1 /* Subroutines for insn-output.c for SPARC.
2 Copyright (C) 1987-2016 Free Software Foundation, Inc.
3 Contributed by Michael Tiemann (tiemann@cygnus.com)
4 64-bit SPARC-V9 support by Michael Tiemann, Jim Wilson, and Doug Evans,
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify
10 it under the terms of the GNU General Public License as published by
11 the Free Software Foundation; either version 3, or (at your option)
14 GCC is distributed in the hope that it will be useful,
15 but WITHOUT ANY WARRANTY; without even the implied warranty of
16 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 GNU General Public License for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
25 #include "coretypes.h"
33 #include "stringpool.h"
39 #include "diagnostic-core.h"
41 #include "fold-const.h"
42 #include "stor-layout.h"
46 #include "insn-attr.h"
50 #include "common/common-target.h"
52 #include "langhooks.h"
55 #include "tree-pass.h"
59 /* This file should be included last. */
60 #include "target-def.h"
64 struct processor_costs
{
68 /* Integer signed load */
71 /* Integer zeroed load */
77 /* fmov, fneg, fabs */
81 const int float_plusminus
;
87 const int float_cmove
;
93 const int float_div_sf
;
96 const int float_div_df
;
99 const int float_sqrt_sf
;
102 const int float_sqrt_df
;
110 /* integer multiply cost for each bit set past the most
111 significant 3, so the formula for multiply cost becomes:
114 highest_bit = highest_clear_bit(rs1);
116 highest_bit = highest_set_bit(rs1);
119 cost = int_mul{,X} + ((highest_bit - 3) / int_mul_bit_factor);
121 A value of zero indicates that the multiply costs is fixed,
123 const int int_mul_bit_factor
;
134 /* penalty for shifts, due to scheduling rules etc. */
135 const int shift_penalty
;
139 struct processor_costs cypress_costs
= {
140 COSTS_N_INSNS (2), /* int load */
141 COSTS_N_INSNS (2), /* int signed load */
142 COSTS_N_INSNS (2), /* int zeroed load */
143 COSTS_N_INSNS (2), /* float load */
144 COSTS_N_INSNS (5), /* fmov, fneg, fabs */
145 COSTS_N_INSNS (5), /* fadd, fsub */
146 COSTS_N_INSNS (1), /* fcmp */
147 COSTS_N_INSNS (1), /* fmov, fmovr */
148 COSTS_N_INSNS (7), /* fmul */
149 COSTS_N_INSNS (37), /* fdivs */
150 COSTS_N_INSNS (37), /* fdivd */
151 COSTS_N_INSNS (63), /* fsqrts */
152 COSTS_N_INSNS (63), /* fsqrtd */
153 COSTS_N_INSNS (1), /* imul */
154 COSTS_N_INSNS (1), /* imulX */
155 0, /* imul bit factor */
156 COSTS_N_INSNS (1), /* idiv */
157 COSTS_N_INSNS (1), /* idivX */
158 COSTS_N_INSNS (1), /* movcc/movr */
159 0, /* shift penalty */
163 struct processor_costs supersparc_costs
= {
164 COSTS_N_INSNS (1), /* int load */
165 COSTS_N_INSNS (1), /* int signed load */
166 COSTS_N_INSNS (1), /* int zeroed load */
167 COSTS_N_INSNS (0), /* float load */
168 COSTS_N_INSNS (3), /* fmov, fneg, fabs */
169 COSTS_N_INSNS (3), /* fadd, fsub */
170 COSTS_N_INSNS (3), /* fcmp */
171 COSTS_N_INSNS (1), /* fmov, fmovr */
172 COSTS_N_INSNS (3), /* fmul */
173 COSTS_N_INSNS (6), /* fdivs */
174 COSTS_N_INSNS (9), /* fdivd */
175 COSTS_N_INSNS (12), /* fsqrts */
176 COSTS_N_INSNS (12), /* fsqrtd */
177 COSTS_N_INSNS (4), /* imul */
178 COSTS_N_INSNS (4), /* imulX */
179 0, /* imul bit factor */
180 COSTS_N_INSNS (4), /* idiv */
181 COSTS_N_INSNS (4), /* idivX */
182 COSTS_N_INSNS (1), /* movcc/movr */
183 1, /* shift penalty */
187 struct processor_costs hypersparc_costs
= {
188 COSTS_N_INSNS (1), /* int load */
189 COSTS_N_INSNS (1), /* int signed load */
190 COSTS_N_INSNS (1), /* int zeroed load */
191 COSTS_N_INSNS (1), /* float load */
192 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
193 COSTS_N_INSNS (1), /* fadd, fsub */
194 COSTS_N_INSNS (1), /* fcmp */
195 COSTS_N_INSNS (1), /* fmov, fmovr */
196 COSTS_N_INSNS (1), /* fmul */
197 COSTS_N_INSNS (8), /* fdivs */
198 COSTS_N_INSNS (12), /* fdivd */
199 COSTS_N_INSNS (17), /* fsqrts */
200 COSTS_N_INSNS (17), /* fsqrtd */
201 COSTS_N_INSNS (17), /* imul */
202 COSTS_N_INSNS (17), /* imulX */
203 0, /* imul bit factor */
204 COSTS_N_INSNS (17), /* idiv */
205 COSTS_N_INSNS (17), /* idivX */
206 COSTS_N_INSNS (1), /* movcc/movr */
207 0, /* shift penalty */
211 struct processor_costs leon_costs
= {
212 COSTS_N_INSNS (1), /* int load */
213 COSTS_N_INSNS (1), /* int signed load */
214 COSTS_N_INSNS (1), /* int zeroed load */
215 COSTS_N_INSNS (1), /* float load */
216 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
217 COSTS_N_INSNS (1), /* fadd, fsub */
218 COSTS_N_INSNS (1), /* fcmp */
219 COSTS_N_INSNS (1), /* fmov, fmovr */
220 COSTS_N_INSNS (1), /* fmul */
221 COSTS_N_INSNS (15), /* fdivs */
222 COSTS_N_INSNS (15), /* fdivd */
223 COSTS_N_INSNS (23), /* fsqrts */
224 COSTS_N_INSNS (23), /* fsqrtd */
225 COSTS_N_INSNS (5), /* imul */
226 COSTS_N_INSNS (5), /* imulX */
227 0, /* imul bit factor */
228 COSTS_N_INSNS (5), /* idiv */
229 COSTS_N_INSNS (5), /* idivX */
230 COSTS_N_INSNS (1), /* movcc/movr */
231 0, /* shift penalty */
235 struct processor_costs leon3_costs
= {
236 COSTS_N_INSNS (1), /* int load */
237 COSTS_N_INSNS (1), /* int signed load */
238 COSTS_N_INSNS (1), /* int zeroed load */
239 COSTS_N_INSNS (1), /* float load */
240 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
241 COSTS_N_INSNS (1), /* fadd, fsub */
242 COSTS_N_INSNS (1), /* fcmp */
243 COSTS_N_INSNS (1), /* fmov, fmovr */
244 COSTS_N_INSNS (1), /* fmul */
245 COSTS_N_INSNS (14), /* fdivs */
246 COSTS_N_INSNS (15), /* fdivd */
247 COSTS_N_INSNS (22), /* fsqrts */
248 COSTS_N_INSNS (23), /* fsqrtd */
249 COSTS_N_INSNS (5), /* imul */
250 COSTS_N_INSNS (5), /* imulX */
251 0, /* imul bit factor */
252 COSTS_N_INSNS (35), /* idiv */
253 COSTS_N_INSNS (35), /* idivX */
254 COSTS_N_INSNS (1), /* movcc/movr */
255 0, /* shift penalty */
259 struct processor_costs sparclet_costs
= {
260 COSTS_N_INSNS (3), /* int load */
261 COSTS_N_INSNS (3), /* int signed load */
262 COSTS_N_INSNS (1), /* int zeroed load */
263 COSTS_N_INSNS (1), /* float load */
264 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
265 COSTS_N_INSNS (1), /* fadd, fsub */
266 COSTS_N_INSNS (1), /* fcmp */
267 COSTS_N_INSNS (1), /* fmov, fmovr */
268 COSTS_N_INSNS (1), /* fmul */
269 COSTS_N_INSNS (1), /* fdivs */
270 COSTS_N_INSNS (1), /* fdivd */
271 COSTS_N_INSNS (1), /* fsqrts */
272 COSTS_N_INSNS (1), /* fsqrtd */
273 COSTS_N_INSNS (5), /* imul */
274 COSTS_N_INSNS (5), /* imulX */
275 0, /* imul bit factor */
276 COSTS_N_INSNS (5), /* idiv */
277 COSTS_N_INSNS (5), /* idivX */
278 COSTS_N_INSNS (1), /* movcc/movr */
279 0, /* shift penalty */
283 struct processor_costs ultrasparc_costs
= {
284 COSTS_N_INSNS (2), /* int load */
285 COSTS_N_INSNS (3), /* int signed load */
286 COSTS_N_INSNS (2), /* int zeroed load */
287 COSTS_N_INSNS (2), /* float load */
288 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
289 COSTS_N_INSNS (4), /* fadd, fsub */
290 COSTS_N_INSNS (1), /* fcmp */
291 COSTS_N_INSNS (2), /* fmov, fmovr */
292 COSTS_N_INSNS (4), /* fmul */
293 COSTS_N_INSNS (13), /* fdivs */
294 COSTS_N_INSNS (23), /* fdivd */
295 COSTS_N_INSNS (13), /* fsqrts */
296 COSTS_N_INSNS (23), /* fsqrtd */
297 COSTS_N_INSNS (4), /* imul */
298 COSTS_N_INSNS (4), /* imulX */
299 2, /* imul bit factor */
300 COSTS_N_INSNS (37), /* idiv */
301 COSTS_N_INSNS (68), /* idivX */
302 COSTS_N_INSNS (2), /* movcc/movr */
303 2, /* shift penalty */
307 struct processor_costs ultrasparc3_costs
= {
308 COSTS_N_INSNS (2), /* int load */
309 COSTS_N_INSNS (3), /* int signed load */
310 COSTS_N_INSNS (3), /* int zeroed load */
311 COSTS_N_INSNS (2), /* float load */
312 COSTS_N_INSNS (3), /* fmov, fneg, fabs */
313 COSTS_N_INSNS (4), /* fadd, fsub */
314 COSTS_N_INSNS (5), /* fcmp */
315 COSTS_N_INSNS (3), /* fmov, fmovr */
316 COSTS_N_INSNS (4), /* fmul */
317 COSTS_N_INSNS (17), /* fdivs */
318 COSTS_N_INSNS (20), /* fdivd */
319 COSTS_N_INSNS (20), /* fsqrts */
320 COSTS_N_INSNS (29), /* fsqrtd */
321 COSTS_N_INSNS (6), /* imul */
322 COSTS_N_INSNS (6), /* imulX */
323 0, /* imul bit factor */
324 COSTS_N_INSNS (40), /* idiv */
325 COSTS_N_INSNS (71), /* idivX */
326 COSTS_N_INSNS (2), /* movcc/movr */
327 0, /* shift penalty */
331 struct processor_costs niagara_costs
= {
332 COSTS_N_INSNS (3), /* int load */
333 COSTS_N_INSNS (3), /* int signed load */
334 COSTS_N_INSNS (3), /* int zeroed load */
335 COSTS_N_INSNS (9), /* float load */
336 COSTS_N_INSNS (8), /* fmov, fneg, fabs */
337 COSTS_N_INSNS (8), /* fadd, fsub */
338 COSTS_N_INSNS (26), /* fcmp */
339 COSTS_N_INSNS (8), /* fmov, fmovr */
340 COSTS_N_INSNS (29), /* fmul */
341 COSTS_N_INSNS (54), /* fdivs */
342 COSTS_N_INSNS (83), /* fdivd */
343 COSTS_N_INSNS (100), /* fsqrts - not implemented in hardware */
344 COSTS_N_INSNS (100), /* fsqrtd - not implemented in hardware */
345 COSTS_N_INSNS (11), /* imul */
346 COSTS_N_INSNS (11), /* imulX */
347 0, /* imul bit factor */
348 COSTS_N_INSNS (72), /* idiv */
349 COSTS_N_INSNS (72), /* idivX */
350 COSTS_N_INSNS (1), /* movcc/movr */
351 0, /* shift penalty */
355 struct processor_costs niagara2_costs
= {
356 COSTS_N_INSNS (3), /* int load */
357 COSTS_N_INSNS (3), /* int signed load */
358 COSTS_N_INSNS (3), /* int zeroed load */
359 COSTS_N_INSNS (3), /* float load */
360 COSTS_N_INSNS (6), /* fmov, fneg, fabs */
361 COSTS_N_INSNS (6), /* fadd, fsub */
362 COSTS_N_INSNS (6), /* fcmp */
363 COSTS_N_INSNS (6), /* fmov, fmovr */
364 COSTS_N_INSNS (6), /* fmul */
365 COSTS_N_INSNS (19), /* fdivs */
366 COSTS_N_INSNS (33), /* fdivd */
367 COSTS_N_INSNS (19), /* fsqrts */
368 COSTS_N_INSNS (33), /* fsqrtd */
369 COSTS_N_INSNS (5), /* imul */
370 COSTS_N_INSNS (5), /* imulX */
371 0, /* imul bit factor */
372 COSTS_N_INSNS (26), /* idiv, average of 12 - 41 cycle range */
373 COSTS_N_INSNS (26), /* idivX, average of 12 - 41 cycle range */
374 COSTS_N_INSNS (1), /* movcc/movr */
375 0, /* shift penalty */
379 struct processor_costs niagara3_costs
= {
380 COSTS_N_INSNS (3), /* int load */
381 COSTS_N_INSNS (3), /* int signed load */
382 COSTS_N_INSNS (3), /* int zeroed load */
383 COSTS_N_INSNS (3), /* float load */
384 COSTS_N_INSNS (9), /* fmov, fneg, fabs */
385 COSTS_N_INSNS (9), /* fadd, fsub */
386 COSTS_N_INSNS (9), /* fcmp */
387 COSTS_N_INSNS (9), /* fmov, fmovr */
388 COSTS_N_INSNS (9), /* fmul */
389 COSTS_N_INSNS (23), /* fdivs */
390 COSTS_N_INSNS (37), /* fdivd */
391 COSTS_N_INSNS (23), /* fsqrts */
392 COSTS_N_INSNS (37), /* fsqrtd */
393 COSTS_N_INSNS (9), /* imul */
394 COSTS_N_INSNS (9), /* imulX */
395 0, /* imul bit factor */
396 COSTS_N_INSNS (31), /* idiv, average of 17 - 45 cycle range */
397 COSTS_N_INSNS (30), /* idivX, average of 16 - 44 cycle range */
398 COSTS_N_INSNS (1), /* movcc/movr */
399 0, /* shift penalty */
403 struct processor_costs niagara4_costs
= {
404 COSTS_N_INSNS (5), /* int load */
405 COSTS_N_INSNS (5), /* int signed load */
406 COSTS_N_INSNS (5), /* int zeroed load */
407 COSTS_N_INSNS (5), /* float load */
408 COSTS_N_INSNS (11), /* fmov, fneg, fabs */
409 COSTS_N_INSNS (11), /* fadd, fsub */
410 COSTS_N_INSNS (11), /* fcmp */
411 COSTS_N_INSNS (11), /* fmov, fmovr */
412 COSTS_N_INSNS (11), /* fmul */
413 COSTS_N_INSNS (24), /* fdivs */
414 COSTS_N_INSNS (37), /* fdivd */
415 COSTS_N_INSNS (24), /* fsqrts */
416 COSTS_N_INSNS (37), /* fsqrtd */
417 COSTS_N_INSNS (12), /* imul */
418 COSTS_N_INSNS (12), /* imulX */
419 0, /* imul bit factor */
420 COSTS_N_INSNS (50), /* idiv, average of 41 - 60 cycle range */
421 COSTS_N_INSNS (35), /* idivX, average of 26 - 44 cycle range */
422 COSTS_N_INSNS (1), /* movcc/movr */
423 0, /* shift penalty */
427 struct processor_costs niagara7_costs
= {
428 COSTS_N_INSNS (5), /* int load */
429 COSTS_N_INSNS (5), /* int signed load */
430 COSTS_N_INSNS (5), /* int zeroed load */
431 COSTS_N_INSNS (5), /* float load */
432 COSTS_N_INSNS (11), /* fmov, fneg, fabs */
433 COSTS_N_INSNS (11), /* fadd, fsub */
434 COSTS_N_INSNS (11), /* fcmp */
435 COSTS_N_INSNS (11), /* fmov, fmovr */
436 COSTS_N_INSNS (11), /* fmul */
437 COSTS_N_INSNS (24), /* fdivs */
438 COSTS_N_INSNS (37), /* fdivd */
439 COSTS_N_INSNS (24), /* fsqrts */
440 COSTS_N_INSNS (37), /* fsqrtd */
441 COSTS_N_INSNS (12), /* imul */
442 COSTS_N_INSNS (12), /* imulX */
443 0, /* imul bit factor */
444 COSTS_N_INSNS (51), /* idiv, average of 42 - 61 cycle range */
445 COSTS_N_INSNS (35), /* idivX, average of 26 - 44 cycle range */
446 COSTS_N_INSNS (1), /* movcc/movr */
447 0, /* shift penalty */
450 static const struct processor_costs
*sparc_costs
= &cypress_costs
;
452 #ifdef HAVE_AS_RELAX_OPTION
453 /* If 'as' and 'ld' are relaxing tail call insns into branch always, use
454 "or %o7,%g0,X; call Y; or X,%g0,%o7" always, so that it can be optimized.
455 With sethi/jmp, neither 'as' nor 'ld' has an easy way how to find out if
456 somebody does not branch between the sethi and jmp. */
457 #define LEAF_SIBCALL_SLOT_RESERVED_P 1
459 #define LEAF_SIBCALL_SLOT_RESERVED_P \
460 ((TARGET_ARCH64 && !TARGET_CM_MEDLOW) || flag_pic)
463 /* Vector to say how input registers are mapped to output registers.
464 HARD_FRAME_POINTER_REGNUM cannot be remapped by this function to
465 eliminate it. You must use -fomit-frame-pointer to get that. */
466 char leaf_reg_remap
[] =
467 { 0, 1, 2, 3, 4, 5, 6, 7,
468 -1, -1, -1, -1, -1, -1, 14, -1,
469 -1, -1, -1, -1, -1, -1, -1, -1,
470 8, 9, 10, 11, 12, 13, -1, 15,
472 32, 33, 34, 35, 36, 37, 38, 39,
473 40, 41, 42, 43, 44, 45, 46, 47,
474 48, 49, 50, 51, 52, 53, 54, 55,
475 56, 57, 58, 59, 60, 61, 62, 63,
476 64, 65, 66, 67, 68, 69, 70, 71,
477 72, 73, 74, 75, 76, 77, 78, 79,
478 80, 81, 82, 83, 84, 85, 86, 87,
479 88, 89, 90, 91, 92, 93, 94, 95,
480 96, 97, 98, 99, 100, 101, 102};
482 /* Vector, indexed by hard register number, which contains 1
483 for a register that is allowable in a candidate for leaf
484 function treatment. */
485 char sparc_leaf_regs
[] =
486 { 1, 1, 1, 1, 1, 1, 1, 1,
487 0, 0, 0, 0, 0, 0, 1, 0,
488 0, 0, 0, 0, 0, 0, 0, 0,
489 1, 1, 1, 1, 1, 1, 0, 1,
490 1, 1, 1, 1, 1, 1, 1, 1,
491 1, 1, 1, 1, 1, 1, 1, 1,
492 1, 1, 1, 1, 1, 1, 1, 1,
493 1, 1, 1, 1, 1, 1, 1, 1,
494 1, 1, 1, 1, 1, 1, 1, 1,
495 1, 1, 1, 1, 1, 1, 1, 1,
496 1, 1, 1, 1, 1, 1, 1, 1,
497 1, 1, 1, 1, 1, 1, 1, 1,
498 1, 1, 1, 1, 1, 1, 1};
500 struct GTY(()) machine_function
502 /* Size of the frame of the function. */
503 HOST_WIDE_INT frame_size
;
505 /* Size of the frame of the function minus the register window save area
506 and the outgoing argument area. */
507 HOST_WIDE_INT apparent_frame_size
;
509 /* Register we pretend the frame pointer is allocated to. Normally, this
510 is %fp, but if we are in a leaf procedure, this is (%sp + offset). We
511 record "offset" separately as it may be too big for (reg + disp). */
513 HOST_WIDE_INT frame_base_offset
;
515 /* Number of global or FP registers to be saved (as 4-byte quantities). */
516 int n_global_fp_regs
;
518 /* True if the current function is leaf and uses only leaf regs,
519 so that the SPARC leaf function optimization can be applied.
520 Private version of crtl->uses_only_leaf_regs, see
521 sparc_expand_prologue for the rationale. */
524 /* True if the prologue saves local or in registers. */
525 bool save_local_in_regs_p
;
527 /* True if the data calculated by sparc_expand_prologue are valid. */
528 bool prologue_data_valid_p
;
531 #define sparc_frame_size cfun->machine->frame_size
532 #define sparc_apparent_frame_size cfun->machine->apparent_frame_size
533 #define sparc_frame_base_reg cfun->machine->frame_base_reg
534 #define sparc_frame_base_offset cfun->machine->frame_base_offset
535 #define sparc_n_global_fp_regs cfun->machine->n_global_fp_regs
536 #define sparc_leaf_function_p cfun->machine->leaf_function_p
537 #define sparc_save_local_in_regs_p cfun->machine->save_local_in_regs_p
538 #define sparc_prologue_data_valid_p cfun->machine->prologue_data_valid_p
540 /* 1 if the next opcode is to be specially indented. */
541 int sparc_indent_opcode
= 0;
543 static void sparc_option_override (void);
544 static void sparc_init_modes (void);
545 static int function_arg_slotno (const CUMULATIVE_ARGS
*, machine_mode
,
546 const_tree
, bool, bool, int *, int *);
548 static int supersparc_adjust_cost (rtx_insn
*, rtx
, rtx_insn
*, int);
549 static int hypersparc_adjust_cost (rtx_insn
*, rtx
, rtx_insn
*, int);
551 static void sparc_emit_set_const32 (rtx
, rtx
);
552 static void sparc_emit_set_const64 (rtx
, rtx
);
553 static void sparc_output_addr_vec (rtx
);
554 static void sparc_output_addr_diff_vec (rtx
);
555 static void sparc_output_deferred_case_vectors (void);
556 static bool sparc_legitimate_address_p (machine_mode
, rtx
, bool);
557 static bool sparc_legitimate_constant_p (machine_mode
, rtx
);
558 static rtx
sparc_builtin_saveregs (void);
559 static int epilogue_renumber (rtx
*, int);
560 static bool sparc_assemble_integer (rtx
, unsigned int, int);
561 static int set_extends (rtx_insn
*);
562 static void sparc_asm_function_prologue (FILE *, HOST_WIDE_INT
);
563 static void sparc_asm_function_epilogue (FILE *, HOST_WIDE_INT
);
564 #ifdef TARGET_SOLARIS
565 static void sparc_solaris_elf_asm_named_section (const char *, unsigned int,
566 tree
) ATTRIBUTE_UNUSED
;
568 static int sparc_adjust_cost (rtx_insn
*, rtx
, rtx_insn
*, int);
569 static int sparc_issue_rate (void);
570 static void sparc_sched_init (FILE *, int, int);
571 static int sparc_use_sched_lookahead (void);
573 static void emit_soft_tfmode_libcall (const char *, int, rtx
*);
574 static void emit_soft_tfmode_binop (enum rtx_code
, rtx
*);
575 static void emit_soft_tfmode_unop (enum rtx_code
, rtx
*);
576 static void emit_soft_tfmode_cvt (enum rtx_code
, rtx
*);
577 static void emit_hard_tfmode_operation (enum rtx_code
, rtx
*);
579 static bool sparc_function_ok_for_sibcall (tree
, tree
);
580 static void sparc_init_libfuncs (void);
581 static void sparc_init_builtins (void);
582 static void sparc_fpu_init_builtins (void);
583 static void sparc_vis_init_builtins (void);
584 static tree
sparc_builtin_decl (unsigned, bool);
585 static rtx
sparc_expand_builtin (tree
, rtx
, rtx
, machine_mode
, int);
586 static tree
sparc_fold_builtin (tree
, int, tree
*, bool);
587 static void sparc_output_mi_thunk (FILE *, tree
, HOST_WIDE_INT
,
588 HOST_WIDE_INT
, tree
);
589 static bool sparc_can_output_mi_thunk (const_tree
, HOST_WIDE_INT
,
590 HOST_WIDE_INT
, const_tree
);
591 static struct machine_function
* sparc_init_machine_status (void);
592 static bool sparc_cannot_force_const_mem (machine_mode
, rtx
);
593 static rtx
sparc_tls_get_addr (void);
594 static rtx
sparc_tls_got (void);
595 static int sparc_register_move_cost (machine_mode
,
596 reg_class_t
, reg_class_t
);
597 static bool sparc_rtx_costs (rtx
, machine_mode
, int, int, int *, bool);
598 static rtx
sparc_function_value (const_tree
, const_tree
, bool);
599 static rtx
sparc_libcall_value (machine_mode
, const_rtx
);
600 static bool sparc_function_value_regno_p (const unsigned int);
601 static rtx
sparc_struct_value_rtx (tree
, int);
602 static machine_mode
sparc_promote_function_mode (const_tree
, machine_mode
,
603 int *, const_tree
, int);
604 static bool sparc_return_in_memory (const_tree
, const_tree
);
605 static bool sparc_strict_argument_naming (cumulative_args_t
);
606 static void sparc_va_start (tree
, rtx
);
607 static tree
sparc_gimplify_va_arg (tree
, tree
, gimple_seq
*, gimple_seq
*);
608 static bool sparc_vector_mode_supported_p (machine_mode
);
609 static bool sparc_tls_referenced_p (rtx
);
610 static rtx
sparc_legitimize_tls_address (rtx
);
611 static rtx
sparc_legitimize_pic_address (rtx
, rtx
);
612 static rtx
sparc_legitimize_address (rtx
, rtx
, machine_mode
);
613 static rtx
sparc_delegitimize_address (rtx
);
614 static bool sparc_mode_dependent_address_p (const_rtx
, addr_space_t
);
615 static bool sparc_pass_by_reference (cumulative_args_t
,
616 machine_mode
, const_tree
, bool);
617 static void sparc_function_arg_advance (cumulative_args_t
,
618 machine_mode
, const_tree
, bool);
619 static rtx
sparc_function_arg_1 (cumulative_args_t
,
620 machine_mode
, const_tree
, bool, bool);
621 static rtx
sparc_function_arg (cumulative_args_t
,
622 machine_mode
, const_tree
, bool);
623 static rtx
sparc_function_incoming_arg (cumulative_args_t
,
624 machine_mode
, const_tree
, bool);
625 static unsigned int sparc_function_arg_boundary (machine_mode
,
627 static int sparc_arg_partial_bytes (cumulative_args_t
,
628 machine_mode
, tree
, bool);
629 static void sparc_output_dwarf_dtprel (FILE *, int, rtx
) ATTRIBUTE_UNUSED
;
630 static void sparc_file_end (void);
631 static bool sparc_frame_pointer_required (void);
632 static bool sparc_can_eliminate (const int, const int);
633 static rtx
sparc_builtin_setjmp_frame_value (void);
634 static void sparc_conditional_register_usage (void);
635 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
636 static const char *sparc_mangle_type (const_tree
);
638 static void sparc_trampoline_init (rtx
, tree
, rtx
);
639 static machine_mode
sparc_preferred_simd_mode (machine_mode
);
640 static reg_class_t
sparc_preferred_reload_class (rtx x
, reg_class_t rclass
);
641 static bool sparc_print_operand_punct_valid_p (unsigned char);
642 static void sparc_print_operand (FILE *, rtx
, int);
643 static void sparc_print_operand_address (FILE *, machine_mode
, rtx
);
644 static reg_class_t
sparc_secondary_reload (bool, rtx
, reg_class_t
,
646 secondary_reload_info
*);
647 static machine_mode
sparc_cstore_mode (enum insn_code icode
);
648 static void sparc_atomic_assign_expand_fenv (tree
*, tree
*, tree
*);
650 #ifdef SUBTARGET_ATTRIBUTE_TABLE
651 /* Table of valid machine attributes. */
652 static const struct attribute_spec sparc_attribute_table
[] =
654 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler,
656 SUBTARGET_ATTRIBUTE_TABLE
,
657 { NULL
, 0, 0, false, false, false, NULL
, false }
661 /* Option handling. */
664 enum cmodel sparc_cmodel
;
666 char sparc_hard_reg_printed
[8];
668 /* Initialize the GCC target structure. */
670 /* The default is to use .half rather than .short for aligned HI objects. */
671 #undef TARGET_ASM_ALIGNED_HI_OP
672 #define TARGET_ASM_ALIGNED_HI_OP "\t.half\t"
674 #undef TARGET_ASM_UNALIGNED_HI_OP
675 #define TARGET_ASM_UNALIGNED_HI_OP "\t.uahalf\t"
676 #undef TARGET_ASM_UNALIGNED_SI_OP
677 #define TARGET_ASM_UNALIGNED_SI_OP "\t.uaword\t"
678 #undef TARGET_ASM_UNALIGNED_DI_OP
679 #define TARGET_ASM_UNALIGNED_DI_OP "\t.uaxword\t"
681 /* The target hook has to handle DI-mode values. */
682 #undef TARGET_ASM_INTEGER
683 #define TARGET_ASM_INTEGER sparc_assemble_integer
685 #undef TARGET_ASM_FUNCTION_PROLOGUE
686 #define TARGET_ASM_FUNCTION_PROLOGUE sparc_asm_function_prologue
687 #undef TARGET_ASM_FUNCTION_EPILOGUE
688 #define TARGET_ASM_FUNCTION_EPILOGUE sparc_asm_function_epilogue
690 #undef TARGET_SCHED_ADJUST_COST
691 #define TARGET_SCHED_ADJUST_COST sparc_adjust_cost
692 #undef TARGET_SCHED_ISSUE_RATE
693 #define TARGET_SCHED_ISSUE_RATE sparc_issue_rate
694 #undef TARGET_SCHED_INIT
695 #define TARGET_SCHED_INIT sparc_sched_init
696 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
697 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD sparc_use_sched_lookahead
699 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
700 #define TARGET_FUNCTION_OK_FOR_SIBCALL sparc_function_ok_for_sibcall
702 #undef TARGET_INIT_LIBFUNCS
703 #define TARGET_INIT_LIBFUNCS sparc_init_libfuncs
705 #undef TARGET_LEGITIMIZE_ADDRESS
706 #define TARGET_LEGITIMIZE_ADDRESS sparc_legitimize_address
707 #undef TARGET_DELEGITIMIZE_ADDRESS
708 #define TARGET_DELEGITIMIZE_ADDRESS sparc_delegitimize_address
709 #undef TARGET_MODE_DEPENDENT_ADDRESS_P
710 #define TARGET_MODE_DEPENDENT_ADDRESS_P sparc_mode_dependent_address_p
712 #undef TARGET_INIT_BUILTINS
713 #define TARGET_INIT_BUILTINS sparc_init_builtins
714 #undef TARGET_BUILTIN_DECL
715 #define TARGET_BUILTIN_DECL sparc_builtin_decl
716 #undef TARGET_EXPAND_BUILTIN
717 #define TARGET_EXPAND_BUILTIN sparc_expand_builtin
718 #undef TARGET_FOLD_BUILTIN
719 #define TARGET_FOLD_BUILTIN sparc_fold_builtin
722 #undef TARGET_HAVE_TLS
723 #define TARGET_HAVE_TLS true
726 #undef TARGET_CANNOT_FORCE_CONST_MEM
727 #define TARGET_CANNOT_FORCE_CONST_MEM sparc_cannot_force_const_mem
729 #undef TARGET_ASM_OUTPUT_MI_THUNK
730 #define TARGET_ASM_OUTPUT_MI_THUNK sparc_output_mi_thunk
731 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
732 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK sparc_can_output_mi_thunk
734 #undef TARGET_RTX_COSTS
735 #define TARGET_RTX_COSTS sparc_rtx_costs
736 #undef TARGET_ADDRESS_COST
737 #define TARGET_ADDRESS_COST hook_int_rtx_mode_as_bool_0
738 #undef TARGET_REGISTER_MOVE_COST
739 #define TARGET_REGISTER_MOVE_COST sparc_register_move_cost
741 #undef TARGET_PROMOTE_FUNCTION_MODE
742 #define TARGET_PROMOTE_FUNCTION_MODE sparc_promote_function_mode
744 #undef TARGET_FUNCTION_VALUE
745 #define TARGET_FUNCTION_VALUE sparc_function_value
746 #undef TARGET_LIBCALL_VALUE
747 #define TARGET_LIBCALL_VALUE sparc_libcall_value
748 #undef TARGET_FUNCTION_VALUE_REGNO_P
749 #define TARGET_FUNCTION_VALUE_REGNO_P sparc_function_value_regno_p
751 #undef TARGET_STRUCT_VALUE_RTX
752 #define TARGET_STRUCT_VALUE_RTX sparc_struct_value_rtx
753 #undef TARGET_RETURN_IN_MEMORY
754 #define TARGET_RETURN_IN_MEMORY sparc_return_in_memory
755 #undef TARGET_MUST_PASS_IN_STACK
756 #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size
757 #undef TARGET_PASS_BY_REFERENCE
758 #define TARGET_PASS_BY_REFERENCE sparc_pass_by_reference
759 #undef TARGET_ARG_PARTIAL_BYTES
760 #define TARGET_ARG_PARTIAL_BYTES sparc_arg_partial_bytes
761 #undef TARGET_FUNCTION_ARG_ADVANCE
762 #define TARGET_FUNCTION_ARG_ADVANCE sparc_function_arg_advance
763 #undef TARGET_FUNCTION_ARG
764 #define TARGET_FUNCTION_ARG sparc_function_arg
765 #undef TARGET_FUNCTION_INCOMING_ARG
766 #define TARGET_FUNCTION_INCOMING_ARG sparc_function_incoming_arg
767 #undef TARGET_FUNCTION_ARG_BOUNDARY
768 #define TARGET_FUNCTION_ARG_BOUNDARY sparc_function_arg_boundary
770 #undef TARGET_EXPAND_BUILTIN_SAVEREGS
771 #define TARGET_EXPAND_BUILTIN_SAVEREGS sparc_builtin_saveregs
772 #undef TARGET_STRICT_ARGUMENT_NAMING
773 #define TARGET_STRICT_ARGUMENT_NAMING sparc_strict_argument_naming
775 #undef TARGET_EXPAND_BUILTIN_VA_START
776 #define TARGET_EXPAND_BUILTIN_VA_START sparc_va_start
777 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
778 #define TARGET_GIMPLIFY_VA_ARG_EXPR sparc_gimplify_va_arg
780 #undef TARGET_VECTOR_MODE_SUPPORTED_P
781 #define TARGET_VECTOR_MODE_SUPPORTED_P sparc_vector_mode_supported_p
783 #undef TARGET_VECTORIZE_PREFERRED_SIMD_MODE
784 #define TARGET_VECTORIZE_PREFERRED_SIMD_MODE sparc_preferred_simd_mode
786 #ifdef SUBTARGET_INSERT_ATTRIBUTES
787 #undef TARGET_INSERT_ATTRIBUTES
788 #define TARGET_INSERT_ATTRIBUTES SUBTARGET_INSERT_ATTRIBUTES
791 #ifdef SUBTARGET_ATTRIBUTE_TABLE
792 #undef TARGET_ATTRIBUTE_TABLE
793 #define TARGET_ATTRIBUTE_TABLE sparc_attribute_table
796 #undef TARGET_OPTION_OVERRIDE
797 #define TARGET_OPTION_OVERRIDE sparc_option_override
799 #if TARGET_GNU_TLS && defined(HAVE_AS_SPARC_UA_PCREL)
800 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
801 #define TARGET_ASM_OUTPUT_DWARF_DTPREL sparc_output_dwarf_dtprel
804 #undef TARGET_ASM_FILE_END
805 #define TARGET_ASM_FILE_END sparc_file_end
807 #undef TARGET_FRAME_POINTER_REQUIRED
808 #define TARGET_FRAME_POINTER_REQUIRED sparc_frame_pointer_required
810 #undef TARGET_BUILTIN_SETJMP_FRAME_VALUE
811 #define TARGET_BUILTIN_SETJMP_FRAME_VALUE sparc_builtin_setjmp_frame_value
813 #undef TARGET_CAN_ELIMINATE
814 #define TARGET_CAN_ELIMINATE sparc_can_eliminate
816 #undef TARGET_PREFERRED_RELOAD_CLASS
817 #define TARGET_PREFERRED_RELOAD_CLASS sparc_preferred_reload_class
819 #undef TARGET_SECONDARY_RELOAD
820 #define TARGET_SECONDARY_RELOAD sparc_secondary_reload
822 #undef TARGET_CONDITIONAL_REGISTER_USAGE
823 #define TARGET_CONDITIONAL_REGISTER_USAGE sparc_conditional_register_usage
825 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
826 #undef TARGET_MANGLE_TYPE
827 #define TARGET_MANGLE_TYPE sparc_mangle_type
830 #undef TARGET_LEGITIMATE_ADDRESS_P
831 #define TARGET_LEGITIMATE_ADDRESS_P sparc_legitimate_address_p
833 #undef TARGET_LEGITIMATE_CONSTANT_P
834 #define TARGET_LEGITIMATE_CONSTANT_P sparc_legitimate_constant_p
836 #undef TARGET_TRAMPOLINE_INIT
837 #define TARGET_TRAMPOLINE_INIT sparc_trampoline_init
839 #undef TARGET_PRINT_OPERAND_PUNCT_VALID_P
840 #define TARGET_PRINT_OPERAND_PUNCT_VALID_P sparc_print_operand_punct_valid_p
841 #undef TARGET_PRINT_OPERAND
842 #define TARGET_PRINT_OPERAND sparc_print_operand
843 #undef TARGET_PRINT_OPERAND_ADDRESS
844 #define TARGET_PRINT_OPERAND_ADDRESS sparc_print_operand_address
846 /* The value stored by LDSTUB. */
847 #undef TARGET_ATOMIC_TEST_AND_SET_TRUEVAL
848 #define TARGET_ATOMIC_TEST_AND_SET_TRUEVAL 0xff
850 #undef TARGET_CSTORE_MODE
851 #define TARGET_CSTORE_MODE sparc_cstore_mode
853 #undef TARGET_ATOMIC_ASSIGN_EXPAND_FENV
854 #define TARGET_ATOMIC_ASSIGN_EXPAND_FENV sparc_atomic_assign_expand_fenv
856 struct gcc_target targetm
= TARGET_INITIALIZER
;
858 /* Return the memory reference contained in X if any, zero otherwise. */
863 if (GET_CODE (x
) == SIGN_EXTEND
|| GET_CODE (x
) == ZERO_EXTEND
)
872 /* We use a machine specific pass to enable workarounds for errata.
