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;
275 /* Register we pretend to think the frame pointer is allocated to.
276 Normally, this is %fp, but if we are in a leaf procedure, this
277 is %sp+"something". We record "something" separately as it may
278 be too big for reg+constant addressing. */
279 static rtx frame_base_reg;
280 static HOST_WIDE_INT frame_base_offset;
282 /* 1 if the next opcode is to be specially indented. */
283 int sparc_indent_opcode = 0;
285 static void sparc_init_modes (void);
286 static void scan_record_type (tree, int *, int *, int *);
287 static int function_arg_slotno (const CUMULATIVE_ARGS *, enum machine_mode,
288 tree, int, int, int *, int *);
290 static int supersparc_adjust_cost (rtx, rtx, rtx, int);
291 static int hypersparc_adjust_cost (rtx, rtx, rtx, int);
293 static void sparc_output_addr_vec (rtx);
294 static void sparc_output_addr_diff_vec (rtx);
295 static void sparc_output_deferred_case_vectors (void);
296 static rtx sparc_builtin_saveregs (void);
297 static int epilogue_renumber (rtx *, int);
298 static bool sparc_assemble_integer (rtx, unsigned int, int);
299 static int set_extends (rtx);
300 static void load_pic_register (void);
301 static int save_or_restore_regs (int, int, rtx, int, int);
302 static void emit_save_regs (void);
303 static void emit_restore_regs (void);
304 static void sparc_asm_function_prologue (FILE *, HOST_WIDE_INT);
305 static void sparc_asm_function_epilogue (FILE *, HOST_WIDE_INT);
306 #ifdef OBJECT_FORMAT_ELF
307 static void sparc_elf_asm_named_section (const char *, unsigned int);
310 static int sparc_adjust_cost (rtx, rtx, rtx, int);
311 static int sparc_issue_rate (void);
312 static void sparc_sched_init (FILE *, int, int);
313 static int sparc_use_sched_lookahead (void);
315 static void emit_soft_tfmode_libcall (const char *, int, rtx *);
316 static void emit_soft_tfmode_binop (enum rtx_code, rtx *);
317 static void emit_soft_tfmode_unop (enum rtx_code, rtx *);
318 static void emit_soft_tfmode_cvt (enum rtx_code, rtx *);
319 static void emit_hard_tfmode_operation (enum rtx_code, rtx *);
321 static bool sparc_function_ok_for_sibcall (tree, tree);
322 static void sparc_init_libfuncs (void);
323 static void sparc_output_mi_thunk (FILE *, tree, HOST_WIDE_INT,
324 HOST_WIDE_INT, tree);
325 static struct machine_function * sparc_init_machine_status (void);
326 static bool sparc_cannot_force_const_mem (rtx);
327 static rtx sparc_tls_get_addr (void);
328 static rtx sparc_tls_got (void);
329 static const char *get_some_local_dynamic_name (void);
330 static int get_some_local_dynamic_name_1 (rtx *, void *);
331 static bool sparc_rtx_costs (rtx, int, int, int *);
332 static bool sparc_promote_prototypes (tree);
333 static rtx sparc_struct_value_rtx (tree, int);
334 static bool sparc_return_in_memory (tree, tree);
335 static bool sparc_strict_argument_naming (CUMULATIVE_ARGS *);
336 static tree sparc_gimplify_va_arg (tree, tree, tree *, tree *);
337 static bool sparc_pass_by_reference (CUMULATIVE_ARGS *,
338 enum machine_mode, tree, bool);
339 #ifdef SUBTARGET_ATTRIBUTE_TABLE
340 const struct attribute_spec sparc_attribute_table[];
343 /* Option handling. */
345 /* Code model option as passed by user. */
346 const char *sparc_cmodel_string;
348 enum cmodel sparc_cmodel;
350 char sparc_hard_reg_printed[8];
352 struct sparc_cpu_select sparc_select[] =
354 /* switch name, tune arch */
355 { (char *)0, "default", 1, 1 },
356 { (char *)0, "-mcpu=", 1, 1 },
357 { (char *)0, "-mtune=", 1, 0 },
361 /* CPU type. This is set from TARGET_CPU_DEFAULT and -m{cpu,tune}=xxx. */
362 enum processor_type sparc_cpu;
364 /* Initialize the GCC target structure. */
366 /* The sparc default is to use .half rather than .short for aligned
367 HI objects. Use .word instead of .long on non-ELF systems. */
368 #undef TARGET_ASM_ALIGNED_HI_OP
369 #define TARGET_ASM_ALIGNED_HI_OP "\t.half\t"
370 #ifndef OBJECT_FORMAT_ELF
371 #undef TARGET_ASM_ALIGNED_SI_OP
372 #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t"
375 #undef TARGET_ASM_UNALIGNED_HI_OP
376 #define TARGET_ASM_UNALIGNED_HI_OP "\t.uahalf\t"
377 #undef TARGET_ASM_UNALIGNED_SI_OP
378 #define TARGET_ASM_UNALIGNED_SI_OP "\t.uaword\t"
379 #undef TARGET_ASM_UNALIGNED_DI_OP
380 #define TARGET_ASM_UNALIGNED_DI_OP "\t.uaxword\t"
382 /* The target hook has to handle DI-mode values. */
383 #undef TARGET_ASM_INTEGER
384 #define TARGET_ASM_INTEGER sparc_assemble_integer
386 #undef TARGET_ASM_FUNCTION_PROLOGUE
387 #define TARGET_ASM_FUNCTION_PROLOGUE sparc_asm_function_prologue
388 #undef TARGET_ASM_FUNCTION_EPILOGUE
389 #define TARGET_ASM_FUNCTION_EPILOGUE sparc_asm_function_epilogue
391 #undef TARGET_SCHED_ADJUST_COST
392 #define TARGET_SCHED_ADJUST_COST sparc_adjust_cost
393 #undef TARGET_SCHED_ISSUE_RATE
394 #define TARGET_SCHED_ISSUE_RATE sparc_issue_rate
395 #undef TARGET_SCHED_INIT
396 #define TARGET_SCHED_INIT sparc_sched_init
397 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
398 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD sparc_use_sched_lookahead
400 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
401 #define TARGET_FUNCTION_OK_FOR_SIBCALL sparc_function_ok_for_sibcall
403 #undef TARGET_INIT_LIBFUNCS
404 #define TARGET_INIT_LIBFUNCS sparc_init_libfuncs
407 #undef TARGET_HAVE_TLS
408 #define TARGET_HAVE_TLS true
410 #undef TARGET_CANNOT_FORCE_CONST_MEM
411 #define TARGET_CANNOT_FORCE_CONST_MEM sparc_cannot_force_const_mem
413 #undef TARGET_ASM_OUTPUT_MI_THUNK
414 #define TARGET_ASM_OUTPUT_MI_THUNK sparc_output_mi_thunk
415 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
416 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK default_can_output_mi_thunk_no_vcall
418 #undef TARGET_RTX_COSTS
419 #define TARGET_RTX_COSTS sparc_rtx_costs
420 #undef TARGET_ADDRESS_COST
421 #define TARGET_ADDRESS_COST hook_int_rtx_0
423 /* This is only needed for TARGET_ARCH64, but since PROMOTE_FUNCTION_MODE is a
424 no-op for TARGET_ARCH32 this is ok. Otherwise we'd need to add a runtime
425 test for this value. */
426 #undef TARGET_PROMOTE_FUNCTION_ARGS
427 #define TARGET_PROMOTE_FUNCTION_ARGS hook_bool_tree_true
429 /* This is only needed for TARGET_ARCH64, but since PROMOTE_FUNCTION_MODE is a
430 no-op for TARGET_ARCH32 this is ok. Otherwise we'd need to add a runtime
431 test for this value. */
432 #undef TARGET_PROMOTE_FUNCTION_RETURN
433 #define TARGET_PROMOTE_FUNCTION_RETURN hook_bool_tree_true
435 #undef TARGET_PROMOTE_PROTOTYPES
436 #define TARGET_PROMOTE_PROTOTYPES sparc_promote_prototypes
438 #undef TARGET_STRUCT_VALUE_RTX
439 #define TARGET_STRUCT_VALUE_RTX sparc_struct_value_rtx
440 #undef TARGET_RETURN_IN_MEMORY
441 #define TARGET_RETURN_IN_MEMORY sparc_return_in_memory
442 #undef TARGET_MUST_PASS_IN_STACK
443 #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size
444 #undef TARGET_PASS_BY_REFERENCE
445 #define TARGET_PASS_BY_REFERENCE sparc_pass_by_reference
447 #undef TARGET_EXPAND_BUILTIN_SAVEREGS
448 #define TARGET_EXPAND_BUILTIN_SAVEREGS sparc_builtin_saveregs
449 #undef TARGET_STRICT_ARGUMENT_NAMING
450 #define TARGET_STRICT_ARGUMENT_NAMING sparc_strict_argument_naming
452 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
453 #define TARGET_GIMPLIFY_VA_ARG_EXPR sparc_gimplify_va_arg
455 #undef TARGET_LATE_RTL_PROLOGUE_EPILOGUE
456 #define TARGET_LATE_RTL_PROLOGUE_EPILOGUE true
458 #ifdef SUBTARGET_INSERT_ATTRIBUTES
459 #undef TARGET_INSERT_ATTRIBUTES
460 #define TARGET_INSERT_ATTRIBUTES SUBTARGET_INSERT_ATTRIBUTES
463 #ifdef SUBTARGET_ATTRIBUTE_TABLE
464 #undef TARGET_ATTRIBUTE_TABLE
465 #define TARGET_ATTRIBUTE_TABLE sparc_attribute_table
468 struct gcc_target targetm = TARGET_INITIALIZER;
470 /* Validate and override various options, and do some machine dependent
474 sparc_override_options (void)
476 static struct code_model {
477 const char *const name;
479 } const cmodels[] = {
481 { "medlow", CM_MEDLOW },
482 { "medmid", CM_MEDMID },
483 { "medany", CM_MEDANY },
484 { "embmedany", CM_EMBMEDANY },
487 const struct code_model *cmodel;
488 /* Map TARGET_CPU_DEFAULT to value for -m{arch,tune}=. */
489 static struct cpu_default {
491 const char *const name;
492 } const cpu_default[] = {
493 /* There must be one entry here for each TARGET_CPU value. */
494 { TARGET_CPU_sparc, "cypress" },
495 { TARGET_CPU_sparclet, "tsc701" },
496 { TARGET_CPU_sparclite, "f930" },
497 { TARGET_CPU_v8, "v8" },
498 { TARGET_CPU_hypersparc, "hypersparc" },
499 { TARGET_CPU_sparclite86x, "sparclite86x" },
500 { TARGET_CPU_supersparc, "supersparc" },
501 { TARGET_CPU_v9, "v9" },
502 { TARGET_CPU_ultrasparc, "ultrasparc" },
503 { TARGET_CPU_ultrasparc3, "ultrasparc3" },
506 const struct cpu_default *def;
507 /* Table of values for -m{cpu,tune}=. */
508 static struct cpu_table {
509 const char *const name;
510 const enum processor_type processor;
513 } const cpu_table[] = {
514 { "v7", PROCESSOR_V7, MASK_ISA, 0 },
515 { "cypress", PROCESSOR_CYPRESS, MASK_ISA, 0 },
516 { "v8", PROCESSOR_V8, MASK_ISA, MASK_V8 },
517 /* TI TMS390Z55 supersparc */
518 { "supersparc", PROCESSOR_SUPERSPARC, MASK_ISA, MASK_V8 },
519 { "sparclite", PROCESSOR_SPARCLITE, MASK_ISA, MASK_SPARCLITE },
520 /* The Fujitsu MB86930 is the original sparclite chip, with no fpu.
521 The Fujitsu MB86934 is the recent sparclite chip, with an fpu. */
522 { "f930", PROCESSOR_F930, MASK_ISA|MASK_FPU, MASK_SPARCLITE },
523 { "f934", PROCESSOR_F934, MASK_ISA, MASK_SPARCLITE|MASK_FPU },
524 { "hypersparc", PROCESSOR_HYPERSPARC, MASK_ISA, MASK_V8|MASK_FPU },
525 { "sparclite86x", PROCESSOR_SPARCLITE86X, MASK_ISA|MASK_FPU,
527 { "sparclet", PROCESSOR_SPARCLET, MASK_ISA, MASK_SPARCLET },
529 { "tsc701", PROCESSOR_TSC701, MASK_ISA, MASK_SPARCLET },
530 { "v9", PROCESSOR_V9, MASK_ISA, MASK_V9 },
531 /* TI ultrasparc I, II, IIi */
532 { "ultrasparc", PROCESSOR_ULTRASPARC, MASK_ISA, MASK_V9
533 /* Although insns using %y are deprecated, it is a clear win on current
535 |MASK_DEPRECATED_V8_INSNS},
536 /* TI ultrasparc III */
537 /* ??? Check if %y issue still holds true in ultra3. */
538 { "ultrasparc3", PROCESSOR_ULTRASPARC3, MASK_ISA, MASK_V9|MASK_DEPRECATED_V8_INSNS},
541 const struct cpu_table *cpu;
542 const struct sparc_cpu_select *sel;
545 #ifndef SPARC_BI_ARCH
546 /* Check for unsupported architecture size. */
547 if (! TARGET_64BIT != DEFAULT_ARCH32_P)
548 error ("%s is not supported by this configuration",
549 DEFAULT_ARCH32_P ? "-m64" : "-m32");
552 /* We force all 64bit archs to use 128 bit long double */
553 if (TARGET_64BIT && ! TARGET_LONG_DOUBLE_128)
555 error ("-mlong-double-64 not allowed with -m64");
556 target_flags |= MASK_LONG_DOUBLE_128;
559 /* Code model selection. */
560 sparc_cmodel = SPARC_DEFAULT_CMODEL;
564 sparc_cmodel = CM_32;
567 if (sparc_cmodel_string != NULL)
571 for (cmodel = &cmodels[0]; cmodel->name; cmodel++)
572 if (strcmp (sparc_cmodel_string, cmodel->name) == 0)
574 if (cmodel->name == NULL)
575 error ("bad value (%s) for -mcmodel= switch", sparc_cmodel_string);
577 sparc_cmodel = cmodel->value;
580 error ("-mcmodel= is not supported on 32 bit systems");
583 fpu = TARGET_FPU; /* save current -mfpu status */
585 /* Set the default CPU. */
586 for (def = &cpu_default[0]; def->name; ++def)
587 if (def->cpu == TARGET_CPU_DEFAULT)
591 sparc_select[0].string = def->name;
593 for (sel = &sparc_select[0]; sel->name; ++sel)
597 for (cpu = &cpu_table[0]; cpu->name; ++cpu)
598 if (! strcmp (sel->string, cpu->name))
601 sparc_cpu = cpu->processor;
605 target_flags &= ~cpu->disable;
606 target_flags |= cpu->enable;
612 error ("bad value (%s) for %s switch", sel->string, sel->name);
616 /* If -mfpu or -mno-fpu was explicitly used, don't override with
617 the processor default. Clear MASK_FPU_SET to avoid confusing
618 the reverse mapping from switch values to names. */
621 target_flags = (target_flags & ~MASK_FPU) | fpu;
622 target_flags &= ~MASK_FPU_SET;
625 /* Don't allow -mvis if FPU is disabled. */
627 target_flags &= ~MASK_VIS;
629 /* -mvis assumes UltraSPARC+, so we are sure v9 instructions
631 -m64 also implies v9. */
632 if (TARGET_VIS || TARGET_ARCH64)
634 target_flags |= MASK_V9;
635 target_flags &= ~(MASK_V8 | MASK_SPARCLET | MASK_SPARCLITE);
638 /* Use the deprecated v8 insns for sparc64 in 32 bit mode. */
639 if (TARGET_V9 && TARGET_ARCH32)
640 target_flags |= MASK_DEPRECATED_V8_INSNS;
642 /* V8PLUS requires V9, makes no sense in 64 bit mode. */
643 if (! TARGET_V9 || TARGET_ARCH64)
644 target_flags &= ~MASK_V8PLUS;
646 /* Don't use stack biasing in 32 bit mode. */
648 target_flags &= ~MASK_STACK_BIAS;
650 /* Supply a default value for align_functions. */
651 if (align_functions == 0
652 && (sparc_cpu == PROCESSOR_ULTRASPARC
653 || sparc_cpu == PROCESSOR_ULTRASPARC3))
654 align_functions = 32;
656 /* Validate PCC_STRUCT_RETURN. */
657 if (flag_pcc_struct_return == DEFAULT_PCC_STRUCT_RETURN)
658 flag_pcc_struct_return = (TARGET_ARCH64 ? 0 : 1);
660 /* Only use .uaxword when compiling for a 64-bit target. */
662 targetm.asm_out.unaligned_op.di = NULL;
664 /* Do various machine dependent initializations. */
667 /* Acquire a unique set number for our register saves and restores. */
668 sparc_sr_alias_set = new_alias_set ();
670 /* Set up function hooks. */
671 init_machine_status = sparc_init_machine_status;
676 case PROCESSOR_CYPRESS:
677 sparc_costs = &cypress_costs;
680 case PROCESSOR_SPARCLITE:
681 case PROCESSOR_SUPERSPARC:
682 sparc_costs = &supersparc_costs;
686 case PROCESSOR_HYPERSPARC:
687 case PROCESSOR_SPARCLITE86X:
688 sparc_costs = &hypersparc_costs;
690 case PROCESSOR_SPARCLET:
691 case PROCESSOR_TSC701:
692 sparc_costs = &sparclet_costs;
695 case PROCESSOR_ULTRASPARC:
696 sparc_costs = &ultrasparc_costs;
698 case PROCESSOR_ULTRASPARC3:
699 sparc_costs = &ultrasparc3_costs;
704 #ifdef SUBTARGET_ATTRIBUTE_TABLE
705 /* Table of valid machine attributes. */
706 const struct attribute_spec sparc_attribute_table[] =
708 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
709 SUBTARGET_ATTRIBUTE_TABLE,
710 { NULL, 0, 0, false, false, false, NULL }
714 /* Miscellaneous utilities. */
716 /* Nonzero if CODE, a comparison, is suitable for use in v9 conditional move
717 or branch on register contents instructions. */
720 v9_regcmp_p (enum rtx_code code)
722 return (code == EQ || code == NE || code == GE || code == LT
723 || code == LE || code == GT);
727 /* Operand constraints. */
729 /* Return nonzero only if OP is a register of mode MODE,
733 reg_or_0_operand (rtx op, enum machine_mode mode)
735 if (register_operand (op, mode))
737 if (op == const0_rtx)
739 if (GET_MODE (op) == VOIDmode && GET_CODE (op) == CONST_DOUBLE
740 && CONST_DOUBLE_HIGH (op) == 0
741 && CONST_DOUBLE_LOW (op) == 0)
743 if (fp_zero_operand (op, mode))
748 /* Return nonzero only if OP is const1_rtx. */
751 const1_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
753 return op == const1_rtx;
756 /* Nonzero if OP is a floating point value with value 0.0. */
759 fp_zero_operand (rtx op, enum machine_mode mode)
761 if (GET_MODE_CLASS (GET_MODE (op)) != MODE_FLOAT)
763 return op == CONST0_RTX (mode);
766 /* Nonzero if OP is a register operand in floating point register. */
769 fp_register_operand (rtx op, enum machine_mode mode)
771 if (! register_operand (op, mode))
773 if (GET_CODE (op) == SUBREG)
774 op = SUBREG_REG (op);
775 return GET_CODE (op) == REG && SPARC_FP_REG_P (REGNO (op));
778 /* Nonzero if OP is a floating point constant which can
779 be loaded into an integer register using a single
780 sethi instruction. */
785 if (GET_CODE (op) == CONST_DOUBLE)
790 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
791 if (REAL_VALUES_EQUAL (r, dconst0) &&
792 ! REAL_VALUE_MINUS_ZERO (r))
794 REAL_VALUE_TO_TARGET_SINGLE (r, i);
795 if (SPARC_SETHI_P (i))
802 /* Nonzero if OP is a floating point constant which can
803 be loaded into an integer register using a single
809 if (GET_CODE (op) == CONST_DOUBLE)
814 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
815 if (REAL_VALUES_EQUAL (r, dconst0) &&
816 ! REAL_VALUE_MINUS_ZERO (r))
818 REAL_VALUE_TO_TARGET_SINGLE (r, i);
819 if (SPARC_SIMM13_P (i))
826 /* Nonzero if OP is a floating point constant which can
827 be loaded into an integer register using a high/losum
828 instruction sequence. */
831 fp_high_losum_p (rtx op)
833 /* The constraints calling this should only be in
834 SFmode move insns, so any constant which cannot
835 be moved using a single insn will do. */
836 if (GET_CODE (op) == CONST_DOUBLE)
841 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
842 if (REAL_VALUES_EQUAL (r, dconst0) &&
843 ! REAL_VALUE_MINUS_ZERO (r))
845 REAL_VALUE_TO_TARGET_SINGLE (r, i);
846 if (! SPARC_SETHI_P (i)
847 && ! SPARC_SIMM13_P (i))
854 /* Nonzero if OP is an integer register. */
857 intreg_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
859 return (register_operand (op, SImode)
860 || (TARGET_ARCH64 && register_operand (op, DImode)));
863 /* Nonzero if OP is a floating point condition code register. */
866 fcc_reg_operand (rtx op, enum machine_mode mode)
868 /* This can happen when recog is called from combine. Op may be a MEM.
869 Fail instead of calling abort in this case. */
870 if (GET_CODE (op) != REG)
873 if (mode != VOIDmode && mode != GET_MODE (op))
876 && (GET_MODE (op) != CCFPmode && GET_MODE (op) != CCFPEmode))
879 #if 0 /* ??? ==> 1 when %fcc0-3 are pseudos first. See gen_compare_reg(). */
880 if (reg_renumber == 0)
881 return REGNO (op) >= FIRST_PSEUDO_REGISTER;
882 return REGNO_OK_FOR_CCFP_P (REGNO (op));
884 return (unsigned) REGNO (op) - SPARC_FIRST_V9_FCC_REG < 4;
888 /* Nonzero if OP is a floating point condition code fcc0 register. */
891 fcc0_reg_operand (rtx op, enum machine_mode mode)
893 /* This can happen when recog is called from combine. Op may be a MEM.
894 Fail instead of calling abort in this case. */
895 if (GET_CODE (op) != REG)
898 if (mode != VOIDmode && mode != GET_MODE (op))
901 && (GET_MODE (op) != CCFPmode && GET_MODE (op) != CCFPEmode))
904 return REGNO (op) == SPARC_FCC_REG;
907 /* Nonzero if OP is an integer or floating point condition code register. */
910 icc_or_fcc_reg_operand (rtx op, enum machine_mode mode)
912 if (GET_CODE (op) == REG && REGNO (op) == SPARC_ICC_REG)
914 if (mode != VOIDmode && mode != GET_MODE (op))
917 && GET_MODE (op) != CCmode && GET_MODE (op) != CCXmode)
922 return fcc_reg_operand (op, mode);
925 /* Call insn on SPARC can take a PC-relative constant address, or any regular
929 call_operand (rtx op, enum machine_mode mode)
931 if (GET_CODE (op) != MEM)
934 return (symbolic_operand (op, mode) || memory_address_p (Pmode, op));
938 call_operand_address (rtx op, enum machine_mode mode)
940 return (symbolic_operand (op, mode) || memory_address_p (Pmode, op));
943 /* If OP is a SYMBOL_REF of a thread-local symbol, return its TLS mode,
944 otherwise return 0. */
947 tls_symbolic_operand (rtx op)
949 if (GET_CODE (op) != SYMBOL_REF)
951 return SYMBOL_REF_TLS_MODEL (op);
955 tgd_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
957 return tls_symbolic_operand (op) == TLS_MODEL_GLOBAL_DYNAMIC;
961 tld_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
963 return tls_symbolic_operand (op) == TLS_MODEL_LOCAL_DYNAMIC;
967 tie_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
969 return tls_symbolic_operand (op) == TLS_MODEL_INITIAL_EXEC;
973 tle_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
975 return tls_symbolic_operand (op) == TLS_MODEL_LOCAL_EXEC;
978 /* Returns 1 if OP is either a symbol reference or a sum of a symbol
979 reference and a constant. */
982 symbolic_operand (register rtx op, enum machine_mode mode)
984 enum machine_mode omode = GET_MODE (op);
986 if (omode != mode && omode != VOIDmode && mode != VOIDmode)
989 switch (GET_CODE (op))
992 return !SYMBOL_REF_TLS_MODEL (op);
999 return (((GET_CODE (XEXP (op, 0)) == SYMBOL_REF
1000 && !SYMBOL_REF_TLS_MODEL (XEXP (op, 0)))
1001 || GET_CODE (XEXP (op, 0)) == LABEL_REF)
1002 && GET_CODE (XEXP (op, 1)) == CONST_INT);
1009 /* Return truth value of statement that OP is a symbolic memory
1010 operand of mode MODE. */
1013 symbolic_memory_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1015 if (GET_CODE (op) == SUBREG)
1016 op = SUBREG_REG (op);
1017 if (GET_CODE (op) != MEM)
1020 return ((GET_CODE (op) == SYMBOL_REF && !SYMBOL_REF_TLS_MODEL (op))
1021 || GET_CODE (op) == CONST || GET_CODE (op) == HIGH
1022 || GET_CODE (op) == LABEL_REF);
1025 /* Return truth value of statement that OP is a LABEL_REF of mode MODE. */
1028 label_ref_operand (rtx op, enum machine_mode mode)
1030 if (GET_CODE (op) != LABEL_REF)
1032 if (GET_MODE (op) != mode)
1037 /* Return 1 if the operand is an argument used in generating pic references
1038 in either the medium/low or medium/anywhere code models of sparc64. */
1041 sp64_medium_pic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1043 /* Check for (const (minus (symbol_ref:GOT)
1044 (const (minus (label) (pc))))). */
1045 if (GET_CODE (op) != CONST)
1048 if (GET_CODE (op) != MINUS)
1050 if (GET_CODE (XEXP (op, 0)) != SYMBOL_REF)
1052 /* ??? Ensure symbol is GOT. */
1053 if (GET_CODE (XEXP (op, 1)) != CONST)
1055 if (GET_CODE (XEXP (XEXP (op, 1), 0)) != MINUS)
1060 /* Return 1 if the operand is a data segment reference. This includes
1061 the readonly data segment, or in other words anything but the text segment.
1062 This is needed in the medium/anywhere code model on v9. These values
1063 are accessed with EMBMEDANY_BASE_REG. */
1066 data_segment_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1068 switch (GET_CODE (op))
1071 return ! SYMBOL_REF_FUNCTION_P (op);
1073 /* Assume canonical format of symbol + constant.
1076 return data_segment_operand (XEXP (op, 0), VOIDmode);
1082 /* Return 1 if the operand is a text segment reference.
1083 This is needed in the medium/anywhere code model on v9. */
1086 text_segment_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1088 switch (GET_CODE (op))
1093 return SYMBOL_REF_FUNCTION_P (op);
1095 /* Assume canonical format of symbol + constant.
1098 return text_segment_operand (XEXP (op, 0), VOIDmode);
1104 /* Return 1 if the operand is either a register or a memory operand that is
1108 reg_or_nonsymb_mem_operand (register rtx op, enum machine_mode mode)
1110 if (register_operand (op, mode))
1113 if (memory_operand (op, mode) && ! symbolic_memory_operand (op, mode))
1120 splittable_symbolic_memory_operand (rtx op,
1121 enum machine_mode mode ATTRIBUTE_UNUSED)
1123 if (GET_CODE (op) != MEM)
1125 if (! symbolic_operand (XEXP (op, 0), Pmode))
1131 splittable_immediate_memory_operand (rtx op,
1132 enum machine_mode mode ATTRIBUTE_UNUSED)
1134 if (GET_CODE (op) != MEM)
1136 if (! immediate_operand (XEXP (op, 0), Pmode))
1141 /* Return truth value of whether OP is EQ or NE. */
1144 eq_or_neq (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1146 return (GET_CODE (op) == EQ || GET_CODE (op) == NE);
1149 /* Return 1 if this is a comparison operator, but not an EQ, NE, GEU,
1150 or LTU for non-floating-point. We handle those specially. */
1153 normal_comp_operator (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1157 if (!COMPARISON_P (op))
1160 if (GET_MODE (XEXP (op, 0)) == CCFPmode
1161 || GET_MODE (XEXP (op, 0)) == CCFPEmode)
1164 code = GET_CODE (op);
1165 return (code != NE && code != EQ && code != GEU && code != LTU);
1168 /* Return 1 if this is a comparison operator. This allows the use of
1169 MATCH_OPERATOR to recognize all the branch insns. */
1172 noov_compare_op (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1176 if (!COMPARISON_P (op))
1179 code = GET_CODE (op);
1180 if (GET_MODE (XEXP (op, 0)) == CC_NOOVmode
1181 || GET_MODE (XEXP (op, 0)) == CCX_NOOVmode)
1182 /* These are the only branches which work with CC_NOOVmode. */
1183 return (code == EQ || code == NE || code == GE || code == LT);
1187 /* Return 1 if this is a 64-bit comparison operator. This allows the use of
1188 MATCH_OPERATOR to recognize all the branch insns. */
1191 noov_compare64_op (register rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1198 if (!COMPARISON_P (op))
1201 code = GET_CODE (op);
1202 if (GET_MODE (XEXP (op, 0)) == CCX_NOOVmode)
1203 /* These are the only branches which work with CCX_NOOVmode. */
1204 return (code == EQ || code == NE || code == GE || code == LT);
1205 return (GET_MODE (XEXP (op, 0)) == CCXmode);
1208 /* Nonzero if OP is a comparison operator suitable for use in v9
1209 conditional move or branch on register contents instructions. */
1212 v9_regcmp_op (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1216 if (!COMPARISON_P (op))
1219 code = GET_CODE (op);
1220 return v9_regcmp_p (code);
1223 /* Return 1 if this is a SIGN_EXTEND or ZERO_EXTEND operation. */
1226 extend_op (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1228 return GET_CODE (op) == SIGN_EXTEND || GET_CODE (op) == ZERO_EXTEND;
1231 /* Return nonzero if OP is an operator of mode MODE which can set
1232 the condition codes explicitly. We do not include PLUS and MINUS
1233 because these require CC_NOOVmode, which we handle explicitly. */
1236 cc_arithop (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1238 if (GET_CODE (op) == AND
1239 || GET_CODE (op) == IOR
1240 || GET_CODE (op) == XOR)
1246 /* Return nonzero if OP is an operator of mode MODE which can bitwise
1247 complement its second operand and set the condition codes explicitly. */
1250 cc_arithopn (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1252 /* XOR is not here because combine canonicalizes (xor (not ...) ...)
1253 and (xor ... (not ...)) to (not (xor ...)). */
1254 return (GET_CODE (op) == AND
1255 || GET_CODE (op) == IOR);
1258 /* Return true if OP is a register, or is a CONST_INT that can fit in a
1259 signed 13 bit immediate field. This is an acceptable SImode operand for
1260 most 3 address instructions. */
1263 arith_operand (rtx op, enum machine_mode mode)
1265 if (register_operand (op, mode))
1267 if (GET_CODE (op) != CONST_INT)
1269 return SMALL_INT32 (op);
1272 /* Return true if OP is a constant 4096 */
1275 arith_4096_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1277 if (GET_CODE (op) != CONST_INT)
1280 return INTVAL (op) == 4096;
1283 /* Return true if OP is suitable as second operand for add/sub */
1286 arith_add_operand (rtx op, enum machine_mode mode)
1288 return arith_operand (op, mode) || arith_4096_operand (op, mode);
1291 /* Return true if OP is a CONST_INT or a CONST_DOUBLE which can fit in the
1292 immediate field of OR and XOR instructions. Used for 64-bit
1293 constant formation patterns. */
1295 const64_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1297 return ((GET_CODE (op) == CONST_INT
1298 && SPARC_SIMM13_P (INTVAL (op)))
1299 #if HOST_BITS_PER_WIDE_INT != 64
1300 || (GET_CODE (op) == CONST_DOUBLE
1301 && SPARC_SIMM13_P (CONST_DOUBLE_LOW (op))
1302 && (CONST_DOUBLE_HIGH (op) ==
1303 ((CONST_DOUBLE_LOW (op) & 0x80000000) != 0 ?
1304 (HOST_WIDE_INT)-1 : 0)))
1309 /* The same, but only for sethi instructions. */
1311 const64_high_operand (rtx op, enum machine_mode mode)
1313 return ((GET_CODE (op) == CONST_INT
1314 && (INTVAL (op) & ~(HOST_WIDE_INT)0x3ff) != 0
1315 && SPARC_SETHI_P (INTVAL (op) & GET_MODE_MASK (mode))
1317 || (GET_CODE (op) == CONST_DOUBLE
1318 && CONST_DOUBLE_HIGH (op) == 0
1319 && (CONST_DOUBLE_LOW (op) & ~(HOST_WIDE_INT)0x3ff) != 0
1320 && SPARC_SETHI_P (CONST_DOUBLE_LOW (op))));
1323 /* Return true if OP is a register, or is a CONST_INT that can fit in a
1324 signed 11 bit immediate field. This is an acceptable SImode operand for
1325 the movcc instructions. */
1328 arith11_operand (rtx op, enum machine_mode mode)
1330 return (register_operand (op, mode)
1331 || (GET_CODE (op) == CONST_INT && SPARC_SIMM11_P (INTVAL (op))));
1334 /* Return true if OP is a register, or is a CONST_INT that can fit in a
1335 signed 10 bit immediate field. This is an acceptable SImode operand for
1336 the movrcc instructions. */
1339 arith10_operand (rtx op, enum machine_mode mode)
1341 return (register_operand (op, mode)
1342 || (GET_CODE (op) == CONST_INT && SPARC_SIMM10_P (INTVAL (op))));
1345 /* Return true if OP is a register, is a CONST_INT that fits in a 13 bit
1346 immediate field, or is a CONST_DOUBLE whose both parts fit in a 13 bit
1348 ARCH64: Return true if OP is a register, or is a CONST_INT or CONST_DOUBLE that
1349 can fit in a 13 bit immediate field. This is an acceptable DImode operand
1350 for most 3 address instructions. */
1353 arith_double_operand (rtx op, enum machine_mode mode)
1355 return (register_operand (op, mode)
1356 || (GET_CODE (op) == CONST_INT && SMALL_INT (op))
1358 && GET_CODE (op) == CONST_DOUBLE
1359 && (unsigned HOST_WIDE_INT) (CONST_DOUBLE_LOW (op) + 0x1000) < 0x2000
1360 && (unsigned HOST_WIDE_INT) (CONST_DOUBLE_HIGH (op) + 0x1000) < 0x2000)
1362 && GET_CODE (op) == CONST_DOUBLE
1363 && (unsigned HOST_WIDE_INT) (CONST_DOUBLE_LOW (op) + 0x1000) < 0x2000
1364 && ((CONST_DOUBLE_HIGH (op) == -1
1365 && (CONST_DOUBLE_LOW (op) & 0x1000) == 0x1000)
1366 || (CONST_DOUBLE_HIGH (op) == 0
1367 && (CONST_DOUBLE_LOW (op) & 0x1000) == 0))));
1370 /* Return true if OP is a constant 4096 for DImode on ARCH64 */
1373 arith_double_4096_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1375 return (TARGET_ARCH64 &&
1376 ((GET_CODE (op) == CONST_INT && INTVAL (op) == 4096) ||
1377 (GET_CODE (op) == CONST_DOUBLE &&
1378 CONST_DOUBLE_LOW (op) == 4096 &&
1379 CONST_DOUBLE_HIGH (op) == 0)));
1382 /* Return true if OP is suitable as second operand for add/sub in DImode */
1385 arith_double_add_operand (rtx op, enum machine_mode mode)
1387 return arith_double_operand (op, mode) || arith_double_4096_operand (op, mode);
1390 /* Return true if OP is a register, or is a CONST_INT or CONST_DOUBLE that
1391 can fit in an 11 bit immediate field. This is an acceptable DImode
1392 operand for the movcc instructions. */
1393 /* ??? Replace with arith11_operand? */
1396 arith11_double_operand (rtx op, enum machine_mode mode)
1398 return (register_operand (op, mode)
1399 || (GET_CODE (op) == CONST_DOUBLE
1400 && (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode)
1401 && (unsigned HOST_WIDE_INT) (CONST_DOUBLE_LOW (op) + 0x400) < 0x800
1402 && ((CONST_DOUBLE_HIGH (op) == -1
1403 && (CONST_DOUBLE_LOW (op) & 0x400) == 0x400)
1404 || (CONST_DOUBLE_HIGH (op) == 0
1405 && (CONST_DOUBLE_LOW (op) & 0x400) == 0)))
1406 || (GET_CODE (op) == CONST_INT
1407 && (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode)
1408 && (unsigned HOST_WIDE_INT) (INTVAL (op) + 0x400) < 0x800));
1411 /* Return true if OP is a register, or is a CONST_INT or CONST_DOUBLE that
1412 can fit in an 10 bit immediate field. This is an acceptable DImode
1413 operand for the movrcc instructions. */
1414 /* ??? Replace with arith10_operand? */
1417 arith10_double_operand (rtx op, enum machine_mode mode)
1419 return (register_operand (op, mode)
1420 || (GET_CODE (op) == CONST_DOUBLE
1421 && (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode)
1422 && (unsigned) (CONST_DOUBLE_LOW (op) + 0x200) < 0x400
1423 && ((CONST_DOUBLE_HIGH (op) == -1
1424 && (CONST_DOUBLE_LOW (op) & 0x200) == 0x200)
1425 || (CONST_DOUBLE_HIGH (op) == 0
1426 && (CONST_DOUBLE_LOW (op) & 0x200) == 0)))
1427 || (GET_CODE (op) == CONST_INT
1428 && (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode)
1429 && (unsigned HOST_WIDE_INT) (INTVAL (op) + 0x200) < 0x400));
1432 /* Return truth value of whether OP is an integer which fits the
1433 range constraining immediate operands in most three-address insns,
1434 which have a 13 bit immediate field. */
1437 small_int (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1439 return (GET_CODE (op) == CONST_INT && SMALL_INT (op));
1443 small_int_or_double (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1445 return ((GET_CODE (op) == CONST_INT && SMALL_INT (op))
1446 || (GET_CODE (op) == CONST_DOUBLE
1447 && CONST_DOUBLE_HIGH (op) == 0
1448 && SPARC_SIMM13_P (CONST_DOUBLE_LOW (op))));
1451 /* Recognize operand values for the umul instruction. That instruction sign
1452 extends immediate values just like all other sparc instructions, but
1453 interprets the extended result as an unsigned number. */
1456 uns_small_int (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1458 #if HOST_BITS_PER_WIDE_INT > 32
1459 /* All allowed constants will fit a CONST_INT. */
1460 return (GET_CODE (op) == CONST_INT
1461 && ((INTVAL (op) >= 0 && INTVAL (op) < 0x1000)
1462 || (INTVAL (op) >= 0xFFFFF000
1463 && INTVAL (op) <= 0xFFFFFFFF)));
1465 return ((GET_CODE (op) == CONST_INT && (unsigned) INTVAL (op) < 0x1000)
1466 || (GET_CODE (op) == CONST_DOUBLE
1467 && CONST_DOUBLE_HIGH (op) == 0
1468 && (unsigned) CONST_DOUBLE_LOW (op) - 0xFFFFF000 < 0x1000));
1473 uns_arith_operand (rtx op, enum machine_mode mode)
1475 return register_operand (op, mode) || uns_small_int (op, mode);
1478 /* Return truth value of statement that OP is a call-clobbered register. */
1480 clobbered_register (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
1482 return (GET_CODE (op) == REG && call_used_regs[REGNO (op)]);
1485 /* Return 1 if OP is a valid operand for the source of a move insn. */
1488 input_operand (rtx op, enum machine_mode mode)
1490 /* If both modes are non-void they must be the same. */
1491 if (mode != VOIDmode && GET_MODE (op) != VOIDmode && mode != GET_MODE (op))
1494 /* Allow any one instruction integer constant, and all CONST_INT
1495 variants when we are working in DImode and !arch64. */
1496 if (GET_MODE_CLASS (mode) == MODE_INT
1497 && ((GET_CODE (op) == CONST_INT
1498 && (SPARC_SETHI_P (INTVAL (op) & GET_MODE_MASK (mode))
1499 || SPARC_SIMM13_P (INTVAL (op))
1501 && ! TARGET_ARCH64)))
1503 && GET_CODE (op) == CONST_DOUBLE
1504 && ((CONST_DOUBLE_HIGH (op) == 0
1505 && SPARC_SETHI_P (CONST_DOUBLE_LOW (op)))
1507 #if HOST_BITS_PER_WIDE_INT == 64
1508 (CONST_DOUBLE_HIGH (op) == 0
1509 && SPARC_SIMM13_P (CONST_DOUBLE_LOW (op)))
1511 (SPARC_SIMM13_P (CONST_DOUBLE_LOW (op))
1512 && (((CONST_DOUBLE_LOW (op) & 0x80000000) == 0
1513 && CONST_DOUBLE_HIGH (op) == 0)
1514 || (CONST_DOUBLE_HIGH (op) == -1
1515 && CONST_DOUBLE_LOW (op) & 0x80000000) != 0))
1520 /* If !arch64 and this is a DImode const, allow it so that
1521 the splits can be generated. */
1524 && GET_CODE (op) == CONST_DOUBLE)
1527 if (register_operand (op, mode))
1530 if (GET_MODE_CLASS (mode) == MODE_FLOAT
1531 && GET_CODE (op) == CONST_DOUBLE)
1534 /* If this is a SUBREG, look inside so that we handle
1535 paradoxical ones. */
1536 if (GET_CODE (op) == SUBREG)
1537 op = SUBREG_REG (op);
1539 /* Check for valid MEM forms. */
1540 if (GET_CODE (op) == MEM)
1542 rtx inside = XEXP (op, 0);
1544 if (GET_CODE (inside) == LO_SUM)
1546 /* We can't allow these because all of the splits
1547 (eventually as they trickle down into DFmode
1548 splits) require offsettable memory references. */
1550 && GET_MODE (op) == TFmode)
1553 return (register_operand (XEXP (inside, 0), Pmode)
1554 && CONSTANT_P (XEXP (inside, 1)));
1556 return memory_address_p (mode, inside);
1562 /* Return 1 if OP is valid for the lhs of a compare insn. */
1565 compare_operand (rtx op, enum machine_mode mode)
1567 if (GET_CODE (op) == ZERO_EXTRACT)
1568 return (register_operand (XEXP (op, 0), mode)
1569 && small_int_or_double (XEXP (op, 1), mode)
1570 && small_int_or_double (XEXP (op, 2), mode)
1571 /* This matches cmp_zero_extract. */
1573 && ((GET_CODE (XEXP (op, 2)) == CONST_INT
1574 && INTVAL (XEXP (op, 2)) > 19)
1575 || (GET_CODE (XEXP (op, 2)) == CONST_DOUBLE
1576 && CONST_DOUBLE_LOW (XEXP (op, 2)) > 19)))
1577 /* This matches cmp_zero_extract_sp64. */
1580 && ((GET_CODE (XEXP (op, 2)) == CONST_INT
1581 && INTVAL (XEXP (op, 2)) > 51)
1582 || (GET_CODE (XEXP (op, 2)) == CONST_DOUBLE
1583 && CONST_DOUBLE_LOW (XEXP (op, 2)) > 51)))));
1585 return register_operand (op, mode);
1589 /* We know it can't be done in one insn when we get here,
1590 the movsi expander guarantees this. */
1592 sparc_emit_set_const32 (rtx op0, rtx op1)
1594 enum machine_mode mode = GET_MODE (op0);
1597 if (GET_CODE (op1) == CONST_INT)
1599 HOST_WIDE_INT value = INTVAL (op1);
1601 if (SPARC_SETHI_P (value & GET_MODE_MASK (mode))
1602 || SPARC_SIMM13_P (value))
1606 /* Full 2-insn decomposition is needed. */
1607 if (reload_in_progress || reload_completed)
1610 temp = gen_reg_rtx (mode);
1612 if (GET_CODE (op1) == CONST_INT)
1614 /* Emit them as real moves instead of a HIGH/LO_SUM,
1615 this way CSE can see everything and reuse intermediate
1616 values if it wants. */
1618 && HOST_BITS_PER_WIDE_INT != 64
1619 && (INTVAL (op1) & 0x80000000) != 0)
1620 emit_insn (gen_rtx_SET
1622 immed_double_const (INTVAL (op1) & ~(HOST_WIDE_INT)0x3ff,
1625 emit_insn (gen_rtx_SET (VOIDmode, temp,
1626 GEN_INT (INTVAL (op1)
1627 & ~(HOST_WIDE_INT)0x3ff)));
1629 emit_insn (gen_rtx_SET (VOIDmode,
1631 gen_rtx_IOR (mode, temp,
1632 GEN_INT (INTVAL (op1) & 0x3ff))));
1636 /* A symbol, emit in the traditional way. */
1637 emit_insn (gen_rtx_SET (VOIDmode, temp,
1638 gen_rtx_HIGH (mode, op1)));
1639 emit_insn (gen_rtx_SET (VOIDmode,
1640 op0, gen_rtx_LO_SUM (mode, temp, op1)));
1646 /* Load OP1, a symbolic 64-bit constant, into OP0, a DImode register.
1647 If TEMP is non-zero, we are forbidden to use any other scratch
1648 registers. Otherwise, we are allowed to generate them as needed.
1650 Note that TEMP may have TImode if the code model is TARGET_CM_MEDANY
1651 or TARGET_CM_EMBMEDANY (see the reload_indi and reload_outdi patterns). */
1653 sparc_emit_set_symbolic_const64 (rtx op0, rtx op1, rtx temp)
1655 rtx temp1, temp2, temp3, temp4, temp5;
1658 if (temp && GET_MODE (temp) == TImode)
1661 temp = gen_rtx_REG (DImode, REGNO (temp));
1664 /* SPARC-V9 code-model support. */
1665 switch (sparc_cmodel)
1668 /* The range spanned by all instructions in the object is less
1669 than 2^31 bytes (2GB) and the distance from any instruction
1670 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1671 than 2^31 bytes (2GB).
1673 The executable must be in the low 4TB of the virtual address
1676 sethi %hi(symbol), %temp1
1677 or %temp1, %lo(symbol), %reg */
1679 temp1 = temp; /* op0 is allowed. */
1681 temp1 = gen_reg_rtx (DImode);
1683 emit_insn (gen_rtx_SET (VOIDmode, temp1, gen_rtx_HIGH (DImode, op1)));
1684 emit_insn (gen_rtx_SET (VOIDmode, op0, gen_rtx_LO_SUM (DImode, temp1, op1)));
1688 /* The range spanned by all instructions in the object is less
1689 than 2^31 bytes (2GB) and the distance from any instruction
1690 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1691 than 2^31 bytes (2GB).
1693 The executable must be in the low 16TB of the virtual address
1696 sethi %h44(symbol), %temp1
1697 or %temp1, %m44(symbol), %temp2
1698 sllx %temp2, 12, %temp3
1699 or %temp3, %l44(symbol), %reg */
1704 temp3 = temp; /* op0 is allowed. */
1708 temp1 = gen_reg_rtx (DImode);
1709 temp2 = gen_reg_rtx (DImode);
1710 temp3 = gen_reg_rtx (DImode);
1713 emit_insn (gen_seth44 (temp1, op1));
1714 emit_insn (gen_setm44 (temp2, temp1, op1));
1715 emit_insn (gen_rtx_SET (VOIDmode, temp3,
1716 gen_rtx_ASHIFT (DImode, temp2, GEN_INT (12))));
1717 emit_insn (gen_setl44 (op0, temp3, op1));
1721 /* The range spanned by all instructions in the object is less
1722 than 2^31 bytes (2GB) and the distance from any instruction
1723 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1724 than 2^31 bytes (2GB).
1726 The executable can be placed anywhere in the virtual address
1729 sethi %hh(symbol), %temp1
1730 sethi %lm(symbol), %temp2
1731 or %temp1, %hm(symbol), %temp3
1732 sllx %temp3, 32, %temp4
1733 or %temp4, %temp2, %temp5
1734 or %temp5, %lo(symbol), %reg */
1737 /* It is possible that one of the registers we got for operands[2]
1738 might coincide with that of operands[0] (which is why we made
1739 it TImode). Pick the other one to use as our scratch. */
1740 if (rtx_equal_p (temp, op0))
1743 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1748 temp2 = temp; /* op0 is _not_ allowed, see above. */
1755 temp1 = gen_reg_rtx (DImode);
1756 temp2 = gen_reg_rtx (DImode);
1757 temp3 = gen_reg_rtx (DImode);
1758 temp4 = gen_reg_rtx (DImode);
1759 temp5 = gen_reg_rtx (DImode);
1762 emit_insn (gen_sethh (temp1, op1));
1763 emit_insn (gen_setlm (temp2, op1));
1764 emit_insn (gen_sethm (temp3, temp1, op1));
1765 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1766 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1767 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1768 gen_rtx_PLUS (DImode, temp4, temp2)));
1769 emit_insn (gen_setlo (op0, temp5, op1));
1773 /* Old old old backwards compatibility kruft here.
1774 Essentially it is MEDLOW with a fixed 64-bit
1775 virtual base added to all data segment addresses.
1776 Text-segment stuff is computed like MEDANY, we can't
1777 reuse the code above because the relocation knobs
1780 Data segment: sethi %hi(symbol), %temp1
1781 add %temp1, EMBMEDANY_BASE_REG, %temp2
1782 or %temp2, %lo(symbol), %reg */
1783 if (data_segment_operand (op1, GET_MODE (op1)))
1787 temp1 = temp; /* op0 is allowed. */
1792 temp1 = gen_reg_rtx (DImode);
1793 temp2 = gen_reg_rtx (DImode);
1796 emit_insn (gen_embmedany_sethi (temp1, op1));
1797 emit_insn (gen_embmedany_brsum (temp2, temp1));
1798 emit_insn (gen_embmedany_losum (op0, temp2, op1));
1801 /* Text segment: sethi %uhi(symbol), %temp1
1802 sethi %hi(symbol), %temp2
1803 or %temp1, %ulo(symbol), %temp3
1804 sllx %temp3, 32, %temp4
1805 or %temp4, %temp2, %temp5
1806 or %temp5, %lo(symbol), %reg */
1811 /* It is possible that one of the registers we got for operands[2]
1812 might coincide with that of operands[0] (which is why we made
1813 it TImode). Pick the other one to use as our scratch. */
1814 if (rtx_equal_p (temp, op0))
1817 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1822 temp2 = temp; /* op0 is _not_ allowed, see above. */
1829 temp1 = gen_reg_rtx (DImode);
1830 temp2 = gen_reg_rtx (DImode);
1831 temp3 = gen_reg_rtx (DImode);
1832 temp4 = gen_reg_rtx (DImode);
1833 temp5 = gen_reg_rtx (DImode);
1836 emit_insn (gen_embmedany_textuhi (temp1, op1));
1837 emit_insn (gen_embmedany_texthi (temp2, op1));
1838 emit_insn (gen_embmedany_textulo (temp3, temp1, op1));
1839 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1840 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1841 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1842 gen_rtx_PLUS (DImode, temp4, temp2)));
1843 emit_insn (gen_embmedany_textlo (op0, temp5, op1));
1852 /* These avoid problems when cross compiling. If we do not
1853 go through all this hair then the optimizer will see
1854 invalid REG_EQUAL notes or in some cases none at all. */
1855 static void sparc_emit_set_safe_HIGH64 (rtx, HOST_WIDE_INT);
1856 static rtx gen_safe_SET64 (rtx, HOST_WIDE_INT);
1857 static rtx gen_safe_OR64 (rtx, HOST_WIDE_INT);
1858 static rtx gen_safe_XOR64 (rtx, HOST_WIDE_INT);
1860 #if HOST_BITS_PER_WIDE_INT == 64
1861 #define GEN_HIGHINT64(__x) GEN_INT ((__x) & ~(HOST_WIDE_INT)0x3ff)
1862 #define GEN_INT64(__x) GEN_INT (__x)
1864 #define GEN_HIGHINT64(__x) \
1865 immed_double_const ((__x) & ~(HOST_WIDE_INT)0x3ff, 0, DImode)
1866 #define GEN_INT64(__x) \
1867 immed_double_const ((__x) & 0xffffffff, \
1868 ((__x) & 0x80000000 ? -1 : 0), DImode)
1871 /* The optimizer is not to assume anything about exactly
1872 which bits are set for a HIGH, they are unspecified.
1873 Unfortunately this leads to many missed optimizations
1874 during CSE. We mask out the non-HIGH bits, and matches
1875 a plain movdi, to alleviate this problem. */
1877 sparc_emit_set_safe_HIGH64 (rtx dest, HOST_WIDE_INT val)
1879 emit_insn (gen_rtx_SET (VOIDmode, dest, GEN_HIGHINT64 (val)));
1883 gen_safe_SET64 (rtx dest, HOST_WIDE_INT val)
1885 return gen_rtx_SET (VOIDmode, dest, GEN_INT64 (val));
1889 gen_safe_OR64 (rtx src, HOST_WIDE_INT val)
1891 return gen_rtx_IOR (DImode, src, GEN_INT64 (val));
1895 gen_safe_XOR64 (rtx src, HOST_WIDE_INT val)
1897 return gen_rtx_XOR (DImode, src, GEN_INT64 (val));
1900 /* Worker routines for 64-bit constant formation on arch64.
1901 One of the key things to be doing in these emissions is
1902 to create as many temp REGs as possible. This makes it
1903 possible for half-built constants to be used later when
1904 such values are similar to something required later on.
1905 Without doing this, the optimizer cannot see such
1908 static void sparc_emit_set_const64_quick1 (rtx, rtx,
1909 unsigned HOST_WIDE_INT, int);
1912 sparc_emit_set_const64_quick1 (rtx op0, rtx temp,
1913 unsigned HOST_WIDE_INT low_bits, int is_neg)
1915 unsigned HOST_WIDE_INT high_bits;
1918 high_bits = (~low_bits) & 0xffffffff;
1920 high_bits = low_bits;
1922 sparc_emit_set_safe_HIGH64 (temp, high_bits);
1925 emit_insn (gen_rtx_SET (VOIDmode, op0,
1926 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1930 /* If we are XOR'ing with -1, then we should emit a one's complement
1931 instead. This way the combiner will notice logical operations
1932 such as ANDN later on and substitute. */
1933 if ((low_bits & 0x3ff) == 0x3ff)
1935 emit_insn (gen_rtx_SET (VOIDmode, op0,
1936 gen_rtx_NOT (DImode, temp)));
1940 emit_insn (gen_rtx_SET (VOIDmode, op0,
1941 gen_safe_XOR64 (temp,
1942 (-(HOST_WIDE_INT)0x400
1943 | (low_bits & 0x3ff)))));
1948 static void sparc_emit_set_const64_quick2 (rtx, rtx, unsigned HOST_WIDE_INT,
1949 unsigned HOST_WIDE_INT, int);
1952 sparc_emit_set_const64_quick2 (rtx op0, rtx temp,
1953 unsigned HOST_WIDE_INT high_bits,
1954 unsigned HOST_WIDE_INT low_immediate,
1959 if ((high_bits & 0xfffffc00) != 0)
1961 sparc_emit_set_safe_HIGH64 (temp, high_bits);
1962 if ((high_bits & ~0xfffffc00) != 0)
1963 emit_insn (gen_rtx_SET (VOIDmode, op0,
1964 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1970 emit_insn (gen_safe_SET64 (temp, high_bits));
1974 /* Now shift it up into place. */
1975 emit_insn (gen_rtx_SET (VOIDmode, op0,
1976 gen_rtx_ASHIFT (DImode, temp2,
1977 GEN_INT (shift_count))));
1979 /* If there is a low immediate part piece, finish up by
1980 putting that in as well. */
1981 if (low_immediate != 0)
1982 emit_insn (gen_rtx_SET (VOIDmode, op0,
1983 gen_safe_OR64 (op0, low_immediate)));
1986 static void sparc_emit_set_const64_longway (rtx, rtx, unsigned HOST_WIDE_INT,
1987 unsigned HOST_WIDE_INT);
1989 /* Full 64-bit constant decomposition. Even though this is the
1990 'worst' case, we still optimize a few things away. */
1992 sparc_emit_set_const64_longway (rtx op0, rtx temp,
1993 unsigned HOST_WIDE_INT high_bits,
1994 unsigned HOST_WIDE_INT low_bits)
1998 if (reload_in_progress || reload_completed)
2001 sub_temp = gen_reg_rtx (DImode);
2003 if ((high_bits & 0xfffffc00) != 0)
2005 sparc_emit_set_safe_HIGH64 (temp, high_bits);
2006 if ((high_bits & ~0xfffffc00) != 0)
2007 emit_insn (gen_rtx_SET (VOIDmode,
2009 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
2015 emit_insn (gen_safe_SET64 (temp, high_bits));
2019 if (!reload_in_progress && !reload_completed)
2021 rtx temp2 = gen_reg_rtx (DImode);
2022 rtx temp3 = gen_reg_rtx (DImode);
2023 rtx temp4 = gen_reg_rtx (DImode);
2025 emit_insn (gen_rtx_SET (VOIDmode, temp4,
2026 gen_rtx_ASHIFT (DImode, sub_temp,
2029 sparc_emit_set_safe_HIGH64 (temp2, low_bits);
2030 if ((low_bits & ~0xfffffc00) != 0)
2032 emit_insn (gen_rtx_SET (VOIDmode, temp3,
2033 gen_safe_OR64 (temp2, (low_bits & 0x3ff))));
2034 emit_insn (gen_rtx_SET (VOIDmode, op0,
2035 gen_rtx_PLUS (DImode, temp4, temp3)));
2039 emit_insn (gen_rtx_SET (VOIDmode, op0,
2040 gen_rtx_PLUS (DImode, temp4, temp2)));
2045 rtx low1 = GEN_INT ((low_bits >> (32 - 12)) & 0xfff);
2046 rtx low2 = GEN_INT ((low_bits >> (32 - 12 - 12)) & 0xfff);
2047 rtx low3 = GEN_INT ((low_bits >> (32 - 12 - 12 - 8)) & 0x0ff);
2050 /* We are in the middle of reload, so this is really
2051 painful. However we do still make an attempt to
2052 avoid emitting truly stupid code. */
2053 if (low1 != const0_rtx)
2055 emit_insn (gen_rtx_SET (VOIDmode, op0,
2056 gen_rtx_ASHIFT (DImode, sub_temp,
2057 GEN_INT (to_shift))));
2058 emit_insn (gen_rtx_SET (VOIDmode, op0,
2059 gen_rtx_IOR (DImode, op0, low1)));
2067 if (low2 != const0_rtx)
2069 emit_insn (gen_rtx_SET (VOIDmode, op0,
2070 gen_rtx_ASHIFT (DImode, sub_temp,
2071 GEN_INT (to_shift))));
2072 emit_insn (gen_rtx_SET (VOIDmode, op0,
2073 gen_rtx_IOR (DImode, op0, low2)));
2081 emit_insn (gen_rtx_SET (VOIDmode, op0,
2082 gen_rtx_ASHIFT (DImode, sub_temp,
2083 GEN_INT (to_shift))));
2084 if (low3 != const0_rtx)
2085 emit_insn (gen_rtx_SET (VOIDmode, op0,
2086 gen_rtx_IOR (DImode, op0, low3)));
2091 /* Analyze a 64-bit constant for certain properties. */
2092 static void analyze_64bit_constant (unsigned HOST_WIDE_INT,
2093 unsigned HOST_WIDE_INT,
2094 int *, int *, int *);
2097 analyze_64bit_constant (unsigned HOST_WIDE_INT high_bits,
2098 unsigned HOST_WIDE_INT low_bits,
2099 int *hbsp, int *lbsp, int *abbasp)
2101 int lowest_bit_set, highest_bit_set, all_bits_between_are_set;
2104 lowest_bit_set = highest_bit_set = -1;
2108 if ((lowest_bit_set == -1)
2109 && ((low_bits >> i) & 1))
2111 if ((highest_bit_set == -1)
2112 && ((high_bits >> (32 - i - 1)) & 1))
2113 highest_bit_set = (64 - i - 1);
2116 && ((highest_bit_set == -1)
2117 || (lowest_bit_set == -1)));
2123 if ((lowest_bit_set == -1)
2124 && ((high_bits >> i) & 1))
2125 lowest_bit_set = i + 32;
2126 if ((highest_bit_set == -1)
2127 && ((low_bits >> (32 - i - 1)) & 1))
2128 highest_bit_set = 32 - i - 1;
2131 && ((highest_bit_set == -1)
2132 || (lowest_bit_set == -1)));
2134 /* If there are no bits set this should have gone out
2135 as one instruction! */
2136 if (lowest_bit_set == -1
2137 || highest_bit_set == -1)
2139 all_bits_between_are_set = 1;
2140 for (i = lowest_bit_set; i <= highest_bit_set; i++)
2144 if ((low_bits & (1 << i)) != 0)
2149 if ((high_bits & (1 << (i - 32))) != 0)
2152 all_bits_between_are_set = 0;
2155 *hbsp = highest_bit_set;
2156 *lbsp = lowest_bit_set;
2157 *abbasp = all_bits_between_are_set;
2160 static int const64_is_2insns (unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT);
2163 const64_is_2insns (unsigned HOST_WIDE_INT high_bits,
2164 unsigned HOST_WIDE_INT low_bits)
2166 int highest_bit_set, lowest_bit_set, all_bits_between_are_set;
2169 || high_bits == 0xffffffff)
2172 analyze_64bit_constant (high_bits, low_bits,
2173 &highest_bit_set, &lowest_bit_set,
2174 &all_bits_between_are_set);
2176 if ((highest_bit_set == 63
2177 || lowest_bit_set == 0)
2178 && all_bits_between_are_set != 0)
2181 if ((highest_bit_set - lowest_bit_set) < 21)
2187 static unsigned HOST_WIDE_INT create_simple_focus_bits (unsigned HOST_WIDE_INT,
2188 unsigned HOST_WIDE_INT,
2191 static unsigned HOST_WIDE_INT
2192 create_simple_focus_bits (unsigned HOST_WIDE_INT high_bits,
2193 unsigned HOST_WIDE_INT low_bits,
2194 int lowest_bit_set, int shift)
2196 HOST_WIDE_INT hi, lo;
2198 if (lowest_bit_set < 32)
2200 lo = (low_bits >> lowest_bit_set) << shift;
2201 hi = ((high_bits << (32 - lowest_bit_set)) << shift);
2206 hi = ((high_bits >> (lowest_bit_set - 32)) << shift);
2213 /* Here we are sure to be arch64 and this is an integer constant
2214 being loaded into a register. Emit the most efficient
2215 insn sequence possible. Detection of all the 1-insn cases
2216 has been done already. */
2218 sparc_emit_set_const64 (rtx op0, rtx op1)
2220 unsigned HOST_WIDE_INT high_bits, low_bits;
2221 int lowest_bit_set, highest_bit_set;
2222 int all_bits_between_are_set;
2225 /* Sanity check that we know what we are working with. */
2226 if (! TARGET_ARCH64)
2229 if (GET_CODE (op0) != SUBREG)
2231 if (GET_CODE (op0) != REG
2232 || (REGNO (op0) >= SPARC_FIRST_FP_REG
2233 && REGNO (op0) <= SPARC_LAST_V9_FP_REG))
2237 if (reload_in_progress || reload_completed)
2240 if (GET_CODE (op1) != CONST_DOUBLE
2241 && GET_CODE (op1) != CONST_INT)
2243 sparc_emit_set_symbolic_const64 (op0, op1, temp);
2248 temp = gen_reg_rtx (DImode);
2250 if (GET_CODE (op1) == CONST_DOUBLE)
2252 #if HOST_BITS_PER_WIDE_INT == 64
2253 high_bits = (CONST_DOUBLE_LOW (op1) >> 32) & 0xffffffff;
2254 low_bits = CONST_DOUBLE_LOW (op1) & 0xffffffff;
2256 high_bits = CONST_DOUBLE_HIGH (op1);
2257 low_bits = CONST_DOUBLE_LOW (op1);
2262 #if HOST_BITS_PER_WIDE_INT == 64
2263 high_bits = ((INTVAL (op1) >> 32) & 0xffffffff);
2264 low_bits = (INTVAL (op1) & 0xffffffff);
2266 high_bits = ((INTVAL (op1) < 0) ?
2269 low_bits = INTVAL (op1);
2273 /* low_bits bits 0 --> 31
2274 high_bits bits 32 --> 63 */
2276 analyze_64bit_constant (high_bits, low_bits,
2277 &highest_bit_set, &lowest_bit_set,
2278 &all_bits_between_are_set);
2280 /* First try for a 2-insn sequence. */
2282 /* These situations are preferred because the optimizer can
2283 * do more things with them:
2285 * sllx %reg, shift, %reg
2287 * srlx %reg, shift, %reg
2288 * 3) mov some_small_const, %reg
2289 * sllx %reg, shift, %reg
2291 if (((highest_bit_set == 63
2292 || lowest_bit_set == 0)
2293 && all_bits_between_are_set != 0)
2294 || ((highest_bit_set - lowest_bit_set) < 12))
2296 HOST_WIDE_INT the_const = -1;
2297 int shift = lowest_bit_set;
2299 if ((highest_bit_set != 63
2300 && lowest_bit_set != 0)
2301 || all_bits_between_are_set == 0)
2304 create_simple_focus_bits (high_bits, low_bits,
2307 else if (lowest_bit_set == 0)
2308 shift = -(63 - highest_bit_set);
2310 if (! SPARC_SIMM13_P (the_const))
2313 emit_insn (gen_safe_SET64 (temp, the_const));
2315 emit_insn (gen_rtx_SET (VOIDmode,
2317 gen_rtx_ASHIFT (DImode,
2321 emit_insn (gen_rtx_SET (VOIDmode,
2323 gen_rtx_LSHIFTRT (DImode,
2325 GEN_INT (-shift))));
2331 /* Now a range of 22 or less bits set somewhere.
2332 * 1) sethi %hi(focus_bits), %reg
2333 * sllx %reg, shift, %reg
2334 * 2) sethi %hi(focus_bits), %reg
2335 * srlx %reg, shift, %reg
2337 if ((highest_bit_set - lowest_bit_set) < 21)
2339 unsigned HOST_WIDE_INT focus_bits =
2340 create_simple_focus_bits (high_bits, low_bits,
2341 lowest_bit_set, 10);
2343 if (! SPARC_SETHI_P (focus_bits))
2346 sparc_emit_set_safe_HIGH64 (temp, focus_bits);
2348 /* If lowest_bit_set == 10 then a sethi alone could have done it. */
2349 if (lowest_bit_set < 10)
2350 emit_insn (gen_rtx_SET (VOIDmode,
2352 gen_rtx_LSHIFTRT (DImode, temp,
2353 GEN_INT (10 - lowest_bit_set))));
2354 else if (lowest_bit_set > 10)
2355 emit_insn (gen_rtx_SET (VOIDmode,
2357 gen_rtx_ASHIFT (DImode, temp,
2358 GEN_INT (lowest_bit_set - 10))));
2364 /* 1) sethi %hi(low_bits), %reg
2365 * or %reg, %lo(low_bits), %reg
2366 * 2) sethi %hi(~low_bits), %reg
2367 * xor %reg, %lo(-0x400 | (low_bits & 0x3ff)), %reg
2370 || high_bits == 0xffffffff)
2372 sparc_emit_set_const64_quick1 (op0, temp, low_bits,
2373 (high_bits == 0xffffffff));
2377 /* Now, try 3-insn sequences. */
2379 /* 1) sethi %hi(high_bits), %reg
2380 * or %reg, %lo(high_bits), %reg
2381 * sllx %reg, 32, %reg
2385 sparc_emit_set_const64_quick2 (op0, temp, high_bits, 0, 32);
2389 /* We may be able to do something quick
2390 when the constant is negated, so try that. */
2391 if (const64_is_2insns ((~high_bits) & 0xffffffff,
2392 (~low_bits) & 0xfffffc00))
2394 /* NOTE: The trailing bits get XOR'd so we need the
2395 non-negated bits, not the negated ones. */
2396 unsigned HOST_WIDE_INT trailing_bits = low_bits & 0x3ff;
2398 if ((((~high_bits) & 0xffffffff) == 0
2399 && ((~low_bits) & 0x80000000) == 0)
2400 || (((~high_bits) & 0xffffffff) == 0xffffffff
2401 && ((~low_bits) & 0x80000000) != 0))
2403 int fast_int = (~low_bits & 0xffffffff);
2405 if ((SPARC_SETHI_P (fast_int)
2406 && (~high_bits & 0xffffffff) == 0)
2407 || SPARC_SIMM13_P (fast_int))
2408 emit_insn (gen_safe_SET64 (temp, fast_int));
2410 sparc_emit_set_const64 (temp, GEN_INT64 (fast_int));
2415 #if HOST_BITS_PER_WIDE_INT == 64
2416 negated_const = GEN_INT (((~low_bits) & 0xfffffc00) |
2417 (((HOST_WIDE_INT)((~high_bits) & 0xffffffff))<<32));
2419 negated_const = immed_double_const ((~low_bits) & 0xfffffc00,
2420 (~high_bits) & 0xffffffff,
2423 sparc_emit_set_const64 (temp, negated_const);
2426 /* If we are XOR'ing with -1, then we should emit a one's complement
2427 instead. This way the combiner will notice logical operations
2428 such as ANDN later on and substitute. */
2429 if (trailing_bits == 0x3ff)
2431 emit_insn (gen_rtx_SET (VOIDmode, op0,
2432 gen_rtx_NOT (DImode, temp)));
2436 emit_insn (gen_rtx_SET (VOIDmode,
2438 gen_safe_XOR64 (temp,
2439 (-0x400 | trailing_bits))));
2444 /* 1) sethi %hi(xxx), %reg
2445 * or %reg, %lo(xxx), %reg
2446 * sllx %reg, yyy, %reg
2448 * ??? This is just a generalized version of the low_bits==0
2449 * thing above, FIXME...
2451 if ((highest_bit_set - lowest_bit_set) < 32)
2453 unsigned HOST_WIDE_INT focus_bits =
2454 create_simple_focus_bits (high_bits, low_bits,
2457 /* We can't get here in this state. */
2458 if (highest_bit_set < 32
2459 || lowest_bit_set >= 32)
2462 /* So what we know is that the set bits straddle the
2463 middle of the 64-bit word. */
2464 sparc_emit_set_const64_quick2 (op0, temp,
2470 /* 1) sethi %hi(high_bits), %reg
2471 * or %reg, %lo(high_bits), %reg
2472 * sllx %reg, 32, %reg
2473 * or %reg, low_bits, %reg
2475 if (SPARC_SIMM13_P(low_bits)
2476 && ((int)low_bits > 0))
2478 sparc_emit_set_const64_quick2 (op0, temp, high_bits, low_bits, 32);
2482 /* The easiest way when all else fails, is full decomposition. */
2484 printf ("sparc_emit_set_const64: Hard constant [%08lx%08lx] neg[%08lx%08lx]\n",
2485 high_bits, low_bits, ~high_bits, ~low_bits);
2487 sparc_emit_set_const64_longway (op0, temp, high_bits, low_bits);
2490 /* Given a comparison code (EQ, NE, etc.) and the first operand of a COMPARE,
2491 return the mode to be used for the comparison. For floating-point,
2492 CCFP[E]mode is used. CC_NOOVmode should be used when the first operand
2493 is a PLUS, MINUS, NEG, or ASHIFT. CCmode should be used when no special
2494 processing is needed. */
2497 select_cc_mode (enum rtx_code op, rtx x, rtx y ATTRIBUTE_UNUSED)
2499 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2525 else if (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS
2526 || GET_CODE (x) == NEG || GET_CODE (x) == ASHIFT)
2528 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
2529 return CCX_NOOVmode;
2535 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
2542 /* X and Y are two things to compare using CODE. Emit the compare insn and
2543 return the rtx for the cc reg in the proper mode. */
2546 gen_compare_reg (enum rtx_code code, rtx x, rtx y)
2548 enum machine_mode mode = SELECT_CC_MODE (code, x, y);
2551 /* ??? We don't have movcc patterns so we cannot generate pseudo regs for the
2552 fcc regs (cse can't tell they're really call clobbered regs and will
2553 remove a duplicate comparison even if there is an intervening function
2554 call - it will then try to reload the cc reg via an int reg which is why
2555 we need the movcc patterns). It is possible to provide the movcc
2556 patterns by using the ldxfsr/stxfsr v9 insns. I tried it: you need two
2557 registers (say %g1,%g5) and it takes about 6 insns. A better fix would be
2558 to tell cse that CCFPE mode registers (even pseudos) are call
2561 /* ??? This is an experiment. Rather than making changes to cse which may
2562 or may not be easy/clean, we do our own cse. This is possible because
2563 we will generate hard registers. Cse knows they're call clobbered (it
2564 doesn't know the same thing about pseudos). If we guess wrong, no big
2565 deal, but if we win, great! */
2567 if (TARGET_V9 && GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2568 #if 1 /* experiment */
2571 /* We cycle through the registers to ensure they're all exercised. */
2572 static int next_fcc_reg = 0;
2573 /* Previous x,y for each fcc reg. */
2574 static rtx prev_args[4][2];
2576 /* Scan prev_args for x,y. */
2577 for (reg = 0; reg < 4; reg++)
2578 if (prev_args[reg][0] == x && prev_args[reg][1] == y)
2583 prev_args[reg][0] = x;
2584 prev_args[reg][1] = y;
2585 next_fcc_reg = (next_fcc_reg + 1) & 3;
2587 cc_reg = gen_rtx_REG (mode, reg + SPARC_FIRST_V9_FCC_REG);
2590 cc_reg = gen_reg_rtx (mode);
2591 #endif /* ! experiment */
2592 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2593 cc_reg = gen_rtx_REG (mode, SPARC_FCC_REG);
2595 cc_reg = gen_rtx_REG (mode, SPARC_ICC_REG);
2597 emit_insn (gen_rtx_SET (VOIDmode, cc_reg,
2598 gen_rtx_COMPARE (mode, x, y)));
2603 /* This function is used for v9 only.
2604 CODE is the code for an Scc's comparison.
2605 OPERANDS[0] is the target of the Scc insn.
2606 OPERANDS[1] is the value we compare against const0_rtx (which hasn't
2607 been generated yet).
2609 This function is needed to turn
2612 (gt (reg:CCX 100 %icc)
2616 (gt:DI (reg:CCX 100 %icc)
2619 IE: The instruction recognizer needs to see the mode of the comparison to
2620 find the right instruction. We could use "gt:DI" right in the
2621 define_expand, but leaving it out allows us to handle DI, SI, etc.
2623 We refer to the global sparc compare operands sparc_compare_op0 and
2624 sparc_compare_op1. */
2627 gen_v9_scc (enum rtx_code compare_code, register rtx *operands)
2632 && (GET_MODE (sparc_compare_op0) == DImode
2633 || GET_MODE (operands[0]) == DImode))
2636 op0 = sparc_compare_op0;
2637 op1 = sparc_compare_op1;
2639 /* Try to use the movrCC insns. */
2641 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT
2642 && op1 == const0_rtx
2643 && v9_regcmp_p (compare_code))
2645 /* Special case for op0 != 0. This can be done with one instruction if
2646 operands[0] == sparc_compare_op0. */
2648 if (compare_code == NE
2649 && GET_MODE (operands[0]) == DImode
2650 && rtx_equal_p (op0, operands[0]))
2652 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
2653 gen_rtx_IF_THEN_ELSE (DImode,
2654 gen_rtx_fmt_ee (compare_code, DImode,
2661 if (reg_overlap_mentioned_p (operands[0], op0))
2663 /* Handle the case where operands[0] == sparc_compare_op0.
2664 We "early clobber" the result. */
2665 op0 = gen_reg_rtx (GET_MODE (sparc_compare_op0));
2666 emit_move_insn (op0, sparc_compare_op0);
2669 emit_insn (gen_rtx_SET (VOIDmode, operands[0], const0_rtx));
2670 if (GET_MODE (op0) != DImode)
2672 temp = gen_reg_rtx (DImode);
2673 convert_move (temp, op0, 0);
2677 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
2678 gen_rtx_IF_THEN_ELSE (GET_MODE (operands[0]),
2679 gen_rtx_fmt_ee (compare_code, DImode,
2687 operands[1] = gen_compare_reg (compare_code, op0, op1);
2689 switch (GET_MODE (operands[1]))
2699 emit_insn (gen_rtx_SET (VOIDmode, operands[0], const0_rtx));
2700 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
2701 gen_rtx_IF_THEN_ELSE (GET_MODE (operands[0]),
2702 gen_rtx_fmt_ee (compare_code,
2703 GET_MODE (operands[1]),
2704 operands[1], const0_rtx),
2705 const1_rtx, operands[0])));
2710 /* Emit a conditional jump insn for the v9 architecture using comparison code
2711 CODE and jump target LABEL.
2712 This function exists to take advantage of the v9 brxx insns. */
2715 emit_v9_brxx_insn (enum rtx_code code, rtx op0, rtx label)
2717 emit_jump_insn (gen_rtx_SET (VOIDmode,
2719 gen_rtx_IF_THEN_ELSE (VOIDmode,
2720 gen_rtx_fmt_ee (code, GET_MODE (op0),
2722 gen_rtx_LABEL_REF (VOIDmode, label),
2726 /* Generate a DFmode part of a hard TFmode register.
2727 REG is the TFmode hard register, LOW is 1 for the
2728 low 64bit of the register and 0 otherwise.
2731 gen_df_reg (rtx reg, int low)
2733 int regno = REGNO (reg);
2735 if ((WORDS_BIG_ENDIAN == 0) ^ (low != 0))
2736 regno += (TARGET_ARCH64 && regno < 32) ? 1 : 2;
2737 return gen_rtx_REG (DFmode, regno);
2740 /* Generate a call to FUNC with OPERANDS. Operand 0 is the return value.
2741 Unlike normal calls, TFmode operands are passed by reference. It is
2742 assumed that no more than 3 operands are required. */
2745 emit_soft_tfmode_libcall (const char *func_name, int nargs, rtx *operands)
2747 rtx ret_slot = NULL, arg[3], func_sym;
2750 /* We only expect to be called for conversions, unary, and binary ops. */
2751 if (nargs < 2 || nargs > 3)
2754 for (i = 0; i < nargs; ++i)
2756 rtx this_arg = operands[i];
2759 /* TFmode arguments and return values are passed by reference. */
2760 if (GET_MODE (this_arg) == TFmode)
2762 int force_stack_temp;
2764 force_stack_temp = 0;
2765 if (TARGET_BUGGY_QP_LIB && i == 0)
2766 force_stack_temp = 1;
2768 if (GET_CODE (this_arg) == MEM
2769 && ! force_stack_temp)
2770 this_arg = XEXP (this_arg, 0);
2771 else if (CONSTANT_P (this_arg)
2772 && ! force_stack_temp)
2774 this_slot = force_const_mem (TFmode, this_arg);
2775 this_arg = XEXP (this_slot, 0);
2779 this_slot = assign_stack_temp (TFmode, GET_MODE_SIZE (TFmode), 0);
2781 /* Operand 0 is the return value. We'll copy it out later. */
2783 emit_move_insn (this_slot, this_arg);
2785 ret_slot = this_slot;
2787 this_arg = XEXP (this_slot, 0);
2794 func_sym = gen_rtx_SYMBOL_REF (Pmode, func_name);
2796 if (GET_MODE (operands[0]) == TFmode)
2799 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 2,
2800 arg[0], GET_MODE (arg[0]),
2801 arg[1], GET_MODE (arg[1]));
2803 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 3,
2804 arg[0], GET_MODE (arg[0]),
2805 arg[1], GET_MODE (arg[1]),
2806 arg[2], GET_MODE (arg[2]));
2809 emit_move_insn (operands[0], ret_slot);
2818 ret = emit_library_call_value (func_sym, operands[0], LCT_NORMAL,
2819 GET_MODE (operands[0]), 1,
2820 arg[1], GET_MODE (arg[1]));
2822 if (ret != operands[0])
2823 emit_move_insn (operands[0], ret);
2827 /* Expand soft-float TFmode calls to sparc abi routines. */
2830 emit_soft_tfmode_binop (enum rtx_code code, rtx *operands)
2852 emit_soft_tfmode_libcall (func, 3, operands);
2856 emit_soft_tfmode_unop (enum rtx_code code, rtx *operands)
2869 emit_soft_tfmode_libcall (func, 2, operands);
2873 emit_soft_tfmode_cvt (enum rtx_code code, rtx *operands)
2880 switch (GET_MODE (operands[1]))
2893 case FLOAT_TRUNCATE:
2894 switch (GET_MODE (operands[0]))
2908 switch (GET_MODE (operands[1]))
2921 case UNSIGNED_FLOAT:
2922 switch (GET_MODE (operands[1]))
2936 switch (GET_MODE (operands[0]))
2950 switch (GET_MODE (operands[0]))
2967 emit_soft_tfmode_libcall (func, 2, operands);
2970 /* Expand a hard-float tfmode operation. All arguments must be in
2974 emit_hard_tfmode_operation (enum rtx_code code, rtx *operands)
2978 if (GET_RTX_CLASS (code) == RTX_UNARY)
2980 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2981 op = gen_rtx_fmt_e (code, GET_MODE (operands[0]), operands[1]);
2985 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2986 operands[2] = force_reg (GET_MODE (operands[2]), operands[2]);
2987 op = gen_rtx_fmt_ee (code, GET_MODE (operands[0]),
2988 operands[1], operands[2]);
2991 if (register_operand (operands[0], VOIDmode))
2994 dest = gen_reg_rtx (GET_MODE (operands[0]));
2996 emit_insn (gen_rtx_SET (VOIDmode, dest, op));
2998 if (dest != operands[0])
2999 emit_move_insn (operands[0], dest);
3003 emit_tfmode_binop (enum rtx_code code, rtx *operands)
3005 if (TARGET_HARD_QUAD)
3006 emit_hard_tfmode_operation (code, operands);
3008 emit_soft_tfmode_binop (code, operands);
3012 emit_tfmode_unop (enum rtx_code code, rtx *operands)
3014 if (TARGET_HARD_QUAD)
3015 emit_hard_tfmode_operation (code, operands);
3017 emit_soft_tfmode_unop (code, operands);
3021 emit_tfmode_cvt (enum rtx_code code, rtx *operands)
3023 if (TARGET_HARD_QUAD)
3024 emit_hard_tfmode_operation (code, operands);
3026 emit_soft_tfmode_cvt (code, operands);
3029 /* Return nonzero if a branch/jump/call instruction will be emitting
3030 nop into its delay slot. */
3033 empty_delay_slot (rtx insn)
3037 /* If no previous instruction (should not happen), return true. */
3038 if (PREV_INSN (insn) == NULL)
3041 seq = NEXT_INSN (PREV_INSN (insn));
3042 if (GET_CODE (PATTERN (seq)) == SEQUENCE)
3048 /* Return nonzero if TRIAL can go into the call delay slot. */
3051 tls_call_delay (rtx trial)
3056 call __tls_get_addr, %tgd_call (foo)
3057 add %l7, %o0, %o0, %tgd_add (foo)
3058 while Sun as/ld does not. */
3059 if (TARGET_GNU_TLS || !TARGET_TLS)
3062 pat = PATTERN (trial);
3063 if (GET_CODE (pat) != SET || GET_CODE (SET_DEST (pat)) != PLUS)
3066 unspec = XEXP (SET_DEST (pat), 1);
3067 if (GET_CODE (unspec) != UNSPEC
3068 || (XINT (unspec, 1) != UNSPEC_TLSGD
3069 && XINT (unspec, 1) != UNSPEC_TLSLDM))
3075 /* Return nonzero if TRIAL, an insn, can be combined with a 'restore'
3076 instruction. RETURN_P is true if the v9 variant 'return' is to be
3077 considered in the test too.
3079 TRIAL must be a SET whose destination is a REG appropriate for the
3080 'restore' instruction or, if RETURN_P is true, for the 'return'
3084 eligible_for_restore_insn (rtx trial, bool return_p)
3086 rtx pat = PATTERN (trial);
3087 rtx src = SET_SRC (pat);
3089 /* The 'restore src,%g0,dest' pattern for word mode and below. */
3090 if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
3091 && arith_operand (src, GET_MODE (src)))
3094 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
3096 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (SImode);
3099 /* The 'restore src,%g0,dest' pattern for double-word mode. */
3100 else if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
3101 && arith_double_operand (src, GET_MODE (src)))
3102 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
3104 /* The 'restore src,%g0,dest' pattern for float if no FPU. */
3105 else if (! TARGET_FPU && register_operand (src, SFmode))
3108 /* The 'restore src,%g0,dest' pattern for double if no FPU. */
3109 else if (! TARGET_FPU && TARGET_ARCH64 && register_operand (src, DFmode))
3112 /* If we have the 'return' instruction, anything that does not use
3113 local or output registers and can go into a delay slot wins. */
3114 else if (return_p && TARGET_V9 && ! epilogue_renumber (&pat, 1)
3115 && (get_attr_in_uncond_branch_delay (trial)
3116 == IN_UNCOND_BRANCH_DELAY_TRUE))
3119 /* The 'restore src1,src2,dest' pattern for SImode. */
3120 else if (GET_CODE (src) == PLUS
3121 && register_operand (XEXP (src, 0), SImode)
3122 && arith_operand (XEXP (src, 1), SImode))
3125 /* The 'restore src1,src2,dest' pattern for DImode. */
3126 else if (GET_CODE (src) == PLUS
3127 && register_operand (XEXP (src, 0), DImode)
3128 && arith_double_operand (XEXP (src, 1), DImode))
3131 /* The 'restore src1,%lo(src2),dest' pattern. */
3132 else if (GET_CODE (src) == LO_SUM
3133 && ! TARGET_CM_MEDMID
3134 && ((register_operand (XEXP (src, 0), SImode)
3135 && immediate_operand (XEXP (src, 1), SImode))
3137 && register_operand (XEXP (src, 0), DImode)
3138 && immediate_operand (XEXP (src, 1), DImode))))
3141 /* The 'restore src,src,dest' pattern. */
3142 else if (GET_CODE (src) == ASHIFT
3143 && (register_operand (XEXP (src, 0), SImode)
3144 || register_operand (XEXP (src, 0), DImode))
3145 && XEXP (src, 1) == const1_rtx)
3151 /* Return nonzero if TRIAL can go into the function return's
3155 eligible_for_return_delay (rtx trial)
3157 int leaf_function_p = current_function_uses_only_leaf_regs;
3160 if (GET_CODE (trial) != INSN || GET_CODE (PATTERN (trial)) != SET)
3163 if (get_attr_length (trial) != 1)
3166 /* If there are any call-saved registers, we should scan TRIAL if it
3167 does not reference them. For now just make it easy. */
3171 /* If the function uses __builtin_eh_return, the eh_return machinery
3172 occupies the delay slot. */
3173 if (current_function_calls_eh_return)
3176 /* In the case of a true leaf function, anything can go into the slot. */
3177 if (leaf_function_p)
3178 return get_attr_in_uncond_branch_delay (trial)
3179 == IN_UNCOND_BRANCH_DELAY_TRUE;
3181 pat = PATTERN (trial);
3183 /* Otherwise, only operations which can be done in tandem with
3184 a `restore' or `return' insn can go into the delay slot. */
3185 if (GET_CODE (SET_DEST (pat)) != REG
3186 || (REGNO (SET_DEST (pat)) >= 8 && REGNO (SET_DEST (pat)) < 24))
3189 /* If this instruction sets up floating point register and we have a return
3190 instruction, it can probably go in. But restore will not work
3192 if (REGNO (SET_DEST (pat)) >= 32)
3194 && ! epilogue_renumber (&pat, 1)
3195 && (get_attr_in_uncond_branch_delay (trial)
3196 == IN_UNCOND_BRANCH_DELAY_TRUE));
3198 return eligible_for_restore_insn (trial, true);
3201 /* Return nonzero if TRIAL can go into the sibling call's
3205 eligible_for_sibcall_delay (rtx trial)
3207 int leaf_function_p = current_function_uses_only_leaf_regs;
3210 if (GET_CODE (trial) != INSN || GET_CODE (PATTERN (trial)) != SET)
3213 if (get_attr_length (trial) != 1)
3216 pat = PATTERN (trial);
3218 if (leaf_function_p)
3220 /* If the tail call is done using the call instruction,
3221 we have to restore %o7 in the delay slot. */
3222 if (LEAF_SIBCALL_SLOT_RESERVED_P)
3225 /* %g1 is used to build the function address */
3226 if (reg_mentioned_p (gen_rtx_REG (Pmode, 1), pat))
3232 /* Otherwise, only operations which can be done in tandem with
3233 a `restore' insn can go into the delay slot. */
3234 if (GET_CODE (SET_DEST (pat)) != REG
3235 || (REGNO (SET_DEST (pat)) >= 8 && REGNO (SET_DEST (pat)) < 24)
3236 || REGNO (SET_DEST (pat)) >= 32)
3239 /* If it mentions %o7, it can't go in, because sibcall will clobber it
3241 if (reg_mentioned_p (gen_rtx_REG (Pmode, 15), pat))
3244 return eligible_for_restore_insn (trial, false);
3248 short_branch (int uid1, int uid2)
3250 int delta = INSN_ADDRESSES (uid1) - INSN_ADDRESSES (uid2);
3252 /* Leave a few words of "slop". */
3253 if (delta >= -1023 && delta <= 1022)
3259 /* Return nonzero if REG is not used after INSN.
3260 We assume REG is a reload reg, and therefore does
3261 not live past labels or calls or jumps. */
3263 reg_unused_after (rtx reg, rtx insn)
3265 enum rtx_code code, prev_code = UNKNOWN;
3267 while ((insn = NEXT_INSN (insn)))
3269 if (prev_code == CALL_INSN && call_used_regs[REGNO (reg)])
3272 code = GET_CODE (insn);
3273 if (GET_CODE (insn) == CODE_LABEL)
3278 rtx set = single_set (insn);
3279 int in_src = set && reg_overlap_mentioned_p (reg, SET_SRC (set));
3282 if (set && reg_overlap_mentioned_p (reg, SET_DEST (set)))
3284 if (set == 0 && reg_overlap_mentioned_p (reg, PATTERN (insn)))
3292 /* Determine if it's legal to put X into the constant pool. This
3293 is not possible if X contains the address of a symbol that is
3294 not constant (TLS) or not known at final link time (PIC). */
3297 sparc_cannot_force_const_mem (rtx x)
3299 switch (GET_CODE (x))
3303 /* Accept all non-symbolic constants. */
3307 /* Labels are OK iff we are non-PIC. */
3308 return flag_pic != 0;
3311 /* 'Naked' TLS symbol references are never OK,
3312 non-TLS symbols are OK iff we are non-PIC. */
3313 if (SYMBOL_REF_TLS_MODEL (x))
3316 return flag_pic != 0;
3319 return sparc_cannot_force_const_mem (XEXP (x, 0));
3322 return sparc_cannot_force_const_mem (XEXP (x, 0))
3323 || sparc_cannot_force_const_mem (XEXP (x, 1));
3331 /* The table we use to reference PIC data. */
3332 static GTY(()) rtx global_offset_table;
3334 /* The function we use to get at it. */
3335 static GTY(()) rtx add_pc_to_pic_symbol;
3336 static GTY(()) char add_pc_to_pic_symbol_name[256];
3338 /* Ensure that we are not using patterns that are not OK with PIC. */
3346 if (GET_CODE (recog_data.operand[i]) == SYMBOL_REF
3347 || (GET_CODE (recog_data.operand[i]) == CONST
3348 && ! (GET_CODE (XEXP (recog_data.operand[i], 0)) == MINUS
3349 && (XEXP (XEXP (recog_data.operand[i], 0), 0)
3350 == global_offset_table)
3351 && (GET_CODE (XEXP (XEXP (recog_data.operand[i], 0), 1))
3360 /* Return true if X is an address which needs a temporary register when
3361 reloaded while generating PIC code. */
3364 pic_address_needs_scratch (rtx x)
3366 /* An address which is a symbolic plus a non SMALL_INT needs a temp reg. */
3367 if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS
3368 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
3369 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
3370 && ! SMALL_INT (XEXP (XEXP (x, 0), 1)))
3376 /* Determine if a given RTX is a valid constant. We already know this
3377 satisfies CONSTANT_P. */
3380 legitimate_constant_p (rtx x)
3384 switch (GET_CODE (x))
3387 /* TLS symbols are not constant. */
3388 if (SYMBOL_REF_TLS_MODEL (x))
3393 inner = XEXP (x, 0);
3395 /* Offsets of TLS symbols are never valid.
3396 Discourage CSE from creating them. */
3397 if (GET_CODE (inner) == PLUS
3398 && tls_symbolic_operand (XEXP (inner, 0)))
3403 if (GET_MODE (x) == VOIDmode)
3406 /* Floating point constants are generally not ok.
3407 The only exception is 0.0 in VIS. */
3409 && (GET_MODE (x) == SFmode
3410 || GET_MODE (x) == DFmode
3411 || GET_MODE (x) == TFmode)
3412 && fp_zero_operand (x, GET_MODE (x)))
3424 /* Determine if a given RTX is a valid constant address. */
3427 constant_address_p (rtx x)
3429 switch (GET_CODE (x))
3437 if (flag_pic && pic_address_needs_scratch (x))
3439 return legitimate_constant_p (x);
3442 return !flag_pic && legitimate_constant_p (x);
3449 /* Nonzero if the constant value X is a legitimate general operand
3450 when generating PIC code. It is given that flag_pic is on and
3451 that X satisfies CONSTANT_P or is a CONST_DOUBLE. */
3454 legitimate_pic_operand_p (rtx x)
3456 if (pic_address_needs_scratch (x))
3458 if (tls_symbolic_operand (x)
3459 || (GET_CODE (x) == CONST
3460 && GET_CODE (XEXP (x, 0)) == PLUS
3461 && tls_symbolic_operand (XEXP (XEXP (x, 0), 0))))
3466 /* Return nonzero if ADDR is a valid memory address.
3467 STRICT specifies whether strict register checking applies. */
3470 legitimate_address_p (enum machine_mode mode, rtx addr, int strict)
3472 rtx rs1 = NULL, rs2 = NULL, imm1 = NULL, imm2;
3474 if (REG_P (addr) || GET_CODE (addr) == SUBREG)
3476 else if (GET_CODE (addr) == PLUS)
3478 rs1 = XEXP (addr, 0);
3479 rs2 = XEXP (addr, 1);
3481 /* Canonicalize. REG comes first, if there are no regs,
3482 LO_SUM comes first. */
3484 && GET_CODE (rs1) != SUBREG
3486 || GET_CODE (rs2) == SUBREG
3487 || (GET_CODE (rs2) == LO_SUM && GET_CODE (rs1) != LO_SUM)))
3489 rs1 = XEXP (addr, 1);
3490 rs2 = XEXP (addr, 0);
3494 && rs1 == pic_offset_table_rtx
3496 && GET_CODE (rs2) != SUBREG
3497 && GET_CODE (rs2) != LO_SUM
3498 && GET_CODE (rs2) != MEM
3499 && !tls_symbolic_operand (rs2)
3500 && (! symbolic_operand (rs2, VOIDmode) || mode == Pmode)
3501 && (GET_CODE (rs2) != CONST_INT || SMALL_INT (rs2)))
3503 || GET_CODE (rs1) == SUBREG)
3504 && RTX_OK_FOR_OFFSET_P (rs2)))
3509 else if ((REG_P (rs1) || GET_CODE (rs1) == SUBREG)
3510 && (REG_P (rs2) || GET_CODE (rs2) == SUBREG))
3512 /* We prohibit REG + REG for TFmode when there are no instructions
3513 which accept REG+REG instructions. We do this because REG+REG
3514 is not an offsetable address. If we get the situation in reload
3515 where source and destination of a movtf pattern are both MEMs with
3516 REG+REG address, then only one of them gets converted to an
3517 offsetable address. */
3519 && !(TARGET_FPU && TARGET_ARCH64 && TARGET_V9
3520 && TARGET_HARD_QUAD))
3523 /* We prohibit REG + REG on ARCH32 if not optimizing for
3524 DFmode/DImode because then mem_min_alignment is likely to be zero
3525 after reload and the forced split would lack a matching splitter
3527 if (TARGET_ARCH32 && !optimize
3528 && (mode == DFmode || mode == DImode))
3531 else if (USE_AS_OFFSETABLE_LO10
3532 && GET_CODE (rs1) == LO_SUM
3534 && ! TARGET_CM_MEDMID
3535 && RTX_OK_FOR_OLO10_P (rs2))
3539 imm1 = XEXP (rs1, 1);
3540 rs1 = XEXP (rs1, 0);
3541 if (! CONSTANT_P (imm1) || tls_symbolic_operand (rs1))
3545 else if (GET_CODE (addr) == LO_SUM)
3547 rs1 = XEXP (addr, 0);
3548 imm1 = XEXP (addr, 1);
3550 if (! CONSTANT_P (imm1) || tls_symbolic_operand (rs1))
3553 /* We can't allow TFmode, because an offset greater than or equal to the
3554 alignment (8) may cause the LO_SUM to overflow if !v9. */
3555 if (mode == TFmode && !TARGET_V9)
3558 else if (GET_CODE (addr) == CONST_INT && SMALL_INT (addr))
3563 if (GET_CODE (rs1) == SUBREG)
3564 rs1 = SUBREG_REG (rs1);
3570 if (GET_CODE (rs2) == SUBREG)
3571 rs2 = SUBREG_REG (rs2);
3578 if (!REGNO_OK_FOR_BASE_P (REGNO (rs1))
3579 || (rs2 && !REGNO_OK_FOR_BASE_P (REGNO (rs2))))
3584 if ((REGNO (rs1) >= 32
3585 && REGNO (rs1) != FRAME_POINTER_REGNUM
3586 && REGNO (rs1) < FIRST_PSEUDO_REGISTER)
3588 && (REGNO (rs2) >= 32
3589 && REGNO (rs2) != FRAME_POINTER_REGNUM
3590 && REGNO (rs2) < FIRST_PSEUDO_REGISTER)))
3596 /* Construct the SYMBOL_REF for the tls_get_offset function. */
3598 static GTY(()) rtx sparc_tls_symbol;
3600 sparc_tls_get_addr (void)
3602 if (!sparc_tls_symbol)
3603 sparc_tls_symbol = gen_rtx_SYMBOL_REF (Pmode, "__tls_get_addr");
3605 return sparc_tls_symbol;
3609 sparc_tls_got (void)
3614 current_function_uses_pic_offset_table = 1;
3615 return pic_offset_table_rtx;
3618 if (!global_offset_table)
3619 global_offset_table = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_");
3620 temp = gen_reg_rtx (Pmode);
3621 emit_move_insn (temp, global_offset_table);
3626 /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute
3627 this (thread-local) address. */
3630 legitimize_tls_address (rtx addr)
3632 rtx temp1, temp2, temp3, ret, o0, got, insn;
3637 if (GET_CODE (addr) == SYMBOL_REF)
3638 switch (SYMBOL_REF_TLS_MODEL (addr))
3640 case TLS_MODEL_GLOBAL_DYNAMIC:
3642 temp1 = gen_reg_rtx (SImode);
3643 temp2 = gen_reg_rtx (SImode);
3644 ret = gen_reg_rtx (Pmode);
3645 o0 = gen_rtx_REG (Pmode, 8);
3646 got = sparc_tls_got ();
3647 emit_insn (gen_tgd_hi22 (temp1, addr));
3648 emit_insn (gen_tgd_lo10 (temp2, temp1, addr));
3651 emit_insn (gen_tgd_add32 (o0, got, temp2, addr));
3652 insn = emit_call_insn (gen_tgd_call32 (o0, sparc_tls_get_addr (),
3657 emit_insn (gen_tgd_add64 (o0, got, temp2, addr));
3658 insn = emit_call_insn (gen_tgd_call64 (o0, sparc_tls_get_addr (),
3661 CALL_INSN_FUNCTION_USAGE (insn)
3662 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_USE (VOIDmode, o0),
3663 CALL_INSN_FUNCTION_USAGE (insn));
3664 insn = get_insns ();
3666 emit_libcall_block (insn, ret, o0, addr);
3669 case TLS_MODEL_LOCAL_DYNAMIC:
3671 temp1 = gen_reg_rtx (SImode);
3672 temp2 = gen_reg_rtx (SImode);
3673 temp3 = gen_reg_rtx (Pmode);
3674 ret = gen_reg_rtx (Pmode);
3675 o0 = gen_rtx_REG (Pmode, 8);
3676 got = sparc_tls_got ();
3677 emit_insn (gen_tldm_hi22 (temp1));
3678 emit_insn (gen_tldm_lo10 (temp2, temp1));
3681 emit_insn (gen_tldm_add32 (o0, got, temp2));
3682 insn = emit_call_insn (gen_tldm_call32 (o0, sparc_tls_get_addr (),
3687 emit_insn (gen_tldm_add64 (o0, got, temp2));
3688 insn = emit_call_insn (gen_tldm_call64 (o0, sparc_tls_get_addr (),
3691 CALL_INSN_FUNCTION_USAGE (insn)
3692 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_USE (VOIDmode, o0),
3693 CALL_INSN_FUNCTION_USAGE (insn));
3694 insn = get_insns ();
3696 emit_libcall_block (insn, temp3, o0,
3697 gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
3698 UNSPEC_TLSLD_BASE));
3699 temp1 = gen_reg_rtx (SImode);
3700 temp2 = gen_reg_rtx (SImode);
3701 emit_insn (gen_tldo_hix22 (temp1, addr));
3702 emit_insn (gen_tldo_lox10 (temp2, temp1, addr));
3704 emit_insn (gen_tldo_add32 (ret, temp3, temp2, addr));
3706 emit_insn (gen_tldo_add64 (ret, temp3, temp2, addr));
3709 case TLS_MODEL_INITIAL_EXEC:
3710 temp1 = gen_reg_rtx (SImode);
3711 temp2 = gen_reg_rtx (SImode);
3712 temp3 = gen_reg_rtx (Pmode);
3713 got = sparc_tls_got ();
3714 emit_insn (gen_tie_hi22 (temp1, addr));
3715 emit_insn (gen_tie_lo10 (temp2, temp1, addr));
3717 emit_insn (gen_tie_ld32 (temp3, got, temp2, addr));
3719 emit_insn (gen_tie_ld64 (temp3, got, temp2, addr));
3722 ret = gen_reg_rtx (Pmode);
3724 emit_insn (gen_tie_add32 (ret, gen_rtx_REG (Pmode, 7),
3727 emit_insn (gen_tie_add64 (ret, gen_rtx_REG (Pmode, 7),
3731 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp3);
3734 case TLS_MODEL_LOCAL_EXEC:
3735 temp1 = gen_reg_rtx (Pmode);
3736 temp2 = gen_reg_rtx (Pmode);
3739 emit_insn (gen_tle_hix22_sp32 (temp1, addr));
3740 emit_insn (gen_tle_lox10_sp32 (temp2, temp1, addr));
3744 emit_insn (gen_tle_hix22_sp64 (temp1, addr));
3745 emit_insn (gen_tle_lox10_sp64 (temp2, temp1, addr));
3747 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp2);
3755 abort (); /* for now ... */
3761 /* Legitimize PIC addresses. If the address is already position-independent,
3762 we return ORIG. Newly generated position-independent addresses go into a
3763 reg. This is REG if nonzero, otherwise we allocate register(s) as
3767 legitimize_pic_address (rtx orig, enum machine_mode mode ATTRIBUTE_UNUSED,
3770 if (GET_CODE (orig) == SYMBOL_REF)
3772 rtx pic_ref, address;
3777 if (reload_in_progress || reload_completed)
3780 reg = gen_reg_rtx (Pmode);
3785 /* If not during reload, allocate another temp reg here for loading
3786 in the address, so that these instructions can be optimized
3788 rtx temp_reg = ((reload_in_progress || reload_completed)
3789 ? reg : gen_reg_rtx (Pmode));
3791 /* Must put the SYMBOL_REF inside an UNSPEC here so that cse
3792 won't get confused into thinking that these two instructions
3793 are loading in the true address of the symbol. If in the
3794 future a PIC rtx exists, that should be used instead. */
3795 if (Pmode == SImode)
3797 emit_insn (gen_movsi_high_pic (temp_reg, orig));
3798 emit_insn (gen_movsi_lo_sum_pic (temp_reg, temp_reg, orig));
3802 emit_insn (gen_movdi_high_pic (temp_reg, orig));
3803 emit_insn (gen_movdi_lo_sum_pic (temp_reg, temp_reg, orig));
3810 pic_ref = gen_const_mem (Pmode,
3811 gen_rtx_PLUS (Pmode,
3812 pic_offset_table_rtx, address));
3813 current_function_uses_pic_offset_table = 1;
3814 insn = emit_move_insn (reg, pic_ref);
3815 /* Put a REG_EQUAL note on this insn, so that it can be optimized
3817 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_EQUAL, orig,
3821 else if (GET_CODE (orig) == CONST)
3825 if (GET_CODE (XEXP (orig, 0)) == PLUS
3826 && XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx)
3831 if (reload_in_progress || reload_completed)
3834 reg = gen_reg_rtx (Pmode);
3837 if (GET_CODE (XEXP (orig, 0)) == PLUS)
3839 base = legitimize_pic_address (XEXP (XEXP (orig, 0), 0), Pmode, reg);
3840 offset = legitimize_pic_address (XEXP (XEXP (orig, 0), 1), Pmode,
3841 base == reg ? 0 : reg);
3846 if (GET_CODE (offset) == CONST_INT)
3848 if (SMALL_INT (offset))
3849 return plus_constant (base, INTVAL (offset));
3850 else if (! reload_in_progress && ! reload_completed)
3851 offset = force_reg (Pmode, offset);
3853 /* If we reach here, then something is seriously wrong. */
3856 return gen_rtx_PLUS (Pmode, base, offset);
3858 else if (GET_CODE (orig) == LABEL_REF)
3859 /* ??? Why do we do this? */
3860 /* Now movsi_pic_label_ref uses it, but we ought to be checking that
3861 the register is live instead, in case it is eliminated. */
3862 current_function_uses_pic_offset_table = 1;
3867 /* Try machine-dependent ways of modifying an illegitimate address X
3868 to be legitimate. If we find one, return the new, valid address.
3870 OLDX is the address as it was before break_out_memory_refs was called.
3871 In some cases it is useful to look at this to decide what needs to be done.
3873 MODE is the mode of the operand pointed to by X. */
3876 legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED, enum machine_mode mode)
3880 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == MULT)
3881 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3882 force_operand (XEXP (x, 0), NULL_RTX));
3883 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == MULT)
3884 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3885 force_operand (XEXP (x, 1), NULL_RTX));
3886 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == PLUS)
3887 x = gen_rtx_PLUS (Pmode, force_operand (XEXP (x, 0), NULL_RTX),
3889 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == PLUS)
3890 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3891 force_operand (XEXP (x, 1), NULL_RTX));
3893 if (x != orig_x && legitimate_address_p (mode, x, FALSE))
3896 if (tls_symbolic_operand (x))
3897 x = legitimize_tls_address (x);
3899 x = legitimize_pic_address (x, mode, 0);
3900 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 1)))
3901 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3902 copy_to_mode_reg (Pmode, XEXP (x, 1)));
3903 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 0)))
3904 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3905 copy_to_mode_reg (Pmode, XEXP (x, 0)));
3906 else if (GET_CODE (x) == SYMBOL_REF
3907 || GET_CODE (x) == CONST
3908 || GET_CODE (x) == LABEL_REF)
3909 x = copy_to_suggested_reg (x, NULL_RTX, Pmode);
3913 /* Emit the special PIC prologue. */
3916 load_pic_register (void)
3918 int orig_flag_pic = flag_pic;
3920 /* If we haven't emitted the special helper function, do so now. */
3921 if (add_pc_to_pic_symbol_name[0] == 0)
3923 const char *pic_name = reg_names[REGNO (pic_offset_table_rtx)];
3926 ASM_GENERATE_INTERNAL_LABEL (add_pc_to_pic_symbol_name, "LADDPC", 0);
3929 align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
3931 ASM_OUTPUT_ALIGN (asm_out_file, align);
3932 ASM_OUTPUT_LABEL (asm_out_file, add_pc_to_pic_symbol_name);
3933 if (flag_delayed_branch)
3934 fprintf (asm_out_file, "\tjmp %%o7+8\n\t add\t%%o7, %s, %s\n",
3935 pic_name, pic_name);
3937 fprintf (asm_out_file, "\tadd\t%%o7, %s, %s\n\tjmp %%o7+8\n\t nop\n",
3938 pic_name, pic_name);
3941 /* Initialize every time through, since we can't easily
3942 know this to be permanent. */
3943 global_offset_table = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_");
3944 add_pc_to_pic_symbol = gen_rtx_SYMBOL_REF (Pmode, add_pc_to_pic_symbol_name);
3947 emit_insn (gen_load_pcrel_sym (pic_offset_table_rtx, global_offset_table,
3948 add_pc_to_pic_symbol));
3949 flag_pic = orig_flag_pic;
3951 /* Need to emit this whether or not we obey regdecls,
3952 since setjmp/longjmp can cause life info to screw up.
3953 ??? In the case where we don't obey regdecls, this is not sufficient
3954 since we may not fall out the bottom. */
3955 emit_insn (gen_rtx_USE (VOIDmode, pic_offset_table_rtx));
3958 /* Return 1 if RTX is a MEM which is known to be aligned to at
3959 least a DESIRED byte boundary. */
3962 mem_min_alignment (rtx mem, int desired)
3964 rtx addr, base, offset;
3966 /* If it's not a MEM we can't accept it. */
3967 if (GET_CODE (mem) != MEM)
3970 addr = XEXP (mem, 0);
3971 base = offset = NULL_RTX;
3972 if (GET_CODE (addr) == PLUS)
3974 if (GET_CODE (XEXP (addr, 0)) == REG)
3976 base = XEXP (addr, 0);
3978 /* What we are saying here is that if the base
3979 REG is aligned properly, the compiler will make
3980 sure any REG based index upon it will be so
3982 if (GET_CODE (XEXP (addr, 1)) == CONST_INT)
3983 offset = XEXP (addr, 1);
3985 offset = const0_rtx;
3988 else if (GET_CODE (addr) == REG)
3991 offset = const0_rtx;
3994 if (base != NULL_RTX)
3996 int regno = REGNO (base);
3998 if (regno != HARD_FRAME_POINTER_REGNUM && regno != STACK_POINTER_REGNUM)
4000 /* Check if the compiler has recorded some information
4001 about the alignment of the base REG. If reload has
4002 completed, we already matched with proper alignments.
4003 If not running global_alloc, reload might give us
4004 unaligned pointer to local stack though. */
4006 && REGNO_POINTER_ALIGN (regno) >= desired * BITS_PER_UNIT)
4007 || (optimize && reload_completed))
4008 && (INTVAL (offset) & (desired - 1)) == 0)
4013 if (((INTVAL (offset) - SPARC_STACK_BIAS) & (desired - 1)) == 0)
4017 else if (! TARGET_UNALIGNED_DOUBLES
4018 || CONSTANT_P (addr)
4019 || GET_CODE (addr) == LO_SUM)
4021 /* Anything else we know is properly aligned unless TARGET_UNALIGNED_DOUBLES
4022 is true, in which case we can only assume that an access is aligned if
4023 it is to a constant address, or the address involves a LO_SUM. */
4027 /* An obviously unaligned address. */
4032 /* Vectors to keep interesting information about registers where it can easily
4033 be got. We used to use the actual mode value as the bit number, but there
4034 are more than 32 modes now. Instead we use two tables: one indexed by
4035 hard register number, and one indexed by mode. */
4037 /* The purpose of sparc_mode_class is to shrink the range of modes so that
4038 they all fit (as bit numbers) in a 32 bit word (again). Each real mode is
4039 mapped into one sparc_mode_class mode. */
4041 enum sparc_mode_class {
4042 S_MODE, D_MODE, T_MODE, O_MODE,
4043 SF_MODE, DF_MODE, TF_MODE, OF_MODE,
4047 /* Modes for single-word and smaller quantities. */
4048 #define S_MODES ((1 << (int) S_MODE) | (1 << (int) SF_MODE))
4050 /* Modes for double-word and smaller quantities. */
4051 #define D_MODES (S_MODES | (1 << (int) D_MODE) | (1 << DF_MODE))
4053 /* Modes for quad-word and smaller quantities. */
4054 #define T_MODES (D_MODES | (1 << (int) T_MODE) | (1 << (int) TF_MODE))
4056 /* Modes for 8-word and smaller quantities. */
4057 #define O_MODES (T_MODES | (1 << (int) O_MODE) | (1 << (int) OF_MODE))
4059 /* Modes for single-float quantities. We must allow any single word or
4060 smaller quantity. This is because the fix/float conversion instructions
4061 take integer inputs/outputs from the float registers. */
4062 #define SF_MODES (S_MODES)
4064 /* Modes for double-float and smaller quantities. */
4065 #define DF_MODES (S_MODES | D_MODES)
4067 /* Modes for double-float only quantities. */
4068 #define DF_MODES_NO_S ((1 << (int) D_MODE) | (1 << (int) DF_MODE))
4070 /* Modes for quad-float only quantities. */
4071 #define TF_ONLY_MODES (1 << (int) TF_MODE)
4073 /* Modes for quad-float and smaller quantities. */
4074 #define TF_MODES (DF_MODES | TF_ONLY_MODES)
4076 /* Modes for quad-float and double-float quantities. */
4077 #define TF_MODES_NO_S (DF_MODES_NO_S | TF_ONLY_MODES)
4079 /* Modes for quad-float pair only quantities. */
4080 #define OF_ONLY_MODES (1 << (int) OF_MODE)
4082 /* Modes for quad-float pairs and smaller quantities. */
4083 #define OF_MODES (TF_MODES | OF_ONLY_MODES)
4085 #define OF_MODES_NO_S (TF_MODES_NO_S | OF_ONLY_MODES)
4087 /* Modes for condition codes. */
4088 #define CC_MODES (1 << (int) CC_MODE)
4089 #define CCFP_MODES (1 << (int) CCFP_MODE)
4091 /* Value is 1 if register/mode pair is acceptable on sparc.
4092 The funny mixture of D and T modes is because integer operations
4093 do not specially operate on tetra quantities, so non-quad-aligned
4094 registers can hold quadword quantities (except %o4 and %i4 because
4095 they cross fixed registers). */
4097 /* This points to either the 32 bit or the 64 bit version. */
4098 const int *hard_regno_mode_classes;
4100 static const int hard_32bit_mode_classes[] = {
4101 S_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
4102 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
4103 T_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
4104 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
4106 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4107 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4108 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4109 OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
4111 /* FP regs f32 to f63. Only the even numbered registers actually exist,
4112 and none can hold SFmode/SImode values. */
4113 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4114 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4115 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4116 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4119 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
4125 static const int hard_64bit_mode_classes[] = {
4126 D_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4127 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4128 T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4129 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4131 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4132 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4133 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4134 OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
4136 /* FP regs f32 to f63. Only the even numbered registers actually exist,
4137 and none can hold SFmode/SImode values. */
4138 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4139 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4140 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4141 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4144 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
4150 int sparc_mode_class [NUM_MACHINE_MODES];
4152 enum reg_class sparc_regno_reg_class[FIRST_PSEUDO_REGISTER];
4155 sparc_init_modes (void)
4159 for (i = 0; i < NUM_MACHINE_MODES; i++)
4161 switch (GET_MODE_CLASS (i))
4164 case MODE_PARTIAL_INT:
4165 case MODE_COMPLEX_INT:
4166 if (GET_MODE_SIZE (i) <= 4)
4167 sparc_mode_class[i] = 1 << (int) S_MODE;
4168 else if (GET_MODE_SIZE (i) == 8)
4169 sparc_mode_class[i] = 1 << (int) D_MODE;
4170 else if (GET_MODE_SIZE (i) == 16)
4171 sparc_mode_class[i] = 1 << (int) T_MODE;
4172 else if (GET_MODE_SIZE (i) == 32)
4173 sparc_mode_class[i] = 1 << (int) O_MODE;
4175 sparc_mode_class[i] = 0;
4178 case MODE_COMPLEX_FLOAT:
4179 if (GET_MODE_SIZE (i) <= 4)
4180 sparc_mode_class[i] = 1 << (int) SF_MODE;
4181 else if (GET_MODE_SIZE (i) == 8)
4182 sparc_mode_class[i] = 1 << (int) DF_MODE;
4183 else if (GET_MODE_SIZE (i) == 16)
4184 sparc_mode_class[i] = 1 << (int) TF_MODE;
4185 else if (GET_MODE_SIZE (i) == 32)
4186 sparc_mode_class[i] = 1 << (int) OF_MODE;
4188 sparc_mode_class[i] = 0;
4191 if (i == (int) CCFPmode || i == (int) CCFPEmode)
4192 sparc_mode_class[i] = 1 << (int) CCFP_MODE;
4194 sparc_mode_class[i] = 1 << (int) CC_MODE;
4197 sparc_mode_class[i] = 0;
4203 hard_regno_mode_classes = hard_64bit_mode_classes;
4205 hard_regno_mode_classes = hard_32bit_mode_classes;
4207 /* Initialize the array used by REGNO_REG_CLASS. */
4208 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
4210 if (i < 16 && TARGET_V8PLUS)
4211 sparc_regno_reg_class[i] = I64_REGS;
4212 else if (i < 32 || i == FRAME_POINTER_REGNUM)
4213 sparc_regno_reg_class[i] = GENERAL_REGS;
4215 sparc_regno_reg_class[i] = FP_REGS;
4217 sparc_regno_reg_class[i] = EXTRA_FP_REGS;
4219 sparc_regno_reg_class[i] = FPCC_REGS;
4221 sparc_regno_reg_class[i] = NO_REGS;
4225 /* Compute the frame size required by the function. This function is called
4226 during the reload pass and also by sparc_expand_prologue. */
4229 sparc_compute_frame_size (HOST_WIDE_INT size, int leaf_function_p)
4231 int outgoing_args_size = (current_function_outgoing_args_size
4232 + REG_PARM_STACK_SPACE (current_function_decl));
4233 int n_regs = 0; /* N_REGS is the number of 4-byte regs saved thus far. */
4238 for (i = 0; i < 8; i++)
4239 if (regs_ever_live[i] && ! call_used_regs[i])
4244 for (i = 0; i < 8; i += 2)
4245 if ((regs_ever_live[i] && ! call_used_regs[i])
4246 || (regs_ever_live[i+1] && ! call_used_regs[i+1]))
4250 for (i = 32; i < (TARGET_V9 ? 96 : 64); i += 2)
4251 if ((regs_ever_live[i] && ! call_used_regs[i])
4252 || (regs_ever_live[i+1] && ! call_used_regs[i+1]))
4255 /* Set up values for use in prologue and epilogue. */
4256 num_gfregs = n_regs;
4261 && current_function_outgoing_args_size == 0)
4262 actual_fsize = apparent_fsize = 0;
4265 /* We subtract STARTING_FRAME_OFFSET, remember it's negative. */
4266 apparent_fsize = (size - STARTING_FRAME_OFFSET + 7) & -8;
4267 apparent_fsize += n_regs * 4;
4268 actual_fsize = apparent_fsize + ((outgoing_args_size + 7) & -8);
4271 /* Make sure nothing can clobber our register windows.
4272 If a SAVE must be done, or there is a stack-local variable,
4273 the register window area must be allocated.
4274 ??? For v8 we apparently need an additional 8 bytes of reserved space. */
4275 if (! leaf_function_p || size > 0)
4276 actual_fsize += (16 * UNITS_PER_WORD) + (TARGET_ARCH64 ? 0 : 8);
4278 return SPARC_STACK_ALIGN (actual_fsize);
4281 /* Output any necessary .register pseudo-ops. */
4284 sparc_output_scratch_registers (FILE *file ATTRIBUTE_UNUSED)
4286 #ifdef HAVE_AS_REGISTER_PSEUDO_OP
4292 /* Check if %g[2367] were used without
4293 .register being printed for them already. */
4294 for (i = 2; i < 8; i++)
4296 if (regs_ever_live [i]
4297 && ! sparc_hard_reg_printed [i])
4299 sparc_hard_reg_printed [i] = 1;
4300 fprintf (file, "\t.register\t%%g%d, #scratch\n", i);
4307 /* Save/restore call-saved registers from LOW to HIGH at BASE+OFFSET
4308 as needed. LOW should be double-word aligned for 32-bit registers.
4309 Return the new OFFSET. */
4312 #define SORR_RESTORE 1
4315 save_or_restore_regs (int low, int high, rtx base, int offset, int action)
4320 if (TARGET_ARCH64 && high <= 32)
4322 for (i = low; i < high; i++)
4324 if (regs_ever_live[i] && ! call_used_regs[i])
4326 mem = gen_rtx_MEM (DImode, plus_constant (base, offset));
4327 set_mem_alias_set (mem, sparc_sr_alias_set);
4328 if (action == SORR_SAVE)
4330 insn = emit_move_insn (mem, gen_rtx_REG (DImode, i));
4331 RTX_FRAME_RELATED_P (insn) = 1;
4333 else /* action == SORR_RESTORE */
4334 emit_move_insn (gen_rtx_REG (DImode, i), mem);
4341 for (i = low; i < high; i += 2)
4343 bool reg0 = regs_ever_live[i] && ! call_used_regs[i];
4344 bool reg1 = regs_ever_live[i+1] && ! call_used_regs[i+1];
4345 enum machine_mode mode;
4350 mode = i < 32 ? DImode : DFmode;
4355 mode = i < 32 ? SImode : SFmode;
4360 mode = i < 32 ? SImode : SFmode;
4367 mem = gen_rtx_MEM (mode, plus_constant (base, offset));
4368 set_mem_alias_set (mem, sparc_sr_alias_set);
4369 if (action == SORR_SAVE)
4371 insn = emit_move_insn (mem, gen_rtx_REG (mode, regno));
4372 RTX_FRAME_RELATED_P (insn) = 1;
4374 else /* action == SORR_RESTORE */
4375 emit_move_insn (gen_rtx_REG (mode, regno), mem);
4377 /* Always preserve double-word alignment. */
4378 offset = (offset + 7) & -8;
4385 /* Emit code to save call-saved registers. */
4388 emit_save_regs (void)
4390 HOST_WIDE_INT offset;
4393 offset = frame_base_offset - apparent_fsize;
4395 if (offset < -4096 || offset + num_gfregs * 4 > 4096)
4397 /* ??? This might be optimized a little as %g1 might already have a
4398 value close enough that a single add insn will do. */
4399 /* ??? Although, all of this is probably only a temporary fix
4400 because if %g1 can hold a function result, then
4401 sparc_expand_epilogue will lose (the result will be
4403 base = gen_rtx_REG (Pmode, 1);
4404 emit_move_insn (base, GEN_INT (offset));
4405 emit_insn (gen_rtx_SET (VOIDmode,
4407 gen_rtx_PLUS (Pmode, frame_base_reg, base)));
4411 base = frame_base_reg;
4413 offset = save_or_restore_regs (0, 8, base, offset, SORR_SAVE);
4414 save_or_restore_regs (32, TARGET_V9 ? 96 : 64, base, offset, SORR_SAVE);
4417 /* Emit code to restore call-saved registers. */
4420 emit_restore_regs (void)
4422 HOST_WIDE_INT offset;
4425 offset = frame_base_offset - apparent_fsize;
4427 if (offset < -4096 || offset + num_gfregs * 4 > 4096 - 8 /*double*/)
4429 base = gen_rtx_REG (Pmode, 1);
4430 emit_move_insn (base, GEN_INT (offset));
4431 emit_insn (gen_rtx_SET (VOIDmode,
4433 gen_rtx_PLUS (Pmode, frame_base_reg, base)));
4437 base = frame_base_reg;
4439 offset = save_or_restore_regs (0, 8, base, offset, SORR_RESTORE);
4440 save_or_restore_regs (32, TARGET_V9 ? 96 : 64, base, offset, SORR_RESTORE);
4443 /* Emit an increment for the stack pointer. */
4446 emit_stack_pointer_increment (rtx increment)
4449 emit_insn (gen_adddi3 (stack_pointer_rtx, stack_pointer_rtx, increment));
4451 emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, increment));
4454 /* Emit a decrement for the stack pointer. */
4457 emit_stack_pointer_decrement (rtx decrement)
4460 emit_insn (gen_subdi3 (stack_pointer_rtx, stack_pointer_rtx, decrement));
4462 emit_insn (gen_subsi3 (stack_pointer_rtx, stack_pointer_rtx, decrement));
4465 /* Expand the function prologue. The prologue is responsible for reserving
4466 storage for the frame, saving the call-saved registers and loading the
4467 PIC register if needed. */
4470 sparc_expand_prologue (void)
4472 int leaf_function_p = current_function_uses_only_leaf_regs;
4474 /* Need to use actual_fsize, since we are also allocating
4475 space for our callee (and our own register save area). */
4476 actual_fsize = sparc_compute_frame_size (get_frame_size(), leaf_function_p);
4478 if (leaf_function_p)
4480 frame_base_reg = stack_pointer_rtx;
4481 frame_base_offset = actual_fsize + SPARC_STACK_BIAS;
4485 frame_base_reg = hard_frame_pointer_rtx;
4486 frame_base_offset = SPARC_STACK_BIAS;
4489 if (actual_fsize == 0)
4491 else if (leaf_function_p)
4493 if (actual_fsize <= 4096)
4494 emit_stack_pointer_increment (GEN_INT (- actual_fsize));
4495 else if (actual_fsize <= 8192)
4497 emit_stack_pointer_increment (GEN_INT (-4096));
4498 emit_stack_pointer_increment (GEN_INT (4096 - actual_fsize));
4502 rtx reg = gen_rtx_REG (Pmode, 1);
4503 emit_move_insn (reg, GEN_INT (-actual_fsize));
4504 emit_stack_pointer_increment (reg);
4509 if (actual_fsize <= 4096)
4510 emit_insn (gen_save_register_window (GEN_INT (-actual_fsize)));
4511 else if (actual_fsize <= 8192)
4513 emit_insn (gen_save_register_window (GEN_INT (-4096)));
4514 emit_stack_pointer_increment (GEN_INT (4096 - actual_fsize));
4518 rtx reg = gen_rtx_REG (Pmode, 1);
4519 emit_move_insn (reg, GEN_INT (-actual_fsize));
4520 emit_insn (gen_save_register_window (reg));
4524 /* Call-saved registers are saved just above the outgoing argument area. */
4528 /* Load the PIC register if needed. */
4529 if (flag_pic && current_function_uses_pic_offset_table)
4530 load_pic_register ();
4533 /* This function generates the assembly code for function entry, which boils
4534 down to emitting the necessary .register directives. It also informs the
4535 DWARF-2 back-end on the layout of the frame.
4537 ??? Historical cruft: "On SPARC, move-double insns between fpu and cpu need
4538 an 8-byte block of memory. If any fpu reg is used in the function, we
4539 allocate such a block here, at the bottom of the frame, just in case it's
4540 needed." Could this explain the -8 in emit_restore_regs? */
4543 sparc_asm_function_prologue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
4545 int leaf_function_p = current_function_uses_only_leaf_regs;
4547 sparc_output_scratch_registers (file);
4549 if (dwarf2out_do_frame () && actual_fsize)
4551 char *label = dwarf2out_cfi_label ();
4553 /* The canonical frame address refers to the top of the frame. */
4554 dwarf2out_def_cfa (label,
4556 ? STACK_POINTER_REGNUM
4557 : HARD_FRAME_POINTER_REGNUM,
4560 if (! leaf_function_p)
4562 /* Note the register window save. This tells the unwinder that
4563 it needs to restore the window registers from the previous
4564 frame's window save area at 0(cfa). */
4565 dwarf2out_window_save (label);
4567 /* The return address (-8) is now in %i7. */
4568 dwarf2out_return_reg (label, 31);
4573 /* Expand the function epilogue, either normal or part of a sibcall.
4574 We emit all the instructions except the return or the call. */
4577 sparc_expand_epilogue (void)
4579 int leaf_function_p = current_function_uses_only_leaf_regs;
4582 emit_restore_regs ();
4584 if (actual_fsize == 0)
4586 else if (leaf_function_p)
4588 if (actual_fsize <= 4096)
4589 emit_stack_pointer_decrement (GEN_INT (- actual_fsize));
4590 else if (actual_fsize <= 8192)
4592 emit_stack_pointer_decrement (GEN_INT (-4096));
4593 emit_stack_pointer_decrement (GEN_INT (4096 - actual_fsize));
4597 rtx reg = gen_rtx_REG (Pmode, 1);
4598 emit_move_insn (reg, GEN_INT (-actual_fsize));
4599 emit_stack_pointer_decrement (reg);
4604 /* This function generates the assembly code for function exit. */
4607 sparc_asm_function_epilogue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
4609 /* If code does not drop into the epilogue, we have to still output
4610 a dummy nop for the sake of sane backtraces. Otherwise, if the
4611 last two instructions of a function were "call foo; dslot;" this
4612 can make the return PC of foo (ie. address of call instruction
4613 plus 8) point to the first instruction in the next function. */
4615 rtx insn, last_real_insn;
4617 insn = get_last_insn ();
4619 last_real_insn = prev_real_insn (insn);
4621 && GET_CODE (last_real_insn) == INSN
4622 && GET_CODE (PATTERN (last_real_insn)) == SEQUENCE)
4623 last_real_insn = XVECEXP (PATTERN (last_real_insn), 0, 0);
4625 if (last_real_insn && GET_CODE (last_real_insn) == CALL_INSN)
4626 fputs("\tnop\n", file);
4628 sparc_output_deferred_case_vectors ();
4631 /* Output a 'restore' instruction. */
4634 output_restore (rtx pat)
4640 fputs ("\t restore\n", asm_out_file);
4644 if (GET_CODE (pat) != SET)
4647 operands[0] = SET_DEST (pat);
4648 pat = SET_SRC (pat);
4650 switch (GET_CODE (pat))
4653 operands[1] = XEXP (pat, 0);
4654 operands[2] = XEXP (pat, 1);
4655 output_asm_insn (" restore %r1, %2, %Y0", operands);
4658 operands[1] = XEXP (pat, 0);
4659 operands[2] = XEXP (pat, 1);
4660 output_asm_insn (" restore %r1, %%lo(%a2), %Y0", operands);
4663 operands[1] = XEXP (pat, 0);
4664 if (XEXP (pat, 1) != const1_rtx)
4666 output_asm_insn (" restore %r1, %r1, %Y0", operands);
4670 output_asm_insn (" restore %%g0, %1, %Y0", operands);
4675 /* Output a return. */
4678 output_return (rtx insn)
4680 if (current_function_uses_only_leaf_regs)
4682 /* This is a leaf function so we don't have to bother restoring the
4683 register window, which frees us from dealing with the convoluted
4684 semantics of restore/return. We simply output the jump to the
4685 return address and the insn in the delay slot (if any). */
4687 if (current_function_calls_eh_return)
4690 return "jmp\t%%o7+%)%#";
4694 /* This is a regular function so we have to restore the register window.
4695 We may have a pending insn for the delay slot, which will be either
4696 combined with the 'restore' instruction or put in the delay slot of
4697 the 'return' instruction. */
4699 if (current_function_calls_eh_return)
4701 /* If the function uses __builtin_eh_return, the eh_return
4702 machinery occupies the delay slot. */
4706 if (! flag_delayed_branch)
4707 fputs ("\tadd\t%fp, %g1, %fp\n", asm_out_file);
4710 fputs ("\treturn\t%i7+8\n", asm_out_file);
4712 fputs ("\trestore\n\tjmp\t%o7+8\n", asm_out_file);
4714 if (flag_delayed_branch)
4715 fputs ("\t add\t%sp, %g1, %sp\n", asm_out_file);
4717 fputs ("\t nop\n", asm_out_file);
4719 else if (final_sequence)
4723 delay = NEXT_INSN (insn);
4727 pat = PATTERN (delay);
4729 if (TARGET_V9 && ! epilogue_renumber (&pat, 1))
4731 epilogue_renumber (&pat, 0);
4732 return "return\t%%i7+%)%#";
4736 output_asm_insn ("jmp\t%%i7+%)", NULL);
4737 output_restore (pat);
4738 PATTERN (delay) = gen_blockage ();
4739 INSN_CODE (delay) = -1;
4744 /* The delay slot is empty. */
4746 return "return\t%%i7+%)\n\t nop";
4747 else if (flag_delayed_branch)
4748 return "jmp\t%%i7+%)\n\t restore";
4750 return "restore\n\tjmp\t%%o7+%)\n\t nop";
4757 /* Output a sibling call. */
4760 output_sibcall (rtx insn, rtx call_operand)
4764 if (! flag_delayed_branch)
4767 operands[0] = call_operand;
4769 if (current_function_uses_only_leaf_regs)
4771 /* This is a leaf function so we don't have to bother restoring the
4772 register window. We simply output the jump to the function and
4773 the insn in the delay slot (if any). */
4775 if (LEAF_SIBCALL_SLOT_RESERVED_P && final_sequence)
4779 output_asm_insn ("sethi\t%%hi(%a0), %%g1\n\tjmp\t%%g1 + %%lo(%a0)%#",
4782 /* Use or with rs2 %%g0 instead of mov, so that as/ld can optimize
4783 it into branch if possible. */
4784 output_asm_insn ("or\t%%o7, %%g0, %%g1\n\tcall\t%a0, 0\n\t or\t%%g1, %%g0, %%o7",
4789 /* This is a regular function so we have to restore the register window.
4790 We may have a pending insn for the delay slot, which will be combined
4791 with the 'restore' instruction. */
4793 output_asm_insn ("call\t%a0, 0", operands);
4797 rtx delay = NEXT_INSN (insn);
4801 output_restore (PATTERN (delay));
4803 PATTERN (delay) = gen_blockage ();
4804 INSN_CODE (delay) = -1;
4807 output_restore (NULL_RTX);
4813 /* Functions for handling argument passing.
4815 For 32-bit, the first 6 args are normally in registers and the rest are
4816 pushed. Any arg that starts within the first 6 words is at least
4817 partially passed in a register unless its data type forbids.
4819 For 64-bit, the argument registers are laid out as an array of 16 elements
4820 and arguments are added sequentially. The first 6 int args and up to the
4821 first 16 fp args (depending on size) are passed in regs.
4823 Slot Stack Integral Float Float in structure Double Long Double
4824 ---- ----- -------- ----- ------------------ ------ -----------
4825 15 [SP+248] %f31 %f30,%f31 %d30
4826 14 [SP+240] %f29 %f28,%f29 %d28 %q28
4827 13 [SP+232] %f27 %f26,%f27 %d26
4828 12 [SP+224] %f25 %f24,%f25 %d24 %q24
4829 11 [SP+216] %f23 %f22,%f23 %d22
4830 10 [SP+208] %f21 %f20,%f21 %d20 %q20
4831 9 [SP+200] %f19 %f18,%f19 %d18
4832 8 [SP+192] %f17 %f16,%f17 %d16 %q16
4833 7 [SP+184] %f15 %f14,%f15 %d14
4834 6 [SP+176] %f13 %f12,%f13 %d12 %q12
4835 5 [SP+168] %o5 %f11 %f10,%f11 %d10
4836 4 [SP+160] %o4 %f9 %f8,%f9 %d8 %q8
4837 3 [SP+152] %o3 %f7 %f6,%f7 %d6
4838 2 [SP+144] %o2 %f5 %f4,%f5 %d4 %q4
4839 1 [SP+136] %o1 %f3 %f2,%f3 %d2
4840 0 [SP+128] %o0 %f1 %f0,%f1 %d0 %q0
4842 Here SP = %sp if -mno-stack-bias or %sp+stack_bias otherwise.
4844 Integral arguments are always passed as 64-bit quantities appropriately
4847 Passing of floating point values is handled as follows.
4848 If a prototype is in scope:
4849 If the value is in a named argument (i.e. not a stdarg function or a
4850 value not part of the `...') then the value is passed in the appropriate
4852 If the value is part of the `...' and is passed in one of the first 6
4853 slots then the value is passed in the appropriate int reg.
4854 If the value is part of the `...' and is not passed in one of the first 6
4855 slots then the value is passed in memory.
4856 If a prototype is not in scope:
4857 If the value is one of the first 6 arguments the value is passed in the
4858 appropriate integer reg and the appropriate fp reg.
4859 If the value is not one of the first 6 arguments the value is passed in
4860 the appropriate fp reg and in memory.
4863 Summary of the calling conventions implemented by GCC on SPARC:
4866 size argument return value
4868 small integer <4 int. reg. int. reg.
4869 word 4 int. reg. int. reg.
4870 double word 8 int. reg. int. reg.
4872 _Complex small integer <8 int. reg. int. reg.
4873 _Complex word 8 int. reg. int. reg.
4874 _Complex double word 16 memory int. reg.
4876 vector integer <=8 int. reg. FP reg.
4877 vector integer >8 memory memory
4879 float 4 int. reg. FP reg.
4880 double 8 int. reg. FP reg.
4881 long double 16 memory memory
4883 _Complex float 8 memory FP reg.
4884 _Complex double 16 memory FP reg.
4885 _Complex long double 32 memory FP reg.
4887 vector float <=32 memory FP reg.
4888 vector float >32 memory memory
4890 aggregate any memory memory
4895 size argument return value
4897 small integer <8 int. reg. int. reg.
4898 word 8 int. reg. int. reg.
4899 double word 16 int. reg. int. reg.
4901 _Complex small integer <16 int. reg. int. reg.
4902 _Complex word 16 int. reg. int. reg.
4903 _Complex double word 32 memory int. reg.
4905 vector integer <=16 FP reg. FP reg.
4906 vector integer 16<s<=32 memory FP reg.
4907 vector integer >32 memory memory
4909 float 4 FP reg. FP reg.
4910 double 8 FP reg. FP reg.
4911 long double 16 FP reg. FP reg.
4913 _Complex float 8 FP reg. FP reg.
4914 _Complex double 16 FP reg. FP reg.
4915 _Complex long double 32 memory FP reg.
4917 vector float <=16 FP reg. FP reg.
4918 vector float 16<s<=32 memory FP reg.
4919 vector float >32 memory memory
4921 aggregate <=16 reg. reg.
4922 aggregate 16<s<=32 memory reg.
4923 aggregate >32 memory memory
4927 Note #1: complex floating-point types follow the extended SPARC ABIs as
4928 implemented by the Sun compiler.
4930 Note #2: integral vector types follow the scalar floating-point types
4931 conventions to match what is implemented by the Sun VIS SDK.
4933 Note #3: floating-point vector types follow the complex floating-point
4934 types conventions. */
4937 /* Maximum number of int regs for args. */
4938 #define SPARC_INT_ARG_MAX 6
4939 /* Maximum number of fp regs for args. */
4940 #define SPARC_FP_ARG_MAX 16
4942 #define ROUND_ADVANCE(SIZE) (((SIZE) + UNITS_PER_WORD - 1) / UNITS_PER_WORD)
4944 /* Handle the INIT_CUMULATIVE_ARGS macro.
4945 Initialize a variable CUM of type CUMULATIVE_ARGS
4946 for a call to a function whose data type is FNTYPE.
4947 For a library call, FNTYPE is 0. */
4950 init_cumulative_args (struct sparc_args *cum, tree fntype,
4951 rtx libname ATTRIBUTE_UNUSED,
4952 tree fndecl ATTRIBUTE_UNUSED)
4955 cum->prototype_p = fntype && TYPE_ARG_TYPES (fntype);
4956 cum->libcall_p = fntype == 0;
4959 /* Handle the TARGET_PROMOTE_PROTOTYPES target hook.
4960 When a prototype says `char' or `short', really pass an `int'. */
4963 sparc_promote_prototypes (tree fntype ATTRIBUTE_UNUSED)
4965 return TARGET_ARCH32 ? true : false;
4968 /* Handle the TARGET_STRICT_ARGUMENT_NAMING target hook. */
4971 sparc_strict_argument_naming (CUMULATIVE_ARGS *ca ATTRIBUTE_UNUSED)
4973 return TARGET_ARCH64 ? true : false;
4976 /* Scan the record type TYPE and return the following predicates:
4977 - INTREGS_P: the record contains at least one field or sub-field
4978 that is eligible for promotion in integer registers.
4979 - FP_REGS_P: the record contains at least one field or sub-field
4980 that is eligible for promotion in floating-point registers.
4981 - PACKED_P: the record contains at least one field that is packed.
4983 Sub-fields are not taken into account for the PACKED_P predicate. */
4986 scan_record_type (tree type, int *intregs_p, int *fpregs_p, int *packed_p)
4990 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
4992 if (TREE_CODE (field) == FIELD_DECL)
4994 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
4995 scan_record_type (TREE_TYPE (field), intregs_p, fpregs_p, 0);
4996 else if (FLOAT_TYPE_P (TREE_TYPE (field)) && TARGET_FPU)
5001 if (packed_p && DECL_PACKED (field))
5007 /* Compute the slot number to pass an argument in.
5008 Return the slot number or -1 if passing on the stack.
5010 CUM is a variable of type CUMULATIVE_ARGS which gives info about
5011 the preceding args and about the function being called.
5012 MODE is the argument's machine mode.
5013 TYPE is the data type of the argument (as a tree).
5014 This is null for libcalls where that information may
5016 NAMED is nonzero if this argument is a named parameter
5017 (otherwise it is an extra parameter matching an ellipsis).
5018 INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG.
5019 *PREGNO records the register number to use if scalar type.
5020 *PPADDING records the amount of padding needed in words. */
5023 function_arg_slotno (const struct sparc_args *cum, enum machine_mode mode,
5024 tree type, int named, int incoming_p,
5025 int *pregno, int *ppadding)
5027 int regbase = (incoming_p
5028 ? SPARC_INCOMING_INT_ARG_FIRST
5029 : SPARC_OUTGOING_INT_ARG_FIRST);
5030 int slotno = cum->words;
5035 if (type && TREE_ADDRESSABLE (type))
5041 && TYPE_ALIGN (type) % PARM_BOUNDARY != 0)
5044 /* For SPARC64, objects requiring 16-byte alignment get it. */
5046 && GET_MODE_ALIGNMENT (mode) >= 2 * BITS_PER_WORD
5047 && (slotno & 1) != 0)
5048 slotno++, *ppadding = 1;
5050 switch (GET_MODE_CLASS (mode))
5053 case MODE_COMPLEX_FLOAT:
5054 case MODE_VECTOR_INT:
5055 case MODE_VECTOR_FLOAT:
5056 if (TARGET_ARCH64 && TARGET_FPU && named)
5058 if (slotno >= SPARC_FP_ARG_MAX)
5060 regno = SPARC_FP_ARG_FIRST + slotno * 2;
5061 /* Arguments filling only one single FP register are
5062 right-justified in the outer double FP register. */
5063 if (GET_MODE_SIZE (mode) <= 4)
5070 case MODE_COMPLEX_INT:
5071 if (slotno >= SPARC_INT_ARG_MAX)
5073 regno = regbase + slotno;
5077 if (mode == VOIDmode)
5078 /* MODE is VOIDmode when generating the actual call. */
5081 if (mode != BLKmode)
5084 /* For SPARC64, objects requiring 16-byte alignment get it. */
5087 && TYPE_ALIGN (type) >= 2 * BITS_PER_WORD
5088 && (slotno & 1) != 0)
5089 slotno++, *ppadding = 1;
5091 if (TARGET_ARCH32 || (type && TREE_CODE (type) == UNION_TYPE))
5093 if (slotno >= SPARC_INT_ARG_MAX)
5095 regno = regbase + slotno;
5097 else /* TARGET_ARCH64 && type && TREE_CODE (type) == RECORD_TYPE */
5099 int intregs_p = 0, fpregs_p = 0, packed_p = 0;
5101 /* First see what kinds of registers we would need. */
5102 scan_record_type (type, &intregs_p, &fpregs_p, &packed_p);
5104 /* The ABI obviously doesn't specify how packed structures
5105 are passed. These are defined to be passed in int regs
5106 if possible, otherwise memory. */
5107 if (packed_p || !named)
5108 fpregs_p = 0, intregs_p = 1;
5110 /* If all arg slots are filled, then must pass on stack. */
5111 if (fpregs_p && slotno >= SPARC_FP_ARG_MAX)
5114 /* If there are only int args and all int arg slots are filled,
5115 then must pass on stack. */
5116 if (!fpregs_p && intregs_p && slotno >= SPARC_INT_ARG_MAX)
5119 /* Note that even if all int arg slots are filled, fp members may
5120 still be passed in regs if such regs are available.
5121 *PREGNO isn't set because there may be more than one, it's up
5122 to the caller to compute them. */
5135 /* Handle recursive register counting for structure field layout. */
5137 struct function_arg_record_value_parms
5139 rtx ret; /* return expression being built. */
5140 int slotno; /* slot number of the argument. */
5141 int named; /* whether the argument is named. */
5142 int regbase; /* regno of the base register. */
5143 int stack; /* 1 if part of the argument is on the stack. */
5144 int intoffset; /* offset of the first pending integer field. */
5145 unsigned int nregs; /* number of words passed in registers. */
5148 static void function_arg_record_value_3
5149 (HOST_WIDE_INT, struct function_arg_record_value_parms *);
5150 static void function_arg_record_value_2
5151 (tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
5152 static void function_arg_record_value_1
5153 (tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
5154 static rtx function_arg_record_value (tree, enum machine_mode, int, int, int);
5155 static rtx function_arg_union_value (int, enum machine_mode, int);
5157 /* A subroutine of function_arg_record_value. Traverse the structure
5158 recursively and determine how many registers will be required. */
5161 function_arg_record_value_1 (tree type, HOST_WIDE_INT startbitpos,
5162 struct function_arg_record_value_parms *parms,
5167 /* We need to compute how many registers are needed so we can
5168 allocate the PARALLEL but before we can do that we need to know
5169 whether there are any packed fields. The ABI obviously doesn't
5170 specify how structures are passed in this case, so they are
5171 defined to be passed in int regs if possible, otherwise memory,
5172 regardless of whether there are fp values present. */
5175 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5177 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
5184 /* Compute how many registers we need. */
5185 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5187 if (TREE_CODE (field) == FIELD_DECL)
5189 HOST_WIDE_INT bitpos = startbitpos;
5191 if (DECL_SIZE (field) != 0
5192 && host_integerp (bit_position (field), 1))
5193 bitpos += int_bit_position (field);
5195 /* ??? FIXME: else assume zero offset. */
5197 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
5198 function_arg_record_value_1 (TREE_TYPE (field),
5202 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
5203 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
5208 if (parms->intoffset != -1)
5210 unsigned int startbit, endbit;
5211 int intslots, this_slotno;
5213 startbit = parms->intoffset & -BITS_PER_WORD;
5214 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
5216 intslots = (endbit - startbit) / BITS_PER_WORD;
5217 this_slotno = parms->slotno + parms->intoffset
5220 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
5222 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
5223 /* We need to pass this field on the stack. */
5227 parms->nregs += intslots;
5228 parms->intoffset = -1;
5231 /* There's no need to check this_slotno < SPARC_FP_ARG MAX.
5232 If it wasn't true we wouldn't be here. */
5234 if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
5239 if (parms->intoffset == -1)
5240 parms->intoffset = bitpos;
5246 /* A subroutine of function_arg_record_value. Assign the bits of the
5247 structure between parms->intoffset and bitpos to integer registers. */
5250 function_arg_record_value_3 (HOST_WIDE_INT bitpos,
5251 struct function_arg_record_value_parms *parms)
5253 enum machine_mode mode;
5255 unsigned int startbit, endbit;
5256 int this_slotno, intslots, intoffset;
5259 if (parms->intoffset == -1)
5262 intoffset = parms->intoffset;
5263 parms->intoffset = -1;
5265 startbit = intoffset & -BITS_PER_WORD;
5266 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
5267 intslots = (endbit - startbit) / BITS_PER_WORD;
5268 this_slotno = parms->slotno + intoffset / BITS_PER_WORD;
5270 intslots = MIN (intslots, SPARC_INT_ARG_MAX - this_slotno);
5274 /* If this is the trailing part of a word, only load that much into
5275 the register. Otherwise load the whole register. Note that in
5276 the latter case we may pick up unwanted bits. It's not a problem
5277 at the moment but may wish to revisit. */
5279 if (intoffset % BITS_PER_WORD != 0)
5280 mode = mode_for_size (BITS_PER_WORD - intoffset % BITS_PER_WORD,
5285 intoffset /= BITS_PER_UNIT;
5288 regno = parms->regbase + this_slotno;
5289 reg = gen_rtx_REG (mode, regno);
5290 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
5291 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (intoffset));
5294 intoffset = (intoffset | (UNITS_PER_WORD-1)) + 1;
5299 while (intslots > 0);
5302 /* A subroutine of function_arg_record_value. Traverse the structure
5303 recursively and assign bits to floating point registers. Track which
5304 bits in between need integer registers; invoke function_arg_record_value_3
5305 to make that happen. */
5308 function_arg_record_value_2 (tree type, HOST_WIDE_INT startbitpos,
5309 struct function_arg_record_value_parms *parms,
5315 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5317 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
5324 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5326 if (TREE_CODE (field) == FIELD_DECL)
5328 HOST_WIDE_INT bitpos = startbitpos;
5330 if (DECL_SIZE (field) != 0
5331 && host_integerp (bit_position (field), 1))
5332 bitpos += int_bit_position (field);
5334 /* ??? FIXME: else assume zero offset. */
5336 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
5337 function_arg_record_value_2 (TREE_TYPE (field),
5341 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
5342 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
5347 int this_slotno = parms->slotno + bitpos / BITS_PER_WORD;
5349 enum machine_mode mode = DECL_MODE (field);
5352 function_arg_record_value_3 (bitpos, parms);
5355 case SCmode: mode = SFmode; break;
5356 case DCmode: mode = DFmode; break;
5357 case TCmode: mode = TFmode; break;
5360 regno = SPARC_FP_ARG_FIRST + this_slotno * 2;
5361 if (GET_MODE_SIZE (mode) <= 4 && (bitpos & 32) != 0)
5363 reg = gen_rtx_REG (mode, regno);
5364 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
5365 = gen_rtx_EXPR_LIST (VOIDmode, reg,
5366 GEN_INT (bitpos / BITS_PER_UNIT));
5368 if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
5370 regno += GET_MODE_SIZE (mode) / 4;
5371 reg = gen_rtx_REG (mode, regno);
5372 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
5373 = gen_rtx_EXPR_LIST (VOIDmode, reg,
5374 GEN_INT ((bitpos + GET_MODE_BITSIZE (mode))
5381 if (parms->intoffset == -1)
5382 parms->intoffset = bitpos;
5388 /* Used by function_arg and function_value to implement the complex
5389 conventions of the 64-bit ABI for passing and returning structures.
5390 Return an expression valid as a return value for the two macros
5391 FUNCTION_ARG and FUNCTION_VALUE.
5393 TYPE is the data type of the argument (as a tree).
5394 This is null for libcalls where that information may
5396 MODE is the argument's machine mode.
5397 SLOTNO is the index number of the argument's slot in the parameter array.
5398 NAMED is nonzero if this argument is a named parameter
5399 (otherwise it is an extra parameter matching an ellipsis).
5400 REGBASE is the regno of the base register for the parameter array. */
5403 function_arg_record_value (tree type, enum machine_mode mode,
5404 int slotno, int named, int regbase)
5406 HOST_WIDE_INT typesize = int_size_in_bytes (type);
5407 struct function_arg_record_value_parms parms;
5410 parms.ret = NULL_RTX;
5411 parms.slotno = slotno;
5412 parms.named = named;
5413 parms.regbase = regbase;
5416 /* Compute how many registers we need. */
5418 parms.intoffset = 0;
5419 function_arg_record_value_1 (type, 0, &parms, false);
5421 /* Take into account pending integer fields. */
5422 if (parms.intoffset != -1)
5424 unsigned int startbit, endbit;
5425 int intslots, this_slotno;
5427 startbit = parms.intoffset & -BITS_PER_WORD;
5428 endbit = (typesize*BITS_PER_UNIT + BITS_PER_WORD - 1) & -BITS_PER_WORD;
5429 intslots = (endbit - startbit) / BITS_PER_WORD;
5430 this_slotno = slotno + parms.intoffset / BITS_PER_WORD;
5432 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
5434 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
5435 /* We need to pass this field on the stack. */
5439 parms.nregs += intslots;
5441 nregs = parms.nregs;
5443 /* Allocate the vector and handle some annoying special cases. */
5446 /* ??? Empty structure has no value? Duh? */
5449 /* Though there's nothing really to store, return a word register
5450 anyway so the rest of gcc doesn't go nuts. Returning a PARALLEL
5451 leads to breakage due to the fact that there are zero bytes to
5453 return gen_rtx_REG (mode, regbase);
5457 /* ??? C++ has structures with no fields, and yet a size. Give up
5458 for now and pass everything back in integer registers. */
5459 nregs = (typesize + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
5461 if (nregs + slotno > SPARC_INT_ARG_MAX)
5462 nregs = SPARC_INT_ARG_MAX - slotno;
5467 parms.ret = gen_rtx_PARALLEL (mode, rtvec_alloc (parms.stack + nregs));
5469 /* If at least one field must be passed on the stack, generate
5470 (parallel [(expr_list (nil) ...) ...]) so that all fields will
5471 also be passed on the stack. We can't do much better because the
5472 semantics of FUNCTION_ARG_PARTIAL_NREGS doesn't handle the case
5473 of structures for which the fields passed exclusively in registers
5474 are not at the beginning of the structure. */
5476 XVECEXP (parms.ret, 0, 0)
5477 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
5479 /* Fill in the entries. */
5481 parms.intoffset = 0;
5482 function_arg_record_value_2 (type, 0, &parms, false);
5483 function_arg_record_value_3 (typesize * BITS_PER_UNIT, &parms);
5485 if (parms.nregs != nregs)
5491 /* Used by function_arg and function_value to implement the conventions
5492 of the 64-bit ABI for passing and returning unions.
5493 Return an expression valid as a return value for the two macros
5494 FUNCTION_ARG and FUNCTION_VALUE.
5496 SIZE is the size in bytes of the union.
5497 MODE is the argument's machine mode.
5498 REGNO is the hard register the union will be passed in. */
5501 function_arg_union_value (int size, enum machine_mode mode, int regno)
5503 int nwords = ROUND_ADVANCE (size), i;
5506 /* Unions are passed left-justified. */
5507 regs = gen_rtx_PARALLEL (mode, rtvec_alloc (nwords));
5509 for (i = 0; i < nwords; i++)
5510 XVECEXP (regs, 0, i)
5511 = gen_rtx_EXPR_LIST (VOIDmode,
5512 gen_rtx_REG (word_mode, regno + i),
5513 GEN_INT (UNITS_PER_WORD * i));
5518 /* Handle the FUNCTION_ARG macro.
5519 Determine where to put an argument to a function.
5520 Value is zero to push the argument on the stack,
5521 or a hard register in which to store the argument.
5523 CUM is a variable of type CUMULATIVE_ARGS which gives info about
5524 the preceding args and about the function being called.
5525 MODE is the argument's machine mode.
5526 TYPE is the data type of the argument (as a tree).
5527 This is null for libcalls where that information may
5529 NAMED is nonzero if this argument is a named parameter
5530 (otherwise it is an extra parameter matching an ellipsis).
5531 INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG. */
5534 function_arg (const struct sparc_args *cum, enum machine_mode mode,
5535 tree type, int named, int incoming_p)
5537 int regbase = (incoming_p
5538 ? SPARC_INCOMING_INT_ARG_FIRST
5539 : SPARC_OUTGOING_INT_ARG_FIRST);
5540 int slotno, regno, padding;
5543 slotno = function_arg_slotno (cum, mode, type, named, incoming_p,
5551 reg = gen_rtx_REG (mode, regno);
5555 if (type && TREE_CODE (type) == RECORD_TYPE)
5557 /* Structures up to 16 bytes in size are passed in arg slots on the
5558 stack and are promoted to registers where possible. */
5560 if (int_size_in_bytes (type) > 16)
5561 abort (); /* shouldn't get here */
5563 return function_arg_record_value (type, mode, slotno, named, regbase);
5565 else if (type && TREE_CODE (type) == UNION_TYPE)
5567 HOST_WIDE_INT size = int_size_in_bytes (type);
5570 abort (); /* shouldn't get here */
5572 return function_arg_union_value (size, mode, regno);
5574 /* v9 fp args in reg slots beyond the int reg slots get passed in regs
5575 but also have the slot allocated for them.
5576 If no prototype is in scope fp values in register slots get passed
5577 in two places, either fp regs and int regs or fp regs and memory. */
5578 else if ((GET_MODE_CLASS (mode) == MODE_FLOAT
5579 || GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT
5580 || GET_MODE_CLASS (mode) == MODE_VECTOR_INT
5581 || GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT)
5582 && SPARC_FP_REG_P (regno))
5584 reg = gen_rtx_REG (mode, regno);
5585 if (cum->prototype_p || cum->libcall_p)
5587 /* "* 2" because fp reg numbers are recorded in 4 byte
5590 /* ??? This will cause the value to be passed in the fp reg and
5591 in the stack. When a prototype exists we want to pass the
5592 value in the reg but reserve space on the stack. That's an
5593 optimization, and is deferred [for a bit]. */
5594 if ((regno - SPARC_FP_ARG_FIRST) >= SPARC_INT_ARG_MAX * 2)
5595 return gen_rtx_PARALLEL (mode,
5597 gen_rtx_EXPR_LIST (VOIDmode,
5598 NULL_RTX, const0_rtx),
5599 gen_rtx_EXPR_LIST (VOIDmode,
5603 /* ??? It seems that passing back a register even when past
5604 the area declared by REG_PARM_STACK_SPACE will allocate
5605 space appropriately, and will not copy the data onto the
5606 stack, exactly as we desire.
5608 This is due to locate_and_pad_parm being called in
5609 expand_call whenever reg_parm_stack_space > 0, which
5610 while beneficial to our example here, would seem to be
5611 in error from what had been intended. Ho hum... -- r~ */
5619 if ((regno - SPARC_FP_ARG_FIRST) < SPARC_INT_ARG_MAX * 2)
5623 /* On incoming, we don't need to know that the value
5624 is passed in %f0 and %i0, and it confuses other parts
5625 causing needless spillage even on the simplest cases. */
5629 intreg = (SPARC_OUTGOING_INT_ARG_FIRST
5630 + (regno - SPARC_FP_ARG_FIRST) / 2);
5632 v0 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
5633 v1 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_REG (mode, intreg),
5635 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
5639 v0 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
5640 v1 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
5641 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
5647 /* Scalar or complex int. */
5648 reg = gen_rtx_REG (mode, regno);
5654 /* Handle the FUNCTION_ARG_PARTIAL_NREGS macro.
5655 For an arg passed partly in registers and partly in memory,
5656 this is the number of registers used.
5657 For args passed entirely in registers or entirely in memory, zero.
5659 Any arg that starts in the first 6 regs but won't entirely fit in them
5660 needs partial registers on v8. On v9, structures with integer
5661 values in arg slots 5,6 will be passed in %o5 and SP+176, and complex fp
5662 values that begin in the last fp reg [where "last fp reg" varies with the
5663 mode] will be split between that reg and memory. */
5666 function_arg_partial_nregs (const struct sparc_args *cum,
5667 enum machine_mode mode, tree type, int named)
5669 int slotno, regno, padding;
5671 /* We pass 0 for incoming_p here, it doesn't matter. */
5672 slotno = function_arg_slotno (cum, mode, type, named, 0, ®no, &padding);
5679 if ((slotno + (mode == BLKmode
5680 ? ROUND_ADVANCE (int_size_in_bytes (type))
5681 : ROUND_ADVANCE (GET_MODE_SIZE (mode))))
5682 > SPARC_INT_ARG_MAX)
5683 return SPARC_INT_ARG_MAX - slotno;
5687 /* We are guaranteed by pass_by_reference that the size of the
5688 argument is not greater than 16 bytes, so we only need to
5689 return 1 if the argument is partially passed in registers. */
5691 if (type && AGGREGATE_TYPE_P (type))
5693 int size = int_size_in_bytes (type);
5695 if (size > UNITS_PER_WORD
5696 && slotno == SPARC_INT_ARG_MAX - 1)
5699 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_INT
5700 || (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT
5701 && ! (TARGET_FPU && named)))
5703 /* The complex types are passed as packed types. */
5704 if (GET_MODE_SIZE (mode) > UNITS_PER_WORD
5705 && slotno == SPARC_INT_ARG_MAX - 1)
5708 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT)
5710 if ((slotno + GET_MODE_SIZE (mode) / UNITS_PER_WORD)
5719 /* Return true if the argument should be passed by reference.
5720 !v9: The SPARC ABI stipulates passing struct arguments (of any size) and
5721 quad-precision floats by invisible reference.
5722 v9: Aggregates greater than 16 bytes are passed by reference.
5723 For Pascal, also pass arrays by reference. */
5726 sparc_pass_by_reference (CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED,
5727 enum machine_mode mode, tree type,
5728 bool named ATTRIBUTE_UNUSED)
5732 return ((type && AGGREGATE_TYPE_P (type))
5733 /* Extended ABI (as implemented by the Sun compiler) says
5734 that all complex floats are passed in memory. */
5736 /* Enforce the 2-word cap for passing arguments in registers.
5737 This affects CDImode, TFmode, DCmode, TCmode and large
5739 || GET_MODE_SIZE (mode) > 8);
5743 return ((type && TREE_CODE (type) == ARRAY_TYPE)
5745 && AGGREGATE_TYPE_P (type)
5746 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 16)
5747 /* Enforce the 2-word cap for passing arguments in registers.
5748 This affects CTImode, TCmode and large vector modes. */
5749 || GET_MODE_SIZE (mode) > 16);
5753 /* Handle the FUNCTION_ARG_ADVANCE macro.
5754 Update the data in CUM to advance over an argument
5755 of mode MODE and data type TYPE.
5756 TYPE is null for libcalls where that information may not be available. */
5759 function_arg_advance (struct sparc_args *cum, enum machine_mode mode,
5760 tree type, int named)
5762 int slotno, regno, padding;
5764 /* We pass 0 for incoming_p here, it doesn't matter. */
5765 slotno = function_arg_slotno (cum, mode, type, named, 0, ®no, &padding);
5767 /* If register required leading padding, add it. */
5769 cum->words += padding;
5773 cum->words += (mode != BLKmode
5774 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
5775 : ROUND_ADVANCE (int_size_in_bytes (type)));
5779 if (type && AGGREGATE_TYPE_P (type))
5781 int size = int_size_in_bytes (type);
5785 else if (size <= 16)
5787 else /* passed by reference */
5792 cum->words += (mode != BLKmode
5793 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
5794 : ROUND_ADVANCE (int_size_in_bytes (type)));
5799 /* Handle the FUNCTION_ARG_PADDING macro.
5800 For the 64 bit ABI structs are always stored left shifted in their
5804 function_arg_padding (enum machine_mode mode, tree type)
5806 if (TARGET_ARCH64 && type != 0 && AGGREGATE_TYPE_P (type))
5809 /* Fall back to the default. */
5810 return DEFAULT_FUNCTION_ARG_PADDING (mode, type);
5813 /* Handle the TARGET_RETURN_IN_MEMORY target hook.
5814 Specify whether to return the return value in memory. */
5817 sparc_return_in_memory (tree type, tree fntype ATTRIBUTE_UNUSED)
5820 /* Original SPARC 32-bit ABI says that quad-precision floats
5821 and all structures are returned in memory. Extended ABI
5822 (as implemented by the Sun compiler) says that all complex
5823 floats are returned in registers (8 FP registers at most
5824 for '_Complex long double'). Return all complex integers
5825 in registers (4 at most for '_Complex long long'). */
5826 return (TYPE_MODE (type) == BLKmode
5827 || TYPE_MODE (type) == TFmode
5828 /* Integral vector types follow the scalar FP types conventions. */
5829 || (GET_MODE_CLASS (TYPE_MODE (type)) == MODE_VECTOR_INT
5830 && GET_MODE_SIZE (TYPE_MODE (type)) > 8)
5831 /* FP vector types follow the complex FP types conventions. */
5832 || (GET_MODE_CLASS (TYPE_MODE (type)) == MODE_VECTOR_FLOAT
5833 && GET_MODE_SIZE (TYPE_MODE (type)) > 32));
5835 /* Original SPARC 64-bit ABI says that structures and unions
5836 smaller than 32 bytes are returned in registers. Extended
5837 ABI (as implemented by the Sun compiler) says that all complex
5838 floats are returned in registers (8 FP registers at most
5839 for '_Complex long double'). Return all complex integers
5840 in registers (4 at most for '_Complex TItype'). */
5841 return ((TYPE_MODE (type) == BLKmode
5842 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 32)
5843 || GET_MODE_SIZE (TYPE_MODE (type)) > 32);
5846 /* Handle the TARGET_STRUCT_VALUE target hook.
5847 Return where to find the structure return value address. */
5850 sparc_struct_value_rtx (tree fndecl ATTRIBUTE_UNUSED, int incoming)
5857 return gen_rtx_MEM (Pmode, plus_constant (frame_pointer_rtx,
5858 STRUCT_VALUE_OFFSET));
5860 return gen_rtx_MEM (Pmode, plus_constant (stack_pointer_rtx,
5861 STRUCT_VALUE_OFFSET));
5865 /* Handle FUNCTION_VALUE, FUNCTION_OUTGOING_VALUE, and LIBCALL_VALUE macros.
5866 For v9, function return values are subject to the same rules as arguments,
5867 except that up to 32 bytes may be returned in registers. */
5870 function_value (tree type, enum machine_mode mode, int incoming_p)
5872 /* Beware that the two values are swapped here wrt function_arg. */
5873 int regbase = (incoming_p
5874 ? SPARC_OUTGOING_INT_ARG_FIRST
5875 : SPARC_INCOMING_INT_ARG_FIRST);
5878 if (TARGET_ARCH64 && type)
5880 if (TREE_CODE (type) == RECORD_TYPE)
5882 /* Structures up to 32 bytes in size are passed in registers,
5883 promoted to fp registers where possible. */
5885 if (int_size_in_bytes (type) > 32)
5886 abort (); /* shouldn't get here */
5888 return function_arg_record_value (type, mode, 0, 1, regbase);
5890 else if (TREE_CODE (type) == UNION_TYPE)
5892 HOST_WIDE_INT size = int_size_in_bytes (type);
5895 abort (); /* shouldn't get here */
5897 return function_arg_union_value (size, mode, regbase);
5899 else if (AGGREGATE_TYPE_P (type))
5901 /* All other aggregate types are passed in an integer register
5902 in a mode corresponding to the size of the type. */
5903 HOST_WIDE_INT bytes = int_size_in_bytes (type);
5906 abort (); /* shouldn't get here */
5908 mode = mode_for_size (bytes * BITS_PER_UNIT, MODE_INT, 0);
5910 /* ??? We probably should have made the same ABI change in
5911 3.4.0 as the one we made for unions. The latter was
5912 required by the SCD though, while the former is not
5913 specified, so we favored compatibility and efficiency.
5915 Now we're stuck for aggregates larger than 16 bytes,
5916 because OImode vanished in the meantime. Let's not
5917 try to be unduly clever, and simply follow the ABI
5918 for unions in that case. */
5919 if (mode == BLKmode)
5920 return function_arg_union_value (bytes, mode, regbase);
5922 else if (GET_MODE_CLASS (mode) == MODE_INT
5923 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5927 if (TARGET_FPU && (FLOAT_MODE_P (mode) || VECTOR_MODE_P (mode)))
5928 regno = SPARC_FP_ARG_FIRST;
5932 return gen_rtx_REG (mode, regno);
5935 /* Do what is necessary for `va_start'. We look at the current function
5936 to determine if stdarg or varargs is used and return the address of
5937 the first unnamed parameter. */
5940 sparc_builtin_saveregs (void)
5942 int first_reg = current_function_args_info.words;
5946 for (regno = first_reg; regno < SPARC_INT_ARG_MAX; regno++)
5947 emit_move_insn (gen_rtx_MEM (word_mode,
5948 gen_rtx_PLUS (Pmode,
5950 GEN_INT (FIRST_PARM_OFFSET (0)
5953 gen_rtx_REG (word_mode,
5954 SPARC_INCOMING_INT_ARG_FIRST + regno));
5956 address = gen_rtx_PLUS (Pmode,
5958 GEN_INT (FIRST_PARM_OFFSET (0)
5959 + UNITS_PER_WORD * first_reg));
5964 /* Implement `va_start' for stdarg. */
5967 sparc_va_start (tree valist, rtx nextarg)
5969 nextarg = expand_builtin_saveregs ();
5970 std_expand_builtin_va_start (valist, nextarg);
5973 /* Implement `va_arg' for stdarg. */
5976 sparc_gimplify_va_arg (tree valist, tree type, tree *pre_p, tree *post_p)
5978 HOST_WIDE_INT size, rsize, align;
5981 tree ptrtype = build_pointer_type (type);
5983 if (pass_by_reference (NULL, TYPE_MODE (type), type, 0))
5986 size = rsize = UNITS_PER_WORD;
5992 size = int_size_in_bytes (type);
5993 rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD;
5998 /* For SPARC64, objects requiring 16-byte alignment get it. */
5999 if (TYPE_ALIGN (type) >= 2 * (unsigned) BITS_PER_WORD)
6000 align = 2 * UNITS_PER_WORD;
6002 /* SPARC-V9 ABI states that structures up to 16 bytes in size
6003 are given whole slots as needed. */
6004 if (AGGREGATE_TYPE_P (type))
6007 size = rsize = UNITS_PER_WORD;
6017 incr = fold (build2 (PLUS_EXPR, ptr_type_node, incr,
6018 ssize_int (align - 1)));
6019 incr = fold (build2 (BIT_AND_EXPR, ptr_type_node, incr,
6020 ssize_int (-align)));
6023 gimplify_expr (&incr, pre_p, post_p, is_gimple_val, fb_rvalue);
6026 if (BYTES_BIG_ENDIAN && size < rsize)
6027 addr = fold (build2 (PLUS_EXPR, ptr_type_node, incr,
6028 ssize_int (rsize - size)));
6032 addr = fold_convert (build_pointer_type (ptrtype), addr);
6033 addr = build_fold_indirect_ref (addr);
6035 /* If the address isn't aligned properly for the type,
6036 we may need to copy to a temporary.
6037 FIXME: This is inefficient. Usually we can do this
6040 && TYPE_ALIGN (type) > BITS_PER_WORD)
6042 tree tmp = create_tmp_var (type, "va_arg_tmp");
6043 tree dest_addr = build_fold_addr_expr (tmp);
6045 tree copy = build_function_call_expr
6046 (implicit_built_in_decls[BUILT_IN_MEMCPY],
6047 tree_cons (NULL_TREE, dest_addr,
6048 tree_cons (NULL_TREE, addr,
6049 tree_cons (NULL_TREE, size_int (rsize),
6052 gimplify_and_add (copy, pre_p);
6056 addr = fold_convert (ptrtype, addr);
6058 incr = fold (build2 (PLUS_EXPR, ptr_type_node, incr, ssize_int (rsize)));
6059 incr = build2 (MODIFY_EXPR, ptr_type_node, valist, incr);
6060 gimplify_and_add (incr, post_p);
6062 return build_fold_indirect_ref (addr);
6065 /* Return the string to output an unconditional branch to LABEL, which is
6066 the operand number of the label.
6068 DEST is the destination insn (i.e. the label), INSN is the source. */
6071 output_ubranch (rtx dest, int label, rtx insn)
6073 static char string[64];
6074 bool v9_form = false;
6077 if (TARGET_V9 && INSN_ADDRESSES_SET_P ())
6079 int delta = (INSN_ADDRESSES (INSN_UID (dest))
6080 - INSN_ADDRESSES (INSN_UID (insn)));
6081 /* Leave some instructions for "slop". */
6082 if (delta >= -260000 && delta < 260000)
6087 strcpy (string, "ba%*,pt\t%%xcc, ");
6089 strcpy (string, "b%*\t");
6091 p = strchr (string, '\0');
6102 /* Return the string to output a conditional branch to LABEL, which is
6103 the operand number of the label. OP is the conditional expression.
6104 XEXP (OP, 0) is assumed to be a condition code register (integer or
6105 floating point) and its mode specifies what kind of comparison we made.
6107 DEST is the destination insn (i.e. the label), INSN is the source.
6109 REVERSED is nonzero if we should reverse the sense of the comparison.
6111 ANNUL is nonzero if we should generate an annulling branch. */
6114 output_cbranch (rtx op, rtx dest, int label, int reversed, int annul,
6117 static char string[64];
6118 enum rtx_code code = GET_CODE (op);
6119 rtx cc_reg = XEXP (op, 0);
6120 enum machine_mode mode = GET_MODE (cc_reg);
6121 const char *labelno, *branch;
6122 int spaces = 8, far;
6125 /* v9 branches are limited to +-1MB. If it is too far away,
6138 fbne,a,pn %fcc2, .LC29
6146 far = TARGET_V9 && (get_attr_length (insn) >= 3);
6149 /* Reversal of FP compares takes care -- an ordered compare
6150 becomes an unordered compare and vice versa. */
6151 if (mode == CCFPmode || mode == CCFPEmode)
6152 code = reverse_condition_maybe_unordered (code);
6154 code = reverse_condition (code);
6157 /* Start by writing the branch condition. */
6158 if (mode == CCFPmode || mode == CCFPEmode)
6209 /* ??? !v9: FP branches cannot be preceded by another floating point
6210 insn. Because there is currently no concept of pre-delay slots,
6211 we can fix this only by always emitting a nop before a floating
6216 strcpy (string, "nop\n\t");
6217 strcat (string, branch);
6230 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
6242 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
6263 strcpy (string, branch);
6265 spaces -= strlen (branch);
6266 p = strchr (string, '\0');
6268 /* Now add the annulling, the label, and a possible noop. */
6281 if (! far && insn && INSN_ADDRESSES_SET_P ())
6283 int delta = (INSN_ADDRESSES (INSN_UID (dest))
6284 - INSN_ADDRESSES (INSN_UID (insn)));
6285 /* Leave some instructions for "slop". */
6286 if (delta < -260000 || delta >= 260000)
6290 if (mode == CCFPmode || mode == CCFPEmode)
6292 static char v9_fcc_labelno[] = "%%fccX, ";
6293 /* Set the char indicating the number of the fcc reg to use. */
6294 v9_fcc_labelno[5] = REGNO (cc_reg) - SPARC_FIRST_V9_FCC_REG + '0';
6295 labelno = v9_fcc_labelno;
6298 if (REGNO (cc_reg) == SPARC_FCC_REG)
6304 else if (mode == CCXmode || mode == CCX_NOOVmode)
6306 labelno = "%%xcc, ";
6312 labelno = "%%icc, ";
6317 if (*labelno && insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
6320 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
6333 strcpy (p, labelno);
6334 p = strchr (p, '\0');
6337 strcpy (p, ".+12\n\t nop\n\tb\t");
6338 /* Skip the next insn if requested or
6339 if we know that it will be a nop. */
6340 if (annul || ! final_sequence)
6354 /* Emit a library call comparison between floating point X and Y.
6355 COMPARISON is the rtl operator to compare with (EQ, NE, GT, etc.).
6356 TARGET_ARCH64 uses _Qp_* functions, which use pointers to TFmode
6357 values as arguments instead of the TFmode registers themselves,
6358 that's why we cannot call emit_float_lib_cmp. */
6360 sparc_emit_float_lib_cmp (rtx x, rtx y, enum rtx_code comparison)
6363 rtx slot0, slot1, result, tem, tem2;
6364 enum machine_mode mode;
6369 qpfunc = (TARGET_ARCH64) ? "_Qp_feq" : "_Q_feq";
6373 qpfunc = (TARGET_ARCH64) ? "_Qp_fne" : "_Q_fne";
6377 qpfunc = (TARGET_ARCH64) ? "_Qp_fgt" : "_Q_fgt";
6381 qpfunc = (TARGET_ARCH64) ? "_Qp_fge" : "_Q_fge";
6385 qpfunc = (TARGET_ARCH64) ? "_Qp_flt" : "_Q_flt";
6389 qpfunc = (TARGET_ARCH64) ? "_Qp_fle" : "_Q_fle";
6400 qpfunc = (TARGET_ARCH64) ? "_Qp_cmp" : "_Q_cmp";
6410 if (GET_CODE (x) != MEM)
6412 slot0 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0);
6413 emit_insn (gen_rtx_SET (VOIDmode, slot0, x));
6418 if (GET_CODE (y) != MEM)
6420 slot1 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0);
6421 emit_insn (gen_rtx_SET (VOIDmode, slot1, y));
6426 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, qpfunc), LCT_NORMAL,
6428 XEXP (slot0, 0), Pmode,
6429 XEXP (slot1, 0), Pmode);
6435 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, qpfunc), LCT_NORMAL,
6437 x, TFmode, y, TFmode);
6443 /* Immediately move the result of the libcall into a pseudo
6444 register so reload doesn't clobber the value if it needs
6445 the return register for a spill reg. */
6446 result = gen_reg_rtx (mode);
6447 emit_move_insn (result, hard_libcall_value (mode));
6452 emit_cmp_insn (result, const0_rtx, NE, NULL_RTX, mode, 0);
6456 emit_cmp_insn (result, GEN_INT(3), comparison == UNORDERED ? EQ : NE,
6461 emit_cmp_insn (result, const1_rtx,
6462 comparison == UNGT ? GT : NE, NULL_RTX, mode, 0);
6465 emit_cmp_insn (result, const2_rtx, NE, NULL_RTX, mode, 0);
6468 tem = gen_reg_rtx (mode);
6470 emit_insn (gen_andsi3 (tem, result, const1_rtx));
6472 emit_insn (gen_anddi3 (tem, result, const1_rtx));
6473 emit_cmp_insn (tem, const0_rtx, NE, NULL_RTX, mode, 0);
6477 tem = gen_reg_rtx (mode);
6479 emit_insn (gen_addsi3 (tem, result, const1_rtx));
6481 emit_insn (gen_adddi3 (tem, result, const1_rtx));
6482 tem2 = gen_reg_rtx (mode);
6484 emit_insn (gen_andsi3 (tem2, tem, const2_rtx));
6486 emit_insn (gen_anddi3 (tem2, tem, const2_rtx));
6487 emit_cmp_insn (tem2, const0_rtx, comparison == UNEQ ? EQ : NE,
6493 /* Generate an unsigned DImode to FP conversion. This is the same code
6494 optabs would emit if we didn't have TFmode patterns. */
6497 sparc_emit_floatunsdi (rtx *operands, enum machine_mode mode)
6499 rtx neglab, donelab, i0, i1, f0, in, out;
6502 in = force_reg (DImode, operands[1]);
6503 neglab = gen_label_rtx ();
6504 donelab = gen_label_rtx ();
6505 i0 = gen_reg_rtx (DImode);
6506 i1 = gen_reg_rtx (DImode);
6507 f0 = gen_reg_rtx (mode);
6509 emit_cmp_and_jump_insns (in, const0_rtx, LT, const0_rtx, DImode, 0, neglab);
6511 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_FLOAT (mode, in)));
6512 emit_jump_insn (gen_jump (donelab));
6515 emit_label (neglab);
6517 emit_insn (gen_lshrdi3 (i0, in, const1_rtx));
6518 emit_insn (gen_anddi3 (i1, in, const1_rtx));
6519 emit_insn (gen_iordi3 (i0, i0, i1));
6520 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_FLOAT (mode, i0)));
6521 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_PLUS (mode, f0, f0)));
6523 emit_label (donelab);
6526 /* Generate an FP to unsigned DImode conversion. This is the same code
6527 optabs would emit if we didn't have TFmode patterns. */
6530 sparc_emit_fixunsdi (rtx *operands, enum machine_mode mode)
6532 rtx neglab, donelab, i0, i1, f0, in, out, limit;
6535 in = force_reg (mode, operands[1]);
6536 neglab = gen_label_rtx ();
6537 donelab = gen_label_rtx ();
6538 i0 = gen_reg_rtx (DImode);
6539 i1 = gen_reg_rtx (DImode);
6540 limit = gen_reg_rtx (mode);
6541 f0 = gen_reg_rtx (mode);
6543 emit_move_insn (limit,
6544 CONST_DOUBLE_FROM_REAL_VALUE (
6545 REAL_VALUE_ATOF ("9223372036854775808.0", mode), mode));
6546 emit_cmp_and_jump_insns (in, limit, GE, NULL_RTX, mode, 0, neglab);
6548 emit_insn (gen_rtx_SET (VOIDmode,
6550 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, in))));
6551 emit_jump_insn (gen_jump (donelab));
6554 emit_label (neglab);
6556 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_MINUS (mode, in, limit)));
6557 emit_insn (gen_rtx_SET (VOIDmode,
6559 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, f0))));
6560 emit_insn (gen_movdi (i1, const1_rtx));
6561 emit_insn (gen_ashldi3 (i1, i1, GEN_INT (63)));
6562 emit_insn (gen_xordi3 (out, i0, i1));
6564 emit_label (donelab);
6567 /* Return the string to output a conditional branch to LABEL, testing
6568 register REG. LABEL is the operand number of the label; REG is the
6569 operand number of the reg. OP is the conditional expression. The mode
6570 of REG says what kind of comparison we made.
6572 DEST is the destination insn (i.e. the label), INSN is the source.
6574 REVERSED is nonzero if we should reverse the sense of the comparison.
6576 ANNUL is nonzero if we should generate an annulling branch. */
6579 output_v9branch (rtx op, rtx dest, int reg, int label, int reversed,
6580 int annul, rtx insn)
6582 static char string[64];
6583 enum rtx_code code = GET_CODE (op);
6584 enum machine_mode mode = GET_MODE (XEXP (op, 0));
6589 /* branch on register are limited to +-128KB. If it is too far away,
6602 brgez,a,pn %o1, .LC29
6608 ba,pt %xcc, .LC29 */
6610 far = get_attr_length (insn) >= 3;
6612 /* If not floating-point or if EQ or NE, we can just reverse the code. */
6614 code = reverse_condition (code);
6616 /* Only 64 bit versions of these instructions exist. */
6620 /* Start by writing the branch condition. */
6625 strcpy (string, "brnz");
6629 strcpy (string, "brz");
6633 strcpy (string, "brgez");
6637 strcpy (string, "brlz");
6641 strcpy (string, "brlez");
6645 strcpy (string, "brgz");
6652 p = strchr (string, '\0');
6654 /* Now add the annulling, reg, label, and nop. */
6661 if (insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
6664 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
6669 *p = p < string + 8 ? '\t' : ' ';
6677 int veryfar = 1, delta;
6679 if (INSN_ADDRESSES_SET_P ())
6681 delta = (INSN_ADDRESSES (INSN_UID (dest))
6682 - INSN_ADDRESSES (INSN_UID (insn)));
6683 /* Leave some instructions for "slop". */
6684 if (delta >= -260000 && delta < 260000)
6688 strcpy (p, ".+12\n\t nop\n\t");
6689 /* Skip the next insn if requested or
6690 if we know that it will be a nop. */
6691 if (annul || ! final_sequence)
6701 strcpy (p, "ba,pt\t%%xcc, ");
6715 /* Return 1, if any of the registers of the instruction are %l[0-7] or %o[0-7].
6716 Such instructions cannot be used in the delay slot of return insn on v9.
6717 If TEST is 0, also rename all %i[0-7] registers to their %o[0-7] counterparts.
6721 epilogue_renumber (register rtx *where, int test)
6723 register const char *fmt;
6725 register enum rtx_code code;
6730 code = GET_CODE (*where);
6735 if (REGNO (*where) >= 8 && REGNO (*where) < 24) /* oX or lX */
6737 if (! test && REGNO (*where) >= 24 && REGNO (*where) < 32)
6738 *where = gen_rtx_REG (GET_MODE (*where), OUTGOING_REGNO (REGNO(*where)));
6746 /* Do not replace the frame pointer with the stack pointer because
6747 it can cause the delayed instruction to load below the stack.
6748 This occurs when instructions like:
6750 (set (reg/i:SI 24 %i0)
6751 (mem/f:SI (plus:SI (reg/f:SI 30 %fp)
6752 (const_int -20 [0xffffffec])) 0))
6754 are in the return delayed slot. */
6756 if (GET_CODE (XEXP (*where, 0)) == REG
6757 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM
6758 && (GET_CODE (XEXP (*where, 1)) != CONST_INT
6759 || INTVAL (XEXP (*where, 1)) < SPARC_STACK_BIAS))
6764 if (SPARC_STACK_BIAS
6765 && GET_CODE (XEXP (*where, 0)) == REG
6766 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM)
6774 fmt = GET_RTX_FORMAT (code);
6776 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6781 for (j = XVECLEN (*where, i) - 1; j >= 0; j--)
6782 if (epilogue_renumber (&(XVECEXP (*where, i, j)), test))
6785 else if (fmt[i] == 'e'
6786 && epilogue_renumber (&(XEXP (*where, i)), test))
6792 /* Leaf functions and non-leaf functions have different needs. */
6795 reg_leaf_alloc_order[] = REG_LEAF_ALLOC_ORDER;
6798 reg_nonleaf_alloc_order[] = REG_ALLOC_ORDER;
6800 static const int *const reg_alloc_orders[] = {
6801 reg_leaf_alloc_order,
6802 reg_nonleaf_alloc_order};
6805 order_regs_for_local_alloc (void)
6807 static int last_order_nonleaf = 1;
6809 if (regs_ever_live[15] != last_order_nonleaf)
6811 last_order_nonleaf = !last_order_nonleaf;
6812 memcpy ((char *) reg_alloc_order,
6813 (const char *) reg_alloc_orders[last_order_nonleaf],
6814 FIRST_PSEUDO_REGISTER * sizeof (int));
6818 /* Return 1 if REG and MEM are legitimate enough to allow the various
6819 mem<-->reg splits to be run. */
6822 sparc_splitdi_legitimate (rtx reg, rtx mem)
6824 /* Punt if we are here by mistake. */
6825 if (! reload_completed)
6828 /* We must have an offsettable memory reference. */
6829 if (! offsettable_memref_p (mem))
6832 /* If we have legitimate args for ldd/std, we do not want
6833 the split to happen. */
6834 if ((REGNO (reg) % 2) == 0
6835 && mem_min_alignment (mem, 8))
6842 /* Return 1 if x and y are some kind of REG and they refer to
6843 different hard registers. This test is guaranteed to be
6844 run after reload. */
6847 sparc_absnegfloat_split_legitimate (rtx x, rtx y)
6849 if (GET_CODE (x) != REG)
6851 if (GET_CODE (y) != REG)
6853 if (REGNO (x) == REGNO (y))
6858 /* Return 1 if REGNO (reg1) is even and REGNO (reg1) == REGNO (reg2) - 1.
6859 This makes them candidates for using ldd and std insns.
6861 Note reg1 and reg2 *must* be hard registers. */
6864 registers_ok_for_ldd_peep (rtx reg1, rtx reg2)
6866 /* We might have been passed a SUBREG. */
6867 if (GET_CODE (reg1) != REG || GET_CODE (reg2) != REG)
6870 if (REGNO (reg1) % 2 != 0)
6873 /* Integer ldd is deprecated in SPARC V9 */
6874 if (TARGET_V9 && REGNO (reg1) < 32)
6877 return (REGNO (reg1) == REGNO (reg2) - 1);
6880 /* Return 1 if the addresses in mem1 and mem2 are suitable for use in
6883 This can only happen when addr1 and addr2, the addresses in mem1
6884 and mem2, are consecutive memory locations (addr1 + 4 == addr2).
6885 addr1 must also be aligned on a 64-bit boundary.
6887 Also iff dependent_reg_rtx is not null it should not be used to
6888 compute the address for mem1, i.e. we cannot optimize a sequence
6900 But, note that the transformation from:
6905 is perfectly fine. Thus, the peephole2 patterns always pass us
6906 the destination register of the first load, never the second one.
6908 For stores we don't have a similar problem, so dependent_reg_rtx is
6912 mems_ok_for_ldd_peep (rtx mem1, rtx mem2, rtx dependent_reg_rtx)
6916 HOST_WIDE_INT offset1;
6918 /* The mems cannot be volatile. */
6919 if (MEM_VOLATILE_P (mem1) || MEM_VOLATILE_P (mem2))
6922 /* MEM1 should be aligned on a 64-bit boundary. */
6923 if (MEM_ALIGN (mem1) < 64)
6926 addr1 = XEXP (mem1, 0);
6927 addr2 = XEXP (mem2, 0);
6929 /* Extract a register number and offset (if used) from the first addr. */
6930 if (GET_CODE (addr1) == PLUS)
6932 /* If not a REG, return zero. */
6933 if (GET_CODE (XEXP (addr1, 0)) != REG)
6937 reg1 = REGNO (XEXP (addr1, 0));
6938 /* The offset must be constant! */
6939 if (GET_CODE (XEXP (addr1, 1)) != CONST_INT)
6941 offset1 = INTVAL (XEXP (addr1, 1));
6944 else if (GET_CODE (addr1) != REG)
6948 reg1 = REGNO (addr1);
6949 /* This was a simple (mem (reg)) expression. Offset is 0. */
6953 /* Make sure the second address is a (mem (plus (reg) (const_int). */
6954 if (GET_CODE (addr2) != PLUS)
6957 if (GET_CODE (XEXP (addr2, 0)) != REG
6958 || GET_CODE (XEXP (addr2, 1)) != CONST_INT)
6961 if (reg1 != REGNO (XEXP (addr2, 0)))
6964 if (dependent_reg_rtx != NULL_RTX && reg1 == REGNO (dependent_reg_rtx))
6967 /* The first offset must be evenly divisible by 8 to ensure the
6968 address is 64 bit aligned. */
6969 if (offset1 % 8 != 0)
6972 /* The offset for the second addr must be 4 more than the first addr. */
6973 if (INTVAL (XEXP (addr2, 1)) != offset1 + 4)
6976 /* All the tests passed. addr1 and addr2 are valid for ldd and std
6981 /* Return 1 if reg is a pseudo, or is the first register in
6982 a hard register pair. This makes it a candidate for use in
6983 ldd and std insns. */
6986 register_ok_for_ldd (rtx reg)
6988 /* We might have been passed a SUBREG. */
6989 if (GET_CODE (reg) != REG)
6992 if (REGNO (reg) < FIRST_PSEUDO_REGISTER)
6993 return (REGNO (reg) % 2 == 0);
6998 /* Print operand X (an rtx) in assembler syntax to file FILE.
6999 CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified.
7000 For `%' followed by punctuation, CODE is the punctuation and X is null. */
7003 print_operand (FILE *file, rtx x, int code)
7008 /* Output an insn in a delay slot. */
7010 sparc_indent_opcode = 1;
7012 fputs ("\n\t nop", file);
7015 /* Output an annul flag if there's nothing for the delay slot and we
7016 are optimizing. This is always used with '(' below.
7017 Sun OS 4.1.1 dbx can't handle an annulled unconditional branch;
7018 this is a dbx bug. So, we only do this when optimizing.
7019 On UltraSPARC, a branch in a delay slot causes a pipeline flush.
7020 Always emit a nop in case the next instruction is a branch. */
7021 if (! final_sequence && (optimize && (int)sparc_cpu < PROCESSOR_V9))
7025 /* Output a 'nop' if there's nothing for the delay slot and we are
7026 not optimizing. This is always used with '*' above. */
7027 if (! final_sequence && ! (optimize && (int)sparc_cpu < PROCESSOR_V9))
7028 fputs ("\n\t nop", file);
7029 else if (final_sequence)
7030 sparc_indent_opcode = 1;
7033 /* Output the right displacement from the saved PC on function return.
7034 The caller may have placed an "unimp" insn immediately after the call
7035 so we have to account for it. This insn is used in the 32-bit ABI
7036 when calling a function that returns a non zero-sized structure. The
7037 64-bit ABI doesn't have it. Be careful to have this test be the same
7038 as that used on the call. */
7040 && current_function_returns_struct
7041 && (TREE_CODE (DECL_SIZE (DECL_RESULT (current_function_decl)))
7043 && ! integer_zerop (DECL_SIZE (DECL_RESULT (current_function_decl))))
7049 /* Output the Embedded Medium/Anywhere code model base register. */
7050 fputs (EMBMEDANY_BASE_REG, file);
7053 /* Print some local dynamic TLS name. */
7054 assemble_name (file, get_some_local_dynamic_name ());
7058 /* Adjust the operand to take into account a RESTORE operation. */
7059 if (GET_CODE (x) == CONST_INT)
7061 else if (GET_CODE (x) != REG)
7062 output_operand_lossage ("invalid %%Y operand");
7063 else if (REGNO (x) < 8)
7064 fputs (reg_names[REGNO (x)], file);
7065 else if (REGNO (x) >= 24 && REGNO (x) < 32)
7066 fputs (reg_names[REGNO (x)-16], file);
7068 output_operand_lossage ("invalid %%Y operand");
7071 /* Print out the low order register name of a register pair. */
7072 if (WORDS_BIG_ENDIAN)
7073 fputs (reg_names[REGNO (x)+1], file);
7075 fputs (reg_names[REGNO (x)], file);
7078 /* Print out the high order register name of a register pair. */
7079 if (WORDS_BIG_ENDIAN)
7080 fputs (reg_names[REGNO (x)], file);
7082 fputs (reg_names[REGNO (x)+1], file);
7085 /* Print out the second register name of a register pair or quad.
7086 I.e., R (%o0) => %o1. */
7087 fputs (reg_names[REGNO (x)+1], file);
7090 /* Print out the third register name of a register quad.
7091 I.e., S (%o0) => %o2. */
7092 fputs (reg_names[REGNO (x)+2], file);
7095 /* Print out the fourth register name of a register quad.
7096 I.e., T (%o0) => %o3. */
7097 fputs (reg_names[REGNO (x)+3], file);
7100 /* Print a condition code register. */
7101 if (REGNO (x) == SPARC_ICC_REG)
7103 /* We don't handle CC[X]_NOOVmode because they're not supposed
7105 if (GET_MODE (x) == CCmode)
7106 fputs ("%icc", file);
7107 else if (GET_MODE (x) == CCXmode)
7108 fputs ("%xcc", file);
7113 /* %fccN register */
7114 fputs (reg_names[REGNO (x)], file);
7117 /* Print the operand's address only. */
7118 output_address (XEXP (x, 0));
7121 /* In this case we need a register. Use %g0 if the
7122 operand is const0_rtx. */
7124 || (GET_MODE (x) != VOIDmode && x == CONST0_RTX (GET_MODE (x))))
7126 fputs ("%g0", file);
7133 switch (GET_CODE (x))
7135 case IOR: fputs ("or", file); break;
7136 case AND: fputs ("and", file); break;
7137 case XOR: fputs ("xor", file); break;
7138 default: output_operand_lossage ("invalid %%A operand");
7143 switch (GET_CODE (x))
7145 case IOR: fputs ("orn", file); break;
7146 case AND: fputs ("andn", file); break;
7147 case XOR: fputs ("xnor", file); break;
7148 default: output_operand_lossage ("invalid %%B operand");
7152 /* These are used by the conditional move instructions. */
7156 enum rtx_code rc = GET_CODE (x);
7160 enum machine_mode mode = GET_MODE (XEXP (x, 0));
7161 if (mode == CCFPmode || mode == CCFPEmode)
7162 rc = reverse_condition_maybe_unordered (GET_CODE (x));
7164 rc = reverse_condition (GET_CODE (x));
7168 case NE: fputs ("ne", file); break;
7169 case EQ: fputs ("e", file); break;
7170 case GE: fputs ("ge", file); break;
7171 case GT: fputs ("g", file); break;
7172 case LE: fputs ("le", file); break;
7173 case LT: fputs ("l", file); break;
7174 case GEU: fputs ("geu", file); break;
7175 case GTU: fputs ("gu", file); break;
7176 case LEU: fputs ("leu", file); break;
7177 case LTU: fputs ("lu", file); break;
7178 case LTGT: fputs ("lg", file); break;
7179 case UNORDERED: fputs ("u", file); break;
7180 case ORDERED: fputs ("o", file); break;
7181 case UNLT: fputs ("ul", file); break;
7182 case UNLE: fputs ("ule", file); break;
7183 case UNGT: fputs ("ug", file); break;
7184 case UNGE: fputs ("uge", file); break;
7185 case UNEQ: fputs ("ue", file); break;
7186 default: output_operand_lossage (code == 'c'
7187 ? "invalid %%c operand"
7188 : "invalid %%C operand");
7193 /* These are used by the movr instruction pattern. */
7197 enum rtx_code rc = (code == 'd'
7198 ? reverse_condition (GET_CODE (x))
7202 case NE: fputs ("ne", file); break;
7203 case EQ: fputs ("e", file); break;
7204 case GE: fputs ("gez", file); break;
7205 case LT: fputs ("lz", file); break;
7206 case LE: fputs ("lez", file); break;
7207 case GT: fputs ("gz", file); break;
7208 default: output_operand_lossage (code == 'd'
7209 ? "invalid %%d operand"
7210 : "invalid %%D operand");
7217 /* Print a sign-extended character. */
7218 int i = trunc_int_for_mode (INTVAL (x), QImode);
7219 fprintf (file, "%d", i);
7224 /* Operand must be a MEM; write its address. */
7225 if (GET_CODE (x) != MEM)
7226 output_operand_lossage ("invalid %%f operand");
7227 output_address (XEXP (x, 0));
7232 /* Print a sign-extended 32-bit value. */
7234 if (GET_CODE(x) == CONST_INT)
7236 else if (GET_CODE(x) == CONST_DOUBLE)
7237 i = CONST_DOUBLE_LOW (x);
7240 output_operand_lossage ("invalid %%s operand");
7243 i = trunc_int_for_mode (i, SImode);
7244 fprintf (file, HOST_WIDE_INT_PRINT_DEC, i);
7249 /* Do nothing special. */
7253 /* Undocumented flag. */
7254 output_operand_lossage ("invalid operand output code");
7257 if (GET_CODE (x) == REG)
7258 fputs (reg_names[REGNO (x)], file);
7259 else if (GET_CODE (x) == MEM)
7262 /* Poor Sun assembler doesn't understand absolute addressing. */
7263 if (CONSTANT_P (XEXP (x, 0)))
7264 fputs ("%g0+", file);
7265 output_address (XEXP (x, 0));
7268 else if (GET_CODE (x) == HIGH)
7270 fputs ("%hi(", file);
7271 output_addr_const (file, XEXP (x, 0));
7274 else if (GET_CODE (x) == LO_SUM)
7276 print_operand (file, XEXP (x, 0), 0);
7277 if (TARGET_CM_MEDMID)
7278 fputs ("+%l44(", file);
7280 fputs ("+%lo(", file);
7281 output_addr_const (file, XEXP (x, 1));
7284 else if (GET_CODE (x) == CONST_DOUBLE
7285 && (GET_MODE (x) == VOIDmode
7286 || GET_MODE_CLASS (GET_MODE (x)) == MODE_INT))
7288 if (CONST_DOUBLE_HIGH (x) == 0)
7289 fprintf (file, "%u", (unsigned int) CONST_DOUBLE_LOW (x));
7290 else if (CONST_DOUBLE_HIGH (x) == -1
7291 && CONST_DOUBLE_LOW (x) < 0)
7292 fprintf (file, "%d", (int) CONST_DOUBLE_LOW (x));
7294 output_operand_lossage ("long long constant not a valid immediate operand");
7296 else if (GET_CODE (x) == CONST_DOUBLE)
7297 output_operand_lossage ("floating point constant not a valid immediate operand");
7298 else { output_addr_const (file, x); }
7301 /* Target hook for assembling integer objects. The sparc version has
7302 special handling for aligned DI-mode objects. */
7305 sparc_assemble_integer (rtx x, unsigned int size, int aligned_p)
7307 /* ??? We only output .xword's for symbols and only then in environments
7308 where the assembler can handle them. */
7309 if (aligned_p && size == 8
7310 && (GET_CODE (x) != CONST_INT && GET_CODE (x) != CONST_DOUBLE))
7314 assemble_integer_with_op ("\t.xword\t", x);
7319 assemble_aligned_integer (4, const0_rtx);
7320 assemble_aligned_integer (4, x);
7324 return default_assemble_integer (x, size, aligned_p);
7327 /* Return the value of a code used in the .proc pseudo-op that says
7328 what kind of result this function returns. For non-C types, we pick
7329 the closest C type. */
7331 #ifndef SHORT_TYPE_SIZE
7332 #define SHORT_TYPE_SIZE (BITS_PER_UNIT * 2)
7335 #ifndef INT_TYPE_SIZE
7336 #define INT_TYPE_SIZE BITS_PER_WORD
7339 #ifndef LONG_TYPE_SIZE
7340 #define LONG_TYPE_SIZE BITS_PER_WORD
7343 #ifndef LONG_LONG_TYPE_SIZE
7344 #define LONG_LONG_TYPE_SIZE (BITS_PER_WORD * 2)
7347 #ifndef FLOAT_TYPE_SIZE
7348 #define FLOAT_TYPE_SIZE BITS_PER_WORD
7351 #ifndef DOUBLE_TYPE_SIZE
7352 #define DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
7355 #ifndef LONG_DOUBLE_TYPE_SIZE
7356 #define LONG_DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
7360 sparc_type_code (register tree type)
7362 register unsigned long qualifiers = 0;
7363 register unsigned shift;
7365 /* Only the first 30 bits of the qualifier are valid. We must refrain from
7366 setting more, since some assemblers will give an error for this. Also,
7367 we must be careful to avoid shifts of 32 bits or more to avoid getting
7368 unpredictable results. */
7370 for (shift = 6; shift < 30; shift += 2, type = TREE_TYPE (type))
7372 switch (TREE_CODE (type))
7378 qualifiers |= (3 << shift);
7383 qualifiers |= (2 << shift);
7387 case REFERENCE_TYPE:
7389 qualifiers |= (1 << shift);
7393 return (qualifiers | 8);
7396 case QUAL_UNION_TYPE:
7397 return (qualifiers | 9);
7400 return (qualifiers | 10);
7403 return (qualifiers | 16);
7406 /* If this is a range type, consider it to be the underlying
7408 if (TREE_TYPE (type) != 0)
7411 /* Carefully distinguish all the standard types of C,
7412 without messing up if the language is not C. We do this by
7413 testing TYPE_PRECISION and TYPE_UNSIGNED. The old code used to
7414 look at both the names and the above fields, but that's redundant.
7415 Any type whose size is between two C types will be considered
7416 to be the wider of the two types. Also, we do not have a
7417 special code to use for "long long", so anything wider than
7418 long is treated the same. Note that we can't distinguish
7419 between "int" and "long" in this code if they are the same
7420 size, but that's fine, since neither can the assembler. */
7422 if (TYPE_PRECISION (type) <= CHAR_TYPE_SIZE)
7423 return (qualifiers | (TYPE_UNSIGNED (type) ? 12 : 2));
7425 else if (TYPE_PRECISION (type) <= SHORT_TYPE_SIZE)
7426 return (qualifiers | (TYPE_UNSIGNED (type) ? 13 : 3));
7428 else if (TYPE_PRECISION (type) <= INT_TYPE_SIZE)
7429 return (qualifiers | (TYPE_UNSIGNED (type) ? 14 : 4));
7432 return (qualifiers | (TYPE_UNSIGNED (type) ? 15 : 5));
7435 /* If this is a range type, consider it to be the underlying
7437 if (TREE_TYPE (type) != 0)
7440 /* Carefully distinguish all the standard types of C,
7441 without messing up if the language is not C. */
7443 if (TYPE_PRECISION (type) == FLOAT_TYPE_SIZE)
7444 return (qualifiers | 6);
7447 return (qualifiers | 7);
7449 case COMPLEX_TYPE: /* GNU Fortran COMPLEX type. */
7450 /* ??? We need to distinguish between double and float complex types,
7451 but I don't know how yet because I can't reach this code from
7452 existing front-ends. */
7453 return (qualifiers | 7); /* Who knows? */
7456 case CHAR_TYPE: /* GNU Pascal CHAR type. Not used in C. */
7457 case BOOLEAN_TYPE: /* GNU Fortran BOOLEAN type. */
7458 case FILE_TYPE: /* GNU Pascal FILE type. */
7459 case SET_TYPE: /* GNU Pascal SET type. */
7460 case LANG_TYPE: /* ? */
7464 abort (); /* Not a type! */
7471 /* Nested function support. */
7473 /* Emit RTL insns to initialize the variable parts of a trampoline.
7474 FNADDR is an RTX for the address of the function's pure code.
7475 CXT is an RTX for the static chain value for the function.
7477 This takes 16 insns: 2 shifts & 2 ands (to split up addresses), 4 sethi
7478 (to load in opcodes), 4 iors (to merge address and opcodes), and 4 writes
7479 (to store insns). This is a bit excessive. Perhaps a different
7480 mechanism would be better here.
7482 Emit enough FLUSH insns to synchronize the data and instruction caches. */
7485 sparc_initialize_trampoline (rtx tramp, rtx fnaddr, rtx cxt)
7487 /* SPARC 32-bit trampoline:
7490 sethi %hi(static), %g2
7492 or %g2, %lo(static), %g2
7494 SETHI i,r = 00rr rrr1 00ii iiii iiii iiii iiii iiii
7495 JMPL r+i,d = 10dd ddd1 1100 0rrr rr1i iiii iiii iiii
7499 (gen_rtx_MEM (SImode, plus_constant (tramp, 0)),
7500 expand_binop (SImode, ior_optab,
7501 expand_shift (RSHIFT_EXPR, SImode, fnaddr,
7502 size_int (10), 0, 1),
7503 GEN_INT (trunc_int_for_mode (0x03000000, SImode)),
7504 NULL_RTX, 1, OPTAB_DIRECT));
7507 (gen_rtx_MEM (SImode, plus_constant (tramp, 4)),
7508 expand_binop (SImode, ior_optab,
7509 expand_shift (RSHIFT_EXPR, SImode, cxt,
7510 size_int (10), 0, 1),
7511 GEN_INT (trunc_int_for_mode (0x05000000, SImode)),
7512 NULL_RTX, 1, OPTAB_DIRECT));
7515 (gen_rtx_MEM (SImode, plus_constant (tramp, 8)),
7516 expand_binop (SImode, ior_optab,
7517 expand_and (SImode, fnaddr, GEN_INT (0x3ff), NULL_RTX),
7518 GEN_INT (trunc_int_for_mode (0x81c06000, SImode)),
7519 NULL_RTX, 1, OPTAB_DIRECT));
7522 (gen_rtx_MEM (SImode, plus_constant (tramp, 12)),
7523 expand_binop (SImode, ior_optab,
7524 expand_and (SImode, cxt, GEN_INT (0x3ff), NULL_RTX),
7525 GEN_INT (trunc_int_for_mode (0x8410a000, SImode)),
7526 NULL_RTX, 1, OPTAB_DIRECT));
7528 /* On UltraSPARC a flush flushes an entire cache line. The trampoline is
7529 aligned on a 16 byte boundary so one flush clears it all. */
7530 emit_insn (gen_flush (validize_mem (gen_rtx_MEM (SImode, tramp))));
7531 if (sparc_cpu != PROCESSOR_ULTRASPARC
7532 && sparc_cpu != PROCESSOR_ULTRASPARC3)
7533 emit_insn (gen_flush (validize_mem (gen_rtx_MEM (SImode,
7534 plus_constant (tramp, 8)))));
7536 /* Call __enable_execute_stack after writing onto the stack to make sure
7537 the stack address is accessible. */
7538 #ifdef ENABLE_EXECUTE_STACK
7539 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
7540 LCT_NORMAL, VOIDmode, 1, tramp, Pmode);
7545 /* The 64-bit version is simpler because it makes more sense to load the
7546 values as "immediate" data out of the trampoline. It's also easier since
7547 we can read the PC without clobbering a register. */
7550 sparc64_initialize_trampoline (rtx tramp, rtx fnaddr, rtx cxt)
7552 /* SPARC 64-bit trampoline:
7561 emit_move_insn (gen_rtx_MEM (SImode, tramp),
7562 GEN_INT (trunc_int_for_mode (0x83414000, SImode)));
7563 emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 4)),
7564 GEN_INT (trunc_int_for_mode (0xca586018, SImode)));
7565 emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 8)),
7566 GEN_INT (trunc_int_for_mode (0x81c14000, SImode)));
7567 emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 12)),
7568 GEN_INT (trunc_int_for_mode (0xca586010, SImode)));
7569 emit_move_insn (gen_rtx_MEM (DImode, plus_constant (tramp, 16)), cxt);
7570 emit_move_insn (gen_rtx_MEM (DImode, plus_constant (tramp, 24)), fnaddr);
7571 emit_insn (gen_flushdi (validize_mem (gen_rtx_MEM (DImode, tramp))));
7573 if (sparc_cpu != PROCESSOR_ULTRASPARC
7574 && sparc_cpu != PROCESSOR_ULTRASPARC3)
7575 emit_insn (gen_flushdi (validize_mem (gen_rtx_MEM (DImode, plus_constant (tramp, 8)))));
7577 /* Call __enable_execute_stack after writing onto the stack to make sure
7578 the stack address is accessible. */
7579 #ifdef ENABLE_EXECUTE_STACK
7580 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
7581 LCT_NORMAL, VOIDmode, 1, tramp, Pmode);
7585 /* Adjust the cost of a scheduling dependency. Return the new cost of
7586 a dependency LINK or INSN on DEP_INSN. COST is the current cost. */
7589 supersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
7591 enum attr_type insn_type;
7593 if (! recog_memoized (insn))
7596 insn_type = get_attr_type (insn);
7598 if (REG_NOTE_KIND (link) == 0)
7600 /* Data dependency; DEP_INSN writes a register that INSN reads some
7603 /* if a load, then the dependence must be on the memory address;
7604 add an extra "cycle". Note that the cost could be two cycles
7605 if the reg was written late in an instruction group; we ca not tell
7607 if (insn_type == TYPE_LOAD || insn_type == TYPE_FPLOAD)
7610 /* Get the delay only if the address of the store is the dependence. */
7611 if (insn_type == TYPE_STORE || insn_type == TYPE_FPSTORE)
7613 rtx pat = PATTERN(insn);
7614 rtx dep_pat = PATTERN (dep_insn);
7616 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
7617 return cost; /* This should not happen! */
7619 /* The dependency between the two instructions was on the data that
7620 is being stored. Assume that this implies that the address of the
7621 store is not dependent. */
7622 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
7625 return cost + 3; /* An approximation. */
7628 /* A shift instruction cannot receive its data from an instruction
7629 in the same cycle; add a one cycle penalty. */
7630 if (insn_type == TYPE_SHIFT)
7631 return cost + 3; /* Split before cascade into shift. */
7635 /* Anti- or output- dependency; DEP_INSN reads/writes a register that
7636 INSN writes some cycles later. */
7638 /* These are only significant for the fpu unit; writing a fp reg before
7639 the fpu has finished with it stalls the processor. */
7641 /* Reusing an integer register causes no problems. */
7642 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
7650 hypersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
7652 enum attr_type insn_type, dep_type;
7653 rtx pat = PATTERN(insn);
7654 rtx dep_pat = PATTERN (dep_insn);
7656 if (recog_memoized (insn) < 0 || recog_memoized (dep_insn) < 0)
7659 insn_type = get_attr_type (insn);
7660 dep_type = get_attr_type (dep_insn);
7662 switch (REG_NOTE_KIND (link))
7665 /* Data dependency; DEP_INSN writes a register that INSN reads some
7672 /* Get the delay iff the address of the store is the dependence. */
7673 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
7676 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
7683 /* If a load, then the dependence must be on the memory address. If
7684 the addresses aren't equal, then it might be a false dependency */
7685 if (dep_type == TYPE_STORE || dep_type == TYPE_FPSTORE)
7687 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET
7688 || GET_CODE (SET_DEST (dep_pat)) != MEM
7689 || GET_CODE (SET_SRC (pat)) != MEM
7690 || ! rtx_equal_p (XEXP (SET_DEST (dep_pat), 0),
7691 XEXP (SET_SRC (pat), 0)))
7699 /* Compare to branch latency is 0. There is no benefit from
7700 separating compare and branch. */
7701 if (dep_type == TYPE_COMPARE)
7703 /* Floating point compare to branch latency is less than
7704 compare to conditional move. */
7705 if (dep_type == TYPE_FPCMP)
7714 /* Anti-dependencies only penalize the fpu unit. */
7715 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
7727 sparc_adjust_cost(rtx insn, rtx link, rtx dep, int cost)
7731 case PROCESSOR_SUPERSPARC:
7732 cost = supersparc_adjust_cost (insn, link, dep, cost);
7734 case PROCESSOR_HYPERSPARC:
7735 case PROCESSOR_SPARCLITE86X:
7736 cost = hypersparc_adjust_cost (insn, link, dep, cost);
7745 sparc_sched_init (FILE *dump ATTRIBUTE_UNUSED,
7746 int sched_verbose ATTRIBUTE_UNUSED,
7747 int max_ready ATTRIBUTE_UNUSED)
7752 sparc_use_sched_lookahead (void)
7754 if (sparc_cpu == PROCESSOR_ULTRASPARC
7755 || sparc_cpu == PROCESSOR_ULTRASPARC3)
7757 if ((1 << sparc_cpu) &
7758 ((1 << PROCESSOR_SUPERSPARC) | (1 << PROCESSOR_HYPERSPARC) |
7759 (1 << PROCESSOR_SPARCLITE86X)))
7765 sparc_issue_rate (void)
7772 /* Assume V9 processors are capable of at least dual-issue. */
7774 case PROCESSOR_SUPERSPARC:
7776 case PROCESSOR_HYPERSPARC:
7777 case PROCESSOR_SPARCLITE86X:
7779 case PROCESSOR_ULTRASPARC:
7780 case PROCESSOR_ULTRASPARC3:
7786 set_extends (rtx insn)
7788 register rtx pat = PATTERN (insn);
7790 switch (GET_CODE (SET_SRC (pat)))
7792 /* Load and some shift instructions zero extend. */
7795 /* sethi clears the high bits */
7797 /* LO_SUM is used with sethi. sethi cleared the high
7798 bits and the values used with lo_sum are positive */
7800 /* Store flag stores 0 or 1 */
7810 rtx op0 = XEXP (SET_SRC (pat), 0);
7811 rtx op1 = XEXP (SET_SRC (pat), 1);
7812 if (GET_CODE (op1) == CONST_INT)
7813 return INTVAL (op1) >= 0;
7814 if (GET_CODE (op0) != REG)
7816 if (sparc_check_64 (op0, insn) == 1)
7818 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
7823 rtx op0 = XEXP (SET_SRC (pat), 0);
7824 rtx op1 = XEXP (SET_SRC (pat), 1);
7825 if (GET_CODE (op0) != REG || sparc_check_64 (op0, insn) <= 0)
7827 if (GET_CODE (op1) == CONST_INT)
7828 return INTVAL (op1) >= 0;
7829 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
7832 return GET_MODE (SET_SRC (pat)) == SImode;
7833 /* Positive integers leave the high bits zero. */
7835 return ! (CONST_DOUBLE_LOW (SET_SRC (pat)) & 0x80000000);
7837 return ! (INTVAL (SET_SRC (pat)) & 0x80000000);
7840 return - (GET_MODE (SET_SRC (pat)) == SImode);
7842 return sparc_check_64 (SET_SRC (pat), insn);
7848 /* We _ought_ to have only one kind per function, but... */
7849 static GTY(()) rtx sparc_addr_diff_list;
7850 static GTY(()) rtx sparc_addr_list;
7853 sparc_defer_case_vector (rtx lab, rtx vec, int diff)
7855 vec = gen_rtx_EXPR_LIST (VOIDmode, lab, vec);
7857 sparc_addr_diff_list
7858 = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_diff_list);
7860 sparc_addr_list = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_list);
7864 sparc_output_addr_vec (rtx vec)
7866 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
7867 int idx, vlen = XVECLEN (body, 0);
7869 #ifdef ASM_OUTPUT_ADDR_VEC_START
7870 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
7873 #ifdef ASM_OUTPUT_CASE_LABEL
7874 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
7877 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
7880 for (idx = 0; idx < vlen; idx++)
7882 ASM_OUTPUT_ADDR_VEC_ELT
7883 (asm_out_file, CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 0, idx), 0)));
7886 #ifdef ASM_OUTPUT_ADDR_VEC_END
7887 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
7892 sparc_output_addr_diff_vec (rtx vec)
7894 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
7895 rtx base = XEXP (XEXP (body, 0), 0);
7896 int idx, vlen = XVECLEN (body, 1);
7898 #ifdef ASM_OUTPUT_ADDR_VEC_START
7899 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
7902 #ifdef ASM_OUTPUT_CASE_LABEL
7903 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
7906 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
7909 for (idx = 0; idx < vlen; idx++)
7911 ASM_OUTPUT_ADDR_DIFF_ELT
7914 CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 1, idx), 0)),
7915 CODE_LABEL_NUMBER (base));
7918 #ifdef ASM_OUTPUT_ADDR_VEC_END
7919 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
7924 sparc_output_deferred_case_vectors (void)
7929 if (sparc_addr_list == NULL_RTX
7930 && sparc_addr_diff_list == NULL_RTX)
7933 /* Align to cache line in the function's code section. */
7934 function_section (current_function_decl);
7936 align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
7938 ASM_OUTPUT_ALIGN (asm_out_file, align);
7940 for (t = sparc_addr_list; t ; t = XEXP (t, 1))
7941 sparc_output_addr_vec (XEXP (t, 0));
7942 for (t = sparc_addr_diff_list; t ; t = XEXP (t, 1))
7943 sparc_output_addr_diff_vec (XEXP (t, 0));
7945 sparc_addr_list = sparc_addr_diff_list = NULL_RTX;
7948 /* Return 0 if the high 32 bits of X (the low word of X, if DImode) are
7949 unknown. Return 1 if the high bits are zero, -1 if the register is
7952 sparc_check_64 (rtx x, rtx insn)
7954 /* If a register is set only once it is safe to ignore insns this
7955 code does not know how to handle. The loop will either recognize
7956 the single set and return the correct value or fail to recognize
7961 if (GET_CODE (x) != REG)
7964 if (GET_MODE (x) == DImode)
7965 y = gen_rtx_REG (SImode, REGNO (x) + WORDS_BIG_ENDIAN);
7967 if (flag_expensive_optimizations
7968 && REG_N_SETS (REGNO (y)) == 1)
7974 insn = get_last_insn_anywhere ();
7979 while ((insn = PREV_INSN (insn)))
7981 switch (GET_CODE (insn))
7994 rtx pat = PATTERN (insn);
7995 if (GET_CODE (pat) != SET)
7997 if (rtx_equal_p (x, SET_DEST (pat)))
7998 return set_extends (insn);
7999 if (y && rtx_equal_p (y, SET_DEST (pat)))
8000 return set_extends (insn);
8001 if (reg_overlap_mentioned_p (SET_DEST (pat), y))
8009 /* Returns assembly code to perform a DImode shift using
8010 a 64-bit global or out register on SPARC-V8+. */
8012 output_v8plus_shift (rtx *operands, rtx insn, const char *opcode)
8014 static char asm_code[60];
8016 /* The scratch register is only required when the destination
8017 register is not a 64-bit global or out register. */
8018 if (which_alternative != 2)
8019 operands[3] = operands[0];
8021 /* We can only shift by constants <= 63. */
8022 if (GET_CODE (operands[2]) == CONST_INT)
8023 operands[2] = GEN_INT (INTVAL (operands[2]) & 0x3f);
8025 if (GET_CODE (operands[1]) == CONST_INT)
8027 output_asm_insn ("mov\t%1, %3", operands);
8031 output_asm_insn ("sllx\t%H1, 32, %3", operands);
8032 if (sparc_check_64 (operands[1], insn) <= 0)
8033 output_asm_insn ("srl\t%L1, 0, %L1", operands);
8034 output_asm_insn ("or\t%L1, %3, %3", operands);
8037 strcpy(asm_code, opcode);
8039 if (which_alternative != 2)
8040 return strcat (asm_code, "\t%0, %2, %L0\n\tsrlx\t%L0, 32, %H0");
8042 return strcat (asm_code, "\t%3, %2, %3\n\tsrlx\t%3, 32, %H0\n\tmov\t%3, %L0");
8045 /* Output rtl to increment the profiler label LABELNO
8046 for profiling a function entry. */
8049 sparc_profile_hook (int labelno)
8054 ASM_GENERATE_INTERNAL_LABEL (buf, "LP", labelno);
8055 lab = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (buf));
8056 fun = gen_rtx_SYMBOL_REF (Pmode, MCOUNT_FUNCTION);
8058 emit_library_call (fun, LCT_NORMAL, VOIDmode, 1, lab, Pmode);
8061 #ifdef OBJECT_FORMAT_ELF
8063 sparc_elf_asm_named_section (const char *name, unsigned int flags)
8065 if (flags & SECTION_MERGE)
8067 /* entsize cannot be expressed in this section attributes
8069 default_elf_asm_named_section (name, flags);
8073 fprintf (asm_out_file, "\t.section\t\"%s\"", name);
8075 if (!(flags & SECTION_DEBUG))
8076 fputs (",#alloc", asm_out_file);
8077 if (flags & SECTION_WRITE)
8078 fputs (",#write", asm_out_file);
8079 if (flags & SECTION_TLS)
8080 fputs (",#tls", asm_out_file);
8081 if (flags & SECTION_CODE)
8082 fputs (",#execinstr", asm_out_file);
8084 /* ??? Handle SECTION_BSS. */
8086 fputc ('\n', asm_out_file);
8088 #endif /* OBJECT_FORMAT_ELF */
8090 /* We do not allow indirect calls to be optimized into sibling calls.
8092 We cannot use sibling calls when delayed branches are disabled
8093 because they will likely require the call delay slot to be filled.
8095 Also, on SPARC 32-bit we cannot emit a sibling call when the
8096 current function returns a structure. This is because the "unimp
8097 after call" convention would cause the callee to return to the
8098 wrong place. The generic code already disallows cases where the
8099 function being called returns a structure.
8101 It may seem strange how this last case could occur. Usually there
8102 is code after the call which jumps to epilogue code which dumps the
8103 return value into the struct return area. That ought to invalidate
8104 the sibling call right? Well, in the C++ case we can end up passing
8105 the pointer to the struct return area to a constructor (which returns
8106 void) and then nothing else happens. Such a sibling call would look
8107 valid without the added check here. */
8109 sparc_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
8112 && flag_delayed_branch
8113 && (TARGET_ARCH64 || ! current_function_returns_struct));
8116 /* libfunc renaming. */
8117 #include "config/gofast.h"
8120 sparc_init_libfuncs (void)
8124 /* Use the subroutines that Sun's library provides for integer
8125 multiply and divide. The `*' prevents an underscore from
8126 being prepended by the compiler. .umul is a little faster
8128 set_optab_libfunc (smul_optab, SImode, "*.umul");
8129 set_optab_libfunc (sdiv_optab, SImode, "*.div");
8130 set_optab_libfunc (udiv_optab, SImode, "*.udiv");
8131 set_optab_libfunc (smod_optab, SImode, "*.rem");
8132 set_optab_libfunc (umod_optab, SImode, "*.urem");
8134 /* TFmode arithmetic. These names are part of the SPARC 32bit ABI. */
8135 set_optab_libfunc (add_optab, TFmode, "_Q_add");
8136 set_optab_libfunc (sub_optab, TFmode, "_Q_sub");
8137 set_optab_libfunc (neg_optab, TFmode, "_Q_neg");
8138 set_optab_libfunc (smul_optab, TFmode, "_Q_mul");
8139 set_optab_libfunc (sdiv_optab, TFmode, "_Q_div");
8141 /* We can define the TFmode sqrt optab only if TARGET_FPU. This
8142 is because with soft-float, the SFmode and DFmode sqrt
8143 instructions will be absent, and the compiler will notice and
8144 try to use the TFmode sqrt instruction for calls to the
8145 builtin function sqrt, but this fails. */
8147 set_optab_libfunc (sqrt_optab, TFmode, "_Q_sqrt");
8149 set_optab_libfunc (eq_optab, TFmode, "_Q_feq");
8150 set_optab_libfunc (ne_optab, TFmode, "_Q_fne");
8151 set_optab_libfunc (gt_optab, TFmode, "_Q_fgt");
8152 set_optab_libfunc (ge_optab, TFmode, "_Q_fge");
8153 set_optab_libfunc (lt_optab, TFmode, "_Q_flt");
8154 set_optab_libfunc (le_optab, TFmode, "_Q_fle");
8156 set_conv_libfunc (sext_optab, TFmode, SFmode, "_Q_stoq");
8157 set_conv_libfunc (sext_optab, TFmode, DFmode, "_Q_dtoq");
8158 set_conv_libfunc (trunc_optab, SFmode, TFmode, "_Q_qtos");
8159 set_conv_libfunc (trunc_optab, DFmode, TFmode, "_Q_qtod");
8161 set_conv_libfunc (sfix_optab, SImode, TFmode, "_Q_qtoi");
8162 set_conv_libfunc (ufix_optab, SImode, TFmode, "_Q_qtou");
8163 set_conv_libfunc (sfloat_optab, TFmode, SImode, "_Q_itoq");
8165 if (DITF_CONVERSION_LIBFUNCS)
8167 set_conv_libfunc (sfix_optab, DImode, TFmode, "_Q_qtoll");
8168 set_conv_libfunc (ufix_optab, DImode, TFmode, "_Q_qtoull");
8169 set_conv_libfunc (sfloat_optab, TFmode, DImode, "_Q_lltoq");
8172 if (SUN_CONVERSION_LIBFUNCS)
8174 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftoll");
8175 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoull");
8176 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtoll");
8177 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoull");
8182 /* In the SPARC 64bit ABI, SImode multiply and divide functions
8183 do not exist in the library. Make sure the compiler does not
8184 emit calls to them by accident. (It should always use the
8185 hardware instructions.) */
8186 set_optab_libfunc (smul_optab, SImode, 0);
8187 set_optab_libfunc (sdiv_optab, SImode, 0);
8188 set_optab_libfunc (udiv_optab, SImode, 0);
8189 set_optab_libfunc (smod_optab, SImode, 0);
8190 set_optab_libfunc (umod_optab, SImode, 0);
8192 if (SUN_INTEGER_MULTIPLY_64)
8194 set_optab_libfunc (smul_optab, DImode, "__mul64");
8195 set_optab_libfunc (sdiv_optab, DImode, "__div64");
8196 set_optab_libfunc (udiv_optab, DImode, "__udiv64");
8197 set_optab_libfunc (smod_optab, DImode, "__rem64");
8198 set_optab_libfunc (umod_optab, DImode, "__urem64");
8201 if (SUN_CONVERSION_LIBFUNCS)
8203 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftol");
8204 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoul");
8205 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtol");
8206 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoul");
8210 gofast_maybe_init_libfuncs ();
8214 sparc_extra_constraint_check (rtx op, int c, int strict)
8219 && (c == 'T' || c == 'U'))
8225 return fp_sethi_p (op);
8228 return fp_mov_p (op);
8231 return fp_high_losum_p (op);
8235 || (GET_CODE (op) == REG
8236 && (REGNO (op) < FIRST_PSEUDO_REGISTER
8237 || reg_renumber[REGNO (op)] >= 0)))
8238 return register_ok_for_ldd (op);
8250 /* Our memory extra constraints have to emulate the
8251 behavior of 'm' and 'o' in order for reload to work
8253 if (GET_CODE (op) == MEM)
8256 if ((TARGET_ARCH64 || mem_min_alignment (op, 8))
8258 || strict_memory_address_p (Pmode, XEXP (op, 0))))
8263 reload_ok_mem = (reload_in_progress
8264 && GET_CODE (op) == REG
8265 && REGNO (op) >= FIRST_PSEUDO_REGISTER
8266 && reg_renumber [REGNO (op)] < 0);
8269 return reload_ok_mem;
8272 /* ??? This duplicates information provided to the compiler by the
8273 ??? scheduler description. Some day, teach genautomata to output
8274 ??? the latencies and then CSE will just use that. */
8277 sparc_rtx_costs (rtx x, int code, int outer_code, int *total)
8279 enum machine_mode mode = GET_MODE (x);
8280 bool float_mode_p = FLOAT_MODE_P (mode);
8285 if (INTVAL (x) < 0x1000 && INTVAL (x) >= -0x1000)
8303 if (GET_MODE (x) == DImode
8304 && ((XINT (x, 3) == 0
8305 && (unsigned HOST_WIDE_INT) XINT (x, 2) < 0x1000)
8306 || (XINT (x, 3) == -1
8308 && XINT (x, 2) >= -0x1000)))
8315 /* If outer-code was a sign or zero extension, a cost
8316 of COSTS_N_INSNS (1) was already added in. This is
8317 why we are subtracting it back out. */
8318 if (outer_code == ZERO_EXTEND)
8320 *total = sparc_costs->int_zload - COSTS_N_INSNS (1);
8322 else if (outer_code == SIGN_EXTEND)
8324 *total = sparc_costs->int_sload - COSTS_N_INSNS (1);
8326 else if (float_mode_p)
8328 *total = sparc_costs->float_load;
8332 *total = sparc_costs->int_load;
8340 *total = sparc_costs->float_plusminus;
8342 *total = COSTS_N_INSNS (1);
8347 *total = sparc_costs->float_mul;
8348 else if (! TARGET_HARD_MUL)
8349 *total = COSTS_N_INSNS (25);
8355 if (sparc_costs->int_mul_bit_factor)
8359 if (GET_CODE (XEXP (x, 1)) == CONST_INT)
8361 unsigned HOST_WIDE_INT value = INTVAL (XEXP (x, 1));
8362 for (nbits = 0; value != 0; value &= value - 1)
8365 else if (GET_CODE (XEXP (x, 1)) == CONST_DOUBLE
8366 && GET_MODE (XEXP (x, 1)) == DImode)
8368 rtx x1 = XEXP (x, 1);
8369 unsigned HOST_WIDE_INT value1 = XINT (x1, 2);
8370 unsigned HOST_WIDE_INT value2 = XINT (x1, 3);
8372 for (nbits = 0; value1 != 0; value1 &= value1 - 1)
8374 for (; value2 != 0; value2 &= value2 - 1)
8382 bit_cost = (nbits - 3) / sparc_costs->int_mul_bit_factor;
8383 bit_cost = COSTS_N_INSNS (bit_cost);
8387 *total = sparc_costs->int_mulX + bit_cost;
8389 *total = sparc_costs->int_mul + bit_cost;
8396 *total = COSTS_N_INSNS (1) + sparc_costs->shift_penalty;
8406 *total = sparc_costs->float_div_df;
8408 *total = sparc_costs->float_div_sf;
8413 *total = sparc_costs->int_divX;
8415 *total = sparc_costs->int_div;
8422 *total = COSTS_N_INSNS (1);
8429 case UNSIGNED_FLOAT:
8433 case FLOAT_TRUNCATE:
8434 *total = sparc_costs->float_move;
8439 *total = sparc_costs->float_sqrt_df;
8441 *total = sparc_costs->float_sqrt_sf;
8446 *total = sparc_costs->float_cmp;
8448 *total = COSTS_N_INSNS (1);
8453 *total = sparc_costs->float_cmove;
8455 *total = sparc_costs->int_cmove;
8463 /* Emit the sequence of insns SEQ while preserving the register REG. */
8466 emit_and_preserve (rtx seq, rtx reg)
8468 rtx slot = gen_rtx_MEM (word_mode,
8469 plus_constant (stack_pointer_rtx, SPARC_STACK_BIAS));
8471 emit_stack_pointer_decrement (GEN_INT (UNITS_PER_WORD));
8472 emit_insn (gen_rtx_SET (VOIDmode, slot, reg));
8474 emit_insn (gen_rtx_SET (VOIDmode, reg, slot));
8475 emit_stack_pointer_increment (GEN_INT (UNITS_PER_WORD));
8478 /* Output code to add DELTA to the first argument, and then jump to FUNCTION.
8479 Used for C++ multiple inheritance. */
8482 sparc_output_mi_thunk (FILE *file, tree thunk_fndecl ATTRIBUTE_UNUSED,
8483 HOST_WIDE_INT delta,
8484 HOST_WIDE_INT vcall_offset ATTRIBUTE_UNUSED,
8487 rtx this, insn, funexp, delta_rtx;
8488 unsigned int int_arg_first;
8490 reload_completed = 1;
8491 epilogue_completed = 1;
8493 reset_block_changes ();
8495 emit_note (NOTE_INSN_PROLOGUE_END);
8497 if (flag_delayed_branch)
8499 /* We will emit a regular sibcall below, so we need to instruct
8500 output_sibcall that we are in a leaf function. */
8501 current_function_uses_only_leaf_regs = 1;
8503 /* This will cause final.c to invoke leaf_renumber_regs so we
8504 must behave as if we were in a not-yet-leafified function. */
8505 int_arg_first = SPARC_INCOMING_INT_ARG_FIRST;
8509 /* We will emit the sibcall manually below, so we will need to
8510 manually spill non-leaf registers. */
8511 current_function_uses_only_leaf_regs = 0;
8513 /* We really are in a leaf function. */
8514 int_arg_first = SPARC_OUTGOING_INT_ARG_FIRST;
8517 /* Find the "this" pointer. Normally in %o0, but in ARCH64 if the function
8518 returns a structure, the structure return pointer is there instead. */
8519 if (TARGET_ARCH64 && aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function))
8520 this = gen_rtx_REG (Pmode, int_arg_first + 1);
8522 this = gen_rtx_REG (Pmode, int_arg_first);
8524 /* Add DELTA. When possible use a plain add, otherwise load it into
8525 a register first. */
8526 delta_rtx = GEN_INT (delta);
8527 if (!SPARC_SIMM13_P (delta))
8529 rtx scratch = gen_rtx_REG (Pmode, 1);
8531 if (input_operand (delta_rtx, GET_MODE (scratch)))
8532 emit_insn (gen_rtx_SET (VOIDmode, scratch, delta_rtx));
8536 sparc_emit_set_const64 (scratch, delta_rtx);
8538 sparc_emit_set_const32 (scratch, delta_rtx);
8541 delta_rtx = scratch;
8544 emit_insn (gen_rtx_SET (VOIDmode,
8546 gen_rtx_PLUS (Pmode, this, delta_rtx)));
8548 /* Generate a tail call to the target function. */
8549 if (! TREE_USED (function))
8551 assemble_external (function);
8552 TREE_USED (function) = 1;
8554 funexp = XEXP (DECL_RTL (function), 0);
8556 if (flag_delayed_branch)
8558 funexp = gen_rtx_MEM (FUNCTION_MODE, funexp);
8559 insn = emit_call_insn (gen_sibcall (funexp));
8560 SIBLING_CALL_P (insn) = 1;
8564 /* The hoops we have to jump through in order to generate a sibcall
8565 without using delay slots... */
8566 rtx spill_reg, seq, scratch = gen_rtx_REG (Pmode, 1);
8570 spill_reg = gen_rtx_REG (word_mode, 15); /* %o7 */
8572 load_pic_register (); /* clobbers %o7 */
8573 scratch = legitimize_pic_address (funexp, Pmode, scratch);
8576 emit_and_preserve (seq, spill_reg);
8578 else if (TARGET_ARCH32)
8580 emit_insn (gen_rtx_SET (VOIDmode,
8582 gen_rtx_HIGH (SImode, funexp)));
8583 emit_insn (gen_rtx_SET (VOIDmode,
8585 gen_rtx_LO_SUM (SImode, scratch, funexp)));
8587 else /* TARGET_ARCH64 */
8589 switch (sparc_cmodel)
8593 /* The destination can serve as a temporary. */
8594 sparc_emit_set_symbolic_const64 (scratch, funexp, scratch);
8599 /* The destination cannot serve as a temporary. */
8600 spill_reg = gen_rtx_REG (DImode, 15); /* %o7 */
8602 sparc_emit_set_symbolic_const64 (scratch, funexp, spill_reg);
8605 emit_and_preserve (seq, spill_reg);
8613 emit_jump_insn (gen_indirect_jump (scratch));
8618 /* Run just enough of rest_of_compilation to get the insns emitted.
8619 There's not really enough bulk here to make other passes such as
8620 instruction scheduling worth while. Note that use_thunk calls
8621 assemble_start_function and assemble_end_function. */
8622 insn = get_insns ();
8623 insn_locators_initialize ();
8624 shorten_branches (insn);
8625 final_start_function (insn, file, 1);
8626 final (insn, file, 1, 0);
8627 final_end_function ();
8629 reload_completed = 0;
8630 epilogue_completed = 0;
8634 /* How to allocate a 'struct machine_function'. */
8636 static struct machine_function *
8637 sparc_init_machine_status (void)
8639 return ggc_alloc_cleared (sizeof (struct machine_function));
8642 /* Locate some local-dynamic symbol still in use by this function
8643 so that we can print its name in local-dynamic base patterns. */
8646 get_some_local_dynamic_name (void)
8650 if (cfun->machine->some_ld_name)
8651 return cfun->machine->some_ld_name;
8653 for (insn = get_insns (); insn ; insn = NEXT_INSN (insn))
8655 && for_each_rtx (&PATTERN (insn), get_some_local_dynamic_name_1, 0))
8656 return cfun->machine->some_ld_name;
8662 get_some_local_dynamic_name_1 (rtx *px, void *data ATTRIBUTE_UNUSED)
8667 && GET_CODE (x) == SYMBOL_REF
8668 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_DYNAMIC)
8670 cfun->machine->some_ld_name = XSTR (x, 0);
8677 /* This is called from dwarf2out.c via ASM_OUTPUT_DWARF_DTPREL.
8678 We need to emit DTP-relative relocations. */
8681 sparc_output_dwarf_dtprel (FILE *file, int size, rtx x)
8686 fputs ("\t.word\t%r_tls_dtpoff32(", file);
8689 fputs ("\t.xword\t%r_tls_dtpoff64(", file);
8694 output_addr_const (file, x);
8698 #include "gt-sparc.h"