode, XEXP (operands[0], 0)); | |
17341 | rtx in = copy_to_mode_reg (SImode, XEXP (operands[1], 0)); | |
17342 | HOST_WIDE_INT len = INTVAL (operands[2]); | |
17343 | HOST_WIDE_INT offset = 0; | |
17344 | ||
17345 | while (len >= 12) | |
17346 | { | |
960f3acf | 17347 | emit_insn (gen_movmem12b (out, in, out, in)); |
cffb2a26 | 17348 | len -= 12; |
17349 | } | |
9e7454d0 | 17350 | |
cffb2a26 | 17351 | if (len >= 8) |
17352 | { | |
960f3acf | 17353 | emit_insn (gen_movmem8b (out, in, out, in)); |
cffb2a26 | 17354 | len -= 8; |
17355 | } | |
9e7454d0 | 17356 | |
cffb2a26 | 17357 | if (len >= 4) |
17358 | { | |
17359 | rtx reg = gen_reg_rtx (SImode); | |
1a83b3ff | 17360 | emit_insn (gen_movsi (reg, gen_rtx_MEM (SImode, in))); |
17361 | emit_insn (gen_movsi (gen_rtx_MEM (SImode, out), reg)); | |
cffb2a26 | 17362 | len -= 4; |
17363 | offset += 4; | |
17364 | } | |
9e7454d0 | 17365 | |
cffb2a26 | 17366 | if (len >= 2) |
17367 | { | |
17368 | rtx reg = gen_reg_rtx (HImode); | |
9e7454d0 | 17369 | emit_insn (gen_movhi (reg, gen_rtx_MEM (HImode, |
1a83b3ff | 17370 | plus_constant (in, offset)))); |
17371 | emit_insn (gen_movhi (gen_rtx_MEM (HImode, plus_constant (out, offset)), | |
cffb2a26 | 17372 | reg)); |
17373 | len -= 2; | |
17374 | offset += 2; | |
17375 | } | |
9e7454d0 | 17376 | |
cffb2a26 | 17377 | if (len) |
17378 | { | |
17379 | rtx reg = gen_reg_rtx (QImode); | |
1a83b3ff | 17380 | emit_insn (gen_movqi (reg, gen_rtx_MEM (QImode, |
17381 | plus_constant (in, offset)))); | |
17382 | emit_insn (gen_movqi (gen_rtx_MEM (QImode, plus_constant (out, offset)), | |
cffb2a26 | 17383 | reg)); |
17384 | } | |
17385 | } | |
17386 | ||
cffb2a26 | 17387 | void |
ebd88a36 | 17388 | thumb_reload_out_hi (rtx *operands) |
cffb2a26 | 17389 | { |
17390 | emit_insn (gen_thumb_movhi_clobber (operands[0], operands[1], operands[2])); | |
17391 | } | |
17392 | ||
9e7454d0 | 17393 | /* Handle reading a half-word from memory during reload. */ |
cffb2a26 | 17394 | void |
ebd88a36 | 17395 | thumb_reload_in_hi (rtx *operands ATTRIBUTE_UNUSED) |
cffb2a26 | 17396 | { |
ed29c566 | 17397 | gcc_unreachable (); |
cffb2a26 | 17398 | } |
17399 | ||
78fe751b | 17400 | /* Return the length of a function name prefix |
17401 | that starts with the character 'c'. */ | |
17402 | static int | |
ebd88a36 | 17403 | arm_get_strip_length (int c) |
78fe751b | 17404 | { |
17405 | switch (c) | |
17406 | { | |
17407 | ARM_NAME_ENCODING_LENGTHS | |
9e7454d0 | 17408 | default: return 0; |
78fe751b | 17409 | } |
17410 | } | |
17411 | ||
17412 | /* Return a pointer to a function's name with any | |
17413 | and all prefix encodings stripped from it. */ | |
17414 | const char * | |
ebd88a36 | 17415 | arm_strip_name_encoding (const char *name) |
78fe751b | 17416 | { |
17417 | int skip; | |
9e7454d0 | 17418 | |
78fe751b | 17419 | while ((skip = arm_get_strip_length (* name))) |
17420 | name += skip; | |
17421 | ||
17422 | return name; | |
17423 | } | |
17424 | ||
d2fcdd8d | 17425 | /* If there is a '*' anywhere in the name's prefix, then |
17426 | emit the stripped name verbatim, otherwise prepend an | |
17427 | underscore if leading underscores are being used. */ | |
d2fcdd8d | 17428 | void |
ebd88a36 | 17429 | arm_asm_output_labelref (FILE *stream, const char *name) |
d2fcdd8d | 17430 | { |
17431 | int skip; | |
17432 | int verbatim = 0; | |
17433 | ||
17434 | while ((skip = arm_get_strip_length (* name))) | |
17435 | { | |
17436 | verbatim |= (*name == '*'); | |
17437 | name += skip; | |
17438 | } | |
17439 | ||
17440 | if (verbatim) | |
17441 | fputs (name, stream); | |
17442 | else | |
17443 | asm_fprintf (stream, "%U%s", name); | |
17444 | } | |
17445 | ||
5fad5123 | 17446 | static void |
17447 | arm_file_start (void) | |
17448 | { | |
17449 | int val; | |
17450 | ||
25f905c2 | 17451 | if (TARGET_UNIFIED_ASM) |
17452 | asm_fprintf (asm_out_file, "\t.syntax unified\n"); | |
17453 | ||
5fad5123 | 17454 | if (TARGET_BPABI) |
17455 | { | |
17456 | const char *fpu_name; | |
17457 | if (arm_select[0].string) | |
17458 | asm_fprintf (asm_out_file, "\t.cpu %s\n", arm_select[0].string); | |
17459 | else if (arm_select[1].string) | |
17460 | asm_fprintf (asm_out_file, "\t.arch %s\n", arm_select[1].string); | |
17461 | else | |
17462 | asm_fprintf (asm_out_file, "\t.cpu %s\n", | |
17463 | all_cores[arm_default_cpu].name); | |
17464 | ||
17465 | if (TARGET_SOFT_FLOAT) | |
17466 | { | |
17467 | if (TARGET_VFP) | |
17468 | fpu_name = "softvfp"; | |
17469 | else | |
17470 | fpu_name = "softfpa"; | |
17471 | } | |
17472 | else | |
17473 | { | |
b5a0636d | 17474 | int set_float_abi_attributes = 0; |
5fad5123 | 17475 | switch (arm_fpu_arch) |
17476 | { | |
17477 | case FPUTYPE_FPA: | |
17478 | fpu_name = "fpa"; | |
17479 | break; | |
17480 | case FPUTYPE_FPA_EMU2: | |
17481 | fpu_name = "fpe2"; | |
17482 | break; | |
17483 | case FPUTYPE_FPA_EMU3: | |
17484 | fpu_name = "fpe3"; | |
17485 | break; | |
17486 | case FPUTYPE_MAVERICK: | |
17487 | fpu_name = "maverick"; | |
17488 | break; | |
17489 | case FPUTYPE_VFP: | |
5fad5123 | 17490 | fpu_name = "vfp"; |
b5a0636d | 17491 | set_float_abi_attributes = 1; |
17492 | break; | |
17493 | case FPUTYPE_VFP3: | |
17494 | fpu_name = "vfp3"; | |
17495 | set_float_abi_attributes = 1; | |
5fad5123 | 17496 | break; |
d98a3884 | 17497 | case FPUTYPE_NEON: |
17498 | fpu_name = "neon"; | |
17499 | set_float_abi_attributes = 1; | |
17500 | break; | |
5fad5123 | 17501 | default: |
17502 | abort(); | |
17503 | } | |
b5a0636d | 17504 | if (set_float_abi_attributes) |
17505 | { | |
17506 | if (TARGET_HARD_FLOAT) | |
17507 | asm_fprintf (asm_out_file, "\t.eabi_attribute 27, 3\n"); | |
17508 | if (TARGET_HARD_FLOAT_ABI) | |
17509 | asm_fprintf (asm_out_file, "\t.eabi_attribute 28, 1\n"); | |
17510 | } | |
5fad5123 | 17511 | } |
17512 | asm_fprintf (asm_out_file, "\t.fpu %s\n", fpu_name); | |
17513 | ||
17514 | /* Some of these attributes only apply when the corresponding features | |
17515 | are used. However we don't have any easy way of figuring this out. | |
17516 | Conservatively record the setting that would have been used. */ | |
17517 | ||
17518 | /* Tag_ABI_PCS_wchar_t. */ | |
17519 | asm_fprintf (asm_out_file, "\t.eabi_attribute 18, %d\n", | |
17520 | (int)WCHAR_TYPE_SIZE / BITS_PER_UNIT); | |
17521 | ||
17522 | /* Tag_ABI_FP_rounding. */ | |
17523 | if (flag_rounding_math) | |
17524 | asm_fprintf (asm_out_file, "\t.eabi_attribute 19, 1\n"); | |
17525 | if (!flag_unsafe_math_optimizations) | |
17526 | { | |
17527 | /* Tag_ABI_FP_denomal. */ | |
17528 | asm_fprintf (asm_out_file, "\t.eabi_attribute 20, 1\n"); | |
17529 | /* Tag_ABI_FP_exceptions. */ | |
17530 | asm_fprintf (asm_out_file, "\t.eabi_attribute 21, 1\n"); | |
17531 | } | |
17532 | /* Tag_ABI_FP_user_exceptions. */ | |
17533 | if (flag_signaling_nans) | |
17534 | asm_fprintf (asm_out_file, "\t.eabi_attribute 22, 1\n"); | |
17535 | /* Tag_ABI_FP_number_model. */ | |
17536 | asm_fprintf (asm_out_file, "\t.eabi_attribute 23, %d\n", | |
17537 | flag_finite_math_only ? 1 : 3); | |
17538 | ||
17539 | /* Tag_ABI_align8_needed. */ | |
17540 | asm_fprintf (asm_out_file, "\t.eabi_attribute 24, 1\n"); | |
17541 | /* Tag_ABI_align8_preserved. */ | |
17542 | asm_fprintf (asm_out_file, "\t.eabi_attribute 25, 1\n"); | |
17543 | /* Tag_ABI_enum_size. */ | |
17544 | asm_fprintf (asm_out_file, "\t.eabi_attribute 26, %d\n", | |
17545 | flag_short_enums ? 1 : 2); | |
17546 | ||