873 We need to have the (essentially) final form of the insn stream in order
874 to properly detect the various hazards. Therefore, this machine specific
875 pass runs as late as possible. The pass is inserted in the pass pipeline
876 at the end of sparc_option_override. */
879 sparc_do_work_around_errata (void)
881 rtx_insn
*insn
, *next
;
883 /* Force all instructions to be split into their final form. */
884 split_all_insns_noflow ();
886 /* Now look for specific patterns in the insn stream. */
887 for (insn
= get_insns (); insn
; insn
= next
)
889 bool insert_nop
= false;
892 /* Look into the instruction in a delay slot. */
893 if (NONJUMP_INSN_P (insn
))
894 if (rtx_sequence
*seq
= dyn_cast
<rtx_sequence
*> (PATTERN (insn
)))
895 insn
= seq
->insn (1);
897 /* Look for a single-word load into an odd-numbered FP register. */
899 && NONJUMP_INSN_P (insn
)
900 && (set
= single_set (insn
)) != NULL_RTX
901 && GET_MODE_SIZE (GET_MODE (SET_SRC (set
))) == 4
902 && MEM_P (SET_SRC (set
))
903 && REG_P (SET_DEST (set
))
904 && REGNO (SET_DEST (set
)) > 31
905 && REGNO (SET_DEST (set
)) % 2 != 0)
907 /* The wrong dependency is on the enclosing double register. */
908 const unsigned int x
= REGNO (SET_DEST (set
)) - 1;
909 unsigned int src1
, src2
, dest
;
912 next
= next_active_insn (insn
);
915 /* If the insn is a branch, then it cannot be problematic. */
916 if (!NONJUMP_INSN_P (next
) || GET_CODE (PATTERN (next
)) == SEQUENCE
)
920 code
= INSN_CODE (next
);
924 case CODE_FOR_adddf3
:
925 case CODE_FOR_subdf3
:
926 case CODE_FOR_muldf3
:
927 case CODE_FOR_divdf3
:
928 dest
= REGNO (recog_data
.operand
[0]);
929 src1
= REGNO (recog_data
.operand
[1]);
930 src2
= REGNO (recog_data
.operand
[2]);
935 FPOPd %f{x,y}, %f{y,x}, %f{x,y} */
936 if ((src1
== x
|| src2
== x
)
937 && (dest
== src1
|| dest
== src2
))
944 FPOPd %fx, %fx, %fx */
947 && (code
== CODE_FOR_adddf3
|| code
== CODE_FOR_muldf3
))
952 case CODE_FOR_sqrtdf2
:
953 dest
= REGNO (recog_data
.operand
[0]);
954 src1
= REGNO (recog_data
.operand
[1]);
958 if (src1
== x
&& dest
== src1
)
967 /* Look for a single-word load into an integer register. */
968 else if (sparc_fix_ut699
969 && NONJUMP_INSN_P (insn
)
970 && (set
= single_set (insn
)) != NULL_RTX
971 && GET_MODE_SIZE (GET_MODE (SET_SRC (set
))) <= 4
972 && mem_ref (SET_SRC (set
)) != NULL_RTX
973 && REG_P (SET_DEST (set
))
974 && REGNO (SET_DEST (set
)) < 32)
976 /* There is no problem if the second memory access has a data
977 dependency on the first single-cycle load. */
978 rtx x
= SET_DEST (set
);
980 next
= next_active_insn (insn
);
983 /* If the insn is a branch, then it cannot be problematic. */
984 if (!NONJUMP_INSN_P (next
) || GET_CODE (PATTERN (next
)) == SEQUENCE
)
987 /* Look for a second memory access to/from an integer register. */
988 if ((set
= single_set (next
)) != NULL_RTX
)
990 rtx src
= SET_SRC (set
);
991 rtx dest
= SET_DEST (set
);
994 /* LDD is affected. */
995 if ((mem
= mem_ref (src
)) != NULL_RTX
998 && !reg_mentioned_p (x
, XEXP (mem
, 0)))
1001 /* STD is *not* affected. */
1002 else if (MEM_P (dest
)
1003 && GET_MODE_SIZE (GET_MODE (dest
)) <= 4
1004 && (src
== CONST0_RTX (GET_MODE (dest
))
1007 && REGNO (src
) != REGNO (x
)))
1008 && !reg_mentioned_p (x
, XEXP (dest
, 0)))
1013 /* Look for a single-word load/operation into an FP register. */
1014 else if (sparc_fix_ut699
1015 && NONJUMP_INSN_P (insn
)
1016 && (set
= single_set (insn
)) != NULL_RTX
1017 && GET_MODE_SIZE (GET_MODE (SET_SRC (set
))) == 4
1018 && REG_P (SET_DEST (set
))
1019 && REGNO (SET_DEST (set
)) > 31)
1021 /* Number of instructions in the problematic window. */
1022 const int n_insns
= 4;
1023 /* The problematic combination is with the sibling FP register. */
1024 const unsigned int x
= REGNO (SET_DEST (set
));
1025 const unsigned int y
= x
^ 1;
1029 next
= next_active_insn (insn
);
1032 /* If the insn is a branch, then it cannot be problematic. */
1033 if (!NONJUMP_INSN_P (next
) || GET_CODE (PATTERN (next
)) == SEQUENCE
)
1036 /* Look for a second load/operation into the sibling FP register. */
1037 if (!((set
= single_set (next
)) != NULL_RTX
1038 && GET_MODE_SIZE (GET_MODE (SET_SRC (set
))) == 4
1039 && REG_P (SET_DEST (set
))
1040 && REGNO (SET_DEST (set
)) == y
))
1043 /* Look for a (possible) store from the FP register in the next N
1044 instructions, but bail out if it is again modified or if there
1045 is a store from the sibling FP register before this store. */
1046 for (after
= next
, i
= 0; i
< n_insns
; i
++)
1050 after
= next_active_insn (after
);
1054 /* This is a branch with an empty delay slot. */
1055 if (!NONJUMP_INSN_P (after
))
1062 /* This is a branch with a filled delay slot. */
1063 else if (rtx_sequence
*seq
=
1064 dyn_cast
<rtx_sequence
*> (PATTERN (after
)))
1069 after
= seq
->insn (1);
1071 /* This is a regular instruction. */
1075 if (after
&& (set
= single_set (after
)) != NULL_RTX
)
1077 const rtx src
= SET_SRC (set
);
1078 const rtx dest
= SET_DEST (set
);
1079 const unsigned int size
= GET_MODE_SIZE (GET_MODE (dest
));
1081 /* If the FP register is again modified before the store,
1082 then the store isn't affected. */
1084 && (REGNO (dest
) == x
1085 || (REGNO (dest
) == y
&& size
== 8)))
1088 if (MEM_P (dest
) && REG_P (src
))
1090 /* If there is a store from the sibling FP register
1091 before the store, then the store is not affected. */
1092 if (REGNO (src
) == y
|| (REGNO (src
) == x
&& size
== 8))
1095 /* Otherwise, the store is affected. */
1096 if (REGNO (src
) == x
&& size
== 4)
1104 /* If we have a branch in the first M instructions, then we
1105 cannot see the (M+2)th instruction so we play safe. */
1106 if (branch_p
&& i
<= (n_insns
- 2))
1115 next
= NEXT_INSN (insn
);
1118 emit_insn_before (gen_nop (), next
);
1126 const pass_data pass_data_work_around_errata
=
1128 RTL_PASS
, /* type */
1129 "errata", /* name */
1130 OPTGROUP_NONE
, /* optinfo_flags */
1131 TV_MACH_DEP
, /* tv_id */
1132 0, /* properties_required */
1133 0, /* properties_provided */
1134 0, /* properties_destroyed */
1135 0, /* todo_flags_start */
1136 0, /* todo_flags_finish */
1139 class pass_work_around_errata
: public rtl_opt_pass
1142 pass_work_around_errata(gcc::context
*ctxt
)
1143 : rtl_opt_pass(pass_data_work_around_errata
, ctxt
)
1146 /* opt_pass methods: */
1147 virtual bool gate (function
*)
1149 /* The only errata we handle are those of the AT697F and UT699. */
1150 return sparc_fix_at697f
!= 0 || sparc_fix_ut699
!= 0;
1153 virtual unsigned int execute (function
*)
1155 return sparc_do_work_around_errata ();
1158 }; // class pass_work_around_errata
1163 make_pass_work_around_errata (gcc::context
*ctxt
)
1165 return new pass_work_around_errata (ctxt
);
1168 /* Helpers for TARGET_DEBUG_OPTIONS. */
1170 dump_target_flag_bits (const int flags
)
1172 if (flags
& MASK_64BIT
)
1173 fprintf (stderr
, "64BIT ");
1174 if (flags
& MASK_APP_REGS
)
1175 fprintf (stderr
, "APP_REGS ");
1176 if (flags
& MASK_FASTER_STRUCTS
)
1177 fprintf (stderr
, "FASTER_STRUCTS ");
1178 if (flags
& MASK_FLAT
)
1179 fprintf (stderr
, "FLAT ");
1180 if (flags
& MASK_FMAF
)
1181 fprintf (stderr
, "FMAF ");
1182 if (flags
& MASK_FPU
)
1183 fprintf (stderr
, "FPU ");
1184 if (flags
& MASK_HARD_QUAD
)
1185 fprintf (stderr
, "HARD_QUAD ");
1186 if (flags
& MASK_POPC
)
1187 fprintf (stderr
, "POPC ");
1188 if (flags
& MASK_PTR64
)
1189 fprintf (stderr
, "PTR64 ");
1190 if (flags
& MASK_STACK_BIAS
)
1191 fprintf (stderr
, "STACK_BIAS ");
1192 if (flags
& MASK_UNALIGNED_DOUBLES
)
1193 fprintf (stderr
, "UNALIGNED_DOUBLES ");
1194 if (flags
& MASK_V8PLUS
)
1195 fprintf (stderr
, "V8PLUS ");
1196 if (flags
& MASK_VIS
)
1197 fprintf (stderr
, "VIS ");
1198 if (flags
& MASK_VIS2
)
1199 fprintf (stderr
, "VIS2 ");
1200 if (flags
& MASK_VIS3
)
1201 fprintf (stderr
, "VIS3 ");
1202 if (flags
& MASK_VIS4
)
1203 fprintf (stderr
, "VIS4 ");
1204 if (flags
& MASK_CBCOND
)
1205 fprintf (stderr
, "CBCOND ");
1206 if (flags
& MASK_DEPRECATED_V8_INSNS
)
1207 fprintf (stderr
, "DEPRECATED_V8_INSNS ");
1208 if (flags
& MASK_SPARCLET
)
1209 fprintf (stderr
, "SPARCLET ");
1210 if (flags
& MASK_SPARCLITE
)
1211 fprintf (stderr
, "SPARCLITE ");
1212 if (flags
& MASK_V8
)
1213 fprintf (stderr
, "V8 ");
1214 if (flags
& MASK_V9
)
1215 fprintf (stderr
, "V9 ");
1219 dump_target_flags (const char *prefix
, const int flags
)
1221 fprintf (stderr
, "%s: (%08x) [ ", prefix
, flags
);
1222 dump_target_flag_bits (flags
);
1223 fprintf(stderr
, "]\n");
1226 /* Validate and override various options, and do some machine dependent
1230 sparc_option_override (void)
1232 static struct code_model
{
1233 const char *const name
;
1234 const enum cmodel value
;
1235 } const cmodels
[] = {
1237 { "medlow", CM_MEDLOW
},
1238 { "medmid", CM_MEDMID
},
1239 { "medany", CM_MEDANY
},
1240 { "embmedany", CM_EMBMEDANY
},
1241 { NULL
, (enum cmodel
) 0 }
1243 const struct code_model
*cmodel
;
1244 /* Map TARGET_CPU_DEFAULT to value for -m{cpu,tune}=. */
1245 static struct cpu_default
{
1247 const enum processor_type processor
;
1248 } const cpu_default
[] = {
1249 /* There must be one entry here for each TARGET_CPU value. */
1250 { TARGET_CPU_sparc
, PROCESSOR_CYPRESS
},
1251 { TARGET_CPU_v8
, PROCESSOR_V8
},
1252 { TARGET_CPU_supersparc
, PROCESSOR_SUPERSPARC
},
1253 { TARGET_CPU_hypersparc
, PROCESSOR_HYPERSPARC
},
1254 { TARGET_CPU_leon
, PROCESSOR_LEON
},
1255 { TARGET_CPU_leon3
, PROCESSOR_LEON3
},
1256 { TARGET_CPU_leon3v7
, PROCESSOR_LEON3V7
},
1257 { TARGET_CPU_sparclite
, PROCESSOR_F930
},
1258 { TARGET_CPU_sparclite86x
, PROCESSOR_SPARCLITE86X
},
1259 { TARGET_CPU_sparclet
, PROCESSOR_TSC701
},
1260 { TARGET_CPU_v9
, PROCESSOR_V9
},
1261 { TARGET_CPU_ultrasparc
, PROCESSOR_ULTRASPARC
},
1262 { TARGET_CPU_ultrasparc3
, PROCESSOR_ULTRASPARC3
},
1263 { TARGET_CPU_niagara
, PROCESSOR_NIAGARA
},
1264 { TARGET_CPU_niagara2
, PROCESSOR_NIAGARA2
},
1265 { TARGET_CPU_niagara3
, PROCESSOR_NIAGARA3
},
1266 { TARGET_CPU_niagara4
, PROCESSOR_NIAGARA4
},
1267 { TARGET_CPU_niagara7
, PROCESSOR_NIAGARA7
},
1268 { -1, PROCESSOR_V7
}
1270 const struct cpu_default
*def
;
1271 /* Table of values for -m{cpu,tune}=. This must match the order of
1272 the enum processor_type in sparc-opts.h. */
1273 static struct cpu_table
{
1274 const char *const name
;
1277 } const cpu_table
[] = {
1278 { "v7", MASK_ISA
, 0 },
1279 { "cypress", MASK_ISA
, 0 },
1280 { "v8", MASK_ISA
, MASK_V8
},
1281 /* TI TMS390Z55 supersparc */
1282 { "supersparc", MASK_ISA
, MASK_V8
},
1283 { "hypersparc", MASK_ISA
, MASK_V8
|MASK_FPU
},
1284 { "leon", MASK_ISA
, MASK_V8
|MASK_LEON
|MASK_FPU
},
1285 { "leon3", MASK_ISA
, MASK_V8
|MASK_LEON3
|MASK_FPU
},
1286 { "leon3v7", MASK_ISA
, MASK_LEON3
|MASK_FPU
},
1287 { "sparclite", MASK_ISA
, MASK_SPARCLITE
},
1288 /* The Fujitsu MB86930 is the original sparclite chip, with no FPU. */
1289 { "f930", MASK_ISA
|MASK_FPU
, MASK_SPARCLITE
},
1290 /* The Fujitsu MB86934 is the recent sparclite chip, with an FPU. */
1291 { "f934", MASK_ISA
, MASK_SPARCLITE
|MASK_FPU
},
1292 { "sparclite86x", MASK_ISA
|MASK_FPU
, MASK_SPARCLITE
},
1293 { "sparclet", MASK_ISA
, MASK_SPARCLET
},
1294 /* TEMIC sparclet */
1295 { "tsc701", MASK_ISA
, MASK_SPARCLET
},
1296 { "v9", MASK_ISA
, MASK_V9
},
1297 /* UltraSPARC I, II, IIi */
1298 { "ultrasparc", MASK_ISA
,
1299 /* Although insns using %y are deprecated, it is a clear win. */
1300 MASK_V9
|MASK_DEPRECATED_V8_INSNS
},
1301 /* UltraSPARC III */
1302 /* ??? Check if %y issue still holds true. */
1303 { "ultrasparc3", MASK_ISA
,
1304 MASK_V9
|MASK_DEPRECATED_V8_INSNS
|MASK_VIS2
},
1306 { "niagara", MASK_ISA
,
1307 MASK_V9
|MASK_DEPRECATED_V8_INSNS
},
1309 { "niagara2", MASK_ISA
,
1310 MASK_V9
|MASK_POPC
|MASK_VIS2
},
1312 { "niagara3", MASK_ISA
,
1313 MASK_V9
|MASK_POPC
|MASK_VIS2
|MASK_VIS3
|MASK_FMAF
},
1315 { "niagara4", MASK_ISA
,
1316 MASK_V9
|MASK_POPC
|MASK_VIS2
|MASK_VIS3
|MASK_FMAF
|MASK_CBCOND
},
1318 { "niagara7", MASK_ISA
,
1319 MASK_V9
|MASK_POPC
|MASK_VIS2
|MASK_VIS3
|MASK_VIS4
|MASK_FMAF
|MASK_CBCOND
},
1321 const struct cpu_table
*cpu
;
1325 if (sparc_debug_string
!= NULL
)
1330 p
= ASTRDUP (sparc_debug_string
);
1331 while ((q
= strtok (p
, ",")) != NULL
)
1345 if (! strcmp (q
, "all"))
1346 mask
= MASK_DEBUG_ALL
;
1347 else if (! strcmp (q
, "options"))
1348 mask
= MASK_DEBUG_OPTIONS
;
1350 error ("unknown -mdebug-%s switch", q
);
1353 sparc_debug
&= ~mask
;
1355 sparc_debug
|= mask
;
1359 if (TARGET_DEBUG_OPTIONS
)
1361 dump_target_flags("Initial target_flags", target_flags
);
1362 dump_target_flags("target_flags_explicit", target_flags_explicit
);
1365 #ifdef SUBTARGET_OVERRIDE_OPTIONS
1366 SUBTARGET_OVERRIDE_OPTIONS
;
1369 #ifndef SPARC_BI_ARCH
1370 /* Check for unsupported architecture size. */
1371 if (! TARGET_64BIT
!= DEFAULT_ARCH32_P
)
1372 error ("%s is not supported by this configuration",
1373 DEFAULT_ARCH32_P
? "-m64" : "-m32");
1376 /* We force all 64bit archs to use 128 bit long double */
1377 if (TARGET_64BIT
&& ! TARGET_LONG_DOUBLE_128
)
1379 error ("-mlong-double-64 not allowed with -m64");
1380 target_flags
|= MASK_LONG_DOUBLE_128
;
1383 /* Code model selection. */
1384 sparc_cmodel
= SPARC_DEFAULT_CMODEL
;
1386 #ifdef SPARC_BI_ARCH
1388 sparc_cmodel
= CM_32
;
1391 if (sparc_cmodel_string
!= NULL
)
1395 for (cmodel
= &cmodels
[0]; cmodel
->name
; cmodel
++)
1396 if (strcmp (sparc_cmodel_string
, cmodel
->name
) == 0)
1398 if (cmodel
->name
== NULL
)
1399 error ("bad value (%s) for -mcmodel= switch", sparc_cmodel_string
);
1401 sparc_cmodel
= cmodel
->value
;
1404 error ("-mcmodel= is not supported on 32 bit systems");
1407 /* Check that -fcall-saved-REG wasn't specified for out registers. */
1408 for (i
= 8; i
< 16; i
++)
1409 if (!call_used_regs
[i
])
1411 error ("-fcall-saved-REG is not supported for out registers");
1412 call_used_regs
[i
] = 1;
1415 fpu
= target_flags
& MASK_FPU
; /* save current -mfpu status */
1417 /* Set the default CPU. */
1418 if (!global_options_set
.x_sparc_cpu_and_features
)
1420 for (def
= &cpu_default
[0]; def
->cpu
!= -1; ++def
)
1421 if (def
->cpu
== TARGET_CPU_DEFAULT
)
1423 gcc_assert (def
->cpu
!= -1);
1424 sparc_cpu_and_features
= def
->processor
;
1427 if (!global_options_set
.x_sparc_cpu
)
1428 sparc_cpu
= sparc_cpu_and_features
;
1430 cpu
= &cpu_table
[(int) sparc_cpu_and_features
];
1432 if (TARGET_DEBUG_OPTIONS
)
1434 fprintf (stderr
, "sparc_cpu_and_features: %s\n", cpu
->name
);
1435 fprintf (stderr
, "sparc_cpu: %s\n",
1436 cpu_table
[(int) sparc_cpu
].name
);
1437 dump_target_flags ("cpu->disable", cpu
->disable
);
1438 dump_target_flags ("cpu->enable", cpu
->enable
);
1441 target_flags
&= ~cpu
->disable
;
1442 target_flags
|= (cpu
->enable
1443 #ifndef HAVE_AS_FMAF_HPC_VIS3
1444 & ~(MASK_FMAF
| MASK_VIS3
)
1446 #ifndef HAVE_AS_SPARC4
1449 #ifndef HAVE_AS_SPARC5_VIS4
1452 #ifndef HAVE_AS_LEON
1453 & ~(MASK_LEON
| MASK_LEON3
)
1457 /* If -mfpu or -mno-fpu was explicitly used, don't override with
1458 the processor default. */
1459 if (target_flags_explicit
& MASK_FPU
)
1460 target_flags
= (target_flags
& ~MASK_FPU
) | fpu
;
1462 /* -mvis2 implies -mvis */
1464 target_flags
|= MASK_VIS
;
1466 /* -mvis3 implies -mvis2 and -mvis */
1468 target_flags
|= MASK_VIS2
| MASK_VIS
;
1470 /* -mvis4 implies -mvis3, -mvis2 and -mvis */
1472 target_flags
|= MASK_VIS3
| MASK_VIS2
| MASK_VIS
;
1474 /* Don't allow -mvis, -mvis2, -mvis3, -mvis4 or -mfmaf if FPU is
1477 target_flags
&= ~(MASK_VIS
| MASK_VIS2
| MASK_VIS3
| MASK_VIS4
1480 /* -mvis assumes UltraSPARC+, so we are sure v9 instructions
1482 -m64 also implies v9. */
1483 if (TARGET_VIS
|| TARGET_ARCH64
)
1485 target_flags
|= MASK_V9
;
1486 target_flags
&= ~(MASK_V8
| MASK_SPARCLET
| MASK_SPARCLITE
);
1489 /* -mvis also implies -mv8plus on 32-bit */
1490 if (TARGET_VIS
&& ! TARGET_ARCH64
)
1491 target_flags
|= MASK_V8PLUS
;
1493 /* Use the deprecated v8 insns for sparc64 in 32 bit mode. */
1494 if (TARGET_V9
&& TARGET_ARCH32
)
1495 target_flags
|= MASK_DEPRECATED_V8_INSNS
;
1497 /* V8PLUS requires V9, makes no sense in 64 bit mode. */
1498 if (! TARGET_V9
|| TARGET_ARCH64
)
1499 target_flags
&= ~MASK_V8PLUS
;
1501 /* Don't use stack biasing in 32 bit mode. */
1503 target_flags
&= ~MASK_STACK_BIAS
;
1505 /* Supply a default value for align_functions. */
1506 if (align_functions
== 0
1507 && (sparc_cpu
== PROCESSOR_ULTRASPARC
1508 || sparc_cpu
== PROCESSOR_ULTRASPARC3
1509 || sparc_cpu
== PROCESSOR_NIAGARA
1510 || sparc_cpu
== PROCESSOR_NIAGARA2
1511 || sparc_cpu
== PROCESSOR_NIAGARA3
1512 || sparc_cpu
== PROCESSOR_NIAGARA4
1513 || sparc_cpu
== PROCESSOR_NIAGARA7
))
1514 align_functions
= 32;
1516 /* Validate PCC_STRUCT_RETURN. */
1517 if (flag_pcc_struct_return
== DEFAULT_PCC_STRUCT_RETURN
)
1518 flag_pcc_struct_return
= (TARGET_ARCH64
? 0 : 1);
1520 /* Only use .uaxword when compiling for a 64-bit target. */
1522 targetm
.asm_out
.unaligned_op
.di
= NULL
;
1524 /* Do various machine dependent initializations. */
1525 sparc_init_modes ();
1527 /* Set up function hooks. */
1528 init_machine_status
= sparc_init_machine_status
;
1533 case PROCESSOR_CYPRESS
:
1534 sparc_costs
= &cypress_costs
;
1537 case PROCESSOR_SPARCLITE
:
1538 case PROCESSOR_SUPERSPARC
:
1539 sparc_costs
= &supersparc_costs
;
1541 case PROCESSOR_F930
:
1542 case PROCESSOR_F934
:
1543 case PROCESSOR_HYPERSPARC
:
1544 case PROCESSOR_SPARCLITE86X
:
1545 sparc_costs
= &hypersparc_costs
;
1547 case PROCESSOR_LEON
:
1548 sparc_costs
= &leon_costs
;
1550 case PROCESSOR_LEON3
:
1551 case PROCESSOR_LEON3V7
:
1552 sparc_costs
= &leon3_costs
;
1554 case PROCESSOR_SPARCLET
:
1555 case PROCESSOR_TSC701
:
1556 sparc_costs
= &sparclet_costs
;
1559 case PROCESSOR_ULTRASPARC
:
1560 sparc_costs
= &ultrasparc_costs
;
1562 case PROCESSOR_ULTRASPARC3
:
1563 sparc_costs
= &ultrasparc3_costs
;
1565 case PROCESSOR_NIAGARA
:
1566 sparc_costs
= &niagara_costs
;
1568 case PROCESSOR_NIAGARA2
:
1569 sparc_costs
= &niagara2_costs
;
1571 case PROCESSOR_NIAGARA3
:
1572 sparc_costs
= &niagara3_costs
;
1574 case PROCESSOR_NIAGARA4
:
1575 sparc_costs
= &niagara4_costs
;
1577 case PROCESSOR_NIAGARA7
:
1578 sparc_costs
= &niagara7_costs
;
1580 case PROCESSOR_NATIVE
:
1584 if (sparc_memory_model
== SMM_DEFAULT
)
1586 /* Choose the memory model for the operating system. */
1587 enum sparc_memory_model_type os_default
= SUBTARGET_DEFAULT_MEMORY_MODEL
;
1588 if (os_default
!= SMM_DEFAULT
)
1589 sparc_memory_model
= os_default
;
1590 /* Choose the most relaxed model for the processor. */
1592 sparc_memory_model
= SMM_RMO
;
1593 else if (TARGET_LEON3
)
1594 sparc_memory_model
= SMM_TSO
;
1595 else if (TARGET_LEON
)
1596 sparc_memory_model
= SMM_SC
;
1598 sparc_memory_model
= SMM_PSO
;
1600 sparc_memory_model
= SMM_SC
;
1603 #ifdef TARGET_DEFAULT_LONG_DOUBLE_128
1604 if (!(target_flags_explicit
& MASK_LONG_DOUBLE_128
))
1605 target_flags
|= MASK_LONG_DOUBLE_128
;
1608 if (TARGET_DEBUG_OPTIONS
)
1609 dump_target_flags ("Final target_flags", target_flags
);
1611 /* PARAM_SIMULTANEOUS_PREFETCHES is the number of prefetches that
1612 can run at the same time. More important, it is the threshold
1613 defining when additional prefetches will be dropped by the
1616 The UltraSPARC-III features a documented prefetch queue with a
1617 size of 8. Additional prefetches issued in the cpu are
1620 Niagara processors are different. In these processors prefetches
1621 are handled much like regular loads. The L1 miss buffer is 32
1622 entries, but prefetches start getting affected when 30 entries
1623 become occupied. That occupation could be a mix of regular loads
1624 and prefetches though. And that buffer is shared by all threads.
1625 Once the threshold is reached, if the core is running a single
1626 thread the prefetch will retry. If more than one thread is
1627 running, the prefetch will be dropped.
1629 All this makes it very difficult to determine how many
1630 simultaneous prefetches can be issued simultaneously, even in a
1631 single-threaded program. Experimental results show that setting
1632 this parameter to 32 works well when the number of threads is not
1634 maybe_set_param_value (PARAM_SIMULTANEOUS_PREFETCHES
,
1635 ((sparc_cpu
== PROCESSOR_ULTRASPARC
1636 || sparc_cpu
== PROCESSOR_NIAGARA
1637 || sparc_cpu
== PROCESSOR_NIAGARA2
1638 || sparc_cpu
== PROCESSOR_NIAGARA3
1639 || sparc_cpu
== PROCESSOR_NIAGARA4
)
1641 : (sparc_cpu
== PROCESSOR_ULTRASPARC3
1642 ? 8 : (sparc_cpu
== PROCESSOR_NIAGARA7
1644 global_options
.x_param_values
,
1645 global_options_set
.x_param_values
);
1647 /* For PARAM_L1_CACHE_LINE_SIZE we use the default 32 bytes (see
1648 params.def), so no maybe_set_param_value is needed.
1650 The Oracle SPARC Architecture (previously the UltraSPARC
1651 Architecture) specification states that when a PREFETCH[A]
1652 instruction is executed an implementation-specific amount of data
1653 is prefetched, and that it is at least 64 bytes long (aligned to
1656 However, this is not correct. The M7 (and implementations prior
1657 to that) does not guarantee a 64B prefetch into a cache if the
1658 line size is smaller. A single cache line is all that is ever
1659 prefetched. So for the M7, where the L1D$ has 32B lines and the
1660 L2D$ and L3 have 64B lines, a prefetch will prefetch 64B into the
1661 L2 and L3, but only 32B are brought into the L1D$. (Assuming it
1662 is a read_n prefetch, which is the only type which allocates to
1665 /* PARAM_L1_CACHE_SIZE is the size of the L1D$ (most SPARC chips use
1666 Hardvard level-1 caches) in kilobytes. Both UltraSPARC and
1667 Niagara processors feature a L1D$ of 16KB. */
1668 maybe_set_param_value (PARAM_L1_CACHE_SIZE
,
1669 ((sparc_cpu
== PROCESSOR_ULTRASPARC
1670 || sparc_cpu
== PROCESSOR_ULTRASPARC3
1671 || sparc_cpu
== PROCESSOR_NIAGARA
1672 || sparc_cpu
== PROCESSOR_NIAGARA2
1673 || sparc_cpu
== PROCESSOR_NIAGARA3
1674 || sparc_cpu
== PROCESSOR_NIAGARA4
1675 || sparc_cpu
== PROCESSOR_NIAGARA7
)
1677 global_options
.x_param_values
,
1678 global_options_set
.x_param_values
);
1681 /* PARAM_L2_CACHE_SIZE is the size fo the L2 in kilobytes. Note
1682 that 512 is the default in params.def. */
1683 maybe_set_param_value (PARAM_L2_CACHE_SIZE
,
1684 (sparc_cpu
== PROCESSOR_NIAGARA4
1685 ? 128 : (sparc_cpu
== PROCESSOR_NIAGARA7
1687 global_options
.x_param_values
,
1688 global_options_set
.x_param_values
);
1691 /* Disable save slot sharing for call-clobbered registers by default.
1692 The IRA sharing algorithm works on single registers only and this
1693 pessimizes for double floating-point registers. */
1694 if (!global_options_set
.x_flag_ira_share_save_slots
)
1695 flag_ira_share_save_slots
= 0;
1697 /* We register a machine specific pass to work around errata, if any.
1698 The pass mut be scheduled as late as possible so that we have the
1699 (essentially) final form of the insn stream to work on.
1700 Registering the pass must be done at start up. It's convenient to
1702 opt_pass
*errata_pass
= make_pass_work_around_errata (g
);
1703 struct register_pass_info insert_pass_work_around_errata
=
1705 errata_pass
, /* pass */
1706 "dbr", /* reference_pass_name */
1707 1, /* ref_pass_instance_number */
1708 PASS_POS_INSERT_AFTER
/* po_op */
1710 register_pass (&insert_pass_work_around_errata
);
1713 /* Miscellaneous utilities. */
1715 /* Nonzero if CODE, a comparison, is suitable for use in v9 conditional move
1716 or branch on register contents instructions. */
1719 v9_regcmp_p (enum rtx_code code
)
1721 return (code
== EQ
|| code
== NE
|| code
== GE
|| code
== LT
1722 || code
== LE
|| code
== GT
);
1725 /* Nonzero if OP is a floating point constant which can
1726 be loaded into an integer register using a single
1727 sethi instruction. */
1732 if (GET_CODE (op
) == CONST_DOUBLE
)
1736 REAL_VALUE_TO_TARGET_SINGLE (*CONST_DOUBLE_REAL_VALUE (op
), i
);
1737 return !SPARC_SIMM13_P (i
) && SPARC_SETHI_P (i
);
1743 /* Nonzero if OP is a floating point constant which can
1744 be loaded into an integer register using a single
1750 if (GET_CODE (op
) == CONST_DOUBLE
)
1754 REAL_VALUE_TO_TARGET_SINGLE (*CONST_DOUBLE_REAL_VALUE (op
), i
);
1755 return SPARC_SIMM13_P (i
);
1761 /* Nonzero if OP is a floating point constant which can
1762 be loaded into an integer register using a high/losum
1763 instruction sequence. */
1766 fp_high_losum_p (rtx op
)
1768 /* The constraints calling this should only be in
1769 SFmode move insns, so any constant which cannot
1770 be moved using a single insn will do. */
1771 if (GET_CODE (op
) == CONST_DOUBLE
)
1775 REAL_VALUE_TO_TARGET_SINGLE (*CONST_DOUBLE_REAL_VALUE (op
), i
);
1776 return !SPARC_SIMM13_P (i
) && !SPARC_SETHI_P (i
);
1782 /* Return true if the address of LABEL can be loaded by means of the
1783 mov{si,di}_pic_label_ref patterns in PIC mode. */
1786 can_use_mov_pic_label_ref (rtx label
)
1788 /* VxWorks does not impose a fixed gap between segments; the run-time
1789 gap can be different from the object-file gap. We therefore can't
1790 assume X - _GLOBAL_OFFSET_TABLE_ is a link-time constant unless we
1791 are absolutely sure that X is in the same segment as the GOT.
1792 Unfortunately, the flexibility of linker scripts means that we
1793 can't be sure of that in general, so assume that GOT-relative
1794 accesses are never valid on VxWorks. */
1795 if (TARGET_VXWORKS_RTP
)
1798 /* Similarly, if the label is non-local, it might end up being placed
1799 in a different section than the current one; now mov_pic_label_ref
1800 requires the label and the code to be in the same section. */
1801 if (LABEL_REF_NONLOCAL_P (label
))
1804 /* Finally, if we are reordering basic blocks and partition into hot
1805 and cold sections, this might happen for any label. */
1806 if (flag_reorder_blocks_and_partition
)
1812 /* Expand a move instruction. Return true if all work is done. */
1815 sparc_expand_move (machine_mode mode
, rtx
*operands
)
1817 /* Handle sets of MEM first. */
1818 if (GET_CODE (operands
[0]) == MEM
)
1820 /* 0 is a register (or a pair of registers) on SPARC. */
1821 if (register_or_zero_operand (operands
[1], mode
))
1824 if (!reload_in_progress
)
1826 operands
[0] = validize_mem (operands
[0]);
1827 operands
[1] = force_reg (mode
, operands
[1]);
1831 /* Fixup TLS cases. */
1833 && CONSTANT_P (operands
[1])
1834 && sparc_tls_referenced_p (operands
[1]))
1836 operands
[1] = sparc_legitimize_tls_address (operands
[1]);
1840 /* Fixup PIC cases. */
1841 if (flag_pic
&& CONSTANT_P (operands
[1]))
1843 if (pic_address_needs_scratch (operands
[1]))
1844 operands
[1] = sparc_legitimize_pic_address (operands
[1], NULL_RTX
);
1846 /* We cannot use the mov{si,di}_pic_label_ref patterns in all cases. */
1847 if (GET_CODE (operands
[1]) == LABEL_REF
1848 && can_use_mov_pic_label_ref (operands
[1]))
1852 emit_insn (gen_movsi_pic_label_ref (operands
[0], operands
[1]));
1858 gcc_assert (TARGET_ARCH64
);
1859 emit_insn (gen_movdi_pic_label_ref (operands
[0], operands
[1]));
1864 if (symbolic_operand (operands
[1], mode
))
1867 = sparc_legitimize_pic_address (operands
[1],
1869 ? operands
[0] : NULL_RTX
);
1874 /* If we are trying to toss an integer constant into FP registers,
1875 or loading a FP or vector constant, force it into memory. */
1876 if (CONSTANT_P (operands
[1])
1877 && REG_P (operands
[0])
1878 && (SPARC_FP_REG_P (REGNO (operands
[0]))
1879 || SCALAR_FLOAT_MODE_P (mode
)
1880 || VECTOR_MODE_P (mode
)))
1882 /* emit_group_store will send such bogosity to us when it is
1883 not storing directly into memory. So fix this up to avoid
1884 crashes in output_constant_pool. */
1885 if (operands
[1] == const0_rtx
)
1886 operands
[1] = CONST0_RTX (mode
);
1888 /* We can clear or set to all-ones FP registers if TARGET_VIS, and
1889 always other regs. */
1890 if ((TARGET_VIS
|| REGNO (operands
[0]) < SPARC_FIRST_FP_REG
)
1891 && (const_zero_operand (operands
[1], mode
)
1892 || const_all_ones_operand (operands
[1], mode
)))
1895 if (REGNO (operands
[0]) < SPARC_FIRST_FP_REG
1896 /* We are able to build any SF constant in integer registers
1897 with at most 2 instructions. */
1899 /* And any DF constant in integer registers. */
1901 && ! can_create_pseudo_p ())))
1904 operands
[1] = force_const_mem (mode
, operands
[1]);
1905 if (!reload_in_progress
)
1906 operands
[1] = validize_mem (operands
[1]);
1910 /* Accept non-constants and valid constants unmodified. */
1911 if (!CONSTANT_P (operands
[1])
1912 || GET_CODE (operands
[1]) == HIGH
1913 || input_operand (operands
[1], mode
))
1919 /* All QImode constants require only one insn, so proceed. */
1924 sparc_emit_set_const32 (operands
[0], operands
[1]);
1928 /* input_operand should have filtered out 32-bit mode. */
1929 sparc_emit_set_const64 (operands
[0], operands
[1]);
1935 /* TImode isn't available in 32-bit mode. */
1936 split_double (operands
[1], &high
, &low
);
1937 emit_insn (gen_movdi (operand_subword (operands
[0], 0, 0, TImode
),
1939 emit_insn (gen_movdi (operand_subword (operands
[0], 1, 0, TImode
),
1951 /* Load OP1, a 32-bit constant, into OP0, a register.
