1 /* Output routines for Visium.
2 Copyright (C) 2002-2022 Free Software Foundation, Inc.
3 Contributed by C.Nettleton, J.P.Parkes and P.Garbett.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it
8 under the terms of the GNU General Public License as published
9 by the Free Software Foundation; either version 3, or (at your
10 option) any later version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT
13 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
14 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public
15 License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 #define IN_TARGET_CODE 1
25 #include "coretypes.h"
30 #include "gimple-expr.h"
34 #include "stringpool.h"
41 #include "diagnostic-core.h"
44 #include "fold-const.h"
45 #include "stor-layout.h"
49 #include "insn-attr.h"
53 #include "langhooks.h"
55 #include "tm-constrs.h"
56 #include "tree-pass.h"
61 /* This file should be included last. */
62 #include "target-def.h"
64 /* Enumeration of indexes into machine_libfunc_table. */
65 enum machine_libfunc_index
75 MLTI_set_trampoline_parity
,
80 struct GTY(()) machine_libfuncs
85 /* The table of Visium-specific libfuncs. */
86 static GTY(()) struct machine_libfuncs visium_libfuncs
;
88 #define vlt visium_libfuncs.table
90 /* Accessor macros for visium_libfuncs. */
91 #define long_int_memcpy_libfunc (vlt[MLTI_long_int_memcpy])
92 #define wrd_memcpy_libfunc (vlt[MLTI_wrd_memcpy])
93 #define byt_memcpy_libfunc (vlt[MLTI_byt_memcpy])
94 #define long_int_memset_libfunc (vlt[MLTI_long_int_memset])
95 #define wrd_memset_libfunc (vlt[MLTI_wrd_memset])
96 #define byt_memset_libfunc (vlt[MLTI_byt_memset])
97 #define set_trampoline_parity_libfunc (vlt[MLTI_set_trampoline_parity])
99 /* Machine specific function data. */
100 struct GTY (()) machine_function
102 /* Size of the frame of the function. */
105 /* Size of the reg parm save area, non-zero only for functions with variable
106 argument list. We cannot use the crtl->args.pretend_args_size machinery
107 for this purpose because this size is added to virtual_incoming_args_rtx
108 to give the location of the first parameter passed by the caller on the
109 stack and virtual_incoming_args_rtx is also the location of the first
110 parameter on the stack. So crtl->args.pretend_args_size can be non-zero
111 only if the first non-register named parameter is not passed entirely on
112 the stack and this runs afoul of the need to have a reg parm save area
113 even with a variable argument list starting on the stack because of the
114 separate handling of general and floating-point registers. */
115 int reg_parm_save_area_size
;
117 /* True if we have created an rtx which relies on the frame pointer. */
120 /* True if we have exposed the flags register. From this moment on, we
121 cannot generate simple operations for integer registers. We could
122 use reload_completed for this purpose, but this would cripple the
123 postreload CSE and GCSE passes which run before postreload split. */
127 #define visium_frame_size cfun->machine->frame_size
128 #define visium_reg_parm_save_area_size cfun->machine->reg_parm_save_area_size
129 #define visium_frame_needed cfun->machine->frame_needed
130 #define visium_flags_exposed cfun->machine->flags_exposed
132 /* 1 if the next opcode is to be specially indented. */
133 int visium_indent_opcode
= 0;
135 /* Register number used for long branches when LR isn't available. It
136 must be a call-used register since it isn't saved on function entry.
137 We do not care whether the branch is predicted or not on the GR6,
138 given how unlikely it is to have a long branch in a leaf function. */
139 static unsigned int long_branch_regnum
= 31;
141 static tree
visium_handle_interrupt_attr (tree
*, tree
, tree
, int, bool *);
142 static inline bool current_function_saves_fp (void);
143 static inline bool current_function_saves_lr (void);
144 static inline bool current_function_has_lr_slot (void);
146 /* Supported attributes:
147 interrupt -- specifies this function is an interrupt handler. */
148 static const struct attribute_spec visium_attribute_table
[] =
150 /* { name, min_len, max_len, decl_req, type_req, fn_type_req,
151 affects_type_identity, handler, exclude } */
152 { "interrupt", 0, 0, true, false, false, false, visium_handle_interrupt_attr
,
154 { NULL
, 0, 0, false, false, false, false, NULL
, NULL
},
157 static struct machine_function
*visium_init_machine_status (void);
159 /* Target hooks and TARGET_INITIALIZER */
161 static bool visium_pass_by_reference (cumulative_args_t
,
162 const function_arg_info
&);
164 static rtx
visium_function_arg (cumulative_args_t
, const function_arg_info
&);
166 static void visium_function_arg_advance (cumulative_args_t
,
167 const function_arg_info
&);
169 static bool visium_return_in_memory (const_tree
, const_tree fntype
);
171 static rtx
visium_function_value (const_tree
, const_tree fn_decl_or_type
,
174 static rtx
visium_libcall_value (machine_mode
, const_rtx
);
176 static void visium_setup_incoming_varargs (cumulative_args_t
,
177 const function_arg_info
&,
180 static void visium_va_start (tree valist
, rtx nextarg
);
182 static tree
visium_gimplify_va_arg (tree
, tree
, gimple_seq
*, gimple_seq
*);
184 static bool visium_function_ok_for_sibcall (tree
, tree
);
186 static bool visium_frame_pointer_required (void);
188 static tree
visium_build_builtin_va_list (void);
190 static rtx_insn
*visium_md_asm_adjust (vec
<rtx
> &, vec
<rtx
> &,
192 vec
<const char *> &, vec
<rtx
> &,
193 HARD_REG_SET
&, location_t
);
195 static bool visium_legitimate_constant_p (machine_mode
, rtx
);
197 static bool visium_legitimate_address_p (machine_mode
, rtx
, bool);
199 static bool visium_print_operand_punct_valid_p (unsigned char);
200 static void visium_print_operand (FILE *, rtx
, int);
201 static void visium_print_operand_address (FILE *, machine_mode
, rtx
);
203 static void visium_conditional_register_usage (void);
205 static rtx
visium_legitimize_address (rtx
, rtx
, machine_mode
);
207 static reg_class_t
visium_secondary_reload (bool, rtx
, reg_class_t
,
209 secondary_reload_info
*);
211 static bool visium_class_likely_spilled_p (reg_class_t
);
213 static void visium_trampoline_init (rtx
, tree
, rtx
);
215 static int visium_issue_rate (void);
217 static int visium_adjust_priority (rtx_insn
*, int);
219 static int visium_adjust_cost (rtx_insn
*, int, rtx_insn
*, int, unsigned int);
221 static int visium_register_move_cost (machine_mode
, reg_class_t
,
224 static int visium_memory_move_cost (machine_mode
, reg_class_t
, bool);
226 static bool visium_rtx_costs (rtx
, machine_mode
, int, int, int *, bool);
228 static void visium_option_override (void);
230 static void visium_init_libfuncs (void);
232 static unsigned int visium_reorg (void);
234 static unsigned int visium_hard_regno_nregs (unsigned int, machine_mode
);
236 static bool visium_hard_regno_mode_ok (unsigned int, machine_mode
);
238 static bool visium_modes_tieable_p (machine_mode
, machine_mode
);
240 static bool visium_can_change_mode_class (machine_mode
, machine_mode
,
243 static HOST_WIDE_INT
visium_constant_alignment (const_tree
, HOST_WIDE_INT
);
245 /* Setup the global target hooks structure. */
247 #undef TARGET_MAX_ANCHOR_OFFSET
248 #define TARGET_MAX_ANCHOR_OFFSET 31
250 #undef TARGET_PASS_BY_REFERENCE
251 #define TARGET_PASS_BY_REFERENCE visium_pass_by_reference
253 #undef TARGET_FUNCTION_ARG
254 #define TARGET_FUNCTION_ARG visium_function_arg
256 #undef TARGET_FUNCTION_ARG_ADVANCE
257 #define TARGET_FUNCTION_ARG_ADVANCE visium_function_arg_advance
259 #undef TARGET_RETURN_IN_MEMORY
260 #define TARGET_RETURN_IN_MEMORY visium_return_in_memory
262 #undef TARGET_FUNCTION_VALUE
263 #define TARGET_FUNCTION_VALUE visium_function_value
265 #undef TARGET_LIBCALL_VALUE
266 #define TARGET_LIBCALL_VALUE visium_libcall_value
268 #undef TARGET_SETUP_INCOMING_VARARGS
269 #define TARGET_SETUP_INCOMING_VARARGS visium_setup_incoming_varargs
271 #undef TARGET_EXPAND_BUILTIN_VA_START
272 #define TARGET_EXPAND_BUILTIN_VA_START visium_va_start
274 #undef TARGET_BUILD_BUILTIN_VA_LIST
275 #define TARGET_BUILD_BUILTIN_VA_LIST visium_build_builtin_va_list
277 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
278 #define TARGET_GIMPLIFY_VA_ARG_EXPR visium_gimplify_va_arg
280 #undef TARGET_LEGITIMATE_CONSTANT_P
281 #define TARGET_LEGITIMATE_CONSTANT_P visium_legitimate_constant_p
284 #define TARGET_LRA_P hook_bool_void_false
286 #undef TARGET_LEGITIMATE_ADDRESS_P
287 #define TARGET_LEGITIMATE_ADDRESS_P visium_legitimate_address_p
289 #undef TARGET_PRINT_OPERAND_PUNCT_VALID_P
290 #define TARGET_PRINT_OPERAND_PUNCT_VALID_P visium_print_operand_punct_valid_p
292 #undef TARGET_PRINT_OPERAND
293 #define TARGET_PRINT_OPERAND visium_print_operand
295 #undef TARGET_PRINT_OPERAND_ADDRESS
296 #define TARGET_PRINT_OPERAND_ADDRESS visium_print_operand_address
298 #undef TARGET_ATTRIBUTE_TABLE
299 #define TARGET_ATTRIBUTE_TABLE visium_attribute_table
301 #undef TARGET_ADDRESS_COST
302 #define TARGET_ADDRESS_COST hook_int_rtx_mode_as_bool_0
304 #undef TARGET_STRICT_ARGUMENT_NAMING
305 #define TARGET_STRICT_ARGUMENT_NAMING hook_bool_CUMULATIVE_ARGS_true
307 #undef TARGET_SCHED_ISSUE_RATE
308 #define TARGET_SCHED_ISSUE_RATE visium_issue_rate
310 #undef TARGET_SCHED_ADJUST_PRIORITY
311 #define TARGET_SCHED_ADJUST_PRIORITY visium_adjust_priority
313 #undef TARGET_SCHED_ADJUST_COST
314 #define TARGET_SCHED_ADJUST_COST visium_adjust_cost
316 #undef TARGET_MEMORY_MOVE_COST
317 #define TARGET_MEMORY_MOVE_COST visium_memory_move_cost
319 #undef TARGET_REGISTER_MOVE_COST
320 #define TARGET_REGISTER_MOVE_COST visium_register_move_cost
322 #undef TARGET_RTX_COSTS
323 #define TARGET_RTX_COSTS visium_rtx_costs
325 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
326 #define TARGET_FUNCTION_OK_FOR_SIBCALL visium_function_ok_for_sibcall
328 #undef TARGET_FRAME_POINTER_REQUIRED
329 #define TARGET_FRAME_POINTER_REQUIRED visium_frame_pointer_required
331 #undef TARGET_SECONDARY_RELOAD
332 #define TARGET_SECONDARY_RELOAD visium_secondary_reload
334 #undef TARGET_CLASS_LIKELY_SPILLED_P
335 #define TARGET_CLASS_LIKELY_SPILLED_P visium_class_likely_spilled_p
337 #undef TARGET_LEGITIMIZE_ADDRESS
338 #define TARGET_LEGITIMIZE_ADDRESS visium_legitimize_address
340 #undef TARGET_OPTION_OVERRIDE
341 #define TARGET_OPTION_OVERRIDE visium_option_override
343 #undef TARGET_INIT_LIBFUNCS
344 #define TARGET_INIT_LIBFUNCS visium_init_libfuncs
346 #undef TARGET_CONDITIONAL_REGISTER_USAGE
347 #define TARGET_CONDITIONAL_REGISTER_USAGE visium_conditional_register_usage
349 #undef TARGET_TRAMPOLINE_INIT
350 #define TARGET_TRAMPOLINE_INIT visium_trampoline_init
352 #undef TARGET_MD_ASM_ADJUST
353 #define TARGET_MD_ASM_ADJUST visium_md_asm_adjust
355 #undef TARGET_FLAGS_REGNUM
356 #define TARGET_FLAGS_REGNUM FLAGS_REGNUM
358 #undef TARGET_HARD_REGNO_NREGS
359 #define TARGET_HARD_REGNO_NREGS visium_hard_regno_nregs
361 #undef TARGET_HARD_REGNO_MODE_OK
362 #define TARGET_HARD_REGNO_MODE_OK visium_hard_regno_mode_ok
364 #undef TARGET_MODES_TIEABLE_P
365 #define TARGET_MODES_TIEABLE_P visium_modes_tieable_p
367 #undef TARGET_CAN_CHANGE_MODE_CLASS
368 #define TARGET_CAN_CHANGE_MODE_CLASS visium_can_change_mode_class
370 #undef TARGET_CONSTANT_ALIGNMENT
371 #define TARGET_CONSTANT_ALIGNMENT visium_constant_alignment
373 #undef TARGET_HAVE_SPECULATION_SAFE_VALUE
374 #define TARGET_HAVE_SPECULATION_SAFE_VALUE speculation_safe_value_not_needed
376 struct gcc_target targetm
= TARGET_INITIALIZER
;
380 const pass_data pass_data_visium_reorg
=
384 OPTGROUP_NONE
, /* optinfo_flags */
385 TV_MACH_DEP
, /* tv_id */
386 0, /* properties_required */
387 0, /* properties_provided */
388 0, /* properties_destroyed */
389 0, /* todo_flags_start */
390 0, /* todo_flags_finish */
393 class pass_visium_reorg
: public rtl_opt_pass
396 pass_visium_reorg(gcc::context
*ctxt
)
397 : rtl_opt_pass(pass_data_visium_reorg
, ctxt
)
400 /* opt_pass methods: */
401 virtual unsigned int execute (function
*)
403 return visium_reorg ();
406 }; // class pass_work_around_errata
411 make_pass_visium_reorg (gcc::context
*ctxt
)
413 return new pass_visium_reorg (ctxt
);
416 /* Options override for Visium. */
419 visium_option_override (void)
422 warning (OPT_fpic
, "%<-fpic%> is not supported");
424 warning (OPT_fPIC
, "%<-fPIC%> is not supported");
426 /* MCM is the default in the GR5/GR6 era. */
427 target_flags
|= MASK_MCM
;
429 /* FPU is the default with MCM, but don't override an explicit option. */
430 if ((target_flags_explicit
& MASK_FPU
) == 0)
431 target_flags
|= MASK_FPU
;
433 /* The supervisor mode is the default. */
434 if ((target_flags_explicit
& MASK_SV_MODE
) == 0)
435 target_flags
|= MASK_SV_MODE
;
437 /* The GR6 has the Block Move Instructions and an IEEE-compliant FPU. */
438 if (visium_cpu_and_features
== PROCESSOR_GR6
)
440 target_flags
|= MASK_BMI
;
441 if (target_flags
& MASK_FPU
)
442 target_flags
|= MASK_FPU_IEEE
;
445 /* Set -mtune from -mcpu if not specified. */
446 if (!OPTION_SET_P (visium_cpu
))
447 visium_cpu
= visium_cpu_and_features
;
449 /* Align functions on 256-byte (32-quadword) for GR5 and 64-byte (8-quadword)
450 boundaries for GR6 so they start a new burst mode window. */
451 if (flag_align_functions
&& !str_align_functions
)
453 if (visium_cpu
== PROCESSOR_GR6
)
454 str_align_functions
= "64";
456 str_align_functions
= "256";
458 /* Allow the size of compilation units to double because of inlining.
459 In practice the global size of the object code is hardly affected
460 because the additional instructions will take up the padding. */
461 SET_OPTION_IF_UNSET (&global_options
, &global_options_set
,
462 param_inline_unit_growth
, 100);
465 /* Likewise for loops. */
466 if (flag_align_loops
&& !str_align_loops
)
468 if (visium_cpu
== PROCESSOR_GR6
)
469 str_align_loops
= "64";
472 /* But not if they are too far away from a 256-byte boundary. */
473 str_align_loops
= "256:32:8";
477 /* Align all jumps on quadword boundaries for the burst mode, and even
478 on 8-quadword boundaries for GR6 so they start a new window. */
479 if (flag_align_jumps
&& !str_align_jumps
)
481 if (visium_cpu
== PROCESSOR_GR6
)
482 str_align_jumps
= "64";
484 str_align_jumps
= "8";
488 /* Register the Visium-specific libfuncs with the middle-end. */
491 visium_init_libfuncs (void)
494 long_int_memcpy_libfunc
= init_one_libfunc ("__long_int_memcpy");
495 wrd_memcpy_libfunc
= init_one_libfunc ("__wrd_memcpy");
496 byt_memcpy_libfunc
= init_one_libfunc ("__byt_memcpy");
498 long_int_memset_libfunc
= init_one_libfunc ("__long_int_memset");
499 wrd_memset_libfunc
= init_one_libfunc ("__wrd_memset");
500 byt_memset_libfunc
= init_one_libfunc ("__byt_memset");
502 set_trampoline_parity_libfunc
= init_one_libfunc ("__set_trampoline_parity");
505 /* Return the number of instructions that can issue on the same cycle. */
508 visium_issue_rate (void)
523 /* Return the adjusted PRIORITY of INSN. */
526 visium_adjust_priority (rtx_insn
*insn
, int priority
)
528 /* On the GR5, we slightly increase the priority of writes in order to avoid
529 scheduling a read on the next cycle. This is necessary in addition to the
530 associated insn reservation because there are no data dependencies.
531 We also slightly increase the priority of reads from ROM in order to group
532 them as much as possible. These reads are a bit problematic because they
533 conflict with the instruction fetches, i.e. the data and instruction buses
534 tread on each other's toes when they are executed. */
535 if (visium_cpu
== PROCESSOR_GR5
538 && recog_memoized (insn
) >= 0)
540 enum attr_type attr_type
= get_attr_type (insn
);
541 if (attr_type
== TYPE_REG_MEM
542 || (attr_type
== TYPE_MEM_REG
543 && MEM_READONLY_P (SET_SRC (PATTERN (insn
)))))
550 /* Adjust the cost of a scheduling dependency. Return the new cost of
551 a dependency LINK of INSN on DEP_INSN. COST is the current cost. */
554 visium_adjust_cost (rtx_insn
*insn
, int dep_type
, rtx_insn
*dep_insn
, int cost
,
557 enum attr_type attr_type
;
559 /* Don't adjust costs for true dependencies as they are described with
560 bypasses. But we make an exception for the first scheduling pass to
561 help the subsequent postreload compare elimination pass. */
562 if (dep_type
== REG_DEP_TRUE
)
564 if (!reload_completed
565 && recog_memoized (insn
) >= 0
566 && get_attr_type (insn
) == TYPE_CMP
)
568 rtx pat
= PATTERN (insn
);
569 gcc_assert (GET_CODE (pat
) == SET
);
570 rtx src
= SET_SRC (pat
);
572 /* Only the branches can be modified by the postreload compare
573 elimination pass, not the cstores because they accept only
574 unsigned comparison operators and they are eliminated if
575 one of the operands is zero. */
576 if (GET_CODE (src
) == IF_THEN_ELSE
577 && XEXP (XEXP (src
, 0), 1) == const0_rtx
578 && recog_memoized (dep_insn
) >= 0)
580 enum attr_type dep_attr_type
= get_attr_type (dep_insn
);
582 /* The logical instructions use CCmode and thus work with any
583 comparison operator, whereas the arithmetic instructions use
584 CCNZmode and thus work with only a small subset. */
585 if (dep_attr_type
== TYPE_LOGIC
586 || (dep_attr_type
== TYPE_ARITH
587 && visium_nz_comparison_operator (XEXP (src
, 0),
597 if (recog_memoized (insn
) < 0)
600 attr_type
= get_attr_type (insn
);
602 /* Anti dependency: DEP_INSN reads a register that INSN writes some
604 if (dep_type
== REG_DEP_ANTI
)
606 /* On the GR5, the latency of FP instructions needs to be taken into
607 account for every dependency involving a write. */
608 if (attr_type
== TYPE_REG_FP
&& visium_cpu
== PROCESSOR_GR5
)
611 rtx pat
= PATTERN (insn
);
612 rtx dep_pat
= PATTERN (dep_insn
);
614 if (GET_CODE (pat
) != SET
|| GET_CODE (dep_pat
) != SET
)
615 /* If this happens, we have to extend this to schedule
616 optimally. Return 0 for now. */
619 if (reg_mentioned_p (SET_DEST (pat
), SET_SRC (dep_pat
)))
621 if (recog_memoized (dep_insn
) < 0)
624 switch (get_attr_type (dep_insn
))
632 /* A fload can't be issued until a preceding arithmetic
633 operation has finished if the target of the fload is
634 any of the sources (or destination) of the arithmetic
635 operation. Note that the latency may be (much)
636 greater than this if the preceding instruction
637 concerned is in a queue. */
638 return insn_default_latency (dep_insn
);
646 /* On the GR6, we try to make sure that the link register is restored
647 sufficiently ahead of the return as to yield a correct prediction
648 from the branch predictor. By default there is no true dependency
649 but an anti dependency between them, so we simply reuse it. */
650 else if (attr_type
== TYPE_RET
&& visium_cpu
== PROCESSOR_GR6
)
652 rtx dep_pat
= PATTERN (dep_insn
);
653 if (GET_CODE (dep_pat
) == SET
654 && REG_P (SET_DEST (dep_pat
))
655 && REGNO (SET_DEST (dep_pat
)) == LINK_REGNUM
)
659 /* For other anti dependencies, the cost is 0. */
663 /* Output dependency: DEP_INSN writes a register that INSN writes some
665 else if (dep_type
== REG_DEP_OUTPUT
)
667 /* On the GR5, the latency of FP instructions needs to be taken into
668 account for every dependency involving a write. */
669 if (attr_type
== TYPE_REG_FP
&& visium_cpu
== PROCESSOR_GR5
)
672 rtx pat
= PATTERN (insn
);
673 rtx dep_pat
= PATTERN (dep_insn
);
675 if (GET_CODE (pat
) != SET
|| GET_CODE (dep_pat
) != SET
)
676 /* If this happens, we have to extend this to schedule
677 optimally. Return 0 for now. */
680 if (reg_mentioned_p (SET_DEST (pat
), SET_DEST (dep_pat
)))
682 if (recog_memoized (dep_insn
) < 0)
685 switch (get_attr_type (dep_insn
))
693 /* A fload can't be issued until a preceding arithmetic
694 operation has finished if the target of the fload is
695 the destination of the arithmetic operation. Note that
696 the latency may be (much) greater than this if the
697 preceding instruction concerned is in a queue. */
698 return insn_default_latency (dep_insn
);
706 /* For other output dependencies, the cost is 0. */
713 /* Handle an "interrupt_handler" attribute; arguments as in
714 struct attribute_spec.handler. */
717 visium_handle_interrupt_attr (tree
*node
, tree name
,
718 tree args ATTRIBUTE_UNUSED
,
719 int flags ATTRIBUTE_UNUSED
,
722 if (TREE_CODE (*node
) != FUNCTION_DECL
)
724 warning (OPT_Wattributes
, "%qE attribute only applies to functions",
726 *no_add_attrs
= true;
728 else if (!TARGET_SV_MODE
)
730 error ("an interrupt handler cannot be compiled with %<-muser-mode%>");
731 *no_add_attrs
= true;
737 /* Return non-zero if the current function is an interrupt function. */
740 visium_interrupt_function_p (void)
743 lookup_attribute ("interrupt",
744 DECL_ATTRIBUTES (current_function_decl
)) != NULL_TREE
;
747 /* Conditionally modify the settings of the register file. */
750 visium_conditional_register_usage (void)
752 /* If the supervisor mode is disabled, mask some general registers. */
755 if (visium_cpu_and_features
== PROCESSOR_GR5
)
762 call_used_regs
[24] = 0;
763 call_used_regs
[25] = 0;
764 call_used_regs
[26] = 0;
765 call_used_regs
[27] = 0;
766 call_used_regs
[28] = 0;
770 call_used_regs
[31] = 0;
772 /* We also need to change the long-branch register. */
773 if (visium_cpu_and_features
== PROCESSOR_GR5
)
774 long_branch_regnum
= 20;
776 long_branch_regnum
= 28;
779 /* If the FPU is disabled, mask the FP registers. */
782 for (int i
= FP_FIRST_REGNUM
; i
<= FP_LAST_REGNUM
; i
++)
785 call_used_regs
[i
] = 0;
790 /* Prepend to CLOBBERS hard registers that are automatically clobbered for
791 an asm We do this for the FLAGS to maintain source compatibility with
792 the original cc0-based compiler. */
795 visium_md_asm_adjust (vec
<rtx
> & /*outputs*/, vec
<rtx
> & /*inputs*/,
796 vec
<machine_mode
> & /*input_modes*/,
797 vec
<const char *> & /*constraints*/, vec
<rtx
> &clobbers
,
798 HARD_REG_SET
&clobbered_regs
, location_t
/*loc*/)
800 clobbers
.safe_push (gen_rtx_REG (CCmode
, FLAGS_REGNUM
));
801 SET_HARD_REG_BIT (clobbered_regs
, FLAGS_REGNUM
);
805 /* Return true if X is a legitimate constant for a MODE immediate operand.
806 X is guaranteed to satisfy the CONSTANT_P predicate. */
809 visium_legitimate_constant_p (machine_mode mode ATTRIBUTE_UNUSED
,
810 rtx x ATTRIBUTE_UNUSED
)
815 /* Compute the alignment for a variable. The alignment of an aggregate is
816 set to be at least that of a scalar less than or equal to it in size. */
819 visium_data_alignment (tree type
, unsigned int align
)
821 if (AGGREGATE_TYPE_P (type
)
823 && TREE_CODE (TYPE_SIZE (type
)) == INTEGER_CST
&& align
< 32)
825 if (TREE_INT_CST_LOW (TYPE_SIZE (type
)) >= 32)
828 if (TREE_INT_CST_LOW (TYPE_SIZE (type
)) >= 16 && align
< 16)
835 /* Implement TARGET_CONSTANT_ALIGNMENT. */
838 visium_constant_alignment (const_tree exp
, HOST_WIDE_INT align
)
840 return visium_data_alignment (TREE_TYPE (exp
), align
);
843 /* Helper function for HARD_REGNO_RENAME_OK (FROM, TO). Return non-zero if
844 it is OK to rename a hard register FROM to another hard register TO. */
847 visium_hard_regno_rename_ok (unsigned int from ATTRIBUTE_UNUSED
,
850 /* If the function doesn't save LR, then the long-branch register will be
851 used for long branches so we need to know whether it is live before the
852 frame layout is computed. */
853 if (!current_function_saves_lr () && to
== long_branch_regnum
)
856 /* Interrupt functions can only use registers that have already been
857 saved by the prologue, even if they would normally be call-clobbered. */
859 && !df_regs_ever_live_p (to
)
860 && visium_interrupt_function_p ())
866 /* Implement TARGET_HARD_REGNO_NREGS. */
869 visium_hard_regno_nregs (unsigned int regno
, machine_mode mode
)
871 if (regno
== MDB_REGNUM
)
872 return CEIL (GET_MODE_SIZE (mode
), 2 * UNITS_PER_WORD
);
873 return CEIL (GET_MODE_SIZE (mode
), UNITS_PER_WORD
);
876 /* Implement TARGET_HARD_REGNO_MODE_OK.
878 Modes with sizes which cross from the one register class to the
879 other cannot be allowed. Only single floats are allowed in the
880 floating point registers, and only fixed point values in the EAM
884 visium_hard_regno_mode_ok (unsigned int regno
, machine_mode mode
)
886 if (GP_REGISTER_P (regno
))
887 return GP_REGISTER_P (end_hard_regno (mode
, regno
) - 1);
889 if (FP_REGISTER_P (regno
))
890 return mode
== SFmode
|| (mode
== SImode
&& TARGET_FPU_IEEE
);
892 return (GET_MODE_CLASS (mode
) == MODE_INT
893 && visium_hard_regno_nregs (regno
, mode
) == 1);
896 /* Implement TARGET_MODES_TIEABLE_P. */
899 visium_modes_tieable_p (machine_mode mode1
, machine_mode mode2
)
901 return (GET_MODE_CLASS (mode1
) == MODE_INT
902 && GET_MODE_CLASS (mode2
) == MODE_INT
);
905 /* Return true if it is ok to do sibling call optimization for the specified
906 call expression EXP. DECL will be the called function, or NULL if this
907 is an indirect call. */
910 visium_function_ok_for_sibcall (tree decl ATTRIBUTE_UNUSED
,
911 tree exp ATTRIBUTE_UNUSED
)
913 return !visium_interrupt_function_p ();
916 /* Prepare operands for a move define_expand in MODE. */
919 prepare_move_operands (rtx
*operands
, machine_mode mode
)
921 /* If the output is not a register, the input must be. */
922 if (GET_CODE (operands
[0]) == MEM
&& !reg_or_0_operand (operands
[1], mode
))
923 operands
[1] = force_reg (mode
, operands
[1]);
926 /* Return true if the operands are valid for a simple move insn. */
929 ok_for_simple_move_operands (rtx
*operands
, machine_mode mode
)
931 /* One of the operands must be a register. */
932 if (!register_operand (operands
[0], mode
)
933 && !reg_or_0_operand (operands
[1], mode
))
936 /* Once the flags are exposed, no simple moves between integer registers. */
937 if (visium_flags_exposed
938 && gpc_reg_operand (operands
[0], mode
)
939 && gpc_reg_operand (operands
[1], mode
))
945 /* Return true if the operands are valid for a simple move strict insn. */
948 ok_for_simple_move_strict_operands (rtx
*operands
, machine_mode mode
)
950 /* Once the flags are exposed, no simple moves between integer registers.
951 Note that, in QImode only, a zero source counts as an integer register
952 since it will be emitted as r0. */
953 if (visium_flags_exposed
954 && gpc_reg_operand (operands
[0], mode
)
955 && (gpc_reg_operand (operands
[1], mode
)
956 || (mode
== QImode
&& operands
[1] == const0_rtx
)))
962 /* Return true if the operands are valid for a simple arithmetic or logical
966 ok_for_simple_arith_logic_operands (rtx
*, machine_mode
)
968 /* Once the flags are exposed, no simple arithmetic or logical operations
969 between integer registers. */
970 return !visium_flags_exposed
;
973 /* Return non-zero if a branch or call instruction will be emitting a nop
974 into its delay slot. */
977 empty_delay_slot (rtx_insn
*insn
)
981 /* If no previous instruction (should not happen), return true. */
982 if (PREV_INSN (insn
) == NULL
)
985 seq
= NEXT_INSN (PREV_INSN (insn
));
986 if (GET_CODE (PATTERN (seq
)) == SEQUENCE
)
992 /* Wrapper around single_set which returns the second SET of a pair if the
993 first SET is to the flags register. */
996 single_set_and_flags (rtx_insn
*insn
)
998 if (multiple_sets (insn
))
1000 rtx pat
= PATTERN (insn
);
1001 if (XVECLEN (pat
, 0) == 2
1002 && GET_CODE (XVECEXP (pat
, 0, 0)) == SET
1003 && REG_P (SET_DEST (XVECEXP (pat
, 0, 0)))
1004 && REGNO (SET_DEST (XVECEXP (pat
, 0, 0))) == FLAGS_REGNUM
)
1005 return XVECEXP (pat
, 0, 1);
1008 return single_set (insn
);
1011 /* This is called with OUT_INSN an instruction setting a (base) register
1012 and IN_INSN a read or a write. Return 1 if these instructions together
1013 constitute a pipeline hazard.
1015 On the original architecture, a pipeline data hazard occurs when the Dest
1016 of one instruction becomes the SrcA for an immediately following READ or
1017 WRITE instruction with a non-zero index (indexing occurs at the decode
1018 stage and so a NOP must be inserted in-between for this to work).
1025 On the MCM, the non-zero index condition is lifted but the hazard is
1026 patched up by the hardware through the injection of wait states:
1031 We nevertheless try to schedule instructions around this. */
1034 gr5_hazard_bypass_p (rtx_insn
*out_insn
, rtx_insn
*in_insn
)
1036 rtx out_set
, in_set
, dest
, memexpr
;
1037 unsigned int out_reg
, in_reg
;
1039 /* A CALL is storage register class, but the link register is of no
1041 if (GET_CODE (out_insn
) == CALL_INSN
)
1044 out_set
= single_set_and_flags (out_insn
);
1045 dest
= SET_DEST (out_set
);
1047 /* Should be no stall/hazard if OUT_INSN is MEM := ???. This only
1048 occurs prior to reload. */
1049 if (GET_CODE (dest
) == MEM
)
1052 if (GET_CODE (dest
) == STRICT_LOW_PART
)
1053 dest
= XEXP (dest
, 0);
1054 if (GET_CODE (dest
) == SUBREG
)
1055 dest
= SUBREG_REG (dest
);
1056 out_reg
= REGNO (dest
);
1058 in_set
= single_set_and_flags (in_insn
);
1060 /* If IN_INSN is MEM := MEM, it's the source that counts. */
1061 if (GET_CODE (SET_SRC (in_set
)) == MEM
)
1062 memexpr
= XEXP (SET_SRC (in_set
), 0);
1064 memexpr
= XEXP (SET_DEST (in_set
), 0);
1066 if (GET_CODE (memexpr
) == PLUS
)
1068 memexpr
= XEXP (memexpr
, 0);
1069 if (GET_CODE (memexpr
) == SUBREG
)
1070 in_reg
= REGNO (SUBREG_REG (memexpr
));
1072 in_reg
= REGNO (memexpr
);
1074 if (in_reg
== out_reg
)
1077 else if (TARGET_MCM
)
1079 if (GET_CODE (memexpr
) == STRICT_LOW_PART
)
1080 memexpr
= XEXP (memexpr
, 0);
1081 if (GET_CODE (memexpr
) == SUBREG
)
1082 memexpr
= SUBREG_REG (memexpr
);
1083 in_reg
= REGNO (memexpr
);
1085 if (in_reg
== out_reg
)
1092 /* Return true if INSN is an empty asm instruction. */
1095 empty_asm_p (rtx insn
)
1097 rtx body
= PATTERN (insn
);
1100 if (GET_CODE (body
) == ASM_INPUT
)
1101 templ
= XSTR (body
, 0);
1102 else if (asm_noperands (body
) >= 0)
1103 templ
= decode_asm_operands (body
, NULL
, NULL
, NULL
, NULL
, NULL
);
1107 return (templ
&& templ
[0] == '\0');
1110 /* Insert a NOP immediately before INSN wherever there is a pipeline hazard.
1111 LAST_REG records the register set in the last insn and LAST_INSN_CALL
1112 records whether the last insn was a call insn. */
1115 gr5_avoid_hazard (rtx_insn
*insn
, unsigned int *last_reg
, bool *last_insn_call
)
1117 unsigned int dest_reg
= 0;
1120 switch (GET_CODE (insn
))
1124 *last_insn_call
= true;
1128 /* If this is an empty asm, just skip it. */
1129 if (!empty_asm_p (insn
))
1132 *last_insn_call
= false;
1137 /* If this is an empty asm, just skip it. */
1138 if (empty_asm_p (insn
))
1146 set
= single_set_and_flags (insn
);
1147 if (set
!= NULL_RTX
)
1149 rtx dest
= SET_DEST (set
);
1150 const bool double_p
= GET_MODE_SIZE (GET_MODE (dest
)) > UNITS_PER_WORD
;
1153 if (GET_CODE (SET_SRC (set
)) == MEM
)
1155 memrtx
= XEXP (SET_SRC (set
), 0);
1156 if (GET_CODE (dest
) == STRICT_LOW_PART
)
1157 dest
= XEXP (dest
, 0);
1159 dest_reg
= REGNO (dest
);
1161 /* If this is a DI or DF mode memory to register
1162 copy, then if rd = rs we get
1167 otherwise the order is
1174 unsigned int base_reg
;
1176 if (GET_CODE (memrtx
) == PLUS
)
1177 base_reg
= REGNO (XEXP (memrtx
, 0));
1179 base_reg
= REGNO (memrtx
);
1181 if (dest_reg
!= base_reg
)
1186 else if (GET_CODE (dest
) == MEM
)
1187 memrtx
= XEXP (dest
, 0);
1189 else if (GET_MODE_CLASS (GET_MODE (dest
)) != MODE_CC
)
1191 if (GET_CODE (dest
) == STRICT_LOW_PART
1192 ||GET_CODE (dest
) == ZERO_EXTRACT
)
1193 dest
= XEXP (dest
, 0);
1194 dest_reg
= REGNO (dest
);
1196 if (GET_CODE (SET_SRC (set
)) == REG
)
1198 unsigned int srcreg
= REGNO (SET_SRC (set
));
1200 /* Check for rs := rs, which will be deleted. */
1201 if (srcreg
== dest_reg
)
1204 /* In the case of a DI or DF mode move from register to
1205 register there is overlap if rd = rs + 1 in which case
1206 the order of the copies is reversed :
1211 if (double_p
&& dest_reg
!= srcreg
+ 1)
1216 /* If this is the delay slot of a call insn, any register it sets
1218 if (*last_insn_call
)
1221 /* If the previous insn sets the value of a register, and this insn
1222 uses a base register, check for the pipeline hazard where it is
1223 the same register in each case. */
1224 if (*last_reg
!= 0 && memrtx
!= NULL_RTX
)
1226 unsigned int reg
= 0;
1228 /* Check for an index (original architecture). */
1229 if (GET_CODE (memrtx
) == PLUS
)
1230 reg
= REGNO (XEXP (memrtx
, 0));
1232 /* Check for an MCM target or rs := [rs], in DI or DF mode. */
1233 else if (TARGET_MCM
|| (double_p
&& REGNO (memrtx
) == dest_reg
))
1234 reg
= REGNO (memrtx
);
1236 /* Remove any pipeline hazard by inserting a NOP. */
1237 if (reg
== *last_reg
)
1241 "inserting nop before insn %d\n", INSN_UID (insn
));
1242 emit_insn_after (gen_hazard_nop (), prev_active_insn (insn
));
1243 emit_insn_after (gen_blockage (), insn
);
1247 *last_reg
= dest_reg
;
1250 *last_insn_call
= false;
1253 /* Go through the instruction stream and insert nops where necessary to avoid
1254 pipeline hazards. There are two cases:
1256 1. On the original architecture, it is invalid to set the value of a
1257 (base) register and then use it in an address with a non-zero index
1258 in the next instruction.
1260 2. On the MCM, setting the value of a (base) register and then using
1261 it in address (including with zero index) in the next instruction
1262 will result in a pipeline stall of 3 cycles. */
1265 gr5_hazard_avoidance (void)
1267 unsigned int last_reg
= 0;
1268 bool last_insn_call
= false;
1271 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
1274 rtx pat
= PATTERN (insn
);
1276 if (GET_CODE (pat
) == SEQUENCE
)
1278 for (int i
= 0; i
< XVECLEN (pat
, 0); i
++)
1279 gr5_avoid_hazard (as_a
<rtx_insn
*> (XVECEXP (pat
, 0, i
)),
1280 &last_reg
, &last_insn_call
);
1283 else if (GET_CODE (insn
) == CALL_INSN
)
1285 /* This call is going to get a nop in its delay slot. */
1287 last_insn_call
= false;
1291 gr5_avoid_hazard (insn
, &last_reg
, &last_insn_call
);
1294 else if (GET_CODE (insn
) == BARRIER
)
1298 /* Perform a target-specific pass over the instruction stream. The compiler
1299 will run it at all optimization levels, just after the point at which it
1300 normally does delayed-branch scheduling. */
1305 if (visium_cpu
== PROCESSOR_GR5
)
1306 gr5_hazard_avoidance ();
1310 /* Return true if an argument must be passed by indirect reference. */
1313 visium_pass_by_reference (cumulative_args_t
, const function_arg_info
&arg
)
1315 tree type
= arg
.type
;
1316 return type
&& (AGGREGATE_TYPE_P (type
) || TREE_CODE (type
) == VECTOR_TYPE
);
1319 /* Define how arguments are passed.
1321 A range of general registers and floating registers is available
1322 for passing arguments. When the class of registers which an
1323 argument would normally use is exhausted, that argument, is passed
1324 in the overflow region of the stack. No argument is split between
1325 registers and stack.
1327 Arguments of type float or _Complex float go in FP registers if FP
1328 hardware is available. If there is no FP hardware, arguments of
1329 type float go in general registers. All other arguments are passed
1330 in general registers. */
1333 visium_function_arg (cumulative_args_t pcum_v
, const function_arg_info
&arg
)
1336 CUMULATIVE_ARGS
*ca
= get_cumulative_args (pcum_v
);
1338 size
= (GET_MODE_SIZE (arg
.mode
) + UNITS_PER_WORD
- 1) / UNITS_PER_WORD
;
1339 if (arg
.end_marker_p ())
1342 /* Scalar or complex single precision floating point arguments are returned
1343 in floating registers. */
1345 && ((GET_MODE_CLASS (arg
.mode
) == MODE_FLOAT
1346 && GET_MODE_SIZE (arg
.mode
) <= UNITS_PER_HWFPVALUE
)
1347 || (GET_MODE_CLASS (arg
.mode
) == MODE_COMPLEX_FLOAT
1348 && GET_MODE_SIZE (arg
.mode
) <= UNITS_PER_HWFPVALUE
* 2)))
1350 if (ca
->frcount
+ size
<= MAX_ARGS_IN_FP_REGISTERS
)
1351 return gen_rtx_REG (arg
.mode
, FP_ARG_FIRST
+ ca
->frcount
);
1356 if (ca
->grcount
+ size
<= MAX_ARGS_IN_GP_REGISTERS
)
1357 return gen_rtx_REG (arg
.mode
, ca
->grcount
+ GP_ARG_FIRST
);
1362 /* Update the summarizer variable pointed to by PCUM_V to advance past
1363 argument ARG. Once this is done, the variable CUM is suitable for
1364 analyzing the _following_ argument with visium_function_arg. */
1367 visium_function_arg_advance (cumulative_args_t pcum_v
,
1368 const function_arg_info
&arg
)
1370 int size
= (GET_MODE_SIZE (arg
.mode
) + UNITS_PER_WORD
- 1) / UNITS_PER_WORD
;
1372 CUMULATIVE_ARGS
*ca
= get_cumulative_args (pcum_v
);
1374 /* Scalar or complex single precision floating point arguments are returned
1375 in floating registers. */
1377 && ((GET_MODE_CLASS (arg
.mode
) == MODE_FLOAT
1378 && GET_MODE_SIZE (arg
.mode
) <= UNITS_PER_HWFPVALUE
)
1379 || (GET_MODE_CLASS (arg
.mode
) == MODE_COMPLEX_FLOAT
1380 && GET_MODE_SIZE (arg
.mode
) <= UNITS_PER_HWFPVALUE
* 2)))
1382 if (ca
->frcount
+ size
<= MAX_ARGS_IN_FP_REGISTERS
)
1383 ca
->frcount
+= size
;
1387 ca
->frcount
= MAX_ARGS_IN_FP_REGISTERS
;
1392 /* Everything else goes in a general register, if enough are
1394 if (ca
->grcount
+ size
<= MAX_ARGS_IN_GP_REGISTERS
)
1395 ca
->grcount
+= size
;
1399 ca
->grcount
= MAX_ARGS_IN_GP_REGISTERS
;
1404 ca
->stack_words
+= stack_size
;
1407 /* Specify whether to return the return value in memory. */
1410 visium_return_in_memory (const_tree type
, const_tree fntype ATTRIBUTE_UNUSED
)
1412 return (AGGREGATE_TYPE_P (type
) || TREE_CODE (type
) == VECTOR_TYPE
);
1415 /* Define how scalar values are returned. */
1418 visium_function_value_1 (machine_mode mode
)
1420 /* Scalar or complex single precision floating point values
1421 are returned in floating register f1. */
1423 && ((GET_MODE_CLASS (mode
) == MODE_FLOAT
1424 && GET_MODE_SIZE (mode
) <= UNITS_PER_HWFPVALUE
)
1425 || (GET_MODE_CLASS (mode
) == MODE_COMPLEX_FLOAT
1426 && GET_MODE_SIZE (mode
) <= UNITS_PER_HWFPVALUE
* 2)))
1427 return gen_rtx_REG (mode
, FP_RETURN_REGNUM
);
1429 /* All others are returned in r1. */
1430 return gen_rtx_REG (mode
, RETURN_REGNUM
);
1433 /* Return an RTX representing the place where a function returns or receives
1434 a value of data type RET_TYPE. */
1437 visium_function_value (const_tree ret_type
,
1438 const_tree fn_decl_or_type ATTRIBUTE_UNUSED
,
1439 bool outgoing ATTRIBUTE_UNUSED
)
1441 return visium_function_value_1 (TYPE_MODE (ret_type
));
1444 /* Return an RTX representing the place where the library function result will
1448 visium_libcall_value (machine_mode mode
, const_rtx fun ATTRIBUTE_UNUSED
)
1450 return visium_function_value_1 (mode
);
1453 /* Store the anonymous register arguments into the stack so that all the
1454 arguments appear to have been passed consecutively on the stack. */
1457 visium_setup_incoming_varargs (cumulative_args_t pcum_v
,
1458 const function_arg_info
&arg
,
1459 int *pretend_size ATTRIBUTE_UNUSED
,
1462 cumulative_args_t local_args_so_far
;
1463 CUMULATIVE_ARGS local_copy
;
1464 CUMULATIVE_ARGS
*locargs
;
1465 int gp_saved
, fp_saved
, size
;
1467 /* Create an internal cumulative_args_t pointer to internally define
1468 storage to ensure calling TARGET_FUNCTION_ARG_ADVANCE does not
1469 make global changes. */
1470 local_args_so_far
.p
= &local_copy
;
1471 locargs
= get_cumulative_args (pcum_v
);
1474 local_args_so_far
.magic
= CUMULATIVE_ARGS_MAGIC
;
1477 local_copy
.grcount
= locargs
->grcount
;
1478 local_copy
.frcount
= locargs
->frcount
;
1479 local_copy
.stack_words
= locargs
->stack_words
;
1481 /* The caller has advanced ARGS_SO_FAR up to, but not beyond, the last named
1482 argument. Advance a local copy of ARGS_SO_FAR past the last "real" named
1483 argument, to find out how many registers are left over. */
1484 TARGET_FUNCTION_ARG_ADVANCE (local_args_so_far
, arg
);
1486 /* Find how many registers we need to save. */
1487 locargs
= get_cumulative_args (local_args_so_far
);
1488 gp_saved
= MAX_ARGS_IN_GP_REGISTERS
- locargs
->grcount
;
1489 fp_saved
= (TARGET_FPU
? MAX_ARGS_IN_FP_REGISTERS
- locargs
->frcount
: 0);
1490 size
= (gp_saved
* UNITS_PER_WORD
) + (fp_saved
* UNITS_PER_HWFPVALUE
);
1492 if (!no_rtl
&& size
> 0)
1494 /* To avoid negative offsets, which are not valid addressing modes on
1495 the Visium, we create a base register for the pretend args. */
1498 plus_constant (Pmode
, virtual_incoming_args_rtx
, -size
));
1503 = gen_rtx_MEM (BLKmode
,
1504 plus_constant (Pmode
,
1506 fp_saved
* UNITS_PER_HWFPVALUE
));
1507 MEM_NOTRAP_P (mem
) = 1;
1508 set_mem_alias_set (mem
, get_varargs_alias_set ());
1509 move_block_from_reg (locargs
->grcount
+ GP_ARG_FIRST
, mem
, gp_saved
);
1514 rtx mem
= gen_rtx_MEM (BLKmode
, ptr
);
1515 MEM_NOTRAP_P (mem
) = 1;
1516 set_mem_alias_set (mem
, get_varargs_alias_set ());
1517 gcc_assert (UNITS_PER_WORD
== UNITS_PER_HWFPVALUE
);
1518 move_block_from_reg (locargs
->frcount
+ FP_ARG_FIRST
, mem
, fp_saved
);
1522 visium_reg_parm_save_area_size
= size
;
1525 /* Define the `__builtin_va_list' type for the ABI. */
1528 visium_build_builtin_va_list (void)
1530 tree f_ovfl
, f_gbase
, f_fbase
, f_gbytes
, f_fbytes
, record
;
1532 record
= (*lang_hooks
.types
.make_type
) (RECORD_TYPE
);
1533 f_ovfl
= build_decl (BUILTINS_LOCATION
, FIELD_DECL
,
1534 get_identifier ("__overflow_argptr"), ptr_type_node
);
1535 f_gbase
= build_decl (BUILTINS_LOCATION
, FIELD_DECL
,
1536 get_identifier ("__gpr_base"), ptr_type_node
);
1537 f_fbase
= build_decl (BUILTINS_LOCATION
, FIELD_DECL
,
1538 get_identifier ("__fpr_base"), ptr_type_node
);
1539 f_gbytes
= build_decl (BUILTINS_LOCATION
, FIELD_DECL
,
1540 get_identifier ("__gpr_bytes"),
1541 short_unsigned_type_node
);
1542 f_fbytes
= build_decl (BUILTINS_LOCATION
, FIELD_DECL
,
1543 get_identifier ("__fpr_bytes"),
1544 short_unsigned_type_node
);
1546 DECL_FIELD_CONTEXT (f_ovfl
) = record
;
1547 DECL_FIELD_CONTEXT (f_gbase
) = record
;
1548 DECL_FIELD_CONTEXT (f_fbase
) = record
;
1549 DECL_FIELD_CONTEXT (f_gbytes
) = record
;
1550 DECL_FIELD_CONTEXT (f_fbytes
) = record
;
1551 TYPE_FIELDS (record
) = f_ovfl
;
1552 TREE_CHAIN (f_ovfl
) = f_gbase
;
1553 TREE_CHAIN (f_gbase
) = f_fbase
;
1554 TREE_CHAIN (f_fbase
) = f_gbytes
;
1555 TREE_CHAIN (f_gbytes
) = f_fbytes
;
1556 layout_type (record
);
1561 /* Implement `va_start' for varargs and stdarg. */
1564 visium_va_start (tree valist
, rtx nextarg ATTRIBUTE_UNUSED
)
1566 const CUMULATIVE_ARGS
*ca
= &crtl
->args
.info
;
1567 int gp_saved
= MAX_ARGS_IN_GP_REGISTERS
- ca
->grcount
;
1568 int fp_saved
= (TARGET_FPU
? MAX_ARGS_IN_FP_REGISTERS
- ca
->frcount
: 0);
1569 int named_stack_size
= ca
->stack_words
* UNITS_PER_WORD
, offset
;
1570 tree f_ovfl
, f_gbase
, f_fbase
, f_gbytes
, f_fbytes
;
1571 tree ovfl
, gbase
, gbytes
, fbase
, fbytes
, t
;
1573 f_ovfl
= TYPE_FIELDS (va_list_type_node
);
1574 f_gbase
= TREE_CHAIN (f_ovfl
);
1575 f_fbase
= TREE_CHAIN (f_gbase
);
1576 f_gbytes
= TREE_CHAIN (f_fbase
);
1577 f_fbytes
= TREE_CHAIN (f_gbytes
);
1578 ovfl
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovfl
), valist
, f_ovfl
, NULL_TREE
);
1579 gbase
= build3 (COMPONENT_REF
, TREE_TYPE (f_gbase
), valist
, f_gbase
,
1581 fbase
= build3 (COMPONENT_REF
, TREE_TYPE (f_fbase
), valist
, f_fbase
,
1583 gbytes
= build3 (COMPONENT_REF
, TREE_TYPE (f_gbytes
), valist
, f_gbytes
,
1585 fbytes
= build3 (COMPONENT_REF
, TREE_TYPE (f_fbytes
), valist
, f_fbytes
,
1588 /* Store the stacked vararg pointer in the OVFL member. */
1589 t
= make_tree (TREE_TYPE (ovfl
), virtual_incoming_args_rtx
);
1590 t
= fold_build_pointer_plus_hwi (t
, named_stack_size
);
1591 t
= build2 (MODIFY_EXPR
, TREE_TYPE (ovfl
), ovfl
, t
);
1592 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
1594 /* Store the base address of the GPR save area into GBASE. */
1595 t
= make_tree (TREE_TYPE (gbase
), virtual_incoming_args_rtx
);
1596 offset
= MAX_ARGS_IN_GP_REGISTERS
* UNITS_PER_WORD
;
1597 t
= fold_build_pointer_plus_hwi (t
, -offset
);
1598 t
= build2 (MODIFY_EXPR
, TREE_TYPE (gbase
), gbase
, t
);
1599 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
1601 /* Store the base address of the FPR save area into FBASE. */
1604 t
= make_tree (TREE_TYPE (fbase
), virtual_incoming_args_rtx
);
1605 offset
= gp_saved
* UNITS_PER_WORD
1606 + MAX_ARGS_IN_FP_REGISTERS
* UNITS_PER_HWFPVALUE
;
1607 t
= fold_build_pointer_plus_hwi (t
, -offset
);
1608 t
= build2 (MODIFY_EXPR
, TREE_TYPE (fbase
), fbase
, t
);
1609 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
1612 /* Fill in the GBYTES member. */
1613 t
= build2 (MODIFY_EXPR
, TREE_TYPE (gbytes
), gbytes
,
1614 size_int (gp_saved
* UNITS_PER_WORD
));
1615 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
1617 /* Fill in the FBYTES member. */
1618 t
= build2 (MODIFY_EXPR
, TREE_TYPE (fbytes
),
1619 fbytes
, size_int (fp_saved
* UNITS_PER_HWFPVALUE
));
1620 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
1623 /* Implement `va_arg'. */
1626 visium_gimplify_va_arg (tree valist
, tree type
, gimple_seq
*pre_p
,
1629 tree f_ovfl
, f_gbase
, f_fbase
, f_gbytes
, f_fbytes
;
1630 tree ovfl
, base
, bytes
;
1631 HOST_WIDE_INT size
, rsize
;
1632 const bool by_reference_p
= pass_va_arg_by_reference (type
);
1633 const bool float_reg_arg_p
1634 = (TARGET_FPU
&& !by_reference_p
1635 && ((GET_MODE_CLASS (TYPE_MODE (type
)) == MODE_FLOAT
1636 && GET_MODE_SIZE (TYPE_MODE (type
)) <= UNITS_PER_HWFPVALUE
)
1637 || (GET_MODE_CLASS (TYPE_MODE (type
)) == MODE_COMPLEX_FLOAT
1638 && (GET_MODE_SIZE (TYPE_MODE (type
))
1639 <= UNITS_PER_HWFPVALUE
* 2))));
1640 const int max_save_area_size
1641 = (float_reg_arg_p
? MAX_ARGS_IN_FP_REGISTERS
* UNITS_PER_HWFPVALUE
1642 : MAX_ARGS_IN_GP_REGISTERS
* UNITS_PER_WORD
);
1644 tree lab_false
, lab_over
, addr
;
1645 tree ptrtype
= build_pointer_type (type
);
1649 t
= visium_gimplify_va_arg (valist
, ptrtype
, pre_p
, post_p
);
1650 return build_va_arg_indirect_ref (t
);
1653 size
= int_size_in_bytes (type
);
1654 rsize
= (size
+ UNITS_PER_WORD
- 1) & -UNITS_PER_WORD
;
1655 f_ovfl
= TYPE_FIELDS (va_list_type_node
);
1656 f_gbase
= TREE_CHAIN (f_ovfl
);
1657 f_fbase
= TREE_CHAIN (f_gbase
);
1658 f_gbytes
= TREE_CHAIN (f_fbase
);
1659 f_fbytes
= TREE_CHAIN (f_gbytes
);
1661 /* We maintain separate pointers and offsets for floating-point and
1662 general registers, but we need similar code in both cases.
1666 BYTES be the number of unused bytes in the register save area.
1667 BASE be the base address of the register save area.
1668 OFFS be the current offset into the register save area. Either
1669 MAX_ARGS_IN_GP_REGISTERS * UNITS_PER_WORD - bytes or
1670 MAX_ARGS_IN_FP_REGISTERS * UNITS_PER_HWFPVALUE - bytes
1671 depending upon whether the argument is in general or floating
1673 ADDR_RTX be the address of the argument.
1674 RSIZE be the size in bytes of the argument.
1675 OVFL be the pointer to the stack overflow area.
1677 The code we want is:
1679 1: if (bytes >= rsize)
1681 3: addr_rtx = base + offs;
1693 addr
= create_tmp_var (ptr_type_node
, "addr");
1694 lab_false
= create_artificial_label (UNKNOWN_LOCATION
);
1695 lab_over
= create_artificial_label (UNKNOWN_LOCATION
);
1696 if (float_reg_arg_p
)
1697 bytes
= build3 (COMPONENT_REF
, TREE_TYPE (f_fbytes
), unshare_expr (valist
),
1698 f_fbytes
, NULL_TREE
);
1700 bytes
= build3 (COMPONENT_REF
, TREE_TYPE (f_gbytes
), unshare_expr (valist
),
1701 f_gbytes
, NULL_TREE
);
1703 /* [1] Emit code to branch if bytes < rsize. */
1704 t
= fold_convert (TREE_TYPE (bytes
), size_int (rsize
));
1705 t
= build2 (LT_EXPR
, boolean_type_node
, bytes
, t
);
1706 u
= build1 (GOTO_EXPR
, void_type_node
, lab_false
);
1707 t
= build3 (COND_EXPR
, void_type_node
, t
, u
, NULL_TREE
);
1708 gimplify_and_add (t
, pre_p
);
1710 /* [3] Emit code for: addr_rtx = base + offs, where
1711 offs = max_save_area_size - bytes. */
1712 t
= fold_convert (sizetype
, bytes
);
1713 offs
= build2 (MINUS_EXPR
, sizetype
, size_int (max_save_area_size
), t
);
1714 if (float_reg_arg_p
)
1715 base
= build3 (COMPONENT_REF
, TREE_TYPE (f_fbase
), valist
, f_fbase
,
1718 base
= build3 (COMPONENT_REF
, TREE_TYPE (f_gbase
), valist
, f_gbase
,
1721 t
= build2 (POINTER_PLUS_EXPR
, TREE_TYPE (base
), base
, offs
);
1722 t
= build2 (MODIFY_EXPR
, void_type_node
, addr
, t
);
1723 gimplify_and_add (t
, pre_p
);
1725 /* [4] Emit code for: bytes -= rsize. */
1726 t
= fold_convert (TREE_TYPE (bytes
), size_int (rsize
));
1727 t
= build2 (MINUS_EXPR
, TREE_TYPE (bytes
), bytes
, t
);
1728 t
= build2 (MODIFY_EXPR
, TREE_TYPE (bytes
), bytes
, t
);
1729 gimplify_and_add (t
, pre_p
);
1731 /* [6] Emit code to branch over the else clause, then the label. */
1732 t
= build1 (GOTO_EXPR
, void_type_node
, lab_over
);
1733 gimplify_and_add (t
, pre_p
);
1734 t
= build1 (LABEL_EXPR
, void_type_node
, lab_false
);
1735 gimplify_and_add (t
, pre_p
);
1737 /* [8] Emit code for: bytes = 0. */
1738 t
= fold_convert (TREE_TYPE (bytes
), size_int (0));
1739 t
= build2 (MODIFY_EXPR
, TREE_TYPE (bytes
), unshare_expr (bytes
), t
);
1740 gimplify_and_add (t
, pre_p
);
1742 /* [9] Emit code for: addr_rtx = ovfl. */
1743 ovfl
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovfl
), valist
, f_ovfl
, NULL_TREE
);
1744 t
= build2 (MODIFY_EXPR
, void_type_node
, addr
, ovfl
);
1745 gimplify_and_add (t
, pre_p
);
1747 /* [10] Emit code for: ovfl += rsize. */
1748 t
= build2 (POINTER_PLUS_EXPR
, TREE_TYPE (ovfl
), ovfl
, size_int (rsize
));
1749 t
= build2 (MODIFY_EXPR
, TREE_TYPE (ovfl
), unshare_expr (ovfl
), t
);
1750 gimplify_and_add (t
, pre_p
);
1751 t
= build1 (LABEL_EXPR
, void_type_node
, lab_over
);
1752 gimplify_and_add (t
, pre_p
);
1754 /* Emit a big-endian correction if size < UNITS_PER_WORD. */
1755 if (size
< UNITS_PER_WORD
)
1757 t
= build2 (POINTER_PLUS_EXPR
, TREE_TYPE (addr
), addr
,
1758 size_int (UNITS_PER_WORD
- size
));
1759 t
= build2 (MODIFY_EXPR
, void_type_node
, addr
, t
);
1760 gimplify_and_add (t
, pre_p
);
1763 addr
= fold_convert (ptrtype
, addr
);
1765 return build_va_arg_indirect_ref (addr
);
1768 /* Return true if OP is an offset suitable for use as a displacement in the
1769 address of a memory access in mode MODE. */
1772 rtx_ok_for_offset_p (machine_mode mode
, rtx op
)
1774 if (!CONST_INT_P (op
) || INTVAL (op
) < 0)
1780 return INTVAL (op
) <= 31;
1783 return (INTVAL (op
) % 2) == 0 && INTVAL (op
) < 63;
1787 return (INTVAL (op
) % 4) == 0 && INTVAL (op
) < 127;
1791 return (INTVAL (op
) % 4) == 0 && INTVAL (op
) < 123;
1798 /* Return whether X is a legitimate memory address for a memory operand
1801 Legitimate addresses are defined in two variants: a strict variant
1802 and a non-strict one. The STRICT parameter chooses which variant
1803 is desired by the caller.
1805 The strict variant is used in the reload pass. It must be defined
1806 so that any pseudo-register that has not been allocated a hard
1807 register is considered a memory reference. This is because in
1808 contexts where some kind of register is required, a
1809 pseudo-register with no hard register must be rejected. For
1810 non-hard registers, the strict variant should look up the
1811 `reg_renumber' array; it should then proceed using the hard
1812 register number in the array, or treat the pseudo as a memory
1813 reference if the array holds `-1'.
1815 The non-strict variant is used in other passes. It must be
1816 defined to accept all pseudo-registers in every context where some
1817 kind of register is required. */
1820 visium_legitimate_address_p (machine_mode mode
, rtx x
, bool strict
)
1825 /* If X is base+disp, check that we have an appropriate offset. */
1826 if (GET_CODE (x
) == PLUS
)
1828 if (!rtx_ok_for_offset_p (mode
, XEXP (x
, 1)))
1835 /* Now check the base: it must be either a register or a subreg thereof. */
1836 if (GET_CODE (base
) == SUBREG
)
1837 base
= SUBREG_REG (base
);
1841 regno
= REGNO (base
);
1843 /* For the strict variant, the register must be REGNO_OK_FOR_BASE_P. */
1845 return REGNO_OK_FOR_BASE_P (regno
);
1847 /* For the non-strict variant, the register may also be a pseudo. */
1848 return BASE_REGISTER_P (regno
) || regno
>= FIRST_PSEUDO_REGISTER
;
1851 /* Try machine-dependent ways of modifying an illegitimate address
1852 to be legitimate. If we find one, return the new, valid address.
1853 This macro is used in only one place: `memory_address' in explow.cc.
1855 OLDX is the address as it was before break_out_memory_refs was called.
1856 In some cases it is useful to look at this to decide what needs to be done.
1858 MODE and WIN are passed so that this macro can use
1859 GO_IF_LEGITIMATE_ADDRESS.
1861 It is always safe for this macro to do nothing. It exists to recognize
1862 opportunities to optimize the output.
1866 memory (reg + <out of range int>)
1870 base_int = <out of range int> & ~mask
1871 ptr_reg = reg + base_int
1872 memory (ptr_reg + <out of range int> - base_int)
1874 Thus ptr_reg is a base register for a range of addresses,
1875 which should help CSE.
1877 For a 1 byte reference mask is 0x1f
1878 for a 2 byte reference mask is 0x3f
1879 For a 4 byte reference mask is 0x7f
1881 This reflects the indexing range of the processor.
1883 For a > 4 byte reference the mask is 0x7f provided all of the words
1884 can be accessed with the base address obtained. Otherwise a mask
1887 On rare occasions an unaligned base register value with an
1888 unaligned offset is generated. Unaligned offsets are left alone for
1892 visium_legitimize_address (rtx x
, rtx oldx ATTRIBUTE_UNUSED
,
1895 if (GET_CODE (x
) == PLUS
1896 && GET_CODE (XEXP (x
, 1)) == CONST_INT
1897 && GET_CODE (XEXP (x
, 0)) == REG
&& mode
!= BLKmode
)
1899 int offset
= INTVAL (XEXP (x
, 1));
1900 int size
= GET_MODE_SIZE (mode
);
1901 int mask
= (size
== 1 ? 0x1f : (size
== 2 ? 0x3f : 0x7f));
1902 int mask1
= (size
== 1 ? 0 : (size
== 2 ? 1 : 3));
1903 int offset_base
= offset
& ~mask
;
1905 /* Check that all of the words can be accessed. */
1906 if (size
> 4 && size
+ offset
- offset_base
> 0x80)
1907 offset_base
= offset
& ~0x3f;
1908 if (offset_base
!= 0 && offset_base
!= offset
&& (offset
& mask1
) == 0)
1910 rtx ptr_reg
= force_reg (Pmode
,
1911 gen_rtx_PLUS (Pmode
,
1913 GEN_INT (offset_base
)));
1915 return plus_constant (Pmode
, ptr_reg
, offset
- offset_base
);
1922 /* Perform a similar function to visium_legitimize_address, but this time
1923 for reload. Generating new registers is not an option here. Parts
1924 that need reloading are indicated by calling push_reload. */
1927 visium_legitimize_reload_address (rtx x
, machine_mode mode
, int opnum
,
1928 int type
, int ind ATTRIBUTE_UNUSED
)
1930 rtx newrtx
, tem
= NULL_RTX
;
1932 if (mode
== BLKmode
)
1935 if (optimize
&& GET_CODE (x
) == PLUS
)
1936 tem
= simplify_binary_operation (PLUS
, GET_MODE (x
), XEXP (x
, 0),
1939 newrtx
= tem
? tem
: x
;
1940 if (GET_CODE (newrtx
) == PLUS
1941 && GET_CODE (XEXP (newrtx
, 1)) == CONST_INT
1942 && GET_CODE (XEXP (newrtx
, 0)) == REG
1943 && BASE_REGISTER_P (REGNO (XEXP (newrtx
, 0))))
1945 int offset
= INTVAL (XEXP (newrtx
, 1));
1946 int size
= GET_MODE_SIZE (mode
);
1947 int mask
= (size
== 1 ? 0x1f : (size
== 2 ? 0x3f : 0x7f));
1948 int mask1
= (size
== 1 ? 0 : (size
== 2 ? 1 : 3));
1949 int offset_base
= offset
& ~mask
;
1951 /* Check that all of the words can be accessed. */
1952 if (size
> 4 && size
+ offset
- offset_base
> 0x80)
1953 offset_base
= offset
& ~0x3f;
1955 if (offset_base
&& (offset
& mask1
) == 0)
1957 rtx temp
= gen_rtx_PLUS (Pmode
,
1958 XEXP (newrtx
, 0), GEN_INT (offset_base
));
1960 x
= gen_rtx_PLUS (Pmode
, temp
, GEN_INT (offset
- offset_base
));
1961 push_reload (XEXP (x
, 0), 0, &XEXP (x
, 0), 0,
1962 BASE_REG_CLASS
, Pmode
, VOIDmode
, 0, 0, opnum
,
1963 (enum reload_type
) type
);
1971 /* Return the cost of moving data of mode MODE from a register in class FROM to
1972 one in class TO. A value of 2 is the default; other values are interpreted
1973 relative to that. */
1976 visium_register_move_cost (machine_mode mode
, reg_class_t from
,
1979 const int numwords
= (GET_MODE_SIZE (mode
) <= UNITS_PER_WORD
) ? 1 : 2;
1981 if (from
== MDB
|| to
== MDB
)
1983 else if (from
== MDC
|| to
== MDC
|| (from
== FP_REGS
) != (to
== FP_REGS
))
1984 return 4 * numwords
;
1986 return 2 * numwords
;
1989 /* Return the cost of moving data of mode MODE between a register of class
1990 CLASS and memory. IN is zero if the value is to be written to memory,
1991 non-zero if it is to be read in. This cost is relative to those in
1992 visium_register_move_cost. */
1995 visium_memory_move_cost (machine_mode mode
,
1996 reg_class_t to ATTRIBUTE_UNUSED
,
1999 /* Moving data in can be from PROM and this is expensive. */
2002 if (GET_MODE_SIZE (mode
) <= UNITS_PER_WORD
)
2008 /* Moving data out is mostly to RAM and should be cheaper. */
2011 if (GET_MODE_SIZE (mode
) <= UNITS_PER_WORD
)
2018 /* Return the relative costs of expression X. */
2021 visium_rtx_costs (rtx x
, machine_mode mode
, int outer_code ATTRIBUTE_UNUSED
,
2022 int opno ATTRIBUTE_UNUSED
, int *total
,
2023 bool speed ATTRIBUTE_UNUSED
)
2025 int code
= GET_CODE (x
);
2030 /* Small integers are as cheap as registers. 4-byte values can
2031 be fetched as immediate constants - let's give that the cost
2032 of an extra insn. */
2033 *total
= COSTS_N_INSNS (!satisfies_constraint_J (x
));
2039 *total
= COSTS_N_INSNS (2);
2045 split_double (x
, &high
, &low
);
2048 (!satisfies_constraint_J (high
) + !satisfies_constraint_J (low
));
2053 *total
= COSTS_N_INSNS (3);
2061 *total
= COSTS_N_INSNS (64);
2063 *total
= COSTS_N_INSNS (32);
2069 /* DImode operations are performed directly on the ALU. */
2071 *total
= COSTS_N_INSNS (2);
2073 *total
= COSTS_N_INSNS (1);
2079 /* DImode operations are performed on the EAM instead. */
2081 *total
= COSTS_N_INSNS (3);
2083 *total
= COSTS_N_INSNS (1);
2087 /* This matches the btst pattern. */
2088 if (GET_CODE (XEXP (x
, 0)) == ZERO_EXTRACT
2089 && XEXP (x
, 1) == const0_rtx
2090 && XEXP (XEXP (x
, 0), 1) == const1_rtx
2091 && satisfies_constraint_K (XEXP (XEXP (x
, 0), 2)))
2092 *total
= COSTS_N_INSNS (1);
2100 /* Split a double move of OPERANDS in MODE. */
2103 visium_split_double_move (rtx
*operands
, machine_mode mode
)
2107 /* Check register to register with overlap. */
2108 if (GET_CODE (operands
[0]) == REG
2109 && GET_CODE (operands
[1]) == REG
2110 && REGNO (operands
[0]) == REGNO (operands
[1]) + 1)
2113 /* Check memory to register where the base reg overlaps the destination. */
2114 if (GET_CODE (operands
[0]) == REG
&& GET_CODE (operands
[1]) == MEM
)
2116 rtx op
= XEXP (operands
[1], 0);
2118 if (GET_CODE (op
) == SUBREG
)
2119 op
= SUBREG_REG (op
);
2121 if (GET_CODE (op
) == REG
&& REGNO (op
) == REGNO (operands
[0]))
2124 if (GET_CODE (op
) == PLUS
)
2126 rtx x
= XEXP (op
, 0);
2127 rtx y
= XEXP (op
, 1);
2129 if (GET_CODE (x
) == REG
&& REGNO (x
) == REGNO (operands
[0]))
2132 if (GET_CODE (y
) == REG
&& REGNO (y
) == REGNO (operands
[0]))
2139 operands
[2] = operand_subword (operands
[0], 1, 1, mode
);
2140 operands
[3] = operand_subword (operands
[1], 1, 1, mode
);
2141 operands
[4] = operand_subword (operands
[0], 0, 1, mode
);
2142 operands
[5] = operand_subword (operands
[1], 0, 1, mode
);
2146 operands
[2] = operand_subword (operands
[0], 0, 1, mode
);
2147 operands
[3] = operand_subword (operands
[1], 0, 1, mode
);
2148 operands
[4] = operand_subword (operands
[0], 1, 1, mode
);
2149 operands
[5] = operand_subword (operands
[1], 1, 1, mode
);
2153 /* Split a double addition or subtraction of operands. */
2156 visium_split_double_add (enum rtx_code code
, rtx op0
, rtx op1
, rtx op2
)
2158 rtx op3
= gen_lowpart (SImode
, op0
);
2159 rtx op4
= gen_lowpart (SImode
, op1
);
2161 rtx op6
= gen_highpart (SImode
, op0
);
2162 rtx op7
= (op1
== const0_rtx
? op1
: gen_highpart (SImode
, op1
));
2166 /* If operand #2 is a small constant, then its high part is null. */
2167 if (CONST_INT_P (op2
))
2169 HOST_WIDE_INT val
= INTVAL (op2
);
2173 code
= (code
== MINUS
? PLUS
: MINUS
);
2177 op5
= gen_int_mode (val
, SImode
);
2182 op5
= gen_lowpart (SImode
, op2
);
2183 op8
= gen_highpart (SImode
, op2
);
2186 if (op4
== const0_rtx
)
2187 pat
= gen_negsi2_insn_set_carry (op3
, op5
);
2188 else if (code
== MINUS
)
2189 pat
= gen_subsi3_insn_set_carry (op3
, op4
, op5
);
2191 pat
= gen_addsi3_insn_set_carry (op3
, op4
, op5
);
2194 /* This is the plus_[plus_]sltu_flags or minus_[minus_]sltu_flags pattern. */
2195 if (op8
== const0_rtx
)
2198 x
= gen_rtx_fmt_ee (code
, SImode
, op7
, op8
);
2199 flags
= gen_rtx_REG (CCCmode
, FLAGS_REGNUM
);
2200 x
= gen_rtx_fmt_ee (code
, SImode
, x
, gen_rtx_LTU (SImode
, flags
, const0_rtx
));
2201 pat
= gen_rtx_PARALLEL (VOIDmode
, rtvec_alloc (2));
2202 XVECEXP (pat
, 0, 0) = gen_rtx_SET (op6
, x
);
2203 flags
= gen_rtx_REG (CCmode
, FLAGS_REGNUM
);
2204 XVECEXP (pat
, 0, 1) = gen_rtx_CLOBBER (VOIDmode
, flags
);
2207 visium_flags_exposed
= true;
2210 /* Expand a copysign of OPERANDS in MODE. */
2213 visium_expand_copysign (rtx
*operands
, machine_mode mode
)
2215 rtx op0
= operands
[0];
2216 rtx op1
= operands
[1];
2217 rtx op2
= operands
[2];
2218 rtx mask
= force_reg (SImode
, GEN_INT (0x7fffffff));
2221 /* We manually handle SFmode because the abs and neg instructions of
2222 the FPU on the MCM have a non-standard behavior wrt NaNs. */
2223 gcc_assert (mode
== SFmode
);
2225 /* First get all the non-sign bits of op1. */
2226 if (GET_CODE (op1
) == CONST_DOUBLE
)
2228 if (real_isneg (CONST_DOUBLE_REAL_VALUE (op1
)))
2229 op1
= simplify_unary_operation (ABS
, mode
, op1
, mode
);
2230 if (op1
!= CONST0_RTX (mode
))
2233 REAL_VALUE_TO_TARGET_SINGLE (*CONST_DOUBLE_REAL_VALUE (op1
), l
);
2234 op1
= force_reg (SImode
, gen_int_mode (l
, SImode
));
2239 op1
= copy_to_mode_reg (SImode
, gen_lowpart (SImode
, op1
));
2240 op1
= force_reg (SImode
, gen_rtx_AND (SImode
, op1
, mask
));
2243 /* Then get the sign bit of op2. */
2244 mask
= force_reg (SImode
, gen_rtx_NOT (SImode
, mask
));
2245 op2
= copy_to_mode_reg (SImode
, gen_lowpart (SImode
, op2
));
2246 op2
= force_reg (SImode
, gen_rtx_AND (SImode
, op2
, mask
));
2248 /* Finally OR the two values. */
2249 if (op1
== CONST0_RTX (SFmode
))
2252 x
= force_reg (SImode
, gen_rtx_IOR (SImode
, op1
, op2
));
2254 /* And move the result to the destination. */
2255 emit_insn (gen_rtx_SET (op0
, gen_lowpart (SFmode
, x
)));
2258 /* Expand a cstore of OPERANDS in MODE for EQ/NE/LTU/GTU/GEU/LEU. We generate
2259 the result in the C flag and use the ADC/SUBC instructions to write it into
2260 the destination register.
2262 It would also be possible to implement support for LT/GT/LE/GE by means of
2263 the RFLAG instruction followed by some shifts, but this can pessimize the
2267 visium_expand_int_cstore (rtx
*operands
, machine_mode mode
)
2269 enum rtx_code code
= GET_CODE (operands
[1]);
2270 rtx op0
= operands
[0], op1
= operands
[2], op2
= operands
[3], sltu
;
2271 bool reverse
= false;
2277 /* We use a special comparison to get the result in the C flag. */
2278 if (op2
!= const0_rtx
)
2279 op1
= force_reg (mode
, gen_rtx_XOR (mode
, op1
, op2
));
2280 op1
= gen_rtx_NOT (mode
, op1
);
2288 /* The result is naturally in the C flag modulo a couple of tricks. */
2289 code
= reverse_condition (code
);
2292 /* ... fall through ... */
2308 /* We need either a single ADC or a SUBC and a PLUS. */
2309 sltu
= gen_rtx_LTU (SImode
, op1
, op2
);
2313 rtx tmp
= copy_to_mode_reg (SImode
, gen_rtx_NEG (SImode
, sltu
));
2314 emit_insn (gen_add3_insn (op0
, tmp
, const1_rtx
));
2317 emit_insn (gen_rtx_SET (op0
, sltu
));
2320 /* Expand a cstore of OPERANDS in MODE for LT/GT/UNGE/UNLE. We generate the
2321 result in the C flag and use the ADC/SUBC instructions to write it into
2322 the destination register. */
2325 visium_expand_fp_cstore (rtx
*operands
,
2326 machine_mode mode ATTRIBUTE_UNUSED
)
2328 enum rtx_code code
= GET_CODE (operands
[1]);
2329 rtx op0
= operands
[0], op1
= operands
[2], op2
= operands
[3], slt
;
2330 bool reverse
= false;
2336 /* The result is naturally in the C flag modulo a couple of tricks. */
2337 code
= reverse_condition_maybe_unordered (code
);
2340 /* ... fall through ... */
2356 /* We need either a single ADC or a SUBC and a PLUS. */
2357 slt
= gen_rtx_LT (SImode
, op1
, op2
);
2361 rtx tmp
= copy_to_mode_reg (SImode
, gen_rtx_NEG (SImode
, slt
));
2362 emit_insn (gen_add3_insn (op0
, tmp
, const1_rtx
));
2365 emit_insn (gen_rtx_SET (op0
, slt
));
2368 /* Split a compare-and-store with CODE, operands OP2 and OP3, combined with
2369 operation with OP_CODE, operands OP0 and OP1. */
2372 visium_split_cstore (enum rtx_code op_code
, rtx op0
, rtx op1
,
2373 enum rtx_code code
, rtx op2
, rtx op3
)
2375 machine_mode cc_mode
= visium_select_cc_mode (code
, op2
, op3
);
2377 /* If a FP cstore was reversed, then it was originally UNGE/UNLE. */
2378 if (cc_mode
== CCFPEmode
&& (op_code
== NEG
|| op_code
== MINUS
))
2381 rtx flags
= gen_rtx_REG (cc_mode
, FLAGS_REGNUM
);
2382 rtx x
= gen_rtx_COMPARE (cc_mode
, op2
, op3
);
2383 x
= gen_rtx_SET (flags
, x
);
2386 x
= gen_rtx_fmt_ee (code
, SImode
, flags
, const0_rtx
);
2392 x
= gen_rtx_NEG (SImode
, x
);
2396 x
= gen_rtx_fmt_ee (op_code
, SImode
, op1
, x
);
2402 rtx pat
= gen_rtx_PARALLEL (VOIDmode
, rtvec_alloc (2));
2403 XVECEXP (pat
, 0, 0) = gen_rtx_SET (op0
, x
);
2404 flags
= gen_rtx_REG (CCmode
, FLAGS_REGNUM
);
2405 XVECEXP (pat
, 0, 1) = gen_rtx_CLOBBER (VOIDmode
, flags
);
2408 visium_flags_exposed
= true;
2411 /* Generate a call to a library function to move BYTES_RTX bytes from SRC with
2412 address SRC_REG to DST with address DST_REG in 4-byte chunks. */
2415 expand_block_move_4 (rtx dst
, rtx dst_reg
, rtx src
, rtx src_reg
, rtx bytes_rtx
)
2417 unsigned HOST_WIDE_INT bytes
= UINTVAL (bytes_rtx
);
2418 unsigned int rem
= bytes
% 4;
2425 emit_move_insn (regno_reg_rtx
[1], dst_reg
);
2426 emit_move_insn (regno_reg_rtx
[2], src_reg
);
2427 emit_move_insn (regno_reg_rtx
[3], bytes_rtx
);
2429 insn
= gen_rtx_PARALLEL (VOIDmode
, rtvec_alloc (8));
2430 XVECEXP (insn
, 0, 0)
2431 = gen_rtx_SET (replace_equiv_address_nv (dst
, regno_reg_rtx
[1]),
2432 replace_equiv_address_nv (src
, regno_reg_rtx
[2]));
2433 XVECEXP (insn
, 0, 1) = gen_rtx_USE (VOIDmode
, regno_reg_rtx
[3]);
2434 for (i
= 1; i
<= 6; i
++)
2435 XVECEXP (insn
, 0, 1 + i
)
2436 = gen_rtx_CLOBBER (VOIDmode
, regno_reg_rtx
[i
]);
2440 emit_library_call (long_int_memcpy_libfunc
, LCT_NORMAL
, VOIDmode
,
2443 convert_to_mode (TYPE_MODE (sizetype
),
2444 GEN_INT (bytes
>> 2),
2445 TYPE_UNSIGNED (sizetype
)),
2446 TYPE_MODE (sizetype
));
2450 dst
= replace_equiv_address_nv (dst
, dst_reg
);
2451 src
= replace_equiv_address_nv (src
, src_reg
);
2456 emit_move_insn (adjust_address_nv (dst
, HImode
, bytes
),
2457 adjust_address_nv (src
, HImode
, bytes
));
2463 emit_move_insn (adjust_address_nv (dst
, QImode
, bytes
),
2464 adjust_address_nv (src
, QImode
, bytes
));
2467 /* Generate a call to a library function to move BYTES_RTX bytes from SRC with
2468 address SRC_REG to DST with address DST_REG in 2-bytes chunks. */
2471 expand_block_move_2 (rtx dst
, rtx dst_reg
, rtx src
, rtx src_reg
, rtx bytes_rtx
)
2473 unsigned HOST_WIDE_INT bytes
= UINTVAL (bytes_rtx
);
2474 unsigned int rem
= bytes
% 2;
2476 emit_library_call (wrd_memcpy_libfunc
, LCT_NORMAL
, VOIDmode
,
2479 convert_to_mode (TYPE_MODE (sizetype
),
2480 GEN_INT (bytes
>> 1),
2481 TYPE_UNSIGNED (sizetype
)),
2482 TYPE_MODE (sizetype
));
2486 dst
= replace_equiv_address_nv (dst
, dst_reg
);
2487 src
= replace_equiv_address_nv (src
, src_reg
);
2490 emit_move_insn (adjust_address_nv (dst
, QImode
, bytes
),
2491 adjust_address_nv (src
, QImode
, bytes
));
2494 /* Generate a call to a library function to move BYTES_RTX bytes from address
2495 SRC_REG to address DST_REG in 1-byte chunks. */
2498 expand_block_move_1 (rtx dst_reg
, rtx src_reg
, rtx bytes_rtx
)
2500 emit_library_call (byt_memcpy_libfunc
, LCT_NORMAL
, VOIDmode
,
2503 convert_to_mode (TYPE_MODE (sizetype
),
2505 TYPE_UNSIGNED (sizetype
)),
2506 TYPE_MODE (sizetype
));
2509 /* Generate a call to a library function to set BYTES_RTX bytes of DST with
2510 address DST_REG to VALUE_RTX in 4-byte chunks. */
2513 expand_block_set_4 (rtx dst
, rtx dst_reg
, rtx value_rtx
, rtx bytes_rtx
)
2515 unsigned HOST_WIDE_INT bytes
= UINTVAL (bytes_rtx
);
2516 unsigned int rem
= bytes
% 4;
2518 value_rtx
= convert_to_mode (Pmode
, value_rtx
, 1);
2519 emit_library_call (long_int_memset_libfunc
, LCT_NORMAL
, VOIDmode
,
2522 convert_to_mode (TYPE_MODE (sizetype
),
2523 GEN_INT (bytes
>> 2),
2524 TYPE_UNSIGNED (sizetype
)),
2525 TYPE_MODE (sizetype
));
2529 dst
= replace_equiv_address_nv (dst
, dst_reg
);
2534 if (CONST_INT_P (value_rtx
))
2536 const unsigned HOST_WIDE_INT value
= UINTVAL (value_rtx
) & 0xff;
2537 emit_move_insn (adjust_address_nv (dst
, HImode
, bytes
),
2538 gen_int_mode ((value
<< 8) | value
, HImode
));
2542 rtx temp
= convert_to_mode (QImode
, value_rtx
, 1);
2543 emit_move_insn (adjust_address_nv (dst
, QImode
, bytes
), temp
);
2544 emit_move_insn (adjust_address_nv (dst
, QImode
, bytes
+ 1), temp
);
2551 emit_move_insn (adjust_address_nv (dst
, QImode
, bytes
),
2552 convert_to_mode (QImode
, value_rtx
, 1));
2555 /* Generate a call to a library function to set BYTES_RTX bytes of DST with
2556 address DST_REG to VALUE_RTX in 2-byte chunks. */
2559 expand_block_set_2 (rtx dst
, rtx dst_reg
, rtx value_rtx
, rtx bytes_rtx
)
2561 unsigned HOST_WIDE_INT bytes
= UINTVAL (bytes_rtx
);
2562 unsigned int rem
= bytes
% 2;
2564 value_rtx
= convert_to_mode (Pmode
, value_rtx
, 1);
2565 emit_library_call (wrd_memset_libfunc
, LCT_NORMAL
, VOIDmode
,
2568 convert_to_mode (TYPE_MODE (sizetype
),
2569 GEN_INT (bytes
>> 1),
2570 TYPE_UNSIGNED (sizetype
)),
2571 TYPE_MODE (sizetype
));
2575 dst
= replace_equiv_address_nv (dst
, dst_reg
);
2578 emit_move_insn (adjust_address_nv (dst
, QImode
, bytes
),
2579 convert_to_mode (QImode
, value_rtx
, 1));
2582 /* Generate a call to a library function to set BYTES_RTX bytes at address
2583 DST_REG to VALUE_RTX in 1-byte chunks. */
2586 expand_block_set_1 (rtx dst_reg
, rtx value_rtx
, rtx bytes_rtx
)
2588 value_rtx
= convert_to_mode (Pmode
, value_rtx
, 1);
2589 emit_library_call (byt_memset_libfunc
, LCT_NORMAL
, VOIDmode
,
2592 convert_to_mode (TYPE_MODE (sizetype
),
2594 TYPE_UNSIGNED (sizetype
)),
2595 TYPE_MODE (sizetype
));
2598 /* Expand string/block move operations.
2600 operands[0] is the pointer to the destination.
2601 operands[1] is the pointer to the source.
2602 operands[2] is the number of bytes to move.
2603 operands[3] is the alignment.
2605 Return 1 upon success, 0 otherwise. */
2608 visium_expand_block_move (rtx
*operands
)
2610 rtx dst
= operands
[0];
2611 rtx src
= operands
[1];
2612 rtx bytes_rtx
= operands
[2];
2613 rtx align_rtx
= operands
[3];
2614 const int align
= INTVAL (align_rtx
);
2615 rtx dst_reg
, src_reg
;
2616 tree dst_expr
, src_expr
;
2618 /* We only handle a fixed number of bytes for now. */
2619 if (!CONST_INT_P (bytes_rtx
) || INTVAL (bytes_rtx
) <= 0)
2622 /* Copy the addresses into scratch registers. */
2623 dst_reg
= copy_addr_to_reg (XEXP (dst
, 0));
2624 src_reg
= copy_addr_to_reg (XEXP (src
, 0));
2626 /* Move the data with the appropriate granularity. */
2628 expand_block_move_4 (dst
, dst_reg
, src
, src_reg
, bytes_rtx
);
2629 else if (align
>= 2)
2630 expand_block_move_2 (dst
, dst_reg
, src
, src_reg
, bytes_rtx
);
2632 expand_block_move_1 (dst_reg
, src_reg
, bytes_rtx
);
2634 /* Since DST and SRC are passed to a libcall, mark the corresponding
2635 tree EXPR as addressable. */
2636 dst_expr
= MEM_EXPR (dst
);
2637 src_expr
= MEM_EXPR (src
);
2639 mark_addressable (dst_expr
);
2641 mark_addressable (src_expr
);
2646 /* Expand string/block set operations.
2648 operands[0] is the pointer to the destination.
2649 operands[1] is the number of bytes to set.
2650 operands[2] is the source value.
2651 operands[3] is the alignment.
2653 Return 1 upon success, 0 otherwise. */
2656 visium_expand_block_set (rtx
*operands
)
2658 rtx dst
= operands
[0];
2659 rtx bytes_rtx
= operands
[1];
2660 rtx value_rtx
= operands
[2];
2661 rtx align_rtx
= operands
[3];
2662 const int align
= INTVAL (align_rtx
);
2666 /* We only handle a fixed number of bytes for now. */
2667 if (!CONST_INT_P (bytes_rtx
) || INTVAL (bytes_rtx
) <= 0)
2670 /* Copy the address into a scratch register. */
2671 dst_reg
= copy_addr_to_reg (XEXP (dst
, 0));
2673 /* Set the data with the appropriate granularity. */
2675 expand_block_set_4 (dst
, dst_reg
, value_rtx
, bytes_rtx
);
2676 else if (align
>= 2)
2677 expand_block_set_2 (dst
, dst_reg
, value_rtx
, bytes_rtx
);
2679 expand_block_set_1 (dst_reg
, value_rtx
, bytes_rtx
);
2681 /* Since DST is passed to a libcall, mark the corresponding
2682 tree EXPR as addressable. */
2683 dst_expr
= MEM_EXPR (dst
);
2685 mark_addressable (dst_expr
);
2690 /* Initialize a trampoline. M_TRAMP is an RTX for the memory block for the
2691 trampoline, FNDECL is the FUNCTION_DECL for the nested function and
2692 STATIC_CHAIN is an RTX for the static chain value that should be passed
2693 to the function when it is called. */
2696 visium_trampoline_init (rtx m_tramp
, tree fndecl
, rtx static_chain
)
2698 rtx fnaddr
= XEXP (DECL_RTL (fndecl
), 0);
2699 rtx addr
= XEXP (m_tramp
, 0);
2701 /* The trampoline initialization sequence is:
2703 moviu r9,%u FUNCTION
2704 movil r9,%l FUNCTION
2710 We don't use r0 as the destination register of the branch because we want
2711 the Branch Pre-decode Logic of the GR6 to use the Address Load Array to
2712 predict the branch target. */
2714 emit_move_insn (gen_rtx_MEM (SImode
, plus_constant (Pmode
, addr
, 0)),
2715 plus_constant (SImode
,
2716 expand_shift (RSHIFT_EXPR
, SImode
, fnaddr
,
2720 emit_move_insn (gen_rtx_MEM (SImode
, plus_constant (Pmode
, addr
, 4)),
2721 plus_constant (SImode
,
2722 expand_and (SImode
, fnaddr
, GEN_INT (0xffff),
2726 if (visium_cpu
== PROCESSOR_GR6
)
2728 /* For the GR6, the BRA insn must be aligned on a 64-bit boundary. */
2729 gcc_assert (TRAMPOLINE_ALIGNMENT
>= 64);
2730 emit_move_insn (gen_rtx_MEM (SImode
, plus_constant (Pmode
, addr
, 12)),
2731 gen_int_mode (0, SImode
));
2734 emit_move_insn (gen_rtx_MEM (SImode
, plus_constant (Pmode
, addr
, 8)),
2735 plus_constant (SImode
,
2736 expand_shift (RSHIFT_EXPR
, SImode
,
2741 emit_move_insn (gen_rtx_MEM (SImode
, plus_constant (Pmode
, addr
, 12)),
2742 gen_int_mode (0xff892404, SImode
));
2744 emit_move_insn (gen_rtx_MEM (SImode
, plus_constant (Pmode
, addr
, 16)),
2745 plus_constant (SImode
,
2746 expand_and (SImode
, static_chain
,
2747 GEN_INT (0xffff), NULL_RTX
),
2750 emit_library_call (set_trampoline_parity_libfunc
, LCT_NORMAL
, VOIDmode
,
2754 /* Return true if the current function must have and use a frame pointer. */
2757 visium_frame_pointer_required (void)
2759 /* The frame pointer is required if the function isn't leaf to be able to
2760 do manual stack unwinding. */
2764 /* If the stack pointer is dynamically modified in the function, it cannot
2765 serve as the frame pointer. */
2766 if (!crtl
->sp_is_unchanging
)
2769 /* If the function receives nonlocal gotos, it needs to save the frame
2770 pointer in the nonlocal_goto_save_area object. */
2771 if (cfun
->has_nonlocal_label
)
2774 /* The frame also needs to be established in some special cases. */
2775 if (visium_frame_needed
)
2781 /* Profiling support. Just a call to MCOUNT is needed. No labelled counter
2782 location is involved. Proper support for __builtin_return_address is also
2783 required, which is fairly straightforward provided a frame gets created. */
2786 visium_profile_hook (void)
2788 visium_frame_needed
= true;
2789 emit_library_call (gen_rtx_SYMBOL_REF (Pmode
, "mcount"), LCT_NORMAL
,
2793 /* A C expression whose value is RTL representing the address in a stack frame
2794 where the pointer to the caller's frame is stored. Assume that FRAMEADDR is
2795 an RTL expression for the address of the stack frame itself.
2797 If you don't define this macro, the default is to return the value of
2798 FRAMEADDR--that is, the stack frame address is also the address of the stack
2799 word that points to the previous frame. */
2802 visium_dynamic_chain_address (rtx frame
)
2804 /* This is the default, but we need to make sure the frame gets created. */
2805 visium_frame_needed
= true;
2809 /* A C expression whose value is RTL representing the value of the return
2810 address for the frame COUNT steps up from the current frame, after the
2811 prologue. FRAMEADDR is the frame pointer of the COUNT frame, or the frame
2812 pointer of the COUNT - 1 frame if `RETURN_ADDR_IN_PREVIOUS_FRAME' is
2815 The value of the expression must always be the correct address when COUNT is
2816 zero, but may be `NULL_RTX' if there is not way to determine the return
2817 address of other frames. */
2820 visium_return_addr_rtx (int count
, rtx frame ATTRIBUTE_UNUSED
)
2822 /* Dont try to compute anything other than frame zero. */
2826 visium_frame_needed
= true;
2828 gen_frame_mem (Pmode
, plus_constant (Pmode
, hard_frame_pointer_rtx
, 4));
2831 /* Helper function for EH_RETURN_HANDLER_RTX. Return the RTX representing a
2832 location in which to store the address of an exception handler to which we
2836 visium_eh_return_handler_rtx (void)
2839 = gen_frame_mem (SImode
, plus_constant (Pmode
, hard_frame_pointer_rtx
, 4));
2840 MEM_VOLATILE_P (mem
) = 1;
2844 static struct machine_function
*
2845 visium_init_machine_status (void)
2847 return ggc_cleared_alloc
<machine_function
> ();
2850 /* The per-function data machinery is needed to indicate when a frame
2854 visium_init_expanders (void)
2856 init_machine_status
= visium_init_machine_status
;
2859 /* Given a comparison code (EQ, NE, etc.) and the operands of a COMPARE,
2860 return the mode to be used for the comparison. */
2863 visium_select_cc_mode (enum rtx_code code
, rtx op0
, rtx op1
)
2865 if (GET_MODE_CLASS (GET_MODE (op0
)) == MODE_FLOAT
)
2885 /* These 2 comparison codes are not supported. */
2893 /* This is for the cmp<mode>_sne pattern. */
2894 if (op1
== constm1_rtx
)
2897 /* This is for the add<mode>3_insn_set_carry pattern. */
2898 if ((code
== LTU
|| code
== GEU
)
2899 && GET_CODE (op0
) == PLUS
2900 && rtx_equal_p (XEXP (op0
, 0), op1
))
2903 /* This is for the {add,sub,neg}<mode>3_insn_set_overflow pattern. */
2904 if ((code
== EQ
|| code
== NE
)
2905 && GET_CODE (op1
) == UNSPEC
2906 && (XINT (op1
, 1) == UNSPEC_ADDV
2907 || XINT (op1
, 1) == UNSPEC_SUBV
2908 || XINT (op1
, 1) == UNSPEC_NEGV
))
2911 if (op1
!= const0_rtx
)
2914 switch (GET_CODE (op0
))
2922 /* The C and V flags may be set differently from a COMPARE with zero.
2923 The consequence is that a comparison operator testing C or V must
2924 be turned into another operator not testing C or V and yielding
2925 the same result for a comparison with zero. That's possible for
2926 GE/LT which become NC/NS respectively, but not for GT/LE for which
2927 the altered operator doesn't exist on the Visium. */
2931 /* This is a btst, the result is in C instead of Z. */
2943 /* Pretend that the flags are set as for a COMPARE with zero.
2944 That's mostly true, except for the 2 right shift insns that
2945 will set the C flag. But the C flag is relevant only for
2946 the unsigned comparison operators and they are eliminated
2947 when applied to a comparison with zero. */
2950 /* ??? Cater to the junk RTXes sent by try_merge_compare. */
2966 /* Split a compare-and-branch with CODE, operands OP0 and OP1, and LABEL. */
2969 visium_split_cbranch (enum rtx_code code
, rtx op0
, rtx op1
, rtx label
)
2971 machine_mode cc_mode
= visium_select_cc_mode (code
, op0
, op1
);
2972 rtx flags
= gen_rtx_REG (cc_mode
, FLAGS_REGNUM
);
2974 rtx x
= gen_rtx_COMPARE (cc_mode
, op0
, op1
);
2975 x
= gen_rtx_SET (flags
, x
);
2978 x
= gen_rtx_fmt_ee (code
, VOIDmode
, flags
, const0_rtx
);
2979 x
= gen_rtx_IF_THEN_ELSE (VOIDmode
, x
, gen_rtx_LABEL_REF (Pmode
, label
),
2981 x
= gen_rtx_SET (pc_rtx
, x
);
2984 visium_flags_exposed
= true;
2987 /* Branch instructions on the Visium.
2989 Setting aside the interrupt-handling specific instructions, the ISA has
2990 two branch instructions: BRR and BRA. The former is used to implement
2991 short branches (+/- 2^17) within functions and its target is encoded in
2992 the instruction. The latter is used to implement all the other types
2993 of control flow changes and its target might not be statically known
2994 or even easily predictable at run time. Here's a complete summary of
2995 the patterns that generate a BRA instruction:
3005 Among these patterns, only the return (5) and the long branch (6) can be
3006 conditional; all the other patterns are always unconditional.
3008 The following algorithm can be used to identify the pattern for which
3009 the BRA instruction was generated and work out its target:
3011 A. If the source is r21 and the destination is r0, this is a return (5)
3012 and the target is the caller (i.e. the value of r21 on function's
3015 B. If the source is rN, N != 21 and the destination is r0, this is either
3016 an indirect jump or a table jump (1, 2) and the target is not easily
3019 C. If the source is rN, N != 21 and the destination is r21, this is a call
3020 (3) and the target is given by the preceding MOVIL/MOVIU pair for rN,
3021 unless this is an indirect call in which case the target is not easily
3024 D. If the source is rN, N != 21 and the destination is also rN, this is
3025 either a sibling call or a trampoline (4, 7) and the target is given
3026 by the preceding MOVIL/MOVIU pair for rN.
3028 E. If the source is r21 and the destination is also r21, this is a long
3029 branch (6) and the target is given by the preceding MOVIL/MOVIU pair
3032 The other combinations are not used. This implementation has been devised
3033 to accommodate the branch predictor of the GR6 but is used unconditionally
3034 by the compiler, i.e. including for earlier processors. */
3036 /* Output a conditional/unconditional branch to LABEL. COND is the string
3037 condition. INSN is the instruction. */
3040 output_branch (rtx label
, const char *cond
, rtx_insn
*insn
)
3046 operands
[0] = label
;
3048 /* If the length of the instruction is greater than 12, then this is a
3049 long branch and we need to work harder to emit it properly. */
3050 if (get_attr_length (insn
) > 12)
3054 /* If the link register has been saved, then we use it. */
3055 if (current_function_saves_lr ())
3057 operands
[1] = regno_reg_rtx
[LINK_REGNUM
];
3061 /* Or else, if the long-branch register isn't live, we use it. */
3062 else if (!df_regs_ever_live_p (long_branch_regnum
))
3064 operands
[1] = regno_reg_rtx
[long_branch_regnum
];
3068 /* Otherwise, we will use the long-branch register but we need to
3069 spill it to the stack and reload it at the end. We should have
3070 reserved the LR slot for this purpose. */
3073 operands
[1] = regno_reg_rtx
[long_branch_regnum
];
3075 gcc_assert (current_function_has_lr_slot ());
3078 /* First emit the spill to the stack:
3081 write.l [1](sp),reg */
3086 rtx_insn
*delay
= NEXT_INSN (insn
);
3089 final_scan_insn (delay
, asm_out_file
, optimize
, 0, NULL
);
3090 PATTERN (delay
) = gen_blockage ();
3091 INSN_CODE (delay
) = -1;
3094 if (current_function_saves_fp ())
3095 output_asm_insn ("write.l 1(sp),%1", operands
);
3097 output_asm_insn ("write.l (sp),%1", operands
);
3100 /* Then emit the core sequence:
3106 We don't use r0 as the destination register of the branch because we
3107 want the Branch Pre-decode Logic of the GR6 to use the Address Load
3108 Array to predict the branch target. */
3109 output_asm_insn ("moviu %1,%%u %0", operands
);
3110 output_asm_insn ("movil %1,%%l %0", operands
);
3111 strcpy (str
, "bra ");
3113 strcat (str
, ",%1,%1");
3116 strcat (str
, "\t\t;long branch");
3117 output_asm_insn (str
, operands
);
3119 /* Finally emit the reload:
3121 read.l reg,[1](sp) */
3124 if (current_function_saves_fp ())
3125 output_asm_insn (" read.l %1,1(sp)", operands
);
3127 output_asm_insn (" read.l %1,(sp)", operands
);
3131 /* Or else, if the label is PC, then this is a return. */
3132 else if (label
== pc_rtx
)
3134 strcpy (str
, "bra ");
3136 strcat (str
, ",r21,r0%#\t\t;return");
3137 output_asm_insn (str
, operands
);
3140 /* Otherwise, this is a short branch. */
3143 strcpy (str
, "brr ");
3145 strcat (str
, ",%0%#");
3146 output_asm_insn (str
, operands
);
3152 /* Output an unconditional branch to LABEL. INSN is the instruction. */
3155 output_ubranch (rtx label
, rtx_insn
*insn
)
3157 return output_branch (label
, "tr", insn
);
3160 /* Output a conditional branch to LABEL. CODE is the comparison code.
3161 CC_MODE is the mode of the CC register. REVERSED is non-zero if we
3162 should reverse the sense of the comparison. INSN is the instruction. */
3165 output_cbranch (rtx label
, enum rtx_code code
, machine_mode cc_mode
,
3166 int reversed
, rtx_insn
*insn
)
3172 if (cc_mode
== CCFPmode
|| cc_mode
== CCFPEmode
)
3173 code
= reverse_condition_maybe_unordered (code
);
3175 code
= reverse_condition (code
);
3181 if (cc_mode
== CCCmode
)
3183 else if (cc_mode
== CCVmode
)
3190 if (cc_mode
== CCCmode
)
3192 else if (cc_mode
== CCVmode
)
3199 if (cc_mode
== CCNZmode
)
3210 if (cc_mode
== CCFPmode
|| cc_mode
== CCFPEmode
)
3217 if (cc_mode
== CCFPmode
|| cc_mode
== CCFPEmode
)
3218 cond
= "cs"; /* or "ns" */
3219 else if (cc_mode
== CCNZmode
)
3250 cond
= "cc"; /* or "nc" */
3265 /* These 2 comparison codes are not supported. */
3272 return output_branch (label
, cond
, insn
);
3275 /* Implement TARGET_PRINT_OPERAND_PUNCT_VALID_P. */
3278 visium_print_operand_punct_valid_p (unsigned char code
)
3283 /* Implement TARGET_PRINT_OPERAND. Output to stdio stream FILE the assembler
3284 syntax for an instruction operand OP subject to the modifier LETTER. */
3287 visium_print_operand (FILE *file
, rtx op
, int letter
)
3292 /* Output an insn in a delay slot. */
3294 visium_indent_opcode
= 1;
3296 fputs ("\n\t nop", file
);
3300 /* Print LS 8 bits of operand. */
3301 fprintf (file
, HOST_WIDE_INT_PRINT_UNSIGNED
, UINTVAL (op
) & 0xff);
3305 /* Print LS 16 bits of operand. */
3306 fprintf (file
, HOST_WIDE_INT_PRINT_UNSIGNED
, UINTVAL (op
) & 0xffff);
3310 /* Print MS 16 bits of operand. */
3312 HOST_WIDE_INT_PRINT_UNSIGNED
, (UINTVAL (op
) >> 16) & 0xffff);
3316 /* It's either a register or zero. */
3317 if (GET_CODE (op
) == REG
)
3318 fputs (reg_names
[REGNO (op
)], file
);
3320 fputs (reg_names
[0], file
);
3324 /* It's either a FP register or zero. */
3325 if (GET_CODE (op
) == REG
)
3326 fputs (reg_names
[REGNO (op
)], file
);
3328 fputs (reg_names
[FP_FIRST_REGNUM
], file
);
3332 switch (GET_CODE (op
))
3336 fputs (reg_names
[REGNO (op
) + 1], file
);
3338 fputs (reg_names
[REGNO (op
)], file
);
3344 output_addr_const (file
, op
);
3348 visium_print_operand_address (file
, GET_MODE (op
), XEXP (op
, 0));
3352 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (op
));
3356 asm_fprintf (file
, "%LL%d", CODE_LABEL_NUMBER (op
));
3360 visium_print_operand (file
, XEXP (op
, 1), letter
);
3364 fatal_insn ("illegal operand ", op
);
3368 /* Implement TARGET_PRINT_OPERAND_ADDRESS. Output to stdio stream FILE the
3369 assembler syntax for an instruction operand that is a memory reference
3370 whose address is ADDR. */
3373 visium_print_operand_address (FILE *file
, machine_mode mode
, rtx addr
)
3375 switch (GET_CODE (addr
))
3379 fprintf (file
, "(%s)", reg_names
[true_regnum (addr
)]);
3384 rtx x
= XEXP (addr
, 0), y
= XEXP (addr
, 1);
3386 switch (GET_CODE (x
))
3390 if (CONST_INT_P (y
))
3392 unsigned int regno
= true_regnum (x
);
3393 HOST_WIDE_INT val
= INTVAL (y
);
3411 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
"(%s)", val
,
3415 fatal_insn ("illegal operand address (1)", addr
);
3419 if (CONSTANT_P (x
) && CONSTANT_P (y
))
3420 output_addr_const (file
, addr
);
3422 fatal_insn ("illegal operand address (2)", addr
);
3432 output_addr_const (file
, addr
);
3436 if (NOTE_KIND (addr
) != NOTE_INSN_DELETED_LABEL
)
3437 fatal_insn ("illegal operand address (3)", addr
);
3441 asm_fprintf (file
, "%LL%d", CODE_LABEL_NUMBER (addr
));
3445 fatal_insn ("illegal operand address (4)", addr
);
3450 /* The Visium stack frames look like:
3452 Before call After call
3453 +-----------------------+ +-----------------------+
3455 high | previous | | previous |
3456 mem | frame | | frame |
3458 +-----------------------+ +-----------------------+
3460 | arguments on stack | | arguments on stack |
3462 SP+0->+-----------------------+ +-----------------------+
3463 | reg parm save area, |
3464 | only created for |
3465 | variable argument |
3467 +-----------------------+
3469 | register save area |
3471 +-----------------------+
3475 FP+8->+-----------------------+
3477 FP+4->+-----------------------+
3479 FP+0->+-----------------------+
3481 | alloca allocations |
3483 +-----------------------+
3485 low | arguments on stack |
3487 SP+0->+-----------------------+
3490 1) The "reg parm save area" does not exist for non variable argument fns.
3491 2) The FP register is not saved if `frame_pointer_needed' is zero and it
3492 is not altered in the current function.
3493 3) The return address is not saved if there is no frame pointer and the
3494 current function is leaf.
3495 4) If the return address is not saved and the static chain register is
3496 live in the function, we allocate the return address slot to be able
3497 to spill the register for a long branch. */
3499 /* Define the register classes for local purposes. */
3500 enum reg_type
{ general
, mdb
, mdc
, floating
, last_type
};
3502 #define GET_REG_TYPE(regno) \
3503 (GP_REGISTER_P (regno) ? general : \
3504 (regno) == MDB_REGNUM ? mdb : \
3505 (regno) == MDC_REGNUM ? mdc : \
3508 /* First regno of each register type. */
3509 const int first_regno
[last_type
] = {0, MDB_REGNUM
, MDC_REGNUM
, FP_FIRST_REGNUM
};
3511 /* Size in bytes of each register type. */
3512 const int reg_type_size
[last_type
] = {4, 8, 4, 4};
3514 /* Structure to be filled in by visium_compute_frame_size. */
3515 struct visium_frame_info
3517 unsigned int save_area_size
; /* # bytes in the reg parm save area. */
3518 unsigned int reg_size1
; /* # bytes to store first block of regs. */
3519 unsigned int reg_size2
; /* # bytes to store second block of regs. */
3520 unsigned int max_reg1
; /* max. regno in first block */
3521 unsigned int var_size
; /* # bytes that variables take up. */
3522 unsigned int save_fp
; /* Nonzero if fp must be saved. */
3523 unsigned int save_lr
; /* Nonzero if lr must be saved. */
3524 unsigned int lr_slot
; /* Nonzero if the lr slot is needed. */
3525 unsigned int combine
; /* Nonzero if we can combine the allocation of
3526 variables and regs. */
3527 unsigned int interrupt
; /* Nonzero if the function is an interrupt
3529 unsigned int mask
[last_type
]; /* Masks of saved regs: gp, mdb, mdc, fp */
3532 /* Current frame information calculated by visium_compute_frame_size. */
3533 static struct visium_frame_info current_frame_info
;
3535 /* Accessor for current_frame_info.save_fp. */
3538 current_function_saves_fp (void)
3540 return current_frame_info
.save_fp
!= 0;
3543 /* Accessor for current_frame_info.save_lr. */
3546 current_function_saves_lr (void)
3548 return current_frame_info
.save_lr
!= 0;
3551 /* Accessor for current_frame_info.lr_slot. */
3554 current_function_has_lr_slot (void)
3556 return current_frame_info
.lr_slot
!= 0;
3559 /* Return non-zero if register REGNO needs to be saved in the frame. */
3562 visium_save_reg_p (int interrupt
, int regno
)
3566 case HARD_FRAME_POINTER_REGNUM
:
3567 /* This register is call-saved but handled specially. */
3571 /* This register is fixed but can be modified. */
3576 /* These registers are fixed and hold the interrupt context. */
3577 return (interrupt
!= 0);
3580 /* The other fixed registers are either immutable or special. */
3581 if (fixed_regs
[regno
])
3590 if (df_regs_ever_live_p (regno
))
3593 else if (call_used_or_fixed_reg_p (regno
))
3596 /* To save mdb requires two temporary registers. To save mdc or
3597 any of the floating registers requires one temporary
3598 register. If this is an interrupt routine, the temporary
3599 registers need to be saved as well. These temporary registers
3600 are call used, so we only need deal with the case of leaf
3602 if (regno
== PROLOGUE_TMP_REGNUM
)
3604 if (df_regs_ever_live_p (MDB_REGNUM
)
3605 || df_regs_ever_live_p (MDC_REGNUM
))
3608 for (int i
= FP_FIRST_REGNUM
; i
<= FP_LAST_REGNUM
; i
++)
3609 if (df_regs_ever_live_p (i
))
3613 else if (regno
== PROLOGUE_TMP_REGNUM
+ 1)
3615 if (df_regs_ever_live_p (MDB_REGNUM
))
3620 return df_regs_ever_live_p (regno
) && !call_used_or_fixed_reg_p (regno
);
3623 /* Compute the frame size required by the function. This function is called
3624 during the reload pass and also by visium_expand_prologue. */
3627 visium_compute_frame_size (int size
)
3629 const int save_area_size
= visium_reg_parm_save_area_size
;
3630 const int var_size
= VISIUM_STACK_ALIGN (size
);
3632 = frame_pointer_needed
|| df_regs_ever_live_p (HARD_FRAME_POINTER_REGNUM
);
3633 const int save_lr
= frame_pointer_needed
|| !crtl
->is_leaf
;
3634 const int lr_slot
= !save_lr
&& df_regs_ever_live_p (long_branch_regnum
);
3635 const int local_frame_offset
3636 = (save_fp
+ save_lr
+ lr_slot
) * UNITS_PER_WORD
;
3637 const int interrupt
= visium_interrupt_function_p ();
3638 unsigned int mask
[last_type
];
3647 memset (mask
, 0, last_type
* sizeof (unsigned int));
3649 /* The registers may need stacking in 2 blocks since only 32 32-bit words
3650 can be indexed from a given base address. */
3651 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
3653 if (visium_save_reg_p (interrupt
, regno
))
3655 enum reg_type reg_type
= GET_REG_TYPE (regno
);
3656 int mask_bit
= 1 << (regno
- first_regno
[reg_type
]);
3657 int nbytes
= reg_type_size
[reg_type
];
3659 if (reg_size1
+ nbytes
> 32 * UNITS_PER_WORD
)
3662 reg_size1
+= nbytes
;
3664 mask
[reg_type
] |= mask_bit
;
3668 for (regno
= max_reg1
+ 1; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
3670 if (visium_save_reg_p (interrupt
, regno
))
3672 enum reg_type reg_type
= GET_REG_TYPE (regno
);
3673 int mask_bit
= 1 << (regno
- first_regno
[reg_type
]);
3674 int nbytes
= reg_type_size
[reg_type
];
3676 reg_size2
+= nbytes
;
3677 mask
[reg_type
] |= mask_bit
;
3681 reg_size
= reg_size2
? reg_size2
: reg_size1
;
3682 combine
= (local_frame_offset
+ var_size
+ reg_size
) <= 32 * UNITS_PER_WORD
;
3684 = local_frame_offset
+ var_size
+ reg_size2
+ reg_size1
+ save_area_size
;
3686 current_frame_info
.save_area_size
= save_area_size
;
3687 current_frame_info
.reg_size1
= reg_size1
;
3688 current_frame_info
.max_reg1
= max_reg1
;
3689 current_frame_info
.reg_size2
= reg_size2
;
3690 current_frame_info
.var_size
= var_size
;
3691 current_frame_info
.save_fp
= save_fp
;
3692 current_frame_info
.save_lr
= save_lr
;
3693 current_frame_info
.lr_slot
= lr_slot
;
3694 current_frame_info
.combine
= combine
;
3695 current_frame_info
.interrupt
= interrupt
;
3697 memcpy (current_frame_info
.mask
, mask
, last_type
* sizeof (unsigned int));
3702 /* Helper function for INITIAL_ELIMINATION_OFFSET(FROM, TO, OFFSET). Define
3703 the offset between two registers, one to be eliminated, and the other its
3704 replacement, at the start of a routine. */
3707 visium_initial_elimination_offset (int from
, int to ATTRIBUTE_UNUSED
)
3709 const int save_fp
= current_frame_info
.save_fp
;
3710 const int save_lr
= current_frame_info
.save_lr
;
3711 const int lr_slot
= current_frame_info
.lr_slot
;
3714 if (from
== FRAME_POINTER_REGNUM
)
3715 offset
= (save_fp
+ save_lr
+ lr_slot
) * UNITS_PER_WORD
;
3716 else if (from
== ARG_POINTER_REGNUM
)
3717 offset
= visium_compute_frame_size (get_frame_size ());
3724 /* For an interrupt handler, we may be saving call-clobbered registers.
3725 Say the epilogue uses these in addition to the link register. */
3728 visium_epilogue_uses (int regno
)
3730 if (regno
== LINK_REGNUM
)
3733 if (reload_completed
)
3735 enum reg_type reg_type
= GET_REG_TYPE (regno
);
3736 int mask_bit
= 1 << (regno
- first_regno
[reg_type
]);
3738 return (current_frame_info
.mask
[reg_type
] & mask_bit
) != 0;
3744 /* Wrapper around emit_insn that sets RTX_FRAME_RELATED_P on the insn. */
3747 emit_frame_insn (rtx x
)
3750 RTX_FRAME_RELATED_P (x
) = 1;
3754 /* Allocate ALLOC bytes on the stack and save the registers LOW_REGNO to
3755 HIGH_REGNO at OFFSET from the stack pointer. */
3758 visium_save_regs (int alloc
, int offset
, int low_regno
, int high_regno
)
3760 /* If this is an interrupt handler function, then mark the register
3761 stores as volatile. This will prevent the instruction scheduler
3762 from scrambling the order of register saves. */
3763 const int volatile_p
= current_frame_info
.interrupt
;
3766 /* Allocate the stack space. */
3767 emit_frame_insn (gen_addsi3_flags (stack_pointer_rtx
, stack_pointer_rtx
,
3770 for (regno
= low_regno
; regno
<= high_regno
; regno
++)
3772 enum reg_type reg_type
= GET_REG_TYPE (regno
);
3773 int mask_bit
= 1 << (regno
- first_regno
[reg_type
]);
3776 if (current_frame_info
.mask
[reg_type
] & mask_bit
)
3778 offset
-= reg_type_size
[reg_type
];
3784 = gen_frame_mem (SImode
,
3785 plus_constant (Pmode
,
3786 stack_pointer_rtx
, offset
));
3787 MEM_VOLATILE_P (mem
) = volatile_p
;
3788 emit_frame_insn (gen_movsi (mem
, gen_rtx_REG (SImode
, regno
)));
3794 rtx tmp
= gen_rtx_REG (DImode
, PROLOGUE_TMP_REGNUM
);
3796 = gen_frame_mem (DImode
,
3797 plus_constant (Pmode
,
3798 stack_pointer_rtx
, offset
));
3799 rtx reg
= gen_rtx_REG (DImode
, regno
);
3800 MEM_VOLATILE_P (mem
) = volatile_p
;
3801 emit_insn (gen_movdi (tmp
, reg
));
3802 /* Do not generate CFI if in interrupt handler. */
3804 emit_insn (gen_movdi (mem
, tmp
));
3807 insn
= emit_frame_insn (gen_movdi (mem
, tmp
));
3808 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
3809 gen_rtx_SET (mem
, reg
));
3816 rtx tmp
= gen_rtx_REG (SImode
, PROLOGUE_TMP_REGNUM
);
3818 = gen_frame_mem (SImode
,
3819 plus_constant (Pmode
,
3820 stack_pointer_rtx
, offset
));
3821 rtx reg
= gen_rtx_REG (SImode
, regno
);
3822 MEM_VOLATILE_P (mem
) = volatile_p
;
3823 emit_insn (gen_movsi (tmp
, reg
));
3824 insn
= emit_frame_insn (gen_movsi (mem
, tmp
));
3825 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
3826 gen_rtx_SET (mem
, reg
));
3832 rtx tmp
= gen_rtx_REG (SFmode
, PROLOGUE_TMP_REGNUM
);
3834 = gen_frame_mem (SFmode
,
3835 plus_constant (Pmode
,
3836 stack_pointer_rtx
, offset
));
3837 rtx reg
= gen_rtx_REG (SFmode
, regno
);
3838 MEM_VOLATILE_P (mem
) = volatile_p
;
3839 emit_insn (gen_movsf (tmp
, reg
));
3840 insn
= emit_frame_insn (gen_movsf (mem
, tmp
));
3841 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
3842 gen_rtx_SET (mem
, reg
));
3853 /* This function generates the code for function entry. */
3856 visium_expand_prologue (void)
3858 const int frame_size
= visium_compute_frame_size (get_frame_size ());
3859 const int save_area_size
= current_frame_info
.save_area_size
;
3860 const int reg_size1
= current_frame_info
.reg_size1
;
3861 const int max_reg1
= current_frame_info
.max_reg1
;
3862 const int reg_size2
= current_frame_info
.reg_size2
;
3863 const int var_size
= current_frame_info
.var_size
;
3864 const int save_fp
= current_frame_info
.save_fp
;
3865 const int save_lr
= current_frame_info
.save_lr
;
3866 const int lr_slot
= current_frame_info
.lr_slot
;
3867 const int local_frame_offset
3868 = (save_fp
+ save_lr
+ lr_slot
) * UNITS_PER_WORD
;
3869 const int combine
= current_frame_info
.combine
;
3874 /* Save the frame size for future references. */
3875 visium_frame_size
= frame_size
;
3877 if (flag_stack_usage_info
)
3878 current_function_static_stack_size
= frame_size
;
3880 /* If the registers have to be stacked in 2 blocks, stack the first one. */
3883 visium_save_regs (reg_size1
+ save_area_size
, reg_size1
, 0, max_reg1
);
3884 reg_size
= reg_size2
;
3885 first_reg
= max_reg1
+ 1;
3886 fsize
= local_frame_offset
+ var_size
+ reg_size2
;
3890 reg_size
= reg_size1
;
3892 fsize
= local_frame_offset
+ var_size
+ reg_size1
+ save_area_size
;
3895 /* If we can't combine register stacking with variable allocation, partially
3896 allocate and stack the (remaining) registers now. */
3897 if (reg_size
&& !combine
)
3898 visium_save_regs (fsize
- local_frame_offset
- var_size
, reg_size
,
3899 first_reg
, FIRST_PSEUDO_REGISTER
- 1);
3901 /* If we can combine register stacking with variable allocation, fully
3902 allocate and stack the (remaining) registers now. */
3903 if (reg_size
&& combine
)
3904 visium_save_regs (fsize
, local_frame_offset
+ var_size
+ reg_size
,
3905 first_reg
, FIRST_PSEUDO_REGISTER
- 1);
3907 /* Otherwise space may still need to be allocated for the variables. */
3910 const int alloc_size
= reg_size
? local_frame_offset
+ var_size
: fsize
;
3912 if (alloc_size
> 65535)
3914 rtx tmp
= gen_rtx_REG (SImode
, PROLOGUE_TMP_REGNUM
), insn
;
3915 emit_insn (gen_movsi (tmp
, GEN_INT (alloc_size
)));
3916 insn
= emit_frame_insn (gen_subsi3_flags (stack_pointer_rtx
,
3919 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
3920 gen_rtx_SET (stack_pointer_rtx
,
3921 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
3922 GEN_INT (-alloc_size
))));
3925 emit_frame_insn (gen_addsi3_flags (stack_pointer_rtx
,
3927 GEN_INT (-alloc_size
)));
3931 emit_frame_insn (gen_movsi (gen_frame_mem (SImode
, stack_pointer_rtx
),
3932 hard_frame_pointer_rtx
));
3934 if (frame_pointer_needed
)
3935 emit_frame_insn (gen_stack_save ());
3941 /* Normally the frame pointer and link register get saved via
3946 Indexing off sp rather than fp to store the link register
3947 avoids presenting the instruction scheduler with an initial
3948 pipeline hazard. If however the frame is needed for eg.
3949 __builtin_return_address which needs to retrieve the saved
3950 value of the link register from the stack at fp + 4 then
3951 indexing from sp can confuse the dataflow, causing the link
3952 register to be retrieved before it has been saved. */
3953 if (cfun
->machine
->frame_needed
)
3954 base_rtx
= hard_frame_pointer_rtx
;
3956 base_rtx
= stack_pointer_rtx
;
3958 mem
= gen_frame_mem (SImode
,
3959 plus_constant (Pmode
,
3960 base_rtx
, save_fp
* UNITS_PER_WORD
));
3961 emit_frame_insn (gen_movsi (mem
, gen_rtx_REG (SImode
, LINK_REGNUM
)));
3965 static GTY(()) rtx cfa_restores
;
3967 /* Queue a REG_CFA_RESTORE note until next stack manipulation insn. */
3970 visium_add_cfa_restore_note (rtx reg
)
3972 cfa_restores
= alloc_reg_note (REG_CFA_RESTORE
, reg
, cfa_restores
);
3975 /* Add queued REG_CFA_RESTORE notes to INSN, if any. */
3978 visium_add_queued_cfa_restore_notes (rtx insn
)
3983 for (last
= cfa_restores
; XEXP (last
, 1); last
= XEXP (last
, 1))
3985 XEXP (last
, 1) = REG_NOTES (insn
);
3986 REG_NOTES (insn
) = cfa_restores
;
3987 cfa_restores
= NULL_RTX
;
3990 /* Restore the registers LOW_REGNO to HIGH_REGNO from the save area at OFFSET
3991 from the stack pointer and pop DEALLOC bytes off the stack. */
3994 visium_restore_regs (int dealloc
, int offset
, int high_regno
, int low_regno
)
3996 /* If this is an interrupt handler function, then mark the register
3997 restores as volatile. This will prevent the instruction scheduler
3998 from scrambling the order of register restores. */
3999 const int volatile_p
= current_frame_info
.interrupt
;
4000 int r30_offset
= -1;
4003 for (regno
= high_regno
; regno
>= low_regno
; --regno
)
4005 enum reg_type reg_type
= GET_REG_TYPE (regno
);
4006 int mask_bit
= 1 << (regno
- first_regno
[reg_type
]);
4008 if (current_frame_info
.mask
[reg_type
] & mask_bit
)
4013 /* Postpone restoring the interrupted context registers
4014 until last, since they need to be preceded by a dsi. */
4017 else if (regno
== 30)
4018 r30_offset
= offset
;
4022 = gen_frame_mem (SImode
,
4023 plus_constant (Pmode
,
4026 rtx reg
= gen_rtx_REG (SImode
, regno
);
4027 MEM_VOLATILE_P (mem
) = volatile_p
;
4028 emit_insn (gen_movsi (reg
, mem
));
4029 visium_add_cfa_restore_note (reg
);
4035 rtx tmp
= gen_rtx_REG (DImode
, PROLOGUE_TMP_REGNUM
);
4037 = gen_frame_mem (DImode
,
4038 plus_constant (Pmode
,
4039 stack_pointer_rtx
, offset
));
4040 rtx reg
= gen_rtx_REG (DImode
, regno
);
4041 MEM_VOLATILE_P (mem
) = volatile_p
;
4042 emit_insn (gen_movdi (tmp
, mem
));
4043 emit_insn (gen_movdi (reg
, tmp
));
4044 /* Do not generate CFI if in interrupt handler. */
4046 visium_add_cfa_restore_note (reg
);
4052 rtx tmp
= gen_rtx_REG (SImode
, PROLOGUE_TMP_REGNUM
);
4054 = gen_frame_mem (SImode
,
4055 plus_constant (Pmode
,
4056 stack_pointer_rtx
, offset
));
4057 rtx reg
= gen_rtx_REG (SImode
, regno
);
4058 MEM_VOLATILE_P (mem
) = volatile_p
;
4059 emit_insn (gen_movsi (tmp
, mem
));
4060 emit_insn (gen_movsi (reg
, tmp
));
4061 visium_add_cfa_restore_note (reg
);
4067 rtx tmp
= gen_rtx_REG (SFmode
, PROLOGUE_TMP_REGNUM
);
4069 = gen_frame_mem (SFmode
,
4070 plus_constant (Pmode
,
4071 stack_pointer_rtx
, offset
));
4072 rtx reg
= gen_rtx_REG (SFmode
, regno
);
4073 MEM_VOLATILE_P (mem
) = volatile_p
;
4074 emit_insn (gen_movsf (tmp
, mem
));
4075 emit_insn (gen_movsf (reg
, tmp
));
4076 visium_add_cfa_restore_note (reg
);
4084 offset
+= reg_type_size
[reg_type
];
4088 /* If the interrupted context needs to be restored, precede the
4089 restores of r29 and r30 by a dsi. */
4090 if (r30_offset
>= 0)
4092 emit_insn (gen_dsi ());
4093 emit_move_insn (gen_rtx_REG (SImode
, 30),
4094 gen_frame_mem (SImode
,
4095 plus_constant (Pmode
,
4098 emit_move_insn (gen_rtx_REG (SImode
, 29),
4099 gen_frame_mem (SImode
,
4100 plus_constant (Pmode
,
4105 /* Deallocate the stack space. */
4106 rtx insn
= emit_frame_insn (gen_stack_pop (GEN_INT (dealloc
)));
4107 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
4108 gen_rtx_SET (stack_pointer_rtx
,
4109 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
4110 GEN_INT (dealloc
))));
4111 visium_add_queued_cfa_restore_notes (insn
);
4114 /* This function generates the code for function exit. */
4117 visium_expand_epilogue (void)
4119 const int save_area_size
= current_frame_info
.save_area_size
;
4120 const int reg_size1
= current_frame_info
.reg_size1
;
4121 const int max_reg1
= current_frame_info
.max_reg1
;
4122 const int reg_size2
= current_frame_info
.reg_size2
;
4123 const int var_size
= current_frame_info
.var_size
;
4124 const int restore_fp
= current_frame_info
.save_fp
;
4125 const int restore_lr
= current_frame_info
.save_lr
;
4126 const int lr_slot
= current_frame_info
.lr_slot
;
4127 const int local_frame_offset
4128 = (restore_fp
+ restore_lr
+ lr_slot
) * UNITS_PER_WORD
;
4129 const int combine
= current_frame_info
.combine
;
4134 /* Do not bother restoring the stack pointer if it hasn't been changed in
4135 the function since it was saved _after_ the allocation of the frame. */
4136 if (!crtl
->sp_is_unchanging
)
4137 emit_insn (gen_stack_restore ());
4139 /* Restore the frame pointer if necessary. The usual code would be:
4144 but for the MCM this constitutes a stall/hazard so it is changed to:
4149 if the stack pointer has actually been restored. */
4154 if (TARGET_MCM
&& !crtl
->sp_is_unchanging
)
4155 src
= gen_frame_mem (SImode
, hard_frame_pointer_rtx
);
4157 src
= gen_frame_mem (SImode
, stack_pointer_rtx
);
4159 rtx insn
= emit_frame_insn (gen_movsi (hard_frame_pointer_rtx
, src
));
4160 add_reg_note (insn
, REG_CFA_ADJUST_CFA
,
4161 gen_rtx_SET (stack_pointer_rtx
,
4162 hard_frame_pointer_rtx
));
4163 visium_add_cfa_restore_note (hard_frame_pointer_rtx
);
4166 /* Restore the link register if necessary. */
4169 rtx mem
= gen_frame_mem (SImode
,
4170 plus_constant (Pmode
,
4172 restore_fp
* UNITS_PER_WORD
));
4173 rtx reg
= gen_rtx_REG (SImode
, LINK_REGNUM
);
4174 emit_insn (gen_movsi (reg
, mem
));
4175 visium_add_cfa_restore_note (reg
);
4178 /* If we have two blocks of registers, deal with the second one first. */
4181 reg_size
= reg_size2
;
4182 last_reg
= max_reg1
+ 1;
4183 fsize
= local_frame_offset
+ var_size
+ reg_size2
;
4187 reg_size
= reg_size1
;
4189 fsize
= local_frame_offset
+ var_size
+ reg_size1
+ save_area_size
;
4192 /* If the variable allocation could be combined with register stacking,
4193 restore the (remaining) registers and fully deallocate now. */
4194 if (reg_size
&& combine
)
4195 visium_restore_regs (fsize
, local_frame_offset
+ var_size
,
4196 FIRST_PSEUDO_REGISTER
- 1, last_reg
);
4198 /* Otherwise deallocate the variables first. */
4201 const int pop_size
= reg_size
? local_frame_offset
+ var_size
: fsize
;
4204 if (pop_size
> 65535)
4206 rtx tmp
= gen_rtx_REG (SImode
, PROLOGUE_TMP_REGNUM
);
4207 emit_move_insn (tmp
, GEN_INT (pop_size
));
4208 insn
= emit_frame_insn (gen_stack_pop (tmp
));
4211 insn
= emit_frame_insn (gen_stack_pop (GEN_INT (pop_size
)));
4212 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
4213 gen_rtx_SET (stack_pointer_rtx
,
4214 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
4215 GEN_INT (pop_size
))));
4216 visium_add_queued_cfa_restore_notes (insn
);
4219 /* If the variable allocation couldn't be combined with register stacking,
4220 restore the (remaining) registers now and partially deallocate. */
4221 if (reg_size
&& !combine
)
4222 visium_restore_regs (fsize
- local_frame_offset
- var_size
, 0,
4223 FIRST_PSEUDO_REGISTER
- 1, last_reg
);
4225 /* If the first block of registers has yet to be restored, do it now. */
4227 visium_restore_regs (reg_size1
+ save_area_size
, 0, max_reg1
, 0);
4229 /* If this is an exception return, make the necessary stack adjustment. */
4230 if (crtl
->calls_eh_return
)
4231 emit_insn (gen_stack_pop (EH_RETURN_STACKADJ_RTX
));
4234 /* Return true if it is appropriate to emit `return' instructions in the
4235 body of a function. */
4238 visium_can_use_return_insn_p (void)
4240 return reload_completed
4241 && visium_frame_size
== 0
4242 && !visium_interrupt_function_p ();
4245 /* Return the register class required for an intermediate register used to
4246 copy a register of RCLASS from/to X. If no such intermediate register is
4247 required, return NO_REGS. If more than one such intermediate register is
4248 required, describe the one that is closest in the copy chain to the reload
4252 visium_secondary_reload (bool in_p ATTRIBUTE_UNUSED
, rtx x
,
4254 machine_mode mode ATTRIBUTE_UNUSED
,
4255 secondary_reload_info
*sri ATTRIBUTE_UNUSED
)
4257 int regno
= true_regnum (x
);
4259 /* For MDB, MDC and FP_REGS, a general register is needed for a move to
4261 if (regno
== -1 && (rclass
== MDB
|| rclass
== MDC
|| rclass
== FP_REGS
))
4262 return GENERAL_REGS
;
4264 /* Moves between MDB, MDC and FP_REGS also require a general register. */
4265 else if (((regno
== R_MDB
|| regno
== R_MDC
) && rclass
== FP_REGS
)
4266 || (FP_REGISTER_P (regno
) && (rclass
== MDB
|| rclass
== MDC
)))
4267 return GENERAL_REGS
;
4269 /* Finally an (unlikely ?) move between MDB and MDC needs a general reg. */
4270 else if ((regno
== R_MDB
&& rclass
== MDC
)
4271 || (rclass
== MDB
&& regno
== R_MDC
))
4272 return GENERAL_REGS
;
4277 /* Return true if pseudos that have been assigned to registers of RCLASS
4278 would likely be spilled because registers of RCLASS are needed for
4282 visium_class_likely_spilled_p (reg_class_t rclass ATTRIBUTE_UNUSED
)
4284 /* Return false for classes R1, R2 and R3, which are intended to be used
4285 only in the source code in conjunction with block move instructions. */
4289 /* Return the register number if OP is a REG or a SUBREG of a REG, and
4290 INVALID_REGNUM in all the other cases. */
4293 reg_or_subreg_regno (rtx op
)
4297 if (GET_CODE (op
) == REG
)
4299 else if (GET_CODE (op
) == SUBREG
&& GET_CODE (SUBREG_REG (op
)) == REG
)
4301 if (REGNO (SUBREG_REG (op
)) < FIRST_PSEUDO_REGISTER
)
4302 regno
= subreg_regno (op
);
4304 regno
= REGNO (SUBREG_REG (op
));
4307 regno
= INVALID_REGNUM
;
4312 /* Implement TARGET_CAN_CHANGE_MODE_CLASS.
4314 It's not obvious from the documentation of the hook that MDB cannot
4315 change mode. However difficulties arise from expressions of the form
4317 (subreg:SI (reg:DI R_MDB) 0)
4319 There is no way to convert that reference to a single machine
4320 register and, without the following definition, reload will quietly
4326 visium_can_change_mode_class (machine_mode from
, machine_mode to
,
4329 return (rclass
!= MDB
|| GET_MODE_SIZE (from
) == GET_MODE_SIZE (to
));
4332 #include "gt-visium.h"