17547 | /* Tag_ABI_optimization_goals. */ | |
17548 | if (optimize_size) | |
17549 | val = 4; | |
17550 | else if (optimize >= 2) | |
17551 | val = 2; | |
17552 | else if (optimize) | |
17553 | val = 1; | |
17554 | else | |
17555 | val = 6; | |
17556 | asm_fprintf (asm_out_file, "\t.eabi_attribute 30, %d\n", val); | |
17557 | } | |
17558 | default_file_start(); | |
17559 | } | |
17560 | ||
afe27f3b | 17561 | static void |
17562 | arm_file_end (void) | |
17563 | { | |
17564 | int regno; | |
17565 | ||
9ffe2f5a | 17566 | if (NEED_INDICATE_EXEC_STACK) |
17567 | /* Add .note.GNU-stack. */ | |
17568 | file_end_indicate_exec_stack (); | |
17569 | ||
afe27f3b | 17570 | if (! thumb_call_reg_needed) |
17571 | return; | |
17572 | ||
2f14b1f9 | 17573 | switch_to_section (text_section); |
afe27f3b | 17574 | asm_fprintf (asm_out_file, "\t.code 16\n"); |
17575 | ASM_OUTPUT_ALIGN (asm_out_file, 1); | |
17576 | ||
a52972b8 | 17577 | for (regno = 0; regno < LR_REGNUM; regno++) |
afe27f3b | 17578 | { |
17579 | rtx label = thumb_call_via_label[regno]; | |
17580 | ||
17581 | if (label != 0) | |
17582 | { | |
17583 | targetm.asm_out.internal_label (asm_out_file, "L", | |
17584 | CODE_LABEL_NUMBER (label)); | |
17585 | asm_fprintf (asm_out_file, "\tbx\t%r\n", regno); | |
17586 | } | |
17587 | } | |
17588 | } | |
17589 | ||
1f3233d1 | 17590 | rtx aof_pic_label; |
17591 | ||
64266206 | 17592 | #ifdef AOF_ASSEMBLER |
66aacf41 | 17593 | /* Special functions only needed when producing AOF syntax assembler. */ |
64266206 | 17594 | |
69120d26 | 17595 | struct pic_chain |
17596 | { | |
bba10fb8 | 17597 | struct pic_chain * next; |
29122112 | 17598 | const char * symname; |
69120d26 | 17599 | }; |
17600 | ||
bba10fb8 | 17601 | static struct pic_chain * aof_pic_chain = NULL; |
69120d26 | 17602 | |
17603 | rtx | |
ebd88a36 | 17604 | aof_pic_entry (rtx x) |
69120d26 | 17605 | { |
bba10fb8 | 17606 | struct pic_chain ** chainp; |
69120d26 | 17607 | int offset; |
17608 | ||
17609 | if (aof_pic_label == NULL_RTX) | |
17610 | { | |
76676c8e | 17611 | aof_pic_label = gen_rtx_SYMBOL_REF (Pmode, "x$adcons"); |
69120d26 | 17612 | } |
17613 | ||
17614 | for (offset = 0, chainp = &aof_pic_chain; *chainp; | |
17615 | offset += 4, chainp = &(*chainp)->next) | |
17616 | if ((*chainp)->symname == XSTR (x, 0)) | |
17617 | return plus_constant (aof_pic_label, offset); | |
17618 | ||
17619 | *chainp = (struct pic_chain *) xmalloc (sizeof (struct pic_chain)); | |
17620 | (*chainp)->next = NULL; | |
17621 | (*chainp)->symname = XSTR (x, 0); | |
17622 | return plus_constant (aof_pic_label, offset); | |
17623 | } | |
17624 | ||
17625 | void | |
ebd88a36 | 17626 | aof_dump_pic_table (FILE *f) |
69120d26 | 17627 | { |
bba10fb8 | 17628 | struct pic_chain * chain; |
69120d26 | 17629 | |
17630 | if (aof_pic_chain == NULL) | |
17631 | return; | |
17632 | ||
caceeb43 | 17633 | asm_fprintf (f, "\tAREA |%r$$adcons|, BASED %r\n", |
17634 | PIC_OFFSET_TABLE_REGNUM, | |
17635 | PIC_OFFSET_TABLE_REGNUM); | |
69120d26 | 17636 | fputs ("|x$adcons|\n", f); |
9e7454d0 | 17637 | |
69120d26 | 17638 | for (chain = aof_pic_chain; chain; chain = chain->next) |
17639 | { | |
17640 | fputs ("\tDCD\t", f); | |
17641 | assemble_name (f, chain->symname); | |
17642 | fputs ("\n", f); | |
17643 | } | |
17644 | } | |
17645 | ||
64266206 | 17646 | int arm_text_section_count = 1; |
17647 | ||
2f14b1f9 | 17648 | /* A get_unnamed_section callback for switching to the text section. */ |
17649 | ||
17650 | static void | |
17651 | aof_output_text_section_asm_op (const void *data ATTRIBUTE_UNUSED) | |
64266206 | 17652 | { |
2f14b1f9 | 17653 | fprintf (asm_out_file, "\tAREA |C$$code%d|, CODE, READONLY", |
64266206 | 17654 | arm_text_section_count++); |
17655 | if (flag_pic) | |
2f14b1f9 | 17656 | fprintf (asm_out_file, ", PIC, REENTRANT"); |
17657 | fprintf (asm_out_file, "\n"); | |
64266206 | 17658 | } |
17659 | ||
17660 | static int arm_data_section_count = 1; | |
17661 | ||
2f14b1f9 | 17662 | /* A get_unnamed_section callback for switching to the data section. */ |
17663 | ||
17664 | static void | |
17665 | aof_output_data_section_asm_op (const void *data ATTRIBUTE_UNUSED) | |
17666 | { | |
17667 | fprintf (asm_out_file, "\tAREA |C$$data%d|, DATA\n", | |
17668 | arm_data_section_count++); | |
17669 | } | |
17670 | ||
17671 | /* Implement TARGET_ASM_INIT_SECTIONS. | |
17672 | ||
17673 | AOF Assembler syntax is a nightmare when it comes to areas, since once | |
17674 | we change from one area to another, we can't go back again. Instead, | |
17675 | we must create a new area with the same attributes and add the new output | |
17676 | to that. Unfortunately, there is nothing we can do here to guarantee that | |
17677 | two areas with the same attributes will be linked adjacently in the | |
17678 | resulting executable, so we have to be careful not to do pc-relative | |
17679 | addressing across such boundaries. */ | |
17680 | ||
17681 | static void | |
17682 | aof_asm_init_sections (void) | |
17683 | { | |
17684 | text_section = get_unnamed_section (SECTION_CODE, | |
17685 | aof_output_text_section_asm_op, NULL); | |
17686 | data_section = get_unnamed_section (SECTION_WRITE, | |
17687 | aof_output_data_section_asm_op, NULL); | |
17688 | readonly_data_section = text_section; | |
17689 | } | |
17690 | ||
17691 | void | |
17692 | zero_init_section (void) | |
64266206 | 17693 | { |
2f14b1f9 | 17694 | static int zero_init_count = 1; |
17695 | ||
17696 | fprintf (asm_out_file, "\tAREA |C$$zidata%d|,NOINIT\n", zero_init_count++); | |
17697 | in_section = NULL; | |
64266206 | 17698 | } |
17699 | ||
17700 | /* The AOF assembler is religiously strict about declarations of | |
17701 | imported and exported symbols, so that it is impossible to declare | |
ad87de1e | 17702 | a function as imported near the beginning of the file, and then to |
64266206 | 17703 | export it later on. It is, however, possible to delay the decision |
17704 | until all the functions in the file have been compiled. To get | |
17705 | around this, we maintain a list of the imports and exports, and | |
17706 | delete from it any that are subsequently defined. At the end of | |
17707 | compilation we spit the remainder of the list out before the END | |
17708 | directive. */ | |
17709 | ||
17710 | struct import | |
17711 | { | |
bba10fb8 | 17712 | struct import * next; |
29122112 | 17713 | const char * name; |
64266206 | 17714 | }; |
17715 | ||
bba10fb8 | 17716 | static struct import * imports_list = NULL; |
64266206 | 17717 | |
17718 | void | |
ebd88a36 | 17719 | aof_add_import (const char *name) |
64266206 | 17720 | { |
bba10fb8 | 17721 | struct import * new; |
64266206 | 17722 | |
17723 | for (new = imports_list; new; new = new->next) | |
17724 | if (new->name == name) | |
17725 | return; | |
17726 | ||
17727 | new = (struct import *) xmalloc (sizeof (struct import)); | |
17728 | new->next = imports_list; | |
17729 | imports_list = new; | |
17730 | new->name = name; | |
17731 | } | |
17732 | ||
17733 | void | |
ebd88a36 | 17734 | aof_delete_import (const char *name) |
64266206 | 17735 | { |
bba10fb8 | 17736 | struct import ** old; |
64266206 | 17737 | |
17738 | for (old = &imports_list; *old; old = & (*old)->next) | |
17739 | { | |
17740 | if ((*old)->name == name) | |
17741 | { | |
17742 | *old = (*old)->next; | |
17743 | return; | |
17744 | } | |
17745 | } | |
17746 | } | |
17747 | ||
17748 | int arm_main_function = 0; | |
17749 | ||