1952 We know it can't be done in one insn when we get
1953 here, the move expander guarantees this. */
1956 sparc_emit_set_const32 (rtx op0
, rtx op1
)
1958 machine_mode mode
= GET_MODE (op0
);
1961 if (can_create_pseudo_p ())
1962 temp
= gen_reg_rtx (mode
);
1964 if (GET_CODE (op1
) == CONST_INT
)
1966 gcc_assert (!small_int_operand (op1
, mode
)
1967 && !const_high_operand (op1
, mode
));
1969 /* Emit them as real moves instead of a HIGH/LO_SUM,
1970 this way CSE can see everything and reuse intermediate
1971 values if it wants. */
1972 emit_insn (gen_rtx_SET (temp
, GEN_INT (INTVAL (op1
)
1973 & ~(HOST_WIDE_INT
) 0x3ff)));
1975 emit_insn (gen_rtx_SET (op0
,
1976 gen_rtx_IOR (mode
, temp
,
1977 GEN_INT (INTVAL (op1
) & 0x3ff))));
1981 /* A symbol, emit in the traditional way. */
1982 emit_insn (gen_rtx_SET (temp
, gen_rtx_HIGH (mode
, op1
)));
1983 emit_insn (gen_rtx_SET (op0
, gen_rtx_LO_SUM (mode
, temp
, op1
)));
1987 /* Load OP1, a symbolic 64-bit constant, into OP0, a DImode register.
1988 If TEMP is nonzero, we are forbidden to use any other scratch
1989 registers. Otherwise, we are allowed to generate them as needed.
1991 Note that TEMP may have TImode if the code model is TARGET_CM_MEDANY
1992 or TARGET_CM_EMBMEDANY (see the reload_indi and reload_outdi patterns). */
1995 sparc_emit_set_symbolic_const64 (rtx op0
, rtx op1
, rtx temp
)
1997 rtx temp1
, temp2
, temp3
, temp4
, temp5
;
2000 if (temp
&& GET_MODE (temp
) == TImode
)
2003 temp
= gen_rtx_REG (DImode
, REGNO (temp
));
2006 /* SPARC-V9 code-model support. */
2007 switch (sparc_cmodel
)
2010 /* The range spanned by all instructions in the object is less
2011 than 2^31 bytes (2GB) and the distance from any instruction
2012 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
2013 than 2^31 bytes (2GB).
2015 The executable must be in the low 4TB of the virtual address
2018 sethi %hi(symbol), %temp1
2019 or %temp1, %lo(symbol), %reg */
2021 temp1
= temp
; /* op0 is allowed. */
2023 temp1
= gen_reg_rtx (DImode
);
2025 emit_insn (gen_rtx_SET (temp1
, gen_rtx_HIGH (DImode
, op1
)));
2026 emit_insn (gen_rtx_SET (op0
, gen_rtx_LO_SUM (DImode
, temp1
, op1
)));
2030 /* The range spanned by all instructions in the object is less
2031 than 2^31 bytes (2GB) and the distance from any instruction
2032 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
2033 than 2^31 bytes (2GB).
2035 The executable must be in the low 16TB of the virtual address
2038 sethi %h44(symbol), %temp1
2039 or %temp1, %m44(symbol), %temp2
2040 sllx %temp2, 12, %temp3
2041 or %temp3, %l44(symbol), %reg */
2046 temp3
= temp
; /* op0 is allowed. */
2050 temp1
= gen_reg_rtx (DImode
);
2051 temp2
= gen_reg_rtx (DImode
);
2052 temp3
= gen_reg_rtx (DImode
);
2055 emit_insn (gen_seth44 (temp1
, op1
));
2056 emit_insn (gen_setm44 (temp2
, temp1
, op1
));
2057 emit_insn (gen_rtx_SET (temp3
,
2058 gen_rtx_ASHIFT (DImode
, temp2
, GEN_INT (12))));
2059 emit_insn (gen_setl44 (op0
, temp3
, op1
));
2063 /* The range spanned by all instructions in the object is less
2064 than 2^31 bytes (2GB) and the distance from any instruction
2065 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
2066 than 2^31 bytes (2GB).
2068 The executable can be placed anywhere in the virtual address
2071 sethi %hh(symbol), %temp1
2072 sethi %lm(symbol), %temp2
2073 or %temp1, %hm(symbol), %temp3
2074 sllx %temp3, 32, %temp4
2075 or %temp4, %temp2, %temp5
2076 or %temp5, %lo(symbol), %reg */
2079 /* It is possible that one of the registers we got for operands[2]
2080 might coincide with that of operands[0] (which is why we made
2081 it TImode). Pick the other one to use as our scratch. */
2082 if (rtx_equal_p (temp
, op0
))
2084 gcc_assert (ti_temp
);
2085 temp
= gen_rtx_REG (DImode
, REGNO (temp
) + 1);
2088 temp2
= temp
; /* op0 is _not_ allowed, see above. */
2095 temp1
= gen_reg_rtx (DImode
);
2096 temp2
= gen_reg_rtx (DImode
);
2097 temp3
= gen_reg_rtx (DImode
);
2098 temp4
= gen_reg_rtx (DImode
);
2099 temp5
= gen_reg_rtx (DImode
);
2102 emit_insn (gen_sethh (temp1
, op1
));
2103 emit_insn (gen_setlm (temp2
, op1
));
2104 emit_insn (gen_sethm (temp3
, temp1
, op1
));
2105 emit_insn (gen_rtx_SET (temp4
,
2106 gen_rtx_ASHIFT (DImode
, temp3
, GEN_INT (32))));
2107 emit_insn (gen_rtx_SET (temp5
, gen_rtx_PLUS (DImode
, temp4
, temp2
)));
2108 emit_insn (gen_setlo (op0
, temp5
, op1
));
2112 /* Old old old backwards compatibility kruft here.
2113 Essentially it is MEDLOW with a fixed 64-bit
2114 virtual base added to all data segment addresses.
2115 Text-segment stuff is computed like MEDANY, we can't
2116 reuse the code above because the relocation knobs
2119 Data segment: sethi %hi(symbol), %temp1
2120 add %temp1, EMBMEDANY_BASE_REG, %temp2
2121 or %temp2, %lo(symbol), %reg */
2122 if (data_segment_operand (op1
, GET_MODE (op1
)))
2126 temp1
= temp
; /* op0 is allowed. */
2131 temp1
= gen_reg_rtx (DImode
);
2132 temp2
= gen_reg_rtx (DImode
);
2135 emit_insn (gen_embmedany_sethi (temp1
, op1
));
2136 emit_insn (gen_embmedany_brsum (temp2
, temp1
));
2137 emit_insn (gen_embmedany_losum (op0
, temp2
, op1
));
2140 /* Text segment: sethi %uhi(symbol), %temp1
2141 sethi %hi(symbol), %temp2
2142 or %temp1, %ulo(symbol), %temp3
2143 sllx %temp3, 32, %temp4
2144 or %temp4, %temp2, %temp5
2145 or %temp5, %lo(symbol), %reg */
2150 /* It is possible that one of the registers we got for operands[2]
2151 might coincide with that of operands[0] (which is why we made
2152 it TImode). Pick the other one to use as our scratch. */
2153 if (rtx_equal_p (temp
, op0
))
2155 gcc_assert (ti_temp
);
2156 temp
= gen_rtx_REG (DImode
, REGNO (temp
) + 1);
2159 temp2
= temp
; /* op0 is _not_ allowed, see above. */
2166 temp1
= gen_reg_rtx (DImode
);
2167 temp2
= gen_reg_rtx (DImode
);
2168 temp3
= gen_reg_rtx (DImode
);
2169 temp4
= gen_reg_rtx (DImode
);
2170 temp5
= gen_reg_rtx (DImode
);
2173 emit_insn (gen_embmedany_textuhi (temp1
, op1
));
2174 emit_insn (gen_embmedany_texthi (temp2
, op1
));
2175 emit_insn (gen_embmedany_textulo (temp3
, temp1
, op1
));
2176 emit_insn (gen_rtx_SET (temp4
,
2177 gen_rtx_ASHIFT (DImode
, temp3
, GEN_INT (32))));
2178 emit_insn (gen_rtx_SET (temp5
, gen_rtx_PLUS (DImode
, temp4
, temp2
)));
2179 emit_insn (gen_embmedany_textlo (op0
, temp5
, op1
));
2188 /* These avoid problems when cross compiling. If we do not
2189 go through all this hair then the optimizer will see
2190 invalid REG_EQUAL notes or in some cases none at all. */
2191 static rtx
gen_safe_HIGH64 (rtx
, HOST_WIDE_INT
);
2192 static rtx
gen_safe_SET64 (rtx
, HOST_WIDE_INT
);
2193 static rtx
gen_safe_OR64 (rtx
, HOST_WIDE_INT
);
2194 static rtx
gen_safe_XOR64 (rtx
, HOST_WIDE_INT
);
2196 /* The optimizer is not to assume anything about exactly
2197 which bits are set for a HIGH, they are unspecified.
2198 Unfortunately this leads to many missed optimizations
2199 during CSE. We mask out the non-HIGH bits, and matches
2200 a plain movdi, to alleviate this problem. */
2202 gen_safe_HIGH64 (rtx dest
, HOST_WIDE_INT val
)
2204 return gen_rtx_SET (dest
, GEN_INT (val
& ~(HOST_WIDE_INT
)0x3ff));
2208 gen_safe_SET64 (rtx dest
, HOST_WIDE_INT val
)
2210 return gen_rtx_SET (dest
, GEN_INT (val
));
2214 gen_safe_OR64 (rtx src
, HOST_WIDE_INT val
)
2216 return gen_rtx_IOR (DImode
, src
, GEN_INT (val
));
2220 gen_safe_XOR64 (rtx src
, HOST_WIDE_INT val
)
2222 return gen_rtx_XOR (DImode
, src
, GEN_INT (val
));
2225 /* Worker routines for 64-bit constant formation on arch64.
2226 One of the key things to be doing in these emissions is
2227 to create as many temp REGs as possible. This makes it
2228 possible for half-built constants to be used later when
2229 such values are similar to something required later on.
2230 Without doing this, the optimizer cannot see such
2233 static void sparc_emit_set_const64_quick1 (rtx
, rtx
,
2234 unsigned HOST_WIDE_INT
, int);
2237 sparc_emit_set_const64_quick1 (rtx op0
, rtx temp
,
2238 unsigned HOST_WIDE_INT low_bits
, int is_neg
)
2240 unsigned HOST_WIDE_INT high_bits
;
2243 high_bits
= (~low_bits
) & 0xffffffff;
2245 high_bits
= low_bits
;
2247 emit_insn (gen_safe_HIGH64 (temp
, high_bits
));
2250 emit_insn (gen_rtx_SET (op0
, gen_safe_OR64 (temp
, (high_bits
& 0x3ff))));
2254 /* If we are XOR'ing with -1, then we should emit a one's complement
2255 instead. This way the combiner will notice logical operations
2256 such as ANDN later on and substitute. */
2257 if ((low_bits
& 0x3ff) == 0x3ff)
2259 emit_insn (gen_rtx_SET (op0
, gen_rtx_NOT (DImode
, temp
)));
2263 emit_insn (gen_rtx_SET (op0
,
2264 gen_safe_XOR64 (temp
,
2265 (-(HOST_WIDE_INT
)0x400
2266 | (low_bits
& 0x3ff)))));
2271 static void sparc_emit_set_const64_quick2 (rtx
, rtx
, unsigned HOST_WIDE_INT
,
2272 unsigned HOST_WIDE_INT
, int);
2275 sparc_emit_set_const64_quick2 (rtx op0
, rtx temp
,
2276 unsigned HOST_WIDE_INT high_bits
,
2277 unsigned HOST_WIDE_INT low_immediate
,
2282 if ((high_bits
& 0xfffffc00) != 0)
2284 emit_insn (gen_safe_HIGH64 (temp
, high_bits
));
2285 if ((high_bits
& ~0xfffffc00) != 0)
2286 emit_insn (gen_rtx_SET (op0
,
2287 gen_safe_OR64 (temp
, (high_bits
& 0x3ff))));
2293 emit_insn (gen_safe_SET64 (temp
, high_bits
));
2297 /* Now shift it up into place. */
2298 emit_insn (gen_rtx_SET (op0
, gen_rtx_ASHIFT (DImode
, temp2
,
2299 GEN_INT (shift_count
))));
2301 /* If there is a low immediate part piece, finish up by
2302 putting that in as well. */
2303 if (low_immediate
!= 0)
2304 emit_insn (gen_rtx_SET (op0
, gen_safe_OR64 (op0
, low_immediate
)));
2307 static void sparc_emit_set_const64_longway (rtx
, rtx
, unsigned HOST_WIDE_INT
,
2308 unsigned HOST_WIDE_INT
);
2310 /* Full 64-bit constant decomposition. Even though this is the
2311 'worst' case, we still optimize a few things away. */
2313 sparc_emit_set_const64_longway (rtx op0
, rtx temp
,
2314 unsigned HOST_WIDE_INT high_bits
,
2315 unsigned HOST_WIDE_INT low_bits
)
2319 if (can_create_pseudo_p ())
2320 sub_temp
= gen_reg_rtx (DImode
);
2322 if ((high_bits
& 0xfffffc00) != 0)
2324 emit_insn (gen_safe_HIGH64 (temp
, high_bits
));
2325 if ((high_bits
& ~0xfffffc00) != 0)
2326 emit_insn (gen_rtx_SET (sub_temp
,
2327 gen_safe_OR64 (temp
, (high_bits
& 0x3ff))));
2333 emit_insn (gen_safe_SET64 (temp
, high_bits
));
2337 if (can_create_pseudo_p ())
2339 rtx temp2
= gen_reg_rtx (DImode
);
2340 rtx temp3
= gen_reg_rtx (DImode
);
2341 rtx temp4
= gen_reg_rtx (DImode
);
2343 emit_insn (gen_rtx_SET (temp4
, gen_rtx_ASHIFT (DImode
, sub_temp
,
2346 emit_insn (gen_safe_HIGH64 (temp2
, low_bits
));
2347 if ((low_bits
& ~0xfffffc00) != 0)
2349 emit_insn (gen_rtx_SET (temp3
,
2350 gen_safe_OR64 (temp2
, (low_bits
& 0x3ff))));
2351 emit_insn (gen_rtx_SET (op0
, gen_rtx_PLUS (DImode
, temp4
, temp3
)));
2355 emit_insn (gen_rtx_SET (op0
, gen_rtx_PLUS (DImode
, temp4
, temp2
)));
2360 rtx low1
= GEN_INT ((low_bits
>> (32 - 12)) & 0xfff);
2361 rtx low2
= GEN_INT ((low_bits
>> (32 - 12 - 12)) & 0xfff);
2362 rtx low3
= GEN_INT ((low_bits
>> (32 - 12 - 12 - 8)) & 0x0ff);
2365 /* We are in the middle of reload, so this is really
2366 painful. However we do still make an attempt to
2367 avoid emitting truly stupid code. */
2368 if (low1
!= const0_rtx
)
2370 emit_insn (gen_rtx_SET (op0
, gen_rtx_ASHIFT (DImode
, sub_temp
,
2371 GEN_INT (to_shift
))));
2372 emit_insn (gen_rtx_SET (op0
, gen_rtx_IOR (DImode
, op0
, low1
)));
2380 if (low2
!= const0_rtx
)
2382 emit_insn (gen_rtx_SET (op0
, gen_rtx_ASHIFT (DImode
, sub_temp
,
2383 GEN_INT (to_shift
))));
2384 emit_insn (gen_rtx_SET (op0
, gen_rtx_IOR (DImode
, op0
, low2
)));
2392 emit_insn (gen_rtx_SET (op0
, gen_rtx_ASHIFT (DImode
, sub_temp
,
2393 GEN_INT (to_shift
))));
2394 if (low3
!= const0_rtx
)
2395 emit_insn (gen_rtx_SET (op0
, gen_rtx_IOR (DImode
, op0
, low3
)));
2400 /* Analyze a 64-bit constant for certain properties. */
2401 static void analyze_64bit_constant (unsigned HOST_WIDE_INT
,
2402 unsigned HOST_WIDE_INT
,
2403 int *, int *, int *);
2406 analyze_64bit_constant (unsigned HOST_WIDE_INT high_bits
,
2407 unsigned HOST_WIDE_INT low_bits
,
2408 int *hbsp
, int *lbsp
, int *abbasp
)
2410 int lowest_bit_set
, highest_bit_set
, all_bits_between_are_set
;
2413 lowest_bit_set
= highest_bit_set
= -1;
2417 if ((lowest_bit_set
== -1)
2418 && ((low_bits
>> i
) & 1))
2420 if ((highest_bit_set
== -1)
2421 && ((high_bits
>> (32 - i
- 1)) & 1))
2422 highest_bit_set
= (64 - i
- 1);
2425 && ((highest_bit_set
== -1)
2426 || (lowest_bit_set
== -1)));
2432 if ((lowest_bit_set
== -1)
2433 && ((high_bits
>> i
) & 1))
2434 lowest_bit_set
= i
+ 32;
2435 if ((highest_bit_set
== -1)
2436 && ((low_bits
>> (32 - i
- 1)) & 1))
2437 highest_bit_set
= 32 - i
- 1;
2440 && ((highest_bit_set
== -1)
2441 || (lowest_bit_set
== -1)));
2443 /* If there are no bits set this should have gone out
2444 as one instruction! */
2445 gcc_assert (lowest_bit_set
!= -1 && highest_bit_set
!= -1);
2446 all_bits_between_are_set
= 1;
2447 for (i
= lowest_bit_set
; i
<= highest_bit_set
; i
++)
2451 if ((low_bits
& (1 << i
)) != 0)
2456 if ((high_bits
& (1 << (i
- 32))) != 0)
2459 all_bits_between_are_set
= 0;
2462 *hbsp
= highest_bit_set
;
2463 *lbsp
= lowest_bit_set
;
2464 *abbasp
= all_bits_between_are_set
;
2467 static int const64_is_2insns (unsigned HOST_WIDE_INT
, unsigned HOST_WIDE_INT
);
2470 const64_is_2insns (unsigned HOST_WIDE_INT high_bits
,
2471 unsigned HOST_WIDE_INT low_bits
)
2473 int highest_bit_set
, lowest_bit_set
, all_bits_between_are_set
;
2476 || high_bits
== 0xffffffff)
2479 analyze_64bit_constant (high_bits
, low_bits
,
2480 &highest_bit_set
, &lowest_bit_set
,
2481 &all_bits_between_are_set
);
2483 if ((highest_bit_set
== 63
2484 || lowest_bit_set
== 0)
2485 && all_bits_between_are_set
!= 0)
2488 if ((highest_bit_set
- lowest_bit_set
) < 21)
2494 static unsigned HOST_WIDE_INT
create_simple_focus_bits (unsigned HOST_WIDE_INT
,
2495 unsigned HOST_WIDE_INT
,
2498 static unsigned HOST_WIDE_INT
2499 create_simple_focus_bits (unsigned HOST_WIDE_INT high_bits
,
2500 unsigned HOST_WIDE_INT low_bits
,
2501 int lowest_bit_set
, int shift
)
2503 HOST_WIDE_INT hi
, lo
;
2505 if (lowest_bit_set
< 32)
2507 lo
= (low_bits
>> lowest_bit_set
) << shift
;
2508 hi
= ((high_bits
<< (32 - lowest_bit_set
)) << shift
);
2513 hi
= ((high_bits
>> (lowest_bit_set
- 32)) << shift
);
2515 gcc_assert (! (hi
& lo
));
2519 /* Here we are sure to be arch64 and this is an integer constant
2520 being loaded into a register. Emit the most efficient
2521 insn sequence possible. Detection of all the 1-insn cases
2522 has been done already. */
2524 sparc_emit_set_const64 (rtx op0
, rtx op1
)
2526 unsigned HOST_WIDE_INT high_bits
, low_bits
;
2527 int lowest_bit_set
, highest_bit_set
;
2528 int all_bits_between_are_set
;
2531 /* Sanity check that we know what we are working with. */
2532 gcc_assert (TARGET_ARCH64
2533 && (GET_CODE (op0
) == SUBREG
2534 || (REG_P (op0
) && ! SPARC_FP_REG_P (REGNO (op0
)))));
2536 if (! can_create_pseudo_p ())
2539 if (GET_CODE (op1
) != CONST_INT
)
2541 sparc_emit_set_symbolic_const64 (op0
, op1
, temp
);
2546 temp
= gen_reg_rtx (DImode
);
2548 high_bits
= ((INTVAL (op1
) >> 32) & 0xffffffff);
2549 low_bits
= (INTVAL (op1
) & 0xffffffff);
2551 /* low_bits bits 0 --> 31
2552 high_bits bits 32 --> 63 */
2554 analyze_64bit_constant (high_bits
, low_bits
,
2555 &highest_bit_set
, &lowest_bit_set
,
2556 &all_bits_between_are_set
);
2558 /* First try for a 2-insn sequence. */
2560 /* These situations are preferred because the optimizer can
2561 * do more things with them:
2563 * sllx %reg, shift, %reg
2565 * srlx %reg, shift, %reg
2566 * 3) mov some_small_const, %reg
2567 * sllx %reg, shift, %reg
2569 if (((highest_bit_set
== 63
2570 || lowest_bit_set
== 0)
2571 && all_bits_between_are_set
!= 0)
2572 || ((highest_bit_set
- lowest_bit_set
) < 12))
2574 HOST_WIDE_INT the_const
= -1;
2575 int shift
= lowest_bit_set
;
2577 if ((highest_bit_set
!= 63
2578 && lowest_bit_set
!= 0)
2579 || all_bits_between_are_set
== 0)
2582 create_simple_focus_bits (high_bits
, low_bits
,
2585 else if (lowest_bit_set
== 0)
2586 shift
= -(63 - highest_bit_set
);
2588 gcc_assert (SPARC_SIMM13_P (the_const
));
2589 gcc_assert (shift
!= 0);
2591 emit_insn (gen_safe_SET64 (temp
, the_const
));
2593 emit_insn (gen_rtx_SET (op0
, gen_rtx_ASHIFT (DImode
, temp
,
2596 emit_insn (gen_rtx_SET (op0
, gen_rtx_LSHIFTRT (DImode
, temp
,
2597 GEN_INT (-shift
))));
2601 /* Now a range of 22 or less bits set somewhere.
2602 * 1) sethi %hi(focus_bits), %reg
2603 * sllx %reg, shift, %reg
2604 * 2) sethi %hi(focus_bits), %reg
2605 * srlx %reg, shift, %reg
2607 if ((highest_bit_set
- lowest_bit_set
) < 21)
2609 unsigned HOST_WIDE_INT focus_bits
=
2610 create_simple_focus_bits (high_bits
, low_bits
,
2611 lowest_bit_set
, 10);
2613 gcc_assert (SPARC_SETHI_P (focus_bits
));
2614 gcc_assert (lowest_bit_set
!= 10);
2616 emit_insn (gen_safe_HIGH64 (temp
, focus_bits
));
2618 /* If lowest_bit_set == 10 then a sethi alone could have done it. */
2619 if (lowest_bit_set
< 10)
2620 emit_insn (gen_rtx_SET (op0
,
2621 gen_rtx_LSHIFTRT (DImode
, temp
,
2622 GEN_INT (10 - lowest_bit_set
))));
2623 else if (lowest_bit_set
> 10)
2624 emit_insn (gen_rtx_SET (op0
,
2625 gen_rtx_ASHIFT (DImode
, temp
,
2626 GEN_INT (lowest_bit_set
- 10))));
2630 /* 1) sethi %hi(low_bits), %reg
2631 * or %reg, %lo(low_bits), %reg
2632 * 2) sethi %hi(~low_bits), %reg
2633 * xor %reg, %lo(-0x400 | (low_bits & 0x3ff)), %reg
2636 || high_bits
== 0xffffffff)
2638 sparc_emit_set_const64_quick1 (op0
, temp
, low_bits
,
2639 (high_bits
== 0xffffffff));
2643 /* Now, try 3-insn sequences. */
2645 /* 1) sethi %hi(high_bits), %reg
2646 * or %reg, %lo(high_bits), %reg
2647 * sllx %reg, 32, %reg
2651 sparc_emit_set_const64_quick2 (op0
, temp
, high_bits
, 0, 32);
2655 /* We may be able to do something quick
2656 when the constant is negated, so try that. */
2657 if (const64_is_2insns ((~high_bits
) & 0xffffffff,
2658 (~low_bits
) & 0xfffffc00))
2660 /* NOTE: The trailing bits get XOR'd so we need the
2661 non-negated bits, not the negated ones. */
2662 unsigned HOST_WIDE_INT trailing_bits
= low_bits
& 0x3ff;
2664 if ((((~high_bits
) & 0xffffffff) == 0
2665 && ((~low_bits
) & 0x80000000) == 0)
2666 || (((~high_bits
) & 0xffffffff) == 0xffffffff
2667 && ((~low_bits
) & 0x80000000) != 0))
2669 unsigned HOST_WIDE_INT fast_int
= (~low_bits
& 0xffffffff);
2671 if ((SPARC_SETHI_P (fast_int
)
2672 && (~high_bits
& 0xffffffff) == 0)
2673 || SPARC_SIMM13_P (fast_int
))
2674 emit_insn (gen_safe_SET64 (temp
, fast_int
));
2676 sparc_emit_set_const64 (temp
, GEN_INT (fast_int
));
2681 negated_const
= GEN_INT (((~low_bits
) & 0xfffffc00) |
2682 (((HOST_WIDE_INT
)((~high_bits
) & 0xffffffff))<<32));
2683 sparc_emit_set_const64 (temp
, negated_const
);
2686 /* If we are XOR'ing with -1, then we should emit a one's complement
2687 instead. This way the combiner will notice logical operations
2688 such as ANDN later on and substitute. */
2689 if (trailing_bits
== 0x3ff)
2691 emit_insn (gen_rtx_SET (op0
, gen_rtx_NOT (DImode
, temp
)));
2695 emit_insn (gen_rtx_SET (op0
,
2696 gen_safe_XOR64 (temp
,
2697 (-0x400 | trailing_bits
))));
2702 /* 1) sethi %hi(xxx), %reg
2703 * or %reg, %lo(xxx), %reg
2704 * sllx %reg, yyy, %reg
2706 * ??? This is just a generalized version of the low_bits==0
2707 * thing above, FIXME...
2709 if ((highest_bit_set
- lowest_bit_set
) < 32)
2711 unsigned HOST_WIDE_INT focus_bits
=
2712 create_simple_focus_bits (high_bits
, low_bits
,
2715 /* We can't get here in this state. */
2716 gcc_assert (highest_bit_set
>= 32 && lowest_bit_set
< 32);
2718 /* So what we know is that the set bits straddle the
2719 middle of the 64-bit word. */
2720 sparc_emit_set_const64_quick2 (op0
, temp
,
2726 /* 1) sethi %hi(high_bits), %reg
2727 * or %reg, %lo(high_bits), %reg
2728 * sllx %reg, 32, %reg
2729 * or %reg, low_bits, %reg
2731 if (SPARC_SIMM13_P (low_bits
) && ((int)low_bits
> 0))
2733 sparc_emit_set_const64_quick2 (op0
, temp
, high_bits
, low_bits
, 32);
2737 /* The easiest way when all else fails, is full decomposition. */
2738 sparc_emit_set_const64_longway (op0
, temp
, high_bits
, low_bits
);
2741 /* Given a comparison code (EQ, NE, etc.) and the first operand of a COMPARE,
2742 return the mode to be used for the comparison. For floating-point,
2743 CCFP[E]mode is used. CC_NOOVmode should be used when the first operand
2744 is a PLUS, MINUS, NEG, or ASHIFT. CCmode should be used when no special
2745 processing is needed. */
2748 select_cc_mode (enum rtx_code op
, rtx x
, rtx y ATTRIBUTE_UNUSED
)
2750 if (GET_MODE_CLASS (GET_MODE (x
)) == MODE_FLOAT
)
2776 else if (GET_CODE (x
) == PLUS
|| GET_CODE (x
) == MINUS
2777 || GET_CODE (x
) == NEG
|| GET_CODE (x
) == ASHIFT
)
2779 if (TARGET_ARCH64
&& GET_MODE (x
) == DImode
)
2780 return CCX_NOOVmode
;
2786 if (TARGET_ARCH64
&& GET_MODE (x
) == DImode
)
2793 /* Emit the compare insn and return the CC reg for a CODE comparison
2794 with operands X and Y. */
2797 gen_compare_reg_1 (enum rtx_code code
, rtx x
, rtx y
)
2802 if (GET_MODE_CLASS (GET_MODE (x
)) == MODE_CC
)
2805 mode
= SELECT_CC_MODE (code
, x
, y
);
2807 /* ??? We don't have movcc patterns so we cannot generate pseudo regs for the
2808 fcc regs (cse can't tell they're really call clobbered regs and will
2809 remove a duplicate comparison even if there is an intervening function
2810 call - it will then try to reload the cc reg via an int reg which is why
2811 we need the movcc patterns). It is possible to provide the movcc
2812 patterns by using the ldxfsr/stxfsr v9 insns. I tried it: you need two
2813 registers (say %g1,%g5) and it takes about 6 insns. A better fix would be
2814 to tell cse that CCFPE mode registers (even pseudos) are call
2817 /* ??? This is an experiment. Rather than making changes to cse which may
2818 or may not be easy/clean, we do our own cse. This is possible because
2819 we will generate hard registers. Cse knows they're call clobbered (it
2820 doesn't know the same thing about pseudos). If we guess wrong, no big
2821 deal, but if we win, great! */
2823 if (TARGET_V9
&& GET_MODE_CLASS (GET_MODE (x
)) == MODE_FLOAT
)
2824 #if 1 /* experiment */
2827 /* We cycle through the registers to ensure they're all exercised. */
2828 static int next_fcc_reg
= 0;
2829 /* Previous x,y for each fcc reg. */
2830 static rtx prev_args
[4][2];
2832 /* Scan prev_args for x,y. */
2833 for (reg
= 0; reg
< 4; reg
++)
2834 if (prev_args
[reg
][0] == x
&& prev_args
[reg
][1] == y
)
2839 prev_args
[reg
][0] = x
;
2840 prev_args
[reg
][1] = y
;
2841 next_fcc_reg
= (next_fcc_reg
+ 1) & 3;
2843 cc_reg
= gen_rtx_REG (mode
, reg
+ SPARC_FIRST_V9_FCC_REG
);
2846 cc_reg
= gen_reg_rtx (mode
);
2847 #endif /* ! experiment */
2848 else if (GET_MODE_CLASS (GET_MODE (x
)) == MODE_FLOAT
)
2849 cc_reg
= gen_rtx_REG (mode
, SPARC_FCC_REG
);
2851 cc_reg
= gen_rtx_REG (mode
, SPARC_ICC_REG
);
2853 /* We shouldn't get there for TFmode if !TARGET_HARD_QUAD. If we do, this
2854 will only result in an unrecognizable insn so no point in asserting. */
2855 emit_insn (gen_rtx_SET (cc_reg
, gen_rtx_COMPARE (mode
, x
, y
)));
2861 /* Emit the compare insn and return the CC reg for the comparison in CMP. */
2864 gen_compare_reg (rtx cmp
)
2866 return gen_compare_reg_1 (GET_CODE (cmp
), XEXP (cmp
, 0), XEXP (cmp
, 1));
2869 /* This function is used for v9 only.
2870 DEST is the target of the Scc insn.
2871 CODE is the code for an Scc's comparison.
2872 X and Y are the values we compare.
2874 This function is needed to turn
2877 (gt (reg:CCX 100 %icc)
2881 (gt:DI (reg:CCX 100 %icc)
2884 IE: The instruction recognizer needs to see the mode of the comparison to
2885 find the right instruction. We could use "gt:DI" right in the
2886 define_expand, but leaving it out allows us to handle DI, SI, etc. */
2889 gen_v9_scc (rtx dest
, enum rtx_code compare_code
, rtx x
, rtx y
)
2892 && (GET_MODE (x
) == DImode
2893 || GET_MODE (dest
) == DImode
))
2896 /* Try to use the movrCC insns. */
2898 && GET_MODE_CLASS (GET_MODE (x
)) == MODE_INT
2900 && v9_regcmp_p (compare_code
))
2905 /* Special case for op0 != 0. This can be done with one instruction if
2908 if (compare_code
== NE
2909 && GET_MODE (dest
) == DImode
2910 && rtx_equal_p (op0
, dest
))
2912 emit_insn (gen_rtx_SET (dest
,
2913 gen_rtx_IF_THEN_ELSE (DImode
,
2914 gen_rtx_fmt_ee (compare_code
, DImode
,
2921 if (reg_overlap_mentioned_p (dest
, op0
))
2923 /* Handle the case where dest == x.
2924 We "early clobber" the result. */
2925 op0
= gen_reg_rtx (GET_MODE (x
));
2926 emit_move_insn (op0
, x
);
2929 emit_insn (gen_rtx_SET (dest
, const0_rtx
));
2930 if (GET_MODE (op0
) != DImode
)
2932 temp
= gen_reg_rtx (DImode
);
2933 convert_move (temp
, op0
, 0);
2937 emit_insn (gen_rtx_SET (dest
,
2938 gen_rtx_IF_THEN_ELSE (GET_MODE (dest
),
2939 gen_rtx_fmt_ee (compare_code
, DImode
,
2947 x
= gen_compare_reg_1 (compare_code
, x
, y
);
2950 gcc_assert (GET_MODE (x
) != CC_NOOVmode
2951 && GET_MODE (x
) != CCX_NOOVmode
);
2953 emit_insn (gen_rtx_SET (dest
, const0_rtx
));
2954 emit_insn (gen_rtx_SET (dest
,
2955 gen_rtx_IF_THEN_ELSE (GET_MODE (dest
),
2956 gen_rtx_fmt_ee (compare_code
,
2957 GET_MODE (x
), x
, y
),
2958 const1_rtx
, dest
)));
2964 /* Emit an scc insn. For seq, sne, sgeu, and sltu, we can do this
2965 without jumps using the addx/subx instructions. */
2968 emit_scc_insn (rtx operands
[])
2975 /* The quad-word fp compare library routines all return nonzero to indicate
2976 true, which is different from the equivalent libgcc routines, so we must
2977 handle them specially here. */
2978 if (GET_MODE (operands
[2]) == TFmode
&& ! TARGET_HARD_QUAD
)
2980 operands
[1] = sparc_emit_float_lib_cmp (operands
[2], operands
[3],
2981 GET_CODE (operands
[1]));
2982 operands
[2] = XEXP (operands
[1], 0);
2983 operands
[3] = XEXP (operands
[1], 1);
2986 code
= GET_CODE (operands
[1]);
2990 /* For seq/sne on v9 we use the same code as v8 (the addx/subx method has
2991 more applications). The exception to this is "reg != 0" which can
2992 be done in one instruction on v9 (so we do it). */
2995 if (GET_MODE (x
) == SImode
)
2999 pat
= gen_seqsidi_special (operands
[0], x
, y
);
3001 pat
= gen_seqsisi_special (operands
[0], x
, y
);
3005 else if (GET_MODE (x
) == DImode
)
3007 rtx pat
= gen_seqdi_special (operands
[0], x
, y
);
3015 if (GET_MODE (x
) == SImode
)
3019 pat
= gen_snesidi_special (operands
[0], x
, y
);
3021 pat
= gen_snesisi_special (operands
[0], x
, y
);
3025 else if (GET_MODE (x
) == DImode
)
3029 pat
= gen_snedi_special_vis3 (operands
[0], x
, y
);
3031 pat
= gen_snedi_special (operands
[0], x
, y
);
3039 && GET_MODE (x
) == DImode
3041 && (code
== GTU
|| code
== LTU
))
3042 && gen_v9_scc (operands
[0], code
, x
, y
))
3045 /* We can do LTU and GEU using the addx/subx instructions too. And
3046 for GTU/LEU, if both operands are registers swap them and fall
3047 back to the easy case. */
3048 if (code
== GTU
|| code
== LEU
)
3050 if ((GET_CODE (x
) == REG
|| GET_CODE (x
) == SUBREG
)
3051 && (GET_CODE (y
) == REG
|| GET_CODE (y
) == SUBREG
))
3056 code
= swap_condition (code
);
3061 || (!TARGET_VIS3
&& code
== GEU
))
3063 emit_insn (gen_rtx_SET (operands
[0],
3064 gen_rtx_fmt_ee (code
, GET_MODE (operands
[0]),
3065 gen_compare_reg_1 (code
, x
, y
),
3070 /* All the posibilities to use addx/subx based sequences has been
3071 exhausted, try for a 3 instruction sequence using v9 conditional
3073 if (TARGET_V9
&& gen_v9_scc (operands
[0], code
, x
, y
))
3076 /* Nope, do branches. */
3080 /* Emit a conditional jump insn for the v9 architecture using comparison code
3081 CODE and jump target LABEL.
3082 This function exists to take advantage of the v9 brxx insns. */
3085 emit_v9_brxx_insn (enum rtx_code code
, rtx op0
, rtx label
)
3087 emit_jump_insn (gen_rtx_SET (pc_rtx
,
3088 gen_rtx_IF_THEN_ELSE (VOIDmode
,
3089 gen_rtx_fmt_ee (code
, GET_MODE (op0
),
3091 gen_rtx_LABEL_REF (VOIDmode
, label
),
3095 /* Emit a conditional jump insn for the UA2011 architecture using
3096 comparison code CODE and jump target LABEL. This function exists
3097 to take advantage of the UA2011 Compare and Branch insns. */
3100 emit_cbcond_insn (enum rtx_code code
, rtx op0
, rtx op1
, rtx label
)
3104 if_then_else
= gen_rtx_IF_THEN_ELSE (VOIDmode
,
3105 gen_rtx_fmt_ee(code
, GET_MODE(op0
),
3107 gen_rtx_LABEL_REF (VOIDmode
, label
),
3110 emit_jump_insn (gen_rtx_SET (pc_rtx
, if_then_else
));
3114 emit_conditional_branch_insn (rtx operands
[])
3116 /* The quad-word fp compare library routines all return nonzero to indicate
3117 true, which is different from the equivalent libgcc routines, so we must
3118 handle them specially here. */
3119 if (GET_MODE (operands
[1]) == TFmode
&& ! TARGET_HARD_QUAD
)
3121 operands
[0] = sparc_emit_float_lib_cmp (operands
[1], operands
[2],
3122 GET_CODE (operands
[0]));
3123 operands
[1] = XEXP (operands
[0], 0);
3124 operands
[2] = XEXP (operands
[0], 1);
3127 /* If we can tell early on that the comparison is against a constant
3128 that won't fit in the 5-bit signed immediate field of a cbcond,
3129 use one of the other v9 conditional branch sequences. */
3131 && GET_CODE (operands
[1]) == REG
3132 && (GET_MODE (operands
[1]) == SImode
3133 || (TARGET_ARCH64
&& GET_MODE (operands
[1]) == DImode
))
3134 && (GET_CODE (operands
[2]) != CONST_INT
3135 || SPARC_SIMM5_P (INTVAL (operands
[2]))))
3137 emit_cbcond_insn (GET_CODE (operands
[0]), operands
[1], operands
[2], operands
[3]);
3141 if (TARGET_ARCH64
&& operands
[2] == const0_rtx
3142 && GET_CODE (operands
[1]) == REG
3143 && GET_MODE (operands
[1]) == DImode
)
3145 emit_v9_brxx_insn (GET_CODE (operands
[0]), operands
[1], operands
[3]);
3149 operands
[1] = gen_compare_reg (operands
[0]);
3150 operands
[2] = const0_rtx
;
3151 operands
[0] = gen_rtx_fmt_ee (GET_CODE (operands
[0]), VOIDmode
,
3152 operands
[1], operands
[2]);
3153 emit_jump_insn (gen_cbranchcc4 (operands
[0], operands
[1], operands
[2],
3158 /* Generate a DFmode part of a hard TFmode register.
3159 REG is the TFmode hard register, LOW is 1 for the
3160 low 64bit of the register and 0 otherwise.
3163 gen_df_reg (rtx reg
, int low
)
3165 int regno
= REGNO (reg
);
3167 if ((WORDS_BIG_ENDIAN
== 0) ^ (low
!= 0))
3168 regno
+= (TARGET_ARCH64
&& SPARC_INT_REG_P (regno
)) ? 1 : 2;
3169 return gen_rtx_REG (DFmode
, regno
);
3172 /* Generate a call to FUNC with OPERANDS. Operand 0 is the return value.
3173 Unlike normal calls, TFmode operands are passed by reference. It is
3174 assumed that no more than 3 operands are required. */
3177 emit_soft_tfmode_libcall (const char *func_name
, int nargs
, rtx
*operands
)
3179 rtx ret_slot
= NULL
, arg
[3], func_sym
;
3182 /* We only expect to be called for conversions, unary, and binary ops. */
3183 gcc_assert (nargs
== 2 || nargs
== 3);
3185 for (i
= 0; i
< nargs
; ++i
)
3187 rtx this_arg
= operands
[i
];
3190 /* TFmode arguments and return values are passed by reference. */
3191 if (GET_MODE (this_arg
) == TFmode
)
3193 int force_stack_temp
;
3195 force_stack_temp
= 0;
3196 if (TARGET_BUGGY_QP_LIB
&& i
== 0)
3197 force_stack_temp
= 1;
3199 if (GET_CODE (this_arg
) == MEM
3200 && ! force_stack_temp
)
3202 tree expr
= MEM_EXPR (this_arg
);
3204 mark_addressable (expr
);
3205 this_arg
= XEXP (this_arg
, 0);
3207 else if (CONSTANT_P (this_arg
)
3208 && ! force_stack_temp
)
3210 this_slot
= force_const_mem (TFmode
, this_arg
);
3211 this_arg
= XEXP (this_slot
, 0);
3215 this_slot
= assign_stack_temp (TFmode
, GET_MODE_SIZE (TFmode
));
3217 /* Operand 0 is the return value. We'll copy it out later. */
3219 emit_move_insn (this_slot
, this_arg
);
3221 ret_slot
= this_slot
;
3223 this_arg
= XEXP (this_slot
, 0);
3230 func_sym
= gen_rtx_SYMBOL_REF (Pmode
, func_name
);
3232 if (GET_MODE (operands
[0]) == TFmode
)
3235 emit_library_call (func_sym
, LCT_NORMAL
, VOIDmode
, 2,
3236 arg
[0], GET_MODE (arg
[0]),
3237 arg
[1], GET_MODE (arg
[1]));
3239 emit_library_call (func_sym
, LCT_NORMAL
, VOIDmode
, 3,
3240 arg
[0], GET_MODE (arg
[0]),
3241 arg
[1], GET_MODE (arg
[1]),
3242 arg
[2], GET_MODE (arg
[2]));
3245 emit_move_insn (operands
[0], ret_slot
);
3251 gcc_assert (nargs
== 2);
3253 ret
= emit_library_call_value (func_sym
, operands
[0], LCT_NORMAL
,
3254 GET_MODE (operands
[0]), 1,
3255 arg
[1], GET_MODE (arg
[1]));
3257 if (ret
!= operands
[0])
3258 emit_move_insn (operands
[0], ret
);
3262 /* Expand soft-float TFmode calls to sparc abi routines. */
3265 emit_soft_tfmode_binop (enum rtx_code code
, rtx
*operands
)
3287 emit_soft_tfmode_libcall (func
, 3, operands
);
3291 emit_soft_tfmode_unop (enum rtx_code code
, rtx
*operands
)
3295 gcc_assert (code
== SQRT
);
3298 emit_soft_tfmode_libcall (func
, 2, operands
);
3302 emit_soft_tfmode_cvt (enum rtx_code code
, rtx
*operands
)
3309 switch (GET_MODE (operands
[1]))
3322 case FLOAT_TRUNCATE
:
3323 switch (GET_MODE (operands
[0]))
3337 switch (GET_MODE (operands
[1]))
3342 operands
[1] = gen_rtx_SIGN_EXTEND (DImode
, operands
[1]);
3352 case UNSIGNED_FLOAT
:
3353 switch (GET_MODE (operands
[1]))
3358 operands
[1] = gen_rtx_ZERO_EXTEND (DImode
, operands
[1]);
3369 switch (GET_MODE (operands
[0]))
3383 switch (GET_MODE (operands
[0]))
3400 emit_soft_tfmode_libcall (func
, 2, operands
);
3403 /* Expand a hard-float tfmode operation. All arguments must be in
3407 emit_hard_tfmode_operation (enum rtx_code code
, rtx
*operands
)
3411 if (GET_RTX_CLASS (code
) == RTX_UNARY
)
3413 operands
[1] = force_reg (GET_MODE (operands
[1]), operands
[1]);
3414 op
= gen_rtx_fmt_e (code
, GET_MODE (operands
[0]), operands
[1]);
3418 operands
[1] = force_reg (GET_MODE (operands
[1]), operands
[1]);
3419 operands
[2] = force_reg (GET_MODE (operands
[2]), operands
[2]);
3420 op
= gen_rtx_fmt_ee (code
, GET_MODE (operands
[0]),
3421 operands
[1], operands
[2]);
3424 if (register_operand (operands
[0], VOIDmode
))
3427 dest
= gen_reg_rtx (GET_MODE (operands
[0]));
3429 emit_insn (gen_rtx_SET (dest
, op
));
3431 if (dest
!= operands
[0])
3432 emit_move_insn (operands
[0], dest
);
3436 emit_tfmode_binop (enum rtx_code code
, rtx
*operands
)
3438 if (TARGET_HARD_QUAD
)
3439 emit_hard_tfmode_operation (code
, operands
);
3441 emit_soft_tfmode_binop (code
, operands
);
3445 emit_tfmode_unop (enum rtx_code code
, rtx
*operands
)
3447 if (TARGET_HARD_QUAD
)
3448 emit_hard_tfmode_operation (code
, operands
);
3450 emit_soft_tfmode_unop (code
, operands
);
3454 emit_tfmode_cvt (enum rtx_code code
, rtx
*operands
)
3456 if (TARGET_HARD_QUAD
)
3457 emit_hard_tfmode_operation (code
, operands
);
3459 emit_soft_tfmode_cvt (code
, operands
);
3462 /* Return nonzero if a branch/jump/call instruction will be emitting
3463 nop into its delay slot. */
3466 empty_delay_slot (rtx_insn
*insn
)
3470 /* If no previous instruction (should not happen), return true. */
3471 if (PREV_INSN (insn
) == NULL
)
3474 seq
= NEXT_INSN (PREV_INSN (insn
));
3475 if (GET_CODE (PATTERN (seq
)) == SEQUENCE
)
3481 /* Return nonzero if we should emit a nop after a cbcond instruction.
3482 The cbcond instruction does not have a delay slot, however there is
3483 a severe performance penalty if a control transfer appears right
3484 after a cbcond. Therefore we emit a nop when we detect this
3488 emit_cbcond_nop (rtx insn
)
3490 rtx next
= next_active_insn (insn
);
3495 if (NONJUMP_INSN_P (next
)
3496 && GET_CODE (PATTERN (next
)) == SEQUENCE
)
3497 next
= XVECEXP (PATTERN (next
), 0, 0);
3498 else if (CALL_P (next
)
3499 && GET_CODE (PATTERN (next
)) == PARALLEL
)
3501 rtx delay
= XVECEXP (PATTERN (next
), 0, 1);
3503 if (GET_CODE (delay
) == RETURN
)
3505 /* It's a sibling call. Do not emit the nop if we're going
3506 to emit something other than the jump itself as the first
3507 instruction of the sibcall sequence. */
3508 if (sparc_leaf_function_p
|| TARGET_FLAT
)
3513 if (NONJUMP_INSN_P (next
))
3519 /* Return nonzero if TRIAL can go into the call delay slot. */
3522 eligible_for_call_delay (rtx_insn
*trial
)
3526 if (get_attr_in_branch_delay (trial
) == IN_BRANCH_DELAY_FALSE
)
3530 call __tls_get_addr, %tgd_call (foo)
3531 add %l7, %o0, %o0, %tgd_add (foo)
3532 while Sun as/ld does not. */
3533 if (TARGET_GNU_TLS
|| !TARGET_TLS
)
3536 pat
= PATTERN (trial
);
3538 /* We must reject tgd_add{32|64}, i.e.
3539 (set (reg) (plus (reg) (unspec [(reg) (symbol_ref)] UNSPEC_TLSGD)))
3540 and tldm_add{32|64}, i.e.
3541 (set (reg) (plus (reg) (unspec [(reg) (symbol_ref)] UNSPEC_TLSLDM)))
3543 if (GET_CODE (pat
) == SET
3544 && GET_CODE (SET_SRC (pat
)) == PLUS
)
3546 rtx unspec
= XEXP (SET_SRC (pat
), 1);
3548 if (GET_CODE (unspec
) == UNSPEC
3549 && (XINT (unspec
, 1) == UNSPEC_TLSGD
3550 || XINT (unspec
, 1) == UNSPEC_TLSLDM
))
3557 /* Return nonzero if TRIAL, an insn, can be combined with a 'restore'
3558 instruction. RETURN_P is true if the v9 variant 'return' is to be
3559 considered in the test too.
3561 TRIAL must be a SET whose destination is a REG appropriate for the
3562 'restore' instruction or, if RETURN_P is true, for the 'return'
3566 eligible_for_restore_insn (rtx trial
, bool return_p
)
3568 rtx pat
= PATTERN (trial
);
3569 rtx src
= SET_SRC (pat
);
3570 bool src_is_freg
= false;
3573 /* Since we now can do moves between float and integer registers when
3574 VIS3 is enabled, we have to catch this case. We can allow such
3575 moves when doing a 'return' however. */
3577 if (GET_CODE (src_reg
) == SUBREG
)
3578 src_reg
= SUBREG_REG (src_reg
);
3579 if (GET_CODE (src_reg
) == REG
3580 && SPARC_FP_REG_P (REGNO (src_reg
)))
3583 /* The 'restore src,%g0,dest' pattern for word mode and below. */
3584 if (GET_MODE_CLASS (GET_MODE (src
)) != MODE_FLOAT
3585 && arith_operand (src
, GET_MODE (src
))
3589 return GET_MODE_SIZE (GET_MODE (src
)) <= GET_MODE_SIZE (DImode
);
3591 return GET_MODE_SIZE (GET_MODE (src
)) <= GET_MODE_SIZE (SImode
);
3594 /* The 'restore src,%g0,dest' pattern for double-word mode. */
3595 else if (GET_MODE_CLASS (GET_MODE (src
)) != MODE_FLOAT
3596 && arith_double_operand (src
, GET_MODE (src
))
3598 return GET_MODE_SIZE (GET_MODE (src
)) <= GET_MODE_SIZE (DImode
);
3600 /* The 'restore src,%g0,dest' pattern for float if no FPU. */
3601 else if (! TARGET_FPU
&& register_operand (src
, SFmode
))
3604 /* The 'restore src,%g0,dest' pattern for double if no FPU. */
3605 else if (! TARGET_FPU
&& TARGET_ARCH64
&& register_operand (src
, DFmode
))
3608 /* If we have the 'return' instruction, anything that does not use
3609 local or output registers and can go into a delay slot wins. */
3610 else if (return_p
&& TARGET_V9
&& !epilogue_renumber (&pat
, 1))
3613 /* The 'restore src1,src2,dest' pattern for SImode. */
3614 else if (GET_CODE (src
) == PLUS
3615 && register_operand (XEXP (src
, 0), SImode
)
3616 && arith_operand (XEXP (src
, 1), SImode
))
3619 /* The 'restore src1,src2,dest' pattern for DImode. */
3620 else if (GET_CODE (src
) == PLUS
3621 && register_operand (XEXP (src
, 0), DImode
)
3622 && arith_double_operand (XEXP (src
, 1), DImode
))
3625 /* The 'restore src1,%lo(src2),dest' pattern. */
3626 else if (GET_CODE (src
) == LO_SUM
3627 && ! TARGET_CM_MEDMID
3628 && ((register_operand (XEXP (src
, 0), SImode
)
3629 && immediate_operand (XEXP (src
, 1), SImode
))
3631 && register_operand (XEXP (src
, 0), DImode
)
3632 && immediate_operand (XEXP (src
, 1), DImode
))))
3635 /* The 'restore src,src,dest' pattern. */
3636 else if (GET_CODE (src
) == ASHIFT
3637 && (register_operand (XEXP (src
, 0), SImode
)
3638 || register_operand (XEXP (src
, 0), DImode
))
3639 && XEXP (src
, 1) == const1_rtx
)
3645 /* Return nonzero if TRIAL can go into the function return's delay slot. */
3648 eligible_for_return_delay (rtx_insn
*trial
)
3653 /* If the function uses __builtin_eh_return, the eh_return machinery
3654 occupies the delay slot. */
3655 if (crtl
->calls_eh_return
)
3658 if (get_attr_in_branch_delay (trial
) == IN_BRANCH_DELAY_FALSE
)
3661 /* In the case of a leaf or flat function, anything can go into the slot. */
3662 if (sparc_leaf_function_p
|| TARGET_FLAT
)
3665 if (!NONJUMP_INSN_P (trial
))
3668 pat
= PATTERN (trial
);
3669 if (GET_CODE (pat
) == PARALLEL
)
3675 for (i
= XVECLEN (pat
, 0) - 1; i
>= 0; i
--)
3677 rtx expr
= XVECEXP (pat
, 0, i
);
3678 if (GET_CODE (expr
) != SET
)
3680 if (GET_CODE (SET_DEST (expr
)) != REG
)
3682 regno
= REGNO (SET_DEST (expr
));
3683 if (regno
>= 8 && regno
< 24)
3686 return !epilogue_renumber (&pat
, 1);
3689 if (GET_CODE (pat
) != SET
)
3692 if (GET_CODE (SET_DEST (pat
)) != REG
)
3695 regno
= REGNO (SET_DEST (pat
));
3697 /* Otherwise, only operations which can be done in tandem with
3698 a `restore' or `return' insn can go into the delay slot. */
3699 if (regno
>= 8 && regno
< 24)
3702 /* If this instruction sets up floating point register and we have a return
3703 instruction, it can probably go in. But restore will not work
3705 if (! SPARC_INT_REG_P (regno
))
3706 return TARGET_V9
&& !epilogue_renumber (&pat
, 1);
3708 return eligible_for_restore_insn (trial
, true);
3711 /* Return nonzero if TRIAL can go into the sibling call's delay slot. */
3714 eligible_for_sibcall_delay (rtx_insn
*trial
)
3718 if (get_attr_in_branch_delay (trial
) == IN_BRANCH_DELAY_FALSE
)
3721 if (!NONJUMP_INSN_P (trial
))
3724 pat
= PATTERN (trial
);
3726 if (sparc_leaf_function_p
|| TARGET_FLAT
)
3728 /* If the tail call is done using the call instruction,
3729 we have to restore %o7 in the delay slot. */
3730 if (LEAF_SIBCALL_SLOT_RESERVED_P
)
3733 /* %g1 is used to build the function address */
3734 if (reg_mentioned_p (gen_rtx_REG (Pmode
, 1), pat
))
3740 if (GET_CODE (pat
) != SET
)
3743 /* Otherwise, only operations which can be done in tandem with
3744 a `restore' insn can go into the delay slot. */
3745 if (GET_CODE (SET_DEST (pat
)) != REG
3746 || (REGNO (SET_DEST (pat
)) >= 8 && REGNO (SET_DEST (pat
)) < 24)
3747 || ! SPARC_INT_REG_P (REGNO (SET_DEST (pat
))))
3750 /* If it mentions %o7, it can't go in, because sibcall will clobber it
3752 if (reg_mentioned_p (gen_rtx_REG (Pmode
, 15), pat
))
3755 return eligible_for_restore_insn (trial
, false);
3758 /* Determine if it's legal to put X into the constant pool. This
3759 is not possible if X contains the address of a symbol that is
3760 not constant (TLS) or not known at final link time (PIC). */
3763 sparc_cannot_force_const_mem (machine_mode mode
, rtx x
)
3765 switch (GET_CODE (x
))
3768 case CONST_WIDE_INT
:
3771 /* Accept all non-symbolic constants. */
3775 /* Labels are OK iff we are non-PIC. */
3776 return flag_pic
!= 0;
3779 /* 'Naked' TLS symbol references are never OK,
3780 non-TLS symbols are OK iff we are non-PIC. */
3781 if (SYMBOL_REF_TLS_MODEL (x
))
3784 return flag_pic
!= 0;
3787 return sparc_cannot_force_const_mem (mode
, XEXP (x
, 0));
3790 return sparc_cannot_force_const_mem (mode
, XEXP (x
, 0))
3791 || sparc_cannot_force_const_mem (mode
, XEXP (x
, 1));
3799 /* Global Offset Table support. */
3800 static GTY(()) rtx got_helper_rtx
= NULL_RTX
;
3801 static GTY(()) rtx global_offset_table_rtx
= NULL_RTX
;
3803 /* Return the SYMBOL_REF for the Global Offset Table. */
3805 static GTY(()) rtx sparc_got_symbol
= NULL_RTX
;
3810 if (!sparc_got_symbol
)
3811 sparc_got_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "_GLOBAL_OFFSET_TABLE_");
3813 return sparc_got_symbol
;
3816 /* Ensure that we are not using patterns that are not OK with PIC. */
3826 op
= recog_data
.operand
[i
];
3827 gcc_assert (GET_CODE (op
) != SYMBOL_REF
3828 && (GET_CODE (op
) != CONST
3829 || (GET_CODE (XEXP (op
, 0)) == MINUS
3830 && XEXP (XEXP (op
, 0), 0) == sparc_got ()
3831 && GET_CODE (XEXP (XEXP (op
, 0), 1)) == CONST
)));
3838 /* Return true if X is an address which needs a temporary register when
3839 reloaded while generating PIC code. */
3842 pic_address_needs_scratch (rtx x
)
3844 /* An address which is a symbolic plus a non SMALL_INT needs a temp reg. */
3845 if (GET_CODE (x
) == CONST
&& GET_CODE (XEXP (x
, 0)) == PLUS
3846 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == SYMBOL_REF
3847 && GET_CODE (XEXP (XEXP (x
, 0), 1)) == CONST_INT
3848 && ! SMALL_INT (XEXP (XEXP (x
, 0), 1)))
3854 /* Determine if a given RTX is a valid constant. We already know this
3855 satisfies CONSTANT_P. */
3858 sparc_legitimate_constant_p (machine_mode mode
, rtx x
)
3860 switch (GET_CODE (x
))
3864 if (sparc_tls_referenced_p (x
))
3869 /* Floating point constants are generally not ok.
3870 The only exception is 0.0 and all-ones in VIS. */
3872 && SCALAR_FLOAT_MODE_P (mode
)
3873 && (const_zero_operand (x
, mode
)
3874 || const_all_ones_operand (x
, mode
)))
3880 /* Vector constants are generally not ok.
3881 The only exception is 0 or -1 in VIS. */
3883 && (const_zero_operand (x
, mode
)
3884 || const_all_ones_operand (x
, mode
)))
3896 /* Determine if a given RTX is a valid constant address. */
3899 constant_address_p (rtx x
)
3901 switch (GET_CODE (x
))
3909 if (flag_pic
&& pic_address_needs_scratch (x
))
3911 return sparc_legitimate_constant_p (Pmode
, x
);
3914 return !flag_pic
&& sparc_legitimate_constant_p (Pmode
, x
);
3921 /* Nonzero if the constant value X is a legitimate general operand
3922 when generating PIC code. It is given that flag_pic is on and
3923 that X satisfies CONSTANT_P. */
3926 legitimate_pic_operand_p (rtx x
)
3928 if (pic_address_needs_scratch (x
))
3930 if (sparc_tls_referenced_p (x
))
3935 #define RTX_OK_FOR_OFFSET_P(X, MODE) \
3937 && INTVAL (X) >= -0x1000 \
3938 && INTVAL (X) <= (0x1000 - GET_MODE_SIZE (MODE)))
3940 #define RTX_OK_FOR_OLO10_P(X, MODE) \
3942 && INTVAL (X) >= -0x1000 \
3943 && INTVAL (X) <= (0xc00 - GET_MODE_SIZE (MODE)))
3945 /* Handle the TARGET_LEGITIMATE_ADDRESS_P target hook.
3947 On SPARC, the actual legitimate addresses must be REG+REG or REG+SMALLINT
3948 ordinarily. This changes a bit when generating PIC. */
3951 sparc_legitimate_address_p (machine_mode mode
, rtx addr
, bool strict
)
3953 rtx rs1
= NULL
, rs2
= NULL
, imm1
= NULL
;
3955 if (REG_P (addr
) || GET_CODE (addr
) == SUBREG
)
3957 else if (GET_CODE (addr
) == PLUS
)
3959 rs1
= XEXP (addr
, 0);
3960 rs2
= XEXP (addr
, 1);
3962 /* Canonicalize. REG comes first, if there are no regs,
3963 LO_SUM comes first. */
3965 && GET_CODE (rs1
) != SUBREG
3967 || GET_CODE (rs2
) == SUBREG
3968 || (GET_CODE (rs2
) == LO_SUM
&& GET_CODE (rs1
) != LO_SUM
)))
3970 rs1
= XEXP (addr
, 1);
3971 rs2
= XEXP (addr
, 0);
3975 && rs1
== pic_offset_table_rtx
3977 && GET_CODE (rs2
) != SUBREG
3978 && GET_CODE (rs2
) != LO_SUM
3979 && GET_CODE (rs2
) != MEM
3980 && !(GET_CODE (rs2
) == SYMBOL_REF
&& SYMBOL_REF_TLS_MODEL (rs2
))
3981 && (! symbolic_operand (rs2
, VOIDmode
) || mode
== Pmode
)
3982 && (GET_CODE (rs2
) != CONST_INT
|| SMALL_INT (rs2
)))
3984 || GET_CODE (rs1
) == SUBREG
)
3985 && RTX_OK_FOR_OFFSET_P (rs2
, mode
)))
3990 else if ((REG_P (rs1
) || GET_CODE (rs1
) == SUBREG
)
3991 && (REG_P (rs2
) || GET_CODE (rs2
) == SUBREG
))
3993 /* We prohibit REG + REG for TFmode when there are no quad move insns
3994 and we consequently need to split. We do this because REG+REG
3995 is not an offsettable address. If we get the situation in reload
3996 where source and destination of a movtf pattern are both MEMs with
3997 REG+REG address, then only one of them gets converted to an
3998 offsettable address. */
4000 && ! (TARGET_ARCH64
&& TARGET_HARD_QUAD
))
4003 /* Likewise for TImode, but in all cases. */
4007 /* We prohibit REG + REG on ARCH32 if not optimizing for
4008 DFmode/DImode because then mem_min_alignment is likely to be zero
4009 after reload and the forced split would lack a matching splitter
4011 if (TARGET_ARCH32
&& !optimize
4012 && (mode
== DFmode
|| mode
== DImode
))
4015 else if (USE_AS_OFFSETABLE_LO10
4016 && GET_CODE (rs1
) == LO_SUM
4018 && ! TARGET_CM_MEDMID
4019 && RTX_OK_FOR_OLO10_P (rs2
, mode
))
4022 imm1
= XEXP (rs1
, 1);
4023 rs1
= XEXP (rs1
, 0);
4024 if (!CONSTANT_P (imm1
)
4025 || (GET_CODE (rs1
) == SYMBOL_REF
&& SYMBOL_REF_TLS_MODEL (rs1
)))
4029 else if (GET_CODE (addr
) == LO_SUM
)
4031 rs1
= XEXP (addr
, 0);
4032 imm1
= XEXP (addr
, 1);
4034 if (!CONSTANT_P (imm1
)
4035 || (GET_CODE (rs1
) == SYMBOL_REF
&& SYMBOL_REF_TLS_MODEL (rs1
)))
4038 /* We can't allow TFmode in 32-bit mode, because an offset greater
4039 than the alignment (8) may cause the LO_SUM to overflow. */
4040 if (mode
== TFmode
&& TARGET_ARCH32
)
4043 else if (GET_CODE (addr
) == CONST_INT
&& SMALL_INT (addr
))
4048 if (GET_CODE (rs1
) == SUBREG
)
4049 rs1
= SUBREG_REG (rs1
);
4055 if (GET_CODE (rs2
) == SUBREG
)
4056 rs2
= SUBREG_REG (rs2
);
4063 if (!REGNO_OK_FOR_BASE_P (REGNO (rs1
))
4064 || (rs2
&& !REGNO_OK_FOR_BASE_P (REGNO (rs2
))))
4069 if ((! SPARC_INT_REG_P (REGNO (rs1
))
4070 && REGNO (rs1
) != FRAME_POINTER_REGNUM
4071 && REGNO (rs1
) < FIRST_PSEUDO_REGISTER
)
4073 && (! SPARC_INT_REG_P (REGNO (rs2
))
4074 && REGNO (rs2
) != FRAME_POINTER_REGNUM
4075 && REGNO (rs2
) < FIRST_PSEUDO_REGISTER
)))
4081 /* Return the SYMBOL_REF for the tls_get_addr function. */
4083 static GTY(()) rtx sparc_tls_symbol
= NULL_RTX
;
4086 sparc_tls_get_addr (void)
4088 if (!sparc_tls_symbol
)
4089 sparc_tls_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "__tls_get_addr");
4091 return sparc_tls_symbol
;
4094 /* Return the Global Offset Table to be used in TLS mode. */
4097 sparc_tls_got (void)
4099 /* In PIC mode, this is just the PIC offset table. */
4102 crtl
->uses_pic_offset_table
= 1;
4103 return pic_offset_table_rtx
;
4106 /* In non-PIC mode, Sun as (unlike GNU as) emits PC-relative relocations for
4107 the GOT symbol with the 32-bit ABI, so we reload the GOT register. */
4108 if (TARGET_SUN_TLS
&& TARGET_ARCH32
)
4110 load_got_register ();
4111 return global_offset_table_rtx
;
4114 /* In all other cases, we load a new pseudo with the GOT symbol. */
4115 return copy_to_reg (sparc_got ());
4118 /* Return true if X contains a thread-local symbol. */
4121 sparc_tls_referenced_p (rtx x
)
4123 if (!TARGET_HAVE_TLS
)
4126 if (GET_CODE (x
) == CONST
&& GET_CODE (XEXP (x
, 0)) == PLUS
)
4127 x
= XEXP (XEXP (x
, 0), 0);
4129 if (GET_CODE (x
) == SYMBOL_REF
&& SYMBOL_REF_TLS_MODEL (x
))
4132 /* That's all we handle in sparc_legitimize_tls_address for now. */
4136 /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute
4137 this (thread-local) address. */
4140 sparc_legitimize_tls_address (rtx addr
)
4142 rtx temp1
, temp2
, temp3
, ret
, o0
, got
;
4145 gcc_assert (can_create_pseudo_p ());
4147 if (GET_CODE (addr
) == SYMBOL_REF
)
4148 switch (SYMBOL_REF_TLS_MODEL (addr
))
4150 case TLS_MODEL_GLOBAL_DYNAMIC
:
4152 temp1
= gen_reg_rtx (SImode
);
4153 temp2
= gen_reg_rtx (SImode
);
4154 ret
= gen_reg_rtx (Pmode
);
4155 o0
= gen_rtx_REG (Pmode
, 8);
4156 got
= sparc_tls_got ();
4157 emit_insn (gen_tgd_hi22 (temp1
, addr
));
4158 emit_insn (gen_tgd_lo10 (temp2
, temp1
, addr
));
4161 emit_insn (gen_tgd_add32 (o0
, got
, temp2
, addr
));
4162 insn
= emit_call_insn (gen_tgd_call32 (o0
, sparc_tls_get_addr (),
4167 emit_insn (gen_tgd_add64 (o0
, got
, temp2
, addr
));
4168 insn
= emit_call_insn (gen_tgd_call64 (o0
, sparc_tls_get_addr (),
4171 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), o0
);
4172 insn
= get_insns ();
4174 emit_libcall_block (insn
, ret
, o0
, addr
);
4177 case TLS_MODEL_LOCAL_DYNAMIC
:
4179 temp1
= gen_reg_rtx (SImode
);
4180 temp2
= gen_reg_rtx (SImode
);
4181 temp3
= gen_reg_rtx (Pmode
);
4182 ret
= gen_reg_rtx (Pmode
);
4183 o0
= gen_rtx_REG (Pmode
, 8);
4184 got
= sparc_tls_got ();
4185 emit_insn (gen_tldm_hi22 (temp1
));
4186 emit_insn (gen_tldm_lo10 (temp2
, temp1
));
4189 emit_insn (gen_tldm_add32 (o0
, got
, temp2
));
4190 insn
= emit_call_insn (gen_tldm_call32 (o0
, sparc_tls_get_addr (),
4195 emit_insn (gen_tldm_add64 (o0
, got
, temp2
));
4196 insn
= emit_call_insn (gen_tldm_call64 (o0
, sparc_tls_get_addr (),
4199 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), o0
);
4200 insn
= get_insns ();
4202 emit_libcall_block (insn
, temp3
, o0
,
4203 gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, const0_rtx
),
4204 UNSPEC_TLSLD_BASE
));
4205 temp1
= gen_reg_rtx (SImode
);
4206 temp2
= gen_reg_rtx (SImode
);
4207 emit_insn (gen_tldo_hix22 (temp1
, addr
));
4208 emit_insn (gen_tldo_lox10 (temp2
, temp1
, addr
));
4210 emit_insn (gen_tldo_add32 (ret
, temp3
, temp2
, addr
));
4212 emit_insn (gen_tldo_add64 (ret
, temp3
, temp2
, addr
));
4215 case TLS_MODEL_INITIAL_EXEC
:
4216 temp1
= gen_reg_rtx (SImode
);
4217 temp2
= gen_reg_rtx (SImode
);
4218 temp3
= gen_reg_rtx (Pmode
);
4219 got
= sparc_tls_got ();
4220 emit_insn (gen_tie_hi22 (temp1
, addr
));
4221 emit_insn (gen_tie_lo10 (temp2
, temp1
, addr
));
4223 emit_insn (gen_tie_ld32 (temp3
, got
, temp2
, addr
));
4225 emit_insn (gen_tie_ld64 (temp3
, got
, temp2
, addr
));
4228 ret
= gen_reg_rtx (Pmode
);
4230 emit_insn (gen_tie_add32 (ret
, gen_rtx_REG (Pmode
, 7),
4233 emit_insn (gen_tie_add64 (ret
, gen_rtx_REG (Pmode
, 7),
4237 ret
= gen_rtx_PLUS (Pmode
, gen_rtx_REG (Pmode
, 7), temp3
);
4240 case TLS_MODEL_LOCAL_EXEC
:
4241 temp1
= gen_reg_rtx (Pmode
);
4242 temp2
= gen_reg_rtx (Pmode
);
4245 emit_insn (gen_tle_hix22_sp32 (temp1
, addr
));
4246 emit_insn (gen_tle_lox10_sp32 (temp2
, temp1
, addr
));
4250 emit_insn (gen_tle_hix22_sp64 (temp1
, addr
));
4251 emit_insn (gen_tle_lox10_sp64 (temp2
, temp1
, addr
));
4253 ret
= gen_rtx_PLUS (Pmode
, gen_rtx_REG (Pmode
, 7), temp2
);
4260 else if (GET_CODE (addr
) == CONST
)
4264 gcc_assert (GET_CODE (XEXP (addr
, 0)) == PLUS
);
4266 base
= sparc_legitimize_tls_address (XEXP (XEXP (addr
, 0), 0));
4267 offset
= XEXP (XEXP (addr
, 0), 1);
4269 base
= force_operand (base
, NULL_RTX
);
4270 if (!(GET_CODE (offset
) == CONST_INT
&& SMALL_INT (offset
)))
4271 offset
= force_reg (Pmode
, offset
);
4272 ret
= gen_rtx_PLUS (Pmode
, base
, offset
);
4276 gcc_unreachable (); /* for now ... */
4281 /* Legitimize PIC addresses. If the address is already position-independent,
4282 we return ORIG. Newly generated position-independent addresses go into a
4283 reg. This is REG if nonzero, otherwise we allocate register(s) as
4287 sparc_legitimize_pic_address (rtx orig
, rtx reg
)
4289 bool gotdata_op
= false;
4291 if (GET_CODE (orig
) == SYMBOL_REF
4292 /* See the comment in sparc_expand_move. */
4293 || (GET_CODE (orig
) == LABEL_REF
&& !can_use_mov_pic_label_ref (orig
)))
4295 rtx pic_ref
, address
;
4300 gcc_assert (can_create_pseudo_p ());
4301 reg
= gen_reg_rtx (Pmode
);
4306 /* If not during reload, allocate another temp reg here for loading
4307 in the address, so that these instructions can be optimized
4309 rtx temp_reg
= (! can_create_pseudo_p ()
4310 ? reg
: gen_reg_rtx (Pmode
));
4312 /* Must put the SYMBOL_REF inside an UNSPEC here so that cse
4313 won't get confused into thinking that these two instructions
4314 are loading in the true address of the symbol. If in the
4315 future a PIC rtx exists, that should be used instead. */
4318 emit_insn (gen_movdi_high_pic (temp_reg
, orig
));
4319 emit_insn (gen_movdi_lo_sum_pic (temp_reg
, temp_reg
, orig
));
4323 emit_insn (gen_movsi_high_pic (temp_reg
, orig
));
4324 emit_insn (gen_movsi_lo_sum_pic (temp_reg
, temp_reg
, orig
));
4332 crtl
->uses_pic_offset_table
= 1;
4336 insn
= emit_insn (gen_movdi_pic_gotdata_op (reg
,
4337 pic_offset_table_rtx
,
4340 insn
= emit_insn (gen_movsi_pic_gotdata_op (reg
,
4341 pic_offset_table_rtx
,
4347 = gen_const_mem (Pmode
,
4348 gen_rtx_PLUS (Pmode
,
4349 pic_offset_table_rtx
, address
));
4350 insn
= emit_move_insn (reg
, pic_ref
);
4353 /* Put a REG_EQUAL note on this insn, so that it can be optimized
4355 set_unique_reg_note (insn
, REG_EQUAL
, orig
);
4358 else if (GET_CODE (orig
) == CONST
)
4362 if (GET_CODE (XEXP (orig
, 0)) == PLUS
4363 && XEXP (XEXP (orig
, 0), 0) == pic_offset_table_rtx
)
4368 gcc_assert (can_create_pseudo_p ());
4369 reg
= gen_reg_rtx (Pmode
);
4372 gcc_assert (GET_CODE (XEXP (orig
, 0)) == PLUS
);
4373 base
= sparc_legitimize_pic_address (XEXP (XEXP (orig
, 0), 0), reg
);
4374 offset
= sparc_legitimize_pic_address (XEXP (XEXP (orig
, 0), 1),
4375 base
== reg
? NULL_RTX
: reg
);
4377 if (GET_CODE (offset
) == CONST_INT
)
4379 if (SMALL_INT (offset
))
4380 return plus_constant (Pmode
, base
, INTVAL (offset
));
4381 else if (can_create_pseudo_p ())
4382 offset
= force_reg (Pmode
, offset
);
4384 /* If we reach here, then something is seriously wrong. */
4387 return gen_rtx_PLUS (Pmode
, base
, offset
);
4389 else if (GET_CODE (orig
) == LABEL_REF
)
4390 /* ??? We ought to be checking that the register is live instead, in case
4391 it is eliminated. */
4392 crtl
->uses_pic_offset_table
= 1;
4397 /* Try machine-dependent ways of modifying an illegitimate address X
4398 to be legitimate. If we find one, return the new, valid address.
4400 OLDX is the address as it was before break_out_memory_refs was called.
4401 In some cases it is useful to look at this to decide what needs to be done.
4403 MODE is the mode of the operand pointed to by X.
4405 On SPARC, change REG+N into REG+REG, and REG+(X*Y) into REG+REG. */
4408 sparc_legitimize_address (rtx x
, rtx oldx ATTRIBUTE_UNUSED
,
4413 if (GET_CODE (x
) == PLUS
&& GET_CODE (XEXP (x
, 0)) == MULT
)
4414 x
= gen_rtx_PLUS (Pmode
, XEXP (x
, 1),
4415 force_operand (XEXP (x
, 0), NULL_RTX
));
4416 if (GET_CODE (x
) == PLUS
&& GET_CODE (XEXP (x
, 1)) == MULT
)
4417 x
= gen_rtx_PLUS (Pmode
, XEXP (x
, 0),
4418 force_operand (XEXP (x
, 1), NULL_RTX
));
4419 if (GET_CODE (x
) == PLUS
&& GET_CODE (XEXP (x
, 0)) == PLUS
)
4420 x
= gen_rtx_PLUS (Pmode
, force_operand (XEXP (x
, 0), NULL_RTX
),
4422 if (GET_CODE (x
) == PLUS
&& GET_CODE (XEXP (x
, 1)) == PLUS
)
4423 x
= gen_rtx_PLUS (Pmode
, XEXP (x
, 0),
4424 force_operand (XEXP (x
, 1), NULL_RTX
));
4426 if (x
!= orig_x
&& sparc_legitimate_address_p (mode
, x
, FALSE
))
4429 if (sparc_tls_referenced_p (x
))
4430 x
= sparc_legitimize_tls_address (x
);
4432 x
= sparc_legitimize_pic_address (x
, NULL_RTX
);
4433 else if (GET_CODE (x
) == PLUS
&& CONSTANT_ADDRESS_P (XEXP (x
, 1)))
4434 x
= gen_rtx_PLUS (Pmode
, XEXP (x
, 0),
4435 copy_to_mode_reg (Pmode
, XEXP (x
, 1)));
4436 else if (GET_CODE (x
) == PLUS
&& CONSTANT_ADDRESS_P (XEXP (x
, 0)))
4437 x
= gen_rtx_PLUS (Pmode
, XEXP (x
, 1),
4438 copy_to_mode_reg (Pmode
, XEXP (x
, 0)));
4439 else if (GET_CODE (x
) == SYMBOL_REF
4440 || GET_CODE (x
) == CONST
4441 || GET_CODE (x
) == LABEL_REF
)
4442 x
= copy_to_suggested_reg (x
, NULL_RTX
, Pmode
);
4447 /* Delegitimize an address that was legitimized by the above function. */
4450 sparc_delegitimize_address (rtx x
)
4452 x
= delegitimize_mem_from_attrs (x
);
4454 if (GET_CODE (x
) == LO_SUM
&& GET_CODE (XEXP (x
, 1)) == UNSPEC
)
4455 switch (XINT (XEXP (x
, 1), 1))
4457 case UNSPEC_MOVE_PIC
:
4459 x
= XVECEXP (XEXP (x
, 1), 0, 0);
4460 gcc_assert (GET_CODE (x
) == SYMBOL_REF
);
4466 /* This is generated by mov{si,di}_pic_label_ref in PIC mode. */
4467 if (GET_CODE (x
) == MINUS
4468 && REG_P (XEXP (x
, 0))
4469 && REGNO (XEXP (x
, 0)) == PIC_OFFSET_TABLE_REGNUM
4470 && GET_CODE (XEXP (x
, 1)) == LO_SUM
4471 && GET_CODE (XEXP (XEXP (x
, 1), 1)) == UNSPEC
4472 && XINT (XEXP (XEXP (x
, 1), 1), 1) == UNSPEC_MOVE_PIC_LABEL
)
4474 x
= XVECEXP (XEXP (XEXP (x
, 1), 1), 0, 0);
4475 gcc_assert (GET_CODE (x
) == LABEL_REF
);
4481 /* SPARC implementation of LEGITIMIZE_RELOAD_ADDRESS. Returns a value to
4482 replace the input X, or the original X if no replacement is called for.
4483 The output parameter *WIN is 1 if the calling macro should goto WIN,
4486 For SPARC, we wish to handle addresses by splitting them into
4487 HIGH+LO_SUM pairs, retaining the LO_SUM in the memory reference.
4488 This cuts the number of extra insns by one.
4490 Do nothing when generating PIC code and the address is a symbolic
4491 operand or requires a scratch register. */
4494 sparc_legitimize_reload_address (rtx x
, machine_mode mode
,
4495 int opnum
, int type
,
4496 int ind_levels ATTRIBUTE_UNUSED
, int *win
)
4498 /* Decompose SImode constants into HIGH+LO_SUM. */
4500 && (mode
!= TFmode
|| TARGET_ARCH64
)
4501 && GET_MODE (x
) == SImode
4502 && GET_CODE (x
) != LO_SUM
4503 && GET_CODE (x
) != HIGH
4504 && sparc_cmodel
<= CM_MEDLOW
4506 && (symbolic_operand (x
, Pmode
) || pic_address_needs_scratch (x
))))
4508 x
= gen_rtx_LO_SUM (GET_MODE (x
), gen_rtx_HIGH (GET_MODE (x
), x
), x
);
4509 push_reload (XEXP (x
, 0), NULL_RTX
, &XEXP (x
, 0), NULL
,
4510 BASE_REG_CLASS
, GET_MODE (x
), VOIDmode
, 0, 0,
4511 opnum
, (enum reload_type
)type
);
4516 /* We have to recognize what we have already generated above. */
4517 if (GET_CODE (x
) == LO_SUM
&& GET_CODE (XEXP (x
, 0)) == HIGH
)
4519 push_reload (XEXP (x
, 0), NULL_RTX
, &XEXP (x
, 0), NULL
,
4520 BASE_REG_CLASS
, GET_MODE (x
), VOIDmode
, 0, 0,
4521 opnum
, (enum reload_type
)type
);
4530 /* Return true if ADDR (a legitimate address expression)
4531 has an effect that depends on the machine mode it is used for.
4537 is not equivalent to
4539 (mem:QI [%l7+a]) (mem:QI [%l7+a+1])
4541 because [%l7+a+1] is interpreted as the address of (a+1). */
4545 sparc_mode_dependent_address_p (const_rtx addr
,
4546 addr_space_t as ATTRIBUTE_UNUSED
)
4548 if (flag_pic
&& GET_CODE (addr
) == PLUS
)
4550 rtx op0
= XEXP (addr
, 0);
4551 rtx op1
= XEXP (addr
, 1);
4552 if (op0
== pic_offset_table_rtx
4553 && symbolic_operand (op1
, VOIDmode
))
4560 #ifdef HAVE_GAS_HIDDEN
4561 # define USE_HIDDEN_LINKONCE 1
4563 # define USE_HIDDEN_LINKONCE 0
4567 get_pc_thunk_name (char name
[32], unsigned int regno
)
4569 const char *reg_name
= reg_names
[regno
];
4571 /* Skip the leading '%' as that cannot be used in a
4575 if (USE_HIDDEN_LINKONCE
)
4576 sprintf (name
, "__sparc_get_pc_thunk.%s", reg_name
);
4578 ASM_GENERATE_INTERNAL_LABEL (name
, "LADDPC", regno
);
4581 /* Wrapper around the load_pcrel_sym{si,di} patterns. */
4584 gen_load_pcrel_sym (rtx op0
, rtx op1
, rtx op2
, rtx op3
)
4586 int orig_flag_pic
= flag_pic
;
4589 /* The load_pcrel_sym{si,di} patterns require absolute addressing. */
4592 insn
= gen_load_pcrel_symdi (op0
, op1
, op2
, op3
);
4594 insn
= gen_load_pcrel_symsi (op0
, op1
, op2
, op3
);
4595 flag_pic
= orig_flag_pic
;
4600 /* Emit code to load the GOT register. */
4603 load_got_register (void)
4605 /* In PIC mode, this will retrieve pic_offset_table_rtx. */
4606 if (!global_offset_table_rtx
)
4607 global_offset_table_rtx
= gen_rtx_REG (Pmode
, GLOBAL_OFFSET_TABLE_REGNUM
);
4609 if (TARGET_VXWORKS_RTP
)
4610 emit_insn (gen_vxworks_load_got ());
4613 /* The GOT symbol is subject to a PC-relative relocation so we need a
4614 helper function to add the PC value and thus get the final value. */
4615 if (!got_helper_rtx
)
4618 get_pc_thunk_name (name
, GLOBAL_OFFSET_TABLE_REGNUM
);
4619 got_helper_rtx
= gen_rtx_SYMBOL_REF (Pmode
, ggc_strdup (name
));
4622 emit_insn (gen_load_pcrel_sym (global_offset_table_rtx
, sparc_got (),
4624 GEN_INT (GLOBAL_OFFSET_TABLE_REGNUM
)));
4627 /* Need to emit this whether or not we obey regdecls,
4628 since setjmp/longjmp can cause life info to screw up.
4629 ??? In the case where we don't obey regdecls, this is not sufficient
4630 since we may not fall out the bottom. */
4631 emit_use (global_offset_table_rtx
);
4634 /* Emit a call instruction with the pattern given by PAT. ADDR is the
4635 address of the call target. */
4638 sparc_emit_call_insn (rtx pat
, rtx addr
)
4642 insn
= emit_call_insn (pat
);
4644 /* The PIC register is live on entry to VxWorks PIC PLT entries. */
4645 if (TARGET_VXWORKS_RTP
4647 && GET_CODE (addr
) == SYMBOL_REF
4648 && (SYMBOL_REF_DECL (addr
)
4649 ? !targetm
.binds_local_p (SYMBOL_REF_DECL (addr
))
4650 : !SYMBOL_REF_LOCAL_P (addr
)))
4652 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), pic_offset_table_rtx
);
4653 crtl
->uses_pic_offset_table
= 1;
4657 /* Return 1 if RTX is a MEM which is known to be aligned to at
4658 least a DESIRED byte boundary. */
4661 mem_min_alignment (rtx mem
, int desired
)
4663 rtx addr
, base
, offset
;
4665 /* If it's not a MEM we can't accept it. */
4666 if (GET_CODE (mem
) != MEM
)
4670 if (!TARGET_UNALIGNED_DOUBLES
4671 && MEM_ALIGN (mem
) / BITS_PER_UNIT
>= (unsigned)desired
)
4674 /* ??? The rest of the function predates MEM_ALIGN so
4675 there is probably a bit of redundancy. */
4676 addr
= XEXP (mem
, 0);
4677 base
= offset
= NULL_RTX
;
4678 if (GET_CODE (addr
) == PLUS
)
4680 if (GET_CODE (XEXP (addr
, 0)) == REG
)
4682 base
= XEXP (addr
, 0);
4684 /* What we are saying here is that if the base
4685 REG is aligned properly, the compiler will make
4686 sure any REG based index upon it will be so
4688 if (GET_CODE (XEXP (addr
, 1)) == CONST_INT
)
4689 offset
= XEXP (addr
, 1);
4691 offset
= const0_rtx
;
4694 else if (GET_CODE (addr
) == REG
)
4697 offset
= const0_rtx
;
4700 if (base
!= NULL_RTX
)
4702 int regno
= REGNO (base
);
4704 if (regno
!= HARD_FRAME_POINTER_REGNUM
&& regno
!= STACK_POINTER_REGNUM
)
4706 /* Check if the compiler has recorded some information
4707 about the alignment of the base REG. If reload has
4708 completed, we already matched with proper alignments.
4709 If not running global_alloc, reload might give us
4710 unaligned pointer to local stack though. */
4712 && REGNO_POINTER_ALIGN (regno
) >= desired
* BITS_PER_UNIT
)
4713 || (optimize
&& reload_completed
))
4714 && (INTVAL (offset
) & (desired
- 1)) == 0)
4719 if (((INTVAL (offset
) - SPARC_STACK_BIAS
) & (desired
- 1)) == 0)
4723 else if (! TARGET_UNALIGNED_DOUBLES
4724 || CONSTANT_P (addr
)
4725 || GET_CODE (addr
) == LO_SUM
)
4727 /* Anything else we know is properly aligned unless TARGET_UNALIGNED_DOUBLES
4728 is true, in which case we can only assume that an access is aligned if
4729 it is to a constant address, or the address involves a LO_SUM. */
4733 /* An obviously unaligned address. */
4738 /* Vectors to keep interesting information about registers where it can easily
4739 be got. We used to use the actual mode value as the bit number, but there
4740 are more than 32 modes now. Instead we use two tables: one indexed by
4741 hard register number, and one indexed by mode. */
4743 /* The purpose of sparc_mode_class is to shrink the range of modes so that
4744 they all fit (as bit numbers) in a 32-bit word (again). Each real mode is
4745 mapped into one sparc_mode_class mode. */
4747 enum sparc_mode_class
{
4748 H_MODE
, S_MODE
, D_MODE
, T_MODE
, O_MODE
,
4749 SF_MODE
, DF_MODE
, TF_MODE
, OF_MODE
,
4753 /* Modes for single-word and smaller quantities. */
4755 ((1 << (int) H_MODE) | (1 << (int) S_MODE) | (1 << (int) SF_MODE))
4757 /* Modes for double-word and smaller quantities. */
4758 #define D_MODES (S_MODES | (1 << (int) D_MODE) | (1 << DF_MODE))
4760 /* Modes for quad-word and smaller quantities. */
4761 #define T_MODES (D_MODES | (1 << (int) T_MODE) | (1 << (int) TF_MODE))
4763 /* Modes for 8-word and smaller quantities. */
4764 #define O_MODES (T_MODES | (1 << (int) O_MODE) | (1 << (int) OF_MODE))
4766 /* Modes for single-float quantities. */
4767 #define SF_MODES ((1 << (int) S_MODE) | (1 << (int) SF_MODE))
4769 /* Modes for double-float and smaller quantities. */
4770 #define DF_MODES (SF_MODES | (1 << (int) D_MODE) | (1 << DF_MODE))
4772 /* Modes for quad-float and smaller quantities. */
4773 #define TF_MODES (DF_MODES | (1 << (int) TF_MODE))
4775 /* Modes for quad-float pairs and smaller quantities. */
4776 #define OF_MODES (TF_MODES | (1 << (int) OF_MODE))
4778 /* Modes for double-float only quantities. */
4779 #define DF_MODES_NO_S ((1 << (int) D_MODE) | (1 << (int) DF_MODE))
4781 /* Modes for quad-float and double-float only quantities. */
4782 #define TF_MODES_NO_S (DF_MODES_NO_S | (1 << (int) TF_MODE))
4784 /* Modes for quad-float pairs and double-float only quantities. */
4785 #define OF_MODES_NO_S (TF_MODES_NO_S | (1 << (int) OF_MODE))
4787 /* Modes for condition codes. */
4788 #define CC_MODES (1 << (int) CC_MODE)
4789 #define CCFP_MODES (1 << (int) CCFP_MODE)
4791 /* Value is 1 if register/mode pair is acceptable on sparc.
4793 The funny mixture of D and T modes is because integer operations
4794 do not specially operate on tetra quantities, so non-quad-aligned
4795 registers can hold quadword quantities (except %o4 and %i4 because
4796 they cross fixed registers).
4798 ??? Note that, despite the settings, non-double-aligned parameter
4799 registers can hold double-word quantities in 32-bit mode. */
4801 /* This points to either the 32 bit or the 64 bit version. */
4802 const int *hard_regno_mode_classes
;
4804 static const int hard_32bit_mode_classes
[] = {
4805 S_MODES
, S_MODES
, T_MODES
, S_MODES
, T_MODES
, S_MODES
, D_MODES
, S_MODES
,
4806 T_MODES
, S_MODES
, T_MODES
, S_MODES
, D_MODES
, S_MODES
, D_MODES
, S_MODES
,
4807 T_MODES
, S_MODES
, T_MODES
, S_MODES
, T_MODES
, S_MODES
, D_MODES
, S_MODES
,
4808 T_MODES
, S_MODES
, T_MODES
, S_MODES
, D_MODES
, S_MODES
, D_MODES
, S_MODES
,
4810 OF_MODES
, SF_MODES
, DF_MODES
, SF_MODES
, OF_MODES
, SF_MODES
, DF_MODES
, SF_MODES
,
4811 OF_MODES
, SF_MODES
, DF_MODES
, SF_MODES
, OF_MODES
, SF_MODES
, DF_MODES
, SF_MODES
,
4812 OF_MODES
, SF_MODES
, DF_MODES
, SF_MODES
, OF_MODES
, SF_MODES
, DF_MODES
, SF_MODES
,
4813 OF_MODES
, SF_MODES
, DF_MODES
, SF_MODES
, TF_MODES
, SF_MODES
, DF_MODES
, SF_MODES
,
4815 /* FP regs f32 to f63. Only the even numbered registers actually exist,
4816 and none can hold SFmode/SImode values. */
4817 OF_MODES_NO_S
, 0, DF_MODES_NO_S
, 0, OF_MODES_NO_S
, 0, DF_MODES_NO_S
, 0,
4818 OF_MODES_NO_S
, 0, DF_MODES_NO_S
, 0, OF_MODES_NO_S
, 0, DF_MODES_NO_S
, 0,
4819 OF_MODES_NO_S
, 0, DF_MODES_NO_S
, 0, OF_MODES_NO_S
, 0, DF_MODES_NO_S
, 0,
4820 OF_MODES_NO_S
, 0, DF_MODES_NO_S
, 0, TF_MODES_NO_S
, 0, DF_MODES_NO_S
, 0,
4823 CCFP_MODES
, CCFP_MODES
, CCFP_MODES
, CCFP_MODES
,
4825 /* %icc, %sfp, %gsr */
4826 CC_MODES
, 0, D_MODES
4829 static const int hard_64bit_mode_classes
[] = {
4830 D_MODES
, D_MODES
, T_MODES
, D_MODES
, T_MODES
, D_MODES
, T_MODES
, D_MODES
,
4831 O_MODES
, D_MODES
, T_MODES
, D_MODES
, T_MODES
, D_MODES
, T_MODES
, D_MODES
,
4832 T_MODES
, D_MODES
, T_MODES
, D_MODES
, T_MODES
, D_MODES
, T_MODES
, D_MODES
,
4833 O_MODES
, D_MODES
, T_MODES
, D_MODES
, T_MODES
, D_MODES
, T_MODES
, D_MODES
,
4835 OF_MODES
, SF_MODES
, DF_MODES
, SF_MODES
, OF_MODES
, SF_MODES
, DF_MODES
, SF_MODES
,
4836 OF_MODES
, SF_MODES
, DF_MODES
, SF_MODES
, OF_MODES
, SF_MODES
, DF_MODES
, SF_MODES
,
4837 OF_MODES
, SF_MODES
, DF_MODES
, SF_MODES
, OF_MODES
, SF_MODES
, DF_MODES
, SF_MODES
,
4838 OF_MODES
, SF_MODES
, DF_MODES
, SF_MODES
, TF_MODES
, SF_MODES
, DF_MODES
, SF_MODES
,
4840 /* FP regs f32 to f63. Only the even numbered registers actually exist,
4841 and none can hold SFmode/SImode values. */
4842 OF_MODES_NO_S
, 0, DF_MODES_NO_S
, 0, OF_MODES_NO_S
, 0, DF_MODES_NO_S
, 0,
4843 OF_MODES_NO_S
, 0, DF_MODES_NO_S
, 0, OF_MODES_NO_S
, 0, DF_MODES_NO_S
, 0,
4844 OF_MODES_NO_S
, 0, DF_MODES_NO_S
, 0, OF_MODES_NO_S
, 0, DF_MODES_NO_S
, 0,
4845 OF_MODES_NO_S
, 0, DF_MODES_NO_S
, 0, TF_MODES_NO_S
, 0, DF_MODES_NO_S
, 0,
4848 CCFP_MODES
, CCFP_MODES
, CCFP_MODES
, CCFP_MODES
,
4850 /* %icc, %sfp, %gsr */
4851 CC_MODES
, 0, D_MODES
4854 int sparc_mode_class
[NUM_MACHINE_MODES
];
4856 enum reg_class sparc_regno_reg_class
[FIRST_PSEUDO_REGISTER
];
4859 sparc_init_modes (void)
4863 for (i
= 0; i
< NUM_MACHINE_MODES
; i
++)
4865 machine_mode m
= (machine_mode
) i
;
4866 unsigned int size
= GET_MODE_SIZE (m
);
4868 switch (GET_MODE_CLASS (m
))
4871 case MODE_PARTIAL_INT
:
4872 case MODE_COMPLEX_INT
:
4874 sparc_mode_class
[i
] = 1 << (int) H_MODE
;
4876 sparc_mode_class
[i
] = 1 << (int) S_MODE
;
4878 sparc_mode_class
[i
] = 1 << (int) D_MODE
;
4879 else if (size
== 16)
4880 sparc_mode_class
[i
] = 1 << (int) T_MODE
;
4881 else if (size
== 32)
4882 sparc_mode_class
[i
] = 1 << (int) O_MODE
;
4884 sparc_mode_class
[i
] = 0;
4886 case MODE_VECTOR_INT
:
4888 sparc_mode_class
[i
] = 1 << (int) SF_MODE
;
4890 sparc_mode_class
[i
] = 1 << (int) DF_MODE
;
4892 sparc_mode_class
[i
] = 0;
4895 case MODE_COMPLEX_FLOAT
:
4897 sparc_mode_class
[i
] = 1 << (int) SF_MODE
;
4899 sparc_mode_class
[i
] = 1 << (int) DF_MODE
;
4900 else if (size
== 16)
4901 sparc_mode_class
[i
] = 1 << (int) TF_MODE
;
4902 else if (size
== 32)
4903 sparc_mode_class
[i
] = 1 << (int) OF_MODE
;
4905 sparc_mode_class
[i
] = 0;
4908 if (m
== CCFPmode
|| m
== CCFPEmode
)
4909 sparc_mode_class
[i
] = 1 << (int) CCFP_MODE
;
4911 sparc_mode_class
[i
] = 1 << (int) CC_MODE
;
4914 sparc_mode_class
[i
] = 0;
4920 hard_regno_mode_classes
= hard_64bit_mode_classes
;
4922 hard_regno_mode_classes
= hard_32bit_mode_classes
;
4924 /* Initialize the array used by REGNO_REG_CLASS. */
4925 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
4927 if (i
< 16 && TARGET_V8PLUS
)
4928 sparc_regno_reg_class
[i
] = I64_REGS
;
4929 else if (i
< 32 || i
== FRAME_POINTER_REGNUM
)
4930 sparc_regno_reg_class
[i
] = GENERAL_REGS
;
4932 sparc_regno_reg_class
[i
] = FP_REGS
;
4934 sparc_regno_reg_class
[i
] = EXTRA_FP_REGS
;
4936 sparc_regno_reg_class
[i
] = FPCC_REGS
;
4938 sparc_regno_reg_class
[i
] = NO_REGS
;
4942 /* Return whether REGNO, a global or FP register, must be saved/restored. */
4945 save_global_or_fp_reg_p (unsigned int regno
,
4946 int leaf_function ATTRIBUTE_UNUSED
)
4948 return !call_used_regs
[regno
] && df_regs_ever_live_p (regno
);
4951 /* Return whether the return address register (%i7) is needed. */
4954 return_addr_reg_needed_p (int leaf_function
)
4956 /* If it is live, for example because of __builtin_return_address (0). */
4957 if (df_regs_ever_live_p (RETURN_ADDR_REGNUM
))
4960 /* Otherwise, it is needed as save register if %o7 is clobbered. */
4962 /* Loading the GOT register clobbers %o7. */
4963 || crtl
->uses_pic_offset_table
4964 || df_regs_ever_live_p (INCOMING_RETURN_ADDR_REGNUM
))
4970 /* Return whether REGNO, a local or in register, must be saved/restored. */
4973 save_local_or_in_reg_p (unsigned int regno
, int leaf_function
)
4975 /* General case: call-saved registers live at some point. */
4976 if (!call_used_regs
[regno
] && df_regs_ever_live_p (regno
))
4979 /* Frame pointer register (%fp) if needed. */
4980 if (regno
== HARD_FRAME_POINTER_REGNUM
&& frame_pointer_needed
)
4983 /* Return address register (%i7) if needed. */
4984 if (regno
== RETURN_ADDR_REGNUM
&& return_addr_reg_needed_p (leaf_function
))
4987 /* GOT register (%l7) if needed. */
4988 if (regno
== PIC_OFFSET_TABLE_REGNUM
&& crtl
->uses_pic_offset_table
)
4991 /* If the function accesses prior frames, the frame pointer and the return
4992 address of the previous frame must be saved on the stack. */
4993 if (crtl
->accesses_prior_frames
4994 && (regno
== HARD_FRAME_POINTER_REGNUM
|| regno
== RETURN_ADDR_REGNUM
))
5000 /* Compute the frame size required by the function. This function is called
5001 during the reload pass and also by sparc_expand_prologue. */
5004 sparc_compute_frame_size (HOST_WIDE_INT size
, int leaf_function
)
5006 HOST_WIDE_INT frame_size
, apparent_frame_size
;
5007 int args_size
, n_global_fp_regs
= 0;
5008 bool save_local_in_regs_p
= false;
5011 /* If the function allocates dynamic stack space, the dynamic offset is
5012 computed early and contains REG_PARM_STACK_SPACE, so we need to cope. */
5013 if (leaf_function
&& !cfun
->calls_alloca
)
5016 args_size
= crtl
->outgoing_args_size
+ REG_PARM_STACK_SPACE (cfun
->decl
);
5018 /* Calculate space needed for global registers. */
5021 for (i
= 0; i
< 8; i
++)
5022 if (save_global_or_fp_reg_p (i
, 0))
5023 n_global_fp_regs
+= 2;
5027 for (i
= 0; i
< 8; i
+= 2)
5028 if (save_global_or_fp_reg_p (i
, 0)
5029 || save_global_or_fp_reg_p (i
+ 1, 0))
5030 n_global_fp_regs
+= 2;
5033 /* In the flat window model, find out which local and in registers need to
5034 be saved. We don't reserve space in the current frame for them as they
5035 will be spilled into the register window save area of the caller's frame.
5036 However, as soon as we use this register window save area, we must create
5037 that of the current frame to make it the live one. */
5039 for (i
= 16; i
< 32; i
++)
5040 if (save_local_or_in_reg_p (i
, leaf_function
))
5042 save_local_in_regs_p
= true;
5046 /* Calculate space needed for FP registers. */
5047 for (i
= 32; i
< (TARGET_V9
? 96 : 64); i
+= 2)
5048 if (save_global_or_fp_reg_p (i
, 0) || save_global_or_fp_reg_p (i
+ 1, 0))
5049 n_global_fp_regs
+= 2;
5052 && n_global_fp_regs
== 0
5054 && !save_local_in_regs_p
)
5055 frame_size
= apparent_frame_size
= 0;
5058 /* We subtract STARTING_FRAME_OFFSET, remember it's negative. */
5059 apparent_frame_size
= ROUND_UP (size
- STARTING_FRAME_OFFSET
, 8);
5060 apparent_frame_size
+= n_global_fp_regs
* 4;
5062 /* We need to add the size of the outgoing argument area. */
5063 frame_size
= apparent_frame_size
+ ROUND_UP (args_size
, 8);
5065 /* And that of the register window save area. */
5066 frame_size
+= FIRST_PARM_OFFSET (cfun
->decl
);
5068 /* Finally, bump to the appropriate alignment. */
5069 frame_size
= SPARC_STACK_ALIGN (frame_size
);
5072 /* Set up values for use in prologue and epilogue. */
5073 sparc_frame_size
= frame_size
;
5074 sparc_apparent_frame_size
= apparent_frame_size
;
5075 sparc_n_global_fp_regs
= n_global_fp_regs
;
5076 sparc_save_local_in_regs_p
= save_local_in_regs_p
;
5081 /* Implement the macro INITIAL_ELIMINATION_OFFSET, return the OFFSET. */
5084 sparc_initial_elimination_offset (int to
)
5088 if (to
== STACK_POINTER_REGNUM
)
5089 offset
= sparc_compute_frame_size (get_frame_size (), crtl
->is_leaf
);
5093 offset
+= SPARC_STACK_BIAS
;
5097 /* Output any necessary .register pseudo-ops. */
5100 sparc_output_scratch_registers (FILE *file ATTRIBUTE_UNUSED
)
5102 #ifdef HAVE_AS_REGISTER_PSEUDO_OP
5108 /* Check if %g[2367] were used without
5109 .register being printed for them already. */
5110 for (i
= 2; i
< 8; i
++)
5112 if (df_regs_ever_live_p (i
)
5113 && ! sparc_hard_reg_printed
[i
])
5115 sparc_hard_reg_printed
[i
] = 1;
5116 /* %g7 is used as TLS base register, use #ignore
5117 for it instead of #scratch. */
5118 fprintf (file
, "\t.register\t%%g%d, #%s\n", i
,
5119 i
== 7 ? "ignore" : "scratch");
5126 #define PROBE_INTERVAL (1 << STACK_CHECK_PROBE_INTERVAL_EXP)
5128 #if PROBE_INTERVAL > 4096
5129 #error Cannot use indexed addressing mode for stack probing
5132 /* Emit code to probe a range of stack addresses from FIRST to FIRST+SIZE,
5133 inclusive. These are offsets from the current stack pointer.
5135 Note that we don't use the REG+REG addressing mode for the probes because
5136 of the stack bias in 64-bit mode. And it doesn't really buy us anything
5137 so the advantages of having a single code win here. */
5140 sparc_emit_probe_stack_range (HOST_WIDE_INT first
, HOST_WIDE_INT size
)
5142 rtx g1
= gen_rtx_REG (Pmode
, 1);
5144 /* See if we have a constant small number of probes to generate. If so,
5145 that's the easy case. */
5146 if (size
<= PROBE_INTERVAL
)
5148 emit_move_insn (g1
, GEN_INT (first
));
5149 emit_insn (gen_rtx_SET (g1
,
5150 gen_rtx_MINUS (Pmode
, stack_pointer_rtx
, g1
)));
5151 emit_stack_probe (plus_constant (Pmode
, g1
, -size
));
5154 /* The run-time loop is made up of 9 insns in the generic case while the
5155 compile-time loop is made up of 4+2*(n-2) insns for n # of intervals. */
5156 else if (size
<= 4 * PROBE_INTERVAL
)
5160 emit_move_insn (g1
, GEN_INT (first
+ PROBE_INTERVAL
));
5161 emit_insn (gen_rtx_SET (g1
,
5162 gen_rtx_MINUS (Pmode
, stack_pointer_rtx
, g1
)));
5163 emit_stack_probe (g1
);
5165 /* Probe at FIRST + N * PROBE_INTERVAL for values of N from 2 until
5166 it exceeds SIZE. If only two probes are needed, this will not
5167 generate any code. Then probe at FIRST + SIZE. */
5168 for (i
= 2 * PROBE_INTERVAL
; i
< size
; i
+= PROBE_INTERVAL
)
5170 emit_insn (gen_rtx_SET (g1
,
5171 plus_constant (Pmode
, g1
, -PROBE_INTERVAL
)));
5172 emit_stack_probe (g1
);
5175 emit_stack_probe (plus_constant (Pmode
, g1
,
5176 (i
- PROBE_INTERVAL
) - size
));
5179 /* Otherwise, do the same as above, but in a loop. Note that we must be
5180 extra careful with variables wrapping around because we might be at
5181 the very top (or the very bottom) of the address space and we have
5182 to be able to handle this case properly; in particular, we use an
5183 equality test for the loop condition. */
5186 HOST_WIDE_INT rounded_size
;
5187 rtx g4
= gen_rtx_REG (Pmode
, 4);
5189 emit_move_insn (g1
, GEN_INT (first
));
5192 /* Step 1: round SIZE to the previous multiple of the interval. */
5194 rounded_size
= ROUND_DOWN (size
, PROBE_INTERVAL
);
5195 emit_move_insn (g4
, GEN_INT (rounded_size
));
5198 /* Step 2: compute initial and final value of the loop counter. */
5200 /* TEST_ADDR = SP + FIRST. */
5201 emit_insn (gen_rtx_SET (g1
,
5202 gen_rtx_MINUS (Pmode
, stack_pointer_rtx
, g1
)));
5204 /* LAST_ADDR = SP + FIRST + ROUNDED_SIZE. */
5205 emit_insn (gen_rtx_SET (g4
, gen_rtx_MINUS (Pmode
, g1
, g4
)));
5210 while (TEST_ADDR != LAST_ADDR)
5212 TEST_ADDR = TEST_ADDR + PROBE_INTERVAL
5216 probes at FIRST + N * PROBE_INTERVAL for values of N from 1
5217 until it is equal to ROUNDED_SIZE. */
5220 emit_insn (gen_probe_stack_rangedi (g1
, g1
, g4
));
5222 emit_insn (gen_probe_stack_rangesi (g1
, g1
, g4
));
5225 /* Step 4: probe at FIRST + SIZE if we cannot assert at compile-time
5226 that SIZE is equal to ROUNDED_SIZE. */
5228 if (size
!= rounded_size
)
5229 emit_stack_probe (plus_constant (Pmode
, g4
, rounded_size
- size
));
5232 /* Make sure nothing is scheduled before we are done. */
5233 emit_insn (gen_blockage ());
5236 /* Probe a range of stack addresses from REG1 to REG2 inclusive. These are
5237 absolute addresses. */
5240 output_probe_stack_range (rtx reg1
, rtx reg2
)
5242 static int labelno
= 0;
5246 ASM_GENERATE_INTERNAL_LABEL (loop_lab
, "LPSRL", labelno
++);
5249 ASM_OUTPUT_INTERNAL_LABEL (asm_out_file
, loop_lab
);
5251 /* TEST_ADDR = TEST_ADDR + PROBE_INTERVAL. */
5253 xops
[1] = GEN_INT (-PROBE_INTERVAL
);
5254 output_asm_insn ("add\t%0, %1, %0", xops
);
5256 /* Test if TEST_ADDR == LAST_ADDR. */
5258 output_asm_insn ("cmp\t%0, %1", xops
);
5260 /* Probe at TEST_ADDR and branch. */
5262 fputs ("\tbne,pt\t%xcc,", asm_out_file
);
5264 fputs ("\tbne\t", asm_out_file
);
5265 assemble_name_raw (asm_out_file
, loop_lab
);
5266 fputc ('\n', asm_out_file
);
5267 xops
[1] = GEN_INT (SPARC_STACK_BIAS
);
5268 output_asm_insn (" st\t%%g0, [%0+%1]", xops
);
5273 /* Emit code to save/restore registers from LOW to HIGH at BASE+OFFSET as
5274 needed. LOW is supposed to be double-word aligned for 32-bit registers.
5275 SAVE_P decides whether a register must be saved/restored. ACTION_TRUE
5276 is the action to be performed if SAVE_P returns true and ACTION_FALSE
5277 the action to be performed if it returns false. Return the new offset. */
5279 typedef bool (*sorr_pred_t
) (unsigned int, int);
5280 typedef enum { SORR_NONE
, SORR_ADVANCE
, SORR_SAVE
, SORR_RESTORE
} sorr_act_t
;
5283 emit_save_or_restore_regs (unsigned int low
, unsigned int high
, rtx base
,
5284 int offset
, int leaf_function
, sorr_pred_t save_p
,
5285 sorr_act_t action_true
, sorr_act_t action_false
)
5291 if (TARGET_ARCH64
&& high
<= 32)
5295 for (i
= low
; i
< high
; i
++)
5297 if (save_p (i
, leaf_function
))
5299 mem
= gen_frame_mem (DImode
, plus_constant (Pmode
,
5301 if (action_true
== SORR_SAVE
)
5303 insn
= emit_move_insn (mem
, gen_rtx_REG (DImode
, i
));
5304 RTX_FRAME_RELATED_P (insn
) = 1;
5306 else /* action_true == SORR_RESTORE */
5308 /* The frame pointer must be restored last since its old
5309 value may be used as base address for the frame. This
5310 is problematic in 64-bit mode only because of the lack
5311 of double-word load instruction. */
5312 if (i
== HARD_FRAME_POINTER_REGNUM
)
5315 emit_move_insn (gen_rtx_REG (DImode
, i
), mem
);
5319 else if (action_false
== SORR_ADVANCE
)
5325 mem
= gen_frame_mem (DImode
, plus_constant (Pmode
, base
, fp_offset
));
5326 emit_move_insn (hard_frame_pointer_rtx
, mem
);
5331 for (i
= low
; i
< high
; i
+= 2)
5333 bool reg0
= save_p (i
, leaf_function
);
5334 bool reg1
= save_p (i
+ 1, leaf_function
);
5340 mode
= SPARC_INT_REG_P (i
) ? DImode
: DFmode
;
5345 mode
= SPARC_INT_REG_P (i
) ? SImode
: SFmode
;
5350 mode
= SPARC_INT_REG_P (i
) ? SImode
: SFmode
;
5356 if (action_false
== SORR_ADVANCE
)
5361 mem
= gen_frame_mem (mode
, plus_constant (Pmode
, base
, offset
));
5362 if (action_true
== SORR_SAVE
)
5364 insn
= emit_move_insn (mem
, gen_rtx_REG (mode
, regno
));
5365 RTX_FRAME_RELATED_P (insn
) = 1;
5369 mem
= gen_frame_mem (SImode
, plus_constant (Pmode
, base
,
5371 set1
= gen_rtx_SET (mem
, gen_rtx_REG (SImode
, regno
));
5372 RTX_FRAME_RELATED_P (set1
) = 1;
5374 = gen_frame_mem (SImode
, plus_constant (Pmode
, base
,
5376 set2
= gen_rtx_SET (mem
, gen_rtx_REG (SImode
, regno
+ 1));
5377 RTX_FRAME_RELATED_P (set2
) = 1;
5378 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
5379 gen_rtx_PARALLEL (VOIDmode
,
5380 gen_rtvec (2, set1
, set2
)));
5383 else /* action_true == SORR_RESTORE */
5384 emit_move_insn (gen_rtx_REG (mode
, regno
), mem
);
5386 /* Bump and round down to double word
5387 in case we already bumped by 4. */
5388 offset
= ROUND_DOWN (offset
+ 8, 8);
5395 /* Emit code to adjust BASE to OFFSET. Return the new base. */
5398 emit_adjust_base_to_offset (rtx base
, int offset
)
5400 /* ??? This might be optimized a little as %g1 might already have a
5401 value close enough that a single add insn will do. */
5402 /* ??? Although, all of this is probably only a temporary fix because
5403 if %g1 can hold a function result, then sparc_expand_epilogue will
5404 lose (the result will be clobbered). */
5405 rtx new_base
= gen_rtx_REG (Pmode
, 1);
5406 emit_move_insn (new_base
, GEN_INT (offset
));
5407 emit_insn (gen_rtx_SET (new_base
, gen_rtx_PLUS (Pmode
, base
, new_base
)));
5411 /* Emit code to save/restore call-saved global and FP registers. */
5414 emit_save_or_restore_global_fp_regs (rtx base
, int offset
, sorr_act_t action
)
5416 if (offset
< -4096 || offset
+ sparc_n_global_fp_regs
* 4 > 4095)
5418 base
= emit_adjust_base_to_offset (base
, offset
);
5423 = emit_save_or_restore_regs (0, 8, base
, offset
, 0,
5424 save_global_or_fp_reg_p
, action
, SORR_NONE
);
5425 emit_save_or_restore_regs (32, TARGET_V9
? 96 : 64, base
, offset
, 0,
5426 save_global_or_fp_reg_p
, action
, SORR_NONE
);
5429 /* Emit code to save/restore call-saved local and in registers. */
5432 emit_save_or_restore_local_in_regs (rtx base
, int offset
, sorr_act_t action
)
5434 if (offset
< -4096 || offset
+ 16 * UNITS_PER_WORD
> 4095)
5436 base
= emit_adjust_base_to_offset (base
, offset
);
5440 emit_save_or_restore_regs (16, 32, base
, offset
, sparc_leaf_function_p
,
5441 save_local_or_in_reg_p
, action
, SORR_ADVANCE
);
5444 /* Emit a window_save insn. */
5447 emit_window_save (rtx increment
)
5449 rtx_insn
*insn
= emit_insn (gen_window_save (increment
));
5450 RTX_FRAME_RELATED_P (insn
) = 1;
5452 /* The incoming return address (%o7) is saved in %i7. */
5453 add_reg_note (insn
, REG_CFA_REGISTER
,
5454 gen_rtx_SET (gen_rtx_REG (Pmode
, RETURN_ADDR_REGNUM
),
5456 INCOMING_RETURN_ADDR_REGNUM
)));
5458 /* The window save event. */
5459 add_reg_note (insn
, REG_CFA_WINDOW_SAVE
, const0_rtx
);
5461 /* The CFA is %fp, the hard frame pointer. */
5462 add_reg_note (insn
, REG_CFA_DEF_CFA
,
5463 plus_constant (Pmode
, hard_frame_pointer_rtx
,
5464 INCOMING_FRAME_SP_OFFSET
));
5469 /* Generate an increment for the stack pointer. */
5472 gen_stack_pointer_inc (rtx increment
)
5474 return gen_rtx_SET (stack_pointer_rtx
,
5475 gen_rtx_PLUS (Pmode
,
5480 /* Expand the function prologue. The prologue is responsible for reserving
5481 storage for the frame, saving the call-saved registers and loading the
5482 GOT register if needed. */
5485 sparc_expand_prologue (void)
5490 /* Compute a snapshot of crtl->uses_only_leaf_regs. Relying
5491 on the final value of the flag means deferring the prologue/epilogue
5492 expansion until just before the second scheduling pass, which is too
5493 late to emit multiple epilogues or return insns.
5495 Of course we are making the assumption that the value of the flag
5496 will not change between now and its final value. Of the three parts
5497 of the formula, only the last one can reasonably vary. Let's take a
5498 closer look, after assuming that the first two ones are set to true
5499 (otherwise the last value is effectively silenced).
5501 If only_leaf_regs_used returns false, the global predicate will also
5502 be false so the actual frame size calculated below will be positive.
5503 As a consequence, the save_register_window insn will be emitted in
5504 the instruction stream; now this insn explicitly references %fp
5505 which is not a leaf register so only_leaf_regs_used will always
5506 return false subsequently.
5508 If only_leaf_regs_used returns true, we hope that the subsequent
5509 optimization passes won't cause non-leaf registers to pop up. For
5510 example, the regrename pass has special provisions to not rename to
5511 non-leaf registers in a leaf function. */
5512 sparc_leaf_function_p
5513 = optimize
> 0 && crtl
->is_leaf
&& only_leaf_regs_used ();
5515 size
= sparc_compute_frame_size (get_frame_size(), sparc_leaf_function_p
);
5517 if (flag_stack_usage_info
)
5518 current_function_static_stack_size
= size
;
5520 if (flag_stack_check
== STATIC_BUILTIN_STACK_CHECK
)
5522 if (crtl
->is_leaf
&& !cfun
->calls_alloca
)
5524 if (size
> PROBE_INTERVAL
&& size
> STACK_CHECK_PROTECT
)
5525 sparc_emit_probe_stack_range (STACK_CHECK_PROTECT
,
5526 size
- STACK_CHECK_PROTECT
);
5529 sparc_emit_probe_stack_range (STACK_CHECK_PROTECT
, size
);
5534 else if (sparc_leaf_function_p
)
5536 rtx size_int_rtx
= GEN_INT (-size
);
5539 insn
= emit_insn (gen_stack_pointer_inc (size_int_rtx
));
5540 else if (size
<= 8192)
5542 insn
= emit_insn (gen_stack_pointer_inc (GEN_INT (-4096)));
5543 RTX_FRAME_RELATED_P (insn
) = 1;
5545 /* %sp is still the CFA register. */
5546 insn
= emit_insn (gen_stack_pointer_inc (GEN_INT (4096 - size
)));
5550 rtx size_rtx
= gen_rtx_REG (Pmode
, 1);
5551 emit_move_insn (size_rtx
, size_int_rtx
);
5552 insn
= emit_insn (gen_stack_pointer_inc (size_rtx
));
5553 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
5554 gen_stack_pointer_inc (size_int_rtx
));
5557 RTX_FRAME_RELATED_P (insn
) = 1;
5561 rtx size_int_rtx
= GEN_INT (-size
);
5564 emit_window_save (size_int_rtx
);
5565 else if (size
<= 8192)
5567 emit_window_save (GEN_INT (-4096));
5569 /* %sp is not the CFA register anymore. */
5570 emit_insn (gen_stack_pointer_inc (GEN_INT (4096 - size
)));
5572 /* Make sure no %fp-based store is issued until after the frame is
5573 established. The offset between the frame pointer and the stack
5574 pointer is calculated relative to the value of the stack pointer
5575 at the end of the function prologue, and moving instructions that
5576 access the stack via the frame pointer between the instructions
5577 that decrement the stack pointer could result in accessing the
5578 register window save area, which is volatile. */
5579 emit_insn (gen_frame_blockage ());
5583 rtx size_rtx
= gen_rtx_REG (Pmode
, 1);
5584 emit_move_insn (size_rtx
, size_int_rtx
);
5585 emit_window_save (size_rtx
);
5589 if (sparc_leaf_function_p
)
5591 sparc_frame_base_reg
= stack_pointer_rtx
;
5592 sparc_frame_base_offset
= size
+ SPARC_STACK_BIAS
;
5596 sparc_frame_base_reg
= hard_frame_pointer_rtx
;
5597 sparc_frame_base_offset
= SPARC_STACK_BIAS
;
5600 if (sparc_n_global_fp_regs
> 0)
5601 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg
,
5602 sparc_frame_base_offset
5603 - sparc_apparent_frame_size
,
5606 /* Load the GOT register if needed. */
5607 if (crtl
->uses_pic_offset_table
)
5608 load_got_register ();
5610 /* Advertise that the data calculated just above are now valid. */
5611 sparc_prologue_data_valid_p
= true;
5614 /* Expand the function prologue. The prologue is responsible for reserving
5615 storage for the frame, saving the call-saved registers and loading the
5616 GOT register if needed. */
5619 sparc_flat_expand_prologue (void)
5624 sparc_leaf_function_p
= optimize
> 0 && crtl
->is_leaf
;
5626 size
= sparc_compute_frame_size (get_frame_size(), sparc_leaf_function_p
);
5628 if (flag_stack_usage_info
)
5629 current_function_static_stack_size
= size
;
5631 if (flag_stack_check
== STATIC_BUILTIN_STACK_CHECK
)
5633 if (crtl
->is_leaf
&& !cfun
->calls_alloca
)
5635 if (size
> PROBE_INTERVAL
&& size
> STACK_CHECK_PROTECT
)
5636 sparc_emit_probe_stack_range (STACK_CHECK_PROTECT
,
5637 size
- STACK_CHECK_PROTECT
);
5640 sparc_emit_probe_stack_range (STACK_CHECK_PROTECT
, size
);
5643 if (sparc_save_local_in_regs_p
)
5644 emit_save_or_restore_local_in_regs (stack_pointer_rtx
, SPARC_STACK_BIAS
,
5651 rtx size_int_rtx
, size_rtx
;
5653 size_rtx
= size_int_rtx
= GEN_INT (-size
);
5655 /* We establish the frame (i.e. decrement the stack pointer) first, even
5656 if we use a frame pointer, because we cannot clobber any call-saved
5657 registers, including the frame pointer, if we haven't created a new
5658 register save area, for the sake of compatibility with the ABI. */
5660 insn
= emit_insn (gen_stack_pointer_inc (size_int_rtx
));
5661 else if (size
<= 8192 && !frame_pointer_needed
)
5663 insn
= emit_insn (gen_stack_pointer_inc (GEN_INT (-4096)));
5664 RTX_FRAME_RELATED_P (insn
) = 1;
5665 insn
= emit_insn (gen_stack_pointer_inc (GEN_INT (4096 - size
)));
5669 size_rtx
= gen_rtx_REG (Pmode
, 1);
5670 emit_move_insn (size_rtx
, size_int_rtx
);
5671 insn
= emit_insn (gen_stack_pointer_inc (size_rtx
));
5672 add_reg_note (insn
, REG_CFA_ADJUST_CFA
,
5673 gen_stack_pointer_inc (size_int_rtx
));
5675 RTX_FRAME_RELATED_P (insn
) = 1;
5677 /* Ensure nothing is scheduled until after the frame is established. */
5678 emit_insn (gen_blockage ());
5680 if (frame_pointer_needed
)
5682 insn
= emit_insn (gen_rtx_SET (hard_frame_pointer_rtx
,
5683 gen_rtx_MINUS (Pmode
,
5686 RTX_FRAME_RELATED_P (insn
) = 1;
5688 add_reg_note (insn
, REG_CFA_ADJUST_CFA
,
5689 gen_rtx_SET (hard_frame_pointer_rtx
,
5690 plus_constant (Pmode
, stack_pointer_rtx
,
5694 if (return_addr_reg_needed_p (sparc_leaf_function_p
))
5696 rtx o7
= gen_rtx_REG (Pmode
, INCOMING_RETURN_ADDR_REGNUM
);
5697 rtx i7
= gen_rtx_REG (Pmode
, RETURN_ADDR_REGNUM
);
5699 insn
= emit_move_insn (i7
, o7
);
5700 RTX_FRAME_RELATED_P (insn
) = 1;
5702 add_reg_note (insn
, REG_CFA_REGISTER
, gen_rtx_SET (i7
, o7
));
5704 /* Prevent this instruction from ever being considered dead,
5705 even if this function has no epilogue. */
5710 if (frame_pointer_needed
)
5712 sparc_frame_base_reg
= hard_frame_pointer_rtx
;
5713 sparc_frame_base_offset
= SPARC_STACK_BIAS
;
5717 sparc_frame_base_reg
= stack_pointer_rtx
;
5718 sparc_frame_base_offset
= size
+ SPARC_STACK_BIAS
;
5721 if (sparc_n_global_fp_regs
> 0)
5722 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg
,
5723 sparc_frame_base_offset
5724 - sparc_apparent_frame_size
,
5727 /* Load the GOT register if needed. */
5728 if (crtl
->uses_pic_offset_table
)
5729 load_got_register ();
5731 /* Advertise that the data calculated just above are now valid. */
5732 sparc_prologue_data_valid_p
= true;
5735 /* This function generates the assembly code for function entry, which boils
5736 down to emitting the necessary .register directives. */
5739 sparc_asm_function_prologue (FILE *file
, HOST_WIDE_INT size ATTRIBUTE_UNUSED
)
5741 /* Check that the assumption we made in sparc_expand_prologue is valid. */
5743 gcc_assert (sparc_leaf_function_p
== crtl
->uses_only_leaf_regs
);
5745 sparc_output_scratch_registers (file
);
5748 /* Expand the function epilogue, either normal or part of a sibcall.
5749 We emit all the instructions except the return or the call. */
5752 sparc_expand_epilogue (bool for_eh
)
5754 HOST_WIDE_INT size
= sparc_frame_size
;
5756 if (sparc_n_global_fp_regs
> 0)
5757 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg
,
5758 sparc_frame_base_offset
5759 - sparc_apparent_frame_size
,
5762 if (size
== 0 || for_eh
)
5764 else if (sparc_leaf_function_p
)
5767 emit_insn (gen_stack_pointer_inc (GEN_INT (size
)));
5768 else if (size
<= 8192)
5770 emit_insn (gen_stack_pointer_inc (GEN_INT (4096)));
5771 emit_insn (gen_stack_pointer_inc (GEN_INT (size
- 4096)));
5775 rtx reg
= gen_rtx_REG (Pmode
, 1);
5776 emit_move_insn (reg
, GEN_INT (size
));
5777 emit_insn (gen_stack_pointer_inc (reg
));
5782 /* Expand the function epilogue, either normal or part of a sibcall.
5783 We emit all the instructions except the return or the call. */
5786 sparc_flat_expand_epilogue (bool for_eh
)
5788 HOST_WIDE_INT size
= sparc_frame_size
;
5790 if (sparc_n_global_fp_regs
> 0)
5791 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg
,
5792 sparc_frame_base_offset
5793 - sparc_apparent_frame_size
,
5796 /* If we have a frame pointer, we'll need both to restore it before the
5797 frame is destroyed and use its current value in destroying the frame.
5798 Since we don't have an atomic way to do that in the flat window model,
5799 we save the current value into a temporary register (%g1). */
5800 if (frame_pointer_needed
&& !for_eh
)
5801 emit_move_insn (gen_rtx_REG (Pmode
, 1), hard_frame_pointer_rtx
);
5803 if (return_addr_reg_needed_p (sparc_leaf_function_p
))
5804 emit_move_insn (gen_rtx_REG (Pmode
, INCOMING_RETURN_ADDR_REGNUM
),
5805 gen_rtx_REG (Pmode
, RETURN_ADDR_REGNUM
));
5807 if (sparc_save_local_in_regs_p
)
5808 emit_save_or_restore_local_in_regs (sparc_frame_base_reg
,
5809 sparc_frame_base_offset
,
5812 if (size
== 0 || for_eh
)
5814 else if (frame_pointer_needed
)
5816 /* Make sure the frame is destroyed after everything else is done. */
5817 emit_insn (gen_blockage ());
5819 emit_move_insn (stack_pointer_rtx
, gen_rtx_REG (Pmode
, 1));
5824 emit_insn (gen_blockage ());
5827 emit_insn (gen_stack_pointer_inc (GEN_INT (size
)));
5828 else if (size
<= 8192)
5830 emit_insn (gen_stack_pointer_inc (GEN_INT (4096)));
5831 emit_insn (gen_stack_pointer_inc (GEN_INT (size
- 4096)));
5835 rtx reg
= gen_rtx_REG (Pmode
, 1);
5836 emit_move_insn (reg
, GEN_INT (size
));
5837 emit_insn (gen_stack_pointer_inc (reg
));
5842 /* Return true if it is appropriate to emit `return' instructions in the
5843 body of a function. */
5846 sparc_can_use_return_insn_p (void)
5848 return sparc_prologue_data_valid_p
5849 && sparc_n_global_fp_regs
== 0
5851 ? (sparc_frame_size
== 0 && !sparc_save_local_in_regs_p
)
5852 : (sparc_frame_size
== 0 || !sparc_leaf_function_p
);
5855 /* This function generates the assembly code for function exit. */
5858 sparc_asm_function_epilogue (FILE *file
, HOST_WIDE_INT size ATTRIBUTE_UNUSED
)
5860 /* If the last two instructions of a function are "call foo; dslot;"
5861 the return address might point to the first instruction in the next
5862 function and we have to output a dummy nop for the sake of sane
5863 backtraces in such cases. This is pointless for sibling calls since
5864 the return address is explicitly adjusted. */
5866 rtx insn
, last_real_insn
;
5868 insn
= get_last_insn ();
5870 last_real_insn
= prev_real_insn (insn
);
5872 && NONJUMP_INSN_P (last_real_insn
)
5873 && GET_CODE (PATTERN (last_real_insn
)) == SEQUENCE
)
5874 last_real_insn
= XVECEXP (PATTERN (last_real_insn
), 0, 0);
5877 && CALL_P (last_real_insn
)
5878 && !SIBLING_CALL_P (last_real_insn
))
5879 fputs("\tnop\n", file
);
5881 sparc_output_deferred_case_vectors ();
5884 /* Output a 'restore' instruction. */
5887 output_restore (rtx pat
)
5893 fputs ("\t restore\n", asm_out_file
);
5897 gcc_assert (GET_CODE (pat
) == SET
);
5899 operands
[0] = SET_DEST (pat
);
5900 pat
= SET_SRC (pat
);
5902 switch (GET_CODE (pat
))
5905 operands
[1] = XEXP (pat
, 0);
5906 operands
[2] = XEXP (pat
, 1);
5907 output_asm_insn (" restore %r1, %2, %Y0", operands
);
5910 operands
[1] = XEXP (pat
, 0);
5911 operands
[2] = XEXP (pat
, 1);
5912 output_asm_insn (" restore %r1, %%lo(%a2), %Y0", operands
);
5915 operands
[1] = XEXP (pat
, 0);
5916 gcc_assert (XEXP (pat
, 1) == const1_rtx
);
5917 output_asm_insn (" restore %r1, %r1, %Y0", operands
);
5921 output_asm_insn (" restore %%g0, %1, %Y0", operands
);
5926 /* Output a return. */
5929 output_return (rtx_insn
*insn
)
5931 if (crtl
->calls_eh_return
)
5933 /* If the function uses __builtin_eh_return, the eh_return
5934 machinery occupies the delay slot. */
5935 gcc_assert (!final_sequence
);
5937 if (flag_delayed_branch
)
5939 if (!TARGET_FLAT
&& TARGET_V9
)
5940 fputs ("\treturn\t%i7+8\n", asm_out_file
);
5944 fputs ("\trestore\n", asm_out_file
);
5946 fputs ("\tjmp\t%o7+8\n", asm_out_file
);
5949 fputs ("\t add\t%sp, %g1, %sp\n", asm_out_file
);
5954 fputs ("\trestore\n", asm_out_file
);
5956 fputs ("\tadd\t%sp, %g1, %sp\n", asm_out_file
);
5957 fputs ("\tjmp\t%o7+8\n\t nop\n", asm_out_file
);
5960 else if (sparc_leaf_function_p
|| TARGET_FLAT
)
5962 /* This is a leaf or flat function so we don't have to bother restoring
5963 the register window, which frees us from dealing with the convoluted
5964 semantics of restore/return. We simply output the jump to the
5965 return address and the insn in the delay slot (if any). */
5967 return "jmp\t%%o7+%)%#";
5971 /* This is a regular function so we have to restore the register window.
5972 We may have a pending insn for the delay slot, which will be either
5973 combined with the 'restore' instruction or put in the delay slot of
5974 the 'return' instruction. */
5980 delay
= NEXT_INSN (insn
);
5983 pat
= PATTERN (delay
);
5985 if (TARGET_V9
&& ! epilogue_renumber (&pat
, 1))
5987 epilogue_renumber (&pat
, 0);
5988 return "return\t%%i7+%)%#";
5992 output_asm_insn ("jmp\t%%i7+%)", NULL
);
5993 output_restore (pat
);
5994 PATTERN (delay
) = gen_blockage ();
5995 INSN_CODE (delay
) = -1;
6000 /* The delay slot is empty. */
6002 return "return\t%%i7+%)\n\t nop";
6003 else if (flag_delayed_branch
)
6004 return "jmp\t%%i7+%)\n\t restore";
6006 return "restore\n\tjmp\t%%o7+%)\n\t nop";
6013 /* Output a sibling call. */
6016 output_sibcall (rtx_insn
*insn
, rtx call_operand
)
6020 gcc_assert (flag_delayed_branch
);
6022 operands
[0] = call_operand
;
6024 if (sparc_leaf_function_p
|| TARGET_FLAT
)
6026 /* This is a leaf or flat function so we don't have to bother restoring
6027 the register window. We simply output the jump to the function and
6028 the insn in the delay slot (if any). */
6030 gcc_assert (!(LEAF_SIBCALL_SLOT_RESERVED_P
&& final_sequence
));
6033 output_asm_insn ("sethi\t%%hi(%a0), %%g1\n\tjmp\t%%g1 + %%lo(%a0)%#",
6036 /* Use or with rs2 %%g0 instead of mov, so that as/ld can optimize
6037 it into branch if possible. */
6038 output_asm_insn ("or\t%%o7, %%g0, %%g1\n\tcall\t%a0, 0\n\t or\t%%g1, %%g0, %%o7",
6043 /* This is a regular function so we have to restore the register window.
6044 We may have a pending insn for the delay slot, which will be combined
6045 with the 'restore' instruction. */
6047 output_asm_insn ("call\t%a0, 0", operands
);
6051 rtx_insn
*delay
= NEXT_INSN (insn
);
6054 output_restore (PATTERN (delay
));
6056 PATTERN (delay
) = gen_blockage ();
6057 INSN_CODE (delay
) = -1;
6060 output_restore (NULL_RTX
);
6066 /* Functions for handling argument passing.
6068 For 32-bit, the first 6 args are normally in registers and the rest are
6069 pushed. Any arg that starts within the first 6 words is at least
6070 partially passed in a register unless its data type forbids.
6072 For 64-bit, the argument registers are laid out as an array of 16 elements
6073 and arguments are added sequentially. The first 6 int args and up to the
6074 first 16 fp args (depending on size) are passed in regs.
6076 Slot Stack Integral Float Float in structure Double Long Double
6077 ---- ----- -------- ----- ------------------ ------ -----------
6078 15 [SP+248] %f31 %f30,%f31 %d30
6079 14 [SP+240] %f29 %f28,%f29 %d28 %q28
6080 13 [SP+232] %f27 %f26,%f27 %d26
6081 12 [SP+224] %f25 %f24,%f25 %d24 %q24
6082 11 [SP+216] %f23 %f22,%f23 %d22
6083 10 [SP+208] %f21 %f20,%f21 %d20 %q20
6084 9 [SP+200] %f19 %f18,%f19 %d18
6085 8 [SP+192] %f17 %f16,%f17 %d16 %q16
6086 7 [SP+184] %f15 %f14,%f15 %d14
6087 6 [SP+176] %f13 %f12,%f13 %d12 %q12
6088 5 [SP+168] %o5 %f11 %f10,%f11 %d10
6089 4 [SP+160] %o4 %f9 %f8,%f9 %d8 %q8
6090 3 [SP+152] %o3 %f7 %f6,%f7 %d6
6091 2 [SP+144] %o2 %f5 %f4,%f5 %d4 %q4
6092 1 [SP+136] %o1 %f3 %f2,%f3 %d2
6093 0 [SP+128] %o0 %f1 %f0,%f1 %d0 %q0
6095 Here SP = %sp if -mno-stack-bias or %sp+stack_bias otherwise.
6097 Integral arguments are always passed as 64-bit quantities appropriately
6100 Passing of floating point values is handled as follows.
6101 If a prototype is in scope:
6102 If the value is in a named argument (i.e. not a stdarg function or a
6103 value not part of the `...') then the value is passed in the appropriate
6105 If the value is part of the `...' and is passed in one of the first 6
6106 slots then the value is passed in the appropriate int reg.
6107 If the value is part of the `...' and is not passed in one of the first 6
6108 slots then the value is passed in memory.
6109 If a prototype is not in scope:
6110 If the value is one of the first 6 arguments the value is passed in the
6111 appropriate integer reg and the appropriate fp reg.
6112 If the value is not one of the first 6 arguments the value is passed in
6113 the appropriate fp reg and in memory.
6116 Summary of the calling conventions implemented by GCC on the SPARC:
6119 size argument return value
6121 small integer <4 int. reg. int. reg.
6122 word 4 int. reg. int. reg.
6123 double word 8 int. reg. int. reg.
6125 _Complex small integer <8 int. reg. int. reg.
6126 _Complex word 8 int. reg. int. reg.
6127 _Complex double word 16 memory int. reg.
6129 vector integer <=8 int. reg. FP reg.
6130 vector integer >8 memory memory
6132 float 4 int. reg. FP reg.
6133 double 8 int. reg. FP reg.
6134 long double 16 memory memory
6136 _Complex float 8 memory FP reg.
6137 _Complex double 16 memory FP reg.
6138 _Complex long double 32 memory FP reg.
6140 vector float any memory memory
6142 aggregate any memory memory
6147 size argument return value
6149 small integer <8 int. reg. int. reg.
6150 word 8 int. reg. int. reg.
6151 double word 16 int. reg. int. reg.
6153 _Complex small integer <16 int. reg. int. reg.
6154 _Complex word 16 int. reg. int. reg.
6155 _Complex double word 32 memory int. reg.
6157 vector integer <=16 FP reg. FP reg.
6158 vector integer 16<s<=32 memory FP reg.
6159 vector integer >32 memory memory
6161 float 4 FP reg. FP reg.
6162 double 8 FP reg. FP reg.
6163 long double 16 FP reg. FP reg.
6165 _Complex float 8 FP reg. FP reg.
6166 _Complex double 16 FP reg. FP reg.
6167 _Complex long double 32 memory FP reg.
6169 vector float <=16 FP reg. FP reg.
6170 vector float 16<s<=32 memory FP reg.
6171 vector float >32 memory memory
6173 aggregate <=16 reg. reg.
6174 aggregate 16<s<=32 memory reg.
6175 aggregate >32 memory memory
6179 Note #1: complex floating-point types follow the extended SPARC ABIs as
6180 implemented by the Sun compiler.
6182 Note #2: integral vector types follow the scalar floating-point types
6183 conventions to match what is implemented by the Sun VIS SDK.
6185 Note #3: floating-point vector types follow the aggregate types
6189 /* Maximum number of int regs for args. */
6190 #define SPARC_INT_ARG_MAX 6
6191 /* Maximum number of fp regs for args. */
6192 #define SPARC_FP_ARG_MAX 16
6193 /* Number of words (partially) occupied for a given size in units. */
6194 #define CEIL_NWORDS(SIZE) CEIL((SIZE), UNITS_PER_WORD)
6196 /* Handle the INIT_CUMULATIVE_ARGS macro.
6197 Initialize a variable CUM of type CUMULATIVE_ARGS
6198 for a call to a function whose data type is FNTYPE.
6199 For a library call, FNTYPE is 0. */
6202 init_cumulative_args (struct sparc_args
*cum
, tree fntype
, rtx
, tree
)
6205 cum
->prototype_p
= fntype
&& prototype_p (fntype
);
6206 cum
->libcall_p
= !fntype
;
6209 /* Handle promotion of pointer and integer arguments. */
6212 sparc_promote_function_mode (const_tree type
, machine_mode mode
,
6213 int *punsignedp
, const_tree
, int)
6215 if (type
&& POINTER_TYPE_P (type
))
6217 *punsignedp
= POINTERS_EXTEND_UNSIGNED
;
6221 /* Integral arguments are passed as full words, as per the ABI. */
6222 if (GET_MODE_CLASS (mode
) == MODE_INT
6223 && GET_MODE_SIZE (mode
) < UNITS_PER_WORD
)
6229 /* Handle the TARGET_STRICT_ARGUMENT_NAMING target hook. */
6232 sparc_strict_argument_naming (cumulative_args_t ca ATTRIBUTE_UNUSED
)
6234 return TARGET_ARCH64
? true : false;
6237 /* Traverse the record TYPE recursively and call FUNC on its fields.
6238 NAMED is true if this is for a named parameter. DATA is passed
6239 to FUNC for each field. OFFSET is the starting position and
6240 PACKED is true if we are inside a packed record. */
6242 template <typename T
, void Func (const_tree
, HOST_WIDE_INT
, bool, T
*)>
6244 traverse_record_type (const_tree type
, bool named
, T
*data
,
6245 HOST_WIDE_INT offset
= 0, bool packed
= false)
6247 /* The ABI obviously doesn't specify how packed structures are passed.
6248 These are passed in integer regs if possible, otherwise memory. */
6250 for (tree field
= TYPE_FIELDS (type
); field
; field
= DECL_CHAIN (field
))
6251 if (TREE_CODE (field
) == FIELD_DECL
&& DECL_PACKED (field
))
6257 /* Walk the real fields, but skip those with no size or a zero size.
6258 ??? Fields with variable offset are handled as having zero offset. */
6259 for (tree field
= TYPE_FIELDS (type
); field
; field
= DECL_CHAIN (field
))
6260 if (TREE_CODE (field
) == FIELD_DECL
)
6262 if (!DECL_SIZE (field
) || integer_zerop (DECL_SIZE (field
)))
6265 HOST_WIDE_INT bitpos
= offset
;
6266 if (TREE_CODE (DECL_FIELD_OFFSET (field
)) == INTEGER_CST
)
6267 bitpos
+= int_bit_position (field
);
6269 tree field_type
= TREE_TYPE (field
);
6270 if (TREE_CODE (field_type
) == RECORD_TYPE
)
6271 traverse_record_type
<T
, Func
> (field_type
, named
, data
, bitpos
,
6276 = FLOAT_TYPE_P (field_type
) || VECTOR_TYPE_P (field_type
);
6277 Func (field
, bitpos
, fp_type
&& named
&& !packed
&& TARGET_FPU
,
6283 /* Handle recursive register classifying for structure layout. */
6287 bool int_regs
; /* true if field eligible to int registers. */
6288 bool fp_regs
; /* true if field eligible to FP registers. */
6289 bool fp_regs_in_first_word
; /* true if such field in first word. */
6292 /* A subroutine of function_arg_slotno. Classify the field. */
6295 classify_registers (const_tree
, HOST_WIDE_INT bitpos
, bool fp
,
6296 classify_data_t
*data
)
6300 data
->fp_regs
= true;
6301 if (bitpos
< BITS_PER_WORD
)
6302 data
->fp_regs_in_first_word
= true;
6305 data
->int_regs
= true;
6308 /* Compute the slot number to pass an argument in.
6309 Return the slot number or -1 if passing on the stack.
6311 CUM is a variable of type CUMULATIVE_ARGS which gives info about
6312 the preceding args and about the function being called.
6313 MODE is the argument's machine mode.
6314 TYPE is the data type of the argument (as a tree).
6315 This is null for libcalls where that information may
6317 NAMED is nonzero if this argument is a named parameter
6318 (otherwise it is an extra parameter matching an ellipsis).
6319 INCOMING is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG.
6320 *PREGNO records the register number to use if scalar type.
6321 *PPADDING records the amount of padding needed in words. */
6324 function_arg_slotno (const struct sparc_args
*cum
, machine_mode mode
,
6325 const_tree type
, bool named
, bool incoming
,
6326 int *pregno
, int *ppadding
)
6328 int regbase
= (incoming
6329 ? SPARC_INCOMING_INT_ARG_FIRST
6330 : SPARC_OUTGOING_INT_ARG_FIRST
);
6331 int slotno
= cum
->words
;
6332 enum mode_class mclass
;
6337 if (type
&& TREE_ADDRESSABLE (type
))
6343 && TYPE_ALIGN (type
) % PARM_BOUNDARY
!= 0)
6346 /* For SPARC64, objects requiring 16-byte alignment get it. */
6348 && (type
? TYPE_ALIGN (type
) : GET_MODE_ALIGNMENT (mode
)) >= 128
6349 && (slotno
& 1) != 0)
6350 slotno
++, *ppadding
= 1;
6352 mclass
= GET_MODE_CLASS (mode
);
6353 if (type
&& TREE_CODE (type
) == VECTOR_TYPE
)
6355 /* Vector types deserve special treatment because they are
6356 polymorphic wrt their mode, depending upon whether VIS
6357 instructions are enabled. */
6358 if (TREE_CODE (TREE_TYPE (type
)) == REAL_TYPE
)
6360 /* The SPARC port defines no floating-point vector modes. */
6361 gcc_assert (mode
== BLKmode
);
6365 /* Integral vector types should either have a vector
6366 mode or an integral mode, because we are guaranteed
6367 by pass_by_reference that their size is not greater
6368 than 16 bytes and TImode is 16-byte wide. */
6369 gcc_assert (mode
!= BLKmode
);
6371 /* Vector integers are handled like floats according to
6373 mclass
= MODE_FLOAT
;
6380 case MODE_COMPLEX_FLOAT
:
6381 case MODE_VECTOR_INT
:
6382 if (TARGET_ARCH64
&& TARGET_FPU
&& named
)
6384 /* If all arg slots are filled, then must pass on stack. */
6385 if (slotno
>= SPARC_FP_ARG_MAX
)
6388 regno
= SPARC_FP_ARG_FIRST
+ slotno
* 2;
6389 /* Arguments filling only one single FP register are
6390 right-justified in the outer double FP register. */
6391 if (GET_MODE_SIZE (mode
) <= 4)
6398 case MODE_COMPLEX_INT
:
6399 /* If all arg slots are filled, then must pass on stack. */
6400 if (slotno
>= SPARC_INT_ARG_MAX
)
6403 regno
= regbase
+ slotno
;
6407 if (mode
== VOIDmode
)
6408 /* MODE is VOIDmode when generating the actual call. */
6411 gcc_assert (mode
== BLKmode
);
6415 || (TREE_CODE (type
) != RECORD_TYPE
6416 && TREE_CODE (type
) != VECTOR_TYPE
))
6418 /* If all arg slots are filled, then must pass on stack. */
6419 if (slotno
>= SPARC_INT_ARG_MAX
)
6422 regno
= regbase
+ slotno
;
6424 else /* TARGET_ARCH64 && type */
6426 /* If all arg slots are filled, then must pass on stack. */
6427 if (slotno
>= SPARC_FP_ARG_MAX
)
6430 if (TREE_CODE (type
) == RECORD_TYPE
)
6432 classify_data_t data
= { false, false, false };
6433 traverse_record_type
<classify_data_t
, classify_registers
>
6434 (type
, named
, &data
);
6436 /* If all slots are filled except for the last one, but there
6437 is no FP field in the first word, then must pass on stack. */
6439 && !data
.fp_regs_in_first_word
6440 && slotno
>= SPARC_FP_ARG_MAX
- 1)
6443 /* If there are only int args and all int slots are filled,
6444 then must pass on stack. */
6447 && slotno
>= SPARC_INT_ARG_MAX
)
6451 /* PREGNO isn't set since both int and FP regs can be used. */
6464 /* Handle recursive register counting/assigning for structure layout. */
6468 int slotno
; /* slot number of the argument. */
6469 int regbase
; /* regno of the base register. */
6470 int intoffset
; /* offset of the first pending integer field. */
6471 int nregs
; /* number of words passed in registers. */
6472 bool stack
; /* true if part of the argument is on the stack. */
6473 rtx ret
; /* return expression being built. */
6476 /* A subroutine of function_arg_record_value. Compute the number of integer
6477 registers to be assigned between PARMS->intoffset and BITPOS. Return
6478 true if at least one integer register is assigned or false otherwise. */
6481 compute_int_layout (HOST_WIDE_INT bitpos
, assign_data_t
*data
, int *pnregs
)
6483 if (data
->intoffset
< 0)
6486 const int intoffset
= data
->intoffset
;
6487 data
->intoffset
= -1;
6489 const int this_slotno
= data
->slotno
+ intoffset
/ BITS_PER_WORD
;
6490 const unsigned int startbit
= ROUND_DOWN (intoffset
, BITS_PER_WORD
);
6491 const unsigned int endbit
= ROUND_UP (bitpos
, BITS_PER_WORD
);
6492 int nregs
= (endbit
- startbit
) / BITS_PER_WORD
;
6494 if (nregs
> 0 && nregs
> SPARC_INT_ARG_MAX
- this_slotno
)
6496 nregs
= SPARC_INT_ARG_MAX
- this_slotno
;
6498 /* We need to pass this field (partly) on the stack. */
6509 /* A subroutine of function_arg_record_value. Compute the number and the mode
6510 of the FP registers to be assigned for FIELD. Return true if at least one
6511 FP register is assigned or false otherwise. */
6514 compute_fp_layout (const_tree field
, HOST_WIDE_INT bitpos
,
6515 assign_data_t
*data
,
6516 int *pnregs
, machine_mode
*pmode
)
6518 const int this_slotno
= data
->slotno
+ bitpos
/ BITS_PER_WORD
;
6519 machine_mode mode
= DECL_MODE (field
);
6522 /* Slots are counted as words while regs are counted as having the size of
6523 the (inner) mode. */
6524 if (TREE_CODE (TREE_TYPE (field
)) == VECTOR_TYPE
&& mode
== BLKmode
)
6526 mode
= TYPE_MODE (TREE_TYPE (TREE_TYPE (field
)));
6527 nregs
= TYPE_VECTOR_SUBPARTS (TREE_TYPE (field
));
6529 else if (TREE_CODE (TREE_TYPE (field
)) == COMPLEX_TYPE
)
6531 mode
= TYPE_MODE (TREE_TYPE (TREE_TYPE (field
)));
6537 nslots
= CEIL_NWORDS (nregs
* GET_MODE_SIZE (mode
));
6539 if (nslots
> SPARC_FP_ARG_MAX
- this_slotno
)
6541 nslots
= SPARC_FP_ARG_MAX
- this_slotno
;
6542 nregs
= (nslots
* UNITS_PER_WORD
) / GET_MODE_SIZE (mode
);
6544 /* We need to pass this field (partly) on the stack. */
6556 /* A subroutine of function_arg_record_value. Count the number of registers
6557 to be assigned for FIELD and between PARMS->intoffset and BITPOS. */
6560 count_registers (const_tree field
, HOST_WIDE_INT bitpos
, bool fp
,
6561 assign_data_t
*data
)
6568 if (compute_int_layout (bitpos
, data
, &nregs
))
6569 data
->nregs
+= nregs
;
6571 if (compute_fp_layout (field
, bitpos
, data
, &nregs
, &mode
))
6572 data
->nregs
+= nregs
;
6576 if (data
->intoffset
< 0)
6577 data
->intoffset
= bitpos
;
6581 /* A subroutine of function_arg_record_value. Assign the bits of the
6582 structure between PARMS->intoffset and BITPOS to integer registers. */
6585 assign_int_registers (HOST_WIDE_INT bitpos
, assign_data_t
*data
)
6587 int intoffset
= data
->intoffset
;
6591 if (!compute_int_layout (bitpos
, data
, &nregs
))
6594 /* If this is the trailing part of a word, only load that much into
6595 the register. Otherwise load the whole register. Note that in
6596 the latter case we may pick up unwanted bits. It's not a problem
6597 at the moment but may wish to revisit. */
6598 if (intoffset
% BITS_PER_WORD
!= 0)
6599 mode
= smallest_mode_for_size (BITS_PER_WORD
- intoffset
% BITS_PER_WORD
,
6604 const int this_slotno
= data
->slotno
+ intoffset
/ BITS_PER_WORD
;
6605 unsigned int regno
= data
->regbase
+ this_slotno
;
6606 intoffset
/= BITS_PER_UNIT
;
6610 rtx reg
= gen_rtx_REG (mode
, regno
);
6611 XVECEXP (data
->ret
, 0, data
->stack
+ data
->nregs
)
6612 = gen_rtx_EXPR_LIST (VOIDmode
, reg
, GEN_INT (intoffset
));
6616 intoffset
= (intoffset
| (UNITS_PER_WORD
- 1)) + 1;
6618 while (--nregs
> 0);
6621 /* A subroutine of function_arg_record_value. Assign FIELD at position
6622 BITPOS to FP registers. */
6625 assign_fp_registers (const_tree field
, HOST_WIDE_INT bitpos
,
6626 assign_data_t
*data
)
6631 if (!compute_fp_layout (field
, bitpos
, data
, &nregs
, &mode
))
6634 const int this_slotno
= data
->slotno
+ bitpos
/ BITS_PER_WORD
;
6635 int regno
= SPARC_FP_ARG_FIRST
+ this_slotno
* 2;
6636 if (GET_MODE_SIZE (mode
) <= 4 && (bitpos
& 32) != 0)
6638 int pos
= bitpos
/ BITS_PER_UNIT
;
6642 rtx reg
= gen_rtx_REG (mode
, regno
);
6643 XVECEXP (data
->ret
, 0, data
->stack
+ data
->nregs
)
6644 = gen_rtx_EXPR_LIST (VOIDmode
, reg
, GEN_INT (pos
));
6646 regno
+= GET_MODE_SIZE (mode
) / 4;
6647 pos
+= GET_MODE_SIZE (mode
);
6649 while (--nregs
> 0);
6652 /* A subroutine of function_arg_record_value. Assign FIELD and the bits of
6653 the structure between PARMS->intoffset and BITPOS to registers. */
6656 assign_registers (const_tree field
, HOST_WIDE_INT bitpos
, bool fp
,
6657 assign_data_t
*data
)
6661 assign_int_registers (bitpos
, data
);
6663 assign_fp_registers (field
, bitpos
, data
);
6667 if (data
->intoffset
< 0)
6668 data
->intoffset
= bitpos
;
6672 /* Used by function_arg and sparc_function_value_1 to implement the complex
6673 conventions of the 64-bit ABI for passing and returning structures.
6674 Return an expression valid as a return value for the FUNCTION_ARG
6675 and TARGET_FUNCTION_VALUE.
6677 TYPE is the data type of the argument (as a tree).
6678 This is null for libcalls where that information may
6680 MODE is the argument's machine mode.
6681 SLOTNO is the index number of the argument's slot in the parameter array.
6682 NAMED is true if this argument is a named parameter
6683 (otherwise it is an extra parameter matching an ellipsis).
6684 REGBASE is the regno of the base register for the parameter array. */
6687 function_arg_record_value (const_tree type
, machine_mode mode
,
6688 int slotno
, bool named
, int regbase
)
6690 HOST_WIDE_INT typesize
= int_size_in_bytes (type
);
6694 data
.slotno
= slotno
;
6695 data
.regbase
= regbase
;
6697 /* Count how many registers we need. */
6701 traverse_record_type
<assign_data_t
, count_registers
> (type
, named
, &data
);
6703 /* Take into account pending integer fields. */
6704 if (compute_int_layout (typesize
* BITS_PER_UNIT
, &data
, &nregs
))
6705 data
.nregs
+= nregs
;
6707 /* Allocate the vector and handle some annoying special cases. */
6712 /* ??? Empty structure has no value? Duh? */
6715 /* Though there's nothing really to store, return a word register
6716 anyway so the rest of gcc doesn't go nuts. Returning a PARALLEL
6717 leads to breakage due to the fact that there are zero bytes to
6719 return gen_rtx_REG (mode
, regbase
);
6722 /* ??? C++ has structures with no fields, and yet a size. Give up
6723 for now and pass everything back in integer registers. */
6724 nregs
= (typesize
+ UNITS_PER_WORD
- 1) / UNITS_PER_WORD
;
6725 if (nregs
+ slotno
> SPARC_INT_ARG_MAX
)
6726 nregs
= SPARC_INT_ARG_MAX
- slotno
;
6729 gcc_assert (nregs
> 0);
6731 data
.ret
= gen_rtx_PARALLEL (mode
, rtvec_alloc (data
.stack
+ nregs
));
6733 /* If at least one field must be passed on the stack, generate
6734 (parallel [(expr_list (nil) ...) ...]) so that all fields will
6735 also be passed on the stack. We can't do much better because the
6736 semantics of TARGET_ARG_PARTIAL_BYTES doesn't handle the case
6737 of structures for which the fields passed exclusively in registers
6738 are not at the beginning of the structure. */
6740 XVECEXP (data
.ret
, 0, 0)
6741 = gen_rtx_EXPR_LIST (VOIDmode
, NULL_RTX
, const0_rtx
);
6743 /* Assign the registers. */
6746 traverse_record_type
<assign_data_t
, assign_registers
> (type
, named
, &data
);
6748 /* Assign pending integer fields. */
6749 assign_int_registers (typesize
* BITS_PER_UNIT
, &data
);
6751 gcc_assert (data
.nregs
== nregs
);
6756 /* Used by function_arg and sparc_function_value_1 to implement the conventions
6757 of the 64-bit ABI for passing and returning unions.
6758 Return an expression valid as a return value for the FUNCTION_ARG
6759 and TARGET_FUNCTION_VALUE.
6761 SIZE is the size in bytes of the union.
6762 MODE is the argument's machine mode.
6763 REGNO is the hard register the union will be passed in. */
6766 function_arg_union_value (int size
, machine_mode mode
, int slotno
,
6769 int nwords
= CEIL_NWORDS (size
), i
;
6772 /* See comment in previous function for empty structures. */
6774 return gen_rtx_REG (mode
, regno
);
6776 if (slotno
== SPARC_INT_ARG_MAX
- 1)
6779 regs
= gen_rtx_PARALLEL (mode
, rtvec_alloc (nwords
));
6781 for (i
= 0; i
< nwords
; i
++)
6783 /* Unions are passed left-justified. */
6784 XVECEXP (regs
, 0, i
)
6785 = gen_rtx_EXPR_LIST (VOIDmode
,
6786 gen_rtx_REG (word_mode
, regno
),
6787 GEN_INT (UNITS_PER_WORD
* i
));
6794 /* Used by function_arg and sparc_function_value_1 to implement the conventions
6795 for passing and returning BLKmode vectors.
6796 Return an expression valid as a return value for the FUNCTION_ARG
6797 and TARGET_FUNCTION_VALUE.
6799 SIZE is the size in bytes of the vector.
6800 REGNO is the FP hard register the vector will be passed in. */
6803 function_arg_vector_value (int size
, int regno
)
6805 const int nregs
= MAX (1, size
/ 8);
6806 rtx regs
= gen_rtx_PARALLEL (BLKmode
, rtvec_alloc (nregs
));
6809 XVECEXP (regs
, 0, 0)
6810 = gen_rtx_EXPR_LIST (VOIDmode
,
6811 gen_rtx_REG (SImode
, regno
),
6814 for (int i
= 0; i
< nregs
; i
++)
6815 XVECEXP (regs
, 0, i
)
6816 = gen_rtx_EXPR_LIST (VOIDmode
,
6817 gen_rtx_REG (DImode
, regno
+ 2*i
),
6823 /* Determine where to put an argument to a function.
6824 Value is zero to push the argument on the stack,
6825 or a hard register in which to store the argument.
6827 CUM is a variable of type CUMULATIVE_ARGS which gives info about
6828 the preceding args and about the function being called.
6829 MODE is the argument's machine mode.
6830 TYPE is the data type of the argument (as a tree).
6831 This is null for libcalls where that information may
6833 NAMED is true if this argument is a named parameter
6834 (otherwise it is an extra parameter matching an ellipsis).
6835 INCOMING_P is false for TARGET_FUNCTION_ARG, true for
6836 TARGET_FUNCTION_INCOMING_ARG. */
6839 sparc_function_arg_1 (cumulative_args_t cum_v
, machine_mode mode
,
6840 const_tree type
, bool named
, bool incoming
)
6842 const CUMULATIVE_ARGS
*cum
= get_cumulative_args (cum_v
);
6844 int regbase
= (incoming
6845 ? SPARC_INCOMING_INT_ARG_FIRST
6846 : SPARC_OUTGOING_INT_ARG_FIRST
);
6847 int slotno
, regno
, padding
;
6848 enum mode_class mclass
= GET_MODE_CLASS (mode
);
6850 slotno
= function_arg_slotno (cum
, mode
, type
, named
, incoming
,
6855 /* Vector types deserve special treatment because they are polymorphic wrt
6856 their mode, depending upon whether VIS instructions are enabled. */
6857 if (type
&& TREE_CODE (type
) == VECTOR_TYPE
)
6859 HOST_WIDE_INT size
= int_size_in_bytes (type
);
6860 gcc_assert ((TARGET_ARCH32
&& size
<= 8)
6861 || (TARGET_ARCH64
&& size
<= 16));
6863 if (mode
== BLKmode
)
6864 return function_arg_vector_value (size
, SPARC_FP_ARG_FIRST
+ 2*slotno
);
6866 mclass
= MODE_FLOAT
;
6870 return gen_rtx_REG (mode
, regno
);
6872 /* Structures up to 16 bytes in size are passed in arg slots on the stack
6873 and are promoted to registers if possible. */
6874 if (type
&& TREE_CODE (type
) == RECORD_TYPE
)
6876 HOST_WIDE_INT size
= int_size_in_bytes (type
);
6877 gcc_assert (size
<= 16);
6879 return function_arg_record_value (type
, mode
, slotno
, named
, regbase
);
6882 /* Unions up to 16 bytes in size are passed in integer registers. */
6883 else if (type
&& TREE_CODE (type
) == UNION_TYPE
)
6885 HOST_WIDE_INT size
= int_size_in_bytes (type
);
6886 gcc_assert (size
<= 16);
6888 return function_arg_union_value (size
, mode
, slotno
, regno
);
6891 /* v9 fp args in reg slots beyond the int reg slots get passed in regs
6892 but also have the slot allocated for them.
6893 If no prototype is in scope fp values in register slots get passed
6894 in two places, either fp regs and int regs or fp regs and memory. */
6895 else if ((mclass
== MODE_FLOAT
|| mclass
== MODE_COMPLEX_FLOAT
)
6896 && SPARC_FP_REG_P (regno
))
6898 rtx reg
= gen_rtx_REG (mode
, regno
);
6899 if (cum
->prototype_p
|| cum
->libcall_p
)
6905 if ((regno
- SPARC_FP_ARG_FIRST
) < SPARC_INT_ARG_MAX
* 2)
6909 /* On incoming, we don't need to know that the value
6910 is passed in %f0 and %i0, and it confuses other parts
6911 causing needless spillage even on the simplest cases. */
6915 intreg
= (SPARC_OUTGOING_INT_ARG_FIRST
6916 + (regno
- SPARC_FP_ARG_FIRST
) / 2);
6918 v0
= gen_rtx_EXPR_LIST (VOIDmode
, reg
, const0_rtx
);
6919 v1
= gen_rtx_EXPR_LIST (VOIDmode
, gen_rtx_REG (mode
, intreg
),
6921 return gen_rtx_PARALLEL (mode
, gen_rtvec (2, v0
, v1
));
6925 v0
= gen_rtx_EXPR_LIST (VOIDmode
, NULL_RTX
, const0_rtx
);
6926 v1
= gen_rtx_EXPR_LIST (VOIDmode
, reg
, const0_rtx
);
6927 return gen_rtx_PARALLEL (mode
, gen_rtvec (2, v0
, v1
));
6932 /* All other aggregate types are passed in an integer register in a mode
6933 corresponding to the size of the type. */
6934 else if (type
&& AGGREGATE_TYPE_P (type
))
6936 HOST_WIDE_INT size
= int_size_in_bytes (type
);
6937 gcc_assert (size
<= 16);
6939 mode
= mode_for_size (size
* BITS_PER_UNIT
, MODE_INT
, 0);
6942 return gen_rtx_REG (mode
, regno
);
6945 /* Handle the TARGET_FUNCTION_ARG target hook. */
6948 sparc_function_arg (cumulative_args_t cum
, machine_mode mode
,
6949 const_tree type
, bool named
)
6951 return sparc_function_arg_1 (cum
, mode
, type
, named
, false);
6954 /* Handle the TARGET_FUNCTION_INCOMING_ARG target hook. */
6957 sparc_function_incoming_arg (cumulative_args_t cum
, machine_mode mode
,
6958 const_tree type
, bool named
)
6960 return sparc_function_arg_1 (cum
, mode
, type
, named
, true);
6963 /* For sparc64, objects requiring 16 byte alignment are passed that way. */
6966 sparc_function_arg_boundary (machine_mode mode
, const_tree type
)
6968 return ((TARGET_ARCH64
6969 && (GET_MODE_ALIGNMENT (mode
) == 128
6970 || (type
&& TYPE_ALIGN (type
) == 128)))
6975 /* For an arg passed partly in registers and partly in memory,
6976 this is the number of bytes of registers used.
6977 For args passed entirely in registers or entirely in memory, zero.
6979 Any arg that starts in the first 6 regs but won't entirely fit in them
6980 needs partial registers on v8. On v9, structures with integer
6981 values in arg slots 5,6 will be passed in %o5 and SP+176, and complex fp
6982 values that begin in the last fp reg [where "last fp reg" varies with the
6983 mode] will be split between that reg and memory. */
6986 sparc_arg_partial_bytes (cumulative_args_t cum
, machine_mode mode
,
6987 tree type
, bool named
)
6989 int slotno
, regno
, padding
;
6991 /* We pass false for incoming here, it doesn't matter. */
6992 slotno
= function_arg_slotno (get_cumulative_args (cum
), mode
, type
, named
,
6993 false, ®no
, &padding
);
7000 if ((slotno
+ (mode
== BLKmode
7001 ? CEIL_NWORDS (int_size_in_bytes (type
))
7002 : CEIL_NWORDS (GET_MODE_SIZE (mode
))))
7003 > SPARC_INT_ARG_MAX
)
7004 return (SPARC_INT_ARG_MAX
- slotno
) * UNITS_PER_WORD
;
7008 /* We are guaranteed by pass_by_reference that the size of the
7009 argument is not greater than 16 bytes, so we only need to return
7010 one word if the argument is partially passed in registers. */
7012 if (type
&& AGGREGATE_TYPE_P (type
))
7014 int size
= int_size_in_bytes (type
);
7016 if (size
> UNITS_PER_WORD
7017 && (slotno
== SPARC_INT_ARG_MAX
- 1
7018 || slotno
== SPARC_FP_ARG_MAX
- 1))
7019 return UNITS_PER_WORD
;
7021 else if (GET_MODE_CLASS (mode
) == MODE_COMPLEX_INT
7022 || (GET_MODE_CLASS (mode
) == MODE_COMPLEX_FLOAT
7023 && ! (TARGET_FPU
&& named
)))
7025 /* The complex types are passed as packed types. */
7026 if (GET_MODE_SIZE (mode
) > UNITS_PER_WORD
7027 && slotno
== SPARC_INT_ARG_MAX
- 1)
7028 return UNITS_PER_WORD
;
7030 else if (GET_MODE_CLASS (mode
) == MODE_COMPLEX_FLOAT
)
7032 if ((slotno
+ GET_MODE_SIZE (mode
) / UNITS_PER_WORD
)
7034 return UNITS_PER_WORD
;
7041 /* Handle the TARGET_PASS_BY_REFERENCE target hook.
7042 Specify whether to pass the argument by reference. */
7045 sparc_pass_by_reference (cumulative_args_t cum ATTRIBUTE_UNUSED
,
7046 machine_mode mode
, const_tree type
,
7047 bool named ATTRIBUTE_UNUSED
)
7050 /* Original SPARC 32-bit ABI says that structures and unions,
7051 and quad-precision floats are passed by reference. For Pascal,
7052 also pass arrays by reference. All other base types are passed
7055 Extended ABI (as implemented by the Sun compiler) says that all
7056 complex floats are passed by reference. Pass complex integers
7057 in registers up to 8 bytes. More generally, enforce the 2-word
7058 cap for passing arguments in registers.
7060 Vector ABI (as implemented by the Sun VIS SDK) says that vector
7061 integers are passed like floats of the same size, that is in
7062 registers up to 8 bytes. Pass all vector floats by reference
7063 like structure and unions. */
7064 return ((type
&& (AGGREGATE_TYPE_P (type
) || VECTOR_FLOAT_TYPE_P (type
)))
7066 /* Catch CDImode, TFmode, DCmode and TCmode. */
7067 || GET_MODE_SIZE (mode
) > 8
7069 && TREE_CODE (type
) == VECTOR_TYPE
7070 && (unsigned HOST_WIDE_INT
) int_size_in_bytes (type
) > 8));
7072 /* Original SPARC 64-bit ABI says that structures and unions
7073 smaller than 16 bytes are passed in registers, as well as
7074 all other base types.
7076 Extended ABI (as implemented by the Sun compiler) says that
7077 complex floats are passed in registers up to 16 bytes. Pass
7078 all complex integers in registers up to 16 bytes. More generally,
7079 enforce the 2-word cap for passing arguments in registers.
7081 Vector ABI (as implemented by the Sun VIS SDK) says that vector
7082 integers are passed like floats of the same size, that is in
7083 registers (up to 16 bytes). Pass all vector floats like structure
7086 && (AGGREGATE_TYPE_P (type
) || TREE_CODE (type
) == VECTOR_TYPE
)
7087 && (unsigned HOST_WIDE_INT
) int_size_in_bytes (type
) > 16)
7088 /* Catch CTImode and TCmode. */
7089 || GET_MODE_SIZE (mode
) > 16);
7092 /* Handle the TARGET_FUNCTION_ARG_ADVANCE hook.
7093 Update the data in CUM to advance over an argument
7094 of mode MODE and data type TYPE.
7095 TYPE is null for libcalls where that information may not be available. */
7098 sparc_function_arg_advance (cumulative_args_t cum_v
, machine_mode mode
,
7099 const_tree type
, bool named
)
7101 CUMULATIVE_ARGS
*cum
= get_cumulative_args (cum_v
);
7104 /* We pass false for incoming here, it doesn't matter. */
7105 function_arg_slotno (cum
, mode
, type
, named
, false, ®no
, &padding
);
7107 /* If argument requires leading padding, add it. */
7108 cum
->words
+= padding
;
7111 cum
->words
+= (mode
== BLKmode
7112 ? CEIL_NWORDS (int_size_in_bytes (type
))
7113 : CEIL_NWORDS (GET_MODE_SIZE (mode
)));
7116 if (type
&& AGGREGATE_TYPE_P (type
))
7118 int size
= int_size_in_bytes (type
);
7122 else if (size
<= 16)
7124 else /* passed by reference */
7128 cum
->words
+= (mode
== BLKmode
7129 ? CEIL_NWORDS (int_size_in_bytes (type
))
7130 : CEIL_NWORDS (GET_MODE_SIZE (mode
)));
7134 /* Handle the FUNCTION_ARG_PADDING macro.
7135 For the 64 bit ABI structs are always stored left shifted in their
7139 function_arg_padding (machine_mode mode
, const_tree type
)
7141 if (TARGET_ARCH64
&& type
&& AGGREGATE_TYPE_P (type
))
7144 /* Fall back to the default. */
7145 return DEFAULT_FUNCTION_ARG_PADDING (mode
, type
);
7148 /* Handle the TARGET_RETURN_IN_MEMORY target hook.
7149 Specify whether to return the return value in memory. */
7152 sparc_return_in_memory (const_tree type
, const_tree fntype ATTRIBUTE_UNUSED
)
7155 /* Original SPARC 32-bit ABI says that structures and unions,
7156 and quad-precision floats are returned in memory. All other
7157 base types are returned in registers.
7159 Extended ABI (as implemented by the Sun compiler) says that
7160 all complex floats are returned in registers (8 FP registers
7161 at most for '_Complex long double'). Return all complex integers
7162 in registers (4 at most for '_Complex long long').
7164 Vector ABI (as implemented by the Sun VIS SDK) says that vector
7165 integers are returned like floats of the same size, that is in
7166 registers up to 8 bytes and in memory otherwise. Return all
7167 vector floats in memory like structure and unions; note that
7168 they always have BLKmode like the latter. */
7169 return (TYPE_MODE (type
) == BLKmode
7170 || TYPE_MODE (type
) == TFmode
7171 || (TREE_CODE (type
) == VECTOR_TYPE
7172 && (unsigned HOST_WIDE_INT
) int_size_in_bytes (type
) > 8));
7174 /* Original SPARC 64-bit ABI says that structures and unions
7175 smaller than 32 bytes are returned in registers, as well as
7176 all other base types.
7178 Extended ABI (as implemented by the Sun compiler) says that all
7179 complex floats are returned in registers (8 FP registers at most
7180 for '_Complex long double'). Return all complex integers in
7181 registers (4 at most for '_Complex TItype').
7183 Vector ABI (as implemented by the Sun VIS SDK) says that vector
7184 integers are returned like floats of the same size, that is in
7185 registers. Return all vector floats like structure and unions;
7186 note that they always have BLKmode like the latter. */
7187 return (TYPE_MODE (type
) == BLKmode
7188 && (unsigned HOST_WIDE_INT
) int_size_in_bytes (type
) > 32);
7191 /* Handle the TARGET_STRUCT_VALUE target hook.
7192 Return where to find the structure return value address. */
7195 sparc_struct_value_rtx (tree fndecl
, int incoming
)
7204 mem
= gen_frame_mem (Pmode
, plus_constant (Pmode
, frame_pointer_rtx
,
7205 STRUCT_VALUE_OFFSET
));
7207 mem
= gen_frame_mem (Pmode
, plus_constant (Pmode
, stack_pointer_rtx
,
7208 STRUCT_VALUE_OFFSET
));
7210 /* Only follow the SPARC ABI for fixed-size structure returns.
7211 Variable size structure returns are handled per the normal
7212 procedures in GCC. This is enabled by -mstd-struct-return */
7214 && sparc_std_struct_return
7215 && TYPE_SIZE_UNIT (TREE_TYPE (fndecl
))
7216 && TREE_CODE (TYPE_SIZE_UNIT (TREE_TYPE (fndecl
))) == INTEGER_CST
)
7218 /* We must check and adjust the return address, as it is optional
7219 as to whether the return object is really provided. */
7220 rtx ret_reg
= gen_rtx_REG (Pmode
, RETURN_ADDR_REGNUM
);
7221 rtx scratch
= gen_reg_rtx (SImode
);
7222 rtx_code_label
*endlab
= gen_label_rtx ();
7224 /* Calculate the return object size. */
7225 tree size
= TYPE_SIZE_UNIT (TREE_TYPE (fndecl
));
7226 rtx size_rtx
= GEN_INT (TREE_INT_CST_LOW (size
) & 0xfff);
7227 /* Construct a temporary return value. */
7229 = assign_stack_local (Pmode
, TREE_INT_CST_LOW (size
), 0);
7231 /* Implement SPARC 32-bit psABI callee return struct checking:
7233 Fetch the instruction where we will return to and see if
7234 it's an unimp instruction (the most significant 10 bits
7236 emit_move_insn (scratch
, gen_rtx_MEM (SImode
,
7237 plus_constant (Pmode
,
7239 /* Assume the size is valid and pre-adjust. */
7240 emit_insn (gen_add3_insn (ret_reg
, ret_reg
, GEN_INT (4)));
7241 emit_cmp_and_jump_insns (scratch
, size_rtx
, EQ
, const0_rtx
, SImode
,
7243 emit_insn (gen_sub3_insn (ret_reg
, ret_reg
, GEN_INT (4)));
7244 /* Write the address of the memory pointed to by temp_val into
7245 the memory pointed to by mem. */
7246 emit_move_insn (mem
, XEXP (temp_val
, 0));
7247 emit_label (endlab
);
7254 /* Handle TARGET_FUNCTION_VALUE, and TARGET_LIBCALL_VALUE target hook.
7255 For v9, function return values are subject to the same rules as arguments,
7256 except that up to 32 bytes may be returned in registers. */
7259 sparc_function_value_1 (const_tree type
, machine_mode mode
,
7262 /* Beware that the two values are swapped here wrt function_arg. */
7263 int regbase
= (outgoing
7264 ? SPARC_INCOMING_INT_ARG_FIRST
7265 : SPARC_OUTGOING_INT_ARG_FIRST
);
7266 enum mode_class mclass
= GET_MODE_CLASS (mode
);
7269 /* Vector types deserve special treatment because they are polymorphic wrt
7270 their mode, depending upon whether VIS instructions are enabled. */
7271 if (type
&& TREE_CODE (type
) == VECTOR_TYPE
)
7273 HOST_WIDE_INT size
= int_size_in_bytes (type
);
7274 gcc_assert ((TARGET_ARCH32
&& size
<= 8)
7275 || (TARGET_ARCH64
&& size
<= 32));
7277 if (mode
== BLKmode
)
7278 return function_arg_vector_value (size
, SPARC_FP_ARG_FIRST
);
7280 mclass
= MODE_FLOAT
;
7283 if (TARGET_ARCH64
&& type
)
7285 /* Structures up to 32 bytes in size are returned in registers. */
7286 if (TREE_CODE (type
) == RECORD_TYPE
)
7288 HOST_WIDE_INT size
= int_size_in_bytes (type
);
7289 gcc_assert (size
<= 32);
7291 return function_arg_record_value (type
, mode
, 0, 1, regbase
);
7294 /* Unions up to 32 bytes in size are returned in integer registers. */
7295 else if (TREE_CODE (type
) == UNION_TYPE
)
7297 HOST_WIDE_INT size
= int_size_in_bytes (type
);
7298 gcc_assert (size
<= 32);
7300 return function_arg_union_value (size
, mode
, 0, regbase
);
7303 /* Objects that require it are returned in FP registers. */
7304 else if (mclass
== MODE_FLOAT
|| mclass
== MODE_COMPLEX_FLOAT
)
7307 /* All other aggregate types are returned in an integer register in a
7308 mode corresponding to the size of the type. */
7309 else if (AGGREGATE_TYPE_P (type
))
7311 /* All other aggregate types are passed in an integer register
7312 in a mode corresponding to the size of the type. */
7313 HOST_WIDE_INT size
= int_size_in_bytes (type
);
7314 gcc_assert (size
<= 32);
7316 mode
= mode_for_size (size
* BITS_PER_UNIT
, MODE_INT
, 0);
7318 /* ??? We probably should have made the same ABI change in
7319 3.4.0 as the one we made for unions. The latter was
7320 required by the SCD though, while the former is not
7321 specified, so we favored compatibility and efficiency.
7323 Now we're stuck for aggregates larger than 16 bytes,
7324 because OImode vanished in the meantime. Let's not
7325 try to be unduly clever, and simply follow the ABI
7326 for unions in that case. */
7327 if (mode
== BLKmode
)
7328 return function_arg_union_value (size
, mode
, 0, regbase
);
7333 /* We should only have pointer and integer types at this point. This
7334 must match sparc_promote_function_mode. */
7335 else if (mclass
== MODE_INT
&& GET_MODE_SIZE (mode
) < UNITS_PER_WORD
)
7339 /* We should only have pointer and integer types at this point, except with
7340 -freg-struct-return. This must match sparc_promote_function_mode. */
7341 else if (TARGET_ARCH32
7342 && !(type
&& AGGREGATE_TYPE_P (type
))
7343 && mclass
== MODE_INT
7344 && GET_MODE_SIZE (mode
) < UNITS_PER_WORD
)
7347 if ((mclass
== MODE_FLOAT
|| mclass
== MODE_COMPLEX_FLOAT
) && TARGET_FPU
)
7348 regno
= SPARC_FP_ARG_FIRST
;
7352 return gen_rtx_REG (mode
, regno
);
7355 /* Handle TARGET_FUNCTION_VALUE.
7356 On the SPARC, the value is found in the first "output" register, but the
7357 called function leaves it in the first "input" register. */
7360 sparc_function_value (const_tree valtype
,
7361 const_tree fn_decl_or_type ATTRIBUTE_UNUSED
,
7364 return sparc_function_value_1 (valtype
, TYPE_MODE (valtype
), outgoing
);
7367 /* Handle TARGET_LIBCALL_VALUE. */
7370 sparc_libcall_value (machine_mode mode
,
7371 const_rtx fun ATTRIBUTE_UNUSED
)
7373 return sparc_function_value_1 (NULL_TREE
, mode
, false);
7376 /* Handle FUNCTION_VALUE_REGNO_P.
7377 On the SPARC, the first "output" reg is used for integer values, and the
7378 first floating point register is used for floating point values. */
7381 sparc_function_value_regno_p (const unsigned int regno
)
7383 return (regno
== 8 || (TARGET_FPU
&& regno
== 32));
7386 /* Do what is necessary for `va_start'. We look at the current function
7387 to determine if stdarg or varargs is used and return the address of
7388 the first unnamed parameter. */
7391 sparc_builtin_saveregs (void)
7393 int first_reg
= crtl
->args
.info
.words
;
7397 for (regno
= first_reg
; regno
< SPARC_INT_ARG_MAX
; regno
++)
7398 emit_move_insn (gen_rtx_MEM (word_mode
,
7399 gen_rtx_PLUS (Pmode
,
7401 GEN_INT (FIRST_PARM_OFFSET (0)
7404 gen_rtx_REG (word_mode
,
7405 SPARC_INCOMING_INT_ARG_FIRST
+ regno
));
7407 address
= gen_rtx_PLUS (Pmode
,
7409 GEN_INT (FIRST_PARM_OFFSET (0)
7410 + UNITS_PER_WORD
* first_reg
));
7415 /* Implement `va_start' for stdarg. */
7418 sparc_va_start (tree valist
, rtx nextarg
)
7420 nextarg
= expand_builtin_saveregs ();
7421 std_expand_builtin_va_start (valist
, nextarg
);
7424 /* Implement `va_arg' for stdarg. */
7427 sparc_gimplify_va_arg (tree valist
, tree type
, gimple_seq
*pre_p
,
7430 HOST_WIDE_INT size
, rsize
, align
;
7433 tree ptrtype
= build_pointer_type (type
);
7435 if (pass_by_reference (NULL
, TYPE_MODE (type
), type
, false))
7438 size
= rsize
= UNITS_PER_WORD
;
7444 size
= int_size_in_bytes (type
);
7445 rsize
= ROUND_UP (size
, UNITS_PER_WORD
);
7450 /* For SPARC64, objects requiring 16-byte alignment get it. */
7451 if (TYPE_ALIGN (type
) >= 2 * (unsigned) BITS_PER_WORD
)
7452 align
= 2 * UNITS_PER_WORD
;
7454 /* SPARC-V9 ABI states that structures up to 16 bytes in size
7455 are left-justified in their slots. */
7456 if (AGGREGATE_TYPE_P (type
))
7459 size
= rsize
= UNITS_PER_WORD
;
7469 incr
= fold_build_pointer_plus_hwi (incr
, align
- 1);
7470 incr
= fold_convert (sizetype
, incr
);
7471 incr
= fold_build2 (BIT_AND_EXPR
, sizetype
, incr
,
7473 incr
= fold_convert (ptr_type_node
, incr
);
7476 gimplify_expr (&incr
, pre_p
, post_p
, is_gimple_val
, fb_rvalue
);
7479 if (BYTES_BIG_ENDIAN
&& size
< rsize
)
7480 addr
= fold_build_pointer_plus_hwi (incr
, rsize
- size
);
7484 addr
= fold_convert (build_pointer_type (ptrtype
), addr
);
7485 addr
= build_va_arg_indirect_ref (addr
);
7488 /* If the address isn't aligned properly for the type, we need a temporary.
7489 FIXME: This is inefficient, usually we can do this in registers. */
7490 else if (align
== 0 && TYPE_ALIGN (type
) > BITS_PER_WORD
)
7492 tree tmp
= create_tmp_var (type
, "va_arg_tmp");
7493 tree dest_addr
= build_fold_addr_expr (tmp
);
7494 tree copy
= build_call_expr (builtin_decl_implicit (BUILT_IN_MEMCPY
),
7495 3, dest_addr
, addr
, size_int (rsize
));
7496 TREE_ADDRESSABLE (tmp
) = 1;
7497 gimplify_and_add (copy
, pre_p
);
7502 addr
= fold_convert (ptrtype
, addr
);
7504 incr
= fold_build_pointer_plus_hwi (incr
, rsize
);
7505 gimplify_assign (valist
, incr
, post_p
);
7507 return build_va_arg_indirect_ref (addr
);
7510 /* Implement the TARGET_VECTOR_MODE_SUPPORTED_P target hook.
7511 Specify whether the vector mode is supported by the hardware. */
7514 sparc_vector_mode_supported_p (machine_mode mode
)
7516 return TARGET_VIS
&& VECTOR_MODE_P (mode
) ? true : false;
7519 /* Implement the TARGET_VECTORIZE_PREFERRED_SIMD_MODE target hook. */
7522 sparc_preferred_simd_mode (machine_mode mode
)
7540 /* Return the string to output an unconditional branch to LABEL, which is
7541 the operand number of the label.
7543 DEST is the destination insn (i.e. the label), INSN is the source. */
7546 output_ubranch (rtx dest
, rtx_insn
*insn
)
7548 static char string
[64];
7549 bool v9_form
= false;
7553 /* Even if we are trying to use cbcond for this, evaluate
7554 whether we can use V9 branches as our backup plan. */
7557 if (INSN_ADDRESSES_SET_P ())
7558 delta
= (INSN_ADDRESSES (INSN_UID (dest
))
7559 - INSN_ADDRESSES (INSN_UID (insn
)));
7561 /* Leave some instructions for "slop". */
7562 if (TARGET_V9
&& delta
>= -260000 && delta
< 260000)
7567 bool emit_nop
= emit_cbcond_nop (insn
);
7571 if (delta
< -500 || delta
> 500)
7577 rval
= "ba,a,pt\t%%xcc, %l0";
7584 rval
= "cwbe\t%%g0, %%g0, %l0\n\tnop";
7586 rval
= "cwbe\t%%g0, %%g0, %l0";
7592 strcpy (string
, "ba%*,pt\t%%xcc, ");
7594 strcpy (string
, "b%*\t");
7596 p
= strchr (string
, '\0');
7607 /* Return the string to output a conditional branch to LABEL, which is
7608 the operand number of the label. OP is the conditional expression.
7609 XEXP (OP, 0) is assumed to be a condition code register (integer or
7610 floating point) and its mode specifies what kind of comparison we made.
7612 DEST is the destination insn (i.e. the label), INSN is the source.
7614 REVERSED is nonzero if we should reverse the sense of the comparison.
7616 ANNUL is nonzero if we should generate an annulling branch. */
7619 output_cbranch (rtx op
, rtx dest
, int label
, int reversed
, int annul
,
7622 static char string
[64];
7623 enum rtx_code code
= GET_CODE (op
);
7624 rtx cc_reg
= XEXP (op
, 0);
7625 machine_mode mode
= GET_MODE (cc_reg
);
7626 const char *labelno
, *branch
;
7627 int spaces
= 8, far
;
7630 /* v9 branches are limited to +-1MB. If it is too far away,
7643 fbne,a,pn %fcc2, .LC29
7651 far
= TARGET_V9
&& (get_attr_length (insn
) >= 3);
7654 /* Reversal of FP compares takes care -- an ordered compare
7655 becomes an unordered compare and vice versa. */
7656 if (mode
== CCFPmode
|| mode
== CCFPEmode
)
7657 code
= reverse_condition_maybe_unordered (code
);
7659 code
= reverse_condition (code
);
7662 /* Start by writing the branch condition. */
7663 if (mode
== CCFPmode
|| mode
== CCFPEmode
)
7714 /* ??? !v9: FP branches cannot be preceded by another floating point
7715 insn. Because there is currently no concept of pre-delay slots,
7716 we can fix this only by always emitting a nop before a floating
7721 strcpy (string
, "nop\n\t");
7722 strcat (string
, branch
);
7735 if (mode
== CC_NOOVmode
|| mode
== CCX_NOOVmode
)
7747 if (mode
== CC_NOOVmode
|| mode
== CCX_NOOVmode
)
7768 strcpy (string
, branch
);
7770 spaces
-= strlen (branch
);
7771 p
= strchr (string
, '\0');
7773 /* Now add the annulling, the label, and a possible noop. */
7786 if (! far
&& insn
&& INSN_ADDRESSES_SET_P ())
7788 int delta
= (INSN_ADDRESSES (INSN_UID (dest
))
7789 - INSN_ADDRESSES (INSN_UID (insn
)));
7790 /* Leave some instructions for "slop". */
7791 if (delta
< -260000 || delta
>= 260000)
7795 if (mode
== CCFPmode
|| mode
== CCFPEmode
)
7797 static char v9_fcc_labelno
[] = "%%fccX, ";
7798 /* Set the char indicating the number of the fcc reg to use. */
7799 v9_fcc_labelno
[5] = REGNO (cc_reg
) - SPARC_FIRST_V9_FCC_REG
+ '0';
7800 labelno
= v9_fcc_labelno
;
7803 gcc_assert (REGNO (cc_reg
) == SPARC_FCC_REG
);
7807 else if (mode
== CCXmode
|| mode
== CCX_NOOVmode
)
7809 labelno
= "%%xcc, ";
7814 labelno
= "%%icc, ";
7819 if (*labelno
&& insn
&& (note
= find_reg_note (insn
, REG_BR_PROB
, NULL_RTX
)))
7822 ((XINT (note
, 0) >= REG_BR_PROB_BASE
/ 2) ^ far
)
7835 strcpy (p
, labelno
);
7836 p
= strchr (p
, '\0');
7839 strcpy (p
, ".+12\n\t nop\n\tb\t");
7840 /* Skip the next insn if requested or
7841 if we know that it will be a nop. */
7842 if (annul
|| ! final_sequence
)
7856 /* Emit a library call comparison between floating point X and Y.
7857 COMPARISON is the operator to compare with (EQ, NE, GT, etc).
7858 Return the new operator to be used in the comparison sequence.
7860 TARGET_ARCH64 uses _Qp_* functions, which use pointers to TFmode
7861 values as arguments instead of the TFmode registers themselves,
7862 that's why we cannot call emit_float_lib_cmp. */
7865 sparc_emit_float_lib_cmp (rtx x
, rtx y
, enum rtx_code comparison
)
7868 rtx slot0
, slot1
, result
, tem
, tem2
, libfunc
;
7870 enum rtx_code new_comparison
;
7875 qpfunc
= (TARGET_ARCH64
? "_Qp_feq" : "_Q_feq");
7879 qpfunc
= (TARGET_ARCH64
? "_Qp_fne" : "_Q_fne");
7883 qpfunc
= (TARGET_ARCH64
? "_Qp_fgt" : "_Q_fgt");
7887 qpfunc
= (TARGET_ARCH64
? "_Qp_fge" : "_Q_fge");
7891 qpfunc
= (TARGET_ARCH64
? "_Qp_flt" : "_Q_flt");
7895 qpfunc
= (TARGET_ARCH64
? "_Qp_fle" : "_Q_fle");
7906 qpfunc
= (TARGET_ARCH64
? "_Qp_cmp" : "_Q_cmp");
7917 tree expr
= MEM_EXPR (x
);
7919 mark_addressable (expr
);
7924 slot0
= assign_stack_temp (TFmode
, GET_MODE_SIZE(TFmode
));
7925 emit_move_insn (slot0
, x
);
7930 tree expr
= MEM_EXPR (y
);
7932 mark_addressable (expr
);
7937 slot1
= assign_stack_temp (TFmode
, GET_MODE_SIZE(TFmode
));
7938 emit_move_insn (slot1
, y
);
7941 libfunc
= gen_rtx_SYMBOL_REF (Pmode
, qpfunc
);
7942 emit_library_call (libfunc
, LCT_NORMAL
,
7944 XEXP (slot0
, 0), Pmode
,
7945 XEXP (slot1
, 0), Pmode
);
7950 libfunc
= gen_rtx_SYMBOL_REF (Pmode
, qpfunc
);
7951 emit_library_call (libfunc
, LCT_NORMAL
,
7953 x
, TFmode
, y
, TFmode
);
7958 /* Immediately move the result of the libcall into a pseudo
7959 register so reload doesn't clobber the value if it needs
7960 the return register for a spill reg. */
7961 result
= gen_reg_rtx (mode
);
7962 emit_move_insn (result
, hard_libcall_value (mode
, libfunc
));
7967 return gen_rtx_NE (VOIDmode
, result
, const0_rtx
);
7970 new_comparison
= (comparison
== UNORDERED
? EQ
: NE
);
7971 return gen_rtx_fmt_ee (new_comparison
, VOIDmode
, result
, GEN_INT(3));
7974 new_comparison
= (comparison
== UNGT
? GT
: NE
);
7975 return gen_rtx_fmt_ee (new_comparison
, VOIDmode
, result
, const1_rtx
);
7977 return gen_rtx_NE (VOIDmode
, result
, const2_rtx
);
7979 tem
= gen_reg_rtx (mode
);
7981 emit_insn (gen_andsi3 (tem
, result
, const1_rtx
));
7983 emit_insn (gen_anddi3 (tem
, result
, const1_rtx
));
7984 return gen_rtx_NE (VOIDmode
, tem
, const0_rtx
);
7987 tem
= gen_reg_rtx (mode
);
7989 emit_insn (gen_addsi3 (tem
, result
, const1_rtx
));
7991 emit_insn (gen_adddi3 (tem
, result
, const1_rtx
));
7992 tem2
= gen_reg_rtx (mode
);
7994 emit_insn (gen_andsi3 (tem2
, tem
, const2_rtx
));
7996 emit_insn (gen_anddi3 (tem2
, tem
, const2_rtx
));
7997 new_comparison
= (comparison
== UNEQ
? EQ
: NE
);
7998 return gen_rtx_fmt_ee (new_comparison
, VOIDmode
, tem2
, const0_rtx
);
8004 /* Generate an unsigned DImode to FP conversion. This is the same code
8005 optabs would emit if we didn't have TFmode patterns. */
8008 sparc_emit_floatunsdi (rtx
*operands
, machine_mode mode
)
8010 rtx i0
, i1
, f0
, in
, out
;
8013 in
= force_reg (DImode
, operands
[1]);
8014 rtx_code_label
*neglab
= gen_label_rtx ();
8015 rtx_code_label
*donelab
= gen_label_rtx ();
8016 i0
= gen_reg_rtx (DImode
);
8017 i1
= gen_reg_rtx (DImode
);
8018 f0
= gen_reg_rtx (mode
);
8020 emit_cmp_and_jump_insns (in
, const0_rtx
, LT
, const0_rtx
, DImode
, 0, neglab
);
8022 emit_insn (gen_rtx_SET (out
, gen_rtx_FLOAT (mode
, in
)));
8023 emit_jump_insn (gen_jump (donelab
));
8026 emit_label (neglab
);
8028 emit_insn (gen_lshrdi3 (i0
, in
, const1_rtx
));
8029 emit_insn (gen_anddi3 (i1
, in
, const1_rtx
));
8030 emit_insn (gen_iordi3 (i0
, i0
, i1
));
8031 emit_insn (gen_rtx_SET (f0
, gen_rtx_FLOAT (mode
, i0
)));
8032 emit_insn (gen_rtx_SET (out
, gen_rtx_PLUS (mode
, f0
, f0
)));
8034 emit_label (donelab
);
8037 /* Generate an FP to unsigned DImode conversion. This is the same code
8038 optabs would emit if we didn't have TFmode patterns. */
8041 sparc_emit_fixunsdi (rtx
*operands
, machine_mode mode
)
8043 rtx i0
, i1
, f0
, in
, out
, limit
;
8046 in
= force_reg (mode
, operands
[1]);
8047 rtx_code_label
*neglab
= gen_label_rtx ();
8048 rtx_code_label
*donelab
= gen_label_rtx ();
8049 i0
= gen_reg_rtx (DImode
);
8050 i1
= gen_reg_rtx (DImode
);
8051 limit
= gen_reg_rtx (mode
);
8052 f0
= gen_reg_rtx (mode
);
8054 emit_move_insn (limit
,
8055 const_double_from_real_value (
8056 REAL_VALUE_ATOF ("9223372036854775808.0", mode
), mode
));
8057 emit_cmp_and_jump_insns (in
, limit
, GE
, NULL_RTX
, mode
, 0, neglab
);
8059 emit_insn (gen_rtx_SET (out
,
8060 gen_rtx_FIX (DImode
, gen_rtx_FIX (mode
, in
))));
8061 emit_jump_insn (gen_jump (donelab
));
8064 emit_label (neglab
);
8066 emit_insn (gen_rtx_SET (f0
, gen_rtx_MINUS (mode
, in
, limit
)));
8067 emit_insn (gen_rtx_SET (i0
,
8068 gen_rtx_FIX (DImode
, gen_rtx_FIX (mode
, f0
))));
8069 emit_insn (gen_movdi (i1
, const1_rtx
));
8070 emit_insn (gen_ashldi3 (i1
, i1
, GEN_INT (63)));
8071 emit_insn (gen_xordi3 (out
, i0
, i1
));
8073 emit_label (donelab
);
8076 /* Return the string to output a compare and branch instruction to DEST.
8077 DEST is the destination insn (i.e. the label), INSN is the source,
8078 and OP is the conditional expression. */
8081 output_cbcond (rtx op
, rtx dest
, rtx_insn
*insn
)
8083 machine_mode mode
= GET_MODE (XEXP (op
, 0));
8084 enum rtx_code code
= GET_CODE (op
);
8085 const char *cond_str
, *tmpl
;
8086 int far
, emit_nop
, len
;
8087 static char string
[64];
8090 /* Compare and Branch is limited to +-2KB. If it is too far away,
8102 len
= get_attr_length (insn
);
8105 emit_nop
= len
== 2;
8108 code
= reverse_condition (code
);
8110 size_char
= ((mode
== SImode
) ? 'w' : 'x');
8123 if (mode
== CC_NOOVmode
|| mode
== CCX_NOOVmode
)
8138 if (mode
== CC_NOOVmode
|| mode
== CCX_NOOVmode
)
8166 int veryfar
= 1, delta
;
8168 if (INSN_ADDRESSES_SET_P ())
8170 delta
= (INSN_ADDRESSES (INSN_UID (dest
))
8171 - INSN_ADDRESSES (INSN_UID (insn
)));
8172 /* Leave some instructions for "slop". */
8173 if (delta
>= -260000 && delta
< 260000)
8178 tmpl
= "c%cb%s\t%%1, %%2, .+16\n\tnop\n\tb\t%%3\n\tnop";
8180 tmpl
= "c%cb%s\t%%1, %%2, .+16\n\tnop\n\tba,pt\t%%%%xcc, %%3\n\tnop";
8185 tmpl
= "c%cb%s\t%%1, %%2, %%3\n\tnop";
8187 tmpl
= "c%cb%s\t%%1, %%2, %%3";
8190 snprintf (string
, sizeof(string
), tmpl
, size_char
, cond_str
);
8195 /* Return the string to output a conditional branch to LABEL, testing
8196 register REG. LABEL is the operand number of the label; REG is the
8197 operand number of the reg. OP is the conditional expression. The mode
8198 of REG says what kind of comparison we made.
8200 DEST is the destination insn (i.e. the label), INSN is the source.
8202 REVERSED is nonzero if we should reverse the sense of the comparison.
8204 ANNUL is nonzero if we should generate an annulling branch. */
8207 output_v9branch (rtx op
, rtx dest
, int reg
, int label
, int reversed
,
8208 int annul
, rtx_insn
*insn
)
8210 static char string
[64];
8211 enum rtx_code code
= GET_CODE (op
);
8212 machine_mode mode
= GET_MODE (XEXP (op
, 0));
8217 /* branch on register are limited to +-128KB. If it is too far away,
8230 brgez,a,pn %o1, .LC29
8236 ba,pt %xcc, .LC29 */
8238 far
= get_attr_length (insn
) >= 3;
8240 /* If not floating-point or if EQ or NE, we can just reverse the code. */
8242 code
= reverse_condition (code
);
8244 /* Only 64 bit versions of these instructions exist. */
8245 gcc_assert (mode
== DImode
);
8247 /* Start by writing the branch condition. */
8252 strcpy (string
, "brnz");
8256 strcpy (string
, "brz");
8260 strcpy (string
, "brgez");
8264 strcpy (string
, "brlz");
8268 strcpy (string
, "brlez");
8272 strcpy (string
, "brgz");
8279 p
= strchr (string
, '\0');
8281 /* Now add the annulling, reg, label, and nop. */
8288 if (insn
&& (note
= find_reg_note (insn
, REG_BR_PROB
, NULL_RTX
)))
8291 ((XINT (note
, 0) >= REG_BR_PROB_BASE
/ 2) ^ far
)
8296 *p
= p
< string
+ 8 ? '\t' : ' ';
8304 int veryfar
= 1, delta
;
8306 if (INSN_ADDRESSES_SET_P ())
8308 delta
= (INSN_ADDRESSES (INSN_UID (dest
))
8309 - INSN_ADDRESSES (INSN_UID (insn
)));
8310 /* Leave some instructions for "slop". */
8311 if (delta
>= -260000 && delta
< 260000)
8315 strcpy (p
, ".+12\n\t nop\n\t");
8316 /* Skip the next insn if requested or
8317 if we know that it will be a nop. */
8318 if (annul
|| ! final_sequence
)
8328 strcpy (p
, "ba,pt\t%%xcc, ");
8342 /* Return 1, if any of the registers of the instruction are %l[0-7] or %o[0-7].
8343 Such instructions cannot be used in the delay slot of return insn on v9.
8344 If TEST is 0, also rename all %i[0-7] registers to their %o[0-7] counterparts.
8348 epilogue_renumber (register rtx
*where
, int test
)
8350 register const char *fmt
;
8352 register enum rtx_code code
;
8357 code
= GET_CODE (*where
);
8362 if (REGNO (*where
) >= 8 && REGNO (*where
) < 24) /* oX or lX */
8364 if (! test
&& REGNO (*where
) >= 24 && REGNO (*where
) < 32)
8365 *where
= gen_rtx_REG (GET_MODE (*where
), OUTGOING_REGNO (REGNO(*where
)));
8370 case CONST_WIDE_INT
:
8374 /* Do not replace the frame pointer with the stack pointer because
8375 it can cause the delayed instruction to load below the stack.
8376 This occurs when instructions like:
8378 (set (reg/i:SI 24 %i0)
8379 (mem/f:SI (plus:SI (reg/f:SI 30 %fp)
8380 (const_int -20 [0xffffffec])) 0))
8382 are in the return delayed slot. */
8384 if (GET_CODE (XEXP (*where
, 0)) == REG
8385 && REGNO (XEXP (*where
, 0)) == HARD_FRAME_POINTER_REGNUM
8386 && (GET_CODE (XEXP (*where
, 1)) != CONST_INT
8387 || INTVAL (XEXP (*where
, 1)) < SPARC_STACK_BIAS
))
8392 if (SPARC_STACK_BIAS
8393 && GET_CODE (XEXP (*where
, 0)) == REG
8394 && REGNO (XEXP (*where
, 0)) == HARD_FRAME_POINTER_REGNUM
)
8402 fmt
= GET_RTX_FORMAT (code
);
8404 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
8409 for (j
= XVECLEN (*where
, i
) - 1; j
>= 0; j
--)
8410 if (epilogue_renumber (&(XVECEXP (*where
, i
, j
)), test
))
8413 else if (fmt
[i
] == 'e'
8414 && epilogue_renumber (&(XEXP (*where
, i
)), test
))
8420 /* Leaf functions and non-leaf functions have different needs. */
8423 reg_leaf_alloc_order
[] = REG_LEAF_ALLOC_ORDER
;
8426 reg_nonleaf_alloc_order
[] = REG_ALLOC_ORDER
;
8428 static const int *const reg_alloc_orders
[] = {
8429 reg_leaf_alloc_order
,
8430 reg_nonleaf_alloc_order
};
8433 order_regs_for_local_alloc (void)
8435 static int last_order_nonleaf
= 1;
8437 if (df_regs_ever_live_p (15) != last_order_nonleaf
)
8439 last_order_nonleaf
= !last_order_nonleaf
;
8440 memcpy ((char *) reg_alloc_order
,
8441 (const char *) reg_alloc_orders
[last_order_nonleaf
],
8442 FIRST_PSEUDO_REGISTER
* sizeof (int));
8446 /* Return 1 if REG and MEM are legitimate enough to allow the various
8447 mem<-->reg splits to be run. */
8450 sparc_splitdi_legitimate (rtx reg
, rtx mem
)
8452 /* Punt if we are here by mistake. */
8453 gcc_assert (reload_completed
);
8455 /* We must have an offsettable memory reference. */
8456 if (! offsettable_memref_p (mem
))
8459 /* If we have legitimate args for ldd/std, we do not want
8460 the split to happen. */
8461 if ((REGNO (reg
) % 2) == 0
8462 && mem_min_alignment (mem
, 8))
8469 /* Like sparc_splitdi_legitimate but for REG <--> REG moves. */
8472 sparc_split_regreg_legitimate (rtx reg1
, rtx reg2
)
8476 if (GET_CODE (reg1
) == SUBREG
)
8477 reg1
= SUBREG_REG (reg1
);
8478 if (GET_CODE (reg1
) != REG
)
8480 regno1
= REGNO (reg1
);
8482 if (GET_CODE (reg2
) == SUBREG
)
8483 reg2
= SUBREG_REG (reg2
);
8484 if (GET_CODE (reg2
) != REG
)
8486 regno2
= REGNO (reg2
);
8488 if (SPARC_INT_REG_P (regno1
) && SPARC_INT_REG_P (regno2
))
8493 if ((SPARC_INT_REG_P (regno1
) && SPARC_FP_REG_P (regno2
))
8494 || (SPARC_FP_REG_P (regno1
) && SPARC_INT_REG_P (regno2
)))
8501 /* Return 1 if REGNO (reg1) is even and REGNO (reg1) == REGNO (reg2) - 1.
8502 This makes them candidates for using ldd and std insns.
8504 Note reg1 and reg2 *must* be hard registers. */
8507 registers_ok_for_ldd_peep (rtx reg1
, rtx reg2
)
8509 /* We might have been passed a SUBREG. */
8510 if (GET_CODE (reg1
) != REG
|| GET_CODE (reg2
) != REG
)
8513 if (REGNO (reg1
) % 2 != 0)
8516 /* Integer ldd is deprecated in SPARC V9 */
8517 if (TARGET_V9
&& SPARC_INT_REG_P (REGNO (reg1
)))
8520 return (REGNO (reg1
) == REGNO (reg2
) - 1);
8523 /* Return 1 if the addresses in mem1 and mem2 are suitable for use in
8526 This can only happen when addr1 and addr2, the addresses in mem1
8527 and mem2, are consecutive memory locations (addr1 + 4 == addr2).
8528 addr1 must also be aligned on a 64-bit boundary.
8530 Also iff dependent_reg_rtx is not null it should not be used to
8531 compute the address for mem1, i.e. we cannot optimize a sequence
8543 But, note that the transformation from:
8548 is perfectly fine. Thus, the peephole2 patterns always pass us
8549 the destination register of the first load, never the second one.
8551 For stores we don't have a similar problem, so dependent_reg_rtx is
8555 mems_ok_for_ldd_peep (rtx mem1
, rtx mem2
, rtx dependent_reg_rtx
)
8559 HOST_WIDE_INT offset1
;
8561 /* The mems cannot be volatile. */
8562 if (MEM_VOLATILE_P (mem1
) || MEM_VOLATILE_P (mem2
))
8565 /* MEM1 should be aligned on a 64-bit boundary. */
8566 if (MEM_ALIGN (mem1
) < 64)
8569 addr1
= XEXP (mem1
, 0);
8570 addr2
= XEXP (mem2
, 0);
8572 /* Extract a register number and offset (if used) from the first addr. */
8573 if (GET_CODE (addr1
) == PLUS
)
8575 /* If not a REG, return zero. */
8576 if (GET_CODE (XEXP (addr1
, 0)) != REG
)
8580 reg1
= REGNO (XEXP (addr1
, 0));
8581 /* The offset must be constant! */
8582 if (GET_CODE (XEXP (addr1
, 1)) != CONST_INT
)
8584 offset1
= INTVAL (XEXP (addr1
, 1));
8587 else if (GET_CODE (addr1
) != REG
)
8591 reg1
= REGNO (addr1
);
8592 /* This was a simple (mem (reg)) expression. Offset is 0. */
8596 /* Make sure the second address is a (mem (plus (reg) (const_int). */
8597 if (GET_CODE (addr2
) != PLUS
)
8600 if (GET_CODE (XEXP (addr2
, 0)) != REG
8601 || GET_CODE (XEXP (addr2
, 1)) != CONST_INT
)
8604 if (reg1
!= REGNO (XEXP (addr2
, 0)))
8607 if (dependent_reg_rtx
!= NULL_RTX
&& reg1
== REGNO (dependent_reg_rtx
))
8610 /* The first offset must be evenly divisible by 8 to ensure the
8611 address is 64 bit aligned. */
8612 if (offset1
% 8 != 0)
8615 /* The offset for the second addr must be 4 more than the first addr. */
8616 if (INTVAL (XEXP (addr2
, 1)) != offset1
+ 4)
8619 /* All the tests passed. addr1 and addr2 are valid for ldd and std
8624 /* Return the widened memory access made of MEM1 and MEM2 in MODE. */
8627 widen_mem_for_ldd_peep (rtx mem1
, rtx mem2
, machine_mode mode
)
8629 rtx x
= widen_memory_access (mem1
, mode
, 0);
8630 MEM_NOTRAP_P (x
) = MEM_NOTRAP_P (mem1
) && MEM_NOTRAP_P (mem2
);
8634 /* Return 1 if reg is a pseudo, or is the first register in
8635 a hard register pair. This makes it suitable for use in
8636 ldd and std insns. */
8639 register_ok_for_ldd (rtx reg
)
8641 /* We might have been passed a SUBREG. */
8645 if (REGNO (reg
) < FIRST_PSEUDO_REGISTER
)
8646 return (REGNO (reg
) % 2 == 0);
8651 /* Return 1 if OP, a MEM, has an address which is known to be
8652 aligned to an 8-byte boundary. */
8655 memory_ok_for_ldd (rtx op
)
8657 /* In 64-bit mode, we assume that the address is word-aligned. */
8658 if (TARGET_ARCH32
&& !mem_min_alignment (op
, 8))
8661 if (! can_create_pseudo_p ()
8662 && !strict_memory_address_p (Pmode
, XEXP (op
, 0)))
8668 /* Implement TARGET_PRINT_OPERAND_PUNCT_VALID_P. */
8671 sparc_print_operand_punct_valid_p (unsigned char code
)
8684 /* Implement TARGET_PRINT_OPERAND.
8685 Print operand X (an rtx) in assembler syntax to file FILE.
8686 CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified.
8687 For `%' followed by punctuation, CODE is the punctuation and X is null. */
8690 sparc_print_operand (FILE *file
, rtx x
, int code
)
8695 /* Output an insn in a delay slot. */
8697 sparc_indent_opcode
= 1;
8699 fputs ("\n\t nop", file
);
8702 /* Output an annul flag if there's nothing for the delay slot and we
8703 are optimizing. This is always used with '(' below.
8704 Sun OS 4.1.1 dbx can't handle an annulled unconditional branch;
8705 this is a dbx bug. So, we only do this when optimizing.
8706 On UltraSPARC, a branch in a delay slot causes a pipeline flush.
8707 Always emit a nop in case the next instruction is a branch. */
8708 if (! final_sequence
&& (optimize
&& (int)sparc_cpu
< PROCESSOR_V9
))
8712 /* Output a 'nop' if there's nothing for the delay slot and we are
8713 not optimizing. This is always used with '*' above. */
8714 if (! final_sequence
&& ! (optimize
&& (int)sparc_cpu
< PROCESSOR_V9
))
8715 fputs ("\n\t nop", file
);
8716 else if (final_sequence
)
8717 sparc_indent_opcode
= 1;
8720 /* Output the right displacement from the saved PC on function return.
8721 The caller may have placed an "unimp" insn immediately after the call
8722 so we have to account for it. This insn is used in the 32-bit ABI
8723 when calling a function that returns a non zero-sized structure. The
8724 64-bit ABI doesn't have it. Be careful to have this test be the same
8725 as that for the call. The exception is when sparc_std_struct_return
8726 is enabled, the psABI is followed exactly and the adjustment is made
8727 by the code in sparc_struct_value_rtx. The call emitted is the same
8728 when sparc_std_struct_return is enabled. */
8730 && cfun
->returns_struct
8731 && !sparc_std_struct_return
8732 && DECL_SIZE (DECL_RESULT (current_function_decl
))
8733 && TREE_CODE (DECL_SIZE (DECL_RESULT (current_function_decl
)))
8735 && !integer_zerop (DECL_SIZE (DECL_RESULT (current_function_decl
))))
8741 /* Output the Embedded Medium/Anywhere code model base register. */
8742 fputs (EMBMEDANY_BASE_REG
, file
);
8745 /* Print some local dynamic TLS name. */
8746 if (const char *name
= get_some_local_dynamic_name ())
8747 assemble_name (file
, name
);
8749 output_operand_lossage ("'%%&' used without any "
8750 "local dynamic TLS references");
8754 /* Adjust the operand to take into account a RESTORE operation. */
8755 if (GET_CODE (x
) == CONST_INT
)
8757 else if (GET_CODE (x
) != REG
)
8758 output_operand_lossage ("invalid %%Y operand");
8759 else if (REGNO (x
) < 8)
8760 fputs (reg_names
[REGNO (x
)], file
);
8761 else if (REGNO (x
) >= 24 && REGNO (x
) < 32)
8762 fputs (reg_names
[REGNO (x
)-16], file
);
8764 output_operand_lossage ("invalid %%Y operand");
8767 /* Print out the low order register name of a register pair. */
8768 if (WORDS_BIG_ENDIAN
)
8769 fputs (reg_names
[REGNO (x
)+1], file
);
8771 fputs (reg_names
[REGNO (x
)], file
);
8774 /* Print out the high order register name of a register pair. */
8775 if (WORDS_BIG_ENDIAN
)
8776 fputs (reg_names
[REGNO (x
)], file
);
8778 fputs (reg_names
[REGNO (x
)+1], file
);
8781 /* Print out the second register name of a register pair or quad.
8782 I.e., R (%o0) => %o1. */
8783 fputs (reg_names
[REGNO (x
)+1], file
);
8786 /* Print out the third register name of a register quad.
8787 I.e., S (%o0) => %o2. */
8788 fputs (reg_names
[REGNO (x
)+2], file
);
8791 /* Print out the fourth register name of a register quad.
8792 I.e., T (%o0) => %o3. */
8793 fputs (reg_names
[REGNO (x
)+3], file
);
8796 /* Print a condition code register. */
8797 if (REGNO (x
) == SPARC_ICC_REG
)
8799 /* We don't handle CC[X]_NOOVmode because they're not supposed
8801 if (GET_MODE (x
) == CCmode
)
8802 fputs ("%icc", file
);
8803 else if (GET_MODE (x
) == CCXmode
)
8804 fputs ("%xcc", file
);
8809 /* %fccN register */
8810 fputs (reg_names
[REGNO (x
)], file
);
8813 /* Print the operand's address only. */
8814 output_address (GET_MODE (x
), XEXP (x
, 0));
8817 /* In this case we need a register. Use %g0 if the
8818 operand is const0_rtx. */
8820 || (GET_MODE (x
) != VOIDmode
&& x
== CONST0_RTX (GET_MODE (x
))))
8822 fputs ("%g0", file
);
8829 switch (GET_CODE (x
))
8831 case IOR
: fputs ("or", file
); break;
8832 case AND
: fputs ("and", file
); break;
8833 case XOR
: fputs ("xor", file
); break;
8834 default: output_operand_lossage ("invalid %%A operand");
8839 switch (GET_CODE (x
))
8841 case IOR
: fputs ("orn", file
); break;
8842 case AND
: fputs ("andn", file
); break;
8843 case XOR
: fputs ("xnor", file
); break;
8844 default: output_operand_lossage ("invalid %%B operand");
8848 /* This is used by the conditional move instructions. */
8851 enum rtx_code rc
= GET_CODE (x
);
8855 case NE
: fputs ("ne", file
); break;
8856 case EQ
: fputs ("e", file
); break;
8857 case GE
: fputs ("ge", file
); break;
8858 case GT
: fputs ("g", file
); break;
8859 case LE
: fputs ("le", file
); break;
8860 case LT
: fputs ("l", file
); break;
8861 case GEU
: fputs ("geu", file
); break;
8862 case GTU
: fputs ("gu", file
); break;
8863 case LEU
: fputs ("leu", file
); break;
8864 case LTU
: fputs ("lu", file
); break;
8865 case LTGT
: fputs ("lg", file
); break;
8866 case UNORDERED
: fputs ("u", file
); break;
8867 case ORDERED
: fputs ("o", file
); break;
8868 case UNLT
: fputs ("ul", file
); break;
8869 case UNLE
: fputs ("ule", file
); break;
8870 case UNGT
: fputs ("ug", file
); break;
8871 case UNGE
: fputs ("uge", file
); break;
8872 case UNEQ
: fputs ("ue", file
); break;
8873 default: output_operand_lossage ("invalid %%C operand");
8878 /* This are used by the movr instruction pattern. */
8881 enum rtx_code rc
= GET_CODE (x
);
8884 case NE
: fputs ("ne", file
); break;
8885 case EQ
: fputs ("e", file
); break;
8886 case GE
: fputs ("gez", file
); break;
8887 case LT
: fputs ("lz", file
); break;
8888 case LE
: fputs ("lez", file
); break;
8889 case GT
: fputs ("gz", file
); break;
8890 default: output_operand_lossage ("invalid %%D operand");
8897 /* Print a sign-extended character. */
8898 int i
= trunc_int_for_mode (INTVAL (x
), QImode
);
8899 fprintf (file
, "%d", i
);
8904 /* Operand must be a MEM; write its address. */
8905 if (GET_CODE (x
) != MEM
)
8906 output_operand_lossage ("invalid %%f operand");
8907 output_address (GET_MODE (x
), XEXP (x
, 0));
8912 /* Print a sign-extended 32-bit value. */
8914 if (GET_CODE(x
) == CONST_INT
)
8918 output_operand_lossage ("invalid %%s operand");
8921 i
= trunc_int_for_mode (i
, SImode
);
8922 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, i
);
8927 /* Do nothing special. */
8931 /* Undocumented flag. */
8932 output_operand_lossage ("invalid operand output code");
8935 if (GET_CODE (x
) == REG
)
8936 fputs (reg_names
[REGNO (x
)], file
);
8937 else if (GET_CODE (x
) == MEM
)
8940 /* Poor Sun assembler doesn't understand absolute addressing. */
8941 if (CONSTANT_P (XEXP (x
, 0)))
8942 fputs ("%g0+", file
);
8943 output_address (GET_MODE (x
), XEXP (x
, 0));
8946 else if (GET_CODE (x
) == HIGH
)
8948 fputs ("%hi(", file
);
8949 output_addr_const (file
, XEXP (x
, 0));
8952 else if (GET_CODE (x
) == LO_SUM
)
8954 sparc_print_operand (file
, XEXP (x
, 0), 0);
8955 if (TARGET_CM_MEDMID
)
8956 fputs ("+%l44(", file
);
8958 fputs ("+%lo(", file
);
8959 output_addr_const (file
, XEXP (x
, 1));
8962 else if (GET_CODE (x
) == CONST_DOUBLE
)
8963 output_operand_lossage ("floating-point constant not a valid immediate operand");
8965 output_addr_const (file
, x
);
8968 /* Implement TARGET_PRINT_OPERAND_ADDRESS. */
8971 sparc_print_operand_address (FILE *file
, machine_mode
/*mode*/, rtx x
)
8973 register rtx base
, index
= 0;
8975 register rtx addr
= x
;
8978 fputs (reg_names
[REGNO (addr
)], file
);
8979 else if (GET_CODE (addr
) == PLUS
)
8981 if (CONST_INT_P (XEXP (addr
, 0)))
8982 offset
= INTVAL (XEXP (addr
, 0)), base
= XEXP (addr
, 1);
8983 else if (CONST_INT_P (XEXP (addr
, 1)))
8984 offset
= INTVAL (XEXP (addr
, 1)), base
= XEXP (addr
, 0);
8986 base
= XEXP (addr
, 0), index
= XEXP (addr
, 1);
8987 if (GET_CODE (base
) == LO_SUM
)
8989 gcc_assert (USE_AS_OFFSETABLE_LO10
8991 && ! TARGET_CM_MEDMID
);
8992 output_operand (XEXP (base
, 0), 0);
8993 fputs ("+%lo(", file
);
8994 output_address (VOIDmode
, XEXP (base
, 1));
8995 fprintf (file
, ")+%d", offset
);
8999 fputs (reg_names
[REGNO (base
)], file
);
9001 fprintf (file
, "%+d", offset
);
9002 else if (REG_P (index
))
9003 fprintf (file
, "+%s", reg_names
[REGNO (index
)]);
9004 else if (GET_CODE (index
) == SYMBOL_REF
9005 || GET_CODE (index
) == LABEL_REF
9006 || GET_CODE (index
) == CONST
)
9007 fputc ('+', file
), output_addr_const (file
, index
);
9008 else gcc_unreachable ();
9011 else if (GET_CODE (addr
) == MINUS
9012 && GET_CODE (XEXP (addr
, 1)) == LABEL_REF
)
9014 output_addr_const (file
, XEXP (addr
, 0));
9016 output_addr_const (file
, XEXP (addr
, 1));
9017 fputs ("-.)", file
);
9019 else if (GET_CODE (addr
) == LO_SUM
)
9021 output_operand (XEXP (addr
, 0), 0);
9022 if (TARGET_CM_MEDMID
)
9023 fputs ("+%l44(", file
);
9025 fputs ("+%lo(", file
);
9026 output_address (VOIDmode
, XEXP (addr
, 1));
9030 && GET_CODE (addr
) == CONST
9031 && GET_CODE (XEXP (addr
, 0)) == MINUS
9032 && GET_CODE (XEXP (XEXP (addr
, 0), 1)) == CONST
9033 && GET_CODE (XEXP (XEXP (XEXP (addr
, 0), 1), 0)) == MINUS
9034 && XEXP (XEXP (XEXP (XEXP (addr
, 0), 1), 0), 1) == pc_rtx
)
9036 addr
= XEXP (addr
, 0);
9037 output_addr_const (file
, XEXP (addr
, 0));
9038 /* Group the args of the second CONST in parenthesis. */
9040 /* Skip past the second CONST--it does nothing for us. */
9041 output_addr_const (file
, XEXP (XEXP (addr
, 1), 0));
9042 /* Close the parenthesis. */
9047 output_addr_const (file
, addr
);
9051 /* Target hook for assembling integer objects. The sparc version has
9052 special handling for aligned DI-mode objects. */
9055 sparc_assemble_integer (rtx x
, unsigned int size
, int aligned_p
)
9057 /* ??? We only output .xword's for symbols and only then in environments
9058 where the assembler can handle them. */
9059 if (aligned_p
&& size
== 8 && GET_CODE (x
) != CONST_INT
)
9063 assemble_integer_with_op ("\t.xword\t", x
);
9068 assemble_aligned_integer (4, const0_rtx
);
9069 assemble_aligned_integer (4, x
);
9073 return default_assemble_integer (x
, size
, aligned_p
);
9076 /* Return the value of a code used in the .proc pseudo-op that says
9077 what kind of result this function returns. For non-C types, we pick
9078 the closest C type. */
9080 #ifndef SHORT_TYPE_SIZE
9081 #define SHORT_TYPE_SIZE (BITS_PER_UNIT * 2)
9084 #ifndef INT_TYPE_SIZE
9085 #define INT_TYPE_SIZE BITS_PER_WORD
9088 #ifndef LONG_TYPE_SIZE
9089 #define LONG_TYPE_SIZE BITS_PER_WORD
9092 #ifndef LONG_LONG_TYPE_SIZE
9093 #define LONG_LONG_TYPE_SIZE (BITS_PER_WORD * 2)
9096 #ifndef FLOAT_TYPE_SIZE
9097 #define FLOAT_TYPE_SIZE BITS_PER_WORD
9100 #ifndef DOUBLE_TYPE_SIZE
9101 #define DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
9104 #ifndef LONG_DOUBLE_TYPE_SIZE
9105 #define LONG_DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
9109 sparc_type_code (register tree type
)
9111 register unsigned long qualifiers
= 0;
9112 register unsigned shift
;
9114 /* Only the first 30 bits of the qualifier are valid. We must refrain from
9115 setting more, since some assemblers will give an error for this. Also,
9116 we must be careful to avoid shifts of 32 bits or more to avoid getting
9117 unpredictable results. */
9119 for (shift
= 6; shift
< 30; shift
+= 2, type
= TREE_TYPE (type
))
9121 switch (TREE_CODE (type
))
9127 qualifiers
|= (3 << shift
);
9132 qualifiers
|= (2 << shift
);
9136 case REFERENCE_TYPE
:
9138 qualifiers
|= (1 << shift
);
9142 return (qualifiers
| 8);
9145 case QUAL_UNION_TYPE
:
9146 return (qualifiers
| 9);
9149 return (qualifiers
| 10);
9152 return (qualifiers
| 16);
9155 /* If this is a range type, consider it to be the underlying
9157 if (TREE_TYPE (type
) != 0)
9160 /* Carefully distinguish all the standard types of C,
9161 without messing up if the language is not C. We do this by
9162 testing TYPE_PRECISION and TYPE_UNSIGNED. The old code used to
9163 look at both the names and the above fields, but that's redundant.
9164 Any type whose size is between two C types will be considered
9165 to be the wider of the two types. Also, we do not have a
9166 special code to use for "long long", so anything wider than
9167 long is treated the same. Note that we can't distinguish
9168 between "int" and "long" in this code if they are the same
9169 size, but that's fine, since neither can the assembler. */
9171 if (TYPE_PRECISION (type
) <= CHAR_TYPE_SIZE
)
9172 return (qualifiers
| (TYPE_UNSIGNED (type
) ? 12 : 2));
9174 else if (TYPE_PRECISION (type
) <= SHORT_TYPE_SIZE
)
9175 return (qualifiers
| (TYPE_UNSIGNED (type
) ? 13 : 3));
9177 else if (TYPE_PRECISION (type
) <= INT_TYPE_SIZE
)
9178 return (qualifiers
| (TYPE_UNSIGNED (type
) ? 14 : 4));
9181 return (qualifiers
| (TYPE_UNSIGNED (type
) ? 15 : 5));
9184 /* If this is a range type, consider it to be the underlying
9186 if (TREE_TYPE (type
) != 0)
9189 /* Carefully distinguish all the standard types of C,
9190 without messing up if the language is not C. */
9192 if (TYPE_PRECISION (type
) == FLOAT_TYPE_SIZE
)
9193 return (qualifiers
| 6);
9196 return (qualifiers
| 7);
9198 case COMPLEX_TYPE
: /* GNU Fortran COMPLEX type. */
9199 /* ??? We need to distinguish between double and float complex types,
9200 but I don't know how yet because I can't reach this code from
9201 existing front-ends. */
9202 return (qualifiers
| 7); /* Who knows? */
9205 case BOOLEAN_TYPE
: /* Boolean truth value type. */
9211 gcc_unreachable (); /* Not a type! */
9218 /* Nested function support. */
9220 /* Emit RTL insns to initialize the variable parts of a trampoline.
9221 FNADDR is an RTX for the address of the function's pure code.
9222 CXT is an RTX for the static chain value for the function.
9224 This takes 16 insns: 2 shifts & 2 ands (to split up addresses), 4 sethi
9225 (to load in opcodes), 4 iors (to merge address and opcodes), and 4 writes
9226 (to store insns). This is a bit excessive. Perhaps a different
9227 mechanism would be better here.
9229 Emit enough FLUSH insns to synchronize the data and instruction caches. */
9232 sparc32_initialize_trampoline (rtx m_tramp
, rtx fnaddr
, rtx cxt
)
9234 /* SPARC 32-bit trampoline:
9237 sethi %hi(static), %g2
9239 or %g2, %lo(static), %g2
9241 SETHI i,r = 00rr rrr1 00ii iiii iiii iiii iiii iiii
9242 JMPL r+i,d = 10dd ddd1 1100 0rrr rr1i iiii iiii iiii
9246 (adjust_address (m_tramp
, SImode
, 0),
9247 expand_binop (SImode
, ior_optab
,
9248 expand_shift (RSHIFT_EXPR
, SImode
, fnaddr
, 10, 0, 1),
9249 GEN_INT (trunc_int_for_mode (0x03000000, SImode
)),
9250 NULL_RTX
, 1, OPTAB_DIRECT
));
9253 (adjust_address (m_tramp
, SImode
, 4),
9254 expand_binop (SImode
, ior_optab
,
9255 expand_shift (RSHIFT_EXPR
, SImode
, cxt
, 10, 0, 1),
9256 GEN_INT (trunc_int_for_mode (0x05000000, SImode
)),
9257 NULL_RTX
, 1, OPTAB_DIRECT
));
9260 (adjust_address (m_tramp
, SImode
, 8),
9261 expand_binop (SImode
, ior_optab
,
9262 expand_and (SImode
, fnaddr
, GEN_INT (0x3ff), NULL_RTX
),
9263 GEN_INT (trunc_int_for_mode (0x81c06000, SImode
)),
9264 NULL_RTX
, 1, OPTAB_DIRECT
));
9267 (adjust_address (m_tramp
, SImode
, 12),
9268 expand_binop (SImode
, ior_optab
,
9269 expand_and (SImode
, cxt
, GEN_INT (0x3ff), NULL_RTX
),
9270 GEN_INT (trunc_int_for_mode (0x8410a000, SImode
)),
9271 NULL_RTX
, 1, OPTAB_DIRECT
));
9273 /* On UltraSPARC a flush flushes an entire cache line. The trampoline is
9274 aligned on a 16 byte boundary so one flush clears it all. */
9275 emit_insn (gen_flushsi (validize_mem (adjust_address (m_tramp
, SImode
, 0))));
9276 if (sparc_cpu
!= PROCESSOR_ULTRASPARC
9277 && sparc_cpu
!= PROCESSOR_ULTRASPARC3
9278 && sparc_cpu
!= PROCESSOR_NIAGARA
9279 && sparc_cpu
!= PROCESSOR_NIAGARA2
9280 && sparc_cpu
!= PROCESSOR_NIAGARA3
9281 && sparc_cpu
!= PROCESSOR_NIAGARA4
9282 && sparc_cpu
!= PROCESSOR_NIAGARA7
)
9283 emit_insn (gen_flushsi (validize_mem (adjust_address (m_tramp
, SImode
, 8))));
9285 /* Call __enable_execute_stack after writing onto the stack to make sure
9286 the stack address is accessible. */
9287 #ifdef HAVE_ENABLE_EXECUTE_STACK
9288 emit_library_call (gen_rtx_SYMBOL_REF (Pmode
, "__enable_execute_stack"),
9289 LCT_NORMAL
, VOIDmode
, 1, XEXP (m_tramp
, 0), Pmode
);
9294 /* The 64-bit version is simpler because it makes more sense to load the
9295 values as "immediate" data out of the trampoline. It's also easier since
9296 we can read the PC without clobbering a register. */
9299 sparc64_initialize_trampