1 /* Subroutines for insn-output.c for SPARC.
2 Copyright (C) 1987, 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998,
3 1999, 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc.
4 Contributed by Michael Tiemann (tiemann@cygnus.com)
5 64-bit SPARC-V9 support by Michael Tiemann, Jim Wilson, and Doug Evans,
8 This file is part of GCC.
10 GCC is free software; you can redistribute it and/or modify
11 it under the terms of the GNU General Public License as published by
12 the Free Software Foundation; either version 2, or (at your option)
15 GCC is distributed in the hope that it will be useful,
16 but WITHOUT ANY WARRANTY; without even the implied warranty of
17 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 GNU General Public License for more details.
20 You should have received a copy of the GNU General Public License
21 along with GCC; see the file COPYING. If not, write to
22 the Free Software Foundation, 59 Temple Place - Suite 330,
23 Boston, MA 02111-1307, USA. */
27 #include "coretypes.h"
32 #include "hard-reg-set.h"
34 #include "insn-config.h"
35 #include "conditions.h"
37 #include "insn-attr.h"
48 #include "target-def.h"
49 #include "cfglayout.h"
50 #include "tree-gimple.h"
54 struct processor_costs cypress_costs = {
55 COSTS_N_INSNS (2), /* int load */
56 COSTS_N_INSNS (2), /* int signed load */
57 COSTS_N_INSNS (2), /* int zeroed load */
58 COSTS_N_INSNS (2), /* float load */
59 COSTS_N_INSNS (5), /* fmov, fneg, fabs */
60 COSTS_N_INSNS (5), /* fadd, fsub */
61 COSTS_N_INSNS (1), /* fcmp */
62 COSTS_N_INSNS (1), /* fmov, fmovr */
63 COSTS_N_INSNS (7), /* fmul */
64 COSTS_N_INSNS (37), /* fdivs */
65 COSTS_N_INSNS (37), /* fdivd */
66 COSTS_N_INSNS (63), /* fsqrts */
67 COSTS_N_INSNS (63), /* fsqrtd */
68 COSTS_N_INSNS (1), /* imul */
69 COSTS_N_INSNS (1), /* imulX */
70 0, /* imul bit factor */
71 COSTS_N_INSNS (1), /* idiv */
72 COSTS_N_INSNS (1), /* idivX */
73 COSTS_N_INSNS (1), /* movcc/movr */
74 0, /* shift penalty */
78 struct processor_costs supersparc_costs = {
79 COSTS_N_INSNS (1), /* int load */
80 COSTS_N_INSNS (1), /* int signed load */
81 COSTS_N_INSNS (1), /* int zeroed load */
82 COSTS_N_INSNS (0), /* float load */
83 COSTS_N_INSNS (3), /* fmov, fneg, fabs */
84 COSTS_N_INSNS (3), /* fadd, fsub */
85 COSTS_N_INSNS (3), /* fcmp */
86 COSTS_N_INSNS (1), /* fmov, fmovr */
87 COSTS_N_INSNS (3), /* fmul */
88 COSTS_N_INSNS (6), /* fdivs */
89 COSTS_N_INSNS (9), /* fdivd */
90 COSTS_N_INSNS (12), /* fsqrts */
91 COSTS_N_INSNS (12), /* fsqrtd */
92 COSTS_N_INSNS (4), /* imul */
93 COSTS_N_INSNS (4), /* imulX */
94 0, /* imul bit factor */
95 COSTS_N_INSNS (4), /* idiv */
96 COSTS_N_INSNS (4), /* idivX */
97 COSTS_N_INSNS (1), /* movcc/movr */
98 1, /* shift penalty */
102 struct processor_costs hypersparc_costs = {
103 COSTS_N_INSNS (1), /* int load */
104 COSTS_N_INSNS (1), /* int signed load */
105 COSTS_N_INSNS (1), /* int zeroed load */
106 COSTS_N_INSNS (1), /* float load */
107 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
108 COSTS_N_INSNS (1), /* fadd, fsub */
109 COSTS_N_INSNS (1), /* fcmp */
110 COSTS_N_INSNS (1), /* fmov, fmovr */
111 COSTS_N_INSNS (1), /* fmul */
112 COSTS_N_INSNS (8), /* fdivs */
113 COSTS_N_INSNS (12), /* fdivd */
114 COSTS_N_INSNS (17), /* fsqrts */
115 COSTS_N_INSNS (17), /* fsqrtd */
116 COSTS_N_INSNS (17), /* imul */
117 COSTS_N_INSNS (17), /* imulX */
118 0, /* imul bit factor */
119 COSTS_N_INSNS (17), /* idiv */
120 COSTS_N_INSNS (17), /* idivX */
121 COSTS_N_INSNS (1), /* movcc/movr */
122 0, /* shift penalty */
126 struct processor_costs sparclet_costs = {
127 COSTS_N_INSNS (3), /* int load */
128 COSTS_N_INSNS (3), /* int signed load */
129 COSTS_N_INSNS (1), /* int zeroed load */
130 COSTS_N_INSNS (1), /* float load */
131 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
132 COSTS_N_INSNS (1), /* fadd, fsub */
133 COSTS_N_INSNS (1), /* fcmp */
134 COSTS_N_INSNS (1), /* fmov, fmovr */
135 COSTS_N_INSNS (1), /* fmul */
136 COSTS_N_INSNS (1), /* fdivs */
137 COSTS_N_INSNS (1), /* fdivd */
138 COSTS_N_INSNS (1), /* fsqrts */
139 COSTS_N_INSNS (1), /* fsqrtd */
140 COSTS_N_INSNS (5), /* imul */
141 COSTS_N_INSNS (5), /* imulX */
142 0, /* imul bit factor */
143 COSTS_N_INSNS (5), /* idiv */
144 COSTS_N_INSNS (5), /* idivX */
145 COSTS_N_INSNS (1), /* movcc/movr */
146 0, /* shift penalty */
150 struct processor_costs ultrasparc_costs = {
151 COSTS_N_INSNS (2), /* int load */
152 COSTS_N_INSNS (3), /* int signed load */
153 COSTS_N_INSNS (2), /* int zeroed load */
154 COSTS_N_INSNS (2), /* float load */
155 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
156 COSTS_N_INSNS (4), /* fadd, fsub */
157 COSTS_N_INSNS (1), /* fcmp */
158 COSTS_N_INSNS (2), /* fmov, fmovr */
159 COSTS_N_INSNS (4), /* fmul */
160 COSTS_N_INSNS (13), /* fdivs */
161 COSTS_N_INSNS (23), /* fdivd */
162 COSTS_N_INSNS (13), /* fsqrts */
163 COSTS_N_INSNS (23), /* fsqrtd */
164 COSTS_N_INSNS (4), /* imul */
165 COSTS_N_INSNS (4), /* imulX */
166 2, /* imul bit factor */
167 COSTS_N_INSNS (37), /* idiv */
168 COSTS_N_INSNS (68), /* idivX */
169 COSTS_N_INSNS (2), /* movcc/movr */
170 2, /* shift penalty */
174 struct processor_costs ultrasparc3_costs = {
175 COSTS_N_INSNS (2), /* int load */
176 COSTS_N_INSNS (3), /* int signed load */
177 COSTS_N_INSNS (3), /* int zeroed load */
178 COSTS_N_INSNS (2), /* float load */
179 COSTS_N_INSNS (3), /* fmov, fneg, fabs */
180 COSTS_N_INSNS (4), /* fadd, fsub */
181 COSTS_N_INSNS (5), /* fcmp */
182 COSTS_N_INSNS (3), /* fmov, fmovr */
183 COSTS_N_INSNS (4), /* fmul */
184 COSTS_N_INSNS (17), /* fdivs */
185 COSTS_N_INSNS (20), /* fdivd */
186 COSTS_N_INSNS (20), /* fsqrts */
187 COSTS_N_INSNS (29), /* fsqrtd */
188 COSTS_N_INSNS (6), /* imul */
189 COSTS_N_INSNS (6), /* imulX */
190 0, /* imul bit factor */
191 COSTS_N_INSNS (40), /* idiv */
192 COSTS_N_INSNS (71), /* idivX */
193 COSTS_N_INSNS (2), /* movcc/movr */
194 0, /* shift penalty */
197 const struct processor_costs *sparc_costs = &cypress_costs;
199 #ifdef HAVE_AS_RELAX_OPTION
200 /* If 'as' and 'ld' are relaxing tail call insns into branch always, use
201 "or %o7,%g0,X; call Y; or X,%g0,%o7" always, so that it can be optimized.
202 With sethi/jmp, neither 'as' nor 'ld' has an easy way how to find out if
203 somebody does not branch between the sethi and jmp. */
204 #define LEAF_SIBCALL_SLOT_RESERVED_P 1
206 #define LEAF_SIBCALL_SLOT_RESERVED_P \
207 ((TARGET_ARCH64 && !TARGET_CM_MEDLOW) || flag_pic)
210 /* Global variables for machine-dependent things. */
212 /* Size of frame. Need to know this to emit return insns from leaf procedures.
213 ACTUAL_FSIZE is set by sparc_compute_frame_size() which is called during the
214 reload pass. This is important as the value is later used for scheduling
215 (to see what can go in a delay slot).
216 APPARENT_FSIZE is the size of the stack less the register save area and less
217 the outgoing argument area. It is used when saving call preserved regs. */
218 static HOST_WIDE_INT apparent_fsize;
219 static HOST_WIDE_INT actual_fsize;
221 /* Number of live general or floating point registers needed to be
222 saved (as 4-byte quantities). */
223 static int num_gfregs;
225 /* The alias set for prologue/epilogue register save/restore. */
226 static GTY(()) int sparc_sr_alias_set;
228 /* Save the operands last given to a compare for use when we
229 generate a scc or bcc insn. */
230 rtx sparc_compare_op0, sparc_compare_op1;
232 /* Vector to say how input registers are mapped to output registers.
233 HARD_FRAME_POINTER_REGNUM cannot be remapped by this function to
234 eliminate it. You must use -fomit-frame-pointer to get that. */
235 char leaf_reg_remap[] =
236 { 0, 1, 2, 3, 4, 5, 6, 7,
237 -1, -1, -1, -1, -1, -1, 14, -1,
238 -1, -1, -1, -1, -1, -1, -1, -1,
239 8, 9, 10, 11, 12, 13, -1, 15,
241 32, 33, 34, 35, 36, 37, 38, 39,
242 40, 41, 42, 43, 44, 45, 46, 47,
243 48, 49, 50, 51, 52, 53, 54, 55,
244 56, 57, 58, 59, 60, 61, 62, 63,
245 64, 65, 66, 67, 68, 69, 70, 71,
246 72, 73, 74, 75, 76, 77, 78, 79,
247 80, 81, 82, 83, 84, 85, 86, 87,
248 88, 89, 90, 91, 92, 93, 94, 95,
249 96, 97, 98, 99, 100};
251 /* Vector, indexed by hard register number, which contains 1
252 for a register that is allowable in a candidate for leaf
253 function treatment. */
254 char sparc_leaf_regs[] =
255 { 1, 1, 1, 1, 1, 1, 1, 1,
256 0, 0, 0, 0, 0, 0, 1, 0,
257 0, 0, 0, 0, 0, 0, 0, 0,
258 1, 1, 1, 1, 1, 1, 0, 1,
259 1, 1, 1, 1, 1, 1, 1, 1,
260 1, 1, 1, 1, 1, 1, 1, 1,
261 1, 1, 1, 1, 1, 1, 1, 1,
262 1, 1, 1, 1, 1, 1, 1, 1,
263 1, 1, 1, 1, 1, 1, 1, 1,
264 1, 1, 1, 1, 1, 1, 1, 1,
265 1, 1, 1, 1, 1, 1, 1, 1,
266 1, 1, 1, 1, 1, 1, 1, 1,
269 struct machine_function GTY(())
271 /* Some local-dynamic TLS symbol name. */
272 const char *some_ld_name;
274 /* True if the current function is leaf and uses only leaf regs,
275 so that the SPARC leaf function optimization can be applied.
276 Private version of current_function_uses_only_leaf_regs, see
277 sparc_expand_prologue for the rationale. */
280 /* True if the data calculated by sparc_expand_prologue are valid. */
281 bool prologue_data_valid_p;
284 #define sparc_leaf_function_p cfun->machine->leaf_function_p
285 #define sparc_prologue_data_valid_p cfun->machine->prologue_data_valid_p
287 /* Register we pretend to think the frame pointer is allocated to.
288 Normally, this is %fp, but if we are in a leaf procedure, this
289 is %sp+"something". We record "something" separately as it may
290 be too big for reg+constant addressing. */
291 static rtx frame_base_reg;
292 static HOST_WIDE_INT frame_base_offset;
294 /* 1 if the next opcode is to be specially indented. */
295 int sparc_indent_opcode = 0;
297 static void sparc_init_modes (void);
298 static void scan_record_type (tree, int *, int *, int *);
299 static int function_arg_slotno (const CUMULATIVE_ARGS *, enum machine_mode,
300 tree, int, int, int *, int *);
302 static int supersparc_adjust_cost (rtx, rtx, rtx, int);
303 static int hypersparc_adjust_cost (rtx, rtx, rtx, int);
305 static void sparc_output_addr_vec (rtx);
306 static void sparc_output_addr_diff_vec (rtx);
307 static void sparc_output_deferred_case_vectors (void);
308 static rtx sparc_builtin_saveregs (void);
309 static int epilogue_renumber (rtx *, int);
310 static bool sparc_assemble_integer (rtx, unsigned int, int);
311 static int set_extends (rtx);
312 static void load_pic_register (void);
313 static int save_or_restore_regs (int, int, rtx, int, int);
314 static void emit_save_regs (void);
315 static void emit_restore_regs (void);
316 static void sparc_asm_function_prologue (FILE *, HOST_WIDE_INT);
317 static void sparc_asm_function_epilogue (FILE *, HOST_WIDE_INT);
318 #ifdef OBJECT_FORMAT_ELF
319 static void sparc_elf_asm_named_section (const char *, unsigned int, tree);
322 static int sparc_adjust_cost (rtx, rtx, rtx, int);
323 static int sparc_issue_rate (void);
324 static void sparc_sched_init (FILE *, int, int);
325 static int sparc_use_sched_lookahead (void);
327 static void emit_soft_tfmode_libcall (const char *, int, rtx *);
328 static void emit_soft_tfmode_binop (enum rtx_code, rtx *);
329 static void emit_soft_tfmode_unop (enum rtx_code, rtx *);
330 static void emit_soft_tfmode_cvt (enum rtx_code, rtx *);
331 static void emit_hard_tfmode_operation (enum rtx_code, rtx *);
333 static bool sparc_function_ok_for_sibcall (tree, tree);
334 static void sparc_init_libfuncs (void);
335 static void sparc_output_mi_thunk (FILE *, tree, HOST_WIDE_INT,
336 HOST_WIDE_INT, tree);
337 static bool sparc_can_output_mi_thunk (tree, HOST_WIDE_INT,
338 HOST_WIDE_INT, tree);
339 static struct machine_function * sparc_init_machine_status (void);
340 static bool sparc_cannot_force_const_mem (rtx);
341 static rtx sparc_tls_get_addr (void);
342 static rtx sparc_tls_got (void);
343 static const char *get_some_local_dynamic_name (void);
344 static int get_some_local_dynamic_name_1 (rtx *, void *);
345 static bool sparc_rtx_costs (rtx, int, int, int *);
346 static bool sparc_promote_prototypes (tree);
347 static rtx sparc_struct_value_rtx (tree, int);
348 static bool sparc_return_in_memory (tree, tree);
349 static bool sparc_strict_argument_naming (CUMULATIVE_ARGS *);
350 static tree sparc_gimplify_va_arg (tree, tree, tree *, tree *);
351 static bool sparc_pass_by_reference (CUMULATIVE_ARGS *,
352 enum machine_mode, tree, bool);
353 #ifdef SUBTARGET_ATTRIBUTE_TABLE
354 const struct attribute_spec sparc_attribute_table[];
357 /* Option handling. */
359 /* Code model option as passed by user. */
360 const char *sparc_cmodel_string;
362 enum cmodel sparc_cmodel;
364 char sparc_hard_reg_printed[8];
366 struct sparc_cpu_select sparc_select[] =
368 /* switch name, tune arch */
369 { (char *)0, "default", 1, 1 },
370 { (char *)0, "-mcpu=", 1, 1 },
371 { (char *)0, "-mtune=", 1, 0 },
375 /* CPU type. This is set from TARGET_CPU_DEFAULT and -m{cpu,tune}=xxx. */
376 enum processor_type sparc_cpu;
378 /* Initialize the GCC target structure. */
380 /* The sparc default is to use .half rather than .short for aligned
381 HI objects. Use .word instead of .long on non-ELF systems. */
382 #undef TARGET_ASM_ALIGNED_HI_OP
383 #define TARGET_ASM_ALIGNED_HI_OP "\t.half\t"
384 #ifndef OBJECT_FORMAT_ELF
385 #undef TARGET_ASM_ALIGNED_SI_OP
386 #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t"
389 #undef TARGET_ASM_UNALIGNED_HI_OP
390 #define TARGET_ASM_UNALIGNED_HI_OP "\t.uahalf\t"
391 #undef TARGET_ASM_UNALIGNED_SI_OP
392 #define TARGET_ASM_UNALIGNED_SI_OP "\t.uaword\t"
393 #undef TARGET_ASM_UNALIGNED_DI_OP
394 #define TARGET_ASM_UNALIGNED_DI_OP "\t.uaxword\t"
396 /* The target hook has to handle DI-mode values. */
397 #undef TARGET_ASM_INTEGER
398 #define TARGET_ASM_INTEGER sparc_assemble_integer
400 #undef TARGET_ASM_FUNCTION_PROLOGUE
401 #define TARGET_ASM_FUNCTION_PROLOGUE sparc_asm_function_prologue
402 #undef TARGET_ASM_FUNCTION_EPILOGUE
403 #define TARGET_ASM_FUNCTION_EPILOGUE sparc_asm_function_epilogue
405 #undef TARGET_SCHED_ADJUST_COST
406 #define TARGET_SCHED_ADJUST_COST sparc_adjust_cost
407 #undef TARGET_SCHED_ISSUE_RATE
408 #define TARGET_SCHED_ISSUE_RATE sparc_issue_rate
409 #undef TARGET_SCHED_INIT
410 #define TARGET_SCHED_INIT sparc_sched_init
411 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
412 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD sparc_use_sched_lookahead
414 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
415 #define TARGET_FUNCTION_OK_FOR_SIBCALL sparc_function_ok_for_sibcall
417 #undef TARGET_INIT_LIBFUNCS
418 #define TARGET_INIT_LIBFUNCS sparc_init_libfuncs
421 #undef TARGET_HAVE_TLS
422 #define TARGET_HAVE_TLS true
424 #undef TARGET_CANNOT_FORCE_CONST_MEM
425 #define TARGET_CANNOT_FORCE_CONST_MEM sparc_cannot_force_const_mem
427 #undef TARGET_ASM_OUTPUT_MI_THUNK
428 #define TARGET_ASM_OUTPUT_MI_THUNK sparc_output_mi_thunk
429 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
430 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK sparc_can_output_mi_thunk
432 #undef TARGET_RTX_COSTS
433 #define TARGET_RTX_COSTS sparc_rtx_costs
434 #undef TARGET_ADDRESS_COST
435 #define TARGET_ADDRESS_COST hook_int_rtx_0
437 /* This is only needed for TARGET_ARCH64, but since PROMOTE_FUNCTION_MODE is a
438 no-op for TARGET_ARCH32 this is ok. Otherwise we'd need to add a runtime
439 test for this value. */
440 #undef TARGET_PROMOTE_FUNCTION_ARGS
441 #define TARGET_PROMOTE_FUNCTION_ARGS hook_bool_tree_true
443 /* This is only needed for TARGET_ARCH64, but since PROMOTE_FUNCTION_MODE is a
444 no-op for TARGET_ARCH32 this is ok. Otherwise we'd need to add a runtime
445 test for this value. */
446 #undef TARGET_PROMOTE_FUNCTION_RETURN
447 #define TARGET_PROMOTE_FUNCTION_RETURN hook_bool_tree_true
449 #undef TARGET_PROMOTE_PROTOTYPES
450 #define TARGET_PROMOTE_PROTOTYPES sparc_promote_prototypes
452 #undef TARGET_STRUCT_VALUE_RTX
453 #define TARGET_STRUCT_VALUE_RTX sparc_struct_value_rtx
454 #undef TARGET_RETURN_IN_MEMORY
455 #define TARGET_RETURN_IN_MEMORY sparc_return_in_memory
456 #undef TARGET_MUST_PASS_IN_STACK
457 #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size
458 #undef TARGET_PASS_BY_REFERENCE
459 #define TARGET_PASS_BY_REFERENCE sparc_pass_by_reference
461 #undef TARGET_EXPAND_BUILTIN_SAVEREGS
462 #define TARGET_EXPAND_BUILTIN_SAVEREGS sparc_builtin_saveregs
463 #undef TARGET_STRICT_ARGUMENT_NAMING
464 #define TARGET_STRICT_ARGUMENT_NAMING sparc_strict_argument_naming
466 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
467 #define TARGET_GIMPLIFY_VA_ARG_EXPR sparc_gimplify_va_arg
469 #ifdef SUBTARGET_INSERT_ATTRIBUTES
470 #undef TARGET_INSERT_ATTRIBUTES
471 #define TARGET_INSERT_ATTRIBUTES SUBTARGET_INSERT_ATTRIBUTES
474 #ifdef SUBTARGET_ATTRIBUTE_TABLE
475 #undef TARGET_ATTRIBUTE_TABLE
476 #define TARGET_ATTRIBUTE_TABLE sparc_attribute_table
479 struct gcc_target targetm = TARGET_INITIALIZER;
481 /* Validate and override various options, and do some machine dependent
485 sparc_override_options (void)
487 static struct code_model {
488 const char *const name;
490 } const cmodels[] = {
492 { "medlow", CM_MEDLOW },
493 { "medmid", CM_MEDMID },
494 { "medany", CM_MEDANY },
495 { "embmedany", CM_EMBMEDANY },
498 const struct code_model *cmodel;
499 /* Map TARGET_CPU_DEFAULT to value for -m{arch,tune}=. */
500 static struct cpu_default {
502 const char *const name;
503 } const cpu_default[] = {
504 /* There must be one entry here for each TARGET_CPU value. */
505 { TARGET_CPU_sparc, "cypress" },
506 { TARGET_CPU_sparclet, "tsc701" },
507 { TARGET_CPU_sparclite, "f930" },
508 { TARGET_CPU_v8, "v8" },
509 { TARGET_CPU_hypersparc, "hypersparc" },
510 { TARGET_CPU_sparclite86x, "sparclite86x" },
511 { TARGET_CPU_supersparc, "supersparc" },
512 { TARGET_CPU_v9, "v9" },
513 { TARGET_CPU_ultrasparc, "ultrasparc" },
514 { TARGET_CPU_ultrasparc3, "ultrasparc3" },
517 const struct cpu_default *def;
518 /* Table of values for -m{cpu,tune}=. */
519 static struct cpu_table {
520 const char *const name;
521 const enum processor_type processor;
524 } const cpu_table[] = {
525 { "v7", PROCESSOR_V7, MASK_ISA, 0 },
526 { "cypress", PROCESSOR_CYPRESS, MASK_ISA, 0 },
527 { "v8", PROCESSOR_V8, MASK_ISA, MASK_V8 },
528 /* TI TMS390Z55 supersparc */
529 { "supersparc", PROCESSOR_SUPERSPARC, MASK_ISA, MASK_V8 },
530 { "sparclite", PROCESSOR_SPARCLITE, MASK_ISA, MASK_SPARCLITE },
531 /* The Fujitsu MB86930 is the original sparclite chip, with no fpu.
532 The Fujitsu MB86934 is the recent sparclite chip, with an fpu. */
533 { "f930", PROCESSOR_F930, MASK_ISA|MASK_FPU, MASK_SPARCLITE },
534 { "f934", PROCESSOR_F934, MASK_ISA, MASK_SPARCLITE|MASK_FPU },
535 { "hypersparc", PROCESSOR_HYPERSPARC, MASK_ISA, MASK_V8|MASK_FPU },
536 { "sparclite86x", PROCESSOR_SPARCLITE86X, MASK_ISA|MASK_FPU,
538 { "sparclet", PROCESSOR_SPARCLET, MASK_ISA, MASK_SPARCLET },
540 { "tsc701", PROCESSOR_TSC701, MASK_ISA, MASK_SPARCLET },
541 { "v9", PROCESSOR_V9, MASK_ISA, MASK_V9 },
542 /* TI ultrasparc I, II, IIi */
543 { "ultrasparc", PROCESSOR_ULTRASPARC, MASK_ISA, MASK_V9
544 /* Although insns using %y are deprecated, it is a clear win on current
546 |MASK_DEPRECATED_V8_INSNS},
547 /* TI ultrasparc III */
548 /* ??? Check if %y issue still holds true in ultra3. */
549 { "ultrasparc3", PROCESSOR_ULTRASPARC3, MASK_ISA, MASK_V9|MASK_DEPRECATED_V8_INSNS},
552 const struct cpu_table *cpu;
553 const struct sparc_cpu_select *sel;
556 #ifndef SPARC_BI_ARCH
557 /* Check for unsupported architecture size. */
558 if (! TARGET_64BIT != DEFAULT_ARCH32_P)
559 error ("%s is not supported by this configuration",
560 DEFAULT_ARCH32_P ? "-m64" : "-m32");
563 /* We force all 64bit archs to use 128 bit long double */
564 if (TARGET_64BIT && ! TARGET_LONG_DOUBLE_128)
566 error ("-mlong-double-64 not allowed with -m64");
567 target_flags |= MASK_LONG_DOUBLE_128;
570 /* Code model selection. */
571 sparc_cmodel = SPARC_DEFAULT_CMODEL;
575 sparc_cmodel = CM_32;
578 if (sparc_cmodel_string != NULL)
582 for (cmodel = &cmodels[0]; cmodel->name; cmodel++)
583 if (strcmp (sparc_cmodel_string, cmodel->name) == 0)
585 if (cmodel->name == NULL)
586 error ("bad value (%s) for -mcmodel= switch", sparc_cmodel_string);
588 sparc_cmodel = cmodel->value;
591 error ("-mcmodel= is not supported on 32 bit systems");
594 fpu = TARGET_FPU; /* save current -mfpu status */
596 /* Set the default CPU. */
597 for (def = &cpu_default[0]; def->name; ++def)
598 if (def->cpu == TARGET_CPU_DEFAULT)
602 sparc_select[0].string = def->name;
604 for (sel = &sparc_select[0]; sel->name; ++sel)
608 for (cpu = &cpu_table[0]; cpu->name; ++cpu)
609 if (! strcmp (sel->string, cpu->name))
612 sparc_cpu = cpu->processor;
616 target_flags &= ~cpu->disable;
617 target_flags |= cpu->enable;
623 error ("bad value (%s) for %s switch", sel->string, sel->name);
627 /* If -mfpu or -mno-fpu was explicitly used, don't override with
628 the processor default. Clear MASK_FPU_SET to avoid confusing
629 the reverse mapping from switch values to names. */
632 target_flags = (target_flags & ~MASK_FPU) | fpu;
633 target_flags &= ~MASK_FPU_SET;
636 /* Don't allow -mvis if FPU is disabled. */
638 target_flags &= ~MASK_VIS;
640 /* -mvis assumes UltraSPARC+, so we are sure v9 instructions
642 -m64 also implies v9. */
643 if (TARGET_VIS || TARGET_ARCH64)
645 target_flags |= MASK_V9;
646 target_flags &= ~(MASK_V8 | MASK_SPARCLET | MASK_SPARCLITE);
649 /* Use the deprecated v8 insns for sparc64 in 32 bit mode. */
650 if (TARGET_V9 && TARGET_ARCH32)
651 target_flags |= MASK_DEPRECATED_V8_INSNS;
653 /* V8PLUS requires V9, makes no sense in 64 bit mode. */
654 if (! TARGET_V9 || TARGET_ARCH64)
655 target_flags &= ~MASK_V8PLUS;
657 /* Don't use stack biasing in 32 bit mode. */
659 target_flags &= ~MASK_STACK_BIAS;
661 /* Supply a default value for align_functions. */
662 if (align_functions == 0
663 && (sparc_cpu == PROCESSOR_ULTRASPARC
664 || sparc_cpu == PROCESSOR_ULTRASPARC3))
665 align_functions = 32;
667 /* Validate PCC_STRUCT_RETURN. */
668 if (flag_pcc_struct_return == DEFAULT_PCC_STRUCT_RETURN)
669 flag_pcc_struct_return = (TARGET_ARCH64 ? 0 : 1);
671 /* Only use .uaxword when compiling for a 64-bit target. */
673 targetm.asm_out.unaligned_op.di = NULL;
675 /* Do various machine dependent initializations. */
678 /* Acquire a unique set number for our register saves and restores. */
679 sparc_sr_alias_set = new_alias_set ();
681 /* Set up function hooks. */
682 init_machine_status = sparc_init_machine_status;
687 case PROCESSOR_CYPRESS:
688 sparc_costs = &cypress_costs;
691 case PROCESSOR_SPARCLITE:
692 case PROCESSOR_SUPERSPARC:
693 sparc_costs = &supersparc_costs;
697 case PROCESSOR_HYPERSPARC:
698 case PROCESSOR_SPARCLITE86X:
699 sparc_costs = &hypersparc_costs;
701 case PROCESSOR_SPARCLET:
702 case PROCESSOR_TSC701:
703 sparc_costs = &sparclet_costs;
706 case PROCESSOR_ULTRASPARC:
707 sparc_costs = &ultrasparc_costs;
709 case PROCESSOR_ULTRASPARC3:
710 sparc_costs = &ultrasparc3_costs;
715 #ifdef SUBTARGET_ATTRIBUTE_TABLE
716 /* Table of valid machine attributes. */
717 const struct attribute_spec sparc_attribute_table[] =
719 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
720 SUBTARGET_ATTRIBUTE_TABLE,
721 { NULL, 0, 0, false, false, false, NULL }
725 /* Miscellaneous utilities. */
727 /* Nonzero if CODE, a comparison, is suitable for use in v9 conditional move
728 or branch on register contents instructions. */
731 v9_regcmp_p (enum rtx_code code)
733 return (code == EQ || code == NE || code == GE || code == LT
734 || code == LE || code == GT);
738 /* Operand constraints. */
740 /* Return nonzero only if OP is a register of mode MODE,
744 reg_or_0_operand (rtx op, enum machine_mode mode)
746 if (register_operand (op, mode))
748 if (op == const0_rtx)
750 if (GET_MODE (op) == VOIDmode && GET_CODE (op) == CONST_DOUBLE
751 && CONST_DOUBLE_HIGH (op) == 0
752 && CONST_DOUBLE_LOW (op) == 0)
754 if (fp_zero_operand (op, mode))
759 /* Return nonzero only if OP is const1_rtx. */
762 const1_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
764 return op == const1_rtx;
767 /* Nonzero if OP is a floating point value with value 0.0. */
770 fp_zero_operand (rtx op, enum machine_mode mode)
772 if (GET_MODE_CLASS (GET_MODE (op)) != MODE_FLOAT)
774 return op == CONST0_RTX (mode);
777 /* Nonzero if OP is a register operand in floating point register. */
780 fp_register_operand (rtx op, enum machine_mode mode)
782 if (! register_operand (op, mode))
784 if (GET_CODE (op) == SUBREG)
785 op = SUBREG_REG (op);
786 return GET_CODE (op) == REG && SPARC_FP_REG_P (REGNO (op));
789 /* Nonzero if OP is a floating point constant which can
790 be loaded into an integer register using a single
791 sethi instruction. */
796 if (GET_CODE (op) == CONST_DOUBLE)
801 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
802 if (REAL_VALUES_EQUAL (r, dconst0) &&
803 ! REAL_VALUE_MINUS_ZERO (r))
805 REAL_VALUE_TO_TARGET_SINGLE (r, i);
806 if (SPARC_SETHI_P (i))
813 /* Nonzero if OP is a floating point constant which can
814 be loaded into an integer register using a single
820 if (GET_CODE (op) == CONST_DOUBLE)
825 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
826 if (REAL_VALUES_EQUAL (r, dconst0) &&
827 ! REAL_VALUE_MINUS_ZERO (r))
829 REAL_VALUE_TO_TARGET_SINGLE (r, i);
830 if (SPARC_SIMM13_P (i))
837 /* Nonzero if OP is a floating point constant which can
838 be loaded into an integer register using a high/losum
839 instruction sequence. */
842 fp_high_losum_p (rtx op)
844 /* The constraints calling this should only be in
845 SFmode move insns, so any constant which cannot
846 be moved using a single insn will do. */
847 if (GET_CODE (op) == CONST_DOUBLE)
852 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
853 if (REAL_VALUES_EQUAL (r, dconst0) &&
854 ! REAL_VALUE_MINUS_ZERO (r))
856 REAL_VALUE_TO_TARGET_SINGLE (r, i);
857 if (! SPARC_SETHI_P (i)
858 && ! SPARC_SIMM13_P (i))
865 /* Nonzero if OP is an integer register. */
868 intreg_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
870 return (register_operand (op, SImode)
871 || (TARGET_ARCH64 && register_operand (op, DImode)));
874 /* Nonzero if OP is a floating point condition code register. */
877 fcc_reg_operand (rtx op, enum machine_mode mode)
879 /* This can happen when recog is called from combine. Op may be a MEM.
880 Fail instead of calling abort in this case. */
881 if (GET_CODE (op) != REG)
884 if (mode != VOIDmode && mode != GET_MODE (op))
887 && (GET_MODE (op) != CCFPmode && GET_MODE (op) != CCFPEmode))
890 #if 0 /* ??? ==> 1 when %fcc0-3 are pseudos first. See gen_compare_reg(). */
891 if (reg_renumber == 0)
892 return REGNO (op) >= FIRST_PSEUDO_REGISTER;
893 return REGNO_OK_FOR_CCFP_P (REGNO (op));
895 return (unsigned) REGNO (op) - SPARC_FIRST_V9_FCC_REG < 4;
899 /* Nonzero if OP is a floating point condition code fcc0 register. */
902 fcc0_reg_operand (rtx op, enum machine_mode mode)
904 /* This can happen when recog is called from combine. Op may be a MEM.
905 Fail instead of calling abort in this case. */
906 if (GET_CODE (op) != REG)
909 if (mode != VOIDmode && mode != GET_MODE (op))
912 && (GET_MODE (op) != CCFPmode && GET_MODE (op) != CCFPEmode))
915 return REGNO (op) == SPARC_FCC_REG;
918 /* Nonzero if OP is an integer or floating point condition code register. */
921 icc_or_fcc_reg_operand (rtx op, enum machine_mode mode)
923 if (GET_CODE (op) == REG && REGNO (op) == SPARC_ICC_REG)
925 if (mode != VOIDmode && mode != GET_MODE (op))
928 && GET_MODE (op) != CCmode && GET_MODE (op) != CCXmode)
933 return fcc_reg_operand (op, mode);
936 /* Call insn on SPARC can take a PC-relative constant address, or any regular
940 call_operand (rtx op, enum machine_mode mode)
942 if (GET_CODE (op) != MEM)
945 return (symbolic_operand (op, mode) || memory_address_p (Pmode, op));
949 call_operand_address (rtx op, enum machine_mode mode)
951 return (symbolic_operand (op, mode) || memory_address_p (Pmode, op));
954 /* If OP is a SYMBOL_REF of a thread-local symbol, return its TLS mode,
955 otherwise return 0. */
958 tls_symbolic_operand (rtx op)
960 if (GET_CODE (op) != SYMBOL_REF)
962 return SYMBOL_REF_TLS_MODEL (op);
966 tgd_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
968 return tls_symbolic_operand (op) == TLS_MODEL_GLOBAL_DYNAMIC;
972 tld_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
974 return tls_symbolic_operand (op) == TLS_MODEL_LOCAL_DYNAMIC;
978 tie_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
980 return tls_symbolic_operand (op) == TLS_MODEL_INITIAL_EXEC;
984 tle_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
986 return tls_symbolic_operand (op) == TLS_MODEL_LOCAL_EXEC;
989 /* Returns 1 if OP is either a symbol reference or a sum of a symbol
990 reference and a constant. */
993 symbolic_operand (register rtx op, enum machine_mode mode)
995 enum machine_mode omode = GET_MODE (op);
997 if (omode != mode && omode != VOIDmode && mode != VOIDmode)
1000 switch (GET_CODE (op))
1003 return !SYMBOL_REF_TLS_MODEL (op);
1010 return (((GET_CODE (XEXP (op, 0)) == SYMBOL_REF
1011 && !SYMBOL_REF_TLS_MODEL (XEXP (op, 0)))
1012 || GET_CODE (XEXP (op, 0)) == LABEL_REF)
1013 && GET_CODE (XEXP (op, 1)) == CONST_INT);
1020 /* Return truth value of statement that OP is a symbolic memory
1021 operand of mode MODE. */
1024 symbolic_memory_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1026 if (GET_CODE (op) == SUBREG)
1027 op = SUBREG_REG (op);
1028 if (GET_CODE (op) != MEM)
1031 return ((GET_CODE (op) == SYMBOL_REF && !SYMBOL_REF_TLS_MODEL (op))
1032 || GET_CODE (op) == CONST || GET_CODE (op) == HIGH
1033 || GET_CODE (op) == LABEL_REF);
1036 /* Return truth value of statement that OP is a LABEL_REF of mode MODE. */
1039 label_ref_operand (rtx op, enum machine_mode mode)
1041 if (GET_CODE (op) != LABEL_REF)
1043 if (GET_MODE (op) != mode)
1048 /* Return 1 if the operand is an argument used in generating pic references
1049 in either the medium/low or medium/anywhere code models of sparc64. */
1052 sp64_medium_pic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1054 /* Check for (const (minus (symbol_ref:GOT)
1055 (const (minus (label) (pc))))). */
1056 if (GET_CODE (op) != CONST)
1059 if (GET_CODE (op) != MINUS)
1061 if (GET_CODE (XEXP (op, 0)) != SYMBOL_REF)
1063 /* ??? Ensure symbol is GOT. */
1064 if (GET_CODE (XEXP (op, 1)) != CONST)
1066 if (GET_CODE (XEXP (XEXP (op, 1), 0)) != MINUS)
1071 /* Return 1 if the operand is a data segment reference. This includes
1072 the readonly data segment, or in other words anything but the text segment.
1073 This is needed in the medium/anywhere code model on v9. These values
1074 are accessed with EMBMEDANY_BASE_REG. */
1077 data_segment_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1079 switch (GET_CODE (op))
1082 return ! SYMBOL_REF_FUNCTION_P (op);
1084 /* Assume canonical format of symbol + constant.
1087 return data_segment_operand (XEXP (op, 0), VOIDmode);
1093 /* Return 1 if the operand is a text segment reference.
1094 This is needed in the medium/anywhere code model on v9. */
1097 text_segment_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1099 switch (GET_CODE (op))
1104 return SYMBOL_REF_FUNCTION_P (op);
1106 /* Assume canonical format of symbol + constant.
1109 return text_segment_operand (XEXP (op, 0), VOIDmode);
1115 /* Return 1 if the operand is either a register or a memory operand that is
1119 reg_or_nonsymb_mem_operand (register rtx op, enum machine_mode mode)
1121 if (register_operand (op, mode))
1124 if (memory_operand (op, mode) && ! symbolic_memory_operand (op, mode))
1131 splittable_symbolic_memory_operand (rtx op,
1132 enum machine_mode mode ATTRIBUTE_UNUSED)
1134 if (GET_CODE (op) != MEM)
1136 if (! symbolic_operand (XEXP (op, 0), Pmode))
1142 splittable_immediate_memory_operand (rtx op,
1143 enum machine_mode mode ATTRIBUTE_UNUSED)
1145 if (GET_CODE (op) != MEM)
1147 if (! immediate_operand (XEXP (op, 0), Pmode))
1152 /* Return truth value of whether OP is EQ or NE. */
1155 eq_or_neq (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1157 return (GET_CODE (op) == EQ || GET_CODE (op) == NE);
1160 /* Return 1 if this is a comparison operator, but not an EQ, NE, GEU,
1161 or LTU for non-floating-point. We handle those specially. */
1164 normal_comp_operator (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1168 if (!COMPARISON_P (op))
1171 if (GET_MODE (XEXP (op, 0)) == CCFPmode
1172 || GET_MODE (XEXP (op, 0)) == CCFPEmode)
1175 code = GET_CODE (op);
1176 return (code != NE && code != EQ && code != GEU && code != LTU);
1179 /* Return 1 if this is a comparison operator. This allows the use of
1180 MATCH_OPERATOR to recognize all the branch insns. */
1183 noov_compare_op (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1187 if (!COMPARISON_P (op))
1190 code = GET_CODE (op);
1191 if (GET_MODE (XEXP (op, 0)) == CC_NOOVmode
1192 || GET_MODE (XEXP (op, 0)) == CCX_NOOVmode)
1193 /* These are the only branches which work with CC_NOOVmode. */
1194 return (code == EQ || code == NE || code == GE || code == LT);
1198 /* Return 1 if this is a 64-bit comparison operator. This allows the use of
1199 MATCH_OPERATOR to recognize all the branch insns. */
1202 noov_compare64_op (register rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1209 if (!COMPARISON_P (op))
1212 code = GET_CODE (op);
1213 if (GET_MODE (XEXP (op, 0)) == CCX_NOOVmode)
1214 /* These are the only branches which work with CCX_NOOVmode. */
1215 return (code == EQ || code == NE || code == GE || code == LT);
1216 return (GET_MODE (XEXP (op, 0)) == CCXmode);
1219 /* Nonzero if OP is a comparison operator suitable for use in v9
1220 conditional move or branch on register contents instructions. */
1223 v9_regcmp_op (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1227 if (!COMPARISON_P (op))
1230 code = GET_CODE (op);
1231 return v9_regcmp_p (code);
1234 /* Return 1 if this is a SIGN_EXTEND or ZERO_EXTEND operation. */
1237 extend_op (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1239 return GET_CODE (op) == SIGN_EXTEND || GET_CODE (op) == ZERO_EXTEND;
1242 /* Return nonzero if OP is an operator of mode MODE which can set
1243 the condition codes explicitly. We do not include PLUS and MINUS
1244 because these require CC_NOOVmode, which we handle explicitly. */
1247 cc_arithop (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1249 if (GET_CODE (op) == AND
1250 || GET_CODE (op) == IOR
1251 || GET_CODE (op) == XOR)
1257 /* Return nonzero if OP is an operator of mode MODE which can bitwise
1258 complement its second operand and set the condition codes explicitly. */
1261 cc_arithopn (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1263 /* XOR is not here because combine canonicalizes (xor (not ...) ...)
1264 and (xor ... (not ...)) to (not (xor ...)). */
1265 return (GET_CODE (op) == AND
1266 || GET_CODE (op) == IOR);
1269 /* Return true if OP is a register, or is a CONST_INT that can fit in a
1270 signed 13 bit immediate field. This is an acceptable SImode operand for
1271 most 3 address instructions. */
1274 arith_operand (rtx op, enum machine_mode mode)
1276 if (register_operand (op, mode))
1278 if (GET_CODE (op) != CONST_INT)
1280 return SMALL_INT32 (op);
1283 /* Return true if OP is a constant 4096 */
1286 arith_4096_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1288 if (GET_CODE (op) != CONST_INT)
1291 return INTVAL (op) == 4096;
1294 /* Return true if OP is suitable as second operand for add/sub */
1297 arith_add_operand (rtx op, enum machine_mode mode)
1299 return arith_operand (op, mode) || arith_4096_operand (op, mode);
1302 /* Return true if OP is a CONST_INT or a CONST_DOUBLE which can fit in the
1303 immediate field of OR and XOR instructions. Used for 64-bit
1304 constant formation patterns. */
1306 const64_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1308 return ((GET_CODE (op) == CONST_INT
1309 && SPARC_SIMM13_P (INTVAL (op)))
1310 #if HOST_BITS_PER_WIDE_INT != 64
1311 || (GET_CODE (op) == CONST_DOUBLE
1312 && SPARC_SIMM13_P (CONST_DOUBLE_LOW (op))
1313 && (CONST_DOUBLE_HIGH (op) ==
1314 ((CONST_DOUBLE_LOW (op) & 0x80000000) != 0 ?
1315 (HOST_WIDE_INT)-1 : 0)))
1320 /* The same, but only for sethi instructions. */
1322 const64_high_operand (rtx op, enum machine_mode mode)
1324 return ((GET_CODE (op) == CONST_INT
1325 && (INTVAL (op) & ~(HOST_WIDE_INT)0x3ff) != 0
1326 && SPARC_SETHI_P (INTVAL (op) & GET_MODE_MASK (mode))
1328 || (GET_CODE (op) == CONST_DOUBLE
1329 && CONST_DOUBLE_HIGH (op) == 0
1330 && (CONST_DOUBLE_LOW (op) & ~(HOST_WIDE_INT)0x3ff) != 0
1331 && SPARC_SETHI_P (CONST_DOUBLE_LOW (op))));
1334 /* Return true if OP is a register, or is a CONST_INT that can fit in a
1335 signed 11 bit immediate field. This is an acceptable SImode operand for
1336 the movcc instructions. */
1339 arith11_operand (rtx op, enum machine_mode mode)
1341 return (register_operand (op, mode)
1342 || (GET_CODE (op) == CONST_INT && SPARC_SIMM11_P (INTVAL (op))));
1345 /* Return true if OP is a register, or is a CONST_INT that can fit in a
1346 signed 10 bit immediate field. This is an acceptable SImode operand for
1347 the movrcc instructions. */
1350 arith10_operand (rtx op, enum machine_mode mode)
1352 return (register_operand (op, mode)
1353 || (GET_CODE (op) == CONST_INT && SPARC_SIMM10_P (INTVAL (op))));
1356 /* Return true if OP is a register, is a CONST_INT that fits in a 13 bit
1357 immediate field, or is a CONST_DOUBLE whose both parts fit in a 13 bit
1359 ARCH64: Return true if OP is a register, or is a CONST_INT or CONST_DOUBLE that
1360 can fit in a 13 bit immediate field. This is an acceptable DImode operand
1361 for most 3 address instructions. */
1364 arith_double_operand (rtx op, enum machine_mode mode)
1366 return (register_operand (op, mode)
1367 || (GET_CODE (op) == CONST_INT && SMALL_INT (op))
1369 && GET_CODE (op) == CONST_DOUBLE
1370 && (unsigned HOST_WIDE_INT) (CONST_DOUBLE_LOW (op) + 0x1000) < 0x2000
1371 && (unsigned HOST_WIDE_INT) (CONST_DOUBLE_HIGH (op) + 0x1000) < 0x2000)
1373 && GET_CODE (op) == CONST_DOUBLE
1374 && (unsigned HOST_WIDE_INT) (CONST_DOUBLE_LOW (op) + 0x1000) < 0x2000
1375 && ((CONST_DOUBLE_HIGH (op) == -1
1376 && (CONST_DOUBLE_LOW (op) & 0x1000) == 0x1000)
1377 || (CONST_DOUBLE_HIGH (op) == 0
1378 && (CONST_DOUBLE_LOW (op) & 0x1000) == 0))));
1381 /* Return true if OP is a constant 4096 for DImode on ARCH64 */
1384 arith_double_4096_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1386 return (TARGET_ARCH64 &&
1387 ((GET_CODE (op) == CONST_INT && INTVAL (op) == 4096) ||
1388 (GET_CODE (op) == CONST_DOUBLE &&
1389 CONST_DOUBLE_LOW (op) == 4096 &&
1390 CONST_DOUBLE_HIGH (op) == 0)));
1393 /* Return true if OP is suitable as second operand for add/sub in DImode */
1396 arith_double_add_operand (rtx op, enum machine_mode mode)
1398 return arith_double_operand (op, mode) || arith_double_4096_operand (op, mode);
1401 /* Return true if OP is a register, or is a CONST_INT or CONST_DOUBLE that
1402 can fit in an 11 bit immediate field. This is an acceptable DImode
1403 operand for the movcc instructions. */
1404 /* ??? Replace with arith11_operand? */
1407 arith11_double_operand (rtx op, enum machine_mode mode)
1409 return (register_operand (op, mode)
1410 || (GET_CODE (op) == CONST_DOUBLE
1411 && (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode)
1412 && (unsigned HOST_WIDE_INT) (CONST_DOUBLE_LOW (op) + 0x400) < 0x800
1413 && ((CONST_DOUBLE_HIGH (op) == -1
1414 && (CONST_DOUBLE_LOW (op) & 0x400) == 0x400)
1415 || (CONST_DOUBLE_HIGH (op) == 0
1416 && (CONST_DOUBLE_LOW (op) & 0x400) == 0)))
1417 || (GET_CODE (op) == CONST_INT
1418 && (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode)
1419 && (unsigned HOST_WIDE_INT) (INTVAL (op) + 0x400) < 0x800));
1422 /* Return true if OP is a register, or is a CONST_INT or CONST_DOUBLE that
1423 can fit in an 10 bit immediate field. This is an acceptable DImode
1424 operand for the movrcc instructions. */
1425 /* ??? Replace with arith10_operand? */
1428 arith10_double_operand (rtx op, enum machine_mode mode)
1430 return (register_operand (op, mode)
1431 || (GET_CODE (op) == CONST_DOUBLE
1432 && (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode)
1433 && (unsigned) (CONST_DOUBLE_LOW (op) + 0x200) < 0x400
1434 && ((CONST_DOUBLE_HIGH (op) == -1
1435 && (CONST_DOUBLE_LOW (op) & 0x200) == 0x200)
1436 || (CONST_DOUBLE_HIGH (op) == 0
1437 && (CONST_DOUBLE_LOW (op) & 0x200) == 0)))
1438 || (GET_CODE (op) == CONST_INT
1439 && (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode)
1440 && (unsigned HOST_WIDE_INT) (INTVAL (op) + 0x200) < 0x400));
1443 /* Return truth value of whether OP is an integer which fits the
1444 range constraining immediate operands in most three-address insns,
1445 which have a 13 bit immediate field. */
1448 small_int (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1450 return (GET_CODE (op) == CONST_INT && SMALL_INT (op));
1454 small_int_or_double (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1456 return ((GET_CODE (op) == CONST_INT && SMALL_INT (op))
1457 || (GET_CODE (op) == CONST_DOUBLE
1458 && CONST_DOUBLE_HIGH (op) == 0
1459 && SPARC_SIMM13_P (CONST_DOUBLE_LOW (op))));
1462 /* Recognize operand values for the umul instruction. That instruction sign
1463 extends immediate values just like all other sparc instructions, but
1464 interprets the extended result as an unsigned number. */
1467 uns_small_int (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1469 #if HOST_BITS_PER_WIDE_INT > 32
1470 /* All allowed constants will fit a CONST_INT. */
1471 return (GET_CODE (op) == CONST_INT
1472 && ((INTVAL (op) >= 0 && INTVAL (op) < 0x1000)
1473 || (INTVAL (op) >= 0xFFFFF000
1474 && INTVAL (op) <= 0xFFFFFFFF)));
1476 return ((GET_CODE (op) == CONST_INT && (unsigned) INTVAL (op) < 0x1000)
1477 || (GET_CODE (op) == CONST_DOUBLE
1478 && CONST_DOUBLE_HIGH (op) == 0
1479 && (unsigned) CONST_DOUBLE_LOW (op) - 0xFFFFF000 < 0x1000));
1484 uns_arith_operand (rtx op, enum machine_mode mode)
1486 return register_operand (op, mode) || uns_small_int (op, mode);
1489 /* Return truth value of statement that OP is a call-clobbered register. */
1491 clobbered_register (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1493 return (GET_CODE (op) == REG && call_used_regs[REGNO (op)]);
1496 /* Return 1 if OP is a valid operand for the source of a move insn. */
1499 input_operand (rtx op, enum machine_mode mode)
1501 /* If both modes are non-void they must be the same. */
1502 if (mode != VOIDmode && GET_MODE (op) != VOIDmode && mode != GET_MODE (op))
1505 /* Allow any one instruction integer constant, and all CONST_INT
1506 variants when we are working in DImode and !arch64. */
1507 if (GET_MODE_CLASS (mode) == MODE_INT
1508 && ((GET_CODE (op) == CONST_INT
1509 && (SPARC_SETHI_P (INTVAL (op) & GET_MODE_MASK (mode))
1510 || SPARC_SIMM13_P (INTVAL (op))
1512 && ! TARGET_ARCH64)))
1514 && GET_CODE (op) == CONST_DOUBLE
1515 && ((CONST_DOUBLE_HIGH (op) == 0
1516 && SPARC_SETHI_P (CONST_DOUBLE_LOW (op)))
1518 #if HOST_BITS_PER_WIDE_INT == 64
1519 (CONST_DOUBLE_HIGH (op) == 0
1520 && SPARC_SIMM13_P (CONST_DOUBLE_LOW (op)))
1522 (SPARC_SIMM13_P (CONST_DOUBLE_LOW (op))
1523 && (((CONST_DOUBLE_LOW (op) & 0x80000000) == 0
1524 && CONST_DOUBLE_HIGH (op) == 0)
1525 || (CONST_DOUBLE_HIGH (op) == -1
1526 && CONST_DOUBLE_LOW (op) & 0x80000000) != 0))
1531 /* If !arch64 and this is a DImode const, allow it so that
1532 the splits can be generated. */
1535 && GET_CODE (op) == CONST_DOUBLE)
1538 if (register_operand (op, mode))
1541 if (GET_MODE_CLASS (mode) == MODE_FLOAT
1542 && GET_CODE (op) == CONST_DOUBLE)
1545 /* If this is a SUBREG, look inside so that we handle
1546 paradoxical ones. */
1547 if (GET_CODE (op) == SUBREG)
1548 op = SUBREG_REG (op);
1550 /* Check for valid MEM forms. */
1551 if (GET_CODE (op) == MEM)
1552 return memory_address_p (mode, XEXP (op, 0));
1557 /* Return 1 if OP is valid for the lhs of a compare insn. */
1560 compare_operand (rtx op, enum machine_mode mode)
1562 if (GET_CODE (op) == ZERO_EXTRACT)
1563 return (register_operand (XEXP (op, 0), mode)
1564 && small_int_or_double (XEXP (op, 1), mode)
1565 && small_int_or_double (XEXP (op, 2), mode)
1566 /* This matches cmp_zero_extract. */
1568 && ((GET_CODE (XEXP (op, 2)) == CONST_INT
1569 && INTVAL (XEXP (op, 2)) > 19)
1570 || (GET_CODE (XEXP (op, 2)) == CONST_DOUBLE
1571 && CONST_DOUBLE_LOW (XEXP (op, 2)) > 19)))
1572 /* This matches cmp_zero_extract_sp64. */
1575 && ((GET_CODE (XEXP (op, 2)) == CONST_INT
1576 && INTVAL (XEXP (op, 2)) > 51)
1577 || (GET_CODE (XEXP (op, 2)) == CONST_DOUBLE
1578 && CONST_DOUBLE_LOW (XEXP (op, 2)) > 51)))));
1580 return register_operand (op, mode);
1584 /* We know it can't be done in one insn when we get here,
1585 the movsi expander guarantees this. */
1587 sparc_emit_set_const32 (rtx op0, rtx op1)
1589 enum machine_mode mode = GET_MODE (op0);
1592 if (GET_CODE (op1) == CONST_INT)
1594 HOST_WIDE_INT value = INTVAL (op1);
1596 if (SPARC_SETHI_P (value & GET_MODE_MASK (mode))
1597 || SPARC_SIMM13_P (value))
1601 /* Full 2-insn decomposition is needed. */
1602 if (reload_in_progress || reload_completed)
1605 temp = gen_reg_rtx (mode);
1607 if (GET_CODE (op1) == CONST_INT)
1609 /* Emit them as real moves instead of a HIGH/LO_SUM,
1610 this way CSE can see everything and reuse intermediate
1611 values if it wants. */
1613 && HOST_BITS_PER_WIDE_INT != 64
1614 && (INTVAL (op1) & 0x80000000) != 0)
1615 emit_insn (gen_rtx_SET
1617 immed_double_const (INTVAL (op1) & ~(HOST_WIDE_INT)0x3ff,
1620 emit_insn (gen_rtx_SET (VOIDmode, temp,
1621 GEN_INT (INTVAL (op1)
1622 & ~(HOST_WIDE_INT)0x3ff)));
1624 emit_insn (gen_rtx_SET (VOIDmode,
1626 gen_rtx_IOR (mode, temp,
1627 GEN_INT (INTVAL (op1) & 0x3ff))));
1631 /* A symbol, emit in the traditional way. */
1632 emit_insn (gen_rtx_SET (VOIDmode, temp,
1633 gen_rtx_HIGH (mode, op1)));
1634 emit_insn (gen_rtx_SET (VOIDmode,
1635 op0, gen_rtx_LO_SUM (mode, temp, op1)));
1641 /* Load OP1, a symbolic 64-bit constant, into OP0, a DImode register.
1642 If TEMP is nonzero, we are forbidden to use any other scratch
1643 registers. Otherwise, we are allowed to generate them as needed.
1645 Note that TEMP may have TImode if the code model is TARGET_CM_MEDANY
1646 or TARGET_CM_EMBMEDANY (see the reload_indi and reload_outdi patterns). */
1648 sparc_emit_set_symbolic_const64 (rtx op0, rtx op1, rtx temp)
1650 rtx temp1, temp2, temp3, temp4, temp5;
1653 if (temp && GET_MODE (temp) == TImode)
1656 temp = gen_rtx_REG (DImode, REGNO (temp));
1659 /* SPARC-V9 code-model support. */
1660 switch (sparc_cmodel)
1663 /* The range spanned by all instructions in the object is less
1664 than 2^31 bytes (2GB) and the distance from any instruction
1665 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1666 than 2^31 bytes (2GB).
1668 The executable must be in the low 4TB of the virtual address
1671 sethi %hi(symbol), %temp1
1672 or %temp1, %lo(symbol), %reg */
1674 temp1 = temp; /* op0 is allowed. */
1676 temp1 = gen_reg_rtx (DImode);
1678 emit_insn (gen_rtx_SET (VOIDmode, temp1, gen_rtx_HIGH (DImode, op1)));
1679 emit_insn (gen_rtx_SET (VOIDmode, op0, gen_rtx_LO_SUM (DImode, temp1, op1)));
1683 /* The range spanned by all instructions in the object is less
1684 than 2^31 bytes (2GB) and the distance from any instruction
1685 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1686 than 2^31 bytes (2GB).
1688 The executable must be in the low 16TB of the virtual address
1691 sethi %h44(symbol), %temp1
1692 or %temp1, %m44(symbol), %temp2
1693 sllx %temp2, 12, %temp3
1694 or %temp3, %l44(symbol), %reg */
1699 temp3 = temp; /* op0 is allowed. */
1703 temp1 = gen_reg_rtx (DImode);
1704 temp2 = gen_reg_rtx (DImode);
1705 temp3 = gen_reg_rtx (DImode);
1708 emit_insn (gen_seth44 (temp1, op1));
1709 emit_insn (gen_setm44 (temp2, temp1, op1));
1710 emit_insn (gen_rtx_SET (VOIDmode, temp3,
1711 gen_rtx_ASHIFT (DImode, temp2, GEN_INT (12))));
1712 emit_insn (gen_setl44 (op0, temp3, op1));
1716 /* The range spanned by all instructions in the object is less
1717 than 2^31 bytes (2GB) and the distance from any instruction
1718 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1719 than 2^31 bytes (2GB).
1721 The executable can be placed anywhere in the virtual address
1724 sethi %hh(symbol), %temp1
1725 sethi %lm(symbol), %temp2
1726 or %temp1, %hm(symbol), %temp3
1727 sllx %temp3, 32, %temp4
1728 or %temp4, %temp2, %temp5
1729 or %temp5, %lo(symbol), %reg */
1732 /* It is possible that one of the registers we got for operands[2]
1733 might coincide with that of operands[0] (which is why we made
1734 it TImode). Pick the other one to use as our scratch. */
1735 if (rtx_equal_p (temp, op0))
1738 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1743 temp2 = temp; /* op0 is _not_ allowed, see above. */
1750 temp1 = gen_reg_rtx (DImode);
1751 temp2 = gen_reg_rtx (DImode);
1752 temp3 = gen_reg_rtx (DImode);
1753 temp4 = gen_reg_rtx (DImode);
1754 temp5 = gen_reg_rtx (DImode);
1757 emit_insn (gen_sethh (temp1, op1));
1758 emit_insn (gen_setlm (temp2, op1));
1759 emit_insn (gen_sethm (temp3, temp1, op1));
1760 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1761 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1762 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1763 gen_rtx_PLUS (DImode, temp4, temp2)));
1764 emit_insn (gen_setlo (op0, temp5, op1));
1768 /* Old old old backwards compatibility kruft here.
1769 Essentially it is MEDLOW with a fixed 64-bit
1770 virtual base added to all data segment addresses.
1771 Text-segment stuff is computed like MEDANY, we can't
1772 reuse the code above because the relocation knobs
1775 Data segment: sethi %hi(symbol), %temp1
1776 add %temp1, EMBMEDANY_BASE_REG, %temp2
1777 or %temp2, %lo(symbol), %reg */
1778 if (data_segment_operand (op1, GET_MODE (op1)))
1782 temp1 = temp; /* op0 is allowed. */
1787 temp1 = gen_reg_rtx (DImode);
1788 temp2 = gen_reg_rtx (DImode);
1791 emit_insn (gen_embmedany_sethi (temp1, op1));
1792 emit_insn (gen_embmedany_brsum (temp2, temp1));
1793 emit_insn (gen_embmedany_losum (op0, temp2, op1));
1796 /* Text segment: sethi %uhi(symbol), %temp1
1797 sethi %hi(symbol), %temp2
1798 or %temp1, %ulo(symbol), %temp3
1799 sllx %temp3, 32, %temp4
1800 or %temp4, %temp2, %temp5
1801 or %temp5, %lo(symbol), %reg */
1806 /* It is possible that one of the registers we got for operands[2]
1807 might coincide with that of operands[0] (which is why we made
1808 it TImode). Pick the other one to use as our scratch. */
1809 if (rtx_equal_p (temp, op0))
1812 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1817 temp2 = temp; /* op0 is _not_ allowed, see above. */
1824 temp1 = gen_reg_rtx (DImode);
1825 temp2 = gen_reg_rtx (DImode);
1826 temp3 = gen_reg_rtx (DImode);
1827 temp4 = gen_reg_rtx (DImode);
1828 temp5 = gen_reg_rtx (DImode);
1831 emit_insn (gen_embmedany_textuhi (temp1, op1));
1832 emit_insn (gen_embmedany_texthi (temp2, op1));
1833 emit_insn (gen_embmedany_textulo (temp3, temp1, op1));
1834 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1835 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1836 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1837 gen_rtx_PLUS (DImode, temp4, temp2)));
1838 emit_insn (gen_embmedany_textlo (op0, temp5, op1));
1847 /* These avoid problems when cross compiling. If we do not
1848 go through all this hair then the optimizer will see
1849 invalid REG_EQUAL notes or in some cases none at all. */
1850 static void sparc_emit_set_safe_HIGH64 (rtx, HOST_WIDE_INT);
1851 static rtx gen_safe_SET64 (rtx, HOST_WIDE_INT);
1852 static rtx gen_safe_OR64 (rtx, HOST_WIDE_INT);
1853 static rtx gen_safe_XOR64 (rtx, HOST_WIDE_INT);
1855 #if HOST_BITS_PER_WIDE_INT == 64
1856 #define GEN_HIGHINT64(__x) GEN_INT ((__x) & ~(HOST_WIDE_INT)0x3ff)
1857 #define GEN_INT64(__x) GEN_INT (__x)
1859 #define GEN_HIGHINT64(__x) \
1860 immed_double_const ((__x) & ~(HOST_WIDE_INT)0x3ff, 0, DImode)
1861 #define GEN_INT64(__x) \
1862 immed_double_const ((__x) & 0xffffffff, \
1863 ((__x) & 0x80000000 ? -1 : 0), DImode)
1866 /* The optimizer is not to assume anything about exactly
1867 which bits are set for a HIGH, they are unspecified.
1868 Unfortunately this leads to many missed optimizations
1869 during CSE. We mask out the non-HIGH bits, and matches
1870 a plain movdi, to alleviate this problem. */
1872 sparc_emit_set_safe_HIGH64 (rtx dest, HOST_WIDE_INT val)
1874 emit_insn (gen_rtx_SET (VOIDmode, dest, GEN_HIGHINT64 (val)));
1878 gen_safe_SET64 (rtx dest, HOST_WIDE_INT val)
1880 return gen_rtx_SET (VOIDmode, dest, GEN_INT64 (val));
1884 gen_safe_OR64 (rtx src, HOST_WIDE_INT val)
1886 return gen_rtx_IOR (DImode, src, GEN_INT64 (val));
1890 gen_safe_XOR64 (rtx src, HOST_WIDE_INT val)
1892 return gen_rtx_XOR (DImode, src, GEN_INT64 (val));
1895 /* Worker routines for 64-bit constant formation on arch64.
1896 One of the key things to be doing in these emissions is
1897 to create as many temp REGs as possible. This makes it
1898 possible for half-built constants to be used later when
1899 such values are similar to something required later on.
1900 Without doing this, the optimizer cannot see such
1903 static void sparc_emit_set_const64_quick1 (rtx, rtx,
1904 unsigned HOST_WIDE_INT, int);
1907 sparc_emit_set_const64_quick1 (rtx op0, rtx temp,
1908 unsigned HOST_WIDE_INT low_bits, int is_neg)
1910 unsigned HOST_WIDE_INT high_bits;
1913 high_bits = (~low_bits) & 0xffffffff;
1915 high_bits = low_bits;
1917 sparc_emit_set_safe_HIGH64 (temp, high_bits);
1920 emit_insn (gen_rtx_SET (VOIDmode, op0,
1921 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1925 /* If we are XOR'ing with -1, then we should emit a one's complement
1926 instead. This way the combiner will notice logical operations
1927 such as ANDN later on and substitute. */
1928 if ((low_bits & 0x3ff) == 0x3ff)
1930 emit_insn (gen_rtx_SET (VOIDmode, op0,
1931 gen_rtx_NOT (DImode, temp)));
1935 emit_insn (gen_rtx_SET (VOIDmode, op0,
1936 gen_safe_XOR64 (temp,
1937 (-(HOST_WIDE_INT)0x400
1938 | (low_bits & 0x3ff)))));
1943 static void sparc_emit_set_const64_quick2 (rtx, rtx, unsigned HOST_WIDE_INT,
1944 unsigned HOST_WIDE_INT, int);
1947 sparc_emit_set_const64_quick2 (rtx op0, rtx temp,
1948 unsigned HOST_WIDE_INT high_bits,
1949 unsigned HOST_WIDE_INT low_immediate,
1954 if ((high_bits & 0xfffffc00) != 0)
1956 sparc_emit_set_safe_HIGH64 (temp, high_bits);
1957 if ((high_bits & ~0xfffffc00) != 0)
1958 emit_insn (gen_rtx_SET (VOIDmode, op0,
1959 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1965 emit_insn (gen_safe_SET64 (temp, high_bits));
1969 /* Now shift it up into place. */
1970 emit_insn (gen_rtx_SET (VOIDmode, op0,
1971 gen_rtx_ASHIFT (DImode, temp2,
1972 GEN_INT (shift_count))));
1974 /* If there is a low immediate part piece, finish up by
1975 putting that in as well. */
1976 if (low_immediate != 0)
1977 emit_insn (gen_rtx_SET (VOIDmode, op0,
1978 gen_safe_OR64 (op0, low_immediate)));
1981 static void sparc_emit_set_const64_longway (rtx, rtx, unsigned HOST_WIDE_INT,
1982 unsigned HOST_WIDE_INT);
1984 /* Full 64-bit constant decomposition. Even though this is the
1985 'worst' case, we still optimize a few things away. */
1987 sparc_emit_set_const64_longway (rtx op0, rtx temp,
1988 unsigned HOST_WIDE_INT high_bits,
1989 unsigned HOST_WIDE_INT low_bits)
1993 if (reload_in_progress || reload_completed)
1996 sub_temp = gen_reg_rtx (DImode);
1998 if ((high_bits & 0xfffffc00) != 0)
2000 sparc_emit_set_safe_HIGH64 (temp, high_bits);
2001 if ((high_bits & ~0xfffffc00) != 0)
2002 emit_insn (gen_rtx_SET (VOIDmode,
2004 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
2010 emit_insn (gen_safe_SET64 (temp, high_bits));
2014 if (!reload_in_progress && !reload_completed)
2016 rtx temp2 = gen_reg_rtx (DImode);
2017 rtx temp3 = gen_reg_rtx (DImode);
2018 rtx temp4 = gen_reg_rtx (DImode);
2020 emit_insn (gen_rtx_SET (VOIDmode, temp4,
2021 gen_rtx_ASHIFT (DImode, sub_temp,
2024 sparc_emit_set_safe_HIGH64 (temp2, low_bits);
2025 if ((low_bits & ~0xfffffc00) != 0)
2027 emit_insn (gen_rtx_SET (VOIDmode, temp3,
2028 gen_safe_OR64 (temp2, (low_bits & 0x3ff))));
2029 emit_insn (gen_rtx_SET (VOIDmode, op0,
2030 gen_rtx_PLUS (DImode, temp4, temp3)));
2034 emit_insn (gen_rtx_SET (VOIDmode, op0,
2035 gen_rtx_PLUS (DImode, temp4, temp2)));
2040 rtx low1 = GEN_INT ((low_bits >> (32 - 12)) & 0xfff);
2041 rtx low2 = GEN_INT ((low_bits >> (32 - 12 - 12)) & 0xfff);
2042 rtx low3 = GEN_INT ((low_bits >> (32 - 12 - 12 - 8)) & 0x0ff);
2045 /* We are in the middle of reload, so this is really
2046 painful. However we do still make an attempt to
2047 avoid emitting truly stupid code. */
2048 if (low1 != const0_rtx)
2050 emit_insn (gen_rtx_SET (VOIDmode, op0,
2051 gen_rtx_ASHIFT (DImode, sub_temp,
2052 GEN_INT (to_shift))));
2053 emit_insn (gen_rtx_SET (VOIDmode, op0,
2054 gen_rtx_IOR (DImode, op0, low1)));
2062 if (low2 != const0_rtx)
2064 emit_insn (gen_rtx_SET (VOIDmode, op0,
2065 gen_rtx_ASHIFT (DImode, sub_temp,
2066 GEN_INT (to_shift))));
2067 emit_insn (gen_rtx_SET (VOIDmode, op0,
2068 gen_rtx_IOR (DImode, op0, low2)));
2076 emit_insn (gen_rtx_SET (VOIDmode, op0,
2077 gen_rtx_ASHIFT (DImode, sub_temp,
2078 GEN_INT (to_shift))));
2079 if (low3 != const0_rtx)
2080 emit_insn (gen_rtx_SET (VOIDmode, op0,
2081 gen_rtx_IOR (DImode, op0, low3)));
2086 /* Analyze a 64-bit constant for certain properties. */
2087 static void analyze_64bit_constant (unsigned HOST_WIDE_INT,
2088 unsigned HOST_WIDE_INT,
2089 int *, int *, int *);
2092 analyze_64bit_constant (unsigned HOST_WIDE_INT high_bits,
2093 unsigned HOST_WIDE_INT low_bits,
2094 int *hbsp, int *lbsp, int *abbasp)
2096 int lowest_bit_set, highest_bit_set, all_bits_between_are_set;
2099 lowest_bit_set = highest_bit_set = -1;
2103 if ((lowest_bit_set == -1)
2104 && ((low_bits >> i) & 1))
2106 if ((highest_bit_set == -1)
2107 && ((high_bits >> (32 - i - 1)) & 1))
2108 highest_bit_set = (64 - i - 1);
2111 && ((highest_bit_set == -1)
2112 || (lowest_bit_set == -1)));
2118 if ((lowest_bit_set == -1)
2119 && ((high_bits >> i) & 1))
2120 lowest_bit_set = i + 32;
2121 if ((highest_bit_set == -1)
2122 && ((low_bits >> (32 - i - 1)) & 1))
2123 highest_bit_set = 32 - i - 1;
2126 && ((highest_bit_set == -1)
2127 || (lowest_bit_set == -1)));
2129 /* If there are no bits set this should have gone out
2130 as one instruction! */
2131 if (lowest_bit_set == -1
2132 || highest_bit_set == -1)
2134 all_bits_between_are_set = 1;
2135 for (i = lowest_bit_set; i <= highest_bit_set; i++)
2139 if ((low_bits & (1 << i)) != 0)
2144 if ((high_bits & (1 << (i - 32))) != 0)
2147 all_bits_between_are_set = 0;
2150 *hbsp = highest_bit_set;
2151 *lbsp = lowest_bit_set;
2152 *abbasp = all_bits_between_are_set;
2155 static int const64_is_2insns (unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT);
2158 const64_is_2insns (unsigned HOST_WIDE_INT high_bits,
2159 unsigned HOST_WIDE_INT low_bits)
2161 int highest_bit_set, lowest_bit_set, all_bits_between_are_set;
2164 || high_bits == 0xffffffff)
2167 analyze_64bit_constant (high_bits, low_bits,
2168 &highest_bit_set, &lowest_bit_set,
2169 &all_bits_between_are_set);
2171 if ((highest_bit_set == 63
2172 || lowest_bit_set == 0)
2173 && all_bits_between_are_set != 0)
2176 if ((highest_bit_set - lowest_bit_set) < 21)
2182 static unsigned HOST_WIDE_INT create_simple_focus_bits (unsigned HOST_WIDE_INT,
2183 unsigned HOST_WIDE_INT,
2186 static unsigned HOST_WIDE_INT
2187 create_simple_focus_bits (unsigned HOST_WIDE_INT high_bits,
2188 unsigned HOST_WIDE_INT low_bits,
2189 int lowest_bit_set, int shift)
2191 HOST_WIDE_INT hi, lo;
2193 if (lowest_bit_set < 32)
2195 lo = (low_bits >> lowest_bit_set) << shift;
2196 hi = ((high_bits << (32 - lowest_bit_set)) << shift);
2201 hi = ((high_bits >> (lowest_bit_set - 32)) << shift);
2208 /* Here we are sure to be arch64 and this is an integer constant
2209 being loaded into a register. Emit the most efficient
2210 insn sequence possible. Detection of all the 1-insn cases
2211 has been done already. */
2213 sparc_emit_set_const64 (rtx op0, rtx op1)
2215 unsigned HOST_WIDE_INT high_bits, low_bits;
2216 int lowest_bit_set, highest_bit_set;
2217 int all_bits_between_are_set;
2220 /* Sanity check that we know what we are working with. */
2221 if (! TARGET_ARCH64)
2224 if (GET_CODE (op0) != SUBREG)
2226 if (GET_CODE (op0) != REG
2227 || (REGNO (op0) >= SPARC_FIRST_FP_REG
2228 && REGNO (op0) <= SPARC_LAST_V9_FP_REG))
2232 if (reload_in_progress || reload_completed)
2235 if (GET_CODE (op1) != CONST_DOUBLE
2236 && GET_CODE (op1) != CONST_INT)
2238 sparc_emit_set_symbolic_const64 (op0, op1, temp);
2243 temp = gen_reg_rtx (DImode);
2245 if (GET_CODE (op1) == CONST_DOUBLE)
2247 #if HOST_BITS_PER_WIDE_INT == 64
2248 high_bits = (CONST_DOUBLE_LOW (op1) >> 32) & 0xffffffff;
2249 low_bits = CONST_DOUBLE_LOW (op1) & 0xffffffff;
2251 high_bits = CONST_DOUBLE_HIGH (op1);
2252 low_bits = CONST_DOUBLE_LOW (op1);
2257 #if HOST_BITS_PER_WIDE_INT == 64
2258 high_bits = ((INTVAL (op1) >> 32) & 0xffffffff);
2259 low_bits = (INTVAL (op1) & 0xffffffff);
2261 high_bits = ((INTVAL (op1) < 0) ?
2264 low_bits = INTVAL (op1);
2268 /* low_bits bits 0 --> 31
2269 high_bits bits 32 --> 63 */
2271 analyze_64bit_constant (high_bits, low_bits,
2272 &highest_bit_set, &lowest_bit_set,
2273 &all_bits_between_are_set);
2275 /* First try for a 2-insn sequence. */
2277 /* These situations are preferred because the optimizer can
2278 * do more things with them:
2280 * sllx %reg, shift, %reg
2282 * srlx %reg, shift, %reg
2283 * 3) mov some_small_const, %reg
2284 * sllx %reg, shift, %reg
2286 if (((highest_bit_set == 63
2287 || lowest_bit_set == 0)
2288 && all_bits_between_are_set != 0)
2289 || ((highest_bit_set - lowest_bit_set) < 12))
2291 HOST_WIDE_INT the_const = -1;
2292 int shift = lowest_bit_set;
2294 if ((highest_bit_set != 63
2295 && lowest_bit_set != 0)
2296 || all_bits_between_are_set == 0)
2299 create_simple_focus_bits (high_bits, low_bits,
2302 else if (lowest_bit_set == 0)
2303 shift = -(63 - highest_bit_set);
2305 if (! SPARC_SIMM13_P (the_const))
2308 emit_insn (gen_safe_SET64 (temp, the_const));
2310 emit_insn (gen_rtx_SET (VOIDmode,
2312 gen_rtx_ASHIFT (DImode,
2316 emit_insn (gen_rtx_SET (VOIDmode,
2318 gen_rtx_LSHIFTRT (DImode,
2320 GEN_INT (-shift))));
2326 /* Now a range of 22 or less bits set somewhere.
2327 * 1) sethi %hi(focus_bits), %reg
2328 * sllx %reg, shift, %reg
2329 * 2) sethi %hi(focus_bits), %reg
2330 * srlx %reg, shift, %reg
2332 if ((highest_bit_set - lowest_bit_set) < 21)
2334 unsigned HOST_WIDE_INT focus_bits =
2335 create_simple_focus_bits (high_bits, low_bits,
2336 lowest_bit_set, 10);
2338 if (! SPARC_SETHI_P (focus_bits))
2341 sparc_emit_set_safe_HIGH64 (temp, focus_bits);
2343 /* If lowest_bit_set == 10 then a sethi alone could have done it. */
2344 if (lowest_bit_set < 10)
2345 emit_insn (gen_rtx_SET (VOIDmode,
2347 gen_rtx_LSHIFTRT (DImode, temp,
2348 GEN_INT (10 - lowest_bit_set))));
2349 else if (lowest_bit_set > 10)
2350 emit_insn (gen_rtx_SET (VOIDmode,
2352 gen_rtx_ASHIFT (DImode, temp,
2353 GEN_INT (lowest_bit_set - 10))));
2359 /* 1) sethi %hi(low_bits), %reg
2360 * or %reg, %lo(low_bits), %reg
2361 * 2) sethi %hi(~low_bits), %reg
2362 * xor %reg, %lo(-0x400 | (low_bits & 0x3ff)), %reg
2365 || high_bits == 0xffffffff)
2367 sparc_emit_set_const64_quick1 (op0, temp, low_bits,
2368 (high_bits == 0xffffffff));
2372 /* Now, try 3-insn sequences. */
2374 /* 1) sethi %hi(high_bits), %reg
2375 * or %reg, %lo(high_bits), %reg
2376 * sllx %reg, 32, %reg
2380 sparc_emit_set_const64_quick2 (op0, temp, high_bits, 0, 32);
2384 /* We may be able to do something quick
2385 when the constant is negated, so try that. */
2386 if (const64_is_2insns ((~high_bits) & 0xffffffff,
2387 (~low_bits) & 0xfffffc00))
2389 /* NOTE: The trailing bits get XOR'd so we need the
2390 non-negated bits, not the negated ones. */
2391 unsigned HOST_WIDE_INT trailing_bits = low_bits & 0x3ff;
2393 if ((((~high_bits) & 0xffffffff) == 0
2394 && ((~low_bits) & 0x80000000) == 0)
2395 || (((~high_bits) & 0xffffffff) == 0xffffffff
2396 && ((~low_bits) & 0x80000000) != 0))
2398 int fast_int = (~low_bits & 0xffffffff);
2400 if ((SPARC_SETHI_P (fast_int)
2401 && (~high_bits & 0xffffffff) == 0)
2402 || SPARC_SIMM13_P (fast_int))
2403 emit_insn (gen_safe_SET64 (temp, fast_int));
2405 sparc_emit_set_const64 (temp, GEN_INT64 (fast_int));
2410 #if HOST_BITS_PER_WIDE_INT == 64
2411 negated_const = GEN_INT (((~low_bits) & 0xfffffc00) |
2412 (((HOST_WIDE_INT)((~high_bits) & 0xffffffff))<<32));
2414 negated_const = immed_double_const ((~low_bits) & 0xfffffc00,
2415 (~high_bits) & 0xffffffff,
2418 sparc_emit_set_const64 (temp, negated_const);
2421 /* If we are XOR'ing with -1, then we should emit a one's complement
2422 instead. This way the combiner will notice logical operations
2423 such as ANDN later on and substitute. */
2424 if (trailing_bits == 0x3ff)
2426 emit_insn (gen_rtx_SET (VOIDmode, op0,
2427 gen_rtx_NOT (DImode, temp)));
2431 emit_insn (gen_rtx_SET (VOIDmode,
2433 gen_safe_XOR64 (temp,
2434 (-0x400 | trailing_bits))));
2439 /* 1) sethi %hi(xxx), %reg
2440 * or %reg, %lo(xxx), %reg
2441 * sllx %reg, yyy, %reg
2443 * ??? This is just a generalized version of the low_bits==0
2444 * thing above, FIXME...
2446 if ((highest_bit_set - lowest_bit_set) < 32)
2448 unsigned HOST_WIDE_INT focus_bits =
2449 create_simple_focus_bits (high_bits, low_bits,
2452 /* We can't get here in this state. */
2453 if (highest_bit_set < 32
2454 || lowest_bit_set >= 32)
2457 /* So what we know is that the set bits straddle the
2458 middle of the 64-bit word. */
2459 sparc_emit_set_const64_quick2 (op0, temp,
2465 /* 1) sethi %hi(high_bits), %reg
2466 * or %reg, %lo(high_bits), %reg
2467 * sllx %reg, 32, %reg
2468 * or %reg, low_bits, %reg
2470 if (SPARC_SIMM13_P(low_bits)
2471 && ((int)low_bits > 0))
2473 sparc_emit_set_const64_quick2 (op0, temp, high_bits, low_bits, 32);
2477 /* The easiest way when all else fails, is full decomposition. */
2479 printf ("sparc_emit_set_const64: Hard constant [%08lx%08lx] neg[%08lx%08lx]\n",
2480 high_bits, low_bits, ~high_bits, ~low_bits);
2482 sparc_emit_set_const64_longway (op0, temp, high_bits, low_bits);
2485 /* Given a comparison code (EQ, NE, etc.) and the first operand of a COMPARE,
2486 return the mode to be used for the comparison. For floating-point,
2487 CCFP[E]mode is used. CC_NOOVmode should be used when the first operand
2488 is a PLUS, MINUS, NEG, or ASHIFT. CCmode should be used when no special
2489 processing is needed. */
2492 select_cc_mode (enum rtx_code op, rtx x, rtx y ATTRIBUTE_UNUSED)
2494 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2520 else if (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS
2521 || GET_CODE (x) == NEG || GET_CODE (x) == ASHIFT)
2523 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
2524 return CCX_NOOVmode;
2530 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
2537 /* X and Y are two things to compare using CODE. Emit the compare insn and
2538 return the rtx for the cc reg in the proper mode. */
2541 gen_compare_reg (enum rtx_code code, rtx x, rtx y)
2543 enum machine_mode mode = SELECT_CC_MODE (code, x, y);
2546 /* ??? We don't have movcc patterns so we cannot generate pseudo regs for the
2547 fcc regs (cse can't tell they're really call clobbered regs and will
2548 remove a duplicate comparison even if there is an intervening function
2549 call - it will then try to reload the cc reg via an int reg which is why
2550 we need the movcc patterns). It is possible to provide the movcc
2551 patterns by using the ldxfsr/stxfsr v9 insns. I tried it: you need two
2552 registers (say %g1,%g5) and it takes about 6 insns. A better fix would be
2553 to tell cse that CCFPE mode registers (even pseudos) are call
2556 /* ??? This is an experiment. Rather than making changes to cse which may
2557 or may not be easy/clean, we do our own cse. This is possible because
2558 we will generate hard registers. Cse knows they're call clobbered (it
2559 doesn't know the same thing about pseudos). If we guess wrong, no big
2560 deal, but if we win, great! */
2562 if (TARGET_V9 && GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2563 #if 1 /* experiment */
2566 /* We cycle through the registers to ensure they're all exercised. */
2567 static int next_fcc_reg = 0;
2568 /* Previous x,y for each fcc reg. */
2569 static rtx prev_args[4][2];
2571 /* Scan prev_args for x,y. */
2572 for (reg = 0; reg < 4; reg++)
2573 if (prev_args[reg][0] == x && prev_args[reg][1] == y)
2578 prev_args[reg][0] = x;
2579 prev_args[reg][1] = y;
2580 next_fcc_reg = (next_fcc_reg + 1) & 3;
2582 cc_reg = gen_rtx_REG (mode, reg + SPARC_FIRST_V9_FCC_REG);
2585 cc_reg = gen_reg_rtx (mode);
2586 #endif /* ! experiment */
2587 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2588 cc_reg = gen_rtx_REG (mode, SPARC_FCC_REG);
2590 cc_reg = gen_rtx_REG (mode, SPARC_ICC_REG);
2592 emit_insn (gen_rtx_SET (VOIDmode, cc_reg,
2593 gen_rtx_COMPARE (mode, x, y)));
2598 /* This function is used for v9 only.
2599 CODE is the code for an Scc's comparison.
2600 OPERANDS[0] is the target of the Scc insn.
2601 OPERANDS[1] is the value we compare against const0_rtx (which hasn't
2602 been generated yet).
2604 This function is needed to turn
2607 (gt (reg:CCX 100 %icc)
2611 (gt:DI (reg:CCX 100 %icc)
2614 IE: The instruction recognizer needs to see the mode of the comparison to
2615 find the right instruction. We could use "gt:DI" right in the
2616 define_expand, but leaving it out allows us to handle DI, SI, etc.
2618 We refer to the global sparc compare operands sparc_compare_op0 and
2619 sparc_compare_op1. */
2622 gen_v9_scc (enum rtx_code compare_code, register rtx *operands)
2627 && (GET_MODE (sparc_compare_op0) == DImode
2628 || GET_MODE (operands[0]) == DImode))
2631 op0 = sparc_compare_op0;
2632 op1 = sparc_compare_op1;
2634 /* Try to use the movrCC insns. */
2636 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT
2637 && op1 == const0_rtx
2638 && v9_regcmp_p (compare_code))
2640 /* Special case for op0 != 0. This can be done with one instruction if
2641 operands[0] == sparc_compare_op0. */
2643 if (compare_code == NE
2644 && GET_MODE (operands[0]) == DImode
2645 && rtx_equal_p (op0, operands[0]))
2647 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
2648 gen_rtx_IF_THEN_ELSE (DImode,
2649 gen_rtx_fmt_ee (compare_code, DImode,
2656 if (reg_overlap_mentioned_p (operands[0], op0))
2658 /* Handle the case where operands[0] == sparc_compare_op0.
2659 We "early clobber" the result. */
2660 op0 = gen_reg_rtx (GET_MODE (sparc_compare_op0));
2661 emit_move_insn (op0, sparc_compare_op0);
2664 emit_insn (gen_rtx_SET (VOIDmode, operands[0], const0_rtx));
2665 if (GET_MODE (op0) != DImode)
2667 temp = gen_reg_rtx (DImode);
2668 convert_move (temp, op0, 0);
2672 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
2673 gen_rtx_IF_THEN_ELSE (GET_MODE (operands[0]),
2674 gen_rtx_fmt_ee (compare_code, DImode,
2682 operands[1] = gen_compare_reg (compare_code, op0, op1);
2684 switch (GET_MODE (operands[1]))
2694 emit_insn (gen_rtx_SET (VOIDmode, operands[0], const0_rtx));
2695 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
2696 gen_rtx_IF_THEN_ELSE (GET_MODE (operands[0]),
2697 gen_rtx_fmt_ee (compare_code,
2698 GET_MODE (operands[1]),
2699 operands[1], const0_rtx),
2700 const1_rtx, operands[0])));
2705 /* Emit a conditional jump insn for the v9 architecture using comparison code
2706 CODE and jump target LABEL.
2707 This function exists to take advantage of the v9 brxx insns. */
2710 emit_v9_brxx_insn (enum rtx_code code, rtx op0, rtx label)
2712 emit_jump_insn (gen_rtx_SET (VOIDmode,
2714 gen_rtx_IF_THEN_ELSE (VOIDmode,
2715 gen_rtx_fmt_ee (code, GET_MODE (op0),
2717 gen_rtx_LABEL_REF (VOIDmode, label),
2721 /* Generate a DFmode part of a hard TFmode register.
2722 REG is the TFmode hard register, LOW is 1 for the
2723 low 64bit of the register and 0 otherwise.
2726 gen_df_reg (rtx reg, int low)
2728 int regno = REGNO (reg);
2730 if ((WORDS_BIG_ENDIAN == 0) ^ (low != 0))
2731 regno += (TARGET_ARCH64 && regno < 32) ? 1 : 2;
2732 return gen_rtx_REG (DFmode, regno);
2735 /* Generate a call to FUNC with OPERANDS. Operand 0 is the return value.
2736 Unlike normal calls, TFmode operands are passed by reference. It is
2737 assumed that no more than 3 operands are required. */
2740 emit_soft_tfmode_libcall (const char *func_name, int nargs, rtx *operands)
2742 rtx ret_slot = NULL, arg[3], func_sym;
2745 /* We only expect to be called for conversions, unary, and binary ops. */
2746 if (nargs < 2 || nargs > 3)
2749 for (i = 0; i < nargs; ++i)
2751 rtx this_arg = operands[i];
2754 /* TFmode arguments and return values are passed by reference. */
2755 if (GET_MODE (this_arg) == TFmode)
2757 int force_stack_temp;
2759 force_stack_temp = 0;
2760 if (TARGET_BUGGY_QP_LIB && i == 0)
2761 force_stack_temp = 1;
2763 if (GET_CODE (this_arg) == MEM
2764 && ! force_stack_temp)
2765 this_arg = XEXP (this_arg, 0);
2766 else if (CONSTANT_P (this_arg)
2767 && ! force_stack_temp)
2769 this_slot = force_const_mem (TFmode, this_arg);
2770 this_arg = XEXP (this_slot, 0);
2774 this_slot = assign_stack_temp (TFmode, GET_MODE_SIZE (TFmode), 0);
2776 /* Operand 0 is the return value. We'll copy it out later. */
2778 emit_move_insn (this_slot, this_arg);
2780 ret_slot = this_slot;
2782 this_arg = XEXP (this_slot, 0);
2789 func_sym = gen_rtx_SYMBOL_REF (Pmode, func_name);
2791 if (GET_MODE (operands[0]) == TFmode)
2794 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 2,
2795 arg[0], GET_MODE (arg[0]),
2796 arg[1], GET_MODE (arg[1]));
2798 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 3,
2799 arg[0], GET_MODE (arg[0]),
2800 arg[1], GET_MODE (arg[1]),
2801 arg[2], GET_MODE (arg[2]));
2804 emit_move_insn (operands[0], ret_slot);
2813 ret = emit_library_call_value (func_sym, operands[0], LCT_NORMAL,
2814 GET_MODE (operands[0]), 1,
2815 arg[1], GET_MODE (arg[1]));
2817 if (ret != operands[0])
2818 emit_move_insn (operands[0], ret);
2822 /* Expand soft-float TFmode calls to sparc abi routines. */
2825 emit_soft_tfmode_binop (enum rtx_code code, rtx *operands)
2847 emit_soft_tfmode_libcall (func, 3, operands);
2851 emit_soft_tfmode_unop (enum rtx_code code, rtx *operands)
2864 emit_soft_tfmode_libcall (func, 2, operands);
2868 emit_soft_tfmode_cvt (enum rtx_code code, rtx *operands)
2875 switch (GET_MODE (operands[1]))
2888 case FLOAT_TRUNCATE:
2889 switch (GET_MODE (operands[0]))
2903 switch (GET_MODE (operands[1]))
2916 case UNSIGNED_FLOAT:
2917 switch (GET_MODE (operands[1]))
2931 switch (GET_MODE (operands[0]))
2945 switch (GET_MODE (operands[0]))
2962 emit_soft_tfmode_libcall (func, 2, operands);
2965 /* Expand a hard-float tfmode operation. All arguments must be in
2969 emit_hard_tfmode_operation (enum rtx_code code, rtx *operands)
2973 if (GET_RTX_CLASS (code) == RTX_UNARY)
2975 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2976 op = gen_rtx_fmt_e (code, GET_MODE (operands[0]), operands[1]);
2980 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2981 operands[2] = force_reg (GET_MODE (operands[2]), operands[2]);
2982 op = gen_rtx_fmt_ee (code, GET_MODE (operands[0]),
2983 operands[1], operands[2]);
2986 if (register_operand (operands[0], VOIDmode))
2989 dest = gen_reg_rtx (GET_MODE (operands[0]));
2991 emit_insn (gen_rtx_SET (VOIDmode, dest, op));
2993 if (dest != operands[0])
2994 emit_move_insn (operands[0], dest);
2998 emit_tfmode_binop (enum rtx_code code, rtx *operands)
3000 if (TARGET_HARD_QUAD)
3001 emit_hard_tfmode_operation (code, operands);
3003 emit_soft_tfmode_binop (code, operands);
3007 emit_tfmode_unop (enum rtx_code code, rtx *operands)
3009 if (TARGET_HARD_QUAD)
3010 emit_hard_tfmode_operation (code, operands);
3012 emit_soft_tfmode_unop (code, operands);
3016 emit_tfmode_cvt (enum rtx_code code, rtx *operands)
3018 if (TARGET_HARD_QUAD)
3019 emit_hard_tfmode_operation (code, operands);
3021 emit_soft_tfmode_cvt (code, operands);
3024 /* Return nonzero if a branch/jump/call instruction will be emitting
3025 nop into its delay slot. */
3028 empty_delay_slot (rtx insn)
3032 /* If no previous instruction (should not happen), return true. */
3033 if (PREV_INSN (insn) == NULL)
3036 seq = NEXT_INSN (PREV_INSN (insn));
3037 if (GET_CODE (PATTERN (seq)) == SEQUENCE)
3043 /* Return nonzero if TRIAL can go into the call delay slot. */
3046 tls_call_delay (rtx trial)
3051 call __tls_get_addr, %tgd_call (foo)
3052 add %l7, %o0, %o0, %tgd_add (foo)
3053 while Sun as/ld does not. */
3054 if (TARGET_GNU_TLS || !TARGET_TLS)
3057 pat = PATTERN (trial);
3058 if (GET_CODE (pat) != SET || GET_CODE (SET_DEST (pat)) != PLUS)
3061 unspec = XEXP (SET_DEST (pat), 1);
3062 if (GET_CODE (unspec) != UNSPEC
3063 || (XINT (unspec, 1) != UNSPEC_TLSGD
3064 && XINT (unspec, 1) != UNSPEC_TLSLDM))
3070 /* Return nonzero if TRIAL, an insn, can be combined with a 'restore'
3071 instruction. RETURN_P is true if the v9 variant 'return' is to be
3072 considered in the test too.
3074 TRIAL must be a SET whose destination is a REG appropriate for the
3075 'restore' instruction or, if RETURN_P is true, for the 'return'
3079 eligible_for_restore_insn (rtx trial, bool return_p)
3081 rtx pat = PATTERN (trial);
3082 rtx src = SET_SRC (pat);
3084 /* The 'restore src,%g0,dest' pattern for word mode and below. */
3085 if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
3086 && arith_operand (src, GET_MODE (src)))
3089 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
3091 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (SImode);
3094 /* The 'restore src,%g0,dest' pattern for double-word mode. */
3095 else if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
3096 && arith_double_operand (src, GET_MODE (src)))
3097 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
3099 /* The 'restore src,%g0,dest' pattern for float if no FPU. */
3100 else if (! TARGET_FPU && register_operand (src, SFmode))
3103 /* The 'restore src,%g0,dest' pattern for double if no FPU. */
3104 else if (! TARGET_FPU && TARGET_ARCH64 && register_operand (src, DFmode))
3107 /* If we have the 'return' instruction, anything that does not use
3108 local or output registers and can go into a delay slot wins. */
3109 else if (return_p && TARGET_V9 && ! epilogue_renumber (&pat, 1)
3110 && (get_attr_in_uncond_branch_delay (trial)
3111 == IN_UNCOND_BRANCH_DELAY_TRUE))
3114 /* The 'restore src1,src2,dest' pattern for SImode. */
3115 else if (GET_CODE (src) == PLUS
3116 && register_operand (XEXP (src, 0), SImode)
3117 && arith_operand (XEXP (src, 1), SImode))
3120 /* The 'restore src1,src2,dest' pattern for DImode. */
3121 else if (GET_CODE (src) == PLUS
3122 && register_operand (XEXP (src, 0), DImode)
3123 && arith_double_operand (XEXP (src, 1), DImode))
3126 /* The 'restore src1,%lo(src2),dest' pattern. */
3127 else if (GET_CODE (src) == LO_SUM
3128 && ! TARGET_CM_MEDMID
3129 && ((register_operand (XEXP (src, 0), SImode)
3130 && immediate_operand (XEXP (src, 1), SImode))
3132 && register_operand (XEXP (src, 0), DImode)
3133 && immediate_operand (XEXP (src, 1), DImode))))
3136 /* The 'restore src,src,dest' pattern. */
3137 else if (GET_CODE (src) == ASHIFT
3138 && (register_operand (XEXP (src, 0), SImode)
3139 || register_operand (XEXP (src, 0), DImode))
3140 && XEXP (src, 1) == const1_rtx)
3146 /* Return nonzero if TRIAL can go into the function return's
3150 eligible_for_return_delay (rtx trial)
3154 if (GET_CODE (trial) != INSN || GET_CODE (PATTERN (trial)) != SET)
3157 if (get_attr_length (trial) != 1)
3160 /* If there are any call-saved registers, we should scan TRIAL if it
3161 does not reference them. For now just make it easy. */
3165 /* If the function uses __builtin_eh_return, the eh_return machinery
3166 occupies the delay slot. */
3167 if (current_function_calls_eh_return)
3170 /* In the case of a true leaf function, anything can go into the slot. */
3171 if (sparc_leaf_function_p)
3172 return get_attr_in_uncond_branch_delay (trial)
3173 == IN_UNCOND_BRANCH_DELAY_TRUE;
3175 pat = PATTERN (trial);
3177 /* Otherwise, only operations which can be done in tandem with
3178 a `restore' or `return' insn can go into the delay slot. */
3179 if (GET_CODE (SET_DEST (pat)) != REG
3180 || (REGNO (SET_DEST (pat)) >= 8 && REGNO (SET_DEST (pat)) < 24))
3183 /* If this instruction sets up floating point register and we have a return
3184 instruction, it can probably go in. But restore will not work
3186 if (REGNO (SET_DEST (pat)) >= 32)
3188 && ! epilogue_renumber (&pat, 1)
3189 && (get_attr_in_uncond_branch_delay (trial)
3190 == IN_UNCOND_BRANCH_DELAY_TRUE));
3192 return eligible_for_restore_insn (trial, true);
3195 /* Return nonzero if TRIAL can go into the sibling call's
3199 eligible_for_sibcall_delay (rtx trial)
3203 if (GET_CODE (trial) != INSN || GET_CODE (PATTERN (trial)) != SET)
3206 if (get_attr_length (trial) != 1)
3209 pat = PATTERN (trial);
3211 if (sparc_leaf_function_p)
3213 /* If the tail call is done using the call instruction,
3214 we have to restore %o7 in the delay slot. */
3215 if (LEAF_SIBCALL_SLOT_RESERVED_P)
3218 /* %g1 is used to build the function address */
3219 if (reg_mentioned_p (gen_rtx_REG (Pmode, 1), pat))
3225 /* Otherwise, only operations which can be done in tandem with
3226 a `restore' insn can go into the delay slot. */
3227 if (GET_CODE (SET_DEST (pat)) != REG
3228 || (REGNO (SET_DEST (pat)) >= 8 && REGNO (SET_DEST (pat)) < 24)
3229 || REGNO (SET_DEST (pat)) >= 32)
3232 /* If it mentions %o7, it can't go in, because sibcall will clobber it
3234 if (reg_mentioned_p (gen_rtx_REG (Pmode, 15), pat))
3237 return eligible_for_restore_insn (trial, false);
3241 short_branch (int uid1, int uid2)
3243 int delta = INSN_ADDRESSES (uid1) - INSN_ADDRESSES (uid2);
3245 /* Leave a few words of "slop". */
3246 if (delta >= -1023 && delta <= 1022)
3252 /* Return nonzero if REG is not used after INSN.
3253 We assume REG is a reload reg, and therefore does
3254 not live past labels or calls or jumps. */
3256 reg_unused_after (rtx reg, rtx insn)
3258 enum rtx_code code, prev_code = UNKNOWN;
3260 while ((insn = NEXT_INSN (insn)))
3262 if (prev_code == CALL_INSN && call_used_regs[REGNO (reg)])
3265 code = GET_CODE (insn);
3266 if (GET_CODE (insn) == CODE_LABEL)
3271 rtx set = single_set (insn);
3272 int in_src = set && reg_overlap_mentioned_p (reg, SET_SRC (set));
3275 if (set && reg_overlap_mentioned_p (reg, SET_DEST (set)))
3277 if (set == 0 && reg_overlap_mentioned_p (reg, PATTERN (insn)))
3285 /* Determine if it's legal to put X into the constant pool. This
3286 is not possible if X contains the address of a symbol that is
3287 not constant (TLS) or not known at final link time (PIC). */
3290 sparc_cannot_force_const_mem (rtx x)
3292 switch (GET_CODE (x))
3296 /* Accept all non-symbolic constants. */
3300 /* Labels are OK iff we are non-PIC. */
3301 return flag_pic != 0;
3304 /* 'Naked' TLS symbol references are never OK,
3305 non-TLS symbols are OK iff we are non-PIC. */
3306 if (SYMBOL_REF_TLS_MODEL (x))
3309 return flag_pic != 0;
3312 return sparc_cannot_force_const_mem (XEXP (x, 0));
3315 return sparc_cannot_force_const_mem (XEXP (x, 0))
3316 || sparc_cannot_force_const_mem (XEXP (x, 1));
3324 /* The table we use to reference PIC data. */
3325 static GTY(()) rtx global_offset_table;
3327 /* The function we use to get at it. */
3328 static GTY(()) rtx add_pc_to_pic_symbol;
3329 static GTY(()) char add_pc_to_pic_symbol_name[256];
3331 /* Ensure that we are not using patterns that are not OK with PIC. */
3339 if (GET_CODE (recog_data.operand[i]) == SYMBOL_REF
3340 || (GET_CODE (recog_data.operand[i]) == CONST
3341 && ! (GET_CODE (XEXP (recog_data.operand[i], 0)) == MINUS
3342 && (XEXP (XEXP (recog_data.operand[i], 0), 0)
3343 == global_offset_table)
3344 && (GET_CODE (XEXP (XEXP (recog_data.operand[i], 0), 1))
3353 /* Return true if X is an address which needs a temporary register when
3354 reloaded while generating PIC code. */
3357 pic_address_needs_scratch (rtx x)
3359 /* An address which is a symbolic plus a non SMALL_INT needs a temp reg. */
3360 if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS
3361 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
3362 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
3363 && ! SMALL_INT (XEXP (XEXP (x, 0), 1)))
3369 /* Determine if a given RTX is a valid constant. We already know this
3370 satisfies CONSTANT_P. */
3373 legitimate_constant_p (rtx x)
3377 switch (GET_CODE (x))
3380 /* TLS symbols are not constant. */
3381 if (SYMBOL_REF_TLS_MODEL (x))
3386 inner = XEXP (x, 0);
3388 /* Offsets of TLS symbols are never valid.
3389 Discourage CSE from creating them. */
3390 if (GET_CODE (inner) == PLUS
3391 && tls_symbolic_operand (XEXP (inner, 0)))
3396 if (GET_MODE (x) == VOIDmode)
3399 /* Floating point constants are generally not ok.
3400 The only exception is 0.0 in VIS. */
3402 && (GET_MODE (x) == SFmode
3403 || GET_MODE (x) == DFmode
3404 || GET_MODE (x) == TFmode)
3405 && fp_zero_operand (x, GET_MODE (x)))
3417 /* Determine if a given RTX is a valid constant address. */
3420 constant_address_p (rtx x)
3422 switch (GET_CODE (x))
3430 if (flag_pic && pic_address_needs_scratch (x))
3432 return legitimate_constant_p (x);
3435 return !flag_pic && legitimate_constant_p (x);
3442 /* Nonzero if the constant value X is a legitimate general operand
3443 when generating PIC code. It is given that flag_pic is on and
3444 that X satisfies CONSTANT_P or is a CONST_DOUBLE. */
3447 legitimate_pic_operand_p (rtx x)
3449 if (pic_address_needs_scratch (x))
3451 if (tls_symbolic_operand (x)
3452 || (GET_CODE (x) == CONST
3453 && GET_CODE (XEXP (x, 0)) == PLUS
3454 && tls_symbolic_operand (XEXP (XEXP (x, 0), 0))))
3459 /* Return nonzero if ADDR is a valid memory address.
3460 STRICT specifies whether strict register checking applies. */
3463 legitimate_address_p (enum machine_mode mode, rtx addr, int strict)
3465 rtx rs1 = NULL, rs2 = NULL, imm1 = NULL, imm2;
3467 if (REG_P (addr) || GET_CODE (addr) == SUBREG)
3469 else if (GET_CODE (addr) == PLUS)
3471 rs1 = XEXP (addr, 0);
3472 rs2 = XEXP (addr, 1);
3474 /* Canonicalize. REG comes first, if there are no regs,
3475 LO_SUM comes first. */
3477 && GET_CODE (rs1) != SUBREG
3479 || GET_CODE (rs2) == SUBREG
3480 || (GET_CODE (rs2) == LO_SUM && GET_CODE (rs1) != LO_SUM)))
3482 rs1 = XEXP (addr, 1);
3483 rs2 = XEXP (addr, 0);
3487 && rs1 == pic_offset_table_rtx
3489 && GET_CODE (rs2) != SUBREG
3490 && GET_CODE (rs2) != LO_SUM
3491 && GET_CODE (rs2) != MEM
3492 && !tls_symbolic_operand (rs2)
3493 && (! symbolic_operand (rs2, VOIDmode) || mode == Pmode)
3494 && (GET_CODE (rs2) != CONST_INT || SMALL_INT (rs2)))
3496 || GET_CODE (rs1) == SUBREG)
3497 && RTX_OK_FOR_OFFSET_P (rs2)))
3502 else if ((REG_P (rs1) || GET_CODE (rs1) == SUBREG)
3503 && (REG_P (rs2) || GET_CODE (rs2) == SUBREG))
3505 /* We prohibit REG + REG for TFmode when there are no quad move insns
3506 and we consequently need to split. We do this because REG+REG
3507 is not an offsettable address. If we get the situation in reload
3508 where source and destination of a movtf pattern are both MEMs with
3509 REG+REG address, then only one of them gets converted to an
3510 offsettable address. */
3512 && ! (TARGET_FPU && TARGET_ARCH64 && TARGET_HARD_QUAD))
3515 /* We prohibit REG + REG on ARCH32 if not optimizing for
3516 DFmode/DImode because then mem_min_alignment is likely to be zero
3517 after reload and the forced split would lack a matching splitter
3519 if (TARGET_ARCH32 && !optimize
3520 && (mode == DFmode || mode == DImode))
3523 else if (USE_AS_OFFSETABLE_LO10
3524 && GET_CODE (rs1) == LO_SUM
3526 && ! TARGET_CM_MEDMID
3527 && RTX_OK_FOR_OLO10_P (rs2))
3531 imm1 = XEXP (rs1, 1);
3532 rs1 = XEXP (rs1, 0);
3533 if (! CONSTANT_P (imm1) || tls_symbolic_operand (rs1))
3537 else if (GET_CODE (addr) == LO_SUM)
3539 rs1 = XEXP (addr, 0);
3540 imm1 = XEXP (addr, 1);
3542 if (! CONSTANT_P (imm1) || tls_symbolic_operand (rs1))
3545 if (USE_AS_OFFSETABLE_LO10)
3547 /* We can't allow TFmode, because an offset greater than or equal to
3548 the alignment (8) may cause the LO_SUM to overflow if !v9. */
3549 if (mode == TFmode && ! TARGET_V9)
3554 /* We prohibit LO_SUM for TFmode when there are no quad move insns
3555 and we consequently need to split. We do this because LO_SUM
3556 is not an offsettable address. If we get the situation in reload
3557 where source and destination of a movtf pattern are both MEMs with
3558 LO_SUM address, then only one of them gets converted to an
3559 offsettable address. */
3561 && ! (TARGET_FPU && TARGET_ARCH64 && TARGET_HARD_QUAD))
3565 else if (GET_CODE (addr) == CONST_INT && SMALL_INT (addr))
3570 if (GET_CODE (rs1) == SUBREG)
3571 rs1 = SUBREG_REG (rs1);
3577 if (GET_CODE (rs2) == SUBREG)
3578 rs2 = SUBREG_REG (rs2);
3585 if (!REGNO_OK_FOR_BASE_P (REGNO (rs1))
3586 || (rs2 && !REGNO_OK_FOR_BASE_P (REGNO (rs2))))
3591 if ((REGNO (rs1) >= 32
3592 && REGNO (rs1) != FRAME_POINTER_REGNUM
3593 && REGNO (rs1) < FIRST_PSEUDO_REGISTER)
3595 && (REGNO (rs2) >= 32
3596 && REGNO (rs2) != FRAME_POINTER_REGNUM
3597 && REGNO (rs2) < FIRST_PSEUDO_REGISTER)))
3603 /* Construct the SYMBOL_REF for the tls_get_offset function. */
3605 static GTY(()) rtx sparc_tls_symbol;
3607 sparc_tls_get_addr (void)
3609 if (!sparc_tls_symbol)
3610 sparc_tls_symbol = gen_rtx_SYMBOL_REF (Pmode, "__tls_get_addr");
3612 return sparc_tls_symbol;
3616 sparc_tls_got (void)
3621 current_function_uses_pic_offset_table = 1;
3622 return pic_offset_table_rtx;
3625 if (!global_offset_table)
3626 global_offset_table = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_");
3627 temp = gen_reg_rtx (Pmode);
3628 emit_move_insn (temp, global_offset_table);
3633 /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute
3634 this (thread-local) address. */
3637 legitimize_tls_address (rtx addr)
3639 rtx temp1, temp2, temp3, ret, o0, got, insn;
3644 if (GET_CODE (addr) == SYMBOL_REF)
3645 switch (SYMBOL_REF_TLS_MODEL (addr))
3647 case TLS_MODEL_GLOBAL_DYNAMIC:
3649 temp1 = gen_reg_rtx (SImode);
3650 temp2 = gen_reg_rtx (SImode);
3651 ret = gen_reg_rtx (Pmode);
3652 o0 = gen_rtx_REG (Pmode, 8);
3653 got = sparc_tls_got ();
3654 emit_insn (gen_tgd_hi22 (temp1, addr));
3655 emit_insn (gen_tgd_lo10 (temp2, temp1, addr));
3658 emit_insn (gen_tgd_add32 (o0, got, temp2, addr));
3659 insn = emit_call_insn (gen_tgd_call32 (o0, sparc_tls_get_addr (),
3664 emit_insn (gen_tgd_add64 (o0, got, temp2, addr));
3665 insn = emit_call_insn (gen_tgd_call64 (o0, sparc_tls_get_addr (),
3668 CALL_INSN_FUNCTION_USAGE (insn)
3669 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_USE (VOIDmode, o0),
3670 CALL_INSN_FUNCTION_USAGE (insn));
3671 insn = get_insns ();
3673 emit_libcall_block (insn, ret, o0, addr);
3676 case TLS_MODEL_LOCAL_DYNAMIC:
3678 temp1 = gen_reg_rtx (SImode);
3679 temp2 = gen_reg_rtx (SImode);
3680 temp3 = gen_reg_rtx (Pmode);
3681 ret = gen_reg_rtx (Pmode);
3682 o0 = gen_rtx_REG (Pmode, 8);
3683 got = sparc_tls_got ();
3684 emit_insn (gen_tldm_hi22 (temp1));
3685 emit_insn (gen_tldm_lo10 (temp2, temp1));
3688 emit_insn (gen_tldm_add32 (o0, got, temp2));
3689 insn = emit_call_insn (gen_tldm_call32 (o0, sparc_tls_get_addr (),
3694 emit_insn (gen_tldm_add64 (o0, got, temp2));
3695 insn = emit_call_insn (gen_tldm_call64 (o0, sparc_tls_get_addr (),
3698 CALL_INSN_FUNCTION_USAGE (insn)
3699 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_USE (VOIDmode, o0),
3700 CALL_INSN_FUNCTION_USAGE (insn));
3701 insn = get_insns ();
3703 emit_libcall_block (insn, temp3, o0,
3704 gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
3705 UNSPEC_TLSLD_BASE));
3706 temp1 = gen_reg_rtx (SImode);
3707 temp2 = gen_reg_rtx (SImode);
3708 emit_insn (gen_tldo_hix22 (temp1, addr));
3709 emit_insn (gen_tldo_lox10 (temp2, temp1, addr));
3711 emit_insn (gen_tldo_add32 (ret, temp3, temp2, addr));
3713 emit_insn (gen_tldo_add64 (ret, temp3, temp2, addr));
3716 case TLS_MODEL_INITIAL_EXEC:
3717 temp1 = gen_reg_rtx (SImode);
3718 temp2 = gen_reg_rtx (SImode);
3719 temp3 = gen_reg_rtx (Pmode);
3720 got = sparc_tls_got ();
3721 emit_insn (gen_tie_hi22 (temp1, addr));
3722 emit_insn (gen_tie_lo10 (temp2, temp1, addr));
3724 emit_insn (gen_tie_ld32 (temp3, got, temp2, addr));
3726 emit_insn (gen_tie_ld64 (temp3, got, temp2, addr));
3729 ret = gen_reg_rtx (Pmode);
3731 emit_insn (gen_tie_add32 (ret, gen_rtx_REG (Pmode, 7),
3734 emit_insn (gen_tie_add64 (ret, gen_rtx_REG (Pmode, 7),
3738 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp3);
3741 case TLS_MODEL_LOCAL_EXEC:
3742 temp1 = gen_reg_rtx (Pmode);
3743 temp2 = gen_reg_rtx (Pmode);
3746 emit_insn (gen_tle_hix22_sp32 (temp1, addr));
3747 emit_insn (gen_tle_lox10_sp32 (temp2, temp1, addr));
3751 emit_insn (gen_tle_hix22_sp64 (temp1, addr));
3752 emit_insn (gen_tle_lox10_sp64 (temp2, temp1, addr));
3754 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp2);
3762 abort (); /* for now ... */
3768 /* Legitimize PIC addresses. If the address is already position-independent,
3769 we return ORIG. Newly generated position-independent addresses go into a
3770 reg. This is REG if nonzero, otherwise we allocate register(s) as
3774 legitimize_pic_address (rtx orig, enum machine_mode mode ATTRIBUTE_UNUSED,
3777 if (GET_CODE (orig) == SYMBOL_REF)
3779 rtx pic_ref, address;
3784 if (reload_in_progress || reload_completed)
3787 reg = gen_reg_rtx (Pmode);
3792 /* If not during reload, allocate another temp reg here for loading
3793 in the address, so that these instructions can be optimized
3795 rtx temp_reg = ((reload_in_progress || reload_completed)
3796 ? reg : gen_reg_rtx (Pmode));
3798 /* Must put the SYMBOL_REF inside an UNSPEC here so that cse
3799 won't get confused into thinking that these two instructions
3800 are loading in the true address of the symbol. If in the
3801 future a PIC rtx exists, that should be used instead. */
3802 if (Pmode == SImode)
3804 emit_insn (gen_movsi_high_pic (temp_reg, orig));
3805 emit_insn (gen_movsi_lo_sum_pic (temp_reg, temp_reg, orig));
3809 emit_insn (gen_movdi_high_pic (temp_reg, orig));
3810 emit_insn (gen_movdi_lo_sum_pic (temp_reg, temp_reg, orig));
3817 pic_ref = gen_const_mem (Pmode,
3818 gen_rtx_PLUS (Pmode,
3819 pic_offset_table_rtx, address));
3820 current_function_uses_pic_offset_table = 1;
3821 insn = emit_move_insn (reg, pic_ref);
3822 /* Put a REG_EQUAL note on this insn, so that it can be optimized
3824 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_EQUAL, orig,
3828 else if (GET_CODE (orig) == CONST)
3832 if (GET_CODE (XEXP (orig, 0)) == PLUS
3833 && XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx)
3838 if (reload_in_progress || reload_completed)
3841 reg = gen_reg_rtx (Pmode);
3844 if (GET_CODE (XEXP (orig, 0)) == PLUS)
3846 base = legitimize_pic_address (XEXP (XEXP (orig, 0), 0), Pmode, reg);
3847 offset = legitimize_pic_address (XEXP (XEXP (orig, 0), 1), Pmode,
3848 base == reg ? 0 : reg);
3853 if (GET_CODE (offset) == CONST_INT)
3855 if (SMALL_INT (offset))
3856 return plus_constant (base, INTVAL (offset));
3857 else if (! reload_in_progress && ! reload_completed)
3858 offset = force_reg (Pmode, offset);
3860 /* If we reach here, then something is seriously wrong. */
3863 return gen_rtx_PLUS (Pmode, base, offset);
3865 else if (GET_CODE (orig) == LABEL_REF)
3866 /* ??? Why do we do this? */
3867 /* Now movsi_pic_label_ref uses it, but we ought to be checking that
3868 the register is live instead, in case it is eliminated. */
3869 current_function_uses_pic_offset_table = 1;
3874 /* Try machine-dependent ways of modifying an illegitimate address X
3875 to be legitimate. If we find one, return the new, valid address.
3877 OLDX is the address as it was before break_out_memory_refs was called.
3878 In some cases it is useful to look at this to decide what needs to be done.
3880 MODE is the mode of the operand pointed to by X. */
3883 legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED, enum machine_mode mode)
3887 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == MULT)
3888 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3889 force_operand (XEXP (x, 0), NULL_RTX));
3890 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == MULT)
3891 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3892 force_operand (XEXP (x, 1), NULL_RTX));
3893 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == PLUS)
3894 x = gen_rtx_PLUS (Pmode, force_operand (XEXP (x, 0), NULL_RTX),
3896 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == PLUS)
3897 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3898 force_operand (XEXP (x, 1), NULL_RTX));
3900 if (x != orig_x && legitimate_address_p (mode, x, FALSE))
3903 if (tls_symbolic_operand (x))
3904 x = legitimize_tls_address (x);
3906 x = legitimize_pic_address (x, mode, 0);
3907 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 1)))
3908 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3909 copy_to_mode_reg (Pmode, XEXP (x, 1)));
3910 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 0)))
3911 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3912 copy_to_mode_reg (Pmode, XEXP (x, 0)));
3913 else if (GET_CODE (x) == SYMBOL_REF
3914 || GET_CODE (x) == CONST
3915 || GET_CODE (x) == LABEL_REF)
3916 x = copy_to_suggested_reg (x, NULL_RTX, Pmode);
3920 /* Emit the special PIC prologue. */
3923 load_pic_register (void)
3925 int orig_flag_pic = flag_pic;
3927 /* If we haven't emitted the special helper function, do so now. */
3928 if (add_pc_to_pic_symbol_name[0] == 0)
3930 const char *pic_name = reg_names[REGNO (pic_offset_table_rtx)];
3933 ASM_GENERATE_INTERNAL_LABEL (add_pc_to_pic_symbol_name, "LADDPC", 0);
3936 align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
3938 ASM_OUTPUT_ALIGN (asm_out_file, align);
3939 ASM_OUTPUT_LABEL (asm_out_file, add_pc_to_pic_symbol_name);
3940 if (flag_delayed_branch)
3941 fprintf (asm_out_file, "\tjmp %%o7+8\n\t add\t%%o7, %s, %s\n",
3942 pic_name, pic_name);
3944 fprintf (asm_out_file, "\tadd\t%%o7, %s, %s\n\tjmp %%o7+8\n\t nop\n",
3945 pic_name, pic_name);
3948 /* Initialize every time through, since we can't easily
3949 know this to be permanent. */
3950 global_offset_table = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_");
3951 add_pc_to_pic_symbol = gen_rtx_SYMBOL_REF (Pmode, add_pc_to_pic_symbol_name);
3954 emit_insn (gen_load_pcrel_sym (pic_offset_table_rtx, global_offset_table,
3955 add_pc_to_pic_symbol));
3956 flag_pic = orig_flag_pic;
3958 /* Need to emit this whether or not we obey regdecls,
3959 since setjmp/longjmp can cause life info to screw up.
3960 ??? In the case where we don't obey regdecls, this is not sufficient
3961 since we may not fall out the bottom. */
3962 emit_insn (gen_rtx_USE (VOIDmode, pic_offset_table_rtx));
3965 /* Return 1 if RTX is a MEM which is known to be aligned to at
3966 least a DESIRED byte boundary. */
3969 mem_min_alignment (rtx mem, int desired)
3971 rtx addr, base, offset;
3973 /* If it's not a MEM we can't accept it. */
3974 if (GET_CODE (mem) != MEM)
3977 addr = XEXP (mem, 0);
3978 base = offset = NULL_RTX;
3979 if (GET_CODE (addr) == PLUS)
3981 if (GET_CODE (XEXP (addr, 0)) == REG)
3983 base = XEXP (addr, 0);
3985 /* What we are saying here is that if the base
3986 REG is aligned properly, the compiler will make
3987 sure any REG based index upon it will be so
3989 if (GET_CODE (XEXP (addr, 1)) == CONST_INT)
3990 offset = XEXP (addr, 1);
3992 offset = const0_rtx;
3995 else if (GET_CODE (addr) == REG)
3998 offset = const0_rtx;
4001 if (base != NULL_RTX)
4003 int regno = REGNO (base);
4005 if (regno != HARD_FRAME_POINTER_REGNUM && regno != STACK_POINTER_REGNUM)
4007 /* Check if the compiler has recorded some information
4008 about the alignment of the base REG. If reload has
4009 completed, we already matched with proper alignments.
4010 If not running global_alloc, reload might give us
4011 unaligned pointer to local stack though. */
4013 && REGNO_POINTER_ALIGN (regno) >= desired * BITS_PER_UNIT)
4014 || (optimize && reload_completed))
4015 && (INTVAL (offset) & (desired - 1)) == 0)
4020 if (((INTVAL (offset) - SPARC_STACK_BIAS) & (desired - 1)) == 0)
4024 else if (! TARGET_UNALIGNED_DOUBLES
4025 || CONSTANT_P (addr)
4026 || GET_CODE (addr) == LO_SUM)
4028 /* Anything else we know is properly aligned unless TARGET_UNALIGNED_DOUBLES
4029 is true, in which case we can only assume that an access is aligned if
4030 it is to a constant address, or the address involves a LO_SUM. */
4034 /* An obviously unaligned address. */
4039 /* Vectors to keep interesting information about registers where it can easily
4040 be got. We used to use the actual mode value as the bit number, but there
4041 are more than 32 modes now. Instead we use two tables: one indexed by
4042 hard register number, and one indexed by mode. */
4044 /* The purpose of sparc_mode_class is to shrink the range of modes so that
4045 they all fit (as bit numbers) in a 32 bit word (again). Each real mode is
4046 mapped into one sparc_mode_class mode. */
4048 enum sparc_mode_class {
4049 S_MODE, D_MODE, T_MODE, O_MODE,
4050 SF_MODE, DF_MODE, TF_MODE, OF_MODE,
4054 /* Modes for single-word and smaller quantities. */
4055 #define S_MODES ((1 << (int) S_MODE) | (1 << (int) SF_MODE))
4057 /* Modes for double-word and smaller quantities. */
4058 #define D_MODES (S_MODES | (1 << (int) D_MODE) | (1 << DF_MODE))
4060 /* Modes for quad-word and smaller quantities. */
4061 #define T_MODES (D_MODES | (1 << (int) T_MODE) | (1 << (int) TF_MODE))
4063 /* Modes for 8-word and smaller quantities. */
4064 #define O_MODES (T_MODES | (1 << (int) O_MODE) | (1 << (int) OF_MODE))
4066 /* Modes for single-float quantities. We must allow any single word or
4067 smaller quantity. This is because the fix/float conversion instructions
4068 take integer inputs/outputs from the float registers. */
4069 #define SF_MODES (S_MODES)
4071 /* Modes for double-float and smaller quantities. */
4072 #define DF_MODES (S_MODES | D_MODES)
4074 /* Modes for double-float only quantities. */
4075 #define DF_MODES_NO_S ((1 << (int) D_MODE) | (1 << (int) DF_MODE))
4077 /* Modes for quad-float only quantities. */
4078 #define TF_ONLY_MODES (1 << (int) TF_MODE)
4080 /* Modes for quad-float and smaller quantities. */
4081 #define TF_MODES (DF_MODES | TF_ONLY_MODES)
4083 /* Modes for quad-float and double-float quantities. */
4084 #define TF_MODES_NO_S (DF_MODES_NO_S | TF_ONLY_MODES)
4086 /* Modes for quad-float pair only quantities. */
4087 #define OF_ONLY_MODES (1 << (int) OF_MODE)
4089 /* Modes for quad-float pairs and smaller quantities. */
4090 #define OF_MODES (TF_MODES | OF_ONLY_MODES)
4092 #define OF_MODES_NO_S (TF_MODES_NO_S | OF_ONLY_MODES)
4094 /* Modes for condition codes. */
4095 #define CC_MODES (1 << (int) CC_MODE)
4096 #define CCFP_MODES (1 << (int) CCFP_MODE)
4098 /* Value is 1 if register/mode pair is acceptable on sparc.
4099 The funny mixture of D and T modes is because integer operations
4100 do not specially operate on tetra quantities, so non-quad-aligned
4101 registers can hold quadword quantities (except %o4 and %i4 because
4102 they cross fixed registers). */
4104 /* This points to either the 32 bit or the 64 bit version. */
4105 const int *hard_regno_mode_classes;
4107 static const int hard_32bit_mode_classes[] = {
4108 S_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
4109 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
4110 T_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
4111 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
4113 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4114 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4115 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4116 OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
4118 /* FP regs f32 to f63. Only the even numbered registers actually exist,
4119 and none can hold SFmode/SImode values. */
4120 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4121 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4122 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4123 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4126 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
4132 static const int hard_64bit_mode_classes[] = {
4133 D_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4134 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4135 T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4136 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4138 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4139 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4140 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4141 OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
4143 /* FP regs f32 to f63. Only the even numbered registers actually exist,
4144 and none can hold SFmode/SImode values. */
4145 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4146 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4147 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4148 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4151 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
4157 int sparc_mode_class [NUM_MACHINE_MODES];
4159 enum reg_class sparc_regno_reg_class[FIRST_PSEUDO_REGISTER];
4162 sparc_init_modes (void)
4166 for (i = 0; i < NUM_MACHINE_MODES; i++)
4168 switch (GET_MODE_CLASS (i))
4171 case MODE_PARTIAL_INT:
4172 case MODE_COMPLEX_INT:
4173 if (GET_MODE_SIZE (i) <= 4)
4174 sparc_mode_class[i] = 1 << (int) S_MODE;
4175 else if (GET_MODE_SIZE (i) == 8)
4176 sparc_mode_class[i] = 1 << (int) D_MODE;
4177 else if (GET_MODE_SIZE (i) == 16)
4178 sparc_mode_class[i] = 1 << (int) T_MODE;
4179 else if (GET_MODE_SIZE (i) == 32)
4180 sparc_mode_class[i] = 1 << (int) O_MODE;
4182 sparc_mode_class[i] = 0;
4185 case MODE_COMPLEX_FLOAT:
4186 if (GET_MODE_SIZE (i) <= 4)
4187 sparc_mode_class[i] = 1 << (int) SF_MODE;
4188 else if (GET_MODE_SIZE (i) == 8)
4189 sparc_mode_class[i] = 1 << (int) DF_MODE;
4190 else if (GET_MODE_SIZE (i) == 16)
4191 sparc_mode_class[i] = 1 << (int) TF_MODE;
4192 else if (GET_MODE_SIZE (i) == 32)
4193 sparc_mode_class[i] = 1 << (int) OF_MODE;
4195 sparc_mode_class[i] = 0;
4198 if (i == (int) CCFPmode || i == (int) CCFPEmode)
4199 sparc_mode_class[i] = 1 << (int) CCFP_MODE;
4201 sparc_mode_class[i] = 1 << (int) CC_MODE;
4204 sparc_mode_class[i] = 0;
4210 hard_regno_mode_classes = hard_64bit_mode_classes;
4212 hard_regno_mode_classes = hard_32bit_mode_classes;
4214 /* Initialize the array used by REGNO_REG_CLASS. */
4215 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
4217 if (i < 16 && TARGET_V8PLUS)
4218 sparc_regno_reg_class[i] = I64_REGS;
4219 else if (i < 32 || i == FRAME_POINTER_REGNUM)
4220 sparc_regno_reg_class[i] = GENERAL_REGS;
4222 sparc_regno_reg_class[i] = FP_REGS;
4224 sparc_regno_reg_class[i] = EXTRA_FP_REGS;
4226 sparc_regno_reg_class[i] = FPCC_REGS;
4228 sparc_regno_reg_class[i] = NO_REGS;
4232 /* Compute the frame size required by the function. This function is called
4233 during the reload pass and also by sparc_expand_prologue. */
4236 sparc_compute_frame_size (HOST_WIDE_INT size, int leaf_function_p)
4238 int outgoing_args_size = (current_function_outgoing_args_size
4239 + REG_PARM_STACK_SPACE (current_function_decl));
4240 int n_regs = 0; /* N_REGS is the number of 4-byte regs saved thus far. */
4245 for (i = 0; i < 8; i++)
4246 if (regs_ever_live[i] && ! call_used_regs[i])
4251 for (i = 0; i < 8; i += 2)
4252 if ((regs_ever_live[i] && ! call_used_regs[i])
4253 || (regs_ever_live[i+1] && ! call_used_regs[i+1]))
4257 for (i = 32; i < (TARGET_V9 ? 96 : 64); i += 2)
4258 if ((regs_ever_live[i] && ! call_used_regs[i])
4259 || (regs_ever_live[i+1] && ! call_used_regs[i+1]))
4262 /* Set up values for use in prologue and epilogue. */
4263 num_gfregs = n_regs;
4268 && current_function_outgoing_args_size == 0)
4269 actual_fsize = apparent_fsize = 0;
4272 /* We subtract STARTING_FRAME_OFFSET, remember it's negative. */
4273 apparent_fsize = (size - STARTING_FRAME_OFFSET + 7) & -8;
4274 apparent_fsize += n_regs * 4;
4275 actual_fsize = apparent_fsize + ((outgoing_args_size + 7) & -8);
4278 /* Make sure nothing can clobber our register windows.
4279 If a SAVE must be done, or there is a stack-local variable,
4280 the register window area must be allocated.
4281 ??? For v8 we apparently need an additional 8 bytes of reserved space. */
4282 if (! leaf_function_p || size > 0)
4283 actual_fsize += (16 * UNITS_PER_WORD) + (TARGET_ARCH64 ? 0 : 8);
4285 return SPARC_STACK_ALIGN (actual_fsize);
4288 /* Output any necessary .register pseudo-ops. */
4291 sparc_output_scratch_registers (FILE *file ATTRIBUTE_UNUSED)
4293 #ifdef HAVE_AS_REGISTER_PSEUDO_OP
4299 /* Check if %g[2367] were used without
4300 .register being printed for them already. */
4301 for (i = 2; i < 8; i++)
4303 if (regs_ever_live [i]
4304 && ! sparc_hard_reg_printed [i])
4306 sparc_hard_reg_printed [i] = 1;
4307 fprintf (file, "\t.register\t%%g%d, #scratch\n", i);
4314 /* Save/restore call-saved registers from LOW to HIGH at BASE+OFFSET
4315 as needed. LOW should be double-word aligned for 32-bit registers.
4316 Return the new OFFSET. */
4319 #define SORR_RESTORE 1
4322 save_or_restore_regs (int low, int high, rtx base, int offset, int action)
4327 if (TARGET_ARCH64 && high <= 32)
4329 for (i = low; i < high; i++)
4331 if (regs_ever_live[i] && ! call_used_regs[i])
4333 mem = gen_rtx_MEM (DImode, plus_constant (base, offset));
4334 set_mem_alias_set (mem, sparc_sr_alias_set);
4335 if (action == SORR_SAVE)
4337 insn = emit_move_insn (mem, gen_rtx_REG (DImode, i));
4338 RTX_FRAME_RELATED_P (insn) = 1;
4340 else /* action == SORR_RESTORE */
4341 emit_move_insn (gen_rtx_REG (DImode, i), mem);
4348 for (i = low; i < high; i += 2)
4350 bool reg0 = regs_ever_live[i] && ! call_used_regs[i];
4351 bool reg1 = regs_ever_live[i+1] && ! call_used_regs[i+1];
4352 enum machine_mode mode;
4357 mode = i < 32 ? DImode : DFmode;
4362 mode = i < 32 ? SImode : SFmode;
4367 mode = i < 32 ? SImode : SFmode;
4374 mem = gen_rtx_MEM (mode, plus_constant (base, offset));
4375 set_mem_alias_set (mem, sparc_sr_alias_set);
4376 if (action == SORR_SAVE)
4378 insn = emit_move_insn (mem, gen_rtx_REG (mode, regno));
4379 RTX_FRAME_RELATED_P (insn) = 1;
4381 else /* action == SORR_RESTORE */
4382 emit_move_insn (gen_rtx_REG (mode, regno), mem);
4384 /* Always preserve double-word alignment. */
4385 offset = (offset + 7) & -8;
4392 /* Emit code to save call-saved registers. */
4395 emit_save_regs (void)
4397 HOST_WIDE_INT offset;
4400 offset = frame_base_offset - apparent_fsize;
4402 if (offset < -4096 || offset + num_gfregs * 4 > 4096)
4404 /* ??? This might be optimized a little as %g1 might already have a
4405 value close enough that a single add insn will do. */
4406 /* ??? Although, all of this is probably only a temporary fix
4407 because if %g1 can hold a function result, then
4408 sparc_expand_epilogue will lose (the result will be
4410 base = gen_rtx_REG (Pmode, 1);
4411 emit_move_insn (base, GEN_INT (offset));
4412 emit_insn (gen_rtx_SET (VOIDmode,
4414 gen_rtx_PLUS (Pmode, frame_base_reg, base)));
4418 base = frame_base_reg;
4420 offset = save_or_restore_regs (0, 8, base, offset, SORR_SAVE);
4421 save_or_restore_regs (32, TARGET_V9 ? 96 : 64, base, offset, SORR_SAVE);
4424 /* Emit code to restore call-saved registers. */
4427 emit_restore_regs (void)
4429 HOST_WIDE_INT offset;
4432 offset = frame_base_offset - apparent_fsize;
4434 if (offset < -4096 || offset + num_gfregs * 4 > 4096 - 8 /*double*/)
4436 base = gen_rtx_REG (Pmode, 1);
4437 emit_move_insn (base, GEN_INT (offset));
4438 emit_insn (gen_rtx_SET (VOIDmode,
4440 gen_rtx_PLUS (Pmode, frame_base_reg, base)));
4444 base = frame_base_reg;
4446 offset = save_or_restore_regs (0, 8, base, offset, SORR_RESTORE);
4447 save_or_restore_regs (32, TARGET_V9 ? 96 : 64, base, offset, SORR_RESTORE);
4450 /* Emit an increment for the stack pointer. */
4453 emit_stack_pointer_increment (rtx increment)
4456 emit_insn (gen_adddi3 (stack_pointer_rtx, stack_pointer_rtx, increment));
4458 emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, increment));
4461 /* Emit a decrement for the stack pointer. */
4464 emit_stack_pointer_decrement (rtx decrement)
4467 emit_insn (gen_subdi3 (stack_pointer_rtx, stack_pointer_rtx, decrement));
4469 emit_insn (gen_subsi3 (stack_pointer_rtx, stack_pointer_rtx, decrement));
4472 /* Expand the function prologue. The prologue is responsible for reserving
4473 storage for the frame, saving the call-saved registers and loading the
4474 PIC register if needed. */
4477 sparc_expand_prologue (void)
4479 /* Compute a snapshot of current_function_uses_only_leaf_regs. Relying
4480 on the final value of the flag means deferring the prologue/epilogue
4481 expansion until just before the second scheduling pass, which is too
4482 late to emit multiple epilogues or return insns.
4484 Of course we are making the assumption that the value of the flag
4485 will not change between now and its final value. Of the three parts
4486 of the formula, only the last one can reasonably vary. Let's take a
4487 closer look, after assuming that the first two ones are set to true
4488 (otherwise the last value is effectively silenced).
4490 If only_leaf_regs_used returns false, the global predicate will also
4491 be false so the actual frame size calculated below will be positive.
4492 As a consequence, the save_register_window insn will be emitted in
4493 the instruction stream; now this insn explicitly references %fp
4494 which is not a leaf register so only_leaf_regs_used will always
4495 return false subsequently.
4497 If only_leaf_regs_used returns true, we hope that the subsequent
4498 optimization passes won't cause non-leaf registers to pop up. For
4499 example, the regrename pass has special provisions to not rename to
4500 non-leaf registers in a leaf function. */
4501 sparc_leaf_function_p
4502 = optimize > 0 && leaf_function_p () && only_leaf_regs_used ();
4504 /* Need to use actual_fsize, since we are also allocating
4505 space for our callee (and our own register save area). */
4507 = sparc_compute_frame_size (get_frame_size(), sparc_leaf_function_p);
4509 /* Advertise that the data calculated just above are now valid. */
4510 sparc_prologue_data_valid_p = true;
4512 if (sparc_leaf_function_p)
4514 frame_base_reg = stack_pointer_rtx;
4515 frame_base_offset = actual_fsize + SPARC_STACK_BIAS;
4519 frame_base_reg = hard_frame_pointer_rtx;
4520 frame_base_offset = SPARC_STACK_BIAS;
4523 if (actual_fsize == 0)
4525 else if (sparc_leaf_function_p)
4527 if (actual_fsize <= 4096)
4528 emit_stack_pointer_increment (GEN_INT (- actual_fsize));
4529 else if (actual_fsize <= 8192)
4531 emit_stack_pointer_increment (GEN_INT (-4096));
4532 emit_stack_pointer_increment (GEN_INT (4096 - actual_fsize));
4536 rtx reg = gen_rtx_REG (Pmode, 1);
4537 emit_move_insn (reg, GEN_INT (-actual_fsize));
4538 emit_stack_pointer_increment (reg);
4543 if (actual_fsize <= 4096)
4544 emit_insn (gen_save_register_window (GEN_INT (-actual_fsize)));
4545 else if (actual_fsize <= 8192)
4547 emit_insn (gen_save_register_window (GEN_INT (-4096)));
4548 emit_stack_pointer_increment (GEN_INT (4096 - actual_fsize));
4552 rtx reg = gen_rtx_REG (Pmode, 1);
4553 emit_move_insn (reg, GEN_INT (-actual_fsize));
4554 emit_insn (gen_save_register_window (reg));
4558 /* Call-saved registers are saved just above the outgoing argument area. */
4562 /* Load the PIC register if needed. */
4563 if (flag_pic && current_function_uses_pic_offset_table)
4564 load_pic_register ();
4567 /* This function generates the assembly code for function entry, which boils
4568 down to emitting the necessary .register directives. It also informs the
4569 DWARF-2 back-end on the layout of the frame.
4571 ??? Historical cruft: "On SPARC, move-double insns between fpu and cpu need
4572 an 8-byte block of memory. If any fpu reg is used in the function, we
4573 allocate such a block here, at the bottom of the frame, just in case it's
4574 needed." Could this explain the -8 in emit_restore_regs? */
4577 sparc_asm_function_prologue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
4579 /* Check that the assumption we made in sparc_expand_prologue is valid. */
4580 if (sparc_leaf_function_p != current_function_uses_only_leaf_regs)
4583 sparc_output_scratch_registers (file);
4585 if (dwarf2out_do_frame () && actual_fsize)
4587 char *label = dwarf2out_cfi_label ();
4589 /* The canonical frame address refers to the top of the frame. */
4590 dwarf2out_def_cfa (label,
4591 sparc_leaf_function_p
4592 ? STACK_POINTER_REGNUM
4593 : HARD_FRAME_POINTER_REGNUM,
4596 if (! sparc_leaf_function_p)
4598 /* Note the register window save. This tells the unwinder that
4599 it needs to restore the window registers from the previous
4600 frame's window save area at 0(cfa). */
4601 dwarf2out_window_save (label);
4603 /* The return address (-8) is now in %i7. */
4604 dwarf2out_return_reg (label, 31);
4609 /* Expand the function epilogue, either normal or part of a sibcall.
4610 We emit all the instructions except the return or the call. */
4613 sparc_expand_epilogue (void)
4616 emit_restore_regs ();
4618 if (actual_fsize == 0)
4620 else if (sparc_leaf_function_p)
4622 if (actual_fsize <= 4096)
4623 emit_stack_pointer_decrement (GEN_INT (- actual_fsize));
4624 else if (actual_fsize <= 8192)
4626 emit_stack_pointer_decrement (GEN_INT (-4096));
4627 emit_stack_pointer_decrement (GEN_INT (4096 - actual_fsize));
4631 rtx reg = gen_rtx_REG (Pmode, 1);
4632 emit_move_insn (reg, GEN_INT (-actual_fsize));
4633 emit_stack_pointer_decrement (reg);
4638 /* Return true if it is appropriate to emit `return' instructions in the
4639 body of a function. */
4642 sparc_can_use_return_insn_p (void)
4644 return sparc_prologue_data_valid_p
4645 && (actual_fsize == 0 || !sparc_leaf_function_p);
4648 /* This function generates the assembly code for function exit. */
4651 sparc_asm_function_epilogue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
4653 /* If code does not drop into the epilogue, we have to still output
4654 a dummy nop for the sake of sane backtraces. Otherwise, if the
4655 last two instructions of a function were "call foo; dslot;" this
4656 can make the return PC of foo (i.e. address of call instruction
4657 plus 8) point to the first instruction in the next function. */
4659 rtx insn, last_real_insn;
4661 insn = get_last_insn ();
4663 last_real_insn = prev_real_insn (insn);
4665 && GET_CODE (last_real_insn) == INSN
4666 && GET_CODE (PATTERN (last_real_insn)) == SEQUENCE)
4667 last_real_insn = XVECEXP (PATTERN (last_real_insn), 0, 0);
4669 if (last_real_insn && GET_CODE (last_real_insn) == CALL_INSN)
4670 fputs("\tnop\n", file);
4672 sparc_output_deferred_case_vectors ();
4675 /* Output a 'restore' instruction. */
4678 output_restore (rtx pat)
4684 fputs ("\t restore\n", asm_out_file);
4688 if (GET_CODE (pat) != SET)
4691 operands[0] = SET_DEST (pat);
4692 pat = SET_SRC (pat);
4694 switch (GET_CODE (pat))
4697 operands[1] = XEXP (pat, 0);
4698 operands[2] = XEXP (pat, 1);
4699 output_asm_insn (" restore %r1, %2, %Y0", operands);
4702 operands[1] = XEXP (pat, 0);
4703 operands[2] = XEXP (pat, 1);
4704 output_asm_insn (" restore %r1, %%lo(%a2), %Y0", operands);
4707 operands[1] = XEXP (pat, 0);
4708 if (XEXP (pat, 1) != const1_rtx)
4710 output_asm_insn (" restore %r1, %r1, %Y0", operands);
4714 output_asm_insn (" restore %%g0, %1, %Y0", operands);
4719 /* Output a return. */
4722 output_return (rtx insn)
4724 if (sparc_leaf_function_p)
4726 /* This is a leaf function so we don't have to bother restoring the
4727 register window, which frees us from dealing with the convoluted
4728 semantics of restore/return. We simply output the jump to the
4729 return address and the insn in the delay slot (if any). */
4731 if (current_function_calls_eh_return)
4734 return "jmp\t%%o7+%)%#";
4738 /* This is a regular function so we have to restore the register window.
4739 We may have a pending insn for the delay slot, which will be either
4740 combined with the 'restore' instruction or put in the delay slot of
4741 the 'return' instruction. */
4743 if (current_function_calls_eh_return)
4745 /* If the function uses __builtin_eh_return, the eh_return
4746 machinery occupies the delay slot. */
4750 if (! flag_delayed_branch)
4751 fputs ("\tadd\t%fp, %g1, %fp\n", asm_out_file);
4754 fputs ("\treturn\t%i7+8\n", asm_out_file);
4756 fputs ("\trestore\n\tjmp\t%o7+8\n", asm_out_file);
4758 if (flag_delayed_branch)
4759 fputs ("\t add\t%sp, %g1, %sp\n", asm_out_file);
4761 fputs ("\t nop\n", asm_out_file);
4763 else if (final_sequence)
4767 delay = NEXT_INSN (insn);
4771 pat = PATTERN (delay);
4773 if (TARGET_V9 && ! epilogue_renumber (&pat, 1))
4775 epilogue_renumber (&pat, 0);
4776 return "return\t%%i7+%)%#";
4780 output_asm_insn ("jmp\t%%i7+%)", NULL);
4781 output_restore (pat);
4782 PATTERN (delay) = gen_blockage ();
4783 INSN_CODE (delay) = -1;
4788 /* The delay slot is empty. */
4790 return "return\t%%i7+%)\n\t nop";
4791 else if (flag_delayed_branch)
4792 return "jmp\t%%i7+%)\n\t restore";
4794 return "restore\n\tjmp\t%%o7+%)\n\t nop";
4801 /* Output a sibling call. */
4804 output_sibcall (rtx insn, rtx call_operand)
4808 if (! flag_delayed_branch)
4811 operands[0] = call_operand;
4813 if (sparc_leaf_function_p)
4815 /* This is a leaf function so we don't have to bother restoring the
4816 register window. We simply output the jump to the function and
4817 the insn in the delay slot (if any). */
4819 if (LEAF_SIBCALL_SLOT_RESERVED_P && final_sequence)
4823 output_asm_insn ("sethi\t%%hi(%a0), %%g1\n\tjmp\t%%g1 + %%lo(%a0)%#",
4826 /* Use or with rs2 %%g0 instead of mov, so that as/ld can optimize
4827 it into branch if possible. */
4828 output_asm_insn ("or\t%%o7, %%g0, %%g1\n\tcall\t%a0, 0\n\t or\t%%g1, %%g0, %%o7",
4833 /* This is a regular function so we have to restore the register window.
4834 We may have a pending insn for the delay slot, which will be combined
4835 with the 'restore' instruction. */
4837 output_asm_insn ("call\t%a0, 0", operands);
4841 rtx delay = NEXT_INSN (insn);
4845 output_restore (PATTERN (delay));
4847 PATTERN (delay) = gen_blockage ();
4848 INSN_CODE (delay) = -1;
4851 output_restore (NULL_RTX);
4857 /* Functions for handling argument passing.
4859 For 32-bit, the first 6 args are normally in registers and the rest are
4860 pushed. Any arg that starts within the first 6 words is at least
4861 partially passed in a register unless its data type forbids.
4863 For 64-bit, the argument registers are laid out as an array of 16 elements
4864 and arguments are added sequentially. The first 6 int args and up to the
4865 first 16 fp args (depending on size) are passed in regs.
4867 Slot Stack Integral Float Float in structure Double Long Double
4868 ---- ----- -------- ----- ------------------ ------ -----------
4869 15 [SP+248] %f31 %f30,%f31 %d30
4870 14 [SP+240] %f29 %f28,%f29 %d28 %q28
4871 13 [SP+232] %f27 %f26,%f27 %d26
4872 12 [SP+224] %f25 %f24,%f25 %d24 %q24
4873 11 [SP+216] %f23 %f22,%f23 %d22
4874 10 [SP+208] %f21 %f20,%f21 %d20 %q20
4875 9 [SP+200] %f19 %f18,%f19 %d18
4876 8 [SP+192] %f17 %f16,%f17 %d16 %q16
4877 7 [SP+184] %f15 %f14,%f15 %d14
4878 6 [SP+176] %f13 %f12,%f13 %d12 %q12
4879 5 [SP+168] %o5 %f11 %f10,%f11 %d10
4880 4 [SP+160] %o4 %f9 %f8,%f9 %d8 %q8
4881 3 [SP+152] %o3 %f7 %f6,%f7 %d6
4882 2 [SP+144] %o2 %f5 %f4,%f5 %d4 %q4
4883 1 [SP+136] %o1 %f3 %f2,%f3 %d2
4884 0 [SP+128] %o0 %f1 %f0,%f1 %d0 %q0
4886 Here SP = %sp if -mno-stack-bias or %sp+stack_bias otherwise.
4888 Integral arguments are always passed as 64-bit quantities appropriately
4891 Passing of floating point values is handled as follows.
4892 If a prototype is in scope:
4893 If the value is in a named argument (i.e. not a stdarg function or a
4894 value not part of the `...') then the value is passed in the appropriate
4896 If the value is part of the `...' and is passed in one of the first 6
4897 slots then the value is passed in the appropriate int reg.
4898 If the value is part of the `...' and is not passed in one of the first 6
4899 slots then the value is passed in memory.
4900 If a prototype is not in scope:
4901 If the value is one of the first 6 arguments the value is passed in the
4902 appropriate integer reg and the appropriate fp reg.
4903 If the value is not one of the first 6 arguments the value is passed in
4904 the appropriate fp reg and in memory.
4907 Summary of the calling conventions implemented by GCC on SPARC:
4910 size argument return value
4912 small integer <4 int. reg. int. reg.
4913 word 4 int. reg. int. reg.
4914 double word 8 int. reg. int. reg.
4916 _Complex small integer <8 int. reg. int. reg.
4917 _Complex word 8 int. reg. int. reg.
4918 _Complex double word 16 memory int. reg.
4920 vector integer <=8 int. reg. FP reg.
4921 vector integer >8 memory memory
4923 float 4 int. reg. FP reg.
4924 double 8 int. reg. FP reg.
4925 long double 16 memory memory
4927 _Complex float 8 memory FP reg.
4928 _Complex double 16 memory FP reg.
4929 _Complex long double 32 memory FP reg.
4931 vector float any memory memory
4933 aggregate any memory memory
4938 size argument return value
4940 small integer <8 int. reg. int. reg.
4941 word 8 int. reg. int. reg.
4942 double word 16 int. reg. int. reg.
4944 _Complex small integer <16 int. reg. int. reg.
4945 _Complex word 16 int. reg. int. reg.
4946 _Complex double word 32 memory int. reg.
4948 vector integer <=16 FP reg. FP reg.
4949 vector integer 16<s<=32 memory FP reg.
4950 vector integer >32 memory memory
4952 float 4 FP reg. FP reg.
4953 double 8 FP reg. FP reg.
4954 long double 16 FP reg. FP reg.
4956 _Complex float 8 FP reg. FP reg.
4957 _Complex double 16 FP reg. FP reg.
4958 _Complex long double 32 memory FP reg.
4960 vector float <=16 FP reg. FP reg.
4961 vector float 16<s<=32 memory FP reg.
4962 vector float >32 memory memory
4964 aggregate <=16 reg. reg.
4965 aggregate 16<s<=32 memory reg.
4966 aggregate >32 memory memory
4970 Note #1: complex floating-point types follow the extended SPARC ABIs as
4971 implemented by the Sun compiler.
4973 Note #2: integral vector types follow the scalar floating-point types
4974 conventions to match what is implemented by the Sun VIS SDK.
4976 Note #3: floating-point vector types follow the aggregate types
4980 /* Maximum number of int regs for args. */
4981 #define SPARC_INT_ARG_MAX 6
4982 /* Maximum number of fp regs for args. */
4983 #define SPARC_FP_ARG_MAX 16
4985 #define ROUND_ADVANCE(SIZE) (((SIZE) + UNITS_PER_WORD - 1) / UNITS_PER_WORD)
4987 /* Handle the INIT_CUMULATIVE_ARGS macro.
4988 Initialize a variable CUM of type CUMULATIVE_ARGS
4989 for a call to a function whose data type is FNTYPE.
4990 For a library call, FNTYPE is 0. */
4993 init_cumulative_args (struct sparc_args *cum, tree fntype,
4994 rtx libname ATTRIBUTE_UNUSED,
4995 tree fndecl ATTRIBUTE_UNUSED)
4998 cum->prototype_p = fntype && TYPE_ARG_TYPES (fntype);
4999 cum->libcall_p = fntype == 0;
5002 /* Handle the TARGET_PROMOTE_PROTOTYPES target hook.
5003 When a prototype says `char' or `short', really pass an `int'. */
5006 sparc_promote_prototypes (tree fntype ATTRIBUTE_UNUSED)
5008 return TARGET_ARCH32 ? true : false;
5011 /* Handle the TARGET_STRICT_ARGUMENT_NAMING target hook. */
5014 sparc_strict_argument_naming (CUMULATIVE_ARGS *ca ATTRIBUTE_UNUSED)
5016 return TARGET_ARCH64 ? true : false;
5019 /* Scan the record type TYPE and return the following predicates:
5020 - INTREGS_P: the record contains at least one field or sub-field
5021 that is eligible for promotion in integer registers.
5022 - FP_REGS_P: the record contains at least one field or sub-field
5023 that is eligible for promotion in floating-point registers.
5024 - PACKED_P: the record contains at least one field that is packed.
5026 Sub-fields are not taken into account for the PACKED_P predicate. */
5029 scan_record_type (tree type, int *intregs_p, int *fpregs_p, int *packed_p)
5033 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5035 if (TREE_CODE (field) == FIELD_DECL)
5037 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
5038 scan_record_type (TREE_TYPE (field), intregs_p, fpregs_p, 0);
5039 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
5040 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
5046 if (packed_p && DECL_PACKED (field))
5052 /* Compute the slot number to pass an argument in.
5053 Return the slot number or -1 if passing on the stack.
5055 CUM is a variable of type CUMULATIVE_ARGS which gives info about
5056 the preceding args and about the function being called.
5057 MODE is the argument's machine mode.
5058 TYPE is the data type of the argument (as a tree).
5059 This is null for libcalls where that information may
5061 NAMED is nonzero if this argument is a named parameter
5062 (otherwise it is an extra parameter matching an ellipsis).
5063 INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG.
5064 *PREGNO records the register number to use if scalar type.
5065 *PPADDING records the amount of padding needed in words. */
5068 function_arg_slotno (const struct sparc_args *cum, enum machine_mode mode,
5069 tree type, int named, int incoming_p,
5070 int *pregno, int *ppadding)
5072 int regbase = (incoming_p
5073 ? SPARC_INCOMING_INT_ARG_FIRST
5074 : SPARC_OUTGOING_INT_ARG_FIRST);
5075 int slotno = cum->words;
5076 enum mode_class mclass;
5081 if (type && TREE_ADDRESSABLE (type))
5087 && TYPE_ALIGN (type) % PARM_BOUNDARY != 0)
5090 /* For SPARC64, objects requiring 16-byte alignment get it. */
5092 && GET_MODE_ALIGNMENT (mode) >= 2 * BITS_PER_WORD
5093 && (slotno & 1) != 0)
5094 slotno++, *ppadding = 1;
5096 mclass = GET_MODE_CLASS (mode);
5097 if (type && TREE_CODE (type) == VECTOR_TYPE)
5099 /* Vector types deserve special treatment because they are
5100 polymorphic wrt their mode, depending upon whether VIS
5101 instructions are enabled. */
5102 if (TREE_CODE (TREE_TYPE (type)) == REAL_TYPE)
5104 /* The SPARC port defines no floating-point vector modes. */
5105 if (mode != BLKmode)
5110 /* Integral vector types should either have a vector
5111 mode or an integral mode, because we are guaranteed
5112 by pass_by_reference that their size is not greater
5113 than 16 bytes and TImode is 16-byte wide. */
5114 if (mode == BLKmode)
5117 /* Vector integers are handled like floats according to
5119 mclass = MODE_FLOAT;
5126 case MODE_COMPLEX_FLOAT:
5127 if (TARGET_ARCH64 && TARGET_FPU && named)
5129 if (slotno >= SPARC_FP_ARG_MAX)
5131 regno = SPARC_FP_ARG_FIRST + slotno * 2;
5132 /* Arguments filling only one single FP register are
5133 right-justified in the outer double FP register. */
5134 if (GET_MODE_SIZE (mode) <= 4)
5141 case MODE_COMPLEX_INT:
5142 if (slotno >= SPARC_INT_ARG_MAX)
5144 regno = regbase + slotno;
5148 if (mode == VOIDmode)
5149 /* MODE is VOIDmode when generating the actual call. */
5152 if (mode != BLKmode)
5155 /* For SPARC64, objects requiring 16-byte alignment get it. */
5158 && TYPE_ALIGN (type) >= 2 * BITS_PER_WORD
5159 && (slotno & 1) != 0)
5160 slotno++, *ppadding = 1;
5162 if (TARGET_ARCH32 || !type || (TREE_CODE (type) == UNION_TYPE))
5164 if (slotno >= SPARC_INT_ARG_MAX)
5166 regno = regbase + slotno;
5168 else /* TARGET_ARCH64 && type */
5170 int intregs_p = 0, fpregs_p = 0, packed_p = 0;
5172 /* First see what kinds of registers we would need. */
5173 if (TREE_CODE (type) == VECTOR_TYPE)
5176 scan_record_type (type, &intregs_p, &fpregs_p, &packed_p);
5178 /* The ABI obviously doesn't specify how packed structures
5179 are passed. These are defined to be passed in int regs
5180 if possible, otherwise memory. */
5181 if (packed_p || !named)
5182 fpregs_p = 0, intregs_p = 1;
5184 /* If all arg slots are filled, then must pass on stack. */
5185 if (fpregs_p && slotno >= SPARC_FP_ARG_MAX)
5188 /* If there are only int args and all int arg slots are filled,
5189 then must pass on stack. */
5190 if (!fpregs_p && intregs_p && slotno >= SPARC_INT_ARG_MAX)
5193 /* Note that even if all int arg slots are filled, fp members may
5194 still be passed in regs if such regs are available.
5195 *PREGNO isn't set because there may be more than one, it's up
5196 to the caller to compute them. */
5209 /* Handle recursive register counting for structure field layout. */
5211 struct function_arg_record_value_parms
5213 rtx ret; /* return expression being built. */
5214 int slotno; /* slot number of the argument. */
5215 int named; /* whether the argument is named. */
5216 int regbase; /* regno of the base register. */
5217 int stack; /* 1 if part of the argument is on the stack. */
5218 int intoffset; /* offset of the first pending integer field. */
5219 unsigned int nregs; /* number of words passed in registers. */
5222 static void function_arg_record_value_3
5223 (HOST_WIDE_INT, struct function_arg_record_value_parms *);
5224 static void function_arg_record_value_2
5225 (tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
5226 static void function_arg_record_value_1
5227 (tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
5228 static rtx function_arg_record_value (tree, enum machine_mode, int, int, int);
5229 static rtx function_arg_union_value (int, enum machine_mode, int);
5231 /* A subroutine of function_arg_record_value. Traverse the structure
5232 recursively and determine how many registers will be required. */
5235 function_arg_record_value_1 (tree type, HOST_WIDE_INT startbitpos,
5236 struct function_arg_record_value_parms *parms,
5241 /* We need to compute how many registers are needed so we can
5242 allocate the PARALLEL but before we can do that we need to know
5243 whether there are any packed fields. The ABI obviously doesn't
5244 specify how structures are passed in this case, so they are
5245 defined to be passed in int regs if possible, otherwise memory,
5246 regardless of whether there are fp values present. */
5249 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5251 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
5258 /* Compute how many registers we need. */
5259 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5261 if (TREE_CODE (field) == FIELD_DECL)
5263 HOST_WIDE_INT bitpos = startbitpos;
5265 if (DECL_SIZE (field) != 0
5266 && host_integerp (bit_position (field), 1))
5267 bitpos += int_bit_position (field);
5269 /* ??? FIXME: else assume zero offset. */
5271 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
5272 function_arg_record_value_1 (TREE_TYPE (field),
5276 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
5277 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
5282 if (parms->intoffset != -1)
5284 unsigned int startbit, endbit;
5285 int intslots, this_slotno;
5287 startbit = parms->intoffset & -BITS_PER_WORD;
5288 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
5290 intslots = (endbit - startbit) / BITS_PER_WORD;
5291 this_slotno = parms->slotno + parms->intoffset
5294 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
5296 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
5297 /* We need to pass this field on the stack. */
5301 parms->nregs += intslots;
5302 parms->intoffset = -1;
5305 /* There's no need to check this_slotno < SPARC_FP_ARG MAX.
5306 If it wasn't true we wouldn't be here. */
5307 if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE
5308 && DECL_MODE (field) == BLKmode)
5309 parms->nregs += TYPE_VECTOR_SUBPARTS (TREE_TYPE (field));
5310 else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
5317 if (parms->intoffset == -1)
5318 parms->intoffset = bitpos;
5324 /* A subroutine of function_arg_record_value. Assign the bits of the
5325 structure between parms->intoffset and bitpos to integer registers. */
5328 function_arg_record_value_3 (HOST_WIDE_INT bitpos,
5329 struct function_arg_record_value_parms *parms)
5331 enum machine_mode mode;
5333 unsigned int startbit, endbit;
5334 int this_slotno, intslots, intoffset;
5337 if (parms->intoffset == -1)
5340 intoffset = parms->intoffset;
5341 parms->intoffset = -1;
5343 startbit = intoffset & -BITS_PER_WORD;
5344 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
5345 intslots = (endbit - startbit) / BITS_PER_WORD;
5346 this_slotno = parms->slotno + intoffset / BITS_PER_WORD;
5348 intslots = MIN (intslots, SPARC_INT_ARG_MAX - this_slotno);
5352 /* If this is the trailing part of a word, only load that much into
5353 the register. Otherwise load the whole register. Note that in
5354 the latter case we may pick up unwanted bits. It's not a problem
5355 at the moment but may wish to revisit. */
5357 if (intoffset % BITS_PER_WORD != 0)
5358 mode = mode_for_size (BITS_PER_WORD - intoffset % BITS_PER_WORD,
5363 intoffset /= BITS_PER_UNIT;
5366 regno = parms->regbase + this_slotno;
5367 reg = gen_rtx_REG (mode, regno);
5368 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
5369 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (intoffset));
5372 intoffset = (intoffset | (UNITS_PER_WORD-1)) + 1;
5377 while (intslots > 0);
5380 /* A subroutine of function_arg_record_value. Traverse the structure
5381 recursively and assign bits to floating point registers. Track which
5382 bits in between need integer registers; invoke function_arg_record_value_3
5383 to make that happen. */
5386 function_arg_record_value_2 (tree type, HOST_WIDE_INT startbitpos,
5387 struct function_arg_record_value_parms *parms,
5393 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5395 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
5402 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5404 if (TREE_CODE (field) == FIELD_DECL)
5406 HOST_WIDE_INT bitpos = startbitpos;
5408 if (DECL_SIZE (field) != 0
5409 && host_integerp (bit_position (field), 1))
5410 bitpos += int_bit_position (field);
5412 /* ??? FIXME: else assume zero offset. */
5414 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
5415 function_arg_record_value_2 (TREE_TYPE (field),
5419 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
5420 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
5425 int this_slotno = parms->slotno + bitpos / BITS_PER_WORD;
5426 int regno, nregs, pos;
5427 enum machine_mode mode = DECL_MODE (field);
5430 function_arg_record_value_3 (bitpos, parms);
5432 if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE
5435 mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field)));
5436 nregs = TYPE_VECTOR_SUBPARTS (TREE_TYPE (field));
5438 else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
5440 mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field)));
5446 regno = SPARC_FP_ARG_FIRST + this_slotno * 2;
5447 if (GET_MODE_SIZE (mode) <= 4 && (bitpos & 32) != 0)
5449 reg = gen_rtx_REG (mode, regno);
5450 pos = bitpos / BITS_PER_UNIT;
5451 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
5452 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos));
5456 regno += GET_MODE_SIZE (mode) / 4;
5457 reg = gen_rtx_REG (mode, regno);
5458 pos += GET_MODE_SIZE (mode);
5459 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
5460 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos));
5466 if (parms->intoffset == -1)
5467 parms->intoffset = bitpos;
5473 /* Used by function_arg and function_value to implement the complex
5474 conventions of the 64-bit ABI for passing and returning structures.
5475 Return an expression valid as a return value for the two macros
5476 FUNCTION_ARG and FUNCTION_VALUE.
5478 TYPE is the data type of the argument (as a tree).
5479 This is null for libcalls where that information may
5481 MODE is the argument's machine mode.
5482 SLOTNO is the index number of the argument's slot in the parameter array.
5483 NAMED is nonzero if this argument is a named parameter
5484 (otherwise it is an extra parameter matching an ellipsis).
5485 REGBASE is the regno of the base register for the parameter array. */
5488 function_arg_record_value (tree type, enum machine_mode mode,
5489 int slotno, int named, int regbase)
5491 HOST_WIDE_INT typesize = int_size_in_bytes (type);
5492 struct function_arg_record_value_parms parms;
5495 parms.ret = NULL_RTX;
5496 parms.slotno = slotno;
5497 parms.named = named;
5498 parms.regbase = regbase;
5501 /* Compute how many registers we need. */
5503 parms.intoffset = 0;
5504 function_arg_record_value_1 (type, 0, &parms, false);
5506 /* Take into account pending integer fields. */
5507 if (parms.intoffset != -1)
5509 unsigned int startbit, endbit;
5510 int intslots, this_slotno;
5512 startbit = parms.intoffset & -BITS_PER_WORD;
5513 endbit = (typesize*BITS_PER_UNIT + BITS_PER_WORD - 1) & -BITS_PER_WORD;
5514 intslots = (endbit - startbit) / BITS_PER_WORD;
5515 this_slotno = slotno + parms.intoffset / BITS_PER_WORD;
5517 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
5519 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
5520 /* We need to pass this field on the stack. */
5524 parms.nregs += intslots;
5526 nregs = parms.nregs;
5528 /* Allocate the vector and handle some annoying special cases. */
5531 /* ??? Empty structure has no value? Duh? */
5534 /* Though there's nothing really to store, return a word register
5535 anyway so the rest of gcc doesn't go nuts. Returning a PARALLEL
5536 leads to breakage due to the fact that there are zero bytes to
5538 return gen_rtx_REG (mode, regbase);
5542 /* ??? C++ has structures with no fields, and yet a size. Give up
5543 for now and pass everything back in integer registers. */
5544 nregs = (typesize + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
5546 if (nregs + slotno > SPARC_INT_ARG_MAX)
5547 nregs = SPARC_INT_ARG_MAX - slotno;
5552 parms.ret = gen_rtx_PARALLEL (mode, rtvec_alloc (parms.stack + nregs));
5554 /* If at least one field must be passed on the stack, generate
5555 (parallel [(expr_list (nil) ...) ...]) so that all fields will
5556 also be passed on the stack. We can't do much better because the
5557 semantics of FUNCTION_ARG_PARTIAL_NREGS doesn't handle the case
5558 of structures for which the fields passed exclusively in registers
5559 are not at the beginning of the structure. */
5561 XVECEXP (parms.ret, 0, 0)
5562 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
5564 /* Fill in the entries. */
5566 parms.intoffset = 0;
5567 function_arg_record_value_2 (type, 0, &parms, false);
5568 function_arg_record_value_3 (typesize * BITS_PER_UNIT, &parms);
5570 if (parms.nregs != nregs)
5576 /* Used by function_arg and function_value to implement the conventions
5577 of the 64-bit ABI for passing and returning unions.
5578 Return an expression valid as a return value for the two macros
5579 FUNCTION_ARG and FUNCTION_VALUE.
5581 SIZE is the size in bytes of the union.
5582 MODE is the argument's machine mode.
5583 REGNO is the hard register the union will be passed in. */
5586 function_arg_union_value (int size, enum machine_mode mode, int regno)
5588 int nwords = ROUND_ADVANCE (size), i;
5591 regs = gen_rtx_PARALLEL (mode, rtvec_alloc (nwords));
5593 for (i = 0; i < nwords; i++)
5595 /* Unions are passed left-justified. */
5596 XVECEXP (regs, 0, i)
5597 = gen_rtx_EXPR_LIST (VOIDmode,
5598 gen_rtx_REG (word_mode, regno),
5599 GEN_INT (UNITS_PER_WORD * i));
5606 /* Used by function_arg and function_value to implement the conventions
5607 for passing and returning large (BLKmode) vectors.
5608 Return an expression valid as a return value for the two macros
5609 FUNCTION_ARG and FUNCTION_VALUE.
5611 SIZE is the size in bytes of the vector.
5612 BASE_MODE is the argument's base machine mode.
5613 REGNO is the FP hard register the vector will be passed in. */
5616 function_arg_vector_value (int size, enum machine_mode base_mode, int regno)
5618 unsigned short base_mode_size = GET_MODE_SIZE (base_mode);
5619 int nregs = size / base_mode_size, i;
5622 regs = gen_rtx_PARALLEL (BLKmode, rtvec_alloc (nregs));
5624 for (i = 0; i < nregs; i++)
5626 XVECEXP (regs, 0, i)
5627 = gen_rtx_EXPR_LIST (VOIDmode,
5628 gen_rtx_REG (base_mode, regno),
5629 GEN_INT (base_mode_size * i));
5630 regno += base_mode_size / 4;
5636 /* Handle the FUNCTION_ARG macro.
5637 Determine where to put an argument to a function.
5638 Value is zero to push the argument on the stack,
5639 or a hard register in which to store the argument.
5641 CUM is a variable of type CUMULATIVE_ARGS which gives info about
5642 the preceding args and about the function being called.
5643 MODE is the argument's machine mode.
5644 TYPE is the data type of the argument (as a tree).
5645 This is null for libcalls where that information may
5647 NAMED is nonzero if this argument is a named parameter
5648 (otherwise it is an extra parameter matching an ellipsis).
5649 INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG. */
5652 function_arg (const struct sparc_args *cum, enum machine_mode mode,
5653 tree type, int named, int incoming_p)
5655 int regbase = (incoming_p
5656 ? SPARC_INCOMING_INT_ARG_FIRST
5657 : SPARC_OUTGOING_INT_ARG_FIRST);
5658 int slotno, regno, padding;
5659 enum mode_class mclass = GET_MODE_CLASS (mode);
5662 slotno = function_arg_slotno (cum, mode, type, named, incoming_p,
5670 reg = gen_rtx_REG (mode, regno);
5674 if (type && TREE_CODE (type) == RECORD_TYPE)
5676 /* Structures up to 16 bytes in size are passed in arg slots on the
5677 stack and are promoted to registers where possible. */
5679 if (int_size_in_bytes (type) > 16)
5680 abort (); /* shouldn't get here */
5682 return function_arg_record_value (type, mode, slotno, named, regbase);
5684 else if (type && TREE_CODE (type) == UNION_TYPE)
5686 HOST_WIDE_INT size = int_size_in_bytes (type);
5689 abort (); /* shouldn't get here */
5691 return function_arg_union_value (size, mode, regno);
5693 else if (type && TREE_CODE (type) == VECTOR_TYPE)
5695 /* Vector types deserve special treatment because they are
5696 polymorphic wrt their mode, depending upon whether VIS
5697 instructions are enabled. */
5698 HOST_WIDE_INT size = int_size_in_bytes (type);
5701 abort (); /* shouldn't get here */
5703 if (mode == BLKmode)
5704 return function_arg_vector_value (size,
5705 TYPE_MODE (TREE_TYPE (type)),
5706 SPARC_FP_ARG_FIRST + 2*slotno);
5708 mclass = MODE_FLOAT;
5711 /* v9 fp args in reg slots beyond the int reg slots get passed in regs
5712 but also have the slot allocated for them.
5713 If no prototype is in scope fp values in register slots get passed
5714 in two places, either fp regs and int regs or fp regs and memory. */
5715 if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT)
5716 && SPARC_FP_REG_P (regno))
5718 reg = gen_rtx_REG (mode, regno);
5719 if (cum->prototype_p || cum->libcall_p)
5721 /* "* 2" because fp reg numbers are recorded in 4 byte
5724 /* ??? This will cause the value to be passed in the fp reg and
5725 in the stack. When a prototype exists we want to pass the
5726 value in the reg but reserve space on the stack. That's an
5727 optimization, and is deferred [for a bit]. */
5728 if ((regno - SPARC_FP_ARG_FIRST) >= SPARC_INT_ARG_MAX * 2)
5729 return gen_rtx_PARALLEL (mode,
5731 gen_rtx_EXPR_LIST (VOIDmode,
5732 NULL_RTX, const0_rtx),
5733 gen_rtx_EXPR_LIST (VOIDmode,
5737 /* ??? It seems that passing back a register even when past
5738 the area declared by REG_PARM_STACK_SPACE will allocate
5739 space appropriately, and will not copy the data onto the
5740 stack, exactly as we desire.
5742 This is due to locate_and_pad_parm being called in
5743 expand_call whenever reg_parm_stack_space > 0, which
5744 while beneficial to our example here, would seem to be
5745 in error from what had been intended. Ho hum... -- r~ */
5753 if ((regno - SPARC_FP_ARG_FIRST) < SPARC_INT_ARG_MAX * 2)
5757 /* On incoming, we don't need to know that the value
5758 is passed in %f0 and %i0, and it confuses other parts
5759 causing needless spillage even on the simplest cases. */
5763 intreg = (SPARC_OUTGOING_INT_ARG_FIRST
5764 + (regno - SPARC_FP_ARG_FIRST) / 2);
5766 v0 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
5767 v1 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_REG (mode, intreg),
5769 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
5773 v0 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
5774 v1 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
5775 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
5781 /* Scalar or complex int. */
5782 reg = gen_rtx_REG (mode, regno);
5788 /* Handle the FUNCTION_ARG_PARTIAL_NREGS macro.
5789 For an arg passed partly in registers and partly in memory,
5790 this is the number of registers used.
5791 For args passed entirely in registers or entirely in memory, zero.
5793 Any arg that starts in the first 6 regs but won't entirely fit in them
5794 needs partial registers on v8. On v9, structures with integer
5795 values in arg slots 5,6 will be passed in %o5 and SP+176, and complex fp
5796 values that begin in the last fp reg [where "last fp reg" varies with the
5797 mode] will be split between that reg and memory. */
5800 function_arg_partial_nregs (const struct sparc_args *cum,
5801 enum machine_mode mode, tree type, int named)
5803 int slotno, regno, padding;
5805 /* We pass 0 for incoming_p here, it doesn't matter. */
5806 slotno = function_arg_slotno (cum, mode, type, named, 0, ®no, &padding);
5813 if ((slotno + (mode == BLKmode
5814 ? ROUND_ADVANCE (int_size_in_bytes (type))
5815 : ROUND_ADVANCE (GET_MODE_SIZE (mode))))
5816 > SPARC_INT_ARG_MAX)
5817 return SPARC_INT_ARG_MAX - slotno;
5821 /* We are guaranteed by pass_by_reference that the size of the
5822 argument is not greater than 16 bytes, so we only need to
5823 return 1 if the argument is partially passed in registers. */
5825 if (type && AGGREGATE_TYPE_P (type))
5827 int size = int_size_in_bytes (type);
5829 if (size > UNITS_PER_WORD
5830 && slotno == SPARC_INT_ARG_MAX - 1)
5833 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_INT
5834 || (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT
5835 && ! (TARGET_FPU && named)))
5837 /* The complex types are passed as packed types. */
5838 if (GET_MODE_SIZE (mode) > UNITS_PER_WORD
5839 && slotno == SPARC_INT_ARG_MAX - 1)
5842 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT)
5844 if ((slotno + GET_MODE_SIZE (mode) / UNITS_PER_WORD)
5853 /* Handle the TARGET_PASS_BY_REFERENCE target hook.
5854 Specify whether to pass the argument by reference. */
5857 sparc_pass_by_reference (CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED,
5858 enum machine_mode mode, tree type,
5859 bool named ATTRIBUTE_UNUSED)
5863 /* Original SPARC 32-bit ABI says that structures and unions,
5864 and quad-precision floats are passed by reference. For Pascal,
5865 also pass arrays by reference. All other base types are passed
5868 Extended ABI (as implemented by the Sun compiler) says that all
5869 complex floats are passed by reference. Pass complex integers
5870 in registers up to 8 bytes. More generally, enforce the 2-word
5871 cap for passing arguments in registers.
5873 Vector ABI (as implemented by the Sun VIS SDK) says that vector
5874 integers are passed like floats of the same size, that is in
5875 registers up to 8 bytes. Pass all vector floats by reference
5876 like structure and unions. */
5877 return ((type && (AGGREGATE_TYPE_P (type) || VECTOR_FLOAT_TYPE_P (type)))
5879 /* Catch CDImode, TFmode, DCmode and TCmode. */
5880 || GET_MODE_SIZE (mode) > 8
5882 && TREE_CODE (type) == VECTOR_TYPE
5883 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8));
5887 /* Original SPARC 64-bit ABI says that structures and unions
5888 smaller than 16 bytes are passed in registers, as well as
5889 all other base types. For Pascal, pass arrays by reference.
5891 Extended ABI (as implemented by the Sun compiler) says that
5892 complex floats are passed in registers up to 16 bytes. Pass
5893 all complex integers in registers up to 16 bytes. More generally,
5894 enforce the 2-word cap for passing arguments in registers.
5896 Vector ABI (as implemented by the Sun VIS SDK) says that vector
5897 integers are passed like floats of the same size, that is in
5898 registers (up to 16 bytes). Pass all vector floats like structure
5900 return ((type && TREE_CODE (type) == ARRAY_TYPE)
5902 && (AGGREGATE_TYPE_P (type) || TREE_CODE (type) == VECTOR_TYPE)
5903 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 16)
5904 /* Catch CTImode and TCmode. */
5905 || GET_MODE_SIZE (mode) > 16);
5909 /* Handle the FUNCTION_ARG_ADVANCE macro.
5910 Update the data in CUM to advance over an argument
5911 of mode MODE and data type TYPE.
5912 TYPE is null for libcalls where that information may not be available. */
5915 function_arg_advance (struct sparc_args *cum, enum machine_mode mode,
5916 tree type, int named)
5918 int slotno, regno, padding;
5920 /* We pass 0 for incoming_p here, it doesn't matter. */
5921 slotno = function_arg_slotno (cum, mode, type, named, 0, ®no, &padding);
5923 /* If register required leading padding, add it. */
5925 cum->words += padding;
5929 cum->words += (mode != BLKmode
5930 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
5931 : ROUND_ADVANCE (int_size_in_bytes (type)));
5935 if (type && AGGREGATE_TYPE_P (type))
5937 int size = int_size_in_bytes (type);
5941 else if (size <= 16)
5943 else /* passed by reference */
5948 cum->words += (mode != BLKmode
5949 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
5950 : ROUND_ADVANCE (int_size_in_bytes (type)));
5955 /* Handle the FUNCTION_ARG_PADDING macro.
5956 For the 64 bit ABI structs are always stored left shifted in their
5960 function_arg_padding (enum machine_mode mode, tree type)
5962 if (TARGET_ARCH64 && type != 0 && AGGREGATE_TYPE_P (type))
5965 /* Fall back to the default. */
5966 return DEFAULT_FUNCTION_ARG_PADDING (mode, type);
5969 /* Handle the TARGET_RETURN_IN_MEMORY target hook.
5970 Specify whether to return the return value in memory. */
5973 sparc_return_in_memory (tree type, tree fntype ATTRIBUTE_UNUSED)
5976 /* Original SPARC 32-bit ABI says that structures and unions,
5977 and quad-precision floats are returned in memory. All other
5978 base types are returned in registers.
5980 Extended ABI (as implemented by the Sun compiler) says that
5981 all complex floats are returned in registers (8 FP registers
5982 at most for '_Complex long double'). Return all complex integers
5983 in registers (4 at most for '_Complex long long').
5985 Vector ABI (as implemented by the Sun VIS SDK) says that vector
5986 integers are returned like floats of the same size, that is in
5987 registers up to 8 bytes and in memory otherwise. Return all
5988 vector floats in memory like structure and unions; note that
5989 they always have BLKmode like the latter. */
5990 return (TYPE_MODE (type) == BLKmode
5991 || TYPE_MODE (type) == TFmode
5992 || (TREE_CODE (type) == VECTOR_TYPE
5993 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8));
5995 /* Original SPARC 64-bit ABI says that structures and unions
5996 smaller than 32 bytes are returned in registers, as well as
5997 all other base types.
5999 Extended ABI (as implemented by the Sun compiler) says that all
6000 complex floats are returned in registers (8 FP registers at most
6001 for '_Complex long double'). Return all complex integers in
6002 registers (4 at most for '_Complex TItype').
6004 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6005 integers are returned like floats of the same size, that is in
6006 registers. Return all vector floats like structure and unions;
6007 note that they always have BLKmode like the latter. */
6008 return ((TYPE_MODE (type) == BLKmode
6009 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 32));
6012 /* Handle the TARGET_STRUCT_VALUE target hook.
6013 Return where to find the structure return value address. */
6016 sparc_struct_value_rtx (tree fndecl ATTRIBUTE_UNUSED, int incoming)
6023 return gen_rtx_MEM (Pmode, plus_constant (frame_pointer_rtx,
6024 STRUCT_VALUE_OFFSET));
6026 return gen_rtx_MEM (Pmode, plus_constant (stack_pointer_rtx,
6027 STRUCT_VALUE_OFFSET));
6031 /* Handle FUNCTION_VALUE, FUNCTION_OUTGOING_VALUE, and LIBCALL_VALUE macros.
6032 For v9, function return values are subject to the same rules as arguments,
6033 except that up to 32 bytes may be returned in registers. */
6036 function_value (tree type, enum machine_mode mode, int incoming_p)
6038 /* Beware that the two values are swapped here wrt function_arg. */
6039 int regbase = (incoming_p
6040 ? SPARC_OUTGOING_INT_ARG_FIRST
6041 : SPARC_INCOMING_INT_ARG_FIRST);
6042 enum mode_class mclass = GET_MODE_CLASS (mode);
6045 if (type && TREE_CODE (type) == VECTOR_TYPE)
6047 /* Vector types deserve special treatment because they are
6048 polymorphic wrt their mode, depending upon whether VIS
6049 instructions are enabled. */
6050 HOST_WIDE_INT size = int_size_in_bytes (type);
6052 if ((TARGET_ARCH32 && size > 8) || (TARGET_ARCH64 && size > 32))
6053 abort (); /* shouldn't get here */
6055 if (mode == BLKmode)
6056 return function_arg_vector_value (size,
6057 TYPE_MODE (TREE_TYPE (type)),
6058 SPARC_FP_ARG_FIRST);
6060 mclass = MODE_FLOAT;
6062 else if (type && TARGET_ARCH64)
6064 if (TREE_CODE (type) == RECORD_TYPE)
6066 /* Structures up to 32 bytes in size are passed in registers,
6067 promoted to fp registers where possible. */
6069 if (int_size_in_bytes (type) > 32)
6070 abort (); /* shouldn't get here */
6072 return function_arg_record_value (type, mode, 0, 1, regbase);
6074 else if (TREE_CODE (type) == UNION_TYPE)
6076 HOST_WIDE_INT size = int_size_in_bytes (type);
6079 abort (); /* shouldn't get here */
6081 return function_arg_union_value (size, mode, regbase);
6083 else if (AGGREGATE_TYPE_P (type))
6085 /* All other aggregate types are passed in an integer register
6086 in a mode corresponding to the size of the type. */
6087 HOST_WIDE_INT bytes = int_size_in_bytes (type);
6090 abort (); /* shouldn't get here */
6092 mode = mode_for_size (bytes * BITS_PER_UNIT, MODE_INT, 0);
6094 /* ??? We probably should have made the same ABI change in
6095 3.4.0 as the one we made for unions. The latter was
6096 required by the SCD though, while the former is not
6097 specified, so we favored compatibility and efficiency.
6099 Now we're stuck for aggregates larger than 16 bytes,
6100 because OImode vanished in the meantime. Let's not
6101 try to be unduly clever, and simply follow the ABI
6102 for unions in that case. */
6103 if (mode == BLKmode)
6104 return function_arg_union_value (bytes, mode, regbase);
6108 else if (mclass == MODE_INT
6109 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
6113 if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT)
6115 regno = SPARC_FP_ARG_FIRST;
6119 return gen_rtx_REG (mode, regno);
6122 /* Do what is necessary for `va_start'. We look at the current function
6123 to determine if stdarg or varargs is used and return the address of
6124 the first unnamed parameter. */
6127 sparc_builtin_saveregs (void)
6129 int first_reg = current_function_args_info.words;
6133 for (regno = first_reg; regno < SPARC_INT_ARG_MAX; regno++)
6134 emit_move_insn (gen_rtx_MEM (word_mode,
6135 gen_rtx_PLUS (Pmode,
6137 GEN_INT (FIRST_PARM_OFFSET (0)
6140 gen_rtx_REG (word_mode,
6141 SPARC_INCOMING_INT_ARG_FIRST + regno));
6143 address = gen_rtx_PLUS (Pmode,
6145 GEN_INT (FIRST_PARM_OFFSET (0)
6146 + UNITS_PER_WORD * first_reg));
6151 /* Implement `va_start' for stdarg. */
6154 sparc_va_start (tree valist, rtx nextarg)
6156 nextarg = expand_builtin_saveregs ();
6157 std_expand_builtin_va_start (valist, nextarg);
6160 /* Implement `va_arg' for stdarg. */
6163 sparc_gimplify_va_arg (tree valist, tree type, tree *pre_p, tree *post_p)
6165 HOST_WIDE_INT size, rsize, align;
6168 tree ptrtype = build_pointer_type (type);
6170 if (pass_by_reference (NULL, TYPE_MODE (type), type, false))
6173 size = rsize = UNITS_PER_WORD;
6179 size = int_size_in_bytes (type);
6180 rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD;
6185 /* For SPARC64, objects requiring 16-byte alignment get it. */
6186 if (TYPE_ALIGN (type) >= 2 * (unsigned) BITS_PER_WORD)
6187 align = 2 * UNITS_PER_WORD;
6189 /* SPARC-V9 ABI states that structures up to 16 bytes in size
6190 are left-justified in their slots. */
6191 if (AGGREGATE_TYPE_P (type))
6194 size = rsize = UNITS_PER_WORD;
6204 incr = fold (build2 (PLUS_EXPR, ptr_type_node, incr,
6205 ssize_int (align - 1)));
6206 incr = fold (build2 (BIT_AND_EXPR, ptr_type_node, incr,
6207 ssize_int (-align)));
6210 gimplify_expr (&incr, pre_p, post_p, is_gimple_val, fb_rvalue);
6213 if (BYTES_BIG_ENDIAN && size < rsize)
6214 addr = fold (build2 (PLUS_EXPR, ptr_type_node, incr,
6215 ssize_int (rsize - size)));
6219 addr = fold_convert (build_pointer_type (ptrtype), addr);
6220 addr = build_va_arg_indirect_ref (addr);
6222 /* If the address isn't aligned properly for the type,
6223 we may need to copy to a temporary.
6224 FIXME: This is inefficient. Usually we can do this
6227 && TYPE_ALIGN (type) > BITS_PER_WORD)
6229 tree tmp = create_tmp_var (type, "va_arg_tmp");
6230 tree dest_addr = build_fold_addr_expr (tmp);
6232 tree copy = build_function_call_expr
6233 (implicit_built_in_decls[BUILT_IN_MEMCPY],
6234 tree_cons (NULL_TREE, dest_addr,
6235 tree_cons (NULL_TREE, addr,
6236 tree_cons (NULL_TREE, size_int (rsize),
6239 gimplify_and_add (copy, pre_p);
6243 addr = fold_convert (ptrtype, addr);
6245 incr = fold (build2 (PLUS_EXPR, ptr_type_node, incr, ssize_int (rsize)));
6246 incr = build2 (MODIFY_EXPR, ptr_type_node, valist, incr);
6247 gimplify_and_add (incr, post_p);
6249 return build_va_arg_indirect_ref (addr);
6252 /* Return the string to output an unconditional branch to LABEL, which is
6253 the operand number of the label.
6255 DEST is the destination insn (i.e. the label), INSN is the source. */
6258 output_ubranch (rtx dest, int label, rtx insn)
6260 static char string[64];
6261 bool v9_form = false;
6264 if (TARGET_V9 && INSN_ADDRESSES_SET_P ())
6266 int delta = (INSN_ADDRESSES (INSN_UID (dest))
6267 - INSN_ADDRESSES (INSN_UID (insn)));
6268 /* Leave some instructions for "slop". */
6269 if (delta >= -260000 && delta < 260000)
6274 strcpy (string, "ba%*,pt\t%%xcc, ");
6276 strcpy (string, "b%*\t");
6278 p = strchr (string, '\0');
6289 /* Return the string to output a conditional branch to LABEL, which is
6290 the operand number of the label. OP is the conditional expression.
6291 XEXP (OP, 0) is assumed to be a condition code register (integer or
6292 floating point) and its mode specifies what kind of comparison we made.
6294 DEST is the destination insn (i.e. the label), INSN is the source.
6296 REVERSED is nonzero if we should reverse the sense of the comparison.
6298 ANNUL is nonzero if we should generate an annulling branch. */
6301 output_cbranch (rtx op, rtx dest, int label, int reversed, int annul,
6304 static char string[64];
6305 enum rtx_code code = GET_CODE (op);
6306 rtx cc_reg = XEXP (op, 0);
6307 enum machine_mode mode = GET_MODE (cc_reg);
6308 const char *labelno, *branch;
6309 int spaces = 8, far;
6312 /* v9 branches are limited to +-1MB. If it is too far away,
6325 fbne,a,pn %fcc2, .LC29
6333 far = TARGET_V9 && (get_attr_length (insn) >= 3);
6336 /* Reversal of FP compares takes care -- an ordered compare
6337 becomes an unordered compare and vice versa. */
6338 if (mode == CCFPmode || mode == CCFPEmode)
6339 code = reverse_condition_maybe_unordered (code);
6341 code = reverse_condition (code);
6344 /* Start by writing the branch condition. */
6345 if (mode == CCFPmode || mode == CCFPEmode)
6396 /* ??? !v9: FP branches cannot be preceded by another floating point
6397 insn. Because there is currently no concept of pre-delay slots,
6398 we can fix this only by always emitting a nop before a floating
6403 strcpy (string, "nop\n\t");
6404 strcat (string, branch);
6417 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
6429 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
6450 strcpy (string, branch);
6452 spaces -= strlen (branch);
6453 p = strchr (string, '\0');
6455 /* Now add the annulling, the label, and a possible noop. */
6468 if (! far && insn && INSN_ADDRESSES_SET_P ())
6470 int delta = (INSN_ADDRESSES (INSN_UID (dest))
6471 - INSN_ADDRESSES (INSN_UID (insn)));
6472 /* Leave some instructions for "slop". */
6473 if (delta < -260000 || delta >= 260000)
6477 if (mode == CCFPmode || mode == CCFPEmode)
6479 static char v9_fcc_labelno[] = "%%fccX, ";
6480 /* Set the char indicating the number of the fcc reg to use. */
6481 v9_fcc_labelno[5] = REGNO (cc_reg) - SPARC_FIRST_V9_FCC_REG + '0';
6482 labelno = v9_fcc_labelno;
6485 if (REGNO (cc_reg) == SPARC_FCC_REG)
6491 else if (mode == CCXmode || mode == CCX_NOOVmode)
6493 labelno = "%%xcc, ";
6499 labelno = "%%icc, ";
6504 if (*labelno && insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
6507 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
6520 strcpy (p, labelno);
6521 p = strchr (p, '\0');
6524 strcpy (p, ".+12\n\t nop\n\tb\t");
6525 /* Skip the next insn if requested or
6526 if we know that it will be a nop. */
6527 if (annul || ! final_sequence)
6541 /* Emit a library call comparison between floating point X and Y.
6542 COMPARISON is the rtl operator to compare with (EQ, NE, GT, etc.).
6543 TARGET_ARCH64 uses _Qp_* functions, which use pointers to TFmode
6544 values as arguments instead of the TFmode registers themselves,
6545 that's why we cannot call emit_float_lib_cmp. */
6547 sparc_emit_float_lib_cmp (rtx x, rtx y, enum rtx_code comparison)
6550 rtx slot0, slot1, result, tem, tem2;
6551 enum machine_mode mode;
6556 qpfunc = (TARGET_ARCH64) ? "_Qp_feq" : "_Q_feq";
6560 qpfunc = (TARGET_ARCH64) ? "_Qp_fne" : "_Q_fne";
6564 qpfunc = (TARGET_ARCH64) ? "_Qp_fgt" : "_Q_fgt";
6568 qpfunc = (TARGET_ARCH64) ? "_Qp_fge" : "_Q_fge";
6572 qpfunc = (TARGET_ARCH64) ? "_Qp_flt" : "_Q_flt";
6576 qpfunc = (TARGET_ARCH64) ? "_Qp_fle" : "_Q_fle";
6587 qpfunc = (TARGET_ARCH64) ? "_Qp_cmp" : "_Q_cmp";
6597 if (GET_CODE (x) != MEM)
6599 slot0 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0);
6600 emit_insn (gen_rtx_SET (VOIDmode, slot0, x));
6605 if (GET_CODE (y) != MEM)
6607 slot1 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0);
6608 emit_insn (gen_rtx_SET (VOIDmode, slot1, y));
6613 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, qpfunc), LCT_NORMAL,
6615 XEXP (slot0, 0), Pmode,
6616 XEXP (slot1, 0), Pmode);
6622 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, qpfunc), LCT_NORMAL,
6624 x, TFmode, y, TFmode);
6630 /* Immediately move the result of the libcall into a pseudo
6631 register so reload doesn't clobber the value if it needs
6632 the return register for a spill reg. */
6633 result = gen_reg_rtx (mode);
6634 emit_move_insn (result, hard_libcall_value (mode));
6639 emit_cmp_insn (result, const0_rtx, NE, NULL_RTX, mode, 0);
6643 emit_cmp_insn (result, GEN_INT(3), comparison == UNORDERED ? EQ : NE,
6648 emit_cmp_insn (result, const1_rtx,
6649 comparison == UNGT ? GT : NE, NULL_RTX, mode, 0);
6652 emit_cmp_insn (result, const2_rtx, NE, NULL_RTX, mode, 0);
6655 tem = gen_reg_rtx (mode);
6657 emit_insn (gen_andsi3 (tem, result, const1_rtx));
6659 emit_insn (gen_anddi3 (tem, result, const1_rtx));
6660 emit_cmp_insn (tem, const0_rtx, NE, NULL_RTX, mode, 0);
6664 tem = gen_reg_rtx (mode);
6666 emit_insn (gen_addsi3 (tem, result, const1_rtx));
6668 emit_insn (gen_adddi3 (tem, result, const1_rtx));
6669 tem2 = gen_reg_rtx (mode);
6671 emit_insn (gen_andsi3 (tem2, tem, const2_rtx));
6673 emit_insn (gen_anddi3 (tem2, tem, const2_rtx));
6674 emit_cmp_insn (tem2, const0_rtx, comparison == UNEQ ? EQ : NE,
6680 /* Generate an unsigned DImode to FP conversion. This is the same code
6681 optabs would emit if we didn't have TFmode patterns. */
6684 sparc_emit_floatunsdi (rtx *operands, enum machine_mode mode)
6686 rtx neglab, donelab, i0, i1, f0, in, out;
6689 in = force_reg (DImode, operands[1]);
6690 neglab = gen_label_rtx ();
6691 donelab = gen_label_rtx ();
6692 i0 = gen_reg_rtx (DImode);
6693 i1 = gen_reg_rtx (DImode);
6694 f0 = gen_reg_rtx (mode);
6696 emit_cmp_and_jump_insns (in, const0_rtx, LT, const0_rtx, DImode, 0, neglab);
6698 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_FLOAT (mode, in)));
6699 emit_jump_insn (gen_jump (donelab));
6702 emit_label (neglab);
6704 emit_insn (gen_lshrdi3 (i0, in, const1_rtx));
6705 emit_insn (gen_anddi3 (i1, in, const1_rtx));
6706 emit_insn (gen_iordi3 (i0, i0, i1));
6707 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_FLOAT (mode, i0)));
6708 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_PLUS (mode, f0, f0)));
6710 emit_label (donelab);
6713 /* Generate an FP to unsigned DImode conversion. This is the same code
6714 optabs would emit if we didn't have TFmode patterns. */
6717 sparc_emit_fixunsdi (rtx *operands, enum machine_mode mode)
6719 rtx neglab, donelab, i0, i1, f0, in, out, limit;
6722 in = force_reg (mode, operands[1]);
6723 neglab = gen_label_rtx ();
6724 donelab = gen_label_rtx ();
6725 i0 = gen_reg_rtx (DImode);
6726 i1 = gen_reg_rtx (DImode);
6727 limit = gen_reg_rtx (mode);
6728 f0 = gen_reg_rtx (mode);
6730 emit_move_insn (limit,
6731 CONST_DOUBLE_FROM_REAL_VALUE (
6732 REAL_VALUE_ATOF ("9223372036854775808.0", mode), mode));
6733 emit_cmp_and_jump_insns (in, limit, GE, NULL_RTX, mode, 0, neglab);
6735 emit_insn (gen_rtx_SET (VOIDmode,
6737 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, in))));
6738 emit_jump_insn (gen_jump (donelab));
6741 emit_label (neglab);
6743 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_MINUS (mode, in, limit)));
6744 emit_insn (gen_rtx_SET (VOIDmode,
6746 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, f0))));
6747 emit_insn (gen_movdi (i1, const1_rtx));
6748 emit_insn (gen_ashldi3 (i1, i1, GEN_INT (63)));
6749 emit_insn (gen_xordi3 (out, i0, i1));
6751 emit_label (donelab);
6754 /* Return the string to output a conditional branch to LABEL, testing
6755 register REG. LABEL is the operand number of the label; REG is the
6756 operand number of the reg. OP is the conditional expression. The mode
6757 of REG says what kind of comparison we made.
6759 DEST is the destination insn (i.e. the label), INSN is the source.
6761 REVERSED is nonzero if we should reverse the sense of the comparison.
6763 ANNUL is nonzero if we should generate an annulling branch. */
6766 output_v9branch (rtx op, rtx dest, int reg, int label, int reversed,
6767 int annul, rtx insn)
6769 static char string[64];
6770 enum rtx_code code = GET_CODE (op);
6771 enum machine_mode mode = GET_MODE (XEXP (op, 0));
6776 /* branch on register are limited to +-128KB. If it is too far away,
6789 brgez,a,pn %o1, .LC29
6795 ba,pt %xcc, .LC29 */
6797 far = get_attr_length (insn) >= 3;
6799 /* If not floating-point or if EQ or NE, we can just reverse the code. */
6801 code = reverse_condition (code);
6803 /* Only 64 bit versions of these instructions exist. */
6807 /* Start by writing the branch condition. */
6812 strcpy (string, "brnz");
6816 strcpy (string, "brz");
6820 strcpy (string, "brgez");
6824 strcpy (string, "brlz");
6828 strcpy (string, "brlez");
6832 strcpy (string, "brgz");
6839 p = strchr (string, '\0');
6841 /* Now add the annulling, reg, label, and nop. */
6848 if (insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
6851 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
6856 *p = p < string + 8 ? '\t' : ' ';
6864 int veryfar = 1, delta;
6866 if (INSN_ADDRESSES_SET_P ())
6868 delta = (INSN_ADDRESSES (INSN_UID (dest))
6869 - INSN_ADDRESSES (INSN_UID (insn)));
6870 /* Leave some instructions for "slop". */
6871 if (delta >= -260000 && delta < 260000)
6875 strcpy (p, ".+12\n\t nop\n\t");
6876 /* Skip the next insn if requested or
6877 if we know that it will be a nop. */
6878 if (annul || ! final_sequence)
6888 strcpy (p, "ba,pt\t%%xcc, ");
6902 /* Return 1, if any of the registers of the instruction are %l[0-7] or %o[0-7].
6903 Such instructions cannot be used in the delay slot of return insn on v9.
6904 If TEST is 0, also rename all %i[0-7] registers to their %o[0-7] counterparts.
6908 epilogue_renumber (register rtx *where, int test)
6910 register const char *fmt;
6912 register enum rtx_code code;
6917 code = GET_CODE (*where);
6922 if (REGNO (*where) >= 8 && REGNO (*where) < 24) /* oX or lX */
6924 if (! test && REGNO (*where) >= 24 && REGNO (*where) < 32)
6925 *where = gen_rtx_REG (GET_MODE (*where), OUTGOING_REGNO (REGNO(*where)));
6933 /* Do not replace the frame pointer with the stack pointer because
6934 it can cause the delayed instruction to load below the stack.
6935 This occurs when instructions like:
6937 (set (reg/i:SI 24 %i0)
6938 (mem/f:SI (plus:SI (reg/f:SI 30 %fp)
6939 (const_int -20 [0xffffffec])) 0))
6941 are in the return delayed slot. */
6943 if (GET_CODE (XEXP (*where, 0)) == REG
6944 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM
6945 && (GET_CODE (XEXP (*where, 1)) != CONST_INT
6946 || INTVAL (XEXP (*where, 1)) < SPARC_STACK_BIAS))
6951 if (SPARC_STACK_BIAS
6952 && GET_CODE (XEXP (*where, 0)) == REG
6953 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM)
6961 fmt = GET_RTX_FORMAT (code);
6963 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6968 for (j = XVECLEN (*where, i) - 1; j >= 0; j--)
6969 if (epilogue_renumber (&(XVECEXP (*where, i, j)), test))
6972 else if (fmt[i] == 'e'
6973 && epilogue_renumber (&(XEXP (*where, i)), test))
6979 /* Leaf functions and non-leaf functions have different needs. */
6982 reg_leaf_alloc_order[] = REG_LEAF_ALLOC_ORDER;
6985 reg_nonleaf_alloc_order[] = REG_ALLOC_ORDER;
6987 static const int *const reg_alloc_orders[] = {
6988 reg_leaf_alloc_order,
6989 reg_nonleaf_alloc_order};
6992 order_regs_for_local_alloc (void)
6994 static int last_order_nonleaf = 1;
6996 if (regs_ever_live[15] != last_order_nonleaf)
6998 last_order_nonleaf = !last_order_nonleaf;
6999 memcpy ((char *) reg_alloc_order,
7000 (const char *) reg_alloc_orders[last_order_nonleaf],
7001 FIRST_PSEUDO_REGISTER * sizeof (int));
7005 /* Return 1 if REG and MEM are legitimate enough to allow the various
7006 mem<-->reg splits to be run. */
7009 sparc_splitdi_legitimate (rtx reg, rtx mem)
7011 /* Punt if we are here by mistake. */
7012 if (! reload_completed)
7015 /* We must have an offsettable memory reference. */
7016 if (! offsettable_memref_p (mem))
7019 /* If we have legitimate args for ldd/std, we do not want
7020 the split to happen. */
7021 if ((REGNO (reg) % 2) == 0
7022 && mem_min_alignment (mem, 8))
7029 /* Return 1 if x and y are some kind of REG and they refer to
7030 different hard registers. This test is guaranteed to be
7031 run after reload. */
7034 sparc_absnegfloat_split_legitimate (rtx x, rtx y)
7036 if (GET_CODE (x) != REG)
7038 if (GET_CODE (y) != REG)
7040 if (REGNO (x) == REGNO (y))
7045 /* Return 1 if REGNO (reg1) is even and REGNO (reg1) == REGNO (reg2) - 1.
7046 This makes them candidates for using ldd and std insns.
7048 Note reg1 and reg2 *must* be hard registers. */
7051 registers_ok_for_ldd_peep (rtx reg1, rtx reg2)
7053 /* We might have been passed a SUBREG. */
7054 if (GET_CODE (reg1) != REG || GET_CODE (reg2) != REG)
7057 if (REGNO (reg1) % 2 != 0)
7060 /* Integer ldd is deprecated in SPARC V9 */
7061 if (TARGET_V9 && REGNO (reg1) < 32)
7064 return (REGNO (reg1) == REGNO (reg2) - 1);
7067 /* Return 1 if the addresses in mem1 and mem2 are suitable for use in
7070 This can only happen when addr1 and addr2, the addresses in mem1
7071 and mem2, are consecutive memory locations (addr1 + 4 == addr2).
7072 addr1 must also be aligned on a 64-bit boundary.
7074 Also iff dependent_reg_rtx is not null it should not be used to
7075 compute the address for mem1, i.e. we cannot optimize a sequence
7087 But, note that the transformation from:
7092 is perfectly fine. Thus, the peephole2 patterns always pass us
7093 the destination register of the first load, never the second one.
7095 For stores we don't have a similar problem, so dependent_reg_rtx is
7099 mems_ok_for_ldd_peep (rtx mem1, rtx mem2, rtx dependent_reg_rtx)
7103 HOST_WIDE_INT offset1;
7105 /* The mems cannot be volatile. */
7106 if (MEM_VOLATILE_P (mem1) || MEM_VOLATILE_P (mem2))
7109 /* MEM1 should be aligned on a 64-bit boundary. */
7110 if (MEM_ALIGN (mem1) < 64)
7113 addr1 = XEXP (mem1, 0);
7114 addr2 = XEXP (mem2, 0);
7116 /* Extract a register number and offset (if used) from the first addr. */
7117 if (GET_CODE (addr1) == PLUS)
7119 /* If not a REG, return zero. */
7120 if (GET_CODE (XEXP (addr1, 0)) != REG)
7124 reg1 = REGNO (XEXP (addr1, 0));
7125 /* The offset must be constant! */
7126 if (GET_CODE (XEXP (addr1, 1)) != CONST_INT)
7128 offset1 = INTVAL (XEXP (addr1, 1));
7131 else if (GET_CODE (addr1) != REG)
7135 reg1 = REGNO (addr1);
7136 /* This was a simple (mem (reg)) expression. Offset is 0. */
7140 /* Make sure the second address is a (mem (plus (reg) (const_int). */
7141 if (GET_CODE (addr2) != PLUS)
7144 if (GET_CODE (XEXP (addr2, 0)) != REG
7145 || GET_CODE (XEXP (addr2, 1)) != CONST_INT)
7148 if (reg1 != REGNO (XEXP (addr2, 0)))
7151 if (dependent_reg_rtx != NULL_RTX && reg1 == REGNO (dependent_reg_rtx))
7154 /* The first offset must be evenly divisible by 8 to ensure the
7155 address is 64 bit aligned. */
7156 if (offset1 % 8 != 0)
7159 /* The offset for the second addr must be 4 more than the first addr. */
7160 if (INTVAL (XEXP (addr2, 1)) != offset1 + 4)
7163 /* All the tests passed. addr1 and addr2 are valid for ldd and std
7168 /* Return 1 if reg is a pseudo, or is the first register in
7169 a hard register pair. This makes it a candidate for use in
7170 ldd and std insns. */
7173 register_ok_for_ldd (rtx reg)
7175 /* We might have been passed a SUBREG. */
7176 if (GET_CODE (reg) != REG)
7179 if (REGNO (reg) < FIRST_PSEUDO_REGISTER)
7180 return (REGNO (reg) % 2 == 0);
7185 /* Print operand X (an rtx) in assembler syntax to file FILE.
7186 CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified.
7187 For `%' followed by punctuation, CODE is the punctuation and X is null. */
7190 print_operand (FILE *file, rtx x, int code)
7195 /* Output an insn in a delay slot. */
7197 sparc_indent_opcode = 1;
7199 fputs ("\n\t nop", file);
7202 /* Output an annul flag if there's nothing for the delay slot and we
7203 are optimizing. This is always used with '(' below.
7204 Sun OS 4.1.1 dbx can't handle an annulled unconditional branch;
7205 this is a dbx bug. So, we only do this when optimizing.
7206 On UltraSPARC, a branch in a delay slot causes a pipeline flush.
7207 Always emit a nop in case the next instruction is a branch. */
7208 if (! final_sequence && (optimize && (int)sparc_cpu < PROCESSOR_V9))
7212 /* Output a 'nop' if there's nothing for the delay slot and we are
7213 not optimizing. This is always used with '*' above. */
7214 if (! final_sequence && ! (optimize && (int)sparc_cpu < PROCESSOR_V9))
7215 fputs ("\n\t nop", file);
7216 else if (final_sequence)
7217 sparc_indent_opcode = 1;
7220 /* Output the right displacement from the saved PC on function return.
7221 The caller may have placed an "unimp" insn immediately after the call
7222 so we have to account for it. This insn is used in the 32-bit ABI
7223 when calling a function that returns a non zero-sized structure. The
7224 64-bit ABI doesn't have it. Be careful to have this test be the same
7225 as that used on the call. */
7227 && current_function_returns_struct
7228 && (TREE_CODE (DECL_SIZE (DECL_RESULT (current_function_decl)))
7230 && ! integer_zerop (DECL_SIZE (DECL_RESULT (current_function_decl))))
7236 /* Output the Embedded Medium/Anywhere code model base register. */
7237 fputs (EMBMEDANY_BASE_REG, file);
7240 /* Print some local dynamic TLS name. */
7241 assemble_name (file, get_some_local_dynamic_name ());
7245 /* Adjust the operand to take into account a RESTORE operation. */
7246 if (GET_CODE (x) == CONST_INT)
7248 else if (GET_CODE (x) != REG)
7249 output_operand_lossage ("invalid %%Y operand");
7250 else if (REGNO (x) < 8)
7251 fputs (reg_names[REGNO (x)], file);
7252 else if (REGNO (x) >= 24 && REGNO (x) < 32)
7253 fputs (reg_names[REGNO (x)-16], file);
7255 output_operand_lossage ("invalid %%Y operand");
7258 /* Print out the low order register name of a register pair. */
7259 if (WORDS_BIG_ENDIAN)
7260 fputs (reg_names[REGNO (x)+1], file);
7262 fputs (reg_names[REGNO (x)], file);
7265 /* Print out the high order register name of a register pair. */
7266 if (WORDS_BIG_ENDIAN)
7267 fputs (reg_names[REGNO (x)], file);
7269 fputs (reg_names[REGNO (x)+1], file);
7272 /* Print out the second register name of a register pair or quad.
7273 I.e., R (%o0) => %o1. */
7274 fputs (reg_names[REGNO (x)+1], file);
7277 /* Print out the third register name of a register quad.
7278 I.e., S (%o0) => %o2. */
7279 fputs (reg_names[REGNO (x)+2], file);
7282 /* Print out the fourth register name of a register quad.
7283 I.e., T (%o0) => %o3. */
7284 fputs (reg_names[REGNO (x)+3], file);
7287 /* Print a condition code register. */
7288 if (REGNO (x) == SPARC_ICC_REG)
7290 /* We don't handle CC[X]_NOOVmode because they're not supposed
7292 if (GET_MODE (x) == CCmode)
7293 fputs ("%icc", file);
7294 else if (GET_MODE (x) == CCXmode)
7295 fputs ("%xcc", file);
7300 /* %fccN register */
7301 fputs (reg_names[REGNO (x)], file);
7304 /* Print the operand's address only. */
7305 output_address (XEXP (x, 0));
7308 /* In this case we need a register. Use %g0 if the
7309 operand is const0_rtx. */
7311 || (GET_MODE (x) != VOIDmode && x == CONST0_RTX (GET_MODE (x))))
7313 fputs ("%g0", file);
7320 switch (GET_CODE (x))
7322 case IOR: fputs ("or", file); break;
7323 case AND: fputs ("and", file); break;
7324 case XOR: fputs ("xor", file); break;
7325 default: output_operand_lossage ("invalid %%A operand");
7330 switch (GET_CODE (x))
7332 case IOR: fputs ("orn", file); break;
7333 case AND: fputs ("andn", file); break;
7334 case XOR: fputs ("xnor", file); break;
7335 default: output_operand_lossage ("invalid %%B operand");
7339 /* These are used by the conditional move instructions. */
7343 enum rtx_code rc = GET_CODE (x);
7347 enum machine_mode mode = GET_MODE (XEXP (x, 0));
7348 if (mode == CCFPmode || mode == CCFPEmode)
7349 rc = reverse_condition_maybe_unordered (GET_CODE (x));
7351 rc = reverse_condition (GET_CODE (x));
7355 case NE: fputs ("ne", file); break;
7356 case EQ: fputs ("e", file); break;
7357 case GE: fputs ("ge", file); break;
7358 case GT: fputs ("g", file); break;
7359 case LE: fputs ("le", file); break;
7360 case LT: fputs ("l", file); break;
7361 case GEU: fputs ("geu", file); break;
7362 case GTU: fputs ("gu", file); break;
7363 case LEU: fputs ("leu", file); break;
7364 case LTU: fputs ("lu", file); break;
7365 case LTGT: fputs ("lg", file); break;
7366 case UNORDERED: fputs ("u", file); break;
7367 case ORDERED: fputs ("o", file); break;
7368 case UNLT: fputs ("ul", file); break;
7369 case UNLE: fputs ("ule", file); break;
7370 case UNGT: fputs ("ug", file); break;
7371 case UNGE: fputs ("uge", file); break;
7372 case UNEQ: fputs ("ue", file); break;
7373 default: output_operand_lossage (code == 'c'
7374 ? "invalid %%c operand"
7375 : "invalid %%C operand");
7380 /* These are used by the movr instruction pattern. */
7384 enum rtx_code rc = (code == 'd'
7385 ? reverse_condition (GET_CODE (x))
7389 case NE: fputs ("ne", file); break;
7390 case EQ: fputs ("e", file); break;
7391 case GE: fputs ("gez", file); break;
7392 case LT: fputs ("lz", file); break;
7393 case LE: fputs ("lez", file); break;
7394 case GT: fputs ("gz", file); break;
7395 default: output_operand_lossage (code == 'd'
7396 ? "invalid %%d operand"
7397 : "invalid %%D operand");
7404 /* Print a sign-extended character. */
7405 int i = trunc_int_for_mode (INTVAL (x), QImode);
7406 fprintf (file, "%d", i);
7411 /* Operand must be a MEM; write its address. */
7412 if (GET_CODE (x) != MEM)
7413 output_operand_lossage ("invalid %%f operand");
7414 output_address (XEXP (x, 0));
7419 /* Print a sign-extended 32-bit value. */
7421 if (GET_CODE(x) == CONST_INT)
7423 else if (GET_CODE(x) == CONST_DOUBLE)
7424 i = CONST_DOUBLE_LOW (x);
7427 output_operand_lossage ("invalid %%s operand");
7430 i = trunc_int_for_mode (i, SImode);
7431 fprintf (file, HOST_WIDE_INT_PRINT_DEC, i);
7436 /* Do nothing special. */
7440 /* Undocumented flag. */
7441 output_operand_lossage ("invalid operand output code");
7444 if (GET_CODE (x) == REG)
7445 fputs (reg_names[REGNO (x)], file);
7446 else if (GET_CODE (x) == MEM)
7449 /* Poor Sun assembler doesn't understand absolute addressing. */
7450 if (CONSTANT_P (XEXP (x, 0)))
7451 fputs ("%g0+", file);
7452 output_address (XEXP (x, 0));
7455 else if (GET_CODE (x) == HIGH)
7457 fputs ("%hi(", file);
7458 output_addr_const (file, XEXP (x, 0));
7461 else if (GET_CODE (x) == LO_SUM)
7463 print_operand (file, XEXP (x, 0), 0);
7464 if (TARGET_CM_MEDMID)
7465 fputs ("+%l44(", file);
7467 fputs ("+%lo(", file);
7468 output_addr_const (file, XEXP (x, 1));
7471 else if (GET_CODE (x) == CONST_DOUBLE
7472 && (GET_MODE (x) == VOIDmode
7473 || GET_MODE_CLASS (GET_MODE (x)) == MODE_INT))
7475 if (CONST_DOUBLE_HIGH (x) == 0)
7476 fprintf (file, "%u", (unsigned int) CONST_DOUBLE_LOW (x));
7477 else if (CONST_DOUBLE_HIGH (x) == -1
7478 && CONST_DOUBLE_LOW (x) < 0)
7479 fprintf (file, "%d", (int) CONST_DOUBLE_LOW (x));
7481 output_operand_lossage ("long long constant not a valid immediate operand");
7483 else if (GET_CODE (x) == CONST_DOUBLE)
7484 output_operand_lossage ("floating point constant not a valid immediate operand");
7485 else { output_addr_const (file, x); }
7488 /* Target hook for assembling integer objects. The sparc version has
7489 special handling for aligned DI-mode objects. */
7492 sparc_assemble_integer (rtx x, unsigned int size, int aligned_p)
7494 /* ??? We only output .xword's for symbols and only then in environments
7495 where the assembler can handle them. */
7496 if (aligned_p && size == 8
7497 && (GET_CODE (x) != CONST_INT && GET_CODE (x) != CONST_DOUBLE))
7501 assemble_integer_with_op ("\t.xword\t", x);
7506 assemble_aligned_integer (4, const0_rtx);
7507 assemble_aligned_integer (4, x);
7511 return default_assemble_integer (x, size, aligned_p);
7514 /* Return the value of a code used in the .proc pseudo-op that says
7515 what kind of result this function returns. For non-C types, we pick
7516 the closest C type. */
7518 #ifndef SHORT_TYPE_SIZE
7519 #define SHORT_TYPE_SIZE (BITS_PER_UNIT * 2)
7522 #ifndef INT_TYPE_SIZE
7523 #define INT_TYPE_SIZE BITS_PER_WORD
7526 #ifndef LONG_TYPE_SIZE
7527 #define LONG_TYPE_SIZE BITS_PER_WORD
7530 #ifndef LONG_LONG_TYPE_SIZE
7531 #define LONG_LONG_TYPE_SIZE (BITS_PER_WORD * 2)
7534 #ifndef FLOAT_TYPE_SIZE
7535 #define FLOAT_TYPE_SIZE BITS_PER_WORD
7538 #ifndef DOUBLE_TYPE_SIZE
7539 #define DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
7542 #ifndef LONG_DOUBLE_TYPE_SIZE
7543 #define LONG_DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
7547 sparc_type_code (register tree type)
7549 register unsigned long qualifiers = 0;
7550 register unsigned shift;
7552 /* Only the first 30 bits of the qualifier are valid. We must refrain from
7553 setting more, since some assemblers will give an error for this. Also,
7554 we must be careful to avoid shifts of 32 bits or more to avoid getting
7555 unpredictable results. */
7557 for (shift = 6; shift < 30; shift += 2, type = TREE_TYPE (type))
7559 switch (TREE_CODE (type))
7565 qualifiers |= (3 << shift);
7570 qualifiers |= (2 << shift);
7574 case REFERENCE_TYPE:
7576 qualifiers |= (1 << shift);
7580 return (qualifiers | 8);
7583 case QUAL_UNION_TYPE:
7584 return (qualifiers | 9);
7587 return (qualifiers | 10);
7590 return (qualifiers | 16);
7593 /* If this is a range type, consider it to be the underlying
7595 if (TREE_TYPE (type) != 0)
7598 /* Carefully distinguish all the standard types of C,
7599 without messing up if the language is not C. We do this by
7600 testing TYPE_PRECISION and TYPE_UNSIGNED. The old code used to
7601 look at both the names and the above fields, but that's redundant.
7602 Any type whose size is between two C types will be considered
7603 to be the wider of the two types. Also, we do not have a
7604 special code to use for "long long", so anything wider than
7605 long is treated the same. Note that we can't distinguish
7606 between "int" and "long" in this code if they are the same
7607 size, but that's fine, since neither can the assembler. */
7609 if (TYPE_PRECISION (type) <= CHAR_TYPE_SIZE)
7610 return (qualifiers | (TYPE_UNSIGNED (type) ? 12 : 2));
7612 else if (TYPE_PRECISION (type) <= SHORT_TYPE_SIZE)
7613 return (qualifiers | (TYPE_UNSIGNED (type) ? 13 : 3));
7615 else if (TYPE_PRECISION (type) <= INT_TYPE_SIZE)
7616 return (qualifiers | (TYPE_UNSIGNED (type) ? 14 : 4));
7619 return (qualifiers | (TYPE_UNSIGNED (type) ? 15 : 5));
7622 /* If this is a range type, consider it to be the underlying
7624 if (TREE_TYPE (type) != 0)
7627 /* Carefully distinguish all the standard types of C,
7628 without messing up if the language is not C. */
7630 if (TYPE_PRECISION (type) == FLOAT_TYPE_SIZE)
7631 return (qualifiers | 6);
7634 return (qualifiers | 7);
7636 case COMPLEX_TYPE: /* GNU Fortran COMPLEX type. */
7637 /* ??? We need to distinguish between double and float complex types,
7638 but I don't know how yet because I can't reach this code from
7639 existing front-ends. */
7640 return (qualifiers | 7); /* Who knows? */
7643 case CHAR_TYPE: /* GNU Pascal CHAR type. Not used in C. */
7644 case BOOLEAN_TYPE: /* GNU Fortran BOOLEAN type. */
7645 case FILE_TYPE: /* GNU Pascal FILE type. */
7646 case SET_TYPE: /* GNU Pascal SET type. */
7647 case LANG_TYPE: /* ? */
7651 abort (); /* Not a type! */
7658 /* Nested function support. */
7660 /* Emit RTL insns to initialize the variable parts of a trampoline.
7661 FNADDR is an RTX for the address of the function's pure code.
7662 CXT is an RTX for the static chain value for the function.
7664 This takes 16 insns: 2 shifts & 2 ands (to split up addresses), 4 sethi
7665 (to load in opcodes), 4 iors (to merge address and opcodes), and 4 writes
7666 (to store insns). This is a bit excessive. Perhaps a different
7667 mechanism would be better here.
7669 Emit enough FLUSH insns to synchronize the data and instruction caches. */
7672 sparc_initialize_trampoline (rtx tramp, rtx fnaddr, rtx cxt)
7674 /* SPARC 32-bit trampoline:
7677 sethi %hi(static), %g2
7679 or %g2, %lo(static), %g2
7681 SETHI i,r = 00rr rrr1 00ii iiii iiii iiii iiii iiii
7682 JMPL r+i,d = 10dd ddd1 1100 0rrr rr1i iiii iiii iiii
7686 (gen_rtx_MEM (SImode, plus_constant (tramp, 0)),
7687 expand_binop (SImode, ior_optab,
7688 expand_shift (RSHIFT_EXPR, SImode, fnaddr,
7689 size_int (10), 0, 1),
7690 GEN_INT (trunc_int_for_mode (0x03000000, SImode)),
7691 NULL_RTX, 1, OPTAB_DIRECT));
7694 (gen_rtx_MEM (SImode, plus_constant (tramp, 4)),
7695 expand_binop (SImode, ior_optab,
7696 expand_shift (RSHIFT_EXPR, SImode, cxt,
7697 size_int (10), 0, 1),
7698 GEN_INT (trunc_int_for_mode (0x05000000, SImode)),
7699 NULL_RTX, 1, OPTAB_DIRECT));
7702 (gen_rtx_MEM (SImode, plus_constant (tramp, 8)),
7703 expand_binop (SImode, ior_optab,
7704 expand_and (SImode, fnaddr, GEN_INT (0x3ff), NULL_RTX),
7705 GEN_INT (trunc_int_for_mode (0x81c06000, SImode)),
7706 NULL_RTX, 1, OPTAB_DIRECT));
7709 (gen_rtx_MEM (SImode, plus_constant (tramp, 12)),
7710 expand_binop (SImode, ior_optab,
7711 expand_and (SImode, cxt, GEN_INT (0x3ff), NULL_RTX),
7712 GEN_INT (trunc_int_for_mode (0x8410a000, SImode)),
7713 NULL_RTX, 1, OPTAB_DIRECT));
7715 /* On UltraSPARC a flush flushes an entire cache line. The trampoline is
7716 aligned on a 16 byte boundary so one flush clears it all. */
7717 emit_insn (gen_flush (validize_mem (gen_rtx_MEM (SImode, tramp))));
7718 if (sparc_cpu != PROCESSOR_ULTRASPARC
7719 && sparc_cpu != PROCESSOR_ULTRASPARC3)
7720 emit_insn (gen_flush (validize_mem (gen_rtx_MEM (SImode,
7721 plus_constant (tramp, 8)))));
7723 /* Call __enable_execute_stack after writing onto the stack to make sure
7724 the stack address is accessible. */
7725 #ifdef ENABLE_EXECUTE_STACK
7726 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
7727 LCT_NORMAL, VOIDmode, 1, tramp, Pmode);
7732 /* The 64-bit version is simpler because it makes more sense to load the
7733 values as "immediate" data out of the trampoline. It's also easier since
7734 we can read the PC without clobbering a register. */
7737 sparc64_initialize_trampoline (rtx tramp, rtx fnaddr, rtx cxt)
7739 /* SPARC 64-bit trampoline:
7748 emit_move_insn (gen_rtx_MEM (SImode, tramp),
7749 GEN_INT (trunc_int_for_mode (0x83414000, SImode)));
7750 emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 4)),
7751 GEN_INT (trunc_int_for_mode (0xca586018, SImode)));
7752 emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 8)),
7753 GEN_INT (trunc_int_for_mode (0x81c14000, SImode)));
7754 emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 12)),
7755 GEN_INT (trunc_int_for_mode (0xca586010, SImode)));
7756 emit_move_insn (gen_rtx_MEM (DImode, plus_constant (tramp, 16)), cxt);
7757 emit_move_insn (gen_rtx_MEM (DImode, plus_constant (tramp, 24)), fnaddr);
7758 emit_insn (gen_flushdi (validize_mem (gen_rtx_MEM (DImode, tramp))));
7760 if (sparc_cpu != PROCESSOR_ULTRASPARC
7761 && sparc_cpu != PROCESSOR_ULTRASPARC3)
7762 emit_insn (gen_flushdi (validize_mem (gen_rtx_MEM (DImode, plus_constant (tramp, 8)))));
7764 /* Call __enable_execute_stack after writing onto the stack to make sure
7765 the stack address is accessible. */
7766 #ifdef ENABLE_EXECUTE_STACK
7767 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
7768 LCT_NORMAL, VOIDmode, 1, tramp, Pmode);
7772 /* Adjust the cost of a scheduling dependency. Return the new cost of
7773 a dependency LINK or INSN on DEP_INSN. COST is the current cost. */
7776 supersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
7778 enum attr_type insn_type;
7780 if (! recog_memoized (insn))
7783 insn_type = get_attr_type (insn);
7785 if (REG_NOTE_KIND (link) == 0)
7787 /* Data dependency; DEP_INSN writes a register that INSN reads some
7790 /* if a load, then the dependence must be on the memory address;
7791 add an extra "cycle". Note that the cost could be two cycles
7792 if the reg was written late in an instruction group; we ca not tell
7794 if (insn_type == TYPE_LOAD || insn_type == TYPE_FPLOAD)
7797 /* Get the delay only if the address of the store is the dependence. */
7798 if (insn_type == TYPE_STORE || insn_type == TYPE_FPSTORE)
7800 rtx pat = PATTERN(insn);
7801 rtx dep_pat = PATTERN (dep_insn);
7803 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
7804 return cost; /* This should not happen! */
7806 /* The dependency between the two instructions was on the data that
7807 is being stored. Assume that this implies that the address of the
7808 store is not dependent. */
7809 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
7812 return cost + 3; /* An approximation. */
7815 /* A shift instruction cannot receive its data from an instruction
7816 in the same cycle; add a one cycle penalty. */
7817 if (insn_type == TYPE_SHIFT)
7818 return cost + 3; /* Split before cascade into shift. */
7822 /* Anti- or output- dependency; DEP_INSN reads/writes a register that
7823 INSN writes some cycles later. */
7825 /* These are only significant for the fpu unit; writing a fp reg before
7826 the fpu has finished with it stalls the processor. */
7828 /* Reusing an integer register causes no problems. */
7829 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
7837 hypersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
7839 enum attr_type insn_type, dep_type;
7840 rtx pat = PATTERN(insn);
7841 rtx dep_pat = PATTERN (dep_insn);
7843 if (recog_memoized (insn) < 0 || recog_memoized (dep_insn) < 0)
7846 insn_type = get_attr_type (insn);
7847 dep_type = get_attr_type (dep_insn);
7849 switch (REG_NOTE_KIND (link))
7852 /* Data dependency; DEP_INSN writes a register that INSN reads some
7859 /* Get the delay iff the address of the store is the dependence. */
7860 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
7863 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
7870 /* If a load, then the dependence must be on the memory address. If
7871 the addresses aren't equal, then it might be a false dependency */
7872 if (dep_type == TYPE_STORE || dep_type == TYPE_FPSTORE)
7874 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET
7875 || GET_CODE (SET_DEST (dep_pat)) != MEM
7876 || GET_CODE (SET_SRC (pat)) != MEM
7877 || ! rtx_equal_p (XEXP (SET_DEST (dep_pat), 0),
7878 XEXP (SET_SRC (pat), 0)))
7886 /* Compare to branch latency is 0. There is no benefit from
7887 separating compare and branch. */
7888 if (dep_type == TYPE_COMPARE)
7890 /* Floating point compare to branch latency is less than
7891 compare to conditional move. */
7892 if (dep_type == TYPE_FPCMP)
7901 /* Anti-dependencies only penalize the fpu unit. */
7902 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
7914 sparc_adjust_cost(rtx insn, rtx link, rtx dep, int cost)
7918 case PROCESSOR_SUPERSPARC:
7919 cost = supersparc_adjust_cost (insn, link, dep, cost);
7921 case PROCESSOR_HYPERSPARC:
7922 case PROCESSOR_SPARCLITE86X:
7923 cost = hypersparc_adjust_cost (insn, link, dep, cost);
7932 sparc_sched_init (FILE *dump ATTRIBUTE_UNUSED,
7933 int sched_verbose ATTRIBUTE_UNUSED,
7934 int max_ready ATTRIBUTE_UNUSED)
7939 sparc_use_sched_lookahead (void)
7941 if (sparc_cpu == PROCESSOR_ULTRASPARC
7942 || sparc_cpu == PROCESSOR_ULTRASPARC3)
7944 if ((1 << sparc_cpu) &
7945 ((1 << PROCESSOR_SUPERSPARC) | (1 << PROCESSOR_HYPERSPARC) |
7946 (1 << PROCESSOR_SPARCLITE86X)))
7952 sparc_issue_rate (void)
7959 /* Assume V9 processors are capable of at least dual-issue. */
7961 case PROCESSOR_SUPERSPARC:
7963 case PROCESSOR_HYPERSPARC:
7964 case PROCESSOR_SPARCLITE86X:
7966 case PROCESSOR_ULTRASPARC:
7967 case PROCESSOR_ULTRASPARC3:
7973 set_extends (rtx insn)
7975 register rtx pat = PATTERN (insn);
7977 switch (GET_CODE (SET_SRC (pat)))
7979 /* Load and some shift instructions zero extend. */
7982 /* sethi clears the high bits */
7984 /* LO_SUM is used with sethi. sethi cleared the high
7985 bits and the values used with lo_sum are positive */
7987 /* Store flag stores 0 or 1 */
7997 rtx op0 = XEXP (SET_SRC (pat), 0);
7998 rtx op1 = XEXP (SET_SRC (pat), 1);
7999 if (GET_CODE (op1) == CONST_INT)
8000 return INTVAL (op1) >= 0;
8001 if (GET_CODE (op0) != REG)
8003 if (sparc_check_64 (op0, insn) == 1)
8005 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
8010 rtx op0 = XEXP (SET_SRC (pat), 0);
8011 rtx op1 = XEXP (SET_SRC (pat), 1);
8012 if (GET_CODE (op0) != REG || sparc_check_64 (op0, insn) <= 0)
8014 if (GET_CODE (op1) == CONST_INT)
8015 return INTVAL (op1) >= 0;
8016 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
8019 return GET_MODE (SET_SRC (pat)) == SImode;
8020 /* Positive integers leave the high bits zero. */
8022 return ! (CONST_DOUBLE_LOW (SET_SRC (pat)) & 0x80000000);
8024 return ! (INTVAL (SET_SRC (pat)) & 0x80000000);
8027 return - (GET_MODE (SET_SRC (pat)) == SImode);
8029 return sparc_check_64 (SET_SRC (pat), insn);
8035 /* We _ought_ to have only one kind per function, but... */
8036 static GTY(()) rtx sparc_addr_diff_list;
8037 static GTY(()) rtx sparc_addr_list;
8040 sparc_defer_case_vector (rtx lab, rtx vec, int diff)
8042 vec = gen_rtx_EXPR_LIST (VOIDmode, lab, vec);
8044 sparc_addr_diff_list
8045 = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_diff_list);
8047 sparc_addr_list = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_list);
8051 sparc_output_addr_vec (rtx vec)
8053 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
8054 int idx, vlen = XVECLEN (body, 0);
8056 #ifdef ASM_OUTPUT_ADDR_VEC_START
8057 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
8060 #ifdef ASM_OUTPUT_CASE_LABEL
8061 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
8064 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
8067 for (idx = 0; idx < vlen; idx++)
8069 ASM_OUTPUT_ADDR_VEC_ELT
8070 (asm_out_file, CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 0, idx), 0)));
8073 #ifdef ASM_OUTPUT_ADDR_VEC_END
8074 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
8079 sparc_output_addr_diff_vec (rtx vec)
8081 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
8082 rtx base = XEXP (XEXP (body, 0), 0);
8083 int idx, vlen = XVECLEN (body, 1);
8085 #ifdef ASM_OUTPUT_ADDR_VEC_START
8086 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
8089 #ifdef ASM_OUTPUT_CASE_LABEL
8090 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
8093 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
8096 for (idx = 0; idx < vlen; idx++)
8098 ASM_OUTPUT_ADDR_DIFF_ELT
8101 CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 1, idx), 0)),
8102 CODE_LABEL_NUMBER (base));
8105 #ifdef ASM_OUTPUT_ADDR_VEC_END
8106 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
8111 sparc_output_deferred_case_vectors (void)
8116 if (sparc_addr_list == NULL_RTX
8117 && sparc_addr_diff_list == NULL_RTX)
8120 /* Align to cache line in the function's code section. */
8121 function_section (current_function_decl);
8123 align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
8125 ASM_OUTPUT_ALIGN (asm_out_file, align);
8127 for (t = sparc_addr_list; t ; t = XEXP (t, 1))
8128 sparc_output_addr_vec (XEXP (t, 0));
8129 for (t = sparc_addr_diff_list; t ; t = XEXP (t, 1))
8130 sparc_output_addr_diff_vec (XEXP (t, 0));
8132 sparc_addr_list = sparc_addr_diff_list = NULL_RTX;
8135 /* Return 0 if the high 32 bits of X (the low word of X, if DImode) are
8136 unknown. Return 1 if the high bits are zero, -1 if the register is
8139 sparc_check_64 (rtx x, rtx insn)
8141 /* If a register is set only once it is safe to ignore insns this
8142 code does not know how to handle. The loop will either recognize
8143 the single set and return the correct value or fail to recognize
8148 if (GET_CODE (x) != REG)
8151 if (GET_MODE (x) == DImode)
8152 y = gen_rtx_REG (SImode, REGNO (x) + WORDS_BIG_ENDIAN);
8154 if (flag_expensive_optimizations
8155 && REG_N_SETS (REGNO (y)) == 1)
8161 insn = get_last_insn_anywhere ();
8166 while ((insn = PREV_INSN (insn)))
8168 switch (GET_CODE (insn))
8181 rtx pat = PATTERN (insn);
8182 if (GET_CODE (pat) != SET)
8184 if (rtx_equal_p (x, SET_DEST (pat)))
8185 return set_extends (insn);
8186 if (y && rtx_equal_p (y, SET_DEST (pat)))
8187 return set_extends (insn);
8188 if (reg_overlap_mentioned_p (SET_DEST (pat), y))
8196 /* Returns assembly code to perform a DImode shift using
8197 a 64-bit global or out register on SPARC-V8+. */
8199 output_v8plus_shift (rtx *operands, rtx insn, const char *opcode)
8201 static char asm_code[60];
8203 /* The scratch register is only required when the destination
8204 register is not a 64-bit global or out register. */
8205 if (which_alternative != 2)
8206 operands[3] = operands[0];
8208 /* We can only shift by constants <= 63. */
8209 if (GET_CODE (operands[2]) == CONST_INT)
8210 operands[2] = GEN_INT (INTVAL (operands[2]) & 0x3f);
8212 if (GET_CODE (operands[1]) == CONST_INT)
8214 output_asm_insn ("mov\t%1, %3", operands);
8218 output_asm_insn ("sllx\t%H1, 32, %3", operands);
8219 if (sparc_check_64 (operands[1], insn) <= 0)
8220 output_asm_insn ("srl\t%L1, 0, %L1", operands);
8221 output_asm_insn ("or\t%L1, %3, %3", operands);
8224 strcpy(asm_code, opcode);
8226 if (which_alternative != 2)
8227 return strcat (asm_code, "\t%0, %2, %L0\n\tsrlx\t%L0, 32, %H0");
8229 return strcat (asm_code, "\t%3, %2, %3\n\tsrlx\t%3, 32, %H0\n\tmov\t%3, %L0");
8232 /* Output rtl to increment the profiler label LABELNO
8233 for profiling a function entry. */
8236 sparc_profile_hook (int labelno)
8241 ASM_GENERATE_INTERNAL_LABEL (buf, "LP", labelno);
8242 lab = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (buf));
8243 fun = gen_rtx_SYMBOL_REF (Pmode, MCOUNT_FUNCTION);
8245 emit_library_call (fun, LCT_NORMAL, VOIDmode, 1, lab, Pmode);
8248 #ifdef OBJECT_FORMAT_ELF
8250 sparc_elf_asm_named_section (const char *name, unsigned int flags,
8253 if (flags & SECTION_MERGE)
8255 /* entsize cannot be expressed in this section attributes
8257 default_elf_asm_named_section (name, flags, decl);
8261 fprintf (asm_out_file, "\t.section\t\"%s\"", name);
8263 if (!(flags & SECTION_DEBUG))
8264 fputs (",#alloc", asm_out_file);
8265 if (flags & SECTION_WRITE)
8266 fputs (",#write", asm_out_file);
8267 if (flags & SECTION_TLS)
8268 fputs (",#tls", asm_out_file);
8269 if (flags & SECTION_CODE)
8270 fputs (",#execinstr", asm_out_file);
8272 /* ??? Handle SECTION_BSS. */
8274 fputc ('\n', asm_out_file);
8276 #endif /* OBJECT_FORMAT_ELF */
8278 /* We do not allow indirect calls to be optimized into sibling calls.
8280 We cannot use sibling calls when delayed branches are disabled
8281 because they will likely require the call delay slot to be filled.
8283 Also, on SPARC 32-bit we cannot emit a sibling call when the
8284 current function returns a structure. This is because the "unimp
8285 after call" convention would cause the callee to return to the
8286 wrong place. The generic code already disallows cases where the
8287 function being called returns a structure.
8289 It may seem strange how this last case could occur. Usually there
8290 is code after the call which jumps to epilogue code which dumps the
8291 return value into the struct return area. That ought to invalidate
8292 the sibling call right? Well, in the C++ case we can end up passing
8293 the pointer to the struct return area to a constructor (which returns
8294 void) and then nothing else happens. Such a sibling call would look
8295 valid without the added check here. */
8297 sparc_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
8300 && flag_delayed_branch
8301 && (TARGET_ARCH64 || ! current_function_returns_struct));
8304 /* libfunc renaming. */
8305 #include "config/gofast.h"
8308 sparc_init_libfuncs (void)
8312 /* Use the subroutines that Sun's library provides for integer
8313 multiply and divide. The `*' prevents an underscore from
8314 being prepended by the compiler. .umul is a little faster
8316 set_optab_libfunc (smul_optab, SImode, "*.umul");
8317 set_optab_libfunc (sdiv_optab, SImode, "*.div");
8318 set_optab_libfunc (udiv_optab, SImode, "*.udiv");
8319 set_optab_libfunc (smod_optab, SImode, "*.rem");
8320 set_optab_libfunc (umod_optab, SImode, "*.urem");
8322 /* TFmode arithmetic. These names are part of the SPARC 32bit ABI. */
8323 set_optab_libfunc (add_optab, TFmode, "_Q_add");
8324 set_optab_libfunc (sub_optab, TFmode, "_Q_sub");
8325 set_optab_libfunc (neg_optab, TFmode, "_Q_neg");
8326 set_optab_libfunc (smul_optab, TFmode, "_Q_mul");
8327 set_optab_libfunc (sdiv_optab, TFmode, "_Q_div");
8329 /* We can define the TFmode sqrt optab only if TARGET_FPU. This
8330 is because with soft-float, the SFmode and DFmode sqrt
8331 instructions will be absent, and the compiler will notice and
8332 try to use the TFmode sqrt instruction for calls to the
8333 builtin function sqrt, but this fails. */
8335 set_optab_libfunc (sqrt_optab, TFmode, "_Q_sqrt");
8337 set_optab_libfunc (eq_optab, TFmode, "_Q_feq");
8338 set_optab_libfunc (ne_optab, TFmode, "_Q_fne");
8339 set_optab_libfunc (gt_optab, TFmode, "_Q_fgt");
8340 set_optab_libfunc (ge_optab, TFmode, "_Q_fge");
8341 set_optab_libfunc (lt_optab, TFmode, "_Q_flt");
8342 set_optab_libfunc (le_optab, TFmode, "_Q_fle");
8344 set_conv_libfunc (sext_optab, TFmode, SFmode, "_Q_stoq");
8345 set_conv_libfunc (sext_optab, TFmode, DFmode, "_Q_dtoq");
8346 set_conv_libfunc (trunc_optab, SFmode, TFmode, "_Q_qtos");
8347 set_conv_libfunc (trunc_optab, DFmode, TFmode, "_Q_qtod");
8349 set_conv_libfunc (sfix_optab, SImode, TFmode, "_Q_qtoi");
8350 set_conv_libfunc (ufix_optab, SImode, TFmode, "_Q_qtou");
8351 set_conv_libfunc (sfloat_optab, TFmode, SImode, "_Q_itoq");
8353 if (DITF_CONVERSION_LIBFUNCS)
8355 set_conv_libfunc (sfix_optab, DImode, TFmode, "_Q_qtoll");
8356 set_conv_libfunc (ufix_optab, DImode, TFmode, "_Q_qtoull");
8357 set_conv_libfunc (sfloat_optab, TFmode, DImode, "_Q_lltoq");
8360 if (SUN_CONVERSION_LIBFUNCS)
8362 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftoll");
8363 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoull");
8364 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtoll");
8365 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoull");
8370 /* In the SPARC 64bit ABI, SImode multiply and divide functions
8371 do not exist in the library. Make sure the compiler does not
8372 emit calls to them by accident. (It should always use the
8373 hardware instructions.) */
8374 set_optab_libfunc (smul_optab, SImode, 0);
8375 set_optab_libfunc (sdiv_optab, SImode, 0);
8376 set_optab_libfunc (udiv_optab, SImode, 0);
8377 set_optab_libfunc (smod_optab, SImode, 0);
8378 set_optab_libfunc (umod_optab, SImode, 0);
8380 if (SUN_INTEGER_MULTIPLY_64)
8382 set_optab_libfunc (smul_optab, DImode, "__mul64");
8383 set_optab_libfunc (sdiv_optab, DImode, "__div64");
8384 set_optab_libfunc (udiv_optab, DImode, "__udiv64");
8385 set_optab_libfunc (smod_optab, DImode, "__rem64");
8386 set_optab_libfunc (umod_optab, DImode, "__urem64");
8389 if (SUN_CONVERSION_LIBFUNCS)
8391 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftol");
8392 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoul");
8393 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtol");
8394 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoul");
8398 gofast_maybe_init_libfuncs ();
8402 sparc_extra_constraint_check (rtx op, int c, int strict)
8407 && (c == 'T' || c == 'U'))
8413 return fp_sethi_p (op);
8416 return fp_mov_p (op);
8419 return fp_high_losum_p (op);
8423 || (GET_CODE (op) == REG
8424 && (REGNO (op) < FIRST_PSEUDO_REGISTER
8425 || reg_renumber[REGNO (op)] >= 0)))
8426 return register_ok_for_ldd (op);
8438 /* Our memory extra constraints have to emulate the
8439 behavior of 'm' and 'o' in order for reload to work
8441 if (GET_CODE (op) == MEM)
8444 if ((TARGET_ARCH64 || mem_min_alignment (op, 8))
8446 || strict_memory_address_p (Pmode, XEXP (op, 0))))
8451 reload_ok_mem = (reload_in_progress
8452 && GET_CODE (op) == REG
8453 && REGNO (op) >= FIRST_PSEUDO_REGISTER
8454 && reg_renumber [REGNO (op)] < 0);
8457 return reload_ok_mem;
8460 /* ??? This duplicates information provided to the compiler by the
8461 ??? scheduler description. Some day, teach genautomata to output
8462 ??? the latencies and then CSE will just use that. */
8465 sparc_rtx_costs (rtx x, int code, int outer_code, int *total)
8467 enum machine_mode mode = GET_MODE (x);
8468 bool float_mode_p = FLOAT_MODE_P (mode);
8473 if (INTVAL (x) < 0x1000 && INTVAL (x) >= -0x1000)
8491 if (GET_MODE (x) == DImode
8492 && ((XINT (x, 3) == 0
8493 && (unsigned HOST_WIDE_INT) XINT (x, 2) < 0x1000)
8494 || (XINT (x, 3) == -1
8496 && XINT (x, 2) >= -0x1000)))
8503 /* If outer-code was a sign or zero extension, a cost
8504 of COSTS_N_INSNS (1) was already added in. This is
8505 why we are subtracting it back out. */
8506 if (outer_code == ZERO_EXTEND)
8508 *total = sparc_costs->int_zload - COSTS_N_INSNS (1);
8510 else if (outer_code == SIGN_EXTEND)
8512 *total = sparc_costs->int_sload - COSTS_N_INSNS (1);
8514 else if (float_mode_p)
8516 *total = sparc_costs->float_load;
8520 *total = sparc_costs->int_load;
8528 *total = sparc_costs->float_plusminus;
8530 *total = COSTS_N_INSNS (1);
8535 *total = sparc_costs->float_mul;
8536 else if (! TARGET_HARD_MUL)
8537 *total = COSTS_N_INSNS (25);
8543 if (sparc_costs->int_mul_bit_factor)
8547 if (GET_CODE (XEXP (x, 1)) == CONST_INT)
8549 unsigned HOST_WIDE_INT value = INTVAL (XEXP (x, 1));
8550 for (nbits = 0; value != 0; value &= value - 1)
8553 else if (GET_CODE (XEXP (x, 1)) == CONST_DOUBLE
8554 && GET_MODE (XEXP (x, 1)) == DImode)
8556 rtx x1 = XEXP (x, 1);
8557 unsigned HOST_WIDE_INT value1 = XINT (x1, 2);
8558 unsigned HOST_WIDE_INT value2 = XINT (x1, 3);
8560 for (nbits = 0; value1 != 0; value1 &= value1 - 1)
8562 for (; value2 != 0; value2 &= value2 - 1)
8570 bit_cost = (nbits - 3) / sparc_costs->int_mul_bit_factor;
8571 bit_cost = COSTS_N_INSNS (bit_cost);
8575 *total = sparc_costs->int_mulX + bit_cost;
8577 *total = sparc_costs->int_mul + bit_cost;
8584 *total = COSTS_N_INSNS (1) + sparc_costs->shift_penalty;
8594 *total = sparc_costs->float_div_df;
8596 *total = sparc_costs->float_div_sf;
8601 *total = sparc_costs->int_divX;
8603 *total = sparc_costs->int_div;
8610 *total = COSTS_N_INSNS (1);
8617 case UNSIGNED_FLOAT:
8621 case FLOAT_TRUNCATE:
8622 *total = sparc_costs->float_move;
8627 *total = sparc_costs->float_sqrt_df;
8629 *total = sparc_costs->float_sqrt_sf;
8634 *total = sparc_costs->float_cmp;
8636 *total = COSTS_N_INSNS (1);
8641 *total = sparc_costs->float_cmove;
8643 *total = sparc_costs->int_cmove;
8651 /* Emit the sequence of insns SEQ while preserving the register REG. */
8654 emit_and_preserve (rtx seq, rtx reg)
8656 rtx slot = gen_rtx_MEM (word_mode,
8657 plus_constant (stack_pointer_rtx, SPARC_STACK_BIAS));
8659 emit_stack_pointer_decrement (GEN_INT (STACK_BOUNDARY/BITS_PER_UNIT));
8660 emit_insn (gen_rtx_SET (VOIDmode, slot, reg));
8662 emit_insn (gen_rtx_SET (VOIDmode, reg, slot));
8663 emit_stack_pointer_increment (GEN_INT (STACK_BOUNDARY/BITS_PER_UNIT));
8666 /* Output the assembler code for a thunk function. THUNK_DECL is the
8667 declaration for the thunk function itself, FUNCTION is the decl for
8668 the target function. DELTA is an immediate constant offset to be
8669 added to THIS. If VCALL_OFFSET is nonzero, the word at address
8670 (*THIS + VCALL_OFFSET) should be additionally added to THIS. */
8673 sparc_output_mi_thunk (FILE *file, tree thunk_fndecl ATTRIBUTE_UNUSED,
8674 HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset,
8677 rtx this, insn, funexp;
8678 unsigned int int_arg_first;
8680 reload_completed = 1;
8681 epilogue_completed = 1;
8683 reset_block_changes ();
8685 emit_note (NOTE_INSN_PROLOGUE_END);
8687 if (flag_delayed_branch)
8689 /* We will emit a regular sibcall below, so we need to instruct
8690 output_sibcall that we are in a leaf function. */
8691 sparc_leaf_function_p = current_function_uses_only_leaf_regs = 1;
8693 /* This will cause final.c to invoke leaf_renumber_regs so we
8694 must behave as if we were in a not-yet-leafified function. */
8695 int_arg_first = SPARC_INCOMING_INT_ARG_FIRST;
8699 /* We will emit the sibcall manually below, so we will need to
8700 manually spill non-leaf registers. */
8701 sparc_leaf_function_p = current_function_uses_only_leaf_regs = 0;
8703 /* We really are in a leaf function. */
8704 int_arg_first = SPARC_OUTGOING_INT_ARG_FIRST;
8707 /* Find the "this" pointer. Normally in %o0, but in ARCH64 if the function
8708 returns a structure, the structure return pointer is there instead. */
8709 if (TARGET_ARCH64 && aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function))
8710 this = gen_rtx_REG (Pmode, int_arg_first + 1);
8712 this = gen_rtx_REG (Pmode, int_arg_first);
8714 /* Add DELTA. When possible use a plain add, otherwise load it into
8715 a register first. */
8718 rtx delta_rtx = GEN_INT (delta);
8720 if (! SPARC_SIMM13_P (delta))
8722 rtx scratch = gen_rtx_REG (Pmode, 1);
8723 emit_move_insn (scratch, delta_rtx);
8724 delta_rtx = scratch;
8727 /* THIS += DELTA. */
8728 emit_insn (gen_add2_insn (this, delta_rtx));
8731 /* Add the word at address (*THIS + VCALL_OFFSET). */
8734 rtx vcall_offset_rtx = GEN_INT (vcall_offset);
8735 rtx scratch = gen_rtx_REG (Pmode, 1);
8737 if (vcall_offset >= 0)
8740 /* SCRATCH = *THIS. */
8741 emit_move_insn (scratch, gen_rtx_MEM (Pmode, this));
8743 /* Prepare for adding VCALL_OFFSET. The difficulty is that we
8744 may not have any available scratch register at this point. */
8745 if (SPARC_SIMM13_P (vcall_offset))
8747 /* This is the case if ARCH64 (unless -ffixed-g5 is passed). */
8748 else if (! fixed_regs[5]
8749 /* The below sequence is made up of at least 2 insns,
8750 while the default method may need only one. */
8751 && vcall_offset < -8192)
8753 rtx scratch2 = gen_rtx_REG (Pmode, 5);
8754 emit_move_insn (scratch2, vcall_offset_rtx);
8755 vcall_offset_rtx = scratch2;
8759 rtx increment = GEN_INT (-4096);
8761 /* VCALL_OFFSET is a negative number whose typical range can be
8762 estimated as -32768..0 in 32-bit mode. In almost all cases
8763 it is therefore cheaper to emit multiple add insns than
8764 spilling and loading the constant into a register (at least
8766 while (! SPARC_SIMM13_P (vcall_offset))
8768 emit_insn (gen_add2_insn (scratch, increment));
8769 vcall_offset += 4096;
8771 vcall_offset_rtx = GEN_INT (vcall_offset); /* cannot be 0 */
8774 /* SCRATCH = *(*THIS + VCALL_OFFSET). */
8775 emit_move_insn (scratch, gen_rtx_MEM (Pmode,
8776 gen_rtx_PLUS (Pmode,
8778 vcall_offset_rtx)));
8780 /* THIS += *(*THIS + VCALL_OFFSET). */
8781 emit_insn (gen_add2_insn (this, scratch));
8784 /* Generate a tail call to the target function. */
8785 if (! TREE_USED (function))
8787 assemble_external (function);
8788 TREE_USED (function) = 1;
8790 funexp = XEXP (DECL_RTL (function), 0);
8792 if (flag_delayed_branch)
8794 funexp = gen_rtx_MEM (FUNCTION_MODE, funexp);
8795 insn = emit_call_insn (gen_sibcall (funexp));
8796 SIBLING_CALL_P (insn) = 1;
8800 /* The hoops we have to jump through in order to generate a sibcall
8801 without using delay slots... */
8802 rtx spill_reg, seq, scratch = gen_rtx_REG (Pmode, 1);
8806 spill_reg = gen_rtx_REG (word_mode, 15); /* %o7 */
8808 load_pic_register (); /* clobbers %o7 */
8809 scratch = legitimize_pic_address (funexp, Pmode, scratch);
8812 emit_and_preserve (seq, spill_reg);
8814 else if (TARGET_ARCH32)
8816 emit_insn (gen_rtx_SET (VOIDmode,
8818 gen_rtx_HIGH (SImode, funexp)));
8819 emit_insn (gen_rtx_SET (VOIDmode,
8821 gen_rtx_LO_SUM (SImode, scratch, funexp)));
8823 else /* TARGET_ARCH64 */
8825 switch (sparc_cmodel)
8829 /* The destination can serve as a temporary. */
8830 sparc_emit_set_symbolic_const64 (scratch, funexp, scratch);
8835 /* The destination cannot serve as a temporary. */
8836 spill_reg = gen_rtx_REG (DImode, 15); /* %o7 */
8838 sparc_emit_set_symbolic_const64 (scratch, funexp, spill_reg);
8841 emit_and_preserve (seq, spill_reg);
8849 emit_jump_insn (gen_indirect_jump (scratch));
8854 /* Run just enough of rest_of_compilation to get the insns emitted.
8855 There's not really enough bulk here to make other passes such as
8856 instruction scheduling worth while. Note that use_thunk calls
8857 assemble_start_function and assemble_end_function. */
8858 insn = get_insns ();
8859 insn_locators_initialize ();
8860 shorten_branches (insn);
8861 final_start_function (insn, file, 1);
8862 final (insn, file, 1, 0);
8863 final_end_function ();
8865 reload_completed = 0;
8866 epilogue_completed = 0;
8870 /* Return true if sparc_output_mi_thunk would be able to output the
8871 assembler code for the thunk function specified by the arguments
8872 it is passed, and false otherwise. */
8874 sparc_can_output_mi_thunk (tree thunk_fndecl ATTRIBUTE_UNUSED,
8875 HOST_WIDE_INT delta ATTRIBUTE_UNUSED,
8876 HOST_WIDE_INT vcall_offset,
8877 tree function ATTRIBUTE_UNUSED)
8879 /* Bound the loop used in the default method above. */
8880 return (vcall_offset >= -32768 || ! fixed_regs[5]);
8883 /* How to allocate a 'struct machine_function'. */
8885 static struct machine_function *
8886 sparc_init_machine_status (void)
8888 return ggc_alloc_cleared (sizeof (struct machine_function));
8891 /* Locate some local-dynamic symbol still in use by this function
8892 so that we can print its name in local-dynamic base patterns. */
8895 get_some_local_dynamic_name (void)
8899 if (cfun->machine->some_ld_name)
8900 return cfun->machine->some_ld_name;
8902 for (insn = get_insns (); insn ; insn = NEXT_INSN (insn))
8904 && for_each_rtx (&PATTERN (insn), get_some_local_dynamic_name_1, 0))
8905 return cfun->machine->some_ld_name;
8911 get_some_local_dynamic_name_1 (rtx *px, void *data ATTRIBUTE_UNUSED)
8916 && GET_CODE (x) == SYMBOL_REF
8917 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_DYNAMIC)
8919 cfun->machine->some_ld_name = XSTR (x, 0);
8926 /* This is called from dwarf2out.c via ASM_OUTPUT_DWARF_DTPREL.
8927 We need to emit DTP-relative relocations. */
8930 sparc_output_dwarf_dtprel (FILE *file, int size, rtx x)
8935 fputs ("\t.word\t%r_tls_dtpoff32(", file);
8938 fputs ("\t.xword\t%r_tls_dtpoff64(", file);
8943 output_addr_const (file, x);
8947 #include "gt-sparc.h"