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_vector_mode_supported_p (enum machine_mode);
352 static bool sparc_pass_by_reference (CUMULATIVE_ARGS *,
353 enum machine_mode, tree, bool);
354 #ifdef SUBTARGET_ATTRIBUTE_TABLE
355 const struct attribute_spec sparc_attribute_table[];
358 /* Option handling. */
360 /* Code model option as passed by user. */
361 const char *sparc_cmodel_string;
363 enum cmodel sparc_cmodel;
365 char sparc_hard_reg_printed[8];
367 struct sparc_cpu_select sparc_select[] =
369 /* switch name, tune arch */
370 { (char *)0, "default", 1, 1 },
371 { (char *)0, "-mcpu=", 1, 1 },
372 { (char *)0, "-mtune=", 1, 0 },
376 /* CPU type. This is set from TARGET_CPU_DEFAULT and -m{cpu,tune}=xxx. */
377 enum processor_type sparc_cpu;
379 /* Initialize the GCC target structure. */
381 /* The sparc default is to use .half rather than .short for aligned
382 HI objects. Use .word instead of .long on non-ELF systems. */
383 #undef TARGET_ASM_ALIGNED_HI_OP
384 #define TARGET_ASM_ALIGNED_HI_OP "\t.half\t"
385 #ifndef OBJECT_FORMAT_ELF
386 #undef TARGET_ASM_ALIGNED_SI_OP
387 #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t"
390 #undef TARGET_ASM_UNALIGNED_HI_OP
391 #define TARGET_ASM_UNALIGNED_HI_OP "\t.uahalf\t"
392 #undef TARGET_ASM_UNALIGNED_SI_OP
393 #define TARGET_ASM_UNALIGNED_SI_OP "\t.uaword\t"
394 #undef TARGET_ASM_UNALIGNED_DI_OP
395 #define TARGET_ASM_UNALIGNED_DI_OP "\t.uaxword\t"
397 /* The target hook has to handle DI-mode values. */
398 #undef TARGET_ASM_INTEGER
399 #define TARGET_ASM_INTEGER sparc_assemble_integer
401 #undef TARGET_ASM_FUNCTION_PROLOGUE
402 #define TARGET_ASM_FUNCTION_PROLOGUE sparc_asm_function_prologue
403 #undef TARGET_ASM_FUNCTION_EPILOGUE
404 #define TARGET_ASM_FUNCTION_EPILOGUE sparc_asm_function_epilogue
406 #undef TARGET_SCHED_ADJUST_COST
407 #define TARGET_SCHED_ADJUST_COST sparc_adjust_cost
408 #undef TARGET_SCHED_ISSUE_RATE
409 #define TARGET_SCHED_ISSUE_RATE sparc_issue_rate
410 #undef TARGET_SCHED_INIT
411 #define TARGET_SCHED_INIT sparc_sched_init
412 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
413 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD sparc_use_sched_lookahead
415 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
416 #define TARGET_FUNCTION_OK_FOR_SIBCALL sparc_function_ok_for_sibcall
418 #undef TARGET_INIT_LIBFUNCS
419 #define TARGET_INIT_LIBFUNCS sparc_init_libfuncs
422 #undef TARGET_HAVE_TLS
423 #define TARGET_HAVE_TLS true
425 #undef TARGET_CANNOT_FORCE_CONST_MEM
426 #define TARGET_CANNOT_FORCE_CONST_MEM sparc_cannot_force_const_mem
428 #undef TARGET_ASM_OUTPUT_MI_THUNK
429 #define TARGET_ASM_OUTPUT_MI_THUNK sparc_output_mi_thunk
430 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
431 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK sparc_can_output_mi_thunk
433 #undef TARGET_RTX_COSTS
434 #define TARGET_RTX_COSTS sparc_rtx_costs
435 #undef TARGET_ADDRESS_COST
436 #define TARGET_ADDRESS_COST hook_int_rtx_0
438 /* This is only needed for TARGET_ARCH64, but since PROMOTE_FUNCTION_MODE is a
439 no-op for TARGET_ARCH32 this is ok. Otherwise we'd need to add a runtime
440 test for this value. */
441 #undef TARGET_PROMOTE_FUNCTION_ARGS
442 #define TARGET_PROMOTE_FUNCTION_ARGS hook_bool_tree_true
444 /* This is only needed for TARGET_ARCH64, but since PROMOTE_FUNCTION_MODE is a
445 no-op for TARGET_ARCH32 this is ok. Otherwise we'd need to add a runtime
446 test for this value. */
447 #undef TARGET_PROMOTE_FUNCTION_RETURN
448 #define TARGET_PROMOTE_FUNCTION_RETURN hook_bool_tree_true
450 #undef TARGET_PROMOTE_PROTOTYPES
451 #define TARGET_PROMOTE_PROTOTYPES sparc_promote_prototypes
453 #undef TARGET_STRUCT_VALUE_RTX
454 #define TARGET_STRUCT_VALUE_RTX sparc_struct_value_rtx
455 #undef TARGET_RETURN_IN_MEMORY
456 #define TARGET_RETURN_IN_MEMORY sparc_return_in_memory
457 #undef TARGET_MUST_PASS_IN_STACK
458 #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size
459 #undef TARGET_PASS_BY_REFERENCE
460 #define TARGET_PASS_BY_REFERENCE sparc_pass_by_reference
462 #undef TARGET_EXPAND_BUILTIN_SAVEREGS
463 #define TARGET_EXPAND_BUILTIN_SAVEREGS sparc_builtin_saveregs
464 #undef TARGET_STRICT_ARGUMENT_NAMING
465 #define TARGET_STRICT_ARGUMENT_NAMING sparc_strict_argument_naming
467 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
468 #define TARGET_GIMPLIFY_VA_ARG_EXPR sparc_gimplify_va_arg
470 #undef TARGET_VECTOR_MODE_SUPPORTED_P
471 #define TARGET_VECTOR_MODE_SUPPORTED_P sparc_vector_mode_supported_p
473 #ifdef SUBTARGET_INSERT_ATTRIBUTES
474 #undef TARGET_INSERT_ATTRIBUTES
475 #define TARGET_INSERT_ATTRIBUTES SUBTARGET_INSERT_ATTRIBUTES
478 #ifdef SUBTARGET_ATTRIBUTE_TABLE
479 #undef TARGET_ATTRIBUTE_TABLE
480 #define TARGET_ATTRIBUTE_TABLE sparc_attribute_table
483 struct gcc_target targetm = TARGET_INITIALIZER;
485 /* Validate and override various options, and do some machine dependent
489 sparc_override_options (void)
491 static struct code_model {
492 const char *const name;
494 } const cmodels[] = {
496 { "medlow", CM_MEDLOW },
497 { "medmid", CM_MEDMID },
498 { "medany", CM_MEDANY },
499 { "embmedany", CM_EMBMEDANY },
502 const struct code_model *cmodel;
503 /* Map TARGET_CPU_DEFAULT to value for -m{arch,tune}=. */
504 static struct cpu_default {
506 const char *const name;
507 } const cpu_default[] = {
508 /* There must be one entry here for each TARGET_CPU value. */
509 { TARGET_CPU_sparc, "cypress" },
510 { TARGET_CPU_sparclet, "tsc701" },
511 { TARGET_CPU_sparclite, "f930" },
512 { TARGET_CPU_v8, "v8" },
513 { TARGET_CPU_hypersparc, "hypersparc" },
514 { TARGET_CPU_sparclite86x, "sparclite86x" },
515 { TARGET_CPU_supersparc, "supersparc" },
516 { TARGET_CPU_v9, "v9" },
517 { TARGET_CPU_ultrasparc, "ultrasparc" },
518 { TARGET_CPU_ultrasparc3, "ultrasparc3" },
521 const struct cpu_default *def;
522 /* Table of values for -m{cpu,tune}=. */
523 static struct cpu_table {
524 const char *const name;
525 const enum processor_type processor;
528 } const cpu_table[] = {
529 { "v7", PROCESSOR_V7, MASK_ISA, 0 },
530 { "cypress", PROCESSOR_CYPRESS, MASK_ISA, 0 },
531 { "v8", PROCESSOR_V8, MASK_ISA, MASK_V8 },
532 /* TI TMS390Z55 supersparc */
533 { "supersparc", PROCESSOR_SUPERSPARC, MASK_ISA, MASK_V8 },
534 { "sparclite", PROCESSOR_SPARCLITE, MASK_ISA, MASK_SPARCLITE },
535 /* The Fujitsu MB86930 is the original sparclite chip, with no fpu.
536 The Fujitsu MB86934 is the recent sparclite chip, with an fpu. */
537 { "f930", PROCESSOR_F930, MASK_ISA|MASK_FPU, MASK_SPARCLITE },
538 { "f934", PROCESSOR_F934, MASK_ISA, MASK_SPARCLITE|MASK_FPU },
539 { "hypersparc", PROCESSOR_HYPERSPARC, MASK_ISA, MASK_V8|MASK_FPU },
540 { "sparclite86x", PROCESSOR_SPARCLITE86X, MASK_ISA|MASK_FPU,
542 { "sparclet", PROCESSOR_SPARCLET, MASK_ISA, MASK_SPARCLET },
544 { "tsc701", PROCESSOR_TSC701, MASK_ISA, MASK_SPARCLET },
545 { "v9", PROCESSOR_V9, MASK_ISA, MASK_V9 },
546 /* TI ultrasparc I, II, IIi */
547 { "ultrasparc", PROCESSOR_ULTRASPARC, MASK_ISA, MASK_V9
548 /* Although insns using %y are deprecated, it is a clear win on current
550 |MASK_DEPRECATED_V8_INSNS},
551 /* TI ultrasparc III */
552 /* ??? Check if %y issue still holds true in ultra3. */
553 { "ultrasparc3", PROCESSOR_ULTRASPARC3, MASK_ISA, MASK_V9|MASK_DEPRECATED_V8_INSNS},
556 const struct cpu_table *cpu;
557 const struct sparc_cpu_select *sel;
560 #ifndef SPARC_BI_ARCH
561 /* Check for unsupported architecture size. */
562 if (! TARGET_64BIT != DEFAULT_ARCH32_P)
563 error ("%s is not supported by this configuration",
564 DEFAULT_ARCH32_P ? "-m64" : "-m32");
567 /* We force all 64bit archs to use 128 bit long double */
568 if (TARGET_64BIT && ! TARGET_LONG_DOUBLE_128)
570 error ("-mlong-double-64 not allowed with -m64");
571 target_flags |= MASK_LONG_DOUBLE_128;
574 /* Code model selection. */
575 sparc_cmodel = SPARC_DEFAULT_CMODEL;
579 sparc_cmodel = CM_32;
582 if (sparc_cmodel_string != NULL)
586 for (cmodel = &cmodels[0]; cmodel->name; cmodel++)
587 if (strcmp (sparc_cmodel_string, cmodel->name) == 0)
589 if (cmodel->name == NULL)
590 error ("bad value (%s) for -mcmodel= switch", sparc_cmodel_string);
592 sparc_cmodel = cmodel->value;
595 error ("-mcmodel= is not supported on 32 bit systems");
598 fpu = TARGET_FPU; /* save current -mfpu status */
600 /* Set the default CPU. */
601 for (def = &cpu_default[0]; def->name; ++def)
602 if (def->cpu == TARGET_CPU_DEFAULT)
606 sparc_select[0].string = def->name;
608 for (sel = &sparc_select[0]; sel->name; ++sel)
612 for (cpu = &cpu_table[0]; cpu->name; ++cpu)
613 if (! strcmp (sel->string, cpu->name))
616 sparc_cpu = cpu->processor;
620 target_flags &= ~cpu->disable;
621 target_flags |= cpu->enable;
627 error ("bad value (%s) for %s switch", sel->string, sel->name);
631 /* If -mfpu or -mno-fpu was explicitly used, don't override with
632 the processor default. Clear MASK_FPU_SET to avoid confusing
633 the reverse mapping from switch values to names. */
636 target_flags = (target_flags & ~MASK_FPU) | fpu;
637 target_flags &= ~MASK_FPU_SET;
640 /* Don't allow -mvis if FPU is disabled. */
642 target_flags &= ~MASK_VIS;
644 /* -mvis assumes UltraSPARC+, so we are sure v9 instructions
646 -m64 also implies v9. */
647 if (TARGET_VIS || TARGET_ARCH64)
649 target_flags |= MASK_V9;
650 target_flags &= ~(MASK_V8 | MASK_SPARCLET | MASK_SPARCLITE);
653 /* Use the deprecated v8 insns for sparc64 in 32 bit mode. */
654 if (TARGET_V9 && TARGET_ARCH32)
655 target_flags |= MASK_DEPRECATED_V8_INSNS;
657 /* V8PLUS requires V9, makes no sense in 64 bit mode. */
658 if (! TARGET_V9 || TARGET_ARCH64)
659 target_flags &= ~MASK_V8PLUS;
661 /* Don't use stack biasing in 32 bit mode. */
663 target_flags &= ~MASK_STACK_BIAS;
665 /* Supply a default value for align_functions. */
666 if (align_functions == 0
667 && (sparc_cpu == PROCESSOR_ULTRASPARC
668 || sparc_cpu == PROCESSOR_ULTRASPARC3))
669 align_functions = 32;
671 /* Validate PCC_STRUCT_RETURN. */
672 if (flag_pcc_struct_return == DEFAULT_PCC_STRUCT_RETURN)
673 flag_pcc_struct_return = (TARGET_ARCH64 ? 0 : 1);
675 /* Only use .uaxword when compiling for a 64-bit target. */
677 targetm.asm_out.unaligned_op.di = NULL;
679 /* Do various machine dependent initializations. */
682 /* Acquire a unique set number for our register saves and restores. */
683 sparc_sr_alias_set = new_alias_set ();
685 /* Set up function hooks. */
686 init_machine_status = sparc_init_machine_status;
691 case PROCESSOR_CYPRESS:
692 sparc_costs = &cypress_costs;
695 case PROCESSOR_SPARCLITE:
696 case PROCESSOR_SUPERSPARC:
697 sparc_costs = &supersparc_costs;
701 case PROCESSOR_HYPERSPARC:
702 case PROCESSOR_SPARCLITE86X:
703 sparc_costs = &hypersparc_costs;
705 case PROCESSOR_SPARCLET:
706 case PROCESSOR_TSC701:
707 sparc_costs = &sparclet_costs;
710 case PROCESSOR_ULTRASPARC:
711 sparc_costs = &ultrasparc_costs;
713 case PROCESSOR_ULTRASPARC3:
714 sparc_costs = &ultrasparc3_costs;
719 #ifdef SUBTARGET_ATTRIBUTE_TABLE
720 /* Table of valid machine attributes. */
721 const struct attribute_spec sparc_attribute_table[] =
723 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
724 SUBTARGET_ATTRIBUTE_TABLE,
725 { NULL, 0, 0, false, false, false, NULL }
729 /* Miscellaneous utilities. */
731 /* Nonzero if CODE, a comparison, is suitable for use in v9 conditional move
732 or branch on register contents instructions. */
735 v9_regcmp_p (enum rtx_code code)
737 return (code == EQ || code == NE || code == GE || code == LT
738 || code == LE || code == GT);
742 /* Operand constraints. */
744 /* Return nonzero only if OP is a register of mode MODE,
748 reg_or_0_operand (rtx op, enum machine_mode mode)
750 if (register_operand (op, mode))
752 if (op == const0_rtx)
754 if (GET_MODE (op) == VOIDmode && GET_CODE (op) == CONST_DOUBLE
755 && CONST_DOUBLE_HIGH (op) == 0
756 && CONST_DOUBLE_LOW (op) == 0)
758 if (fp_zero_operand (op, mode))
763 /* Return nonzero only if OP is const1_rtx. */
766 const1_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
768 return op == const1_rtx;
771 /* Nonzero if OP is a floating point value with value 0.0. */
774 fp_zero_operand (rtx op, enum machine_mode mode)
776 enum mode_class mclass = GET_MODE_CLASS (GET_MODE (op));
777 if (mclass != MODE_FLOAT && mclass != MODE_VECTOR_INT)
779 return op == CONST0_RTX (mode);
782 /* Nonzero if OP is a register operand in floating point register. */
785 fp_register_operand (rtx op, enum machine_mode mode)
787 if (! register_operand (op, mode))
789 if (GET_CODE (op) == SUBREG)
790 op = SUBREG_REG (op);
791 return GET_CODE (op) == REG && SPARC_FP_REG_P (REGNO (op));
794 /* Nonzero if OP is a floating point constant which can
795 be loaded into an integer register using a single
796 sethi instruction. */
801 if (GET_CODE (op) == CONST_DOUBLE)
806 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
807 if (REAL_VALUES_EQUAL (r, dconst0) &&
808 ! REAL_VALUE_MINUS_ZERO (r))
810 REAL_VALUE_TO_TARGET_SINGLE (r, i);
811 if (SPARC_SETHI_P (i))
818 /* Nonzero if OP is a floating point constant which can
819 be loaded into an integer register using a single
825 if (GET_CODE (op) == CONST_DOUBLE)
830 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
831 if (REAL_VALUES_EQUAL (r, dconst0) &&
832 ! REAL_VALUE_MINUS_ZERO (r))
834 REAL_VALUE_TO_TARGET_SINGLE (r, i);
835 if (SPARC_SIMM13_P (i))
842 /* Nonzero if OP is a floating point constant which can
843 be loaded into an integer register using a high/losum
844 instruction sequence. */
847 fp_high_losum_p (rtx op)
849 /* The constraints calling this should only be in
850 SFmode move insns, so any constant which cannot
851 be moved using a single insn will do. */
852 if (GET_CODE (op) == CONST_DOUBLE)
857 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
858 if (REAL_VALUES_EQUAL (r, dconst0) &&
859 ! REAL_VALUE_MINUS_ZERO (r))
861 REAL_VALUE_TO_TARGET_SINGLE (r, i);
862 if (! SPARC_SETHI_P (i)
863 && ! SPARC_SIMM13_P (i))
870 /* Nonzero if OP is an integer register. */
873 intreg_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
875 return (register_operand (op, SImode)
876 || (TARGET_ARCH64 && register_operand (op, DImode)));
879 /* Nonzero if OP is a floating point condition code register. */
882 fcc_reg_operand (rtx op, enum machine_mode mode)
884 /* This can happen when recog is called from combine. Op may be a MEM.
885 Fail instead of calling abort in this case. */
886 if (GET_CODE (op) != REG)
889 if (mode != VOIDmode && mode != GET_MODE (op))
892 && (GET_MODE (op) != CCFPmode && GET_MODE (op) != CCFPEmode))
895 #if 0 /* ??? ==> 1 when %fcc0-3 are pseudos first. See gen_compare_reg(). */
896 if (reg_renumber == 0)
897 return REGNO (op) >= FIRST_PSEUDO_REGISTER;
898 return REGNO_OK_FOR_CCFP_P (REGNO (op));
900 return (unsigned) REGNO (op) - SPARC_FIRST_V9_FCC_REG < 4;
904 /* Nonzero if OP is a floating point condition code fcc0 register. */
907 fcc0_reg_operand (rtx op, enum machine_mode mode)
909 /* This can happen when recog is called from combine. Op may be a MEM.
910 Fail instead of calling abort in this case. */
911 if (GET_CODE (op) != REG)
914 if (mode != VOIDmode && mode != GET_MODE (op))
917 && (GET_MODE (op) != CCFPmode && GET_MODE (op) != CCFPEmode))
920 return REGNO (op) == SPARC_FCC_REG;
923 /* Nonzero if OP is an integer or floating point condition code register. */
926 icc_or_fcc_reg_operand (rtx op, enum machine_mode mode)
928 if (GET_CODE (op) == REG && REGNO (op) == SPARC_ICC_REG)
930 if (mode != VOIDmode && mode != GET_MODE (op))
933 && GET_MODE (op) != CCmode && GET_MODE (op) != CCXmode)
938 return fcc_reg_operand (op, mode);
941 /* Call insn on SPARC can take a PC-relative constant address, or any regular
945 call_operand (rtx op, enum machine_mode mode)
947 if (GET_CODE (op) != MEM)
950 return (symbolic_operand (op, mode) || memory_address_p (Pmode, op));
954 call_operand_address (rtx op, enum machine_mode mode)
956 return (symbolic_operand (op, mode) || memory_address_p (Pmode, op));
959 /* If OP is a SYMBOL_REF of a thread-local symbol, return its TLS mode,
960 otherwise return 0. */
963 tls_symbolic_operand (rtx op)
965 if (GET_CODE (op) != SYMBOL_REF)
967 return SYMBOL_REF_TLS_MODEL (op);
971 tgd_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
973 return tls_symbolic_operand (op) == TLS_MODEL_GLOBAL_DYNAMIC;
977 tld_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
979 return tls_symbolic_operand (op) == TLS_MODEL_LOCAL_DYNAMIC;
983 tie_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
985 return tls_symbolic_operand (op) == TLS_MODEL_INITIAL_EXEC;
989 tle_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
991 return tls_symbolic_operand (op) == TLS_MODEL_LOCAL_EXEC;
994 /* Returns 1 if OP is either a symbol reference or a sum of a symbol
995 reference and a constant. */
998 symbolic_operand (register rtx op, enum machine_mode mode)
1000 enum machine_mode omode = GET_MODE (op);
1002 if (omode != mode && omode != VOIDmode && mode != VOIDmode)
1005 switch (GET_CODE (op))
1008 return !SYMBOL_REF_TLS_MODEL (op);
1015 return (((GET_CODE (XEXP (op, 0)) == SYMBOL_REF
1016 && !SYMBOL_REF_TLS_MODEL (XEXP (op, 0)))
1017 || GET_CODE (XEXP (op, 0)) == LABEL_REF)
1018 && GET_CODE (XEXP (op, 1)) == CONST_INT);
1025 /* Return truth value of statement that OP is a symbolic memory
1026 operand of mode MODE. */
1029 symbolic_memory_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1031 if (GET_CODE (op) == SUBREG)
1032 op = SUBREG_REG (op);
1033 if (GET_CODE (op) != MEM)
1036 return ((GET_CODE (op) == SYMBOL_REF && !SYMBOL_REF_TLS_MODEL (op))
1037 || GET_CODE (op) == CONST || GET_CODE (op) == HIGH
1038 || GET_CODE (op) == LABEL_REF);
1041 /* Return truth value of statement that OP is a LABEL_REF of mode MODE. */
1044 label_ref_operand (rtx op, enum machine_mode mode)
1046 if (GET_CODE (op) != LABEL_REF)
1048 if (GET_MODE (op) != mode)
1053 /* Return 1 if the operand is an argument used in generating pic references
1054 in either the medium/low or medium/anywhere code models of sparc64. */
1057 sp64_medium_pic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1059 /* Check for (const (minus (symbol_ref:GOT)
1060 (const (minus (label) (pc))))). */
1061 if (GET_CODE (op) != CONST)
1064 if (GET_CODE (op) != MINUS)
1066 if (GET_CODE (XEXP (op, 0)) != SYMBOL_REF)
1068 /* ??? Ensure symbol is GOT. */
1069 if (GET_CODE (XEXP (op, 1)) != CONST)
1071 if (GET_CODE (XEXP (XEXP (op, 1), 0)) != MINUS)
1076 /* Return 1 if the operand is a data segment reference. This includes
1077 the readonly data segment, or in other words anything but the text segment.
1078 This is needed in the medium/anywhere code model on v9. These values
1079 are accessed with EMBMEDANY_BASE_REG. */
1082 data_segment_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1084 switch (GET_CODE (op))
1087 return ! SYMBOL_REF_FUNCTION_P (op);
1089 /* Assume canonical format of symbol + constant.
1092 return data_segment_operand (XEXP (op, 0), VOIDmode);
1098 /* Return 1 if the operand is a text segment reference.
1099 This is needed in the medium/anywhere code model on v9. */
1102 text_segment_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1104 switch (GET_CODE (op))
1109 return SYMBOL_REF_FUNCTION_P (op);
1111 /* Assume canonical format of symbol + constant.
1114 return text_segment_operand (XEXP (op, 0), VOIDmode);
1120 /* Return 1 if the operand is either a register or a memory operand that is
1124 reg_or_nonsymb_mem_operand (register rtx op, enum machine_mode mode)
1126 if (register_operand (op, mode))
1129 if (memory_operand (op, mode) && ! symbolic_memory_operand (op, mode))
1136 splittable_symbolic_memory_operand (rtx op,
1137 enum machine_mode mode ATTRIBUTE_UNUSED)
1139 if (GET_CODE (op) != MEM)
1141 if (! symbolic_operand (XEXP (op, 0), Pmode))
1147 splittable_immediate_memory_operand (rtx op,
1148 enum machine_mode mode ATTRIBUTE_UNUSED)
1150 if (GET_CODE (op) != MEM)
1152 if (! immediate_operand (XEXP (op, 0), Pmode))
1157 /* Return truth value of whether OP is EQ or NE. */
1160 eq_or_neq (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1162 return (GET_CODE (op) == EQ || GET_CODE (op) == NE);
1165 /* Return 1 if this is a comparison operator, but not an EQ, NE, GEU,
1166 or LTU for non-floating-point. We handle those specially. */
1169 normal_comp_operator (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1173 if (!COMPARISON_P (op))
1176 if (GET_MODE (XEXP (op, 0)) == CCFPmode
1177 || GET_MODE (XEXP (op, 0)) == CCFPEmode)
1180 code = GET_CODE (op);
1181 return (code != NE && code != EQ && code != GEU && code != LTU);
1184 /* Return 1 if this is a comparison operator. This allows the use of
1185 MATCH_OPERATOR to recognize all the branch insns. */
1188 noov_compare_op (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1192 if (!COMPARISON_P (op))
1195 code = GET_CODE (op);
1196 if (GET_MODE (XEXP (op, 0)) == CC_NOOVmode
1197 || GET_MODE (XEXP (op, 0)) == CCX_NOOVmode)
1198 /* These are the only branches which work with CC_NOOVmode. */
1199 return (code == EQ || code == NE || code == GE || code == LT);
1203 /* Return 1 if this is a 64-bit comparison operator. This allows the use of
1204 MATCH_OPERATOR to recognize all the branch insns. */
1207 noov_compare64_op (register rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1214 if (!COMPARISON_P (op))
1217 code = GET_CODE (op);
1218 if (GET_MODE (XEXP (op, 0)) == CCX_NOOVmode)
1219 /* These are the only branches which work with CCX_NOOVmode. */
1220 return (code == EQ || code == NE || code == GE || code == LT);
1221 return (GET_MODE (XEXP (op, 0)) == CCXmode);
1224 /* Nonzero if OP is a comparison operator suitable for use in v9
1225 conditional move or branch on register contents instructions. */
1228 v9_regcmp_op (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1232 if (!COMPARISON_P (op))
1235 code = GET_CODE (op);
1236 return v9_regcmp_p (code);
1239 /* Return 1 if this is a SIGN_EXTEND or ZERO_EXTEND operation. */
1242 extend_op (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1244 return GET_CODE (op) == SIGN_EXTEND || GET_CODE (op) == ZERO_EXTEND;
1247 /* Return nonzero if OP is an operator of mode MODE which can set
1248 the condition codes explicitly. We do not include PLUS and MINUS
1249 because these require CC_NOOVmode, which we handle explicitly. */
1252 cc_arithop (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1254 if (GET_CODE (op) == AND
1255 || GET_CODE (op) == IOR
1256 || GET_CODE (op) == XOR)
1262 /* Return nonzero if OP is an operator of mode MODE which can bitwise
1263 complement its second operand and set the condition codes explicitly. */
1266 cc_arithopn (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1268 /* XOR is not here because combine canonicalizes (xor (not ...) ...)
1269 and (xor ... (not ...)) to (not (xor ...)). */
1270 return (GET_CODE (op) == AND
1271 || GET_CODE (op) == IOR);
1274 /* Return true if OP is a register, or is a CONST_INT that can fit in a
1275 signed 13 bit immediate field. This is an acceptable SImode operand for
1276 most 3 address instructions. */
1279 arith_operand (rtx op, enum machine_mode mode)
1281 if (register_operand (op, mode))
1283 if (GET_CODE (op) != CONST_INT)
1285 return SMALL_INT32 (op);
1288 /* Return true if OP is a constant 4096 */
1291 arith_4096_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1293 if (GET_CODE (op) != CONST_INT)
1296 return INTVAL (op) == 4096;
1299 /* Return true if OP is suitable as second operand for add/sub */
1302 arith_add_operand (rtx op, enum machine_mode mode)
1304 return arith_operand (op, mode) || arith_4096_operand (op, mode);
1307 /* Return true if OP is a CONST_INT or a CONST_DOUBLE which can fit in the
1308 immediate field of OR and XOR instructions. Used for 64-bit
1309 constant formation patterns. */
1311 const64_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1313 return ((GET_CODE (op) == CONST_INT
1314 && SPARC_SIMM13_P (INTVAL (op)))
1315 #if HOST_BITS_PER_WIDE_INT != 64
1316 || (GET_CODE (op) == CONST_DOUBLE
1317 && SPARC_SIMM13_P (CONST_DOUBLE_LOW (op))
1318 && (CONST_DOUBLE_HIGH (op) ==
1319 ((CONST_DOUBLE_LOW (op) & 0x80000000) != 0 ?
1320 (HOST_WIDE_INT)-1 : 0)))
1325 /* The same, but only for sethi instructions. */
1327 const64_high_operand (rtx op, enum machine_mode mode)
1329 return ((GET_CODE (op) == CONST_INT
1330 && (INTVAL (op) & ~(HOST_WIDE_INT)0x3ff) != 0
1331 && SPARC_SETHI_P (INTVAL (op) & GET_MODE_MASK (mode))
1333 || (GET_CODE (op) == CONST_DOUBLE
1334 && CONST_DOUBLE_HIGH (op) == 0
1335 && (CONST_DOUBLE_LOW (op) & ~(HOST_WIDE_INT)0x3ff) != 0
1336 && SPARC_SETHI_P (CONST_DOUBLE_LOW (op))));
1339 /* Return true if OP is a register, or is a CONST_INT that can fit in a
1340 signed 11 bit immediate field. This is an acceptable SImode operand for
1341 the movcc instructions. */
1344 arith11_operand (rtx op, enum machine_mode mode)
1346 return (register_operand (op, mode)
1347 || (GET_CODE (op) == CONST_INT && SPARC_SIMM11_P (INTVAL (op))));
1350 /* Return true if OP is a register, or is a CONST_INT that can fit in a
1351 signed 10 bit immediate field. This is an acceptable SImode operand for
1352 the movrcc instructions. */
1355 arith10_operand (rtx op, enum machine_mode mode)
1357 return (register_operand (op, mode)
1358 || (GET_CODE (op) == CONST_INT && SPARC_SIMM10_P (INTVAL (op))));
1361 /* Return true if OP is a register, is a CONST_INT that fits in a 13 bit
1362 immediate field, or is a CONST_DOUBLE whose both parts fit in a 13 bit
1364 ARCH64: Return true if OP is a register, or is a CONST_INT or CONST_DOUBLE that
1365 can fit in a 13 bit immediate field. This is an acceptable DImode operand
1366 for most 3 address instructions. */
1369 arith_double_operand (rtx op, enum machine_mode mode)
1371 return (register_operand (op, mode)
1372 || (GET_CODE (op) == CONST_INT && SMALL_INT (op))
1374 && GET_CODE (op) == CONST_DOUBLE
1375 && (unsigned HOST_WIDE_INT) (CONST_DOUBLE_LOW (op) + 0x1000) < 0x2000
1376 && (unsigned HOST_WIDE_INT) (CONST_DOUBLE_HIGH (op) + 0x1000) < 0x2000)
1378 && GET_CODE (op) == CONST_DOUBLE
1379 && (unsigned HOST_WIDE_INT) (CONST_DOUBLE_LOW (op) + 0x1000) < 0x2000
1380 && ((CONST_DOUBLE_HIGH (op) == -1
1381 && (CONST_DOUBLE_LOW (op) & 0x1000) == 0x1000)
1382 || (CONST_DOUBLE_HIGH (op) == 0
1383 && (CONST_DOUBLE_LOW (op) & 0x1000) == 0))));
1386 /* Return true if OP is a constant 4096 for DImode on ARCH64 */
1389 arith_double_4096_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1391 return (TARGET_ARCH64 &&
1392 ((GET_CODE (op) == CONST_INT && INTVAL (op) == 4096) ||
1393 (GET_CODE (op) == CONST_DOUBLE &&
1394 CONST_DOUBLE_LOW (op) == 4096 &&
1395 CONST_DOUBLE_HIGH (op) == 0)));
1398 /* Return true if OP is suitable as second operand for add/sub in DImode */
1401 arith_double_add_operand (rtx op, enum machine_mode mode)
1403 return arith_double_operand (op, mode) || arith_double_4096_operand (op, mode);
1406 /* Return true if OP is a register, or is a CONST_INT or CONST_DOUBLE that
1407 can fit in an 11 bit immediate field. This is an acceptable DImode
1408 operand for the movcc instructions. */
1409 /* ??? Replace with arith11_operand? */
1412 arith11_double_operand (rtx op, enum machine_mode mode)
1414 return (register_operand (op, mode)
1415 || (GET_CODE (op) == CONST_DOUBLE
1416 && (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode)
1417 && (unsigned HOST_WIDE_INT) (CONST_DOUBLE_LOW (op) + 0x400) < 0x800
1418 && ((CONST_DOUBLE_HIGH (op) == -1
1419 && (CONST_DOUBLE_LOW (op) & 0x400) == 0x400)
1420 || (CONST_DOUBLE_HIGH (op) == 0
1421 && (CONST_DOUBLE_LOW (op) & 0x400) == 0)))
1422 || (GET_CODE (op) == CONST_INT
1423 && (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode)
1424 && (unsigned HOST_WIDE_INT) (INTVAL (op) + 0x400) < 0x800));
1427 /* Return true if OP is a register, or is a CONST_INT or CONST_DOUBLE that
1428 can fit in an 10 bit immediate field. This is an acceptable DImode
1429 operand for the movrcc instructions. */
1430 /* ??? Replace with arith10_operand? */
1433 arith10_double_operand (rtx op, enum machine_mode mode)
1435 return (register_operand (op, mode)
1436 || (GET_CODE (op) == CONST_DOUBLE
1437 && (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode)
1438 && (unsigned) (CONST_DOUBLE_LOW (op) + 0x200) < 0x400
1439 && ((CONST_DOUBLE_HIGH (op) == -1
1440 && (CONST_DOUBLE_LOW (op) & 0x200) == 0x200)
1441 || (CONST_DOUBLE_HIGH (op) == 0
1442 && (CONST_DOUBLE_LOW (op) & 0x200) == 0)))
1443 || (GET_CODE (op) == CONST_INT
1444 && (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode)
1445 && (unsigned HOST_WIDE_INT) (INTVAL (op) + 0x200) < 0x400));
1448 /* Return truth value of whether OP is an integer which fits the
1449 range constraining immediate operands in most three-address insns,
1450 which have a 13 bit immediate field. */
1453 small_int (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1455 return (GET_CODE (op) == CONST_INT && SMALL_INT (op));
1459 small_int_or_double (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1461 return ((GET_CODE (op) == CONST_INT && SMALL_INT (op))
1462 || (GET_CODE (op) == CONST_DOUBLE
1463 && CONST_DOUBLE_HIGH (op) == 0
1464 && SPARC_SIMM13_P (CONST_DOUBLE_LOW (op))));
1467 /* Recognize operand values for the umul instruction. That instruction sign
1468 extends immediate values just like all other sparc instructions, but
1469 interprets the extended result as an unsigned number. */
1472 uns_small_int (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1474 #if HOST_BITS_PER_WIDE_INT > 32
1475 /* All allowed constants will fit a CONST_INT. */
1476 return (GET_CODE (op) == CONST_INT
1477 && ((INTVAL (op) >= 0 && INTVAL (op) < 0x1000)
1478 || (INTVAL (op) >= 0xFFFFF000
1479 && INTVAL (op) <= 0xFFFFFFFF)));
1481 return ((GET_CODE (op) == CONST_INT && (unsigned) INTVAL (op) < 0x1000)
1482 || (GET_CODE (op) == CONST_DOUBLE
1483 && CONST_DOUBLE_HIGH (op) == 0
1484 && (unsigned) CONST_DOUBLE_LOW (op) - 0xFFFFF000 < 0x1000));
1489 uns_arith_operand (rtx op, enum machine_mode mode)
1491 return register_operand (op, mode) || uns_small_int (op, mode);
1494 /* Return truth value of statement that OP is a call-clobbered register. */
1496 clobbered_register (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1498 return (GET_CODE (op) == REG && call_used_regs[REGNO (op)]);
1501 /* Return 1 if OP is a valid operand for the source of a move insn. */
1504 input_operand (rtx op, enum machine_mode mode)
1506 enum mode_class mclass;
1508 /* If both modes are non-void they must be the same. */
1509 if (mode != VOIDmode && GET_MODE (op) != VOIDmode && mode != GET_MODE (op))
1512 /* Allow any one instruction integer constant, and all CONST_INT
1513 variants when we are working in DImode and !arch64. */
1514 if (GET_MODE_CLASS (mode) == MODE_INT
1515 && ((GET_CODE (op) == CONST_INT
1516 && (SPARC_SETHI_P (INTVAL (op) & GET_MODE_MASK (mode))
1517 || SPARC_SIMM13_P (INTVAL (op))
1519 && ! TARGET_ARCH64)))
1521 && GET_CODE (op) == CONST_DOUBLE
1522 && ((CONST_DOUBLE_HIGH (op) == 0
1523 && SPARC_SETHI_P (CONST_DOUBLE_LOW (op)))
1525 #if HOST_BITS_PER_WIDE_INT == 64
1526 (CONST_DOUBLE_HIGH (op) == 0
1527 && SPARC_SIMM13_P (CONST_DOUBLE_LOW (op)))
1529 (SPARC_SIMM13_P (CONST_DOUBLE_LOW (op))
1530 && (((CONST_DOUBLE_LOW (op) & 0x80000000) == 0
1531 && CONST_DOUBLE_HIGH (op) == 0)
1532 || (CONST_DOUBLE_HIGH (op) == -1
1533 && CONST_DOUBLE_LOW (op) & 0x80000000) != 0))
1538 /* If !arch64 and this is a DImode const, allow it so that
1539 the splits can be generated. */
1542 && GET_CODE (op) == CONST_DOUBLE)
1545 if (register_operand (op, mode))
1548 mclass = GET_MODE_CLASS (mode);
1549 if ((mclass == MODE_FLOAT && GET_CODE (op) == CONST_DOUBLE)
1550 || (mclass == MODE_VECTOR_INT && GET_CODE (op) == CONST_VECTOR))
1553 /* If this is a SUBREG, look inside so that we handle
1554 paradoxical ones. */
1555 if (GET_CODE (op) == SUBREG)
1556 op = SUBREG_REG (op);
1558 /* Check for valid MEM forms. */
1559 if (GET_CODE (op) == MEM)
1560 return memory_address_p (mode, XEXP (op, 0));
1565 /* Return 1 if OP is valid for the lhs of a compare insn. */
1568 compare_operand (rtx op, enum machine_mode mode)
1570 if (GET_CODE (op) == ZERO_EXTRACT)
1571 return (register_operand (XEXP (op, 0), mode)
1572 && small_int_or_double (XEXP (op, 1), mode)
1573 && small_int_or_double (XEXP (op, 2), mode)
1574 /* This matches cmp_zero_extract. */
1576 && ((GET_CODE (XEXP (op, 2)) == CONST_INT
1577 && INTVAL (XEXP (op, 2)) > 19)
1578 || (GET_CODE (XEXP (op, 2)) == CONST_DOUBLE
1579 && CONST_DOUBLE_LOW (XEXP (op, 2)) > 19)))
1580 /* This matches cmp_zero_extract_sp64. */
1583 && ((GET_CODE (XEXP (op, 2)) == CONST_INT
1584 && INTVAL (XEXP (op, 2)) > 51)
1585 || (GET_CODE (XEXP (op, 2)) == CONST_DOUBLE
1586 && CONST_DOUBLE_LOW (XEXP (op, 2)) > 51)))));
1588 return register_operand (op, mode);
1592 /* We know it can't be done in one insn when we get here,
1593 the movsi expander guarantees this. */
1595 sparc_emit_set_const32 (rtx op0, rtx op1)
1597 enum machine_mode mode = GET_MODE (op0);
1600 if (GET_CODE (op1) == CONST_INT)
1602 HOST_WIDE_INT value = INTVAL (op1);
1604 if (SPARC_SETHI_P (value & GET_MODE_MASK (mode))
1605 || SPARC_SIMM13_P (value))
1609 /* Full 2-insn decomposition is needed. */
1610 if (reload_in_progress || reload_completed)
1613 temp = gen_reg_rtx (mode);
1615 if (GET_CODE (op1) == CONST_INT)
1617 /* Emit them as real moves instead of a HIGH/LO_SUM,
1618 this way CSE can see everything and reuse intermediate
1619 values if it wants. */
1621 && HOST_BITS_PER_WIDE_INT != 64
1622 && (INTVAL (op1) & 0x80000000) != 0)
1623 emit_insn (gen_rtx_SET
1625 immed_double_const (INTVAL (op1) & ~(HOST_WIDE_INT)0x3ff,
1628 emit_insn (gen_rtx_SET (VOIDmode, temp,
1629 GEN_INT (INTVAL (op1)
1630 & ~(HOST_WIDE_INT)0x3ff)));
1632 emit_insn (gen_rtx_SET (VOIDmode,
1634 gen_rtx_IOR (mode, temp,
1635 GEN_INT (INTVAL (op1) & 0x3ff))));
1639 /* A symbol, emit in the traditional way. */
1640 emit_insn (gen_rtx_SET (VOIDmode, temp,
1641 gen_rtx_HIGH (mode, op1)));
1642 emit_insn (gen_rtx_SET (VOIDmode,
1643 op0, gen_rtx_LO_SUM (mode, temp, op1)));
1649 /* Load OP1, a symbolic 64-bit constant, into OP0, a DImode register.
1650 If TEMP is nonzero, we are forbidden to use any other scratch
1651 registers. Otherwise, we are allowed to generate them as needed.
1653 Note that TEMP may have TImode if the code model is TARGET_CM_MEDANY
1654 or TARGET_CM_EMBMEDANY (see the reload_indi and reload_outdi patterns). */
1656 sparc_emit_set_symbolic_const64 (rtx op0, rtx op1, rtx temp)
1658 rtx temp1, temp2, temp3, temp4, temp5;
1661 if (temp && GET_MODE (temp) == TImode)
1664 temp = gen_rtx_REG (DImode, REGNO (temp));
1667 /* SPARC-V9 code-model support. */
1668 switch (sparc_cmodel)
1671 /* The range spanned by all instructions in the object is less
1672 than 2^31 bytes (2GB) and the distance from any instruction
1673 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1674 than 2^31 bytes (2GB).
1676 The executable must be in the low 4TB of the virtual address
1679 sethi %hi(symbol), %temp1
1680 or %temp1, %lo(symbol), %reg */
1682 temp1 = temp; /* op0 is allowed. */
1684 temp1 = gen_reg_rtx (DImode);
1686 emit_insn (gen_rtx_SET (VOIDmode, temp1, gen_rtx_HIGH (DImode, op1)));
1687 emit_insn (gen_rtx_SET (VOIDmode, op0, gen_rtx_LO_SUM (DImode, temp1, op1)));
1691 /* The range spanned by all instructions in the object is less
1692 than 2^31 bytes (2GB) and the distance from any instruction
1693 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1694 than 2^31 bytes (2GB).
1696 The executable must be in the low 16TB of the virtual address
1699 sethi %h44(symbol), %temp1
1700 or %temp1, %m44(symbol), %temp2
1701 sllx %temp2, 12, %temp3
1702 or %temp3, %l44(symbol), %reg */
1707 temp3 = temp; /* op0 is allowed. */
1711 temp1 = gen_reg_rtx (DImode);
1712 temp2 = gen_reg_rtx (DImode);
1713 temp3 = gen_reg_rtx (DImode);
1716 emit_insn (gen_seth44 (temp1, op1));
1717 emit_insn (gen_setm44 (temp2, temp1, op1));
1718 emit_insn (gen_rtx_SET (VOIDmode, temp3,
1719 gen_rtx_ASHIFT (DImode, temp2, GEN_INT (12))));
1720 emit_insn (gen_setl44 (op0, temp3, op1));
1724 /* The range spanned by all instructions in the object is less
1725 than 2^31 bytes (2GB) and the distance from any instruction
1726 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1727 than 2^31 bytes (2GB).
1729 The executable can be placed anywhere in the virtual address
1732 sethi %hh(symbol), %temp1
1733 sethi %lm(symbol), %temp2
1734 or %temp1, %hm(symbol), %temp3
1735 sllx %temp3, 32, %temp4
1736 or %temp4, %temp2, %temp5
1737 or %temp5, %lo(symbol), %reg */
1740 /* It is possible that one of the registers we got for operands[2]
1741 might coincide with that of operands[0] (which is why we made
1742 it TImode). Pick the other one to use as our scratch. */
1743 if (rtx_equal_p (temp, op0))
1746 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1751 temp2 = temp; /* op0 is _not_ allowed, see above. */
1758 temp1 = gen_reg_rtx (DImode);
1759 temp2 = gen_reg_rtx (DImode);
1760 temp3 = gen_reg_rtx (DImode);
1761 temp4 = gen_reg_rtx (DImode);
1762 temp5 = gen_reg_rtx (DImode);
1765 emit_insn (gen_sethh (temp1, op1));
1766 emit_insn (gen_setlm (temp2, op1));
1767 emit_insn (gen_sethm (temp3, temp1, op1));
1768 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1769 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1770 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1771 gen_rtx_PLUS (DImode, temp4, temp2)));
1772 emit_insn (gen_setlo (op0, temp5, op1));
1776 /* Old old old backwards compatibility kruft here.
1777 Essentially it is MEDLOW with a fixed 64-bit
1778 virtual base added to all data segment addresses.
1779 Text-segment stuff is computed like MEDANY, we can't
1780 reuse the code above because the relocation knobs
1783 Data segment: sethi %hi(symbol), %temp1
1784 add %temp1, EMBMEDANY_BASE_REG, %temp2
1785 or %temp2, %lo(symbol), %reg */
1786 if (data_segment_operand (op1, GET_MODE (op1)))
1790 temp1 = temp; /* op0 is allowed. */
1795 temp1 = gen_reg_rtx (DImode);
1796 temp2 = gen_reg_rtx (DImode);
1799 emit_insn (gen_embmedany_sethi (temp1, op1));
1800 emit_insn (gen_embmedany_brsum (temp2, temp1));
1801 emit_insn (gen_embmedany_losum (op0, temp2, op1));
1804 /* Text segment: sethi %uhi(symbol), %temp1
1805 sethi %hi(symbol), %temp2
1806 or %temp1, %ulo(symbol), %temp3
1807 sllx %temp3, 32, %temp4
1808 or %temp4, %temp2, %temp5
1809 or %temp5, %lo(symbol), %reg */
1814 /* It is possible that one of the registers we got for operands[2]
1815 might coincide with that of operands[0] (which is why we made
1816 it TImode). Pick the other one to use as our scratch. */
1817 if (rtx_equal_p (temp, op0))
1820 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1825 temp2 = temp; /* op0 is _not_ allowed, see above. */
1832 temp1 = gen_reg_rtx (DImode);
1833 temp2 = gen_reg_rtx (DImode);
1834 temp3 = gen_reg_rtx (DImode);
1835 temp4 = gen_reg_rtx (DImode);
1836 temp5 = gen_reg_rtx (DImode);
1839 emit_insn (gen_embmedany_textuhi (temp1, op1));
1840 emit_insn (gen_embmedany_texthi (temp2, op1));
1841 emit_insn (gen_embmedany_textulo (temp3, temp1, op1));
1842 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1843 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1844 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1845 gen_rtx_PLUS (DImode, temp4, temp2)));
1846 emit_insn (gen_embmedany_textlo (op0, temp5, op1));
1855 /* These avoid problems when cross compiling. If we do not
1856 go through all this hair then the optimizer will see
1857 invalid REG_EQUAL notes or in some cases none at all. */
1858 static void sparc_emit_set_safe_HIGH64 (rtx, HOST_WIDE_INT);
1859 static rtx gen_safe_SET64 (rtx, HOST_WIDE_INT);
1860 static rtx gen_safe_OR64 (rtx, HOST_WIDE_INT);
1861 static rtx gen_safe_XOR64 (rtx, HOST_WIDE_INT);
1863 #if HOST_BITS_PER_WIDE_INT == 64
1864 #define GEN_HIGHINT64(__x) GEN_INT ((__x) & ~(HOST_WIDE_INT)0x3ff)
1865 #define GEN_INT64(__x) GEN_INT (__x)
1867 #define GEN_HIGHINT64(__x) \
1868 immed_double_const ((__x) & ~(HOST_WIDE_INT)0x3ff, 0, DImode)
1869 #define GEN_INT64(__x) \
1870 immed_double_const ((__x) & 0xffffffff, \
1871 ((__x) & 0x80000000 ? -1 : 0), DImode)
1874 /* The optimizer is not to assume anything about exactly
1875 which bits are set for a HIGH, they are unspecified.
1876 Unfortunately this leads to many missed optimizations
1877 during CSE. We mask out the non-HIGH bits, and matches
1878 a plain movdi, to alleviate this problem. */
1880 sparc_emit_set_safe_HIGH64 (rtx dest, HOST_WIDE_INT val)
1882 emit_insn (gen_rtx_SET (VOIDmode, dest, GEN_HIGHINT64 (val)));
1886 gen_safe_SET64 (rtx dest, HOST_WIDE_INT val)
1888 return gen_rtx_SET (VOIDmode, dest, GEN_INT64 (val));
1892 gen_safe_OR64 (rtx src, HOST_WIDE_INT val)
1894 return gen_rtx_IOR (DImode, src, GEN_INT64 (val));
1898 gen_safe_XOR64 (rtx src, HOST_WIDE_INT val)
1900 return gen_rtx_XOR (DImode, src, GEN_INT64 (val));
1903 /* Worker routines for 64-bit constant formation on arch64.
1904 One of the key things to be doing in these emissions is
1905 to create as many temp REGs as possible. This makes it
1906 possible for half-built constants to be used later when
1907 such values are similar to something required later on.
1908 Without doing this, the optimizer cannot see such
1911 static void sparc_emit_set_const64_quick1 (rtx, rtx,
1912 unsigned HOST_WIDE_INT, int);
1915 sparc_emit_set_const64_quick1 (rtx op0, rtx temp,
1916 unsigned HOST_WIDE_INT low_bits, int is_neg)
1918 unsigned HOST_WIDE_INT high_bits;
1921 high_bits = (~low_bits) & 0xffffffff;
1923 high_bits = low_bits;
1925 sparc_emit_set_safe_HIGH64 (temp, high_bits);
1928 emit_insn (gen_rtx_SET (VOIDmode, op0,
1929 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1933 /* If we are XOR'ing with -1, then we should emit a one's complement
1934 instead. This way the combiner will notice logical operations
1935 such as ANDN later on and substitute. */
1936 if ((low_bits & 0x3ff) == 0x3ff)
1938 emit_insn (gen_rtx_SET (VOIDmode, op0,
1939 gen_rtx_NOT (DImode, temp)));
1943 emit_insn (gen_rtx_SET (VOIDmode, op0,
1944 gen_safe_XOR64 (temp,
1945 (-(HOST_WIDE_INT)0x400
1946 | (low_bits & 0x3ff)))));
1951 static void sparc_emit_set_const64_quick2 (rtx, rtx, unsigned HOST_WIDE_INT,
1952 unsigned HOST_WIDE_INT, int);
1955 sparc_emit_set_const64_quick2 (rtx op0, rtx temp,
1956 unsigned HOST_WIDE_INT high_bits,
1957 unsigned HOST_WIDE_INT low_immediate,
1962 if ((high_bits & 0xfffffc00) != 0)
1964 sparc_emit_set_safe_HIGH64 (temp, high_bits);
1965 if ((high_bits & ~0xfffffc00) != 0)
1966 emit_insn (gen_rtx_SET (VOIDmode, op0,
1967 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1973 emit_insn (gen_safe_SET64 (temp, high_bits));
1977 /* Now shift it up into place. */
1978 emit_insn (gen_rtx_SET (VOIDmode, op0,
1979 gen_rtx_ASHIFT (DImode, temp2,
1980 GEN_INT (shift_count))));
1982 /* If there is a low immediate part piece, finish up by
1983 putting that in as well. */
1984 if (low_immediate != 0)
1985 emit_insn (gen_rtx_SET (VOIDmode, op0,
1986 gen_safe_OR64 (op0, low_immediate)));
1989 static void sparc_emit_set_const64_longway (rtx, rtx, unsigned HOST_WIDE_INT,
1990 unsigned HOST_WIDE_INT);
1992 /* Full 64-bit constant decomposition. Even though this is the
1993 'worst' case, we still optimize a few things away. */
1995 sparc_emit_set_const64_longway (rtx op0, rtx temp,
1996 unsigned HOST_WIDE_INT high_bits,
1997 unsigned HOST_WIDE_INT low_bits)
2001 if (reload_in_progress || reload_completed)
2004 sub_temp = gen_reg_rtx (DImode);
2006 if ((high_bits & 0xfffffc00) != 0)
2008 sparc_emit_set_safe_HIGH64 (temp, high_bits);
2009 if ((high_bits & ~0xfffffc00) != 0)
2010 emit_insn (gen_rtx_SET (VOIDmode,
2012 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
2018 emit_insn (gen_safe_SET64 (temp, high_bits));
2022 if (!reload_in_progress && !reload_completed)
2024 rtx temp2 = gen_reg_rtx (DImode);
2025 rtx temp3 = gen_reg_rtx (DImode);
2026 rtx temp4 = gen_reg_rtx (DImode);
2028 emit_insn (gen_rtx_SET (VOIDmode, temp4,
2029 gen_rtx_ASHIFT (DImode, sub_temp,
2032 sparc_emit_set_safe_HIGH64 (temp2, low_bits);
2033 if ((low_bits & ~0xfffffc00) != 0)
2035 emit_insn (gen_rtx_SET (VOIDmode, temp3,
2036 gen_safe_OR64 (temp2, (low_bits & 0x3ff))));
2037 emit_insn (gen_rtx_SET (VOIDmode, op0,
2038 gen_rtx_PLUS (DImode, temp4, temp3)));
2042 emit_insn (gen_rtx_SET (VOIDmode, op0,
2043 gen_rtx_PLUS (DImode, temp4, temp2)));
2048 rtx low1 = GEN_INT ((low_bits >> (32 - 12)) & 0xfff);
2049 rtx low2 = GEN_INT ((low_bits >> (32 - 12 - 12)) & 0xfff);
2050 rtx low3 = GEN_INT ((low_bits >> (32 - 12 - 12 - 8)) & 0x0ff);
2053 /* We are in the middle of reload, so this is really
2054 painful. However we do still make an attempt to
2055 avoid emitting truly stupid code. */
2056 if (low1 != const0_rtx)
2058 emit_insn (gen_rtx_SET (VOIDmode, op0,
2059 gen_rtx_ASHIFT (DImode, sub_temp,
2060 GEN_INT (to_shift))));
2061 emit_insn (gen_rtx_SET (VOIDmode, op0,
2062 gen_rtx_IOR (DImode, op0, low1)));
2070 if (low2 != const0_rtx)
2072 emit_insn (gen_rtx_SET (VOIDmode, op0,
2073 gen_rtx_ASHIFT (DImode, sub_temp,
2074 GEN_INT (to_shift))));
2075 emit_insn (gen_rtx_SET (VOIDmode, op0,
2076 gen_rtx_IOR (DImode, op0, low2)));
2084 emit_insn (gen_rtx_SET (VOIDmode, op0,
2085 gen_rtx_ASHIFT (DImode, sub_temp,
2086 GEN_INT (to_shift))));
2087 if (low3 != const0_rtx)
2088 emit_insn (gen_rtx_SET (VOIDmode, op0,
2089 gen_rtx_IOR (DImode, op0, low3)));
2094 /* Analyze a 64-bit constant for certain properties. */
2095 static void analyze_64bit_constant (unsigned HOST_WIDE_INT,
2096 unsigned HOST_WIDE_INT,
2097 int *, int *, int *);
2100 analyze_64bit_constant (unsigned HOST_WIDE_INT high_bits,
2101 unsigned HOST_WIDE_INT low_bits,
2102 int *hbsp, int *lbsp, int *abbasp)
2104 int lowest_bit_set, highest_bit_set, all_bits_between_are_set;
2107 lowest_bit_set = highest_bit_set = -1;
2111 if ((lowest_bit_set == -1)
2112 && ((low_bits >> i) & 1))
2114 if ((highest_bit_set == -1)
2115 && ((high_bits >> (32 - i - 1)) & 1))
2116 highest_bit_set = (64 - i - 1);
2119 && ((highest_bit_set == -1)
2120 || (lowest_bit_set == -1)));
2126 if ((lowest_bit_set == -1)
2127 && ((high_bits >> i) & 1))
2128 lowest_bit_set = i + 32;
2129 if ((highest_bit_set == -1)
2130 && ((low_bits >> (32 - i - 1)) & 1))
2131 highest_bit_set = 32 - i - 1;
2134 && ((highest_bit_set == -1)
2135 || (lowest_bit_set == -1)));
2137 /* If there are no bits set this should have gone out
2138 as one instruction! */
2139 if (lowest_bit_set == -1
2140 || highest_bit_set == -1)
2142 all_bits_between_are_set = 1;
2143 for (i = lowest_bit_set; i <= highest_bit_set; i++)
2147 if ((low_bits & (1 << i)) != 0)
2152 if ((high_bits & (1 << (i - 32))) != 0)
2155 all_bits_between_are_set = 0;
2158 *hbsp = highest_bit_set;
2159 *lbsp = lowest_bit_set;
2160 *abbasp = all_bits_between_are_set;
2163 static int const64_is_2insns (unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT);
2166 const64_is_2insns (unsigned HOST_WIDE_INT high_bits,
2167 unsigned HOST_WIDE_INT low_bits)
2169 int highest_bit_set, lowest_bit_set, all_bits_between_are_set;
2172 || high_bits == 0xffffffff)
2175 analyze_64bit_constant (high_bits, low_bits,
2176 &highest_bit_set, &lowest_bit_set,
2177 &all_bits_between_are_set);
2179 if ((highest_bit_set == 63
2180 || lowest_bit_set == 0)
2181 && all_bits_between_are_set != 0)
2184 if ((highest_bit_set - lowest_bit_set) < 21)
2190 static unsigned HOST_WIDE_INT create_simple_focus_bits (unsigned HOST_WIDE_INT,
2191 unsigned HOST_WIDE_INT,
2194 static unsigned HOST_WIDE_INT
2195 create_simple_focus_bits (unsigned HOST_WIDE_INT high_bits,
2196 unsigned HOST_WIDE_INT low_bits,
2197 int lowest_bit_set, int shift)
2199 HOST_WIDE_INT hi, lo;
2201 if (lowest_bit_set < 32)
2203 lo = (low_bits >> lowest_bit_set) << shift;
2204 hi = ((high_bits << (32 - lowest_bit_set)) << shift);
2209 hi = ((high_bits >> (lowest_bit_set - 32)) << shift);
2216 /* Here we are sure to be arch64 and this is an integer constant
2217 being loaded into a register. Emit the most efficient
2218 insn sequence possible. Detection of all the 1-insn cases
2219 has been done already. */
2221 sparc_emit_set_const64 (rtx op0, rtx op1)
2223 unsigned HOST_WIDE_INT high_bits, low_bits;
2224 int lowest_bit_set, highest_bit_set;
2225 int all_bits_between_are_set;
2228 /* Sanity check that we know what we are working with. */
2229 if (! TARGET_ARCH64)
2232 if (GET_CODE (op0) != SUBREG)
2234 if (GET_CODE (op0) != REG
2235 || (REGNO (op0) >= SPARC_FIRST_FP_REG
2236 && REGNO (op0) <= SPARC_LAST_V9_FP_REG))
2240 if (reload_in_progress || reload_completed)
2243 if (GET_CODE (op1) != CONST_DOUBLE
2244 && GET_CODE (op1) != CONST_INT)
2246 sparc_emit_set_symbolic_const64 (op0, op1, temp);
2251 temp = gen_reg_rtx (DImode);
2253 if (GET_CODE (op1) == CONST_DOUBLE)
2255 #if HOST_BITS_PER_WIDE_INT == 64
2256 high_bits = (CONST_DOUBLE_LOW (op1) >> 32) & 0xffffffff;
2257 low_bits = CONST_DOUBLE_LOW (op1) & 0xffffffff;
2259 high_bits = CONST_DOUBLE_HIGH (op1);
2260 low_bits = CONST_DOUBLE_LOW (op1);
2265 #if HOST_BITS_PER_WIDE_INT == 64
2266 high_bits = ((INTVAL (op1) >> 32) & 0xffffffff);
2267 low_bits = (INTVAL (op1) & 0xffffffff);
2269 high_bits = ((INTVAL (op1) < 0) ?
2272 low_bits = INTVAL (op1);
2276 /* low_bits bits 0 --> 31
2277 high_bits bits 32 --> 63 */
2279 analyze_64bit_constant (high_bits, low_bits,
2280 &highest_bit_set, &lowest_bit_set,
2281 &all_bits_between_are_set);
2283 /* First try for a 2-insn sequence. */
2285 /* These situations are preferred because the optimizer can
2286 * do more things with them:
2288 * sllx %reg, shift, %reg
2290 * srlx %reg, shift, %reg
2291 * 3) mov some_small_const, %reg
2292 * sllx %reg, shift, %reg
2294 if (((highest_bit_set == 63
2295 || lowest_bit_set == 0)
2296 && all_bits_between_are_set != 0)
2297 || ((highest_bit_set - lowest_bit_set) < 12))
2299 HOST_WIDE_INT the_const = -1;
2300 int shift = lowest_bit_set;
2302 if ((highest_bit_set != 63
2303 && lowest_bit_set != 0)
2304 || all_bits_between_are_set == 0)
2307 create_simple_focus_bits (high_bits, low_bits,
2310 else if (lowest_bit_set == 0)
2311 shift = -(63 - highest_bit_set);
2313 if (! SPARC_SIMM13_P (the_const))
2316 emit_insn (gen_safe_SET64 (temp, the_const));
2318 emit_insn (gen_rtx_SET (VOIDmode,
2320 gen_rtx_ASHIFT (DImode,
2324 emit_insn (gen_rtx_SET (VOIDmode,
2326 gen_rtx_LSHIFTRT (DImode,
2328 GEN_INT (-shift))));
2334 /* Now a range of 22 or less bits set somewhere.
2335 * 1) sethi %hi(focus_bits), %reg
2336 * sllx %reg, shift, %reg
2337 * 2) sethi %hi(focus_bits), %reg
2338 * srlx %reg, shift, %reg
2340 if ((highest_bit_set - lowest_bit_set) < 21)
2342 unsigned HOST_WIDE_INT focus_bits =
2343 create_simple_focus_bits (high_bits, low_bits,
2344 lowest_bit_set, 10);
2346 if (! SPARC_SETHI_P (focus_bits))
2349 sparc_emit_set_safe_HIGH64 (temp, focus_bits);
2351 /* If lowest_bit_set == 10 then a sethi alone could have done it. */
2352 if (lowest_bit_set < 10)
2353 emit_insn (gen_rtx_SET (VOIDmode,
2355 gen_rtx_LSHIFTRT (DImode, temp,
2356 GEN_INT (10 - lowest_bit_set))));
2357 else if (lowest_bit_set > 10)
2358 emit_insn (gen_rtx_SET (VOIDmode,
2360 gen_rtx_ASHIFT (DImode, temp,
2361 GEN_INT (lowest_bit_set - 10))));
2367 /* 1) sethi %hi(low_bits), %reg
2368 * or %reg, %lo(low_bits), %reg
2369 * 2) sethi %hi(~low_bits), %reg
2370 * xor %reg, %lo(-0x400 | (low_bits & 0x3ff)), %reg
2373 || high_bits == 0xffffffff)
2375 sparc_emit_set_const64_quick1 (op0, temp, low_bits,
2376 (high_bits == 0xffffffff));
2380 /* Now, try 3-insn sequences. */
2382 /* 1) sethi %hi(high_bits), %reg
2383 * or %reg, %lo(high_bits), %reg
2384 * sllx %reg, 32, %reg
2388 sparc_emit_set_const64_quick2 (op0, temp, high_bits, 0, 32);
2392 /* We may be able to do something quick
2393 when the constant is negated, so try that. */
2394 if (const64_is_2insns ((~high_bits) & 0xffffffff,
2395 (~low_bits) & 0xfffffc00))
2397 /* NOTE: The trailing bits get XOR'd so we need the
2398 non-negated bits, not the negated ones. */
2399 unsigned HOST_WIDE_INT trailing_bits = low_bits & 0x3ff;
2401 if ((((~high_bits) & 0xffffffff) == 0
2402 && ((~low_bits) & 0x80000000) == 0)
2403 || (((~high_bits) & 0xffffffff) == 0xffffffff
2404 && ((~low_bits) & 0x80000000) != 0))
2406 int fast_int = (~low_bits & 0xffffffff);
2408 if ((SPARC_SETHI_P (fast_int)
2409 && (~high_bits & 0xffffffff) == 0)
2410 || SPARC_SIMM13_P (fast_int))
2411 emit_insn (gen_safe_SET64 (temp, fast_int));
2413 sparc_emit_set_const64 (temp, GEN_INT64 (fast_int));
2418 #if HOST_BITS_PER_WIDE_INT == 64
2419 negated_const = GEN_INT (((~low_bits) & 0xfffffc00) |
2420 (((HOST_WIDE_INT)((~high_bits) & 0xffffffff))<<32));
2422 negated_const = immed_double_const ((~low_bits) & 0xfffffc00,
2423 (~high_bits) & 0xffffffff,
2426 sparc_emit_set_const64 (temp, negated_const);
2429 /* If we are XOR'ing with -1, then we should emit a one's complement
2430 instead. This way the combiner will notice logical operations
2431 such as ANDN later on and substitute. */
2432 if (trailing_bits == 0x3ff)
2434 emit_insn (gen_rtx_SET (VOIDmode, op0,
2435 gen_rtx_NOT (DImode, temp)));
2439 emit_insn (gen_rtx_SET (VOIDmode,
2441 gen_safe_XOR64 (temp,
2442 (-0x400 | trailing_bits))));
2447 /* 1) sethi %hi(xxx), %reg
2448 * or %reg, %lo(xxx), %reg
2449 * sllx %reg, yyy, %reg
2451 * ??? This is just a generalized version of the low_bits==0
2452 * thing above, FIXME...
2454 if ((highest_bit_set - lowest_bit_set) < 32)
2456 unsigned HOST_WIDE_INT focus_bits =
2457 create_simple_focus_bits (high_bits, low_bits,
2460 /* We can't get here in this state. */
2461 if (highest_bit_set < 32
2462 || lowest_bit_set >= 32)
2465 /* So what we know is that the set bits straddle the
2466 middle of the 64-bit word. */
2467 sparc_emit_set_const64_quick2 (op0, temp,
2473 /* 1) sethi %hi(high_bits), %reg
2474 * or %reg, %lo(high_bits), %reg
2475 * sllx %reg, 32, %reg
2476 * or %reg, low_bits, %reg
2478 if (SPARC_SIMM13_P(low_bits)
2479 && ((int)low_bits > 0))
2481 sparc_emit_set_const64_quick2 (op0, temp, high_bits, low_bits, 32);
2485 /* The easiest way when all else fails, is full decomposition. */
2487 printf ("sparc_emit_set_const64: Hard constant [%08lx%08lx] neg[%08lx%08lx]\n",
2488 high_bits, low_bits, ~high_bits, ~low_bits);
2490 sparc_emit_set_const64_longway (op0, temp, high_bits, low_bits);
2493 /* Given a comparison code (EQ, NE, etc.) and the first operand of a COMPARE,
2494 return the mode to be used for the comparison. For floating-point,
2495 CCFP[E]mode is used. CC_NOOVmode should be used when the first operand
2496 is a PLUS, MINUS, NEG, or ASHIFT. CCmode should be used when no special
2497 processing is needed. */
2500 select_cc_mode (enum rtx_code op, rtx x, rtx y ATTRIBUTE_UNUSED)
2502 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2528 else if (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS
2529 || GET_CODE (x) == NEG || GET_CODE (x) == ASHIFT)
2531 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
2532 return CCX_NOOVmode;
2538 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
2545 /* X and Y are two things to compare using CODE. Emit the compare insn and
2546 return the rtx for the cc reg in the proper mode. */
2549 gen_compare_reg (enum rtx_code code, rtx x, rtx y)
2551 enum machine_mode mode = SELECT_CC_MODE (code, x, y);
2554 /* ??? We don't have movcc patterns so we cannot generate pseudo regs for the
2555 fcc regs (cse can't tell they're really call clobbered regs and will
2556 remove a duplicate comparison even if there is an intervening function
2557 call - it will then try to reload the cc reg via an int reg which is why
2558 we need the movcc patterns). It is possible to provide the movcc
2559 patterns by using the ldxfsr/stxfsr v9 insns. I tried it: you need two
2560 registers (say %g1,%g5) and it takes about 6 insns. A better fix would be
2561 to tell cse that CCFPE mode registers (even pseudos) are call
2564 /* ??? This is an experiment. Rather than making changes to cse which may
2565 or may not be easy/clean, we do our own cse. This is possible because
2566 we will generate hard registers. Cse knows they're call clobbered (it
2567 doesn't know the same thing about pseudos). If we guess wrong, no big
2568 deal, but if we win, great! */
2570 if (TARGET_V9 && GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2571 #if 1 /* experiment */
2574 /* We cycle through the registers to ensure they're all exercised. */
2575 static int next_fcc_reg = 0;
2576 /* Previous x,y for each fcc reg. */
2577 static rtx prev_args[4][2];
2579 /* Scan prev_args for x,y. */
2580 for (reg = 0; reg < 4; reg++)
2581 if (prev_args[reg][0] == x && prev_args[reg][1] == y)
2586 prev_args[reg][0] = x;
2587 prev_args[reg][1] = y;
2588 next_fcc_reg = (next_fcc_reg + 1) & 3;
2590 cc_reg = gen_rtx_REG (mode, reg + SPARC_FIRST_V9_FCC_REG);
2593 cc_reg = gen_reg_rtx (mode);
2594 #endif /* ! experiment */
2595 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2596 cc_reg = gen_rtx_REG (mode, SPARC_FCC_REG);
2598 cc_reg = gen_rtx_REG (mode, SPARC_ICC_REG);
2600 emit_insn (gen_rtx_SET (VOIDmode, cc_reg,
2601 gen_rtx_COMPARE (mode, x, y)));
2606 /* This function is used for v9 only.
2607 CODE is the code for an Scc's comparison.
2608 OPERANDS[0] is the target of the Scc insn.
2609 OPERANDS[1] is the value we compare against const0_rtx (which hasn't
2610 been generated yet).
2612 This function is needed to turn
2615 (gt (reg:CCX 100 %icc)
2619 (gt:DI (reg:CCX 100 %icc)
2622 IE: The instruction recognizer needs to see the mode of the comparison to
2623 find the right instruction. We could use "gt:DI" right in the
2624 define_expand, but leaving it out allows us to handle DI, SI, etc.
2626 We refer to the global sparc compare operands sparc_compare_op0 and
2627 sparc_compare_op1. */
2630 gen_v9_scc (enum rtx_code compare_code, register rtx *operands)
2635 && (GET_MODE (sparc_compare_op0) == DImode
2636 || GET_MODE (operands[0]) == DImode))
2639 op0 = sparc_compare_op0;
2640 op1 = sparc_compare_op1;
2642 /* Try to use the movrCC insns. */
2644 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT
2645 && op1 == const0_rtx
2646 && v9_regcmp_p (compare_code))
2648 /* Special case for op0 != 0. This can be done with one instruction if
2649 operands[0] == sparc_compare_op0. */
2651 if (compare_code == NE
2652 && GET_MODE (operands[0]) == DImode
2653 && rtx_equal_p (op0, operands[0]))
2655 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
2656 gen_rtx_IF_THEN_ELSE (DImode,
2657 gen_rtx_fmt_ee (compare_code, DImode,
2664 if (reg_overlap_mentioned_p (operands[0], op0))
2666 /* Handle the case where operands[0] == sparc_compare_op0.
2667 We "early clobber" the result. */
2668 op0 = gen_reg_rtx (GET_MODE (sparc_compare_op0));
2669 emit_move_insn (op0, sparc_compare_op0);
2672 emit_insn (gen_rtx_SET (VOIDmode, operands[0], const0_rtx));
2673 if (GET_MODE (op0) != DImode)
2675 temp = gen_reg_rtx (DImode);
2676 convert_move (temp, op0, 0);
2680 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
2681 gen_rtx_IF_THEN_ELSE (GET_MODE (operands[0]),
2682 gen_rtx_fmt_ee (compare_code, DImode,
2690 operands[1] = gen_compare_reg (compare_code, op0, op1);
2692 switch (GET_MODE (operands[1]))
2702 emit_insn (gen_rtx_SET (VOIDmode, operands[0], const0_rtx));
2703 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
2704 gen_rtx_IF_THEN_ELSE (GET_MODE (operands[0]),
2705 gen_rtx_fmt_ee (compare_code,
2706 GET_MODE (operands[1]),
2707 operands[1], const0_rtx),
2708 const1_rtx, operands[0])));
2713 /* Emit a conditional jump insn for the v9 architecture using comparison code
2714 CODE and jump target LABEL.
2715 This function exists to take advantage of the v9 brxx insns. */
2718 emit_v9_brxx_insn (enum rtx_code code, rtx op0, rtx label)
2720 emit_jump_insn (gen_rtx_SET (VOIDmode,
2722 gen_rtx_IF_THEN_ELSE (VOIDmode,
2723 gen_rtx_fmt_ee (code, GET_MODE (op0),
2725 gen_rtx_LABEL_REF (VOIDmode, label),
2729 /* Generate a DFmode part of a hard TFmode register.
2730 REG is the TFmode hard register, LOW is 1 for the
2731 low 64bit of the register and 0 otherwise.
2734 gen_df_reg (rtx reg, int low)
2736 int regno = REGNO (reg);
2738 if ((WORDS_BIG_ENDIAN == 0) ^ (low != 0))
2739 regno += (TARGET_ARCH64 && regno < 32) ? 1 : 2;
2740 return gen_rtx_REG (DFmode, regno);
2743 /* Generate a call to FUNC with OPERANDS. Operand 0 is the return value.
2744 Unlike normal calls, TFmode operands are passed by reference. It is
2745 assumed that no more than 3 operands are required. */
2748 emit_soft_tfmode_libcall (const char *func_name, int nargs, rtx *operands)
2750 rtx ret_slot = NULL, arg[3], func_sym;
2753 /* We only expect to be called for conversions, unary, and binary ops. */
2754 if (nargs < 2 || nargs > 3)
2757 for (i = 0; i < nargs; ++i)
2759 rtx this_arg = operands[i];
2762 /* TFmode arguments and return values are passed by reference. */
2763 if (GET_MODE (this_arg) == TFmode)
2765 int force_stack_temp;
2767 force_stack_temp = 0;
2768 if (TARGET_BUGGY_QP_LIB && i == 0)
2769 force_stack_temp = 1;
2771 if (GET_CODE (this_arg) == MEM
2772 && ! force_stack_temp)
2773 this_arg = XEXP (this_arg, 0);
2774 else if (CONSTANT_P (this_arg)
2775 && ! force_stack_temp)
2777 this_slot = force_const_mem (TFmode, this_arg);
2778 this_arg = XEXP (this_slot, 0);
2782 this_slot = assign_stack_temp (TFmode, GET_MODE_SIZE (TFmode), 0);
2784 /* Operand 0 is the return value. We'll copy it out later. */
2786 emit_move_insn (this_slot, this_arg);
2788 ret_slot = this_slot;
2790 this_arg = XEXP (this_slot, 0);
2797 func_sym = gen_rtx_SYMBOL_REF (Pmode, func_name);
2799 if (GET_MODE (operands[0]) == TFmode)
2802 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 2,
2803 arg[0], GET_MODE (arg[0]),
2804 arg[1], GET_MODE (arg[1]));
2806 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 3,
2807 arg[0], GET_MODE (arg[0]),
2808 arg[1], GET_MODE (arg[1]),
2809 arg[2], GET_MODE (arg[2]));
2812 emit_move_insn (operands[0], ret_slot);
2821 ret = emit_library_call_value (func_sym, operands[0], LCT_NORMAL,
2822 GET_MODE (operands[0]), 1,
2823 arg[1], GET_MODE (arg[1]));
2825 if (ret != operands[0])
2826 emit_move_insn (operands[0], ret);
2830 /* Expand soft-float TFmode calls to sparc abi routines. */
2833 emit_soft_tfmode_binop (enum rtx_code code, rtx *operands)
2855 emit_soft_tfmode_libcall (func, 3, operands);
2859 emit_soft_tfmode_unop (enum rtx_code code, rtx *operands)
2872 emit_soft_tfmode_libcall (func, 2, operands);
2876 emit_soft_tfmode_cvt (enum rtx_code code, rtx *operands)
2883 switch (GET_MODE (operands[1]))
2896 case FLOAT_TRUNCATE:
2897 switch (GET_MODE (operands[0]))
2911 switch (GET_MODE (operands[1]))
2924 case UNSIGNED_FLOAT:
2925 switch (GET_MODE (operands[1]))
2939 switch (GET_MODE (operands[0]))
2953 switch (GET_MODE (operands[0]))
2970 emit_soft_tfmode_libcall (func, 2, operands);
2973 /* Expand a hard-float tfmode operation. All arguments must be in
2977 emit_hard_tfmode_operation (enum rtx_code code, rtx *operands)
2981 if (GET_RTX_CLASS (code) == RTX_UNARY)
2983 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2984 op = gen_rtx_fmt_e (code, GET_MODE (operands[0]), operands[1]);
2988 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2989 operands[2] = force_reg (GET_MODE (operands[2]), operands[2]);
2990 op = gen_rtx_fmt_ee (code, GET_MODE (operands[0]),
2991 operands[1], operands[2]);
2994 if (register_operand (operands[0], VOIDmode))
2997 dest = gen_reg_rtx (GET_MODE (operands[0]));
2999 emit_insn (gen_rtx_SET (VOIDmode, dest, op));
3001 if (dest != operands[0])
3002 emit_move_insn (operands[0], dest);
3006 emit_tfmode_binop (enum rtx_code code, rtx *operands)
3008 if (TARGET_HARD_QUAD)
3009 emit_hard_tfmode_operation (code, operands);
3011 emit_soft_tfmode_binop (code, operands);
3015 emit_tfmode_unop (enum rtx_code code, rtx *operands)
3017 if (TARGET_HARD_QUAD)
3018 emit_hard_tfmode_operation (code, operands);
3020 emit_soft_tfmode_unop (code, operands);
3024 emit_tfmode_cvt (enum rtx_code code, rtx *operands)
3026 if (TARGET_HARD_QUAD)
3027 emit_hard_tfmode_operation (code, operands);
3029 emit_soft_tfmode_cvt (code, operands);
3032 /* Return nonzero if a branch/jump/call instruction will be emitting
3033 nop into its delay slot. */
3036 empty_delay_slot (rtx insn)
3040 /* If no previous instruction (should not happen), return true. */
3041 if (PREV_INSN (insn) == NULL)
3044 seq = NEXT_INSN (PREV_INSN (insn));
3045 if (GET_CODE (PATTERN (seq)) == SEQUENCE)
3051 /* Return nonzero if TRIAL can go into the call delay slot. */
3054 tls_call_delay (rtx trial)
3059 call __tls_get_addr, %tgd_call (foo)
3060 add %l7, %o0, %o0, %tgd_add (foo)
3061 while Sun as/ld does not. */
3062 if (TARGET_GNU_TLS || !TARGET_TLS)
3065 pat = PATTERN (trial);
3066 if (GET_CODE (pat) != SET || GET_CODE (SET_DEST (pat)) != PLUS)
3069 unspec = XEXP (SET_DEST (pat), 1);
3070 if (GET_CODE (unspec) != UNSPEC
3071 || (XINT (unspec, 1) != UNSPEC_TLSGD
3072 && XINT (unspec, 1) != UNSPEC_TLSLDM))
3078 /* Return nonzero if TRIAL, an insn, can be combined with a 'restore'
3079 instruction. RETURN_P is true if the v9 variant 'return' is to be
3080 considered in the test too.
3082 TRIAL must be a SET whose destination is a REG appropriate for the
3083 'restore' instruction or, if RETURN_P is true, for the 'return'
3087 eligible_for_restore_insn (rtx trial, bool return_p)
3089 rtx pat = PATTERN (trial);
3090 rtx src = SET_SRC (pat);
3092 /* The 'restore src,%g0,dest' pattern for word mode and below. */
3093 if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
3094 && arith_operand (src, GET_MODE (src)))
3097 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
3099 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (SImode);
3102 /* The 'restore src,%g0,dest' pattern for double-word mode. */
3103 else if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
3104 && arith_double_operand (src, GET_MODE (src)))
3105 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
3107 /* The 'restore src,%g0,dest' pattern for float if no FPU. */
3108 else if (! TARGET_FPU && register_operand (src, SFmode))
3111 /* The 'restore src,%g0,dest' pattern for double if no FPU. */
3112 else if (! TARGET_FPU && TARGET_ARCH64 && register_operand (src, DFmode))
3115 /* If we have the 'return' instruction, anything that does not use
3116 local or output registers and can go into a delay slot wins. */
3117 else if (return_p && TARGET_V9 && ! epilogue_renumber (&pat, 1)
3118 && (get_attr_in_uncond_branch_delay (trial)
3119 == IN_UNCOND_BRANCH_DELAY_TRUE))
3122 /* The 'restore src1,src2,dest' pattern for SImode. */
3123 else if (GET_CODE (src) == PLUS
3124 && register_operand (XEXP (src, 0), SImode)
3125 && arith_operand (XEXP (src, 1), SImode))
3128 /* The 'restore src1,src2,dest' pattern for DImode. */
3129 else if (GET_CODE (src) == PLUS
3130 && register_operand (XEXP (src, 0), DImode)
3131 && arith_double_operand (XEXP (src, 1), DImode))
3134 /* The 'restore src1,%lo(src2),dest' pattern. */
3135 else if (GET_CODE (src) == LO_SUM
3136 && ! TARGET_CM_MEDMID
3137 && ((register_operand (XEXP (src, 0), SImode)
3138 && immediate_operand (XEXP (src, 1), SImode))
3140 && register_operand (XEXP (src, 0), DImode)
3141 && immediate_operand (XEXP (src, 1), DImode))))
3144 /* The 'restore src,src,dest' pattern. */
3145 else if (GET_CODE (src) == ASHIFT
3146 && (register_operand (XEXP (src, 0), SImode)
3147 || register_operand (XEXP (src, 0), DImode))
3148 && XEXP (src, 1) == const1_rtx)
3154 /* Return nonzero if TRIAL can go into the function return's
3158 eligible_for_return_delay (rtx trial)
3162 if (GET_CODE (trial) != INSN || GET_CODE (PATTERN (trial)) != SET)
3165 if (get_attr_length (trial) != 1)
3168 /* If there are any call-saved registers, we should scan TRIAL if it
3169 does not reference them. For now just make it easy. */
3173 /* If the function uses __builtin_eh_return, the eh_return machinery
3174 occupies the delay slot. */
3175 if (current_function_calls_eh_return)
3178 /* In the case of a true leaf function, anything can go into the slot. */
3179 if (sparc_leaf_function_p)
3180 return get_attr_in_uncond_branch_delay (trial)
3181 == IN_UNCOND_BRANCH_DELAY_TRUE;
3183 pat = PATTERN (trial);
3185 /* Otherwise, only operations which can be done in tandem with
3186 a `restore' or `return' insn can go into the delay slot. */
3187 if (GET_CODE (SET_DEST (pat)) != REG
3188 || (REGNO (SET_DEST (pat)) >= 8 && REGNO (SET_DEST (pat)) < 24))
3191 /* If this instruction sets up floating point register and we have a return
3192 instruction, it can probably go in. But restore will not work
3194 if (REGNO (SET_DEST (pat)) >= 32)
3196 && ! epilogue_renumber (&pat, 1)
3197 && (get_attr_in_uncond_branch_delay (trial)
3198 == IN_UNCOND_BRANCH_DELAY_TRUE));
3200 return eligible_for_restore_insn (trial, true);
3203 /* Return nonzero if TRIAL can go into the sibling call's
3207 eligible_for_sibcall_delay (rtx trial)
3211 if (GET_CODE (trial) != INSN || GET_CODE (PATTERN (trial)) != SET)
3214 if (get_attr_length (trial) != 1)
3217 pat = PATTERN (trial);
3219 if (sparc_leaf_function_p)
3221 /* If the tail call is done using the call instruction,
3222 we have to restore %o7 in the delay slot. */
3223 if (LEAF_SIBCALL_SLOT_RESERVED_P)
3226 /* %g1 is used to build the function address */
3227 if (reg_mentioned_p (gen_rtx_REG (Pmode, 1), pat))
3233 /* Otherwise, only operations which can be done in tandem with
3234 a `restore' insn can go into the delay slot. */
3235 if (GET_CODE (SET_DEST (pat)) != REG
3236 || (REGNO (SET_DEST (pat)) >= 8 && REGNO (SET_DEST (pat)) < 24)
3237 || REGNO (SET_DEST (pat)) >= 32)
3240 /* If it mentions %o7, it can't go in, because sibcall will clobber it
3242 if (reg_mentioned_p (gen_rtx_REG (Pmode, 15), pat))
3245 return eligible_for_restore_insn (trial, false);
3249 short_branch (int uid1, int uid2)
3251 int delta = INSN_ADDRESSES (uid1) - INSN_ADDRESSES (uid2);
3253 /* Leave a few words of "slop". */
3254 if (delta >= -1023 && delta <= 1022)
3260 /* Return nonzero if REG is not used after INSN.
3261 We assume REG is a reload reg, and therefore does
3262 not live past labels or calls or jumps. */
3264 reg_unused_after (rtx reg, rtx insn)
3266 enum rtx_code code, prev_code = UNKNOWN;
3268 while ((insn = NEXT_INSN (insn)))
3270 if (prev_code == CALL_INSN && call_used_regs[REGNO (reg)])
3273 code = GET_CODE (insn);
3274 if (GET_CODE (insn) == CODE_LABEL)
3279 rtx set = single_set (insn);
3280 int in_src = set && reg_overlap_mentioned_p (reg, SET_SRC (set));
3283 if (set && reg_overlap_mentioned_p (reg, SET_DEST (set)))
3285 if (set == 0 && reg_overlap_mentioned_p (reg, PATTERN (insn)))
3293 /* Determine if it's legal to put X into the constant pool. This
3294 is not possible if X contains the address of a symbol that is
3295 not constant (TLS) or not known at final link time (PIC). */
3298 sparc_cannot_force_const_mem (rtx x)
3300 switch (GET_CODE (x))
3305 /* Accept all non-symbolic constants. */
3309 /* Labels are OK iff we are non-PIC. */
3310 return flag_pic != 0;
3313 /* 'Naked' TLS symbol references are never OK,
3314 non-TLS symbols are OK iff we are non-PIC. */
3315 if (SYMBOL_REF_TLS_MODEL (x))
3318 return flag_pic != 0;
3321 return sparc_cannot_force_const_mem (XEXP (x, 0));
3324 return sparc_cannot_force_const_mem (XEXP (x, 0))
3325 || sparc_cannot_force_const_mem (XEXP (x, 1));
3333 /* The table we use to reference PIC data. */
3334 static GTY(()) rtx global_offset_table;
3336 /* The function we use to get at it. */
3337 static GTY(()) rtx add_pc_to_pic_symbol;
3338 static GTY(()) char add_pc_to_pic_symbol_name[256];
3340 /* Ensure that we are not using patterns that are not OK with PIC. */
3348 if (GET_CODE (recog_data.operand[i]) == SYMBOL_REF
3349 || (GET_CODE (recog_data.operand[i]) == CONST
3350 && ! (GET_CODE (XEXP (recog_data.operand[i], 0)) == MINUS
3351 && (XEXP (XEXP (recog_data.operand[i], 0), 0)
3352 == global_offset_table)
3353 && (GET_CODE (XEXP (XEXP (recog_data.operand[i], 0), 1))
3362 /* Return true if X is an address which needs a temporary register when
3363 reloaded while generating PIC code. */
3366 pic_address_needs_scratch (rtx x)
3368 /* An address which is a symbolic plus a non SMALL_INT needs a temp reg. */
3369 if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS
3370 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
3371 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
3372 && ! SMALL_INT (XEXP (XEXP (x, 0), 1)))
3378 /* Determine if a given RTX is a valid constant. We already know this
3379 satisfies CONSTANT_P. */
3382 legitimate_constant_p (rtx x)
3386 switch (GET_CODE (x))
3389 /* TLS symbols are not constant. */
3390 if (SYMBOL_REF_TLS_MODEL (x))
3395 inner = XEXP (x, 0);
3397 /* Offsets of TLS symbols are never valid.
3398 Discourage CSE from creating them. */
3399 if (GET_CODE (inner) == PLUS
3400 && tls_symbolic_operand (XEXP (inner, 0)))
3405 if (GET_MODE (x) == VOIDmode)
3408 /* Floating point constants are generally not ok.
3409 The only exception is 0.0 in VIS. */
3411 && (GET_MODE (x) == SFmode
3412 || GET_MODE (x) == DFmode
3413 || GET_MODE (x) == TFmode)
3414 && fp_zero_operand (x, GET_MODE (x)))
3426 /* Determine if a given RTX is a valid constant address. */
3429 constant_address_p (rtx x)
3431 switch (GET_CODE (x))
3439 if (flag_pic && pic_address_needs_scratch (x))
3441 return legitimate_constant_p (x);
3444 return !flag_pic && legitimate_constant_p (x);
3451 /* Nonzero if the constant value X is a legitimate general operand
3452 when generating PIC code. It is given that flag_pic is on and
3453 that X satisfies CONSTANT_P or is a CONST_DOUBLE. */
3456 legitimate_pic_operand_p (rtx x)
3458 if (pic_address_needs_scratch (x))
3460 if (tls_symbolic_operand (x)
3461 || (GET_CODE (x) == CONST
3462 && GET_CODE (XEXP (x, 0)) == PLUS
3463 && tls_symbolic_operand (XEXP (XEXP (x, 0), 0))))
3468 /* Return nonzero if ADDR is a valid memory address.
3469 STRICT specifies whether strict register checking applies. */
3472 legitimate_address_p (enum machine_mode mode, rtx addr, int strict)
3474 rtx rs1 = NULL, rs2 = NULL, imm1 = NULL, imm2;
3476 if (REG_P (addr) || GET_CODE (addr) == SUBREG)
3478 else if (GET_CODE (addr) == PLUS)
3480 rs1 = XEXP (addr, 0);
3481 rs2 = XEXP (addr, 1);
3483 /* Canonicalize. REG comes first, if there are no regs,
3484 LO_SUM comes first. */
3486 && GET_CODE (rs1) != SUBREG
3488 || GET_CODE (rs2) == SUBREG
3489 || (GET_CODE (rs2) == LO_SUM && GET_CODE (rs1) != LO_SUM)))
3491 rs1 = XEXP (addr, 1);
3492 rs2 = XEXP (addr, 0);
3496 && rs1 == pic_offset_table_rtx
3498 && GET_CODE (rs2) != SUBREG
3499 && GET_CODE (rs2) != LO_SUM
3500 && GET_CODE (rs2) != MEM
3501 && !tls_symbolic_operand (rs2)
3502 && (! symbolic_operand (rs2, VOIDmode) || mode == Pmode)
3503 && (GET_CODE (rs2) != CONST_INT || SMALL_INT (rs2)))
3505 || GET_CODE (rs1) == SUBREG)
3506 && RTX_OK_FOR_OFFSET_P (rs2)))
3511 else if ((REG_P (rs1) || GET_CODE (rs1) == SUBREG)
3512 && (REG_P (rs2) || GET_CODE (rs2) == SUBREG))
3514 /* We prohibit REG + REG for TFmode when there are no quad move insns
3515 and we consequently need to split. We do this because REG+REG
3516 is not an offsettable address. If we get the situation in reload
3517 where source and destination of a movtf pattern are both MEMs with
3518 REG+REG address, then only one of them gets converted to an
3519 offsettable address. */
3521 && ! (TARGET_FPU && TARGET_ARCH64 && TARGET_HARD_QUAD))
3524 /* We prohibit REG + REG on ARCH32 if not optimizing for
3525 DFmode/DImode because then mem_min_alignment is likely to be zero
3526 after reload and the forced split would lack a matching splitter
3528 if (TARGET_ARCH32 && !optimize
3529 && (mode == DFmode || mode == DImode))
3532 else if (USE_AS_OFFSETABLE_LO10
3533 && GET_CODE (rs1) == LO_SUM
3535 && ! TARGET_CM_MEDMID
3536 && RTX_OK_FOR_OLO10_P (rs2))
3540 imm1 = XEXP (rs1, 1);
3541 rs1 = XEXP (rs1, 0);
3542 if (! CONSTANT_P (imm1) || tls_symbolic_operand (rs1))
3546 else if (GET_CODE (addr) == LO_SUM)
3548 rs1 = XEXP (addr, 0);
3549 imm1 = XEXP (addr, 1);
3551 if (! CONSTANT_P (imm1) || tls_symbolic_operand (rs1))
3554 if (USE_AS_OFFSETABLE_LO10)
3556 /* We can't allow TFmode, because an offset greater than or equal to
3557 the alignment (8) may cause the LO_SUM to overflow if !v9. */
3558 if (mode == TFmode && ! TARGET_V9)
3563 /* We prohibit LO_SUM for TFmode when there are no quad move insns
3564 and we consequently need to split. We do this because LO_SUM
3565 is not an offsettable address. If we get the situation in reload
3566 where source and destination of a movtf pattern are both MEMs with
3567 LO_SUM address, then only one of them gets converted to an
3568 offsettable address. */
3570 && ! (TARGET_FPU && TARGET_ARCH64 && TARGET_HARD_QUAD))
3574 else if (GET_CODE (addr) == CONST_INT && SMALL_INT (addr))
3579 if (GET_CODE (rs1) == SUBREG)
3580 rs1 = SUBREG_REG (rs1);
3586 if (GET_CODE (rs2) == SUBREG)
3587 rs2 = SUBREG_REG (rs2);
3594 if (!REGNO_OK_FOR_BASE_P (REGNO (rs1))
3595 || (rs2 && !REGNO_OK_FOR_BASE_P (REGNO (rs2))))
3600 if ((REGNO (rs1) >= 32
3601 && REGNO (rs1) != FRAME_POINTER_REGNUM
3602 && REGNO (rs1) < FIRST_PSEUDO_REGISTER)
3604 && (REGNO (rs2) >= 32
3605 && REGNO (rs2) != FRAME_POINTER_REGNUM
3606 && REGNO (rs2) < FIRST_PSEUDO_REGISTER)))
3612 /* Construct the SYMBOL_REF for the tls_get_offset function. */
3614 static GTY(()) rtx sparc_tls_symbol;
3616 sparc_tls_get_addr (void)
3618 if (!sparc_tls_symbol)
3619 sparc_tls_symbol = gen_rtx_SYMBOL_REF (Pmode, "__tls_get_addr");
3621 return sparc_tls_symbol;
3625 sparc_tls_got (void)
3630 current_function_uses_pic_offset_table = 1;
3631 return pic_offset_table_rtx;
3634 if (!global_offset_table)
3635 global_offset_table = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_");
3636 temp = gen_reg_rtx (Pmode);
3637 emit_move_insn (temp, global_offset_table);
3642 /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute
3643 this (thread-local) address. */
3646 legitimize_tls_address (rtx addr)
3648 rtx temp1, temp2, temp3, ret, o0, got, insn;
3653 if (GET_CODE (addr) == SYMBOL_REF)
3654 switch (SYMBOL_REF_TLS_MODEL (addr))
3656 case TLS_MODEL_GLOBAL_DYNAMIC:
3658 temp1 = gen_reg_rtx (SImode);
3659 temp2 = gen_reg_rtx (SImode);
3660 ret = gen_reg_rtx (Pmode);
3661 o0 = gen_rtx_REG (Pmode, 8);
3662 got = sparc_tls_got ();
3663 emit_insn (gen_tgd_hi22 (temp1, addr));
3664 emit_insn (gen_tgd_lo10 (temp2, temp1, addr));
3667 emit_insn (gen_tgd_add32 (o0, got, temp2, addr));
3668 insn = emit_call_insn (gen_tgd_call32 (o0, sparc_tls_get_addr (),
3673 emit_insn (gen_tgd_add64 (o0, got, temp2, addr));
3674 insn = emit_call_insn (gen_tgd_call64 (o0, sparc_tls_get_addr (),
3677 CALL_INSN_FUNCTION_USAGE (insn)
3678 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_USE (VOIDmode, o0),
3679 CALL_INSN_FUNCTION_USAGE (insn));
3680 insn = get_insns ();
3682 emit_libcall_block (insn, ret, o0, addr);
3685 case TLS_MODEL_LOCAL_DYNAMIC:
3687 temp1 = gen_reg_rtx (SImode);
3688 temp2 = gen_reg_rtx (SImode);
3689 temp3 = gen_reg_rtx (Pmode);
3690 ret = gen_reg_rtx (Pmode);
3691 o0 = gen_rtx_REG (Pmode, 8);
3692 got = sparc_tls_got ();
3693 emit_insn (gen_tldm_hi22 (temp1));
3694 emit_insn (gen_tldm_lo10 (temp2, temp1));
3697 emit_insn (gen_tldm_add32 (o0, got, temp2));
3698 insn = emit_call_insn (gen_tldm_call32 (o0, sparc_tls_get_addr (),
3703 emit_insn (gen_tldm_add64 (o0, got, temp2));
3704 insn = emit_call_insn (gen_tldm_call64 (o0, sparc_tls_get_addr (),
3707 CALL_INSN_FUNCTION_USAGE (insn)
3708 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_USE (VOIDmode, o0),
3709 CALL_INSN_FUNCTION_USAGE (insn));
3710 insn = get_insns ();
3712 emit_libcall_block (insn, temp3, o0,
3713 gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
3714 UNSPEC_TLSLD_BASE));
3715 temp1 = gen_reg_rtx (SImode);
3716 temp2 = gen_reg_rtx (SImode);
3717 emit_insn (gen_tldo_hix22 (temp1, addr));
3718 emit_insn (gen_tldo_lox10 (temp2, temp1, addr));
3720 emit_insn (gen_tldo_add32 (ret, temp3, temp2, addr));
3722 emit_insn (gen_tldo_add64 (ret, temp3, temp2, addr));
3725 case TLS_MODEL_INITIAL_EXEC:
3726 temp1 = gen_reg_rtx (SImode);
3727 temp2 = gen_reg_rtx (SImode);
3728 temp3 = gen_reg_rtx (Pmode);
3729 got = sparc_tls_got ();
3730 emit_insn (gen_tie_hi22 (temp1, addr));
3731 emit_insn (gen_tie_lo10 (temp2, temp1, addr));
3733 emit_insn (gen_tie_ld32 (temp3, got, temp2, addr));
3735 emit_insn (gen_tie_ld64 (temp3, got, temp2, addr));
3738 ret = gen_reg_rtx (Pmode);
3740 emit_insn (gen_tie_add32 (ret, gen_rtx_REG (Pmode, 7),
3743 emit_insn (gen_tie_add64 (ret, gen_rtx_REG (Pmode, 7),
3747 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp3);
3750 case TLS_MODEL_LOCAL_EXEC:
3751 temp1 = gen_reg_rtx (Pmode);
3752 temp2 = gen_reg_rtx (Pmode);
3755 emit_insn (gen_tle_hix22_sp32 (temp1, addr));
3756 emit_insn (gen_tle_lox10_sp32 (temp2, temp1, addr));
3760 emit_insn (gen_tle_hix22_sp64 (temp1, addr));
3761 emit_insn (gen_tle_lox10_sp64 (temp2, temp1, addr));
3763 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp2);
3771 abort (); /* for now ... */
3777 /* Legitimize PIC addresses. If the address is already position-independent,
3778 we return ORIG. Newly generated position-independent addresses go into a
3779 reg. This is REG if nonzero, otherwise we allocate register(s) as
3783 legitimize_pic_address (rtx orig, enum machine_mode mode ATTRIBUTE_UNUSED,
3786 if (GET_CODE (orig) == SYMBOL_REF)
3788 rtx pic_ref, address;
3793 if (reload_in_progress || reload_completed)
3796 reg = gen_reg_rtx (Pmode);
3801 /* If not during reload, allocate another temp reg here for loading
3802 in the address, so that these instructions can be optimized
3804 rtx temp_reg = ((reload_in_progress || reload_completed)
3805 ? reg : gen_reg_rtx (Pmode));
3807 /* Must put the SYMBOL_REF inside an UNSPEC here so that cse
3808 won't get confused into thinking that these two instructions
3809 are loading in the true address of the symbol. If in the
3810 future a PIC rtx exists, that should be used instead. */
3811 if (Pmode == SImode)
3813 emit_insn (gen_movsi_high_pic (temp_reg, orig));
3814 emit_insn (gen_movsi_lo_sum_pic (temp_reg, temp_reg, orig));
3818 emit_insn (gen_movdi_high_pic (temp_reg, orig));
3819 emit_insn (gen_movdi_lo_sum_pic (temp_reg, temp_reg, orig));
3826 pic_ref = gen_const_mem (Pmode,
3827 gen_rtx_PLUS (Pmode,
3828 pic_offset_table_rtx, address));
3829 current_function_uses_pic_offset_table = 1;
3830 insn = emit_move_insn (reg, pic_ref);
3831 /* Put a REG_EQUAL note on this insn, so that it can be optimized
3833 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_EQUAL, orig,
3837 else if (GET_CODE (orig) == CONST)
3841 if (GET_CODE (XEXP (orig, 0)) == PLUS
3842 && XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx)
3847 if (reload_in_progress || reload_completed)
3850 reg = gen_reg_rtx (Pmode);
3853 if (GET_CODE (XEXP (orig, 0)) == PLUS)
3855 base = legitimize_pic_address (XEXP (XEXP (orig, 0), 0), Pmode, reg);
3856 offset = legitimize_pic_address (XEXP (XEXP (orig, 0), 1), Pmode,
3857 base == reg ? 0 : reg);
3862 if (GET_CODE (offset) == CONST_INT)
3864 if (SMALL_INT (offset))
3865 return plus_constant (base, INTVAL (offset));
3866 else if (! reload_in_progress && ! reload_completed)
3867 offset = force_reg (Pmode, offset);
3869 /* If we reach here, then something is seriously wrong. */
3872 return gen_rtx_PLUS (Pmode, base, offset);
3874 else if (GET_CODE (orig) == LABEL_REF)
3875 /* ??? Why do we do this? */
3876 /* Now movsi_pic_label_ref uses it, but we ought to be checking that
3877 the register is live instead, in case it is eliminated. */
3878 current_function_uses_pic_offset_table = 1;
3883 /* Try machine-dependent ways of modifying an illegitimate address X
3884 to be legitimate. If we find one, return the new, valid address.
3886 OLDX is the address as it was before break_out_memory_refs was called.
3887 In some cases it is useful to look at this to decide what needs to be done.
3889 MODE is the mode of the operand pointed to by X. */
3892 legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED, enum machine_mode mode)
3896 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == MULT)
3897 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3898 force_operand (XEXP (x, 0), NULL_RTX));
3899 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == MULT)
3900 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3901 force_operand (XEXP (x, 1), NULL_RTX));
3902 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == PLUS)
3903 x = gen_rtx_PLUS (Pmode, force_operand (XEXP (x, 0), NULL_RTX),
3905 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == PLUS)
3906 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3907 force_operand (XEXP (x, 1), NULL_RTX));
3909 if (x != orig_x && legitimate_address_p (mode, x, FALSE))
3912 if (tls_symbolic_operand (x))
3913 x = legitimize_tls_address (x);
3915 x = legitimize_pic_address (x, mode, 0);
3916 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 1)))
3917 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3918 copy_to_mode_reg (Pmode, XEXP (x, 1)));
3919 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 0)))
3920 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3921 copy_to_mode_reg (Pmode, XEXP (x, 0)));
3922 else if (GET_CODE (x) == SYMBOL_REF
3923 || GET_CODE (x) == CONST
3924 || GET_CODE (x) == LABEL_REF)
3925 x = copy_to_suggested_reg (x, NULL_RTX, Pmode);
3929 /* Emit the special PIC prologue. */
3932 load_pic_register (void)
3934 int orig_flag_pic = flag_pic;
3936 /* If we haven't emitted the special helper function, do so now. */
3937 if (add_pc_to_pic_symbol_name[0] == 0)
3939 const char *pic_name = reg_names[REGNO (pic_offset_table_rtx)];
3942 ASM_GENERATE_INTERNAL_LABEL (add_pc_to_pic_symbol_name, "LADDPC", 0);
3945 align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
3947 ASM_OUTPUT_ALIGN (asm_out_file, align);
3948 ASM_OUTPUT_LABEL (asm_out_file, add_pc_to_pic_symbol_name);
3949 if (flag_delayed_branch)
3950 fprintf (asm_out_file, "\tjmp %%o7+8\n\t add\t%%o7, %s, %s\n",
3951 pic_name, pic_name);
3953 fprintf (asm_out_file, "\tadd\t%%o7, %s, %s\n\tjmp %%o7+8\n\t nop\n",
3954 pic_name, pic_name);
3957 /* Initialize every time through, since we can't easily
3958 know this to be permanent. */
3959 global_offset_table = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_");
3960 add_pc_to_pic_symbol = gen_rtx_SYMBOL_REF (Pmode, add_pc_to_pic_symbol_name);
3963 emit_insn (gen_load_pcrel_sym (pic_offset_table_rtx, global_offset_table,
3964 add_pc_to_pic_symbol));
3965 flag_pic = orig_flag_pic;
3967 /* Need to emit this whether or not we obey regdecls,
3968 since setjmp/longjmp can cause life info to screw up.
3969 ??? In the case where we don't obey regdecls, this is not sufficient
3970 since we may not fall out the bottom. */
3971 emit_insn (gen_rtx_USE (VOIDmode, pic_offset_table_rtx));
3974 /* Return 1 if RTX is a MEM which is known to be aligned to at
3975 least a DESIRED byte boundary. */
3978 mem_min_alignment (rtx mem, int desired)
3980 rtx addr, base, offset;
3982 /* If it's not a MEM we can't accept it. */
3983 if (GET_CODE (mem) != MEM)
3986 addr = XEXP (mem, 0);
3987 base = offset = NULL_RTX;
3988 if (GET_CODE (addr) == PLUS)
3990 if (GET_CODE (XEXP (addr, 0)) == REG)
3992 base = XEXP (addr, 0);
3994 /* What we are saying here is that if the base
3995 REG is aligned properly, the compiler will make
3996 sure any REG based index upon it will be so
3998 if (GET_CODE (XEXP (addr, 1)) == CONST_INT)
3999 offset = XEXP (addr, 1);
4001 offset = const0_rtx;
4004 else if (GET_CODE (addr) == REG)
4007 offset = const0_rtx;
4010 if (base != NULL_RTX)
4012 int regno = REGNO (base);
4014 if (regno != HARD_FRAME_POINTER_REGNUM && regno != STACK_POINTER_REGNUM)
4016 /* Check if the compiler has recorded some information
4017 about the alignment of the base REG. If reload has
4018 completed, we already matched with proper alignments.
4019 If not running global_alloc, reload might give us
4020 unaligned pointer to local stack though. */
4022 && REGNO_POINTER_ALIGN (regno) >= desired * BITS_PER_UNIT)
4023 || (optimize && reload_completed))
4024 && (INTVAL (offset) & (desired - 1)) == 0)
4029 if (((INTVAL (offset) - SPARC_STACK_BIAS) & (desired - 1)) == 0)
4033 else if (! TARGET_UNALIGNED_DOUBLES
4034 || CONSTANT_P (addr)
4035 || GET_CODE (addr) == LO_SUM)
4037 /* Anything else we know is properly aligned unless TARGET_UNALIGNED_DOUBLES
4038 is true, in which case we can only assume that an access is aligned if
4039 it is to a constant address, or the address involves a LO_SUM. */
4043 /* An obviously unaligned address. */
4048 /* Vectors to keep interesting information about registers where it can easily
4049 be got. We used to use the actual mode value as the bit number, but there
4050 are more than 32 modes now. Instead we use two tables: one indexed by
4051 hard register number, and one indexed by mode. */
4053 /* The purpose of sparc_mode_class is to shrink the range of modes so that
4054 they all fit (as bit numbers) in a 32 bit word (again). Each real mode is
4055 mapped into one sparc_mode_class mode. */
4057 enum sparc_mode_class {
4058 S_MODE, D_MODE, T_MODE, O_MODE,
4059 SF_MODE, DF_MODE, TF_MODE, OF_MODE,
4063 /* Modes for single-word and smaller quantities. */
4064 #define S_MODES ((1 << (int) S_MODE) | (1 << (int) SF_MODE))
4066 /* Modes for double-word and smaller quantities. */
4067 #define D_MODES (S_MODES | (1 << (int) D_MODE) | (1 << DF_MODE))
4069 /* Modes for quad-word and smaller quantities. */
4070 #define T_MODES (D_MODES | (1 << (int) T_MODE) | (1 << (int) TF_MODE))
4072 /* Modes for 8-word and smaller quantities. */
4073 #define O_MODES (T_MODES | (1 << (int) O_MODE) | (1 << (int) OF_MODE))
4075 /* Modes for single-float quantities. We must allow any single word or
4076 smaller quantity. This is because the fix/float conversion instructions
4077 take integer inputs/outputs from the float registers. */
4078 #define SF_MODES (S_MODES)
4080 /* Modes for double-float and smaller quantities. */
4081 #define DF_MODES (S_MODES | D_MODES)
4083 /* Modes for double-float only quantities. */
4084 #define DF_MODES_NO_S ((1 << (int) D_MODE) | (1 << (int) DF_MODE))
4086 /* Modes for quad-float only quantities. */
4087 #define TF_ONLY_MODES (1 << (int) TF_MODE)
4089 /* Modes for quad-float and smaller quantities. */
4090 #define TF_MODES (DF_MODES | TF_ONLY_MODES)
4092 /* Modes for quad-float and double-float quantities. */
4093 #define TF_MODES_NO_S (DF_MODES_NO_S | TF_ONLY_MODES)
4095 /* Modes for quad-float pair only quantities. */
4096 #define OF_ONLY_MODES (1 << (int) OF_MODE)
4098 /* Modes for quad-float pairs and smaller quantities. */
4099 #define OF_MODES (TF_MODES | OF_ONLY_MODES)
4101 #define OF_MODES_NO_S (TF_MODES_NO_S | OF_ONLY_MODES)
4103 /* Modes for condition codes. */
4104 #define CC_MODES (1 << (int) CC_MODE)
4105 #define CCFP_MODES (1 << (int) CCFP_MODE)
4107 /* Value is 1 if register/mode pair is acceptable on sparc.
4108 The funny mixture of D and T modes is because integer operations
4109 do not specially operate on tetra quantities, so non-quad-aligned
4110 registers can hold quadword quantities (except %o4 and %i4 because
4111 they cross fixed registers). */
4113 /* This points to either the 32 bit or the 64 bit version. */
4114 const int *hard_regno_mode_classes;
4116 static const int hard_32bit_mode_classes[] = {
4117 S_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
4118 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
4119 T_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
4120 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
4122 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4123 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4124 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4125 OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
4127 /* FP regs f32 to f63. Only the even numbered registers actually exist,
4128 and none can hold SFmode/SImode values. */
4129 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4130 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4131 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4132 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4135 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
4141 static const int hard_64bit_mode_classes[] = {
4142 D_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4143 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4144 T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4145 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4147 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4148 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4149 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4150 OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
4152 /* FP regs f32 to f63. Only the even numbered registers actually exist,
4153 and none can hold SFmode/SImode values. */
4154 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4155 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4156 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4157 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4160 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
4166 int sparc_mode_class [NUM_MACHINE_MODES];
4168 enum reg_class sparc_regno_reg_class[FIRST_PSEUDO_REGISTER];
4171 sparc_init_modes (void)
4175 for (i = 0; i < NUM_MACHINE_MODES; i++)
4177 switch (GET_MODE_CLASS (i))
4180 case MODE_PARTIAL_INT:
4181 case MODE_COMPLEX_INT:
4182 if (GET_MODE_SIZE (i) <= 4)
4183 sparc_mode_class[i] = 1 << (int) S_MODE;
4184 else if (GET_MODE_SIZE (i) == 8)
4185 sparc_mode_class[i] = 1 << (int) D_MODE;
4186 else if (GET_MODE_SIZE (i) == 16)
4187 sparc_mode_class[i] = 1 << (int) T_MODE;
4188 else if (GET_MODE_SIZE (i) == 32)
4189 sparc_mode_class[i] = 1 << (int) O_MODE;
4191 sparc_mode_class[i] = 0;
4193 case MODE_VECTOR_INT:
4194 if (GET_MODE_SIZE (i) <= 4)
4195 sparc_mode_class[i] = 1 << (int)SF_MODE;
4196 else if (GET_MODE_SIZE (i) == 8)
4197 sparc_mode_class[i] = 1 << (int)DF_MODE;
4200 case MODE_COMPLEX_FLOAT:
4201 if (GET_MODE_SIZE (i) <= 4)
4202 sparc_mode_class[i] = 1 << (int) SF_MODE;
4203 else if (GET_MODE_SIZE (i) == 8)
4204 sparc_mode_class[i] = 1 << (int) DF_MODE;
4205 else if (GET_MODE_SIZE (i) == 16)
4206 sparc_mode_class[i] = 1 << (int) TF_MODE;
4207 else if (GET_MODE_SIZE (i) == 32)
4208 sparc_mode_class[i] = 1 << (int) OF_MODE;
4210 sparc_mode_class[i] = 0;
4213 if (i == (int) CCFPmode || i == (int) CCFPEmode)
4214 sparc_mode_class[i] = 1 << (int) CCFP_MODE;
4216 sparc_mode_class[i] = 1 << (int) CC_MODE;
4219 sparc_mode_class[i] = 0;
4225 hard_regno_mode_classes = hard_64bit_mode_classes;
4227 hard_regno_mode_classes = hard_32bit_mode_classes;
4229 /* Initialize the array used by REGNO_REG_CLASS. */
4230 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
4232 if (i < 16 && TARGET_V8PLUS)
4233 sparc_regno_reg_class[i] = I64_REGS;
4234 else if (i < 32 || i == FRAME_POINTER_REGNUM)
4235 sparc_regno_reg_class[i] = GENERAL_REGS;
4237 sparc_regno_reg_class[i] = FP_REGS;
4239 sparc_regno_reg_class[i] = EXTRA_FP_REGS;
4241 sparc_regno_reg_class[i] = FPCC_REGS;
4243 sparc_regno_reg_class[i] = NO_REGS;
4247 /* Compute the frame size required by the function. This function is called
4248 during the reload pass and also by sparc_expand_prologue. */
4251 sparc_compute_frame_size (HOST_WIDE_INT size, int leaf_function_p)
4253 int outgoing_args_size = (current_function_outgoing_args_size
4254 + REG_PARM_STACK_SPACE (current_function_decl));
4255 int n_regs = 0; /* N_REGS is the number of 4-byte regs saved thus far. */
4260 for (i = 0; i < 8; i++)
4261 if (regs_ever_live[i] && ! call_used_regs[i])
4266 for (i = 0; i < 8; i += 2)
4267 if ((regs_ever_live[i] && ! call_used_regs[i])
4268 || (regs_ever_live[i+1] && ! call_used_regs[i+1]))
4272 for (i = 32; i < (TARGET_V9 ? 96 : 64); i += 2)
4273 if ((regs_ever_live[i] && ! call_used_regs[i])
4274 || (regs_ever_live[i+1] && ! call_used_regs[i+1]))
4277 /* Set up values for use in prologue and epilogue. */
4278 num_gfregs = n_regs;
4283 && current_function_outgoing_args_size == 0)
4284 actual_fsize = apparent_fsize = 0;
4287 /* We subtract STARTING_FRAME_OFFSET, remember it's negative. */
4288 apparent_fsize = (size - STARTING_FRAME_OFFSET + 7) & -8;
4289 apparent_fsize += n_regs * 4;
4290 actual_fsize = apparent_fsize + ((outgoing_args_size + 7) & -8);
4293 /* Make sure nothing can clobber our register windows.
4294 If a SAVE must be done, or there is a stack-local variable,
4295 the register window area must be allocated.
4296 ??? For v8 we apparently need an additional 8 bytes of reserved space. */
4297 if (! leaf_function_p || size > 0)
4298 actual_fsize += (16 * UNITS_PER_WORD) + (TARGET_ARCH64 ? 0 : 8);
4300 return SPARC_STACK_ALIGN (actual_fsize);
4303 /* Output any necessary .register pseudo-ops. */
4306 sparc_output_scratch_registers (FILE *file ATTRIBUTE_UNUSED)
4308 #ifdef HAVE_AS_REGISTER_PSEUDO_OP
4314 /* Check if %g[2367] were used without
4315 .register being printed for them already. */
4316 for (i = 2; i < 8; i++)
4318 if (regs_ever_live [i]
4319 && ! sparc_hard_reg_printed [i])
4321 sparc_hard_reg_printed [i] = 1;
4322 fprintf (file, "\t.register\t%%g%d, #scratch\n", i);
4329 /* Save/restore call-saved registers from LOW to HIGH at BASE+OFFSET
4330 as needed. LOW should be double-word aligned for 32-bit registers.
4331 Return the new OFFSET. */
4334 #define SORR_RESTORE 1
4337 save_or_restore_regs (int low, int high, rtx base, int offset, int action)
4342 if (TARGET_ARCH64 && high <= 32)
4344 for (i = low; i < high; i++)
4346 if (regs_ever_live[i] && ! call_used_regs[i])
4348 mem = gen_rtx_MEM (DImode, plus_constant (base, offset));
4349 set_mem_alias_set (mem, sparc_sr_alias_set);
4350 if (action == SORR_SAVE)
4352 insn = emit_move_insn (mem, gen_rtx_REG (DImode, i));
4353 RTX_FRAME_RELATED_P (insn) = 1;
4355 else /* action == SORR_RESTORE */
4356 emit_move_insn (gen_rtx_REG (DImode, i), mem);
4363 for (i = low; i < high; i += 2)
4365 bool reg0 = regs_ever_live[i] && ! call_used_regs[i];
4366 bool reg1 = regs_ever_live[i+1] && ! call_used_regs[i+1];
4367 enum machine_mode mode;
4372 mode = i < 32 ? DImode : DFmode;
4377 mode = i < 32 ? SImode : SFmode;
4382 mode = i < 32 ? SImode : SFmode;
4389 mem = gen_rtx_MEM (mode, plus_constant (base, offset));
4390 set_mem_alias_set (mem, sparc_sr_alias_set);
4391 if (action == SORR_SAVE)
4393 insn = emit_move_insn (mem, gen_rtx_REG (mode, regno));
4394 RTX_FRAME_RELATED_P (insn) = 1;
4396 else /* action == SORR_RESTORE */
4397 emit_move_insn (gen_rtx_REG (mode, regno), mem);
4399 /* Always preserve double-word alignment. */
4400 offset = (offset + 7) & -8;
4407 /* Emit code to save call-saved registers. */
4410 emit_save_regs (void)
4412 HOST_WIDE_INT offset;
4415 offset = frame_base_offset - apparent_fsize;
4417 if (offset < -4096 || offset + num_gfregs * 4 > 4096)
4419 /* ??? This might be optimized a little as %g1 might already have a
4420 value close enough that a single add insn will do. */
4421 /* ??? Although, all of this is probably only a temporary fix
4422 because if %g1 can hold a function result, then
4423 sparc_expand_epilogue will lose (the result will be
4425 base = gen_rtx_REG (Pmode, 1);
4426 emit_move_insn (base, GEN_INT (offset));
4427 emit_insn (gen_rtx_SET (VOIDmode,
4429 gen_rtx_PLUS (Pmode, frame_base_reg, base)));
4433 base = frame_base_reg;
4435 offset = save_or_restore_regs (0, 8, base, offset, SORR_SAVE);
4436 save_or_restore_regs (32, TARGET_V9 ? 96 : 64, base, offset, SORR_SAVE);
4439 /* Emit code to restore call-saved registers. */
4442 emit_restore_regs (void)
4444 HOST_WIDE_INT offset;
4447 offset = frame_base_offset - apparent_fsize;
4449 if (offset < -4096 || offset + num_gfregs * 4 > 4096 - 8 /*double*/)
4451 base = gen_rtx_REG (Pmode, 1);
4452 emit_move_insn (base, GEN_INT (offset));
4453 emit_insn (gen_rtx_SET (VOIDmode,
4455 gen_rtx_PLUS (Pmode, frame_base_reg, base)));
4459 base = frame_base_reg;
4461 offset = save_or_restore_regs (0, 8, base, offset, SORR_RESTORE);
4462 save_or_restore_regs (32, TARGET_V9 ? 96 : 64, base, offset, SORR_RESTORE);
4465 /* Emit an increment for the stack pointer. */
4468 emit_stack_pointer_increment (rtx increment)
4471 emit_insn (gen_adddi3 (stack_pointer_rtx, stack_pointer_rtx, increment));
4473 emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, increment));
4476 /* Emit a decrement for the stack pointer. */
4479 emit_stack_pointer_decrement (rtx decrement)
4482 emit_insn (gen_subdi3 (stack_pointer_rtx, stack_pointer_rtx, decrement));
4484 emit_insn (gen_subsi3 (stack_pointer_rtx, stack_pointer_rtx, decrement));
4487 /* Expand the function prologue. The prologue is responsible for reserving
4488 storage for the frame, saving the call-saved registers and loading the
4489 PIC register if needed. */
4492 sparc_expand_prologue (void)
4494 /* Compute a snapshot of current_function_uses_only_leaf_regs. Relying
4495 on the final value of the flag means deferring the prologue/epilogue
4496 expansion until just before the second scheduling pass, which is too
4497 late to emit multiple epilogues or return insns.
4499 Of course we are making the assumption that the value of the flag
4500 will not change between now and its final value. Of the three parts
4501 of the formula, only the last one can reasonably vary. Let's take a
4502 closer look, after assuming that the first two ones are set to true
4503 (otherwise the last value is effectively silenced).
4505 If only_leaf_regs_used returns false, the global predicate will also
4506 be false so the actual frame size calculated below will be positive.
4507 As a consequence, the save_register_window insn will be emitted in
4508 the instruction stream; now this insn explicitly references %fp
4509 which is not a leaf register so only_leaf_regs_used will always
4510 return false subsequently.
4512 If only_leaf_regs_used returns true, we hope that the subsequent
4513 optimization passes won't cause non-leaf registers to pop up. For
4514 example, the regrename pass has special provisions to not rename to
4515 non-leaf registers in a leaf function. */
4516 sparc_leaf_function_p
4517 = optimize > 0 && leaf_function_p () && only_leaf_regs_used ();
4519 /* Need to use actual_fsize, since we are also allocating
4520 space for our callee (and our own register save area). */
4522 = sparc_compute_frame_size (get_frame_size(), sparc_leaf_function_p);
4524 /* Advertise that the data calculated just above are now valid. */
4525 sparc_prologue_data_valid_p = true;
4527 if (sparc_leaf_function_p)
4529 frame_base_reg = stack_pointer_rtx;
4530 frame_base_offset = actual_fsize + SPARC_STACK_BIAS;
4534 frame_base_reg = hard_frame_pointer_rtx;
4535 frame_base_offset = SPARC_STACK_BIAS;
4538 if (actual_fsize == 0)
4540 else if (sparc_leaf_function_p)
4542 if (actual_fsize <= 4096)
4543 emit_stack_pointer_increment (GEN_INT (- actual_fsize));
4544 else if (actual_fsize <= 8192)
4546 emit_stack_pointer_increment (GEN_INT (-4096));
4547 emit_stack_pointer_increment (GEN_INT (4096 - actual_fsize));
4551 rtx reg = gen_rtx_REG (Pmode, 1);
4552 emit_move_insn (reg, GEN_INT (-actual_fsize));
4553 emit_stack_pointer_increment (reg);
4558 if (actual_fsize <= 4096)
4559 emit_insn (gen_save_register_window (GEN_INT (-actual_fsize)));
4560 else if (actual_fsize <= 8192)
4562 emit_insn (gen_save_register_window (GEN_INT (-4096)));
4563 emit_stack_pointer_increment (GEN_INT (4096 - actual_fsize));
4567 rtx reg = gen_rtx_REG (Pmode, 1);
4568 emit_move_insn (reg, GEN_INT (-actual_fsize));
4569 emit_insn (gen_save_register_window (reg));
4573 /* Call-saved registers are saved just above the outgoing argument area. */
4577 /* Load the PIC register if needed. */
4578 if (flag_pic && current_function_uses_pic_offset_table)
4579 load_pic_register ();
4582 /* This function generates the assembly code for function entry, which boils
4583 down to emitting the necessary .register directives. It also informs the
4584 DWARF-2 back-end on the layout of the frame.
4586 ??? Historical cruft: "On SPARC, move-double insns between fpu and cpu need
4587 an 8-byte block of memory. If any fpu reg is used in the function, we
4588 allocate such a block here, at the bottom of the frame, just in case it's
4589 needed." Could this explain the -8 in emit_restore_regs? */
4592 sparc_asm_function_prologue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
4594 /* Check that the assumption we made in sparc_expand_prologue is valid. */
4595 if (sparc_leaf_function_p != current_function_uses_only_leaf_regs)
4598 sparc_output_scratch_registers (file);
4600 if (dwarf2out_do_frame () && actual_fsize)
4602 char *label = dwarf2out_cfi_label ();
4604 /* The canonical frame address refers to the top of the frame. */
4605 dwarf2out_def_cfa (label,
4606 sparc_leaf_function_p
4607 ? STACK_POINTER_REGNUM
4608 : HARD_FRAME_POINTER_REGNUM,
4611 if (! sparc_leaf_function_p)
4613 /* Note the register window save. This tells the unwinder that
4614 it needs to restore the window registers from the previous
4615 frame's window save area at 0(cfa). */
4616 dwarf2out_window_save (label);
4618 /* The return address (-8) is now in %i7. */
4619 dwarf2out_return_reg (label, 31);
4624 /* Expand the function epilogue, either normal or part of a sibcall.
4625 We emit all the instructions except the return or the call. */
4628 sparc_expand_epilogue (void)
4631 emit_restore_regs ();
4633 if (actual_fsize == 0)
4635 else if (sparc_leaf_function_p)
4637 if (actual_fsize <= 4096)
4638 emit_stack_pointer_decrement (GEN_INT (- actual_fsize));
4639 else if (actual_fsize <= 8192)
4641 emit_stack_pointer_decrement (GEN_INT (-4096));
4642 emit_stack_pointer_decrement (GEN_INT (4096 - actual_fsize));
4646 rtx reg = gen_rtx_REG (Pmode, 1);
4647 emit_move_insn (reg, GEN_INT (-actual_fsize));
4648 emit_stack_pointer_decrement (reg);
4653 /* Return true if it is appropriate to emit `return' instructions in the
4654 body of a function. */
4657 sparc_can_use_return_insn_p (void)
4659 return sparc_prologue_data_valid_p
4660 && (actual_fsize == 0 || !sparc_leaf_function_p);
4663 /* This function generates the assembly code for function exit. */
4666 sparc_asm_function_epilogue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
4668 /* If code does not drop into the epilogue, we have to still output
4669 a dummy nop for the sake of sane backtraces. Otherwise, if the
4670 last two instructions of a function were "call foo; dslot;" this
4671 can make the return PC of foo (i.e. address of call instruction
4672 plus 8) point to the first instruction in the next function. */
4674 rtx insn, last_real_insn;
4676 insn = get_last_insn ();
4678 last_real_insn = prev_real_insn (insn);
4680 && GET_CODE (last_real_insn) == INSN
4681 && GET_CODE (PATTERN (last_real_insn)) == SEQUENCE)
4682 last_real_insn = XVECEXP (PATTERN (last_real_insn), 0, 0);
4684 if (last_real_insn && GET_CODE (last_real_insn) == CALL_INSN)
4685 fputs("\tnop\n", file);
4687 sparc_output_deferred_case_vectors ();
4690 /* Output a 'restore' instruction. */
4693 output_restore (rtx pat)
4699 fputs ("\t restore\n", asm_out_file);
4703 if (GET_CODE (pat) != SET)
4706 operands[0] = SET_DEST (pat);
4707 pat = SET_SRC (pat);
4709 switch (GET_CODE (pat))
4712 operands[1] = XEXP (pat, 0);
4713 operands[2] = XEXP (pat, 1);
4714 output_asm_insn (" restore %r1, %2, %Y0", operands);
4717 operands[1] = XEXP (pat, 0);
4718 operands[2] = XEXP (pat, 1);
4719 output_asm_insn (" restore %r1, %%lo(%a2), %Y0", operands);
4722 operands[1] = XEXP (pat, 0);
4723 if (XEXP (pat, 1) != const1_rtx)
4725 output_asm_insn (" restore %r1, %r1, %Y0", operands);
4729 output_asm_insn (" restore %%g0, %1, %Y0", operands);
4734 /* Output a return. */
4737 output_return (rtx insn)
4739 if (sparc_leaf_function_p)
4741 /* This is a leaf function so we don't have to bother restoring the
4742 register window, which frees us from dealing with the convoluted
4743 semantics of restore/return. We simply output the jump to the
4744 return address and the insn in the delay slot (if any). */
4746 if (current_function_calls_eh_return)
4749 return "jmp\t%%o7+%)%#";
4753 /* This is a regular function so we have to restore the register window.
4754 We may have a pending insn for the delay slot, which will be either
4755 combined with the 'restore' instruction or put in the delay slot of
4756 the 'return' instruction. */
4758 if (current_function_calls_eh_return)
4760 /* If the function uses __builtin_eh_return, the eh_return
4761 machinery occupies the delay slot. */
4765 if (! flag_delayed_branch)
4766 fputs ("\tadd\t%fp, %g1, %fp\n", asm_out_file);
4769 fputs ("\treturn\t%i7+8\n", asm_out_file);
4771 fputs ("\trestore\n\tjmp\t%o7+8\n", asm_out_file);
4773 if (flag_delayed_branch)
4774 fputs ("\t add\t%sp, %g1, %sp\n", asm_out_file);
4776 fputs ("\t nop\n", asm_out_file);
4778 else if (final_sequence)
4782 delay = NEXT_INSN (insn);
4786 pat = PATTERN (delay);
4788 if (TARGET_V9 && ! epilogue_renumber (&pat, 1))
4790 epilogue_renumber (&pat, 0);
4791 return "return\t%%i7+%)%#";
4795 output_asm_insn ("jmp\t%%i7+%)", NULL);
4796 output_restore (pat);
4797 PATTERN (delay) = gen_blockage ();
4798 INSN_CODE (delay) = -1;
4803 /* The delay slot is empty. */
4805 return "return\t%%i7+%)\n\t nop";
4806 else if (flag_delayed_branch)
4807 return "jmp\t%%i7+%)\n\t restore";
4809 return "restore\n\tjmp\t%%o7+%)\n\t nop";
4816 /* Output a sibling call. */
4819 output_sibcall (rtx insn, rtx call_operand)
4823 if (! flag_delayed_branch)
4826 operands[0] = call_operand;
4828 if (sparc_leaf_function_p)
4830 /* This is a leaf function so we don't have to bother restoring the
4831 register window. We simply output the jump to the function and
4832 the insn in the delay slot (if any). */
4834 if (LEAF_SIBCALL_SLOT_RESERVED_P && final_sequence)
4838 output_asm_insn ("sethi\t%%hi(%a0), %%g1\n\tjmp\t%%g1 + %%lo(%a0)%#",
4841 /* Use or with rs2 %%g0 instead of mov, so that as/ld can optimize
4842 it into branch if possible. */
4843 output_asm_insn ("or\t%%o7, %%g0, %%g1\n\tcall\t%a0, 0\n\t or\t%%g1, %%g0, %%o7",
4848 /* This is a regular function so we have to restore the register window.
4849 We may have a pending insn for the delay slot, which will be combined
4850 with the 'restore' instruction. */
4852 output_asm_insn ("call\t%a0, 0", operands);
4856 rtx delay = NEXT_INSN (insn);
4860 output_restore (PATTERN (delay));
4862 PATTERN (delay) = gen_blockage ();
4863 INSN_CODE (delay) = -1;
4866 output_restore (NULL_RTX);
4872 /* Functions for handling argument passing.
4874 For 32-bit, the first 6 args are normally in registers and the rest are
4875 pushed. Any arg that starts within the first 6 words is at least
4876 partially passed in a register unless its data type forbids.
4878 For 64-bit, the argument registers are laid out as an array of 16 elements
4879 and arguments are added sequentially. The first 6 int args and up to the
4880 first 16 fp args (depending on size) are passed in regs.
4882 Slot Stack Integral Float Float in structure Double Long Double
4883 ---- ----- -------- ----- ------------------ ------ -----------
4884 15 [SP+248] %f31 %f30,%f31 %d30
4885 14 [SP+240] %f29 %f28,%f29 %d28 %q28
4886 13 [SP+232] %f27 %f26,%f27 %d26
4887 12 [SP+224] %f25 %f24,%f25 %d24 %q24
4888 11 [SP+216] %f23 %f22,%f23 %d22
4889 10 [SP+208] %f21 %f20,%f21 %d20 %q20
4890 9 [SP+200] %f19 %f18,%f19 %d18
4891 8 [SP+192] %f17 %f16,%f17 %d16 %q16
4892 7 [SP+184] %f15 %f14,%f15 %d14
4893 6 [SP+176] %f13 %f12,%f13 %d12 %q12
4894 5 [SP+168] %o5 %f11 %f10,%f11 %d10
4895 4 [SP+160] %o4 %f9 %f8,%f9 %d8 %q8
4896 3 [SP+152] %o3 %f7 %f6,%f7 %d6
4897 2 [SP+144] %o2 %f5 %f4,%f5 %d4 %q4
4898 1 [SP+136] %o1 %f3 %f2,%f3 %d2
4899 0 [SP+128] %o0 %f1 %f0,%f1 %d0 %q0
4901 Here SP = %sp if -mno-stack-bias or %sp+stack_bias otherwise.
4903 Integral arguments are always passed as 64-bit quantities appropriately
4906 Passing of floating point values is handled as follows.
4907 If a prototype is in scope:
4908 If the value is in a named argument (i.e. not a stdarg function or a
4909 value not part of the `...') then the value is passed in the appropriate
4911 If the value is part of the `...' and is passed in one of the first 6
4912 slots then the value is passed in the appropriate int reg.
4913 If the value is part of the `...' and is not passed in one of the first 6
4914 slots then the value is passed in memory.
4915 If a prototype is not in scope:
4916 If the value is one of the first 6 arguments the value is passed in the
4917 appropriate integer reg and the appropriate fp reg.
4918 If the value is not one of the first 6 arguments the value is passed in
4919 the appropriate fp reg and in memory.
4922 Summary of the calling conventions implemented by GCC on SPARC:
4925 size argument return value
4927 small integer <4 int. reg. int. reg.
4928 word 4 int. reg. int. reg.
4929 double word 8 int. reg. int. reg.
4931 _Complex small integer <8 int. reg. int. reg.
4932 _Complex word 8 int. reg. int. reg.
4933 _Complex double word 16 memory int. reg.
4935 vector integer <=8 int. reg. FP reg.
4936 vector integer >8 memory memory
4938 float 4 int. reg. FP reg.
4939 double 8 int. reg. FP reg.
4940 long double 16 memory memory
4942 _Complex float 8 memory FP reg.
4943 _Complex double 16 memory FP reg.
4944 _Complex long double 32 memory FP reg.
4946 vector float any memory memory
4948 aggregate any memory memory
4953 size argument return value
4955 small integer <8 int. reg. int. reg.
4956 word 8 int. reg. int. reg.
4957 double word 16 int. reg. int. reg.
4959 _Complex small integer <16 int. reg. int. reg.
4960 _Complex word 16 int. reg. int. reg.
4961 _Complex double word 32 memory int. reg.
4963 vector integer <=16 FP reg. FP reg.
4964 vector integer 16<s<=32 memory FP reg.
4965 vector integer >32 memory memory
4967 float 4 FP reg. FP reg.
4968 double 8 FP reg. FP reg.
4969 long double 16 FP reg. FP reg.
4971 _Complex float 8 FP reg. FP reg.
4972 _Complex double 16 FP reg. FP reg.
4973 _Complex long double 32 memory FP reg.
4975 vector float <=16 FP reg. FP reg.
4976 vector float 16<s<=32 memory FP reg.
4977 vector float >32 memory memory
4979 aggregate <=16 reg. reg.
4980 aggregate 16<s<=32 memory reg.
4981 aggregate >32 memory memory
4985 Note #1: complex floating-point types follow the extended SPARC ABIs as
4986 implemented by the Sun compiler.
4988 Note #2: integral vector types follow the scalar floating-point types
4989 conventions to match what is implemented by the Sun VIS SDK.
4991 Note #3: floating-point vector types follow the aggregate types
4995 /* Maximum number of int regs for args. */
4996 #define SPARC_INT_ARG_MAX 6
4997 /* Maximum number of fp regs for args. */
4998 #define SPARC_FP_ARG_MAX 16
5000 #define ROUND_ADVANCE(SIZE) (((SIZE) + UNITS_PER_WORD - 1) / UNITS_PER_WORD)
5002 /* Handle the INIT_CUMULATIVE_ARGS macro.
5003 Initialize a variable CUM of type CUMULATIVE_ARGS
5004 for a call to a function whose data type is FNTYPE.
5005 For a library call, FNTYPE is 0. */
5008 init_cumulative_args (struct sparc_args *cum, tree fntype,
5009 rtx libname ATTRIBUTE_UNUSED,
5010 tree fndecl ATTRIBUTE_UNUSED)
5013 cum->prototype_p = fntype && TYPE_ARG_TYPES (fntype);
5014 cum->libcall_p = fntype == 0;
5017 /* Handle the TARGET_PROMOTE_PROTOTYPES target hook.
5018 When a prototype says `char' or `short', really pass an `int'. */
5021 sparc_promote_prototypes (tree fntype ATTRIBUTE_UNUSED)
5023 return TARGET_ARCH32 ? true : false;
5026 /* Handle the TARGET_STRICT_ARGUMENT_NAMING target hook. */
5029 sparc_strict_argument_naming (CUMULATIVE_ARGS *ca ATTRIBUTE_UNUSED)
5031 return TARGET_ARCH64 ? true : false;
5034 /* Scan the record type TYPE and return the following predicates:
5035 - INTREGS_P: the record contains at least one field or sub-field
5036 that is eligible for promotion in integer registers.
5037 - FP_REGS_P: the record contains at least one field or sub-field
5038 that is eligible for promotion in floating-point registers.
5039 - PACKED_P: the record contains at least one field that is packed.
5041 Sub-fields are not taken into account for the PACKED_P predicate. */
5044 scan_record_type (tree type, int *intregs_p, int *fpregs_p, int *packed_p)
5048 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5050 if (TREE_CODE (field) == FIELD_DECL)
5052 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
5053 scan_record_type (TREE_TYPE (field), intregs_p, fpregs_p, 0);
5054 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
5055 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
5061 if (packed_p && DECL_PACKED (field))
5067 /* Compute the slot number to pass an argument in.
5068 Return the slot number or -1 if passing on the stack.
5070 CUM is a variable of type CUMULATIVE_ARGS which gives info about
5071 the preceding args and about the function being called.
5072 MODE is the argument's machine mode.
5073 TYPE is the data type of the argument (as a tree).
5074 This is null for libcalls where that information may
5076 NAMED is nonzero if this argument is a named parameter
5077 (otherwise it is an extra parameter matching an ellipsis).
5078 INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG.
5079 *PREGNO records the register number to use if scalar type.
5080 *PPADDING records the amount of padding needed in words. */
5083 function_arg_slotno (const struct sparc_args *cum, enum machine_mode mode,
5084 tree type, int named, int incoming_p,
5085 int *pregno, int *ppadding)
5087 int regbase = (incoming_p
5088 ? SPARC_INCOMING_INT_ARG_FIRST
5089 : SPARC_OUTGOING_INT_ARG_FIRST);
5090 int slotno = cum->words;
5091 enum mode_class mclass;
5096 if (type && TREE_ADDRESSABLE (type))
5102 && TYPE_ALIGN (type) % PARM_BOUNDARY != 0)
5105 /* For SPARC64, objects requiring 16-byte alignment get it. */
5107 && GET_MODE_ALIGNMENT (mode) >= 2 * BITS_PER_WORD
5108 && (slotno & 1) != 0)
5109 slotno++, *ppadding = 1;
5111 mclass = GET_MODE_CLASS (mode);
5112 if (type && TREE_CODE (type) == VECTOR_TYPE)
5114 /* Vector types deserve special treatment because they are
5115 polymorphic wrt their mode, depending upon whether VIS
5116 instructions are enabled. */
5117 if (TREE_CODE (TREE_TYPE (type)) == REAL_TYPE)
5119 /* The SPARC port defines no floating-point vector modes. */
5120 if (mode != BLKmode)
5125 /* Integral vector types should either have a vector
5126 mode or an integral mode, because we are guaranteed
5127 by pass_by_reference that their size is not greater
5128 than 16 bytes and TImode is 16-byte wide. */
5129 if (mode == BLKmode)
5132 /* Vector integers are handled like floats according to
5134 mclass = MODE_FLOAT;
5141 case MODE_COMPLEX_FLOAT:
5142 if (TARGET_ARCH64 && TARGET_FPU && named)
5144 if (slotno >= SPARC_FP_ARG_MAX)
5146 regno = SPARC_FP_ARG_FIRST + slotno * 2;
5147 /* Arguments filling only one single FP register are
5148 right-justified in the outer double FP register. */
5149 if (GET_MODE_SIZE (mode) <= 4)
5156 case MODE_COMPLEX_INT:
5157 if (slotno >= SPARC_INT_ARG_MAX)
5159 regno = regbase + slotno;
5163 if (mode == VOIDmode)
5164 /* MODE is VOIDmode when generating the actual call. */
5167 if (mode != BLKmode)
5170 /* For SPARC64, objects requiring 16-byte alignment get it. */
5173 && TYPE_ALIGN (type) >= 2 * BITS_PER_WORD
5174 && (slotno & 1) != 0)
5175 slotno++, *ppadding = 1;
5177 if (TARGET_ARCH32 || !type || (TREE_CODE (type) == UNION_TYPE))
5179 if (slotno >= SPARC_INT_ARG_MAX)
5181 regno = regbase + slotno;
5183 else /* TARGET_ARCH64 && type */
5185 int intregs_p = 0, fpregs_p = 0, packed_p = 0;
5187 /* First see what kinds of registers we would need. */
5188 if (TREE_CODE (type) == VECTOR_TYPE)
5191 scan_record_type (type, &intregs_p, &fpregs_p, &packed_p);
5193 /* The ABI obviously doesn't specify how packed structures
5194 are passed. These are defined to be passed in int regs
5195 if possible, otherwise memory. */
5196 if (packed_p || !named)
5197 fpregs_p = 0, intregs_p = 1;
5199 /* If all arg slots are filled, then must pass on stack. */
5200 if (fpregs_p && slotno >= SPARC_FP_ARG_MAX)
5203 /* If there are only int args and all int arg slots are filled,
5204 then must pass on stack. */
5205 if (!fpregs_p && intregs_p && slotno >= SPARC_INT_ARG_MAX)
5208 /* Note that even if all int arg slots are filled, fp members may
5209 still be passed in regs if such regs are available.
5210 *PREGNO isn't set because there may be more than one, it's up
5211 to the caller to compute them. */
5224 /* Handle recursive register counting for structure field layout. */
5226 struct function_arg_record_value_parms
5228 rtx ret; /* return expression being built. */
5229 int slotno; /* slot number of the argument. */
5230 int named; /* whether the argument is named. */
5231 int regbase; /* regno of the base register. */
5232 int stack; /* 1 if part of the argument is on the stack. */
5233 int intoffset; /* offset of the first pending integer field. */
5234 unsigned int nregs; /* number of words passed in registers. */
5237 static void function_arg_record_value_3
5238 (HOST_WIDE_INT, struct function_arg_record_value_parms *);
5239 static void function_arg_record_value_2
5240 (tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
5241 static void function_arg_record_value_1
5242 (tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
5243 static rtx function_arg_record_value (tree, enum machine_mode, int, int, int);
5244 static rtx function_arg_union_value (int, enum machine_mode, int, int);
5246 /* A subroutine of function_arg_record_value. Traverse the structure
5247 recursively and determine how many registers will be required. */
5250 function_arg_record_value_1 (tree type, HOST_WIDE_INT startbitpos,
5251 struct function_arg_record_value_parms *parms,
5256 /* We need to compute how many registers are needed so we can
5257 allocate the PARALLEL but before we can do that we need to know
5258 whether there are any packed fields. The ABI obviously doesn't
5259 specify how structures are passed in this case, so they are
5260 defined to be passed in int regs if possible, otherwise memory,
5261 regardless of whether there are fp values present. */
5264 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5266 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
5273 /* Compute how many registers we need. */
5274 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5276 if (TREE_CODE (field) == FIELD_DECL)
5278 HOST_WIDE_INT bitpos = startbitpos;
5280 if (DECL_SIZE (field) != 0)
5282 if (integer_zerop (DECL_SIZE (field)))
5285 if (host_integerp (bit_position (field), 1))
5286 bitpos += int_bit_position (field);
5289 /* ??? FIXME: else assume zero offset. */
5291 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
5292 function_arg_record_value_1 (TREE_TYPE (field),
5296 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
5297 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
5302 if (parms->intoffset != -1)
5304 unsigned int startbit, endbit;
5305 int intslots, this_slotno;
5307 startbit = parms->intoffset & -BITS_PER_WORD;
5308 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
5310 intslots = (endbit - startbit) / BITS_PER_WORD;
5311 this_slotno = parms->slotno + parms->intoffset
5314 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
5316 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
5317 /* We need to pass this field on the stack. */
5321 parms->nregs += intslots;
5322 parms->intoffset = -1;
5325 /* There's no need to check this_slotno < SPARC_FP_ARG MAX.
5326 If it wasn't true we wouldn't be here. */
5327 if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE
5328 && DECL_MODE (field) == BLKmode)
5329 parms->nregs += TYPE_VECTOR_SUBPARTS (TREE_TYPE (field));
5330 else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
5337 if (parms->intoffset == -1)
5338 parms->intoffset = bitpos;
5344 /* A subroutine of function_arg_record_value. Assign the bits of the
5345 structure between parms->intoffset and bitpos to integer registers. */
5348 function_arg_record_value_3 (HOST_WIDE_INT bitpos,
5349 struct function_arg_record_value_parms *parms)
5351 enum machine_mode mode;
5353 unsigned int startbit, endbit;
5354 int this_slotno, intslots, intoffset;
5357 if (parms->intoffset == -1)
5360 intoffset = parms->intoffset;
5361 parms->intoffset = -1;
5363 startbit = intoffset & -BITS_PER_WORD;
5364 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
5365 intslots = (endbit - startbit) / BITS_PER_WORD;
5366 this_slotno = parms->slotno + intoffset / BITS_PER_WORD;
5368 intslots = MIN (intslots, SPARC_INT_ARG_MAX - this_slotno);
5372 /* If this is the trailing part of a word, only load that much into
5373 the register. Otherwise load the whole register. Note that in
5374 the latter case we may pick up unwanted bits. It's not a problem
5375 at the moment but may wish to revisit. */
5377 if (intoffset % BITS_PER_WORD != 0)
5378 mode = smallest_mode_for_size (BITS_PER_WORD - intoffset % BITS_PER_WORD,
5383 intoffset /= BITS_PER_UNIT;
5386 regno = parms->regbase + this_slotno;
5387 reg = gen_rtx_REG (mode, regno);
5388 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
5389 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (intoffset));
5392 intoffset = (intoffset | (UNITS_PER_WORD-1)) + 1;
5397 while (intslots > 0);
5400 /* A subroutine of function_arg_record_value. Traverse the structure
5401 recursively and assign bits to floating point registers. Track which
5402 bits in between need integer registers; invoke function_arg_record_value_3
5403 to make that happen. */
5406 function_arg_record_value_2 (tree type, HOST_WIDE_INT startbitpos,
5407 struct function_arg_record_value_parms *parms,
5413 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5415 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
5422 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5424 if (TREE_CODE (field) == FIELD_DECL)
5426 HOST_WIDE_INT bitpos = startbitpos;
5428 if (DECL_SIZE (field) != 0)
5430 if (integer_zerop (DECL_SIZE (field)))
5433 if (host_integerp (bit_position (field), 1))
5434 bitpos += int_bit_position (field);
5437 /* ??? FIXME: else assume zero offset. */
5439 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
5440 function_arg_record_value_2 (TREE_TYPE (field),
5444 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
5445 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
5450 int this_slotno = parms->slotno + bitpos / BITS_PER_WORD;
5451 int regno, nregs, pos;
5452 enum machine_mode mode = DECL_MODE (field);
5455 function_arg_record_value_3 (bitpos, parms);
5457 if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE
5460 mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field)));
5461 nregs = TYPE_VECTOR_SUBPARTS (TREE_TYPE (field));
5463 else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
5465 mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field)));
5471 regno = SPARC_FP_ARG_FIRST + this_slotno * 2;
5472 if (GET_MODE_SIZE (mode) <= 4 && (bitpos & 32) != 0)
5474 reg = gen_rtx_REG (mode, regno);
5475 pos = bitpos / BITS_PER_UNIT;
5476 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
5477 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos));
5481 regno += GET_MODE_SIZE (mode) / 4;
5482 reg = gen_rtx_REG (mode, regno);
5483 pos += GET_MODE_SIZE (mode);
5484 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
5485 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos));
5491 if (parms->intoffset == -1)
5492 parms->intoffset = bitpos;
5498 /* Used by function_arg and function_value to implement the complex
5499 conventions of the 64-bit ABI for passing and returning structures.
5500 Return an expression valid as a return value for the two macros
5501 FUNCTION_ARG and FUNCTION_VALUE.
5503 TYPE is the data type of the argument (as a tree).
5504 This is null for libcalls where that information may
5506 MODE is the argument's machine mode.
5507 SLOTNO is the index number of the argument's slot in the parameter array.
5508 NAMED is nonzero if this argument is a named parameter
5509 (otherwise it is an extra parameter matching an ellipsis).
5510 REGBASE is the regno of the base register for the parameter array. */
5513 function_arg_record_value (tree type, enum machine_mode mode,
5514 int slotno, int named, int regbase)
5516 HOST_WIDE_INT typesize = int_size_in_bytes (type);
5517 struct function_arg_record_value_parms parms;
5520 parms.ret = NULL_RTX;
5521 parms.slotno = slotno;
5522 parms.named = named;
5523 parms.regbase = regbase;
5526 /* Compute how many registers we need. */
5528 parms.intoffset = 0;
5529 function_arg_record_value_1 (type, 0, &parms, false);
5531 /* Take into account pending integer fields. */
5532 if (parms.intoffset != -1)
5534 unsigned int startbit, endbit;
5535 int intslots, this_slotno;
5537 startbit = parms.intoffset & -BITS_PER_WORD;
5538 endbit = (typesize*BITS_PER_UNIT + BITS_PER_WORD - 1) & -BITS_PER_WORD;
5539 intslots = (endbit - startbit) / BITS_PER_WORD;
5540 this_slotno = slotno + parms.intoffset / BITS_PER_WORD;
5542 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
5544 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
5545 /* We need to pass this field on the stack. */
5549 parms.nregs += intslots;
5551 nregs = parms.nregs;
5553 /* Allocate the vector and handle some annoying special cases. */
5556 /* ??? Empty structure has no value? Duh? */
5559 /* Though there's nothing really to store, return a word register
5560 anyway so the rest of gcc doesn't go nuts. Returning a PARALLEL
5561 leads to breakage due to the fact that there are zero bytes to
5563 return gen_rtx_REG (mode, regbase);
5567 /* ??? C++ has structures with no fields, and yet a size. Give up
5568 for now and pass everything back in integer registers. */
5569 nregs = (typesize + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
5571 if (nregs + slotno > SPARC_INT_ARG_MAX)
5572 nregs = SPARC_INT_ARG_MAX - slotno;
5577 parms.ret = gen_rtx_PARALLEL (mode, rtvec_alloc (parms.stack + nregs));
5579 /* If at least one field must be passed on the stack, generate
5580 (parallel [(expr_list (nil) ...) ...]) so that all fields will
5581 also be passed on the stack. We can't do much better because the
5582 semantics of FUNCTION_ARG_PARTIAL_NREGS doesn't handle the case
5583 of structures for which the fields passed exclusively in registers
5584 are not at the beginning of the structure. */
5586 XVECEXP (parms.ret, 0, 0)
5587 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
5589 /* Fill in the entries. */
5591 parms.intoffset = 0;
5592 function_arg_record_value_2 (type, 0, &parms, false);
5593 function_arg_record_value_3 (typesize * BITS_PER_UNIT, &parms);
5595 if (parms.nregs != nregs)
5601 /* Used by function_arg and function_value to implement the conventions
5602 of the 64-bit ABI for passing and returning unions.
5603 Return an expression valid as a return value for the two macros
5604 FUNCTION_ARG and FUNCTION_VALUE.
5606 SIZE is the size in bytes of the union.
5607 MODE is the argument's machine mode.
5608 REGNO is the hard register the union will be passed in. */
5611 function_arg_union_value (int size, enum machine_mode mode, int slotno,
5614 int nwords = ROUND_ADVANCE (size), i;
5617 /* See comment in previous function for empty structures. */
5619 return gen_rtx_REG (mode, regno);
5621 if (slotno == SPARC_INT_ARG_MAX - 1)
5624 regs = gen_rtx_PARALLEL (mode, rtvec_alloc (nwords));
5626 for (i = 0; i < nwords; i++)
5628 /* Unions are passed left-justified. */
5629 XVECEXP (regs, 0, i)
5630 = gen_rtx_EXPR_LIST (VOIDmode,
5631 gen_rtx_REG (word_mode, regno),
5632 GEN_INT (UNITS_PER_WORD * i));
5639 /* Used by function_arg and function_value to implement the conventions
5640 for passing and returning large (BLKmode) vectors.
5641 Return an expression valid as a return value for the two macros
5642 FUNCTION_ARG and FUNCTION_VALUE.
5644 SIZE is the size in bytes of the vector.
5645 BASE_MODE is the argument's base machine mode.
5646 REGNO is the FP hard register the vector will be passed in. */
5649 function_arg_vector_value (int size, enum machine_mode base_mode, int regno)
5651 unsigned short base_mode_size = GET_MODE_SIZE (base_mode);
5652 int nregs = size / base_mode_size, i;
5655 regs = gen_rtx_PARALLEL (BLKmode, rtvec_alloc (nregs));
5657 for (i = 0; i < nregs; i++)
5659 XVECEXP (regs, 0, i)
5660 = gen_rtx_EXPR_LIST (VOIDmode,
5661 gen_rtx_REG (base_mode, regno),
5662 GEN_INT (base_mode_size * i));
5663 regno += base_mode_size / 4;
5669 /* Handle the FUNCTION_ARG macro.
5670 Determine where to put an argument to a function.
5671 Value is zero to push the argument on the stack,
5672 or a hard register in which to store the argument.
5674 CUM is a variable of type CUMULATIVE_ARGS which gives info about
5675 the preceding args and about the function being called.
5676 MODE is the argument's machine mode.
5677 TYPE is the data type of the argument (as a tree).
5678 This is null for libcalls where that information may
5680 NAMED is nonzero if this argument is a named parameter
5681 (otherwise it is an extra parameter matching an ellipsis).
5682 INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG. */
5685 function_arg (const struct sparc_args *cum, enum machine_mode mode,
5686 tree type, int named, int incoming_p)
5688 int regbase = (incoming_p
5689 ? SPARC_INCOMING_INT_ARG_FIRST
5690 : SPARC_OUTGOING_INT_ARG_FIRST);
5691 int slotno, regno, padding;
5692 enum mode_class mclass = GET_MODE_CLASS (mode);
5695 slotno = function_arg_slotno (cum, mode, type, named, incoming_p,
5703 reg = gen_rtx_REG (mode, regno);
5707 if (type && TREE_CODE (type) == RECORD_TYPE)
5709 /* Structures up to 16 bytes in size are passed in arg slots on the
5710 stack and are promoted to registers where possible. */
5712 if (int_size_in_bytes (type) > 16)
5713 abort (); /* shouldn't get here */
5715 return function_arg_record_value (type, mode, slotno, named, regbase);
5717 else if (type && TREE_CODE (type) == UNION_TYPE)
5719 HOST_WIDE_INT size = int_size_in_bytes (type);
5722 abort (); /* shouldn't get here */
5724 return function_arg_union_value (size, mode, slotno, regno);
5726 else if (type && TREE_CODE (type) == VECTOR_TYPE)
5728 /* Vector types deserve special treatment because they are
5729 polymorphic wrt their mode, depending upon whether VIS
5730 instructions are enabled. */
5731 HOST_WIDE_INT size = int_size_in_bytes (type);
5734 abort (); /* shouldn't get here */
5736 if (mode == BLKmode)
5737 return function_arg_vector_value (size,
5738 TYPE_MODE (TREE_TYPE (type)),
5739 SPARC_FP_ARG_FIRST + 2*slotno);
5741 mclass = MODE_FLOAT;
5744 /* v9 fp args in reg slots beyond the int reg slots get passed in regs
5745 but also have the slot allocated for them.
5746 If no prototype is in scope fp values in register slots get passed
5747 in two places, either fp regs and int regs or fp regs and memory. */
5748 if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT)
5749 && SPARC_FP_REG_P (regno))
5751 reg = gen_rtx_REG (mode, regno);
5752 if (cum->prototype_p || cum->libcall_p)
5754 /* "* 2" because fp reg numbers are recorded in 4 byte
5757 /* ??? This will cause the value to be passed in the fp reg and
5758 in the stack. When a prototype exists we want to pass the
5759 value in the reg but reserve space on the stack. That's an
5760 optimization, and is deferred [for a bit]. */
5761 if ((regno - SPARC_FP_ARG_FIRST) >= SPARC_INT_ARG_MAX * 2)
5762 return gen_rtx_PARALLEL (mode,
5764 gen_rtx_EXPR_LIST (VOIDmode,
5765 NULL_RTX, const0_rtx),
5766 gen_rtx_EXPR_LIST (VOIDmode,
5770 /* ??? It seems that passing back a register even when past
5771 the area declared by REG_PARM_STACK_SPACE will allocate
5772 space appropriately, and will not copy the data onto the
5773 stack, exactly as we desire.
5775 This is due to locate_and_pad_parm being called in
5776 expand_call whenever reg_parm_stack_space > 0, which
5777 while beneficial to our example here, would seem to be
5778 in error from what had been intended. Ho hum... -- r~ */
5786 if ((regno - SPARC_FP_ARG_FIRST) < SPARC_INT_ARG_MAX * 2)
5790 /* On incoming, we don't need to know that the value
5791 is passed in %f0 and %i0, and it confuses other parts
5792 causing needless spillage even on the simplest cases. */
5796 intreg = (SPARC_OUTGOING_INT_ARG_FIRST
5797 + (regno - SPARC_FP_ARG_FIRST) / 2);
5799 v0 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
5800 v1 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_REG (mode, intreg),
5802 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
5806 v0 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
5807 v1 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
5808 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
5814 /* Scalar or complex int. */
5815 reg = gen_rtx_REG (mode, regno);
5821 /* Handle the FUNCTION_ARG_PARTIAL_NREGS macro.
5822 For an arg passed partly in registers and partly in memory,
5823 this is the number of registers used.
5824 For args passed entirely in registers or entirely in memory, zero.
5826 Any arg that starts in the first 6 regs but won't entirely fit in them
5827 needs partial registers on v8. On v9, structures with integer
5828 values in arg slots 5,6 will be passed in %o5 and SP+176, and complex fp
5829 values that begin in the last fp reg [where "last fp reg" varies with the
5830 mode] will be split between that reg and memory. */
5833 function_arg_partial_nregs (const struct sparc_args *cum,
5834 enum machine_mode mode, tree type, int named)
5836 int slotno, regno, padding;
5838 /* We pass 0 for incoming_p here, it doesn't matter. */
5839 slotno = function_arg_slotno (cum, mode, type, named, 0, ®no, &padding);
5846 if ((slotno + (mode == BLKmode
5847 ? ROUND_ADVANCE (int_size_in_bytes (type))
5848 : ROUND_ADVANCE (GET_MODE_SIZE (mode))))
5849 > SPARC_INT_ARG_MAX)
5850 return SPARC_INT_ARG_MAX - slotno;
5854 /* We are guaranteed by pass_by_reference that the size of the
5855 argument is not greater than 16 bytes, so we only need to
5856 return 1 if the argument is partially passed in registers. */
5858 if (type && AGGREGATE_TYPE_P (type))
5860 int size = int_size_in_bytes (type);
5862 if (size > UNITS_PER_WORD
5863 && slotno == SPARC_INT_ARG_MAX - 1)
5866 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_INT
5867 || (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT
5868 && ! (TARGET_FPU && named)))
5870 /* The complex types are passed as packed types. */
5871 if (GET_MODE_SIZE (mode) > UNITS_PER_WORD
5872 && slotno == SPARC_INT_ARG_MAX - 1)
5875 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT)
5877 if ((slotno + GET_MODE_SIZE (mode) / UNITS_PER_WORD)
5886 /* Handle the TARGET_PASS_BY_REFERENCE target hook.
5887 Specify whether to pass the argument by reference. */
5890 sparc_pass_by_reference (CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED,
5891 enum machine_mode mode, tree type,
5892 bool named ATTRIBUTE_UNUSED)
5896 /* Original SPARC 32-bit ABI says that structures and unions,
5897 and quad-precision floats are passed by reference. For Pascal,
5898 also pass arrays by reference. All other base types are passed
5901 Extended ABI (as implemented by the Sun compiler) says that all
5902 complex floats are passed by reference. Pass complex integers
5903 in registers up to 8 bytes. More generally, enforce the 2-word
5904 cap for passing arguments in registers.
5906 Vector ABI (as implemented by the Sun VIS SDK) says that vector
5907 integers are passed like floats of the same size, that is in
5908 registers up to 8 bytes. Pass all vector floats by reference
5909 like structure and unions. */
5910 return ((type && (AGGREGATE_TYPE_P (type) || VECTOR_FLOAT_TYPE_P (type)))
5912 /* Catch CDImode, TFmode, DCmode and TCmode. */
5913 || GET_MODE_SIZE (mode) > 8
5915 && TREE_CODE (type) == VECTOR_TYPE
5916 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8));
5920 /* Original SPARC 64-bit ABI says that structures and unions
5921 smaller than 16 bytes are passed in registers, as well as
5922 all other base types. For Pascal, pass arrays by reference.
5924 Extended ABI (as implemented by the Sun compiler) says that
5925 complex floats are passed in registers up to 16 bytes. Pass
5926 all complex integers in registers up to 16 bytes. More generally,
5927 enforce the 2-word cap for passing arguments in registers.
5929 Vector ABI (as implemented by the Sun VIS SDK) says that vector
5930 integers are passed like floats of the same size, that is in
5931 registers (up to 16 bytes). Pass all vector floats like structure
5933 return ((type && TREE_CODE (type) == ARRAY_TYPE)
5935 && (AGGREGATE_TYPE_P (type) || TREE_CODE (type) == VECTOR_TYPE)
5936 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 16)
5937 /* Catch CTImode and TCmode. */
5938 || GET_MODE_SIZE (mode) > 16);
5942 /* Handle the FUNCTION_ARG_ADVANCE macro.
5943 Update the data in CUM to advance over an argument
5944 of mode MODE and data type TYPE.
5945 TYPE is null for libcalls where that information may not be available. */
5948 function_arg_advance (struct sparc_args *cum, enum machine_mode mode,
5949 tree type, int named)
5951 int slotno, regno, padding;
5953 /* We pass 0 for incoming_p here, it doesn't matter. */
5954 slotno = function_arg_slotno (cum, mode, type, named, 0, ®no, &padding);
5956 /* If register required leading padding, add it. */
5958 cum->words += padding;
5962 cum->words += (mode != BLKmode
5963 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
5964 : ROUND_ADVANCE (int_size_in_bytes (type)));
5968 if (type && AGGREGATE_TYPE_P (type))
5970 int size = int_size_in_bytes (type);
5974 else if (size <= 16)
5976 else /* passed by reference */
5981 cum->words += (mode != BLKmode
5982 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
5983 : ROUND_ADVANCE (int_size_in_bytes (type)));
5988 /* Handle the FUNCTION_ARG_PADDING macro.
5989 For the 64 bit ABI structs are always stored left shifted in their
5993 function_arg_padding (enum machine_mode mode, tree type)
5995 if (TARGET_ARCH64 && type != 0 && AGGREGATE_TYPE_P (type))
5998 /* Fall back to the default. */
5999 return DEFAULT_FUNCTION_ARG_PADDING (mode, type);
6002 /* Handle the TARGET_RETURN_IN_MEMORY target hook.
6003 Specify whether to return the return value in memory. */
6006 sparc_return_in_memory (tree type, tree fntype ATTRIBUTE_UNUSED)
6009 /* Original SPARC 32-bit ABI says that structures and unions,
6010 and quad-precision floats are returned in memory. All other
6011 base types are returned in registers.
6013 Extended ABI (as implemented by the Sun compiler) says that
6014 all complex floats are returned in registers (8 FP registers
6015 at most for '_Complex long double'). Return all complex integers
6016 in registers (4 at most for '_Complex long long').
6018 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6019 integers are returned like floats of the same size, that is in
6020 registers up to 8 bytes and in memory otherwise. Return all
6021 vector floats in memory like structure and unions; note that
6022 they always have BLKmode like the latter. */
6023 return (TYPE_MODE (type) == BLKmode
6024 || TYPE_MODE (type) == TFmode
6025 || (TREE_CODE (type) == VECTOR_TYPE
6026 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8));
6028 /* Original SPARC 64-bit ABI says that structures and unions
6029 smaller than 32 bytes are returned in registers, as well as
6030 all other base types.
6032 Extended ABI (as implemented by the Sun compiler) says that all
6033 complex floats are returned in registers (8 FP registers at most
6034 for '_Complex long double'). Return all complex integers in
6035 registers (4 at most for '_Complex TItype').
6037 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6038 integers are returned like floats of the same size, that is in
6039 registers. Return all vector floats like structure and unions;
6040 note that they always have BLKmode like the latter. */
6041 return ((TYPE_MODE (type) == BLKmode
6042 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 32));
6045 /* Handle the TARGET_STRUCT_VALUE target hook.
6046 Return where to find the structure return value address. */
6049 sparc_struct_value_rtx (tree fndecl ATTRIBUTE_UNUSED, int incoming)
6056 return gen_rtx_MEM (Pmode, plus_constant (frame_pointer_rtx,
6057 STRUCT_VALUE_OFFSET));
6059 return gen_rtx_MEM (Pmode, plus_constant (stack_pointer_rtx,
6060 STRUCT_VALUE_OFFSET));
6064 /* Handle FUNCTION_VALUE, FUNCTION_OUTGOING_VALUE, and LIBCALL_VALUE macros.
6065 For v9, function return values are subject to the same rules as arguments,
6066 except that up to 32 bytes may be returned in registers. */
6069 function_value (tree type, enum machine_mode mode, int incoming_p)
6071 /* Beware that the two values are swapped here wrt function_arg. */
6072 int regbase = (incoming_p
6073 ? SPARC_OUTGOING_INT_ARG_FIRST
6074 : SPARC_INCOMING_INT_ARG_FIRST);
6075 enum mode_class mclass = GET_MODE_CLASS (mode);
6078 if (type && TREE_CODE (type) == VECTOR_TYPE)
6080 /* Vector types deserve special treatment because they are
6081 polymorphic wrt their mode, depending upon whether VIS
6082 instructions are enabled. */
6083 HOST_WIDE_INT size = int_size_in_bytes (type);
6085 if ((TARGET_ARCH32 && size > 8) || (TARGET_ARCH64 && size > 32))
6086 abort (); /* shouldn't get here */
6088 if (mode == BLKmode)
6089 return function_arg_vector_value (size,
6090 TYPE_MODE (TREE_TYPE (type)),
6091 SPARC_FP_ARG_FIRST);
6093 mclass = MODE_FLOAT;
6095 else if (type && TARGET_ARCH64)
6097 if (TREE_CODE (type) == RECORD_TYPE)
6099 /* Structures up to 32 bytes in size are passed in registers,
6100 promoted to fp registers where possible. */
6102 if (int_size_in_bytes (type) > 32)
6103 abort (); /* shouldn't get here */
6105 return function_arg_record_value (type, mode, 0, 1, regbase);
6107 else if (TREE_CODE (type) == UNION_TYPE)
6109 HOST_WIDE_INT size = int_size_in_bytes (type);
6112 abort (); /* shouldn't get here */
6114 return function_arg_union_value (size, mode, 0, regbase);
6116 else if (AGGREGATE_TYPE_P (type))
6118 /* All other aggregate types are passed in an integer register
6119 in a mode corresponding to the size of the type. */
6120 HOST_WIDE_INT bytes = int_size_in_bytes (type);
6123 abort (); /* shouldn't get here */
6125 mode = mode_for_size (bytes * BITS_PER_UNIT, MODE_INT, 0);
6127 /* ??? We probably should have made the same ABI change in
6128 3.4.0 as the one we made for unions. The latter was
6129 required by the SCD though, while the former is not
6130 specified, so we favored compatibility and efficiency.
6132 Now we're stuck for aggregates larger than 16 bytes,
6133 because OImode vanished in the meantime. Let's not
6134 try to be unduly clever, and simply follow the ABI
6135 for unions in that case. */
6136 if (mode == BLKmode)
6137 return function_arg_union_value (bytes, mode, 0, regbase);
6141 else if (mclass == MODE_INT
6142 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
6146 if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT)
6148 regno = SPARC_FP_ARG_FIRST;
6152 return gen_rtx_REG (mode, regno);
6155 /* Do what is necessary for `va_start'. We look at the current function
6156 to determine if stdarg or varargs is used and return the address of
6157 the first unnamed parameter. */
6160 sparc_builtin_saveregs (void)
6162 int first_reg = current_function_args_info.words;
6166 for (regno = first_reg; regno < SPARC_INT_ARG_MAX; regno++)
6167 emit_move_insn (gen_rtx_MEM (word_mode,
6168 gen_rtx_PLUS (Pmode,
6170 GEN_INT (FIRST_PARM_OFFSET (0)
6173 gen_rtx_REG (word_mode,
6174 SPARC_INCOMING_INT_ARG_FIRST + regno));
6176 address = gen_rtx_PLUS (Pmode,
6178 GEN_INT (FIRST_PARM_OFFSET (0)
6179 + UNITS_PER_WORD * first_reg));
6184 /* Implement `va_start' for stdarg. */
6187 sparc_va_start (tree valist, rtx nextarg)
6189 nextarg = expand_builtin_saveregs ();
6190 std_expand_builtin_va_start (valist, nextarg);
6193 /* Implement `va_arg' for stdarg. */
6196 sparc_gimplify_va_arg (tree valist, tree type, tree *pre_p, tree *post_p)
6198 HOST_WIDE_INT size, rsize, align;
6201 tree ptrtype = build_pointer_type (type);
6203 if (pass_by_reference (NULL, TYPE_MODE (type), type, false))
6206 size = rsize = UNITS_PER_WORD;
6212 size = int_size_in_bytes (type);
6213 rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD;
6218 /* For SPARC64, objects requiring 16-byte alignment get it. */
6219 if (TYPE_ALIGN (type) >= 2 * (unsigned) BITS_PER_WORD)
6220 align = 2 * UNITS_PER_WORD;
6222 /* SPARC-V9 ABI states that structures up to 16 bytes in size
6223 are left-justified in their slots. */
6224 if (AGGREGATE_TYPE_P (type))
6227 size = rsize = UNITS_PER_WORD;
6237 incr = fold (build2 (PLUS_EXPR, ptr_type_node, incr,
6238 ssize_int (align - 1)));
6239 incr = fold (build2 (BIT_AND_EXPR, ptr_type_node, incr,
6240 ssize_int (-align)));
6243 gimplify_expr (&incr, pre_p, post_p, is_gimple_val, fb_rvalue);
6246 if (BYTES_BIG_ENDIAN && size < rsize)
6247 addr = fold (build2 (PLUS_EXPR, ptr_type_node, incr,
6248 ssize_int (rsize - size)));
6252 addr = fold_convert (build_pointer_type (ptrtype), addr);
6253 addr = build_va_arg_indirect_ref (addr);
6255 /* If the address isn't aligned properly for the type,
6256 we may need to copy to a temporary.
6257 FIXME: This is inefficient. Usually we can do this
6260 && TYPE_ALIGN (type) > BITS_PER_WORD)
6262 tree tmp = create_tmp_var (type, "va_arg_tmp");
6263 tree dest_addr = build_fold_addr_expr (tmp);
6265 tree copy = build_function_call_expr
6266 (implicit_built_in_decls[BUILT_IN_MEMCPY],
6267 tree_cons (NULL_TREE, dest_addr,
6268 tree_cons (NULL_TREE, addr,
6269 tree_cons (NULL_TREE, size_int (rsize),
6272 gimplify_and_add (copy, pre_p);
6276 addr = fold_convert (ptrtype, addr);
6278 incr = fold (build2 (PLUS_EXPR, ptr_type_node, incr, ssize_int (rsize)));
6279 incr = build2 (MODIFY_EXPR, ptr_type_node, valist, incr);
6280 gimplify_and_add (incr, post_p);
6282 return build_va_arg_indirect_ref (addr);
6285 /* Implement the TARGET_VECTOR_MODE_SUPPORTED_P target hook.
6286 Specify whether the vector mode is supported by the hardware. */
6289 sparc_vector_mode_supported_p (enum machine_mode mode)
6291 return TARGET_VIS && VECTOR_MODE_P (mode) ? true : false;
6294 /* Return the string to output an unconditional branch to LABEL, which is
6295 the operand number of the label.
6297 DEST is the destination insn (i.e. the label), INSN is the source. */
6300 output_ubranch (rtx dest, int label, rtx insn)
6302 static char string[64];
6303 bool v9_form = false;
6306 if (TARGET_V9 && INSN_ADDRESSES_SET_P ())
6308 int delta = (INSN_ADDRESSES (INSN_UID (dest))
6309 - INSN_ADDRESSES (INSN_UID (insn)));
6310 /* Leave some instructions for "slop". */
6311 if (delta >= -260000 && delta < 260000)
6316 strcpy (string, "ba%*,pt\t%%xcc, ");
6318 strcpy (string, "b%*\t");
6320 p = strchr (string, '\0');
6331 /* Return the string to output a conditional branch to LABEL, which is
6332 the operand number of the label. OP is the conditional expression.
6333 XEXP (OP, 0) is assumed to be a condition code register (integer or
6334 floating point) and its mode specifies what kind of comparison we made.
6336 DEST is the destination insn (i.e. the label), INSN is the source.
6338 REVERSED is nonzero if we should reverse the sense of the comparison.
6340 ANNUL is nonzero if we should generate an annulling branch. */
6343 output_cbranch (rtx op, rtx dest, int label, int reversed, int annul,
6346 static char string[64];
6347 enum rtx_code code = GET_CODE (op);
6348 rtx cc_reg = XEXP (op, 0);
6349 enum machine_mode mode = GET_MODE (cc_reg);
6350 const char *labelno, *branch;
6351 int spaces = 8, far;
6354 /* v9 branches are limited to +-1MB. If it is too far away,
6367 fbne,a,pn %fcc2, .LC29
6375 far = TARGET_V9 && (get_attr_length (insn) >= 3);
6378 /* Reversal of FP compares takes care -- an ordered compare
6379 becomes an unordered compare and vice versa. */
6380 if (mode == CCFPmode || mode == CCFPEmode)
6381 code = reverse_condition_maybe_unordered (code);
6383 code = reverse_condition (code);
6386 /* Start by writing the branch condition. */
6387 if (mode == CCFPmode || mode == CCFPEmode)
6438 /* ??? !v9: FP branches cannot be preceded by another floating point
6439 insn. Because there is currently no concept of pre-delay slots,
6440 we can fix this only by always emitting a nop before a floating
6445 strcpy (string, "nop\n\t");
6446 strcat (string, branch);
6459 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
6471 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
6492 strcpy (string, branch);
6494 spaces -= strlen (branch);
6495 p = strchr (string, '\0');
6497 /* Now add the annulling, the label, and a possible noop. */
6510 if (! far && insn && INSN_ADDRESSES_SET_P ())
6512 int delta = (INSN_ADDRESSES (INSN_UID (dest))
6513 - INSN_ADDRESSES (INSN_UID (insn)));
6514 /* Leave some instructions for "slop". */
6515 if (delta < -260000 || delta >= 260000)
6519 if (mode == CCFPmode || mode == CCFPEmode)
6521 static char v9_fcc_labelno[] = "%%fccX, ";
6522 /* Set the char indicating the number of the fcc reg to use. */
6523 v9_fcc_labelno[5] = REGNO (cc_reg) - SPARC_FIRST_V9_FCC_REG + '0';
6524 labelno = v9_fcc_labelno;
6527 if (REGNO (cc_reg) == SPARC_FCC_REG)
6533 else if (mode == CCXmode || mode == CCX_NOOVmode)
6535 labelno = "%%xcc, ";
6541 labelno = "%%icc, ";
6546 if (*labelno && insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
6549 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
6562 strcpy (p, labelno);
6563 p = strchr (p, '\0');
6566 strcpy (p, ".+12\n\t nop\n\tb\t");
6567 /* Skip the next insn if requested or
6568 if we know that it will be a nop. */
6569 if (annul || ! final_sequence)
6583 /* Emit a library call comparison between floating point X and Y.
6584 COMPARISON is the rtl operator to compare with (EQ, NE, GT, etc.).
6585 TARGET_ARCH64 uses _Qp_* functions, which use pointers to TFmode
6586 values as arguments instead of the TFmode registers themselves,
6587 that's why we cannot call emit_float_lib_cmp. */
6589 sparc_emit_float_lib_cmp (rtx x, rtx y, enum rtx_code comparison)
6592 rtx slot0, slot1, result, tem, tem2;
6593 enum machine_mode mode;
6598 qpfunc = (TARGET_ARCH64) ? "_Qp_feq" : "_Q_feq";
6602 qpfunc = (TARGET_ARCH64) ? "_Qp_fne" : "_Q_fne";
6606 qpfunc = (TARGET_ARCH64) ? "_Qp_fgt" : "_Q_fgt";
6610 qpfunc = (TARGET_ARCH64) ? "_Qp_fge" : "_Q_fge";
6614 qpfunc = (TARGET_ARCH64) ? "_Qp_flt" : "_Q_flt";
6618 qpfunc = (TARGET_ARCH64) ? "_Qp_fle" : "_Q_fle";
6629 qpfunc = (TARGET_ARCH64) ? "_Qp_cmp" : "_Q_cmp";
6639 if (GET_CODE (x) != MEM)
6641 slot0 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0);
6642 emit_insn (gen_rtx_SET (VOIDmode, slot0, x));
6647 if (GET_CODE (y) != MEM)
6649 slot1 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0);
6650 emit_insn (gen_rtx_SET (VOIDmode, slot1, y));
6655 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, qpfunc), LCT_NORMAL,
6657 XEXP (slot0, 0), Pmode,
6658 XEXP (slot1, 0), Pmode);
6664 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, qpfunc), LCT_NORMAL,
6666 x, TFmode, y, TFmode);
6672 /* Immediately move the result of the libcall into a pseudo
6673 register so reload doesn't clobber the value if it needs
6674 the return register for a spill reg. */
6675 result = gen_reg_rtx (mode);
6676 emit_move_insn (result, hard_libcall_value (mode));
6681 emit_cmp_insn (result, const0_rtx, NE, NULL_RTX, mode, 0);
6685 emit_cmp_insn (result, GEN_INT(3), comparison == UNORDERED ? EQ : NE,
6690 emit_cmp_insn (result, const1_rtx,
6691 comparison == UNGT ? GT : NE, NULL_RTX, mode, 0);
6694 emit_cmp_insn (result, const2_rtx, NE, NULL_RTX, mode, 0);
6697 tem = gen_reg_rtx (mode);
6699 emit_insn (gen_andsi3 (tem, result, const1_rtx));
6701 emit_insn (gen_anddi3 (tem, result, const1_rtx));
6702 emit_cmp_insn (tem, const0_rtx, NE, NULL_RTX, mode, 0);
6706 tem = gen_reg_rtx (mode);
6708 emit_insn (gen_addsi3 (tem, result, const1_rtx));
6710 emit_insn (gen_adddi3 (tem, result, const1_rtx));
6711 tem2 = gen_reg_rtx (mode);
6713 emit_insn (gen_andsi3 (tem2, tem, const2_rtx));
6715 emit_insn (gen_anddi3 (tem2, tem, const2_rtx));
6716 emit_cmp_insn (tem2, const0_rtx, comparison == UNEQ ? EQ : NE,
6722 /* Generate an unsigned DImode to FP conversion. This is the same code
6723 optabs would emit if we didn't have TFmode patterns. */
6726 sparc_emit_floatunsdi (rtx *operands, enum machine_mode mode)
6728 rtx neglab, donelab, i0, i1, f0, in, out;
6731 in = force_reg (DImode, operands[1]);
6732 neglab = gen_label_rtx ();
6733 donelab = gen_label_rtx ();
6734 i0 = gen_reg_rtx (DImode);
6735 i1 = gen_reg_rtx (DImode);
6736 f0 = gen_reg_rtx (mode);
6738 emit_cmp_and_jump_insns (in, const0_rtx, LT, const0_rtx, DImode, 0, neglab);
6740 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_FLOAT (mode, in)));
6741 emit_jump_insn (gen_jump (donelab));
6744 emit_label (neglab);
6746 emit_insn (gen_lshrdi3 (i0, in, const1_rtx));
6747 emit_insn (gen_anddi3 (i1, in, const1_rtx));
6748 emit_insn (gen_iordi3 (i0, i0, i1));
6749 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_FLOAT (mode, i0)));
6750 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_PLUS (mode, f0, f0)));
6752 emit_label (donelab);
6755 /* Generate an FP to unsigned DImode conversion. This is the same code
6756 optabs would emit if we didn't have TFmode patterns. */
6759 sparc_emit_fixunsdi (rtx *operands, enum machine_mode mode)
6761 rtx neglab, donelab, i0, i1, f0, in, out, limit;
6764 in = force_reg (mode, operands[1]);
6765 neglab = gen_label_rtx ();
6766 donelab = gen_label_rtx ();
6767 i0 = gen_reg_rtx (DImode);
6768 i1 = gen_reg_rtx (DImode);
6769 limit = gen_reg_rtx (mode);
6770 f0 = gen_reg_rtx (mode);
6772 emit_move_insn (limit,
6773 CONST_DOUBLE_FROM_REAL_VALUE (
6774 REAL_VALUE_ATOF ("9223372036854775808.0", mode), mode));
6775 emit_cmp_and_jump_insns (in, limit, GE, NULL_RTX, mode, 0, neglab);
6777 emit_insn (gen_rtx_SET (VOIDmode,
6779 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, in))));
6780 emit_jump_insn (gen_jump (donelab));
6783 emit_label (neglab);
6785 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_MINUS (mode, in, limit)));
6786 emit_insn (gen_rtx_SET (VOIDmode,
6788 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, f0))));
6789 emit_insn (gen_movdi (i1, const1_rtx));
6790 emit_insn (gen_ashldi3 (i1, i1, GEN_INT (63)));
6791 emit_insn (gen_xordi3 (out, i0, i1));
6793 emit_label (donelab);
6796 /* Return the string to output a conditional branch to LABEL, testing
6797 register REG. LABEL is the operand number of the label; REG is the
6798 operand number of the reg. OP is the conditional expression. The mode
6799 of REG says what kind of comparison we made.
6801 DEST is the destination insn (i.e. the label), INSN is the source.
6803 REVERSED is nonzero if we should reverse the sense of the comparison.
6805 ANNUL is nonzero if we should generate an annulling branch. */
6808 output_v9branch (rtx op, rtx dest, int reg, int label, int reversed,
6809 int annul, rtx insn)
6811 static char string[64];
6812 enum rtx_code code = GET_CODE (op);
6813 enum machine_mode mode = GET_MODE (XEXP (op, 0));
6818 /* branch on register are limited to +-128KB. If it is too far away,
6831 brgez,a,pn %o1, .LC29
6837 ba,pt %xcc, .LC29 */
6839 far = get_attr_length (insn) >= 3;
6841 /* If not floating-point or if EQ or NE, we can just reverse the code. */
6843 code = reverse_condition (code);
6845 /* Only 64 bit versions of these instructions exist. */
6849 /* Start by writing the branch condition. */
6854 strcpy (string, "brnz");
6858 strcpy (string, "brz");
6862 strcpy (string, "brgez");
6866 strcpy (string, "brlz");
6870 strcpy (string, "brlez");
6874 strcpy (string, "brgz");
6881 p = strchr (string, '\0');
6883 /* Now add the annulling, reg, label, and nop. */
6890 if (insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
6893 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
6898 *p = p < string + 8 ? '\t' : ' ';
6906 int veryfar = 1, delta;
6908 if (INSN_ADDRESSES_SET_P ())
6910 delta = (INSN_ADDRESSES (INSN_UID (dest))
6911 - INSN_ADDRESSES (INSN_UID (insn)));
6912 /* Leave some instructions for "slop". */
6913 if (delta >= -260000 && delta < 260000)
6917 strcpy (p, ".+12\n\t nop\n\t");
6918 /* Skip the next insn if requested or
6919 if we know that it will be a nop. */
6920 if (annul || ! final_sequence)
6930 strcpy (p, "ba,pt\t%%xcc, ");
6944 /* Return 1, if any of the registers of the instruction are %l[0-7] or %o[0-7].
6945 Such instructions cannot be used in the delay slot of return insn on v9.
6946 If TEST is 0, also rename all %i[0-7] registers to their %o[0-7] counterparts.
6950 epilogue_renumber (register rtx *where, int test)
6952 register const char *fmt;
6954 register enum rtx_code code;
6959 code = GET_CODE (*where);
6964 if (REGNO (*where) >= 8 && REGNO (*where) < 24) /* oX or lX */
6966 if (! test && REGNO (*where) >= 24 && REGNO (*where) < 32)
6967 *where = gen_rtx_REG (GET_MODE (*where), OUTGOING_REGNO (REGNO(*where)));
6975 /* Do not replace the frame pointer with the stack pointer because
6976 it can cause the delayed instruction to load below the stack.
6977 This occurs when instructions like:
6979 (set (reg/i:SI 24 %i0)
6980 (mem/f:SI (plus:SI (reg/f:SI 30 %fp)
6981 (const_int -20 [0xffffffec])) 0))
6983 are in the return delayed slot. */
6985 if (GET_CODE (XEXP (*where, 0)) == REG
6986 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM
6987 && (GET_CODE (XEXP (*where, 1)) != CONST_INT
6988 || INTVAL (XEXP (*where, 1)) < SPARC_STACK_BIAS))
6993 if (SPARC_STACK_BIAS
6994 && GET_CODE (XEXP (*where, 0)) == REG
6995 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM)
7003 fmt = GET_RTX_FORMAT (code);
7005 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7010 for (j = XVECLEN (*where, i) - 1; j >= 0; j--)
7011 if (epilogue_renumber (&(XVECEXP (*where, i, j)), test))
7014 else if (fmt[i] == 'e'
7015 && epilogue_renumber (&(XEXP (*where, i)), test))
7021 /* Leaf functions and non-leaf functions have different needs. */
7024 reg_leaf_alloc_order[] = REG_LEAF_ALLOC_ORDER;
7027 reg_nonleaf_alloc_order[] = REG_ALLOC_ORDER;
7029 static const int *const reg_alloc_orders[] = {
7030 reg_leaf_alloc_order,
7031 reg_nonleaf_alloc_order};
7034 order_regs_for_local_alloc (void)
7036 static int last_order_nonleaf = 1;
7038 if (regs_ever_live[15] != last_order_nonleaf)
7040 last_order_nonleaf = !last_order_nonleaf;
7041 memcpy ((char *) reg_alloc_order,
7042 (const char *) reg_alloc_orders[last_order_nonleaf],
7043 FIRST_PSEUDO_REGISTER * sizeof (int));
7047 /* Return 1 if REG and MEM are legitimate enough to allow the various
7048 mem<-->reg splits to be run. */
7051 sparc_splitdi_legitimate (rtx reg, rtx mem)
7053 /* Punt if we are here by mistake. */
7054 if (! reload_completed)
7057 /* We must have an offsettable memory reference. */
7058 if (! offsettable_memref_p (mem))
7061 /* If we have legitimate args for ldd/std, we do not want
7062 the split to happen. */
7063 if ((REGNO (reg) % 2) == 0
7064 && mem_min_alignment (mem, 8))
7071 /* Return 1 if x and y are some kind of REG and they refer to
7072 different hard registers. This test is guaranteed to be
7073 run after reload. */
7076 sparc_absnegfloat_split_legitimate (rtx x, rtx y)
7078 if (GET_CODE (x) != REG)
7080 if (GET_CODE (y) != REG)
7082 if (REGNO (x) == REGNO (y))
7087 /* Return 1 if REGNO (reg1) is even and REGNO (reg1) == REGNO (reg2) - 1.
7088 This makes them candidates for using ldd and std insns.
7090 Note reg1 and reg2 *must* be hard registers. */
7093 registers_ok_for_ldd_peep (rtx reg1, rtx reg2)
7095 /* We might have been passed a SUBREG. */
7096 if (GET_CODE (reg1) != REG || GET_CODE (reg2) != REG)
7099 if (REGNO (reg1) % 2 != 0)
7102 /* Integer ldd is deprecated in SPARC V9 */
7103 if (TARGET_V9 && REGNO (reg1) < 32)
7106 return (REGNO (reg1) == REGNO (reg2) - 1);
7109 /* Return 1 if the addresses in mem1 and mem2 are suitable for use in
7112 This can only happen when addr1 and addr2, the addresses in mem1
7113 and mem2, are consecutive memory locations (addr1 + 4 == addr2).
7114 addr1 must also be aligned on a 64-bit boundary.
7116 Also iff dependent_reg_rtx is not null it should not be used to
7117 compute the address for mem1, i.e. we cannot optimize a sequence
7129 But, note that the transformation from:
7134 is perfectly fine. Thus, the peephole2 patterns always pass us
7135 the destination register of the first load, never the second one.
7137 For stores we don't have a similar problem, so dependent_reg_rtx is
7141 mems_ok_for_ldd_peep (rtx mem1, rtx mem2, rtx dependent_reg_rtx)
7145 HOST_WIDE_INT offset1;
7147 /* The mems cannot be volatile. */
7148 if (MEM_VOLATILE_P (mem1) || MEM_VOLATILE_P (mem2))
7151 /* MEM1 should be aligned on a 64-bit boundary. */
7152 if (MEM_ALIGN (mem1) < 64)
7155 addr1 = XEXP (mem1, 0);
7156 addr2 = XEXP (mem2, 0);
7158 /* Extract a register number and offset (if used) from the first addr. */
7159 if (GET_CODE (addr1) == PLUS)
7161 /* If not a REG, return zero. */
7162 if (GET_CODE (XEXP (addr1, 0)) != REG)
7166 reg1 = REGNO (XEXP (addr1, 0));
7167 /* The offset must be constant! */
7168 if (GET_CODE (XEXP (addr1, 1)) != CONST_INT)
7170 offset1 = INTVAL (XEXP (addr1, 1));
7173 else if (GET_CODE (addr1) != REG)
7177 reg1 = REGNO (addr1);
7178 /* This was a simple (mem (reg)) expression. Offset is 0. */
7182 /* Make sure the second address is a (mem (plus (reg) (const_int). */
7183 if (GET_CODE (addr2) != PLUS)
7186 if (GET_CODE (XEXP (addr2, 0)) != REG
7187 || GET_CODE (XEXP (addr2, 1)) != CONST_INT)
7190 if (reg1 != REGNO (XEXP (addr2, 0)))
7193 if (dependent_reg_rtx != NULL_RTX && reg1 == REGNO (dependent_reg_rtx))
7196 /* The first offset must be evenly divisible by 8 to ensure the
7197 address is 64 bit aligned. */
7198 if (offset1 % 8 != 0)
7201 /* The offset for the second addr must be 4 more than the first addr. */
7202 if (INTVAL (XEXP (addr2, 1)) != offset1 + 4)
7205 /* All the tests passed. addr1 and addr2 are valid for ldd and std
7210 /* Return 1 if reg is a pseudo, or is the first register in
7211 a hard register pair. This makes it a candidate for use in
7212 ldd and std insns. */
7215 register_ok_for_ldd (rtx reg)
7217 /* We might have been passed a SUBREG. */
7218 if (GET_CODE (reg) != REG)
7221 if (REGNO (reg) < FIRST_PSEUDO_REGISTER)
7222 return (REGNO (reg) % 2 == 0);
7227 /* Print operand X (an rtx) in assembler syntax to file FILE.
7228 CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified.
7229 For `%' followed by punctuation, CODE is the punctuation and X is null. */
7232 print_operand (FILE *file, rtx x, int code)
7237 /* Output an insn in a delay slot. */
7239 sparc_indent_opcode = 1;
7241 fputs ("\n\t nop", file);
7244 /* Output an annul flag if there's nothing for the delay slot and we
7245 are optimizing. This is always used with '(' below.
7246 Sun OS 4.1.1 dbx can't handle an annulled unconditional branch;
7247 this is a dbx bug. So, we only do this when optimizing.
7248 On UltraSPARC, a branch in a delay slot causes a pipeline flush.
7249 Always emit a nop in case the next instruction is a branch. */
7250 if (! final_sequence && (optimize && (int)sparc_cpu < PROCESSOR_V9))
7254 /* Output a 'nop' if there's nothing for the delay slot and we are
7255 not optimizing. This is always used with '*' above. */
7256 if (! final_sequence && ! (optimize && (int)sparc_cpu < PROCESSOR_V9))
7257 fputs ("\n\t nop", file);
7258 else if (final_sequence)
7259 sparc_indent_opcode = 1;
7262 /* Output the right displacement from the saved PC on function return.
7263 The caller may have placed an "unimp" insn immediately after the call
7264 so we have to account for it. This insn is used in the 32-bit ABI
7265 when calling a function that returns a non zero-sized structure. The
7266 64-bit ABI doesn't have it. Be careful to have this test be the same
7267 as that used on the call. */
7269 && current_function_returns_struct
7270 && (TREE_CODE (DECL_SIZE (DECL_RESULT (current_function_decl)))
7272 && ! integer_zerop (DECL_SIZE (DECL_RESULT (current_function_decl))))
7278 /* Output the Embedded Medium/Anywhere code model base register. */
7279 fputs (EMBMEDANY_BASE_REG, file);
7282 /* Print some local dynamic TLS name. */
7283 assemble_name (file, get_some_local_dynamic_name ());
7287 /* Adjust the operand to take into account a RESTORE operation. */
7288 if (GET_CODE (x) == CONST_INT)
7290 else if (GET_CODE (x) != REG)
7291 output_operand_lossage ("invalid %%Y operand");
7292 else if (REGNO (x) < 8)
7293 fputs (reg_names[REGNO (x)], file);
7294 else if (REGNO (x) >= 24 && REGNO (x) < 32)
7295 fputs (reg_names[REGNO (x)-16], file);
7297 output_operand_lossage ("invalid %%Y operand");
7300 /* Print out the low order register name of a register pair. */
7301 if (WORDS_BIG_ENDIAN)
7302 fputs (reg_names[REGNO (x)+1], file);
7304 fputs (reg_names[REGNO (x)], file);
7307 /* Print out the high order register name of a register pair. */
7308 if (WORDS_BIG_ENDIAN)
7309 fputs (reg_names[REGNO (x)], file);
7311 fputs (reg_names[REGNO (x)+1], file);
7314 /* Print out the second register name of a register pair or quad.
7315 I.e., R (%o0) => %o1. */
7316 fputs (reg_names[REGNO (x)+1], file);
7319 /* Print out the third register name of a register quad.
7320 I.e., S (%o0) => %o2. */
7321 fputs (reg_names[REGNO (x)+2], file);
7324 /* Print out the fourth register name of a register quad.
7325 I.e., T (%o0) => %o3. */
7326 fputs (reg_names[REGNO (x)+3], file);
7329 /* Print a condition code register. */
7330 if (REGNO (x) == SPARC_ICC_REG)
7332 /* We don't handle CC[X]_NOOVmode because they're not supposed
7334 if (GET_MODE (x) == CCmode)
7335 fputs ("%icc", file);
7336 else if (GET_MODE (x) == CCXmode)
7337 fputs ("%xcc", file);
7342 /* %fccN register */
7343 fputs (reg_names[REGNO (x)], file);
7346 /* Print the operand's address only. */
7347 output_address (XEXP (x, 0));
7350 /* In this case we need a register. Use %g0 if the
7351 operand is const0_rtx. */
7353 || (GET_MODE (x) != VOIDmode && x == CONST0_RTX (GET_MODE (x))))
7355 fputs ("%g0", file);
7362 switch (GET_CODE (x))
7364 case IOR: fputs ("or", file); break;
7365 case AND: fputs ("and", file); break;
7366 case XOR: fputs ("xor", file); break;
7367 default: output_operand_lossage ("invalid %%A operand");
7372 switch (GET_CODE (x))
7374 case IOR: fputs ("orn", file); break;
7375 case AND: fputs ("andn", file); break;
7376 case XOR: fputs ("xnor", file); break;
7377 default: output_operand_lossage ("invalid %%B operand");
7381 /* These are used by the conditional move instructions. */
7385 enum rtx_code rc = GET_CODE (x);
7389 enum machine_mode mode = GET_MODE (XEXP (x, 0));
7390 if (mode == CCFPmode || mode == CCFPEmode)
7391 rc = reverse_condition_maybe_unordered (GET_CODE (x));
7393 rc = reverse_condition (GET_CODE (x));
7397 case NE: fputs ("ne", file); break;
7398 case EQ: fputs ("e", file); break;
7399 case GE: fputs ("ge", file); break;
7400 case GT: fputs ("g", file); break;
7401 case LE: fputs ("le", file); break;
7402 case LT: fputs ("l", file); break;
7403 case GEU: fputs ("geu", file); break;
7404 case GTU: fputs ("gu", file); break;
7405 case LEU: fputs ("leu", file); break;
7406 case LTU: fputs ("lu", file); break;
7407 case LTGT: fputs ("lg", file); break;
7408 case UNORDERED: fputs ("u", file); break;
7409 case ORDERED: fputs ("o", file); break;
7410 case UNLT: fputs ("ul", file); break;
7411 case UNLE: fputs ("ule", file); break;
7412 case UNGT: fputs ("ug", file); break;
7413 case UNGE: fputs ("uge", file); break;
7414 case UNEQ: fputs ("ue", file); break;
7415 default: output_operand_lossage (code == 'c'
7416 ? "invalid %%c operand"
7417 : "invalid %%C operand");
7422 /* These are used by the movr instruction pattern. */
7426 enum rtx_code rc = (code == 'd'
7427 ? reverse_condition (GET_CODE (x))
7431 case NE: fputs ("ne", file); break;
7432 case EQ: fputs ("e", file); break;
7433 case GE: fputs ("gez", file); break;
7434 case LT: fputs ("lz", file); break;
7435 case LE: fputs ("lez", file); break;
7436 case GT: fputs ("gz", file); break;
7437 default: output_operand_lossage (code == 'd'
7438 ? "invalid %%d operand"
7439 : "invalid %%D operand");
7446 /* Print a sign-extended character. */
7447 int i = trunc_int_for_mode (INTVAL (x), QImode);
7448 fprintf (file, "%d", i);
7453 /* Operand must be a MEM; write its address. */
7454 if (GET_CODE (x) != MEM)
7455 output_operand_lossage ("invalid %%f operand");
7456 output_address (XEXP (x, 0));
7461 /* Print a sign-extended 32-bit value. */
7463 if (GET_CODE(x) == CONST_INT)
7465 else if (GET_CODE(x) == CONST_DOUBLE)
7466 i = CONST_DOUBLE_LOW (x);
7469 output_operand_lossage ("invalid %%s operand");
7472 i = trunc_int_for_mode (i, SImode);
7473 fprintf (file, HOST_WIDE_INT_PRINT_DEC, i);
7478 /* Do nothing special. */
7482 /* Undocumented flag. */
7483 output_operand_lossage ("invalid operand output code");
7486 if (GET_CODE (x) == REG)
7487 fputs (reg_names[REGNO (x)], file);
7488 else if (GET_CODE (x) == MEM)
7491 /* Poor Sun assembler doesn't understand absolute addressing. */
7492 if (CONSTANT_P (XEXP (x, 0)))
7493 fputs ("%g0+", file);
7494 output_address (XEXP (x, 0));
7497 else if (GET_CODE (x) == HIGH)
7499 fputs ("%hi(", file);
7500 output_addr_const (file, XEXP (x, 0));
7503 else if (GET_CODE (x) == LO_SUM)
7505 print_operand (file, XEXP (x, 0), 0);
7506 if (TARGET_CM_MEDMID)
7507 fputs ("+%l44(", file);
7509 fputs ("+%lo(", file);
7510 output_addr_const (file, XEXP (x, 1));
7513 else if (GET_CODE (x) == CONST_DOUBLE
7514 && (GET_MODE (x) == VOIDmode
7515 || GET_MODE_CLASS (GET_MODE (x)) == MODE_INT))
7517 if (CONST_DOUBLE_HIGH (x) == 0)
7518 fprintf (file, "%u", (unsigned int) CONST_DOUBLE_LOW (x));
7519 else if (CONST_DOUBLE_HIGH (x) == -1
7520 && CONST_DOUBLE_LOW (x) < 0)
7521 fprintf (file, "%d", (int) CONST_DOUBLE_LOW (x));
7523 output_operand_lossage ("long long constant not a valid immediate operand");
7525 else if (GET_CODE (x) == CONST_DOUBLE)
7526 output_operand_lossage ("floating point constant not a valid immediate operand");
7527 else { output_addr_const (file, x); }
7530 /* Target hook for assembling integer objects. The sparc version has
7531 special handling for aligned DI-mode objects. */
7534 sparc_assemble_integer (rtx x, unsigned int size, int aligned_p)
7536 /* ??? We only output .xword's for symbols and only then in environments
7537 where the assembler can handle them. */
7538 if (aligned_p && size == 8
7539 && (GET_CODE (x) != CONST_INT && GET_CODE (x) != CONST_DOUBLE))
7543 assemble_integer_with_op ("\t.xword\t", x);
7548 assemble_aligned_integer (4, const0_rtx);
7549 assemble_aligned_integer (4, x);
7553 return default_assemble_integer (x, size, aligned_p);
7556 /* Return the value of a code used in the .proc pseudo-op that says
7557 what kind of result this function returns. For non-C types, we pick
7558 the closest C type. */
7560 #ifndef SHORT_TYPE_SIZE
7561 #define SHORT_TYPE_SIZE (BITS_PER_UNIT * 2)
7564 #ifndef INT_TYPE_SIZE
7565 #define INT_TYPE_SIZE BITS_PER_WORD
7568 #ifndef LONG_TYPE_SIZE
7569 #define LONG_TYPE_SIZE BITS_PER_WORD
7572 #ifndef LONG_LONG_TYPE_SIZE
7573 #define LONG_LONG_TYPE_SIZE (BITS_PER_WORD * 2)
7576 #ifndef FLOAT_TYPE_SIZE
7577 #define FLOAT_TYPE_SIZE BITS_PER_WORD
7580 #ifndef DOUBLE_TYPE_SIZE
7581 #define DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
7584 #ifndef LONG_DOUBLE_TYPE_SIZE
7585 #define LONG_DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
7589 sparc_type_code (register tree type)
7591 register unsigned long qualifiers = 0;
7592 register unsigned shift;
7594 /* Only the first 30 bits of the qualifier are valid. We must refrain from
7595 setting more, since some assemblers will give an error for this. Also,
7596 we must be careful to avoid shifts of 32 bits or more to avoid getting
7597 unpredictable results. */
7599 for (shift = 6; shift < 30; shift += 2, type = TREE_TYPE (type))
7601 switch (TREE_CODE (type))
7607 qualifiers |= (3 << shift);
7612 qualifiers |= (2 << shift);
7616 case REFERENCE_TYPE:
7618 qualifiers |= (1 << shift);
7622 return (qualifiers | 8);
7625 case QUAL_UNION_TYPE:
7626 return (qualifiers | 9);
7629 return (qualifiers | 10);
7632 return (qualifiers | 16);
7635 /* If this is a range type, consider it to be the underlying
7637 if (TREE_TYPE (type) != 0)
7640 /* Carefully distinguish all the standard types of C,
7641 without messing up if the language is not C. We do this by
7642 testing TYPE_PRECISION and TYPE_UNSIGNED. The old code used to
7643 look at both the names and the above fields, but that's redundant.
7644 Any type whose size is between two C types will be considered
7645 to be the wider of the two types. Also, we do not have a
7646 special code to use for "long long", so anything wider than
7647 long is treated the same. Note that we can't distinguish
7648 between "int" and "long" in this code if they are the same
7649 size, but that's fine, since neither can the assembler. */
7651 if (TYPE_PRECISION (type) <= CHAR_TYPE_SIZE)
7652 return (qualifiers | (TYPE_UNSIGNED (type) ? 12 : 2));
7654 else if (TYPE_PRECISION (type) <= SHORT_TYPE_SIZE)
7655 return (qualifiers | (TYPE_UNSIGNED (type) ? 13 : 3));
7657 else if (TYPE_PRECISION (type) <= INT_TYPE_SIZE)
7658 return (qualifiers | (TYPE_UNSIGNED (type) ? 14 : 4));
7661 return (qualifiers | (TYPE_UNSIGNED (type) ? 15 : 5));
7664 /* If this is a range type, consider it to be the underlying
7666 if (TREE_TYPE (type) != 0)
7669 /* Carefully distinguish all the standard types of C,
7670 without messing up if the language is not C. */
7672 if (TYPE_PRECISION (type) == FLOAT_TYPE_SIZE)
7673 return (qualifiers | 6);
7676 return (qualifiers | 7);
7678 case COMPLEX_TYPE: /* GNU Fortran COMPLEX type. */
7679 /* ??? We need to distinguish between double and float complex types,
7680 but I don't know how yet because I can't reach this code from
7681 existing front-ends. */
7682 return (qualifiers | 7); /* Who knows? */
7685 case CHAR_TYPE: /* GNU Pascal CHAR type. Not used in C. */
7686 case BOOLEAN_TYPE: /* GNU Fortran BOOLEAN type. */
7687 case FILE_TYPE: /* GNU Pascal FILE type. */
7688 case SET_TYPE: /* GNU Pascal SET type. */
7689 case LANG_TYPE: /* ? */
7693 abort (); /* Not a type! */
7700 /* Nested function support. */
7702 /* Emit RTL insns to initialize the variable parts of a trampoline.
7703 FNADDR is an RTX for the address of the function's pure code.
7704 CXT is an RTX for the static chain value for the function.
7706 This takes 16 insns: 2 shifts & 2 ands (to split up addresses), 4 sethi
7707 (to load in opcodes), 4 iors (to merge address and opcodes), and 4 writes
7708 (to store insns). This is a bit excessive. Perhaps a different
7709 mechanism would be better here.
7711 Emit enough FLUSH insns to synchronize the data and instruction caches. */
7714 sparc_initialize_trampoline (rtx tramp, rtx fnaddr, rtx cxt)
7716 /* SPARC 32-bit trampoline:
7719 sethi %hi(static), %g2
7721 or %g2, %lo(static), %g2
7723 SETHI i,r = 00rr rrr1 00ii iiii iiii iiii iiii iiii
7724 JMPL r+i,d = 10dd ddd1 1100 0rrr rr1i iiii iiii iiii
7728 (gen_rtx_MEM (SImode, plus_constant (tramp, 0)),
7729 expand_binop (SImode, ior_optab,
7730 expand_shift (RSHIFT_EXPR, SImode, fnaddr,
7731 size_int (10), 0, 1),
7732 GEN_INT (trunc_int_for_mode (0x03000000, SImode)),
7733 NULL_RTX, 1, OPTAB_DIRECT));
7736 (gen_rtx_MEM (SImode, plus_constant (tramp, 4)),
7737 expand_binop (SImode, ior_optab,
7738 expand_shift (RSHIFT_EXPR, SImode, cxt,
7739 size_int (10), 0, 1),
7740 GEN_INT (trunc_int_for_mode (0x05000000, SImode)),
7741 NULL_RTX, 1, OPTAB_DIRECT));
7744 (gen_rtx_MEM (SImode, plus_constant (tramp, 8)),
7745 expand_binop (SImode, ior_optab,
7746 expand_and (SImode, fnaddr, GEN_INT (0x3ff), NULL_RTX),
7747 GEN_INT (trunc_int_for_mode (0x81c06000, SImode)),
7748 NULL_RTX, 1, OPTAB_DIRECT));
7751 (gen_rtx_MEM (SImode, plus_constant (tramp, 12)),
7752 expand_binop (SImode, ior_optab,
7753 expand_and (SImode, cxt, GEN_INT (0x3ff), NULL_RTX),
7754 GEN_INT (trunc_int_for_mode (0x8410a000, SImode)),
7755 NULL_RTX, 1, OPTAB_DIRECT));
7757 /* On UltraSPARC a flush flushes an entire cache line. The trampoline is
7758 aligned on a 16 byte boundary so one flush clears it all. */
7759 emit_insn (gen_flush (validize_mem (gen_rtx_MEM (SImode, tramp))));
7760 if (sparc_cpu != PROCESSOR_ULTRASPARC
7761 && sparc_cpu != PROCESSOR_ULTRASPARC3)
7762 emit_insn (gen_flush (validize_mem (gen_rtx_MEM (SImode,
7763 plus_constant (tramp, 8)))));
7765 /* Call __enable_execute_stack after writing onto the stack to make sure
7766 the stack address is accessible. */
7767 #ifdef ENABLE_EXECUTE_STACK
7768 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
7769 LCT_NORMAL, VOIDmode, 1, tramp, Pmode);
7774 /* The 64-bit version is simpler because it makes more sense to load the
7775 values as "immediate" data out of the trampoline. It's also easier since
7776 we can read the PC without clobbering a register. */
7779 sparc64_initialize_trampoline (rtx tramp, rtx fnaddr, rtx cxt)
7781 /* SPARC 64-bit trampoline:
7790 emit_move_insn (gen_rtx_MEM (SImode, tramp),
7791 GEN_INT (trunc_int_for_mode (0x83414000, SImode)));
7792 emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 4)),
7793 GEN_INT (trunc_int_for_mode (0xca586018, SImode)));
7794 emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 8)),
7795 GEN_INT (trunc_int_for_mode (0x81c14000, SImode)));
7796 emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 12)),
7797 GEN_INT (trunc_int_for_mode (0xca586010, SImode)));
7798 emit_move_insn (gen_rtx_MEM (DImode, plus_constant (tramp, 16)), cxt);
7799 emit_move_insn (gen_rtx_MEM (DImode, plus_constant (tramp, 24)), fnaddr);
7800 emit_insn (gen_flushdi (validize_mem (gen_rtx_MEM (DImode, tramp))));
7802 if (sparc_cpu != PROCESSOR_ULTRASPARC
7803 && sparc_cpu != PROCESSOR_ULTRASPARC3)
7804 emit_insn (gen_flushdi (validize_mem (gen_rtx_MEM (DImode, plus_constant (tramp, 8)))));
7806 /* Call __enable_execute_stack after writing onto the stack to make sure
7807 the stack address is accessible. */
7808 #ifdef ENABLE_EXECUTE_STACK
7809 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
7810 LCT_NORMAL, VOIDmode, 1, tramp, Pmode);
7814 /* Adjust the cost of a scheduling dependency. Return the new cost of
7815 a dependency LINK or INSN on DEP_INSN. COST is the current cost. */
7818 supersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
7820 enum attr_type insn_type;
7822 if (! recog_memoized (insn))
7825 insn_type = get_attr_type (insn);
7827 if (REG_NOTE_KIND (link) == 0)
7829 /* Data dependency; DEP_INSN writes a register that INSN reads some
7832 /* if a load, then the dependence must be on the memory address;
7833 add an extra "cycle". Note that the cost could be two cycles
7834 if the reg was written late in an instruction group; we ca not tell
7836 if (insn_type == TYPE_LOAD || insn_type == TYPE_FPLOAD)
7839 /* Get the delay only if the address of the store is the dependence. */
7840 if (insn_type == TYPE_STORE || insn_type == TYPE_FPSTORE)
7842 rtx pat = PATTERN(insn);
7843 rtx dep_pat = PATTERN (dep_insn);
7845 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
7846 return cost; /* This should not happen! */
7848 /* The dependency between the two instructions was on the data that
7849 is being stored. Assume that this implies that the address of the
7850 store is not dependent. */
7851 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
7854 return cost + 3; /* An approximation. */
7857 /* A shift instruction cannot receive its data from an instruction
7858 in the same cycle; add a one cycle penalty. */
7859 if (insn_type == TYPE_SHIFT)
7860 return cost + 3; /* Split before cascade into shift. */
7864 /* Anti- or output- dependency; DEP_INSN reads/writes a register that
7865 INSN writes some cycles later. */
7867 /* These are only significant for the fpu unit; writing a fp reg before
7868 the fpu has finished with it stalls the processor. */
7870 /* Reusing an integer register causes no problems. */
7871 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
7879 hypersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
7881 enum attr_type insn_type, dep_type;
7882 rtx pat = PATTERN(insn);
7883 rtx dep_pat = PATTERN (dep_insn);
7885 if (recog_memoized (insn) < 0 || recog_memoized (dep_insn) < 0)
7888 insn_type = get_attr_type (insn);
7889 dep_type = get_attr_type (dep_insn);
7891 switch (REG_NOTE_KIND (link))
7894 /* Data dependency; DEP_INSN writes a register that INSN reads some
7901 /* Get the delay iff the address of the store is the dependence. */
7902 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
7905 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
7912 /* If a load, then the dependence must be on the memory address. If
7913 the addresses aren't equal, then it might be a false dependency */
7914 if (dep_type == TYPE_STORE || dep_type == TYPE_FPSTORE)
7916 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET
7917 || GET_CODE (SET_DEST (dep_pat)) != MEM
7918 || GET_CODE (SET_SRC (pat)) != MEM
7919 || ! rtx_equal_p (XEXP (SET_DEST (dep_pat), 0),
7920 XEXP (SET_SRC (pat), 0)))
7928 /* Compare to branch latency is 0. There is no benefit from
7929 separating compare and branch. */
7930 if (dep_type == TYPE_COMPARE)
7932 /* Floating point compare to branch latency is less than
7933 compare to conditional move. */
7934 if (dep_type == TYPE_FPCMP)
7943 /* Anti-dependencies only penalize the fpu unit. */
7944 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
7956 sparc_adjust_cost(rtx insn, rtx link, rtx dep, int cost)
7960 case PROCESSOR_SUPERSPARC:
7961 cost = supersparc_adjust_cost (insn, link, dep, cost);
7963 case PROCESSOR_HYPERSPARC:
7964 case PROCESSOR_SPARCLITE86X:
7965 cost = hypersparc_adjust_cost (insn, link, dep, cost);
7974 sparc_sched_init (FILE *dump ATTRIBUTE_UNUSED,
7975 int sched_verbose ATTRIBUTE_UNUSED,
7976 int max_ready ATTRIBUTE_UNUSED)
7981 sparc_use_sched_lookahead (void)
7983 if (sparc_cpu == PROCESSOR_ULTRASPARC
7984 || sparc_cpu == PROCESSOR_ULTRASPARC3)
7986 if ((1 << sparc_cpu) &
7987 ((1 << PROCESSOR_SUPERSPARC) | (1 << PROCESSOR_HYPERSPARC) |
7988 (1 << PROCESSOR_SPARCLITE86X)))
7994 sparc_issue_rate (void)
8001 /* Assume V9 processors are capable of at least dual-issue. */
8003 case PROCESSOR_SUPERSPARC:
8005 case PROCESSOR_HYPERSPARC:
8006 case PROCESSOR_SPARCLITE86X:
8008 case PROCESSOR_ULTRASPARC:
8009 case PROCESSOR_ULTRASPARC3:
8015 set_extends (rtx insn)
8017 register rtx pat = PATTERN (insn);
8019 switch (GET_CODE (SET_SRC (pat)))
8021 /* Load and some shift instructions zero extend. */
8024 /* sethi clears the high bits */
8026 /* LO_SUM is used with sethi. sethi cleared the high
8027 bits and the values used with lo_sum are positive */
8029 /* Store flag stores 0 or 1 */
8039 rtx op0 = XEXP (SET_SRC (pat), 0);
8040 rtx op1 = XEXP (SET_SRC (pat), 1);
8041 if (GET_CODE (op1) == CONST_INT)
8042 return INTVAL (op1) >= 0;
8043 if (GET_CODE (op0) != REG)
8045 if (sparc_check_64 (op0, insn) == 1)
8047 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
8052 rtx op0 = XEXP (SET_SRC (pat), 0);
8053 rtx op1 = XEXP (SET_SRC (pat), 1);
8054 if (GET_CODE (op0) != REG || sparc_check_64 (op0, insn) <= 0)
8056 if (GET_CODE (op1) == CONST_INT)
8057 return INTVAL (op1) >= 0;
8058 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
8061 return GET_MODE (SET_SRC (pat)) == SImode;
8062 /* Positive integers leave the high bits zero. */
8064 return ! (CONST_DOUBLE_LOW (SET_SRC (pat)) & 0x80000000);
8066 return ! (INTVAL (SET_SRC (pat)) & 0x80000000);
8069 return - (GET_MODE (SET_SRC (pat)) == SImode);
8071 return sparc_check_64 (SET_SRC (pat), insn);
8077 /* We _ought_ to have only one kind per function, but... */
8078 static GTY(()) rtx sparc_addr_diff_list;
8079 static GTY(()) rtx sparc_addr_list;
8082 sparc_defer_case_vector (rtx lab, rtx vec, int diff)
8084 vec = gen_rtx_EXPR_LIST (VOIDmode, lab, vec);
8086 sparc_addr_diff_list
8087 = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_diff_list);
8089 sparc_addr_list = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_list);
8093 sparc_output_addr_vec (rtx vec)
8095 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
8096 int idx, vlen = XVECLEN (body, 0);
8098 #ifdef ASM_OUTPUT_ADDR_VEC_START
8099 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
8102 #ifdef ASM_OUTPUT_CASE_LABEL
8103 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
8106 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
8109 for (idx = 0; idx < vlen; idx++)
8111 ASM_OUTPUT_ADDR_VEC_ELT
8112 (asm_out_file, CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 0, idx), 0)));
8115 #ifdef ASM_OUTPUT_ADDR_VEC_END
8116 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
8121 sparc_output_addr_diff_vec (rtx vec)
8123 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
8124 rtx base = XEXP (XEXP (body, 0), 0);
8125 int idx, vlen = XVECLEN (body, 1);
8127 #ifdef ASM_OUTPUT_ADDR_VEC_START
8128 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
8131 #ifdef ASM_OUTPUT_CASE_LABEL
8132 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
8135 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
8138 for (idx = 0; idx < vlen; idx++)
8140 ASM_OUTPUT_ADDR_DIFF_ELT
8143 CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 1, idx), 0)),
8144 CODE_LABEL_NUMBER (base));
8147 #ifdef ASM_OUTPUT_ADDR_VEC_END
8148 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
8153 sparc_output_deferred_case_vectors (void)
8158 if (sparc_addr_list == NULL_RTX
8159 && sparc_addr_diff_list == NULL_RTX)
8162 /* Align to cache line in the function's code section. */
8163 function_section (current_function_decl);
8165 align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
8167 ASM_OUTPUT_ALIGN (asm_out_file, align);
8169 for (t = sparc_addr_list; t ; t = XEXP (t, 1))
8170 sparc_output_addr_vec (XEXP (t, 0));
8171 for (t = sparc_addr_diff_list; t ; t = XEXP (t, 1))
8172 sparc_output_addr_diff_vec (XEXP (t, 0));
8174 sparc_addr_list = sparc_addr_diff_list = NULL_RTX;
8177 /* Return 0 if the high 32 bits of X (the low word of X, if DImode) are
8178 unknown. Return 1 if the high bits are zero, -1 if the register is
8181 sparc_check_64 (rtx x, rtx insn)
8183 /* If a register is set only once it is safe to ignore insns this
8184 code does not know how to handle. The loop will either recognize
8185 the single set and return the correct value or fail to recognize
8190 if (GET_CODE (x) != REG)
8193 if (GET_MODE (x) == DImode)
8194 y = gen_rtx_REG (SImode, REGNO (x) + WORDS_BIG_ENDIAN);
8196 if (flag_expensive_optimizations
8197 && REG_N_SETS (REGNO (y)) == 1)
8203 insn = get_last_insn_anywhere ();
8208 while ((insn = PREV_INSN (insn)))
8210 switch (GET_CODE (insn))
8223 rtx pat = PATTERN (insn);
8224 if (GET_CODE (pat) != SET)
8226 if (rtx_equal_p (x, SET_DEST (pat)))
8227 return set_extends (insn);
8228 if (y && rtx_equal_p (y, SET_DEST (pat)))
8229 return set_extends (insn);
8230 if (reg_overlap_mentioned_p (SET_DEST (pat), y))
8238 /* Returns assembly code to perform a DImode shift using
8239 a 64-bit global or out register on SPARC-V8+. */
8241 output_v8plus_shift (rtx *operands, rtx insn, const char *opcode)
8243 static char asm_code[60];
8245 /* The scratch register is only required when the destination
8246 register is not a 64-bit global or out register. */
8247 if (which_alternative != 2)
8248 operands[3] = operands[0];
8250 /* We can only shift by constants <= 63. */
8251 if (GET_CODE (operands[2]) == CONST_INT)
8252 operands[2] = GEN_INT (INTVAL (operands[2]) & 0x3f);
8254 if (GET_CODE (operands[1]) == CONST_INT)
8256 output_asm_insn ("mov\t%1, %3", operands);
8260 output_asm_insn ("sllx\t%H1, 32, %3", operands);
8261 if (sparc_check_64 (operands[1], insn) <= 0)
8262 output_asm_insn ("srl\t%L1, 0, %L1", operands);
8263 output_asm_insn ("or\t%L1, %3, %3", operands);
8266 strcpy(asm_code, opcode);
8268 if (which_alternative != 2)
8269 return strcat (asm_code, "\t%0, %2, %L0\n\tsrlx\t%L0, 32, %H0");
8271 return strcat (asm_code, "\t%3, %2, %3\n\tsrlx\t%3, 32, %H0\n\tmov\t%3, %L0");
8274 /* Output rtl to increment the profiler label LABELNO
8275 for profiling a function entry. */
8278 sparc_profile_hook (int labelno)
8283 ASM_GENERATE_INTERNAL_LABEL (buf, "LP", labelno);
8284 lab = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (buf));
8285 fun = gen_rtx_SYMBOL_REF (Pmode, MCOUNT_FUNCTION);
8287 emit_library_call (fun, LCT_NORMAL, VOIDmode, 1, lab, Pmode);
8290 #ifdef OBJECT_FORMAT_ELF
8292 sparc_elf_asm_named_section (const char *name, unsigned int flags,
8295 if (flags & SECTION_MERGE)
8297 /* entsize cannot be expressed in this section attributes
8299 default_elf_asm_named_section (name, flags, decl);
8303 fprintf (asm_out_file, "\t.section\t\"%s\"", name);
8305 if (!(flags & SECTION_DEBUG))
8306 fputs (",#alloc", asm_out_file);
8307 if (flags & SECTION_WRITE)
8308 fputs (",#write", asm_out_file);
8309 if (flags & SECTION_TLS)
8310 fputs (",#tls", asm_out_file);
8311 if (flags & SECTION_CODE)
8312 fputs (",#execinstr", asm_out_file);
8314 /* ??? Handle SECTION_BSS. */
8316 fputc ('\n', asm_out_file);
8318 #endif /* OBJECT_FORMAT_ELF */
8320 /* We do not allow indirect calls to be optimized into sibling calls.
8322 We cannot use sibling calls when delayed branches are disabled
8323 because they will likely require the call delay slot to be filled.
8325 Also, on SPARC 32-bit we cannot emit a sibling call when the
8326 current function returns a structure. This is because the "unimp
8327 after call" convention would cause the callee to return to the
8328 wrong place. The generic code already disallows cases where the
8329 function being called returns a structure.
8331 It may seem strange how this last case could occur. Usually there
8332 is code after the call which jumps to epilogue code which dumps the
8333 return value into the struct return area. That ought to invalidate
8334 the sibling call right? Well, in the C++ case we can end up passing
8335 the pointer to the struct return area to a constructor (which returns
8336 void) and then nothing else happens. Such a sibling call would look
8337 valid without the added check here. */
8339 sparc_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
8342 && flag_delayed_branch
8343 && (TARGET_ARCH64 || ! current_function_returns_struct));
8346 /* libfunc renaming. */
8347 #include "config/gofast.h"
8350 sparc_init_libfuncs (void)
8354 /* Use the subroutines that Sun's library provides for integer
8355 multiply and divide. The `*' prevents an underscore from
8356 being prepended by the compiler. .umul is a little faster
8358 set_optab_libfunc (smul_optab, SImode, "*.umul");
8359 set_optab_libfunc (sdiv_optab, SImode, "*.div");
8360 set_optab_libfunc (udiv_optab, SImode, "*.udiv");
8361 set_optab_libfunc (smod_optab, SImode, "*.rem");
8362 set_optab_libfunc (umod_optab, SImode, "*.urem");
8364 /* TFmode arithmetic. These names are part of the SPARC 32bit ABI. */
8365 set_optab_libfunc (add_optab, TFmode, "_Q_add");
8366 set_optab_libfunc (sub_optab, TFmode, "_Q_sub");
8367 set_optab_libfunc (neg_optab, TFmode, "_Q_neg");
8368 set_optab_libfunc (smul_optab, TFmode, "_Q_mul");
8369 set_optab_libfunc (sdiv_optab, TFmode, "_Q_div");
8371 /* We can define the TFmode sqrt optab only if TARGET_FPU. This
8372 is because with soft-float, the SFmode and DFmode sqrt
8373 instructions will be absent, and the compiler will notice and
8374 try to use the TFmode sqrt instruction for calls to the
8375 builtin function sqrt, but this fails. */
8377 set_optab_libfunc (sqrt_optab, TFmode, "_Q_sqrt");
8379 set_optab_libfunc (eq_optab, TFmode, "_Q_feq");
8380 set_optab_libfunc (ne_optab, TFmode, "_Q_fne");
8381 set_optab_libfunc (gt_optab, TFmode, "_Q_fgt");
8382 set_optab_libfunc (ge_optab, TFmode, "_Q_fge");
8383 set_optab_libfunc (lt_optab, TFmode, "_Q_flt");
8384 set_optab_libfunc (le_optab, TFmode, "_Q_fle");
8386 set_conv_libfunc (sext_optab, TFmode, SFmode, "_Q_stoq");
8387 set_conv_libfunc (sext_optab, TFmode, DFmode, "_Q_dtoq");
8388 set_conv_libfunc (trunc_optab, SFmode, TFmode, "_Q_qtos");
8389 set_conv_libfunc (trunc_optab, DFmode, TFmode, "_Q_qtod");
8391 set_conv_libfunc (sfix_optab, SImode, TFmode, "_Q_qtoi");
8392 set_conv_libfunc (ufix_optab, SImode, TFmode, "_Q_qtou");
8393 set_conv_libfunc (sfloat_optab, TFmode, SImode, "_Q_itoq");
8395 if (DITF_CONVERSION_LIBFUNCS)
8397 set_conv_libfunc (sfix_optab, DImode, TFmode, "_Q_qtoll");
8398 set_conv_libfunc (ufix_optab, DImode, TFmode, "_Q_qtoull");
8399 set_conv_libfunc (sfloat_optab, TFmode, DImode, "_Q_lltoq");
8402 if (SUN_CONVERSION_LIBFUNCS)
8404 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftoll");
8405 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoull");
8406 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtoll");
8407 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoull");
8412 /* In the SPARC 64bit ABI, SImode multiply and divide functions
8413 do not exist in the library. Make sure the compiler does not
8414 emit calls to them by accident. (It should always use the
8415 hardware instructions.) */
8416 set_optab_libfunc (smul_optab, SImode, 0);
8417 set_optab_libfunc (sdiv_optab, SImode, 0);
8418 set_optab_libfunc (udiv_optab, SImode, 0);
8419 set_optab_libfunc (smod_optab, SImode, 0);
8420 set_optab_libfunc (umod_optab, SImode, 0);
8422 if (SUN_INTEGER_MULTIPLY_64)
8424 set_optab_libfunc (smul_optab, DImode, "__mul64");
8425 set_optab_libfunc (sdiv_optab, DImode, "__div64");
8426 set_optab_libfunc (udiv_optab, DImode, "__udiv64");
8427 set_optab_libfunc (smod_optab, DImode, "__rem64");
8428 set_optab_libfunc (umod_optab, DImode, "__urem64");
8431 if (SUN_CONVERSION_LIBFUNCS)
8433 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftol");
8434 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoul");
8435 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtol");
8436 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoul");
8440 gofast_maybe_init_libfuncs ();
8444 sparc_extra_constraint_check (rtx op, int c, int strict)
8449 && (c == 'T' || c == 'U'))
8455 return fp_sethi_p (op);
8458 return fp_mov_p (op);
8461 return fp_high_losum_p (op);
8465 || (GET_CODE (op) == REG
8466 && (REGNO (op) < FIRST_PSEUDO_REGISTER
8467 || reg_renumber[REGNO (op)] >= 0)))
8468 return register_ok_for_ldd (op);
8477 return fp_zero_operand (op, GET_MODE (op));
8483 /* Our memory extra constraints have to emulate the
8484 behavior of 'm' and 'o' in order for reload to work
8486 if (GET_CODE (op) == MEM)
8489 if ((TARGET_ARCH64 || mem_min_alignment (op, 8))
8491 || strict_memory_address_p (Pmode, XEXP (op, 0))))
8496 reload_ok_mem = (reload_in_progress
8497 && GET_CODE (op) == REG
8498 && REGNO (op) >= FIRST_PSEUDO_REGISTER
8499 && reg_renumber [REGNO (op)] < 0);
8502 return reload_ok_mem;
8505 /* ??? This duplicates information provided to the compiler by the
8506 ??? scheduler description. Some day, teach genautomata to output
8507 ??? the latencies and then CSE will just use that. */
8510 sparc_rtx_costs (rtx x, int code, int outer_code, int *total)
8512 enum machine_mode mode = GET_MODE (x);
8513 bool float_mode_p = FLOAT_MODE_P (mode);
8518 if (INTVAL (x) < 0x1000 && INTVAL (x) >= -0x1000)
8536 if (GET_MODE (x) == DImode
8537 && ((XINT (x, 3) == 0
8538 && (unsigned HOST_WIDE_INT) XINT (x, 2) < 0x1000)
8539 || (XINT (x, 3) == -1
8541 && XINT (x, 2) >= -0x1000)))
8548 /* If outer-code was a sign or zero extension, a cost
8549 of COSTS_N_INSNS (1) was already added in. This is
8550 why we are subtracting it back out. */
8551 if (outer_code == ZERO_EXTEND)
8553 *total = sparc_costs->int_zload - COSTS_N_INSNS (1);
8555 else if (outer_code == SIGN_EXTEND)
8557 *total = sparc_costs->int_sload - COSTS_N_INSNS (1);
8559 else if (float_mode_p)
8561 *total = sparc_costs->float_load;
8565 *total = sparc_costs->int_load;
8573 *total = sparc_costs->float_plusminus;
8575 *total = COSTS_N_INSNS (1);
8580 *total = sparc_costs->float_mul;
8581 else if (! TARGET_HARD_MUL)
8582 *total = COSTS_N_INSNS (25);
8588 if (sparc_costs->int_mul_bit_factor)
8592 if (GET_CODE (XEXP (x, 1)) == CONST_INT)
8594 unsigned HOST_WIDE_INT value = INTVAL (XEXP (x, 1));
8595 for (nbits = 0; value != 0; value &= value - 1)
8598 else if (GET_CODE (XEXP (x, 1)) == CONST_DOUBLE
8599 && GET_MODE (XEXP (x, 1)) == DImode)
8601 rtx x1 = XEXP (x, 1);
8602 unsigned HOST_WIDE_INT value1 = XINT (x1, 2);
8603 unsigned HOST_WIDE_INT value2 = XINT (x1, 3);
8605 for (nbits = 0; value1 != 0; value1 &= value1 - 1)
8607 for (; value2 != 0; value2 &= value2 - 1)
8615 bit_cost = (nbits - 3) / sparc_costs->int_mul_bit_factor;
8616 bit_cost = COSTS_N_INSNS (bit_cost);
8620 *total = sparc_costs->int_mulX + bit_cost;
8622 *total = sparc_costs->int_mul + bit_cost;
8629 *total = COSTS_N_INSNS (1) + sparc_costs->shift_penalty;
8639 *total = sparc_costs->float_div_df;
8641 *total = sparc_costs->float_div_sf;
8646 *total = sparc_costs->int_divX;
8648 *total = sparc_costs->int_div;
8655 *total = COSTS_N_INSNS (1);
8662 case UNSIGNED_FLOAT:
8666 case FLOAT_TRUNCATE:
8667 *total = sparc_costs->float_move;
8672 *total = sparc_costs->float_sqrt_df;
8674 *total = sparc_costs->float_sqrt_sf;
8679 *total = sparc_costs->float_cmp;
8681 *total = COSTS_N_INSNS (1);
8686 *total = sparc_costs->float_cmove;
8688 *total = sparc_costs->int_cmove;
8696 /* Emit the sequence of insns SEQ while preserving the register REG. */
8699 emit_and_preserve (rtx seq, rtx reg)
8701 rtx slot = gen_rtx_MEM (word_mode,
8702 plus_constant (stack_pointer_rtx, SPARC_STACK_BIAS));
8704 emit_stack_pointer_decrement (GEN_INT (STACK_BOUNDARY/BITS_PER_UNIT));
8705 emit_insn (gen_rtx_SET (VOIDmode, slot, reg));
8707 emit_insn (gen_rtx_SET (VOIDmode, reg, slot));
8708 emit_stack_pointer_increment (GEN_INT (STACK_BOUNDARY/BITS_PER_UNIT));
8711 /* Output the assembler code for a thunk function. THUNK_DECL is the
8712 declaration for the thunk function itself, FUNCTION is the decl for
8713 the target function. DELTA is an immediate constant offset to be
8714 added to THIS. If VCALL_OFFSET is nonzero, the word at address
8715 (*THIS + VCALL_OFFSET) should be additionally added to THIS. */
8718 sparc_output_mi_thunk (FILE *file, tree thunk_fndecl ATTRIBUTE_UNUSED,
8719 HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset,
8722 rtx this, insn, funexp;
8723 unsigned int int_arg_first;
8725 reload_completed = 1;
8726 epilogue_completed = 1;
8728 reset_block_changes ();
8730 emit_note (NOTE_INSN_PROLOGUE_END);
8732 if (flag_delayed_branch)
8734 /* We will emit a regular sibcall below, so we need to instruct
8735 output_sibcall that we are in a leaf function. */
8736 sparc_leaf_function_p = current_function_uses_only_leaf_regs = 1;
8738 /* This will cause final.c to invoke leaf_renumber_regs so we
8739 must behave as if we were in a not-yet-leafified function. */
8740 int_arg_first = SPARC_INCOMING_INT_ARG_FIRST;
8744 /* We will emit the sibcall manually below, so we will need to
8745 manually spill non-leaf registers. */
8746 sparc_leaf_function_p = current_function_uses_only_leaf_regs = 0;
8748 /* We really are in a leaf function. */
8749 int_arg_first = SPARC_OUTGOING_INT_ARG_FIRST;
8752 /* Find the "this" pointer. Normally in %o0, but in ARCH64 if the function
8753 returns a structure, the structure return pointer is there instead. */
8754 if (TARGET_ARCH64 && aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function))
8755 this = gen_rtx_REG (Pmode, int_arg_first + 1);
8757 this = gen_rtx_REG (Pmode, int_arg_first);
8759 /* Add DELTA. When possible use a plain add, otherwise load it into
8760 a register first. */
8763 rtx delta_rtx = GEN_INT (delta);
8765 if (! SPARC_SIMM13_P (delta))
8767 rtx scratch = gen_rtx_REG (Pmode, 1);
8768 emit_move_insn (scratch, delta_rtx);
8769 delta_rtx = scratch;
8772 /* THIS += DELTA. */
8773 emit_insn (gen_add2_insn (this, delta_rtx));
8776 /* Add the word at address (*THIS + VCALL_OFFSET). */
8779 rtx vcall_offset_rtx = GEN_INT (vcall_offset);
8780 rtx scratch = gen_rtx_REG (Pmode, 1);
8782 if (vcall_offset >= 0)
8785 /* SCRATCH = *THIS. */
8786 emit_move_insn (scratch, gen_rtx_MEM (Pmode, this));
8788 /* Prepare for adding VCALL_OFFSET. The difficulty is that we
8789 may not have any available scratch register at this point. */
8790 if (SPARC_SIMM13_P (vcall_offset))
8792 /* This is the case if ARCH64 (unless -ffixed-g5 is passed). */
8793 else if (! fixed_regs[5]
8794 /* The below sequence is made up of at least 2 insns,
8795 while the default method may need only one. */
8796 && vcall_offset < -8192)
8798 rtx scratch2 = gen_rtx_REG (Pmode, 5);
8799 emit_move_insn (scratch2, vcall_offset_rtx);
8800 vcall_offset_rtx = scratch2;
8804 rtx increment = GEN_INT (-4096);
8806 /* VCALL_OFFSET is a negative number whose typical range can be
8807 estimated as -32768..0 in 32-bit mode. In almost all cases
8808 it is therefore cheaper to emit multiple add insns than
8809 spilling and loading the constant into a register (at least
8811 while (! SPARC_SIMM13_P (vcall_offset))
8813 emit_insn (gen_add2_insn (scratch, increment));
8814 vcall_offset += 4096;
8816 vcall_offset_rtx = GEN_INT (vcall_offset); /* cannot be 0 */
8819 /* SCRATCH = *(*THIS + VCALL_OFFSET). */
8820 emit_move_insn (scratch, gen_rtx_MEM (Pmode,
8821 gen_rtx_PLUS (Pmode,
8823 vcall_offset_rtx)));
8825 /* THIS += *(*THIS + VCALL_OFFSET). */
8826 emit_insn (gen_add2_insn (this, scratch));
8829 /* Generate a tail call to the target function. */
8830 if (! TREE_USED (function))
8832 assemble_external (function);
8833 TREE_USED (function) = 1;
8835 funexp = XEXP (DECL_RTL (function), 0);
8837 if (flag_delayed_branch)
8839 funexp = gen_rtx_MEM (FUNCTION_MODE, funexp);
8840 insn = emit_call_insn (gen_sibcall (funexp));
8841 SIBLING_CALL_P (insn) = 1;
8845 /* The hoops we have to jump through in order to generate a sibcall
8846 without using delay slots... */
8847 rtx spill_reg, seq, scratch = gen_rtx_REG (Pmode, 1);
8851 spill_reg = gen_rtx_REG (word_mode, 15); /* %o7 */
8853 load_pic_register (); /* clobbers %o7 */
8854 scratch = legitimize_pic_address (funexp, Pmode, scratch);
8857 emit_and_preserve (seq, spill_reg);
8859 else if (TARGET_ARCH32)
8861 emit_insn (gen_rtx_SET (VOIDmode,
8863 gen_rtx_HIGH (SImode, funexp)));
8864 emit_insn (gen_rtx_SET (VOIDmode,
8866 gen_rtx_LO_SUM (SImode, scratch, funexp)));
8868 else /* TARGET_ARCH64 */
8870 switch (sparc_cmodel)
8874 /* The destination can serve as a temporary. */
8875 sparc_emit_set_symbolic_const64 (scratch, funexp, scratch);
8880 /* The destination cannot serve as a temporary. */
8881 spill_reg = gen_rtx_REG (DImode, 15); /* %o7 */
8883 sparc_emit_set_symbolic_const64 (scratch, funexp, spill_reg);
8886 emit_and_preserve (seq, spill_reg);
8894 emit_jump_insn (gen_indirect_jump (scratch));
8899 /* Run just enough of rest_of_compilation to get the insns emitted.
8900 There's not really enough bulk here to make other passes such as
8901 instruction scheduling worth while. Note that use_thunk calls
8902 assemble_start_function and assemble_end_function. */
8903 insn = get_insns ();
8904 insn_locators_initialize ();
8905 shorten_branches (insn);
8906 final_start_function (insn, file, 1);
8907 final (insn, file, 1, 0);
8908 final_end_function ();
8910 reload_completed = 0;
8911 epilogue_completed = 0;
8915 /* Return true if sparc_output_mi_thunk would be able to output the
8916 assembler code for the thunk function specified by the arguments
8917 it is passed, and false otherwise. */
8919 sparc_can_output_mi_thunk (tree thunk_fndecl ATTRIBUTE_UNUSED,
8920 HOST_WIDE_INT delta ATTRIBUTE_UNUSED,
8921 HOST_WIDE_INT vcall_offset,
8922 tree function ATTRIBUTE_UNUSED)
8924 /* Bound the loop used in the default method above. */
8925 return (vcall_offset >= -32768 || ! fixed_regs[5]);
8928 /* How to allocate a 'struct machine_function'. */
8930 static struct machine_function *
8931 sparc_init_machine_status (void)
8933 return ggc_alloc_cleared (sizeof (struct machine_function));
8936 /* Locate some local-dynamic symbol still in use by this function
8937 so that we can print its name in local-dynamic base patterns. */
8940 get_some_local_dynamic_name (void)
8944 if (cfun->machine->some_ld_name)
8945 return cfun->machine->some_ld_name;
8947 for (insn = get_insns (); insn ; insn = NEXT_INSN (insn))
8949 && for_each_rtx (&PATTERN (insn), get_some_local_dynamic_name_1, 0))
8950 return cfun->machine->some_ld_name;
8956 get_some_local_dynamic_name_1 (rtx *px, void *data ATTRIBUTE_UNUSED)
8961 && GET_CODE (x) == SYMBOL_REF
8962 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_DYNAMIC)
8964 cfun->machine->some_ld_name = XSTR (x, 0);
8971 /* This is called from dwarf2out.c via ASM_OUTPUT_DWARF_DTPREL.
8972 We need to emit DTP-relative relocations. */
8975 sparc_output_dwarf_dtprel (FILE *file, int size, rtx x)
8980 fputs ("\t.word\t%r_tls_dtpoff32(", file);
8983 fputs ("\t.xword\t%r_tls_dtpoff64(", file);
8988 output_addr_const (file, x);
8992 #include "gt-sparc.h"