f6940372 | 17750 | static void |
ebd88a36 | 17751 | aof_dump_imports (FILE *f) |
64266206 | 17752 | { |
17753 | /* The AOF assembler needs this to cause the startup code to be extracted | |
17754 | from the library. Brining in __main causes the whole thing to work | |
17755 | automagically. */ | |
17756 | if (arm_main_function) | |
17757 | { | |
2f14b1f9 | 17758 | switch_to_section (text_section); |
64266206 | 17759 | fputs ("\tIMPORT __main\n", f); |
17760 | fputs ("\tDCD __main\n", f); | |
17761 | } | |
17762 | ||
17763 | /* Now dump the remaining imports. */ | |
17764 | while (imports_list) | |
17765 | { | |
17766 | fprintf (f, "\tIMPORT\t"); | |
17767 | assemble_name (f, imports_list->name); | |
17768 | fputc ('\n', f); | |
17769 | imports_list = imports_list->next; | |
17770 | } | |
17771 | } | |
67c1e638 | 17772 | |
17773 | static void | |
ebd88a36 | 17774 | aof_globalize_label (FILE *stream, const char *name) |
67c1e638 | 17775 | { |
17776 | default_globalize_label (stream, name); | |
17777 | if (! strcmp (name, "main")) | |
17778 | arm_main_function = 1; | |
17779 | } | |
f6940372 | 17780 | |
92c473b8 | 17781 | static void |
3285410a | 17782 | aof_file_start (void) |
92c473b8 | 17783 | { |
17784 | fputs ("__r0\tRN\t0\n", asm_out_file); | |
17785 | fputs ("__a1\tRN\t0\n", asm_out_file); | |
17786 | fputs ("__a2\tRN\t1\n", asm_out_file); | |
17787 | fputs ("__a3\tRN\t2\n", asm_out_file); | |
17788 | fputs ("__a4\tRN\t3\n", asm_out_file); | |
17789 | fputs ("__v1\tRN\t4\n", asm_out_file); | |
17790 | fputs ("__v2\tRN\t5\n", asm_out_file); | |
17791 | fputs ("__v3\tRN\t6\n", asm_out_file); | |
17792 | fputs ("__v4\tRN\t7\n", asm_out_file); | |
17793 | fputs ("__v5\tRN\t8\n", asm_out_file); | |
17794 | fputs ("__v6\tRN\t9\n", asm_out_file); | |
17795 | fputs ("__sl\tRN\t10\n", asm_out_file); | |
17796 | fputs ("__fp\tRN\t11\n", asm_out_file); | |
17797 | fputs ("__ip\tRN\t12\n", asm_out_file); | |
17798 | fputs ("__sp\tRN\t13\n", asm_out_file); | |
17799 | fputs ("__lr\tRN\t14\n", asm_out_file); | |
17800 | fputs ("__pc\tRN\t15\n", asm_out_file); | |
17801 | fputs ("__f0\tFN\t0\n", asm_out_file); | |
17802 | fputs ("__f1\tFN\t1\n", asm_out_file); | |
17803 | fputs ("__f2\tFN\t2\n", asm_out_file); | |
17804 | fputs ("__f3\tFN\t3\n", asm_out_file); | |
17805 | fputs ("__f4\tFN\t4\n", asm_out_file); | |
17806 | fputs ("__f5\tFN\t5\n", asm_out_file); | |
17807 | fputs ("__f6\tFN\t6\n", asm_out_file); | |
17808 | fputs ("__f7\tFN\t7\n", asm_out_file); | |
2f14b1f9 | 17809 | switch_to_section (text_section); |
92c473b8 | 17810 | } |
17811 | ||
f6940372 | 17812 | static void |
ebd88a36 | 17813 | aof_file_end (void) |
f6940372 | 17814 | { |
17815 | if (flag_pic) | |
17816 | aof_dump_pic_table (asm_out_file); | |
afe27f3b | 17817 | arm_file_end (); |
f6940372 | 17818 | aof_dump_imports (asm_out_file); |
17819 | fputs ("\tEND\n", asm_out_file); | |
17820 | } | |
64266206 | 17821 | #endif /* AOF_ASSEMBLER */ |
2cb4ac60 | 17822 | |
7811991d | 17823 | #ifndef ARM_PE |
17824 | /* Symbols in the text segment can be accessed without indirecting via the | |
17825 | constant pool; it may take an extra binary operation, but this is still | |
17826 | faster than indirecting via memory. Don't do this when not optimizing, | |
17827 | since we won't be calculating al of the offsets necessary to do this | |
17828 | simplification. */ | |
17829 | ||
17830 | static void | |
ebd88a36 | 17831 | arm_encode_section_info (tree decl, rtx rtl, int first) |
7811991d | 17832 | { |
17833 | /* This doesn't work with AOF syntax, since the string table may be in | |
17834 | a different AREA. */ | |
17835 | #ifndef AOF_ASSEMBLER | |
a814bad5 | 17836 | if (optimize > 0 && TREE_CONSTANT (decl)) |
2c129d70 | 17837 | SYMBOL_REF_FLAG (XEXP (rtl, 0)) = 1; |
7811991d | 17838 | #endif |
17839 | ||
f655717d | 17840 | default_encode_section_info (decl, rtl, first); |
7811991d | 17841 | } |
17842 | #endif /* !ARM_PE */ | |
c6933ba6 | 17843 | |
805e22b2 | 17844 | static void |
ebd88a36 | 17845 | arm_internal_label (FILE *stream, const char *prefix, unsigned long labelno) |
805e22b2 | 17846 | { |
17847 | if (arm_ccfsm_state == 3 && (unsigned) arm_target_label == labelno | |
17848 | && !strcmp (prefix, "L")) | |
17849 | { | |
17850 | arm_ccfsm_state = 0; | |
17851 | arm_target_insn = NULL; | |
17852 | } | |
17853 | default_internal_label (stream, prefix, labelno); | |
17854 | } | |
17855 | ||
6988553d | 17856 | /* Output code to add DELTA to the first argument, and then jump |
17857 | to FUNCTION. Used for C++ multiple inheritance. */ | |
6988553d | 17858 | static void |
ebd88a36 | 17859 | arm_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED, |
17860 | HOST_WIDE_INT delta, | |
17861 | HOST_WIDE_INT vcall_offset ATTRIBUTE_UNUSED, | |
17862 | tree function) | |
c6933ba6 | 17863 | { |
a2cd141b | 17864 | static int thunk_label = 0; |
17865 | char label[256]; | |
ca1ede88 | 17866 | char labelpc[256]; |
c6933ba6 | 17867 | int mi_delta = delta; |
17868 | const char *const mi_op = mi_delta < 0 ? "sub" : "add"; | |
17869 | int shift = 0; | |
45550790 | 17870 | int this_regno = (aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function) |
c6933ba6 | 17871 | ? 1 : 0); |
17872 | if (mi_delta < 0) | |
17873 | mi_delta = - mi_delta; | |
25f905c2 | 17874 | /* When generating 16-bit thumb code, thunks are entered in arm mode. */ |
17875 | if (TARGET_THUMB1) | |
a2cd141b | 17876 | { |
17877 | int labelno = thunk_label++; | |
17878 | ASM_GENERATE_INTERNAL_LABEL (label, "LTHUMBFUNC", labelno); | |
17879 | fputs ("\tldr\tr12, ", file); | |
17880 | assemble_name (file, label); | |
17881 | fputc ('\n', file); | |
ca1ede88 | 17882 | if (flag_pic) |
17883 | { | |
17884 | /* If we are generating PIC, the ldr instruction below loads | |
17885 | "(target - 7) - .LTHUNKPCn" into r12. The pc reads as | |
17886 | the address of the add + 8, so we have: | |
17887 | ||
17888 | r12 = (target - 7) - .LTHUNKPCn + (.LTHUNKPCn + 8) | |
17889 | = target + 1. | |
17890 | ||
17891 | Note that we have "+ 1" because some versions of GNU ld | |
17892 | don't set the low bit of the result for R_ARM_REL32 | |
17893 | relocations against thumb function symbols. */ | |
17894 | ASM_GENERATE_INTERNAL_LABEL (labelpc, "LTHUNKPC", labelno); | |
17895 | assemble_name (file, labelpc); | |
17896 | fputs (":\n", file); | |
17897 | fputs ("\tadd\tr12, pc, r12\n", file); | |
17898 | } | |
a2cd141b | 17899 | } |
25f905c2 | 17900 | /* TODO: Use movw/movt for large constants when available. */ |
c6933ba6 | 17901 | while (mi_delta != 0) |
17902 | { | |
17903 | if ((mi_delta & (3 << shift)) == 0) | |
17904 | shift += 2; | |
17905 | else | |
17906 | { | |
17907 | asm_fprintf (file, "\t%s\t%r, %r, #%d\n", | |
17908 | mi_op, this_regno, this_regno, | |
17909 | mi_delta & (0xff << shift)); | |
17910 | mi_delta &= ~(0xff << shift); | |
17911 | shift += 8; | |
17912 | } | |
17913 | } | |
25f905c2 | 17914 | if (TARGET_THUMB1) |
a2cd141b | 17915 | { |
17916 | fprintf (file, "\tbx\tr12\n"); | |
17917 | ASM_OUTPUT_ALIGN (file, 2); | |
17918 | assemble_name (file, label); | |
17919 | fputs (":\n", file); | |
ca1ede88 | 17920 | if (flag_pic) |
17921 | { | |
17922 | /* Output ".word .LTHUNKn-7-.LTHUNKPCn". */ | |
17923 | rtx tem = XEXP (DECL_RTL (function), 0); | |
17924 | tem = gen_rtx_PLUS (GET_MODE (tem), tem, GEN_INT (-7)); | |
17925 | tem = gen_rtx_MINUS (GET_MODE (tem), | |
17926 | tem, | |
17927 | gen_rtx_SYMBOL_REF (Pmode, | |
17928 | ggc_strdup (labelpc))); | |
17929 | assemble_integer (tem, 4, BITS_PER_WORD, 1); | |
17930 | } | |
17931 | else | |
17932 | /* Output ".word .LTHUNKn". */ | |
17933 | assemble_integer (XEXP (DECL_RTL (function), 0), 4, BITS_PER_WORD, 1); | |
a2cd141b | 17934 | } |
17935 | else | |
17936 | { | |
17937 | fputs ("\tb\t", file); | |
17938 | assemble_name (file, XSTR (XEXP (DECL_RTL (function), 0), 0)); | |
17939 | if (NEED_PLT_RELOC) | |
17940 | fputs ("(PLT)", file); | |
17941 | fputc ('\n', file); | |
17942 | } | |
c6933ba6 | 17943 | } |
755eb2b4 | 17944 | |
17945 | int | |
068c4bec | 17946 | arm_emit_vector_const (FILE *file, rtx x) |
755eb2b4 | 17947 | { |
17948 | int i; | |
17949 | const char * pattern; | |
17950 | ||
ed29c566 | 17951 | gcc_assert (GET_CODE (x) == CONST_VECTOR); |
755eb2b4 | 17952 | |
17953 | switch (GET_MODE (x)) | |
17954 | { | |
17955 | case V2SImode: pattern = "%08x"; break; | |
17956 | case V4HImode: pattern = "%04x"; break; | |
17957 | case V8QImode: pattern = "%02x"; break; | |
ed29c566 | 17958 | default: gcc_unreachable (); |
755eb2b4 | 17959 | } |
17960 | ||
17961 | fprintf (file, "0x"); | |
17962 | for (i = CONST_VECTOR_NUNITS (x); i--;) | |
17963 | { | |
17964 | rtx element; | |
17965 | ||
17966 | element = CONST_VECTOR_ELT (x, i); | |
17967 | fprintf (file, pattern, INTVAL (element)); | |
17968 | } | |
17969 | ||
17970 | return 1; | |
17971 | } | |
17972 | ||
17973 | const char * | |
068c4bec | 17974 | arm_output_load_gr (rtx *operands) |
755eb2b4 | 17975 | { |
17976 | rtx reg; | |
17977 | rtx offset; | |
17978 | rtx wcgr; | |
17979 | rtx sum; | |
9e7454d0 | 17980 | |
755eb2b4 | 17981 | if (GET_CODE (operands [1]) != MEM |
17982 | || GET_CODE (sum = XEXP (operands [1], 0)) != PLUS | |
17983 | || GET_CODE (reg = XEXP (sum, 0)) != REG | |
17984 | || GET_CODE (offset = XEXP (sum, 1)) != CONST_INT | |
17985 | || ((INTVAL (offset) < 1024) && (INTVAL (offset) > -1024))) | |
17986 | return "wldrw%?\t%0, %1"; | |
9e7454d0 | 17987 | |
17988 | /* Fix up an out-of-range load of a GR register. */ | |
755eb2b4 | 17989 | output_asm_insn ("str%?\t%0, [sp, #-4]!\t@ Start of GR load expansion", & reg); |
17990 | wcgr = operands[0]; | |
17991 | operands[0] = reg; | |
17992 | output_asm_insn ("ldr%?\t%0, %1", operands); | |
17993 | ||
17994 | operands[0] = wcgr; | |
17995 | operands[1] = reg; | |
17996 | output_asm_insn ("tmcr%?\t%0, %1", operands); | |
17997 | output_asm_insn ("ldr%?\t%0, [sp], #4\t@ End of GR load expansion", & reg); | |
17998 | ||
17999 | return ""; | |
18000 | } | |
f2eb2b4d | 18001 | |
4c66acf1 | 18002 | /* Worker function for TARGET_SETUP_INCOMING_VARARGS. |
18003 | ||
18004 | On the ARM, PRETEND_SIZE is set in order to have the prologue push the last | |
18005 | named arg and all anonymous args onto the stack. | |
18006 | XXX I know the prologue shouldn't be pushing registers, but it is faster | |
18007 | that way. */ | |
18008 | ||
18009 | static void | |
18010 | arm_setup_incoming_varargs (CUMULATIVE_ARGS *cum, | |
18011 | enum machine_mode mode ATTRIBUTE_UNUSED, | |
18012 | tree type ATTRIBUTE_UNUSED, | |
18013 | int *pretend_size, | |
18014 | int second_time ATTRIBUTE_UNUSED) | |
18015 | { | |
18016 | cfun->machine->uses_anonymous_args = 1; | |
18017 | if (cum->nregs < NUM_ARG_REGS) | |
18018 | *pretend_size = (NUM_ARG_REGS - cum->nregs) * UNITS_PER_WORD; | |
18019 | } | |
a2cd141b | 18020 | |
58552ed0 | 18021 | /* Return nonzero if the CONSUMER instruction (a store) does not need |
a2cd141b | 18022 | PRODUCER's value to calculate the address. */ |
18023 | ||
18024 | int | |
18025 | arm_no_early_store_addr_dep (rtx producer, rtx consumer) | |
18026 | { | |
18027 | rtx value = PATTERN (producer); | |
18028 | rtx addr = PATTERN (consumer); | |
18029 | ||
18030 | if (GET_CODE (value) == COND_EXEC) | |
18031 | value = COND_EXEC_CODE (value); | |
18032 | if (GET_CODE (value) == PARALLEL) | |
18033 | value = XVECEXP (value, 0, 0); | |
18034 | value = XEXP (value, 0); | |
18035 | if (GET_CODE (addr) == COND_EXEC) | |
18036 | addr = COND_EXEC_CODE (addr); | |
18037 | if (GET_CODE (addr) == PARALLEL) | |
18038 | addr = XVECEXP (addr, 0, 0); | |
18039 | addr = XEXP (addr, 0); | |
9e7454d0 | 18040 | |
a2cd141b | 18041 | return !reg_overlap_mentioned_p (value, addr); |
18042 | } | |
18043 | ||
58552ed0 | 18044 | /* Return nonzero if the CONSUMER instruction (an ALU op) does not |
a2cd141b | 18045 | have an early register shift value or amount dependency on the |
18046 | result of PRODUCER. */ | |
18047 | ||
18048 | int | |
18049 | arm_no_early_alu_shift_dep (rtx producer, rtx consumer) | |
18050 | { | |
18051 | rtx value = PATTERN (producer); | |
18052 | rtx op = PATTERN (consumer); | |
18053 | rtx early_op; | |
18054 | ||
18055 | if (GET_CODE (value) == COND_EXEC) | |
18056 | value = COND_EXEC_CODE (value); | |
18057 | if (GET_CODE (value) == PARALLEL) | |
18058 | value = XVECEXP (value, 0, 0); | |
18059 | value = XEXP (value, 0); | |
18060 | if (GET_CODE (op) == COND_EXEC) | |
18061 | op = COND_EXEC_CODE (op); | |
18062 | if (GET_CODE (op) == PARALLEL) | |
18063 | op = XVECEXP (op, 0, 0); | |
18064 | op = XEXP (op, 1); | |
9e7454d0 | 18065 | |
a2cd141b | 18066 | early_op = XEXP (op, 0); |
18067 | /* This is either an actual independent shift, or a shift applied to | |
18068 | the first operand of another operation. We want the whole shift | |
18069 | operation. */ | |
18070 | if (GET_CODE (early_op) == REG) | |
18071 | early_op = op; | |
18072 | ||
18073 | return !reg_overlap_mentioned_p (value, early_op); | |
18074 | } | |
18075 | ||
58552ed0 | 18076 | /* Return nonzero if the CONSUMER instruction (an ALU op) does not |
a2cd141b | 18077 | have an early register shift value dependency on the result of |
18078 | PRODUCER. */ | |
18079 | ||
18080 | int | |
18081 | arm_no_early_alu_shift_value_dep (rtx producer, rtx consumer) | |
18082 | { | |
18083 | rtx value = PATTERN (producer); | |
18084 | rtx op = PATTERN (consumer); | |
18085 | rtx early_op; | |
18086 | ||
18087 | if (GET_CODE (value) == COND_EXEC) | |
18088 | value = COND_EXEC_CODE (value); | |
18089 | if (GET_CODE (value) == PARALLEL) | |
18090 | value = XVECEXP (value, 0, 0); | |
18091 | value = XEXP (value, 0); | |
18092 | if (GET_CODE (op) == COND_EXEC) | |
18093 | op = COND_EXEC_CODE (op); | |
18094 | if (GET_CODE (op) == PARALLEL) | |
18095 | op = XVECEXP (op, 0, 0); | |
18096 | op = XEXP (op, 1); | |
9e7454d0 | 18097 | |
a2cd141b | 18098 | early_op = XEXP (op, 0); |
18099 | ||
18100 | /* This is either an actual independent shift, or a shift applied to | |
18101 | the first operand of another operation. We want the value being | |
18102 | shifted, in either case. */ | |
18103 | if (GET_CODE (early_op) != REG) | |
18104 | early_op = XEXP (early_op, 0); | |
9e7454d0 | 18105 | |
a2cd141b | 18106 | return !reg_overlap_mentioned_p (value, early_op); |
18107 | } | |
18108 | ||
58552ed0 | 18109 | /* Return nonzero if the CONSUMER (a mul or mac op) does not |
a2cd141b | 18110 | have an early register mult dependency on the result of |
18111 | PRODUCER. */ | |
18112 | ||
18113 | int | |
18114 | arm_no_early_mul_dep (rtx producer, rtx consumer) | |
18115 | { | |
18116 | rtx value = PATTERN (producer); | |
18117 | rtx op = PATTERN (consumer); | |
18118 | ||
18119 | if (GET_CODE (value) == COND_EXEC) | |
18120 | value = COND_EXEC_CODE (value); | |
18121 | if (GET_CODE (value) == PARALLEL) | |
18122 | value = XVECEXP (value, 0, 0); | |
18123 | value = XEXP (value, 0); | |
18124 | if (GET_CODE (op) == COND_EXEC) | |
18125 | op = COND_EXEC_CODE (op); | |
18126 | if (GET_CODE (op) == PARALLEL) | |
18127 | op = XVECEXP (op, 0, 0); | |
18128 | op = XEXP (op, 1); | |
9e7454d0 | 18129 | |
a2cd141b | 18130 | return (GET_CODE (op) == PLUS |
18131 | && !reg_overlap_mentioned_p (value, XEXP (op, 0))); | |
18132 | } | |
18133 | ||
72f1246c | 18134 | /* We can't rely on the caller doing the proper promotion when |
18135 | using APCS or ATPCS. */ | |
18136 | ||
18137 | static bool | |
18138 | arm_promote_prototypes (tree t ATTRIBUTE_UNUSED) | |
18139 | { | |
50459f30 | 18140 | return !TARGET_AAPCS_BASED; |
72f1246c | 18141 | } |
18142 | ||
7faec15f | 18143 | |
18144 | /* AAPCS based ABIs use short enums by default. */ | |
18145 | ||
18146 | static bool | |
18147 | arm_default_short_enums (void) | |
18148 | { | |
6783d878 | 18149 | return TARGET_AAPCS_BASED && arm_abi != ARM_ABI_AAPCS_LINUX; |
7faec15f | 18150 | } |
25ba5be6 | 18151 | |
18152 | ||
18153 | /* AAPCS requires that anonymous bitfields affect structure alignment. */ | |
18154 | ||
18155 | static bool | |
18156 | arm_align_anon_bitfield (void) | |
18157 | { | |
18158 | return TARGET_AAPCS_BASED; | |
18159 | } | |
c1dc02de | 18160 | |
18161 | ||
18162 | /* The generic C++ ABI says 64-bit (long long). The EABI says 32-bit. */ | |
18163 | ||
18164 | static tree | |
18165 | arm_cxx_guard_type (void) | |
18166 | { | |
18167 | return TARGET_AAPCS_BASED ? integer_type_node : long_long_integer_type_node; | |
18168 | } | |
18169 | ||
bcaec148 | 18170 | /* Return non-zero if the consumer (a multiply-accumulate instruction) |
18171 | has an accumulator dependency on the result of the producer (a | |
18172 | multiplication instruction) and no other dependency on that result. */ | |
18173 | int | |
18174 | arm_mac_accumulator_is_mul_result (rtx producer, rtx consumer) | |
18175 | { | |
18176 | rtx mul = PATTERN (producer); | |
18177 | rtx mac = PATTERN (consumer); | |
18178 | rtx mul_result; | |
18179 | rtx mac_op0, mac_op1, mac_acc; | |
18180 | ||
18181 | if (GET_CODE (mul) == COND_EXEC) | |
18182 | mul = COND_EXEC_CODE (mul); | |
18183 | if (GET_CODE (mac) == COND_EXEC) | |
18184 | mac = COND_EXEC_CODE (mac); | |
18185 | ||
18186 | /* Check that mul is of the form (set (...) (mult ...)) | |
18187 | and mla is of the form (set (...) (plus (mult ...) (...))). */ | |
18188 | if ((GET_CODE (mul) != SET || GET_CODE (XEXP (mul, 1)) != MULT) | |
18189 | || (GET_CODE (mac) != SET || GET_CODE (XEXP (mac, 1)) != PLUS | |
18190 | || GET_CODE (XEXP (XEXP (mac, 1), 0)) != MULT)) | |
18191 | return 0; | |
18192 | ||
18193 | mul_result = XEXP (mul, 0); | |
18194 | mac_op0 = XEXP (XEXP (XEXP (mac, 1), 0), 0); | |
18195 | mac_op1 = XEXP (XEXP (XEXP (mac, 1), 0), 1); | |
18196 | mac_acc = XEXP (XEXP (mac, 1), 1); | |
18197 | ||
18198 | return (reg_overlap_mentioned_p (mul_result, mac_acc) | |
18199 | && !reg_overlap_mentioned_p (mul_result, mac_op0) | |
18200 | && !reg_overlap_mentioned_p (mul_result, mac_op1)); | |
18201 | } | |
18202 | ||
c1dc02de | 18203 | |
442e3cb9 | 18204 | /* The EABI says test the least significant bit of a guard variable. */ |
c1dc02de | 18205 | |
18206 | static bool | |
18207 | arm_cxx_guard_mask_bit (void) | |
18208 | { | |
18209 | return TARGET_AAPCS_BASED; | |
18210 | } | |
600f4be7 | 18211 | |
18212 | ||
18213 | /* The EABI specifies that all array cookies are 8 bytes long. */ | |
18214 | ||
18215 | static tree | |
18216 | arm_get_cookie_size (tree type) | |
18217 | { | |
18218 | tree size; | |
18219 | ||
18220 | if (!TARGET_AAPCS_BASED) | |
18221 | return default_cxx_get_cookie_size (type); | |
18222 | ||
7016c612 | 18223 | size = build_int_cst (sizetype, 8); |
600f4be7 | 18224 | return size; |
18225 | } | |
18226 | ||
18227 | ||
18228 | /* The EABI says that array cookies should also contain the element size. */ | |
18229 | ||
18230 | static bool | |
18231 | arm_cookie_has_size (void) | |
18232 | { | |
18233 | return TARGET_AAPCS_BASED; | |
18234 | } | |
853b7640 | 18235 | |
18236 | ||
18237 | /* The EABI says constructors and destructors should return a pointer to | |
18238 | the object constructed/destroyed. */ | |
18239 | ||
18240 | static bool | |
18241 | arm_cxx_cdtor_returns_this (void) | |
18242 | { | |
18243 | return TARGET_AAPCS_BASED; | |
18244 | } | |
4c44712e | 18245 | |
4fc9a8ec | 18246 | /* The EABI says that an inline function may never be the key |
18247 | method. */ | |
18248 | ||
18249 | static bool | |
18250 | arm_cxx_key_method_may_be_inline (void) | |
18251 | { | |
18252 | return !TARGET_AAPCS_BASED; | |
18253 | } | |
18254 | ||
7908506d | 18255 | static void |
18256 | arm_cxx_determine_class_data_visibility (tree decl) | |
18257 | { | |
18258 | if (!TARGET_AAPCS_BASED) | |
18259 | return; | |
4fc9a8ec | 18260 | |
7908506d | 18261 | /* In general, \S 3.2.5.5 of the ARM EABI requires that class data |
18262 | is exported. However, on systems without dynamic vague linkage, | |
18263 | \S 3.2.5.6 says that COMDAT class data has hidden linkage. */ | |
18264 | if (!TARGET_ARM_DYNAMIC_VAGUE_LINKAGE_P && DECL_COMDAT (decl)) | |
18265 | DECL_VISIBILITY (decl) = VISIBILITY_HIDDEN; | |
18266 | else | |
18267 | DECL_VISIBILITY (decl) = VISIBILITY_DEFAULT; | |
18268 | DECL_VISIBILITY_SPECIFIED (decl) = 1; | |
18269 | } | |
a60f1840 | 18270 | |
4fc9a8ec | 18271 | static bool |
7908506d | 18272 | arm_cxx_class_data_always_comdat (void) |
4fc9a8ec | 18273 | { |
7908506d | 18274 | /* \S 3.2.5.4 of the ARM C++ ABI says that class data only have |
18275 | vague linkage if the class has no key function. */ | |
18276 | return !TARGET_AAPCS_BASED; | |
4fc9a8ec | 18277 | } |
4c44712e | 18278 | |
46b9ff5e | 18279 | |
18280 | /* The EABI says __aeabi_atexit should be used to register static | |
18281 | destructors. */ | |
18282 | ||
18283 | static bool | |
18284 | arm_cxx_use_aeabi_atexit (void) | |
18285 | { | |
18286 | return TARGET_AAPCS_BASED; | |
18287 | } | |
18288 | ||
18289 | ||
4c44712e | 18290 | void |
18291 | arm_set_return_address (rtx source, rtx scratch) | |
18292 | { | |
18293 | arm_stack_offsets *offsets; | |
18294 | HOST_WIDE_INT delta; | |
18295 | rtx addr; | |
18296 | unsigned long saved_regs; | |
18297 | ||
18298 | saved_regs = arm_compute_save_reg_mask (); | |
18299 | ||
18300 | if ((saved_regs & (1 << LR_REGNUM)) == 0) | |
18301 | emit_move_insn (gen_rtx_REG (Pmode, LR_REGNUM), source); | |
18302 | else | |
18303 | { | |
18304 | if (frame_pointer_needed) | |
18305 | addr = plus_constant(hard_frame_pointer_rtx, -4); | |
18306 | else | |
18307 | { | |
18308 | /* LR will be the first saved register. */ | |
18309 | offsets = arm_get_frame_offsets (); | |
18310 | delta = offsets->outgoing_args - (offsets->frame + 4); | |
18311 | ||
9e7454d0 | 18312 | |
4c44712e | 18313 | if (delta >= 4096) |
18314 | { | |
18315 | emit_insn (gen_addsi3 (scratch, stack_pointer_rtx, | |
18316 | GEN_INT (delta & ~4095))); | |
18317 | addr = scratch; | |
18318 | delta &= 4095; | |
18319 | } | |
18320 | else | |
18321 | addr = stack_pointer_rtx; | |
18322 | ||
18323 | addr = plus_constant (addr, delta); | |
18324 | } | |
788fcce0 | 18325 | emit_move_insn (gen_frame_mem (Pmode, addr), source); |
4c44712e | 18326 | } |
18327 | } | |
18328 | ||
18329 | ||
18330 | void | |
18331 | thumb_set_return_address (rtx source, rtx scratch) | |
18332 | { | |
18333 | arm_stack_offsets *offsets; | |
4c44712e | 18334 | HOST_WIDE_INT delta; |
25f905c2 | 18335 | HOST_WIDE_INT limit; |
4c44712e | 18336 | int reg; |
18337 | rtx addr; | |
237533cc | 18338 | unsigned long mask; |
4c44712e | 18339 | |
18340 | emit_insn (gen_rtx_USE (VOIDmode, source)); | |
4c44712e | 18341 | |
25f905c2 | 18342 | mask = thumb1_compute_save_reg_mask (); |
237533cc | 18343 | if (mask & (1 << LR_REGNUM)) |
4c44712e | 18344 | { |
18345 | offsets = arm_get_frame_offsets (); | |
18346 | ||
25f905c2 | 18347 | limit = 1024; |
4c44712e | 18348 | /* Find the saved regs. */ |
18349 | if (frame_pointer_needed) | |
18350 | { | |
18351 | delta = offsets->soft_frame - offsets->saved_args; | |
18352 | reg = THUMB_HARD_FRAME_POINTER_REGNUM; | |
25f905c2 | 18353 | if (TARGET_THUMB1) |
18354 | limit = 128; | |
4c44712e | 18355 | } |
18356 | else | |
18357 | { | |
18358 | delta = offsets->outgoing_args - offsets->saved_args; | |
18359 | reg = SP_REGNUM; | |
18360 | } | |
18361 | /* Allow for the stack frame. */ | |
25f905c2 | 18362 | if (TARGET_THUMB1 && TARGET_BACKTRACE) |
4c44712e | 18363 | delta -= 16; |
18364 | /* The link register is always the first saved register. */ | |
18365 | delta -= 4; | |
9e7454d0 | 18366 | |
4c44712e | 18367 | /* Construct the address. */ |
18368 | addr = gen_rtx_REG (SImode, reg); | |
25f905c2 | 18369 | if (delta > limit) |
4c44712e | 18370 | { |
18371 | emit_insn (gen_movsi (scratch, GEN_INT (delta))); | |
18372 | emit_insn (gen_addsi3 (scratch, scratch, stack_pointer_rtx)); | |
18373 | addr = scratch; | |
18374 | } | |
18375 | else | |
18376 | addr = plus_constant (addr, delta); | |
18377 | ||
788fcce0 | 18378 | emit_move_insn (gen_frame_mem (Pmode, addr), source); |
4c44712e | 18379 | } |
18380 | else | |
18381 | emit_move_insn (gen_rtx_REG (Pmode, LR_REGNUM), source); | |
18382 | } | |
18383 | ||
9e7454d0 | 18384 | /* Implements target hook vector_mode_supported_p. */ |
18385 | bool | |
18386 | arm_vector_mode_supported_p (enum machine_mode mode) | |
18387 | { | |
d98a3884 | 18388 | /* Neon also supports V2SImode, etc. listed in the clause below. */ |
18389 | if (TARGET_NEON && (mode == V2SFmode || mode == V4SImode || mode == V8HImode | |
18390 | || mode == V16QImode || mode == V4SFmode || mode == V2DImode)) | |
18391 | return true; | |
18392 | ||
9e7454d0 | 18393 | if ((mode == V2SImode) |
18394 | || (mode == V4HImode) | |
18395 | || (mode == V8QImode)) | |
18396 | return true; | |
18397 | ||
18398 | return false; | |
18399 | } | |
c49547c4 | 18400 | |
18401 | /* Implement TARGET_SHIFT_TRUNCATION_MASK. SImode shifts use normal | |
18402 | ARM insns and therefore guarantee that the shift count is modulo 256. | |
18403 | DImode shifts (those implemented by lib1funcs.asm or by optabs.c) | |
18404 | guarantee no particular behavior for out-of-range counts. */ | |
18405 | ||
18406 | static unsigned HOST_WIDE_INT | |
18407 | arm_shift_truncation_mask (enum machine_mode mode) | |
18408 | { | |
18409 | return mode == SImode ? 255 : 0; | |
18410 | } | |
5f9d1097 | 18411 | |
18412 | ||
18413 | /* Map internal gcc register numbers to DWARF2 register numbers. */ | |
18414 | ||
18415 | unsigned int | |
18416 | arm_dbx_register_number (unsigned int regno) | |
18417 | { | |
18418 | if (regno < 16) | |
18419 | return regno; | |
18420 | ||
18421 | /* TODO: Legacy targets output FPA regs as registers 16-23 for backwards | |
18422 | compatibility. The EABI defines them as registers 96-103. */ | |
18423 | if (IS_FPA_REGNUM (regno)) | |
18424 | return (TARGET_AAPCS_BASED ? 96 : 16) + regno - FIRST_FPA_REGNUM; | |
18425 | ||
b5a0636d | 18426 | /* FIXME: VFPv3 register numbering. */ |
5f9d1097 | 18427 | if (IS_VFP_REGNUM (regno)) |
18428 | return 64 + regno - FIRST_VFP_REGNUM; | |
18429 | ||
18430 | if (IS_IWMMXT_GR_REGNUM (regno)) | |
18431 | return 104 + regno - FIRST_IWMMXT_GR_REGNUM; | |
18432 | ||
18433 | if (IS_IWMMXT_REGNUM (regno)) | |
18434 | return 112 + regno - FIRST_IWMMXT_REGNUM; | |
18435 | ||
ed29c566 | 18436 | gcc_unreachable (); |
5f9d1097 | 18437 | } |
18438 | ||
1774763d | 18439 | |
18440 | #ifdef TARGET_UNWIND_INFO | |
25f905c2 | 18441 | /* Emit unwind directives for a store-multiple instruction or stack pointer |
18442 | push during alignment. | |
18443 | These should only ever be generated by the function prologue code, so | |
18444 | expect them to have a particular form. */ | |
1774763d | 18445 | |
18446 | static void | |
25f905c2 | 18447 | arm_unwind_emit_sequence (FILE * asm_out_file, rtx p) |
1774763d | 18448 | { |
18449 | int i; | |
18450 | HOST_WIDE_INT offset; | |
18451 | HOST_WIDE_INT nregs; | |
18452 | int reg_size; | |
18453 | unsigned reg; | |
18454 | unsigned lastreg; | |
18455 | rtx e; | |
18456 | ||
1774763d | 18457 | e = XVECEXP (p, 0, 0); |
25f905c2 | 18458 | if (GET_CODE (e) != SET) |
18459 | abort (); | |
18460 | ||
18461 | /* First insn will adjust the stack pointer. */ | |
1774763d | 18462 | if (GET_CODE (e) != SET |
18463 | || GET_CODE (XEXP (e, 0)) != REG | |
18464 | || REGNO (XEXP (e, 0)) != SP_REGNUM | |
18465 | || GET_CODE (XEXP (e, 1)) != PLUS) | |
18466 | abort (); | |
18467 | ||
18468 | offset = -INTVAL (XEXP (XEXP (e, 1), 1)); | |
18469 | nregs = XVECLEN (p, 0) - 1; | |
18470 | ||
18471 | reg = REGNO (XEXP (XVECEXP (p, 0, 1), 1)); | |
18472 | if (reg < 16) | |
18473 | { | |
18474 | /* The function prologue may also push pc, but not annotate it as it is | |
77aa6362 | 18475 | never restored. We turn this into a stack pointer adjustment. */ |
1774763d | 18476 | if (nregs * 4 == offset - 4) |
18477 | { | |
18478 | fprintf (asm_out_file, "\t.pad #4\n"); | |
18479 | offset -= 4; | |
18480 | } | |
18481 | reg_size = 4; | |
382a2a57 | 18482 | fprintf (asm_out_file, "\t.save {"); |
1774763d | 18483 | } |
18484 | else if (IS_VFP_REGNUM (reg)) | |
18485 | { | |
1774763d | 18486 | reg_size = 8; |
382a2a57 | 18487 | fprintf (asm_out_file, "\t.vsave {"); |
1774763d | 18488 | } |
18489 | else if (reg >= FIRST_FPA_REGNUM && reg <= LAST_FPA_REGNUM) | |
18490 | { | |
18491 | /* FPA registers are done differently. */ | |
7781aa77 | 18492 | asm_fprintf (asm_out_file, "\t.save %r, %wd\n", reg, nregs); |
1774763d | 18493 | return; |
18494 | } | |
18495 | else | |
18496 | /* Unknown register type. */ | |
18497 | abort (); | |
18498 | ||
18499 | /* If the stack increment doesn't match the size of the saved registers, | |
18500 | something has gone horribly wrong. */ | |
18501 | if (offset != nregs * reg_size) | |
18502 | abort (); | |
18503 | ||
1774763d | 18504 | offset = 0; |
18505 | lastreg = 0; | |
18506 | /* The remaining insns will describe the stores. */ | |
18507 | for (i = 1; i <= nregs; i++) | |
18508 | { | |
18509 | /* Expect (set (mem <addr>) (reg)). | |
18510 | Where <addr> is (reg:SP) or (plus (reg:SP) (const_int)). */ | |
18511 | e = XVECEXP (p, 0, i); | |
18512 | if (GET_CODE (e) != SET | |
18513 | || GET_CODE (XEXP (e, 0)) != MEM | |
18514 | || GET_CODE (XEXP (e, 1)) != REG) | |
18515 | abort (); | |
a60f1840 | 18516 | |
1774763d | 18517 | reg = REGNO (XEXP (e, 1)); |
18518 | if (reg < lastreg) | |
18519 | abort (); | |
a60f1840 | 18520 | |
1774763d | 18521 | if (i != 1) |
18522 | fprintf (asm_out_file, ", "); | |
18523 | /* We can't use %r for vfp because we need to use the | |
18524 | double precision register names. */ | |
18525 | if (IS_VFP_REGNUM (reg)) | |
18526 | asm_fprintf (asm_out_file, "d%d", (reg - FIRST_VFP_REGNUM) / 2); | |
18527 | else | |
18528 | asm_fprintf (asm_out_file, "%r", reg); | |
18529 | ||
18530 | #ifdef ENABLE_CHECKING | |
18531 | /* Check that the addresses are consecutive. */ | |
18532 | e = XEXP (XEXP (e, 0), 0); | |
18533 | if (GET_CODE (e) == PLUS) | |
18534 | { | |
18535 | offset += reg_size; | |
18536 | if (GET_CODE (XEXP (e, 0)) != REG | |
18537 | || REGNO (XEXP (e, 0)) != SP_REGNUM | |
18538 | || GET_CODE (XEXP (e, 1)) != CONST_INT | |
18539 | || offset != INTVAL (XEXP (e, 1))) | |
18540 | abort (); | |
18541 | } | |
18542 | else if (i != 1 | |
18543 | || GET_CODE (e) != REG | |
18544 | || REGNO (e) != SP_REGNUM) | |
18545 | abort (); | |
18546 | #endif | |
18547 | } | |
18548 | fprintf (asm_out_file, "}\n"); | |
18549 | } | |
18550 | ||
18551 | /* Emit unwind directives for a SET. */ | |
18552 | ||
18553 | static void | |
18554 | arm_unwind_emit_set (FILE * asm_out_file, rtx p) | |
18555 | { | |
18556 | rtx e0; | |
18557 | rtx e1; | |
25f905c2 | 18558 | unsigned reg; |
1774763d | 18559 | |
18560 | e0 = XEXP (p, 0); | |
18561 | e1 = XEXP (p, 1); | |
18562 | switch (GET_CODE (e0)) | |
18563 | { | |
18564 | case MEM: | |
18565 | /* Pushing a single register. */ | |
18566 | if (GET_CODE (XEXP (e0, 0)) != PRE_DEC | |
18567 | || GET_CODE (XEXP (XEXP (e0, 0), 0)) != REG | |
18568 | || REGNO (XEXP (XEXP (e0, 0), 0)) != SP_REGNUM) | |
18569 | abort (); | |
18570 | ||
18571 | asm_fprintf (asm_out_file, "\t.save "); | |
18572 | if (IS_VFP_REGNUM (REGNO (e1))) | |
18573 | asm_fprintf(asm_out_file, "{d%d}\n", | |
18574 | (REGNO (e1) - FIRST_VFP_REGNUM) / 2); | |
18575 | else | |
18576 | asm_fprintf(asm_out_file, "{%r}\n", REGNO (e1)); | |
18577 | break; | |
18578 | ||
18579 | case REG: | |
18580 | if (REGNO (e0) == SP_REGNUM) | |
18581 | { | |
18582 | /* A stack increment. */ | |
18583 | if (GET_CODE (e1) != PLUS | |
18584 | || GET_CODE (XEXP (e1, 0)) != REG | |
18585 | || REGNO (XEXP (e1, 0)) != SP_REGNUM | |
18586 | || GET_CODE (XEXP (e1, 1)) != CONST_INT) | |
18587 | abort (); | |
18588 | ||
7781aa77 | 18589 | asm_fprintf (asm_out_file, "\t.pad #%wd\n", |
1774763d | 18590 | -INTVAL (XEXP (e1, 1))); |
18591 | } | |
18592 | else if (REGNO (e0) == HARD_FRAME_POINTER_REGNUM) | |
18593 | { | |
18594 | HOST_WIDE_INT offset; | |
a60f1840 | 18595 | |
1774763d | 18596 | if (GET_CODE (e1) == PLUS) |
18597 | { | |
18598 | if (GET_CODE (XEXP (e1, 0)) != REG | |
18599 | || GET_CODE (XEXP (e1, 1)) != CONST_INT) | |
18600 | abort (); | |
18601 | reg = REGNO (XEXP (e1, 0)); | |
18602 | offset = INTVAL (XEXP (e1, 1)); | |
7781aa77 | 18603 | asm_fprintf (asm_out_file, "\t.setfp %r, %r, #%wd\n", |
1774763d | 18604 | HARD_FRAME_POINTER_REGNUM, reg, |
18605 | INTVAL (XEXP (e1, 1))); | |
18606 | } | |
18607 | else if (GET_CODE (e1) == REG) | |
18608 | { | |
18609 | reg = REGNO (e1); | |
18610 | asm_fprintf (asm_out_file, "\t.setfp %r, %r\n", | |
18611 | HARD_FRAME_POINTER_REGNUM, reg); | |
18612 | } | |
18613 | else | |
18614 | abort (); | |
18615 | } | |
18616 | else if (GET_CODE (e1) == REG && REGNO (e1) == SP_REGNUM) | |
18617 | { | |
18618 | /* Move from sp to reg. */ | |
18619 | asm_fprintf (asm_out_file, "\t.movsp %r\n", REGNO (e0)); | |
18620 | } | |
d105e3e2 | 18621 | else if (GET_CODE (e1) == PLUS |
18622 | && GET_CODE (XEXP (e1, 0)) == REG | |
18623 | && REGNO (XEXP (e1, 0)) == SP_REGNUM | |
18624 | && GET_CODE (XEXP (e1, 1)) == CONST_INT) | |
18625 | { | |
18626 | /* Set reg to offset from sp. */ | |
18627 | asm_fprintf (asm_out_file, "\t.movsp %r, #%d\n", | |
18628 | REGNO (e0), (int)INTVAL(XEXP (e1, 1))); | |
18629 | } | |
25f905c2 | 18630 | else if (GET_CODE (e1) == UNSPEC && XINT (e1, 1) == UNSPEC_STACK_ALIGN) |
18631 | { | |
18632 | /* Stack pointer save before alignment. */ | |
18633 | reg = REGNO (e0); | |
18634 | asm_fprintf (asm_out_file, "\t.unwind_raw 0, 0x%x @ vsp = r%d\n", | |
18635 | reg + 0x90, reg); | |
18636 | } | |
1774763d | 18637 | else |
18638 | abort (); | |
18639 | break; | |
18640 | ||
18641 | default: | |
18642 | abort (); | |
18643 | } | |
18644 | } | |
18645 | ||
18646 | ||
18647 | /* Emit unwind directives for the given insn. */ | |
18648 | ||
18649 | static void | |
18650 | arm_unwind_emit (FILE * asm_out_file, rtx insn) | |
18651 | { | |
18652 | rtx pat; | |
18653 | ||
18654 | if (!ARM_EABI_UNWIND_TABLES) | |
18655 | return; | |
18656 | ||
18657 | if (GET_CODE (insn) == NOTE || !RTX_FRAME_RELATED_P (insn)) | |
18658 | return; | |
18659 | ||
18660 | pat = find_reg_note (insn, REG_FRAME_RELATED_EXPR, NULL_RTX); | |
18661 | if (pat) | |
18662 | pat = XEXP (pat, 0); | |
18663 | else | |
18664 | pat = PATTERN (insn); | |
18665 | ||
18666 | switch (GET_CODE (pat)) | |
18667 | { | |
18668 | case SET: | |
18669 | arm_unwind_emit_set (asm_out_file, pat); | |
18670 | break; | |
18671 | ||
18672 | case SEQUENCE: | |
18673 | /* Store multiple. */ | |
25f905c2 | 18674 | arm_unwind_emit_sequence (asm_out_file, pat); |
1774763d | 18675 | break; |
18676 | ||
18677 | default: | |
18678 | abort(); | |
18679 | } | |
18680 | } | |
18681 | ||
18682 | ||
18683 | /* Output a reference from a function exception table to the type_info | |
18684 | object X. The EABI specifies that the symbol should be relocated by | |
18685 | an R_ARM_TARGET2 relocation. */ | |
18686 | ||
18687 | static bool | |
18688 | arm_output_ttype (rtx x) | |
18689 | { | |
18690 | fputs ("\t.word\t", asm_out_file); | |
18691 | output_addr_const (asm_out_file, x); | |
18692 | /* Use special relocations for symbol references. */ | |
18693 | if (GET_CODE (x) != CONST_INT) | |
18694 | fputs ("(TARGET2)", asm_out_file); | |
18695 | fputc ('\n', asm_out_file); | |
18696 | ||
18697 | return TRUE; | |
18698 | } | |
18699 | #endif /* TARGET_UNWIND_INFO */ | |
18700 | ||
18701 | ||
25f905c2 | 18702 | /* Handle UNSPEC DWARF call frame instructions. These are needed for dynamic |
18703 | stack alignment. */ | |
18704 | ||
18705 | static void | |
18706 | arm_dwarf_handle_frame_unspec (const char *label, rtx pattern, int index) | |
18707 | { | |
18708 | rtx unspec = SET_SRC (pattern); | |
18709 | gcc_assert (GET_CODE (unspec) == UNSPEC); | |
18710 | ||
18711 | switch (index) | |
18712 | { | |
18713 | case UNSPEC_STACK_ALIGN: | |
18714 | /* ??? We should set the CFA = (SP & ~7). At this point we haven't | |
18715 | put anything on the stack, so hopefully it won't matter. | |
18716 | CFA = SP will be correct after alignment. */ | |
18717 | dwarf2out_reg_save_reg (label, stack_pointer_rtx, | |
18718 | SET_DEST (pattern)); | |
18719 | break; | |
18720 | default: | |
18721 | gcc_unreachable (); | |
18722 | } | |
18723 | } | |
18724 | ||
18725 | ||
1774763d | 18726 | /* Output unwind directives for the start/end of a function. */ |
18727 | ||
18728 | void | |
18729 | arm_output_fn_unwind (FILE * f, bool prologue) | |
18730 | { | |
18731 | if (!ARM_EABI_UNWIND_TABLES) | |
18732 | return; | |
18733 | ||
18734 | if (prologue) | |
18735 | fputs ("\t.fnstart\n", f); | |
18736 | else | |
18737 | fputs ("\t.fnend\n", f); | |
18738 | } | |
f655717d | 18739 | |
18740 | static bool | |
18741 | arm_emit_tls_decoration (FILE *fp, rtx x) | |
18742 | { | |
18743 | enum tls_reloc reloc; | |
18744 | rtx val; | |
18745 | ||
18746 | val = XVECEXP (x, 0, 0); | |
18747 | reloc = INTVAL (XVECEXP (x, 0, 1)); | |
18748 | ||
18749 | output_addr_const (fp, val); | |
18750 | ||
18751 | switch (reloc) | |
18752 | { | |
18753 | case TLS_GD32: | |
18754 | fputs ("(tlsgd)", fp); | |
18755 | break; | |
18756 | case TLS_LDM32: | |
18757 | fputs ("(tlsldm)", fp); | |
18758 | break; | |
18759 | case TLS_LDO32: | |
18760 | fputs ("(tlsldo)", fp); | |
18761 | break; | |
18762 | case TLS_IE32: | |
18763 | fputs ("(gottpoff)", fp); | |
18764 | break; | |
18765 | case TLS_LE32: | |
18766 | fputs ("(tpoff)", fp); | |
18767 | break; | |
18768 | default: | |
18769 | gcc_unreachable (); | |
18770 | } | |
18771 | ||
18772 | switch (reloc) | |
18773 | { | |
18774 | case TLS_GD32: | |
18775 | case TLS_LDM32: | |
18776 | case TLS_IE32: | |
18777 | fputs (" + (. - ", fp); | |
18778 | output_addr_const (fp, XVECEXP (x, 0, 2)); | |
18779 | fputs (" - ", fp); | |
18780 | output_addr_const (fp, XVECEXP (x, 0, 3)); | |
18781 | fputc (')', fp); | |
18782 | break; | |
18783 | default: | |
18784 | break; | |
18785 | } | |
18786 | ||
18787 | return TRUE; | |
18788 | } | |
18789 | ||
4471a482 | 18790 | /* ARM implementation of TARGET_ASM_OUTPUT_DWARF_DTPREL. */ |
18791 | ||
18792 | static void | |
18793 | arm_output_dwarf_dtprel (FILE *file, int size, rtx x) | |
18794 | { | |
18795 | gcc_assert (size == 4); | |
18796 | fputs ("\t.word\t", file); | |
18797 | output_addr_const (file, x); | |
18798 | fputs ("(tlsldo)", file); | |
18799 | } | |
18800 | ||
f655717d | 18801 | bool |
18802 | arm_output_addr_const_extra (FILE *fp, rtx x) | |
18803 | { | |
18804 | if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLS) | |
18805 | return arm_emit_tls_decoration (fp, x); | |
6cdcb15c | 18806 | else if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_PIC_LABEL) |
18807 | { | |
18808 | char label[256]; | |
18809 | int labelno = INTVAL (XVECEXP (x, 0, 0)); | |
18810 | ||
18811 | ASM_GENERATE_INTERNAL_LABEL (label, "LPIC", labelno); | |
18812 | assemble_name_raw (fp, label); | |
18813 | ||
18814 | return TRUE; | |
18815 | } | |
f655717d | 18816 | else if (GET_CODE (x) == CONST_VECTOR) |
18817 | return arm_emit_vector_const (fp, x); | |
18818 | ||
18819 | return FALSE; | |
18820 | } | |
18821 | ||
25f905c2 | 18822 | /* Output assembly for a shift instruction. |
18823 | SET_FLAGS determines how the instruction modifies the condition codes. | |
542d5028 | 18824 | 0 - Do not set condition codes. |
25f905c2 | 18825 | 1 - Set condition codes. |
18826 | 2 - Use smallest instruction. */ | |
18827 | const char * | |
18828 | arm_output_shift(rtx * operands, int set_flags) | |
18829 | { | |
18830 | char pattern[100]; | |
18831 | static const char flag_chars[3] = {'?', '.', '!'}; | |
18832 | const char *shift; | |
18833 | HOST_WIDE_INT val; | |
18834 | char c; | |
18835 | ||
18836 | c = flag_chars[set_flags]; | |
18837 | if (TARGET_UNIFIED_ASM) | |
18838 | { | |
18839 | shift = shift_op(operands[3], &val); | |
18840 | if (shift) | |
18841 | { | |
18842 | if (val != -1) | |
18843 | operands[2] = GEN_INT(val); | |
18844 | sprintf (pattern, "%s%%%c\t%%0, %%1, %%2", shift, c); | |
18845 | } | |
18846 | else | |
18847 | sprintf (pattern, "mov%%%c\t%%0, %%1", c); | |
18848 | } | |
18849 | else | |
18850 | sprintf (pattern, "mov%%%c\t%%0, %%1%%S3", c); | |
18851 | output_asm_insn (pattern, operands); | |
18852 | return ""; | |
18853 | } | |
18854 | ||
18855 | /* Output a Thumb-2 casesi instruction. */ | |
18856 | const char * | |
18857 | thumb2_output_casesi (rtx *operands) | |
18858 | { | |
18859 | rtx diff_vec = PATTERN (next_real_insn (operands[2])); | |
18860 | ||
18861 | gcc_assert (GET_CODE (diff_vec) == ADDR_DIFF_VEC); | |
18862 | ||
18863 | output_asm_insn ("cmp\t%0, %1", operands); | |
18864 | output_asm_insn ("bhi\t%l3", operands); | |
18865 | switch (GET_MODE(diff_vec)) | |
18866 | { | |
18867 | case QImode: | |
18868 | return "tbb\t[%|pc, %0]"; | |
18869 | case HImode: | |
18870 | return "tbh\t[%|pc, %0, lsl #1]"; | |
18871 | case SImode: | |
18872 | if (flag_pic) | |
18873 | { | |
18874 | output_asm_insn ("adr\t%4, %l2", operands); | |
18875 | output_asm_insn ("ldr\t%5, [%4, %0, lsl #2]", operands); | |
18876 | output_asm_insn ("add\t%4, %4, %5", operands); | |
18877 | return "bx\t%4"; | |
18878 | } | |
18879 | else | |
18880 | { | |
18881 | output_asm_insn ("adr\t%4, %l2", operands); | |
18882 | return "ldr\t%|pc, [%4, %0, lsl #2]"; | |
18883 | } | |
18884 | default: | |
18885 | gcc_unreachable (); | |
18886 | } | |
18887 | } | |
18888 | ||
eddcdde1 | 18889 | /* A table and a function to perform ARM-specific name mangling for |
18890 | NEON vector types in order to conform to the AAPCS (see "Procedure | |
18891 | Call Standard for the ARM Architecture", Appendix A). To qualify | |
18892 | for emission with the mangled names defined in that document, a | |
18893 | vector type must not only be of the correct mode but also be | |
18894 | composed of NEON vector element types (e.g. __builtin_neon_qi). */ | |
18895 | typedef struct | |
18896 | { | |
18897 | enum machine_mode mode; | |
18898 | const char *element_type_name; | |
18899 | const char *aapcs_name; | |
18900 | } arm_mangle_map_entry; | |
18901 | ||
18902 | static arm_mangle_map_entry arm_mangle_map[] = { | |
18903 | /* 64-bit containerized types. */ | |
18904 | { V8QImode, "__builtin_neon_qi", "15__simd64_int8_t" }, | |
18905 | { V8QImode, "__builtin_neon_uqi", "16__simd64_uint8_t" }, | |
18906 | { V4HImode, "__builtin_neon_hi", "16__simd64_int16_t" }, | |
18907 | { V4HImode, "__builtin_neon_uhi", "17__simd64_uint16_t" }, | |
18908 | { V2SImode, "__builtin_neon_si", "16__simd64_int32_t" }, | |
18909 | { V2SImode, "__builtin_neon_usi", "17__simd64_uint32_t" }, | |
18910 | { V2SFmode, "__builtin_neon_sf", "18__simd64_float32_t" }, | |
18911 | { V8QImode, "__builtin_neon_poly8", "16__simd64_poly8_t" }, | |
18912 | { V4HImode, "__builtin_neon_poly16", "17__simd64_poly16_t" }, | |
18913 | /* 128-bit containerized types. */ | |
18914 | { V16QImode, "__builtin_neon_qi", "16__simd128_int8_t" }, | |
18915 | { V16QImode, "__builtin_neon_uqi", "17__simd128_uint8_t" }, | |
18916 | { V8HImode, "__builtin_neon_hi", "17__simd128_int16_t" }, | |
18917 | { V8HImode, "__builtin_neon_uhi", "18__simd128_uint16_t" }, | |
18918 | { V4SImode, "__builtin_neon_si", "17__simd128_int32_t" }, | |
18919 | { V4SImode, "__builtin_neon_usi", "18__simd128_uint32_t" }, | |
18920 | { V4SFmode, "__builtin_neon_sf", "19__simd128_float32_t" }, | |
18921 | { V16QImode, "__builtin_neon_poly8", "17__simd128_poly8_t" }, | |
18922 | { V8HImode, "__builtin_neon_poly16", "18__simd128_poly16_t" }, | |
18923 | { VOIDmode, NULL, NULL } | |
18924 | }; | |
18925 | ||
18926 | const char * | |
18927 | arm_mangle_type (tree type) | |
18928 | { | |
18929 | arm_mangle_map_entry *pos = arm_mangle_map; | |
18930 | ||
18931 | if (TREE_CODE (type) != VECTOR_TYPE) | |
18932 | return NULL; | |
18933 | ||
18934 | /* Check the mode of the vector type, and the name of the vector | |
18935 | element type, against the table. */ | |
18936 | while (pos->mode != VOIDmode) | |
18937 | { | |
18938 | tree elt_type = TREE_TYPE (type); | |
18939 | ||
18940 | if (pos->mode == TYPE_MODE (type) | |
18941 | && TREE_CODE (TYPE_NAME (elt_type)) == TYPE_DECL | |
18942 | && !strcmp (IDENTIFIER_POINTER (DECL_NAME (TYPE_NAME (elt_type))), | |
18943 | pos->element_type_name)) | |
18944 | return pos->aapcs_name; | |
18945 | ||
18946 | pos++; | |
18947 | } | |
18948 | ||
18949 | /* Use the default mangling for unrecognized (possibly user-defined) | |
18950 | vector types. */ | |
18951 | return NULL; | |
18952 | } | |
18953 | ||
f655717d | 18954 | #include "gt-arm.h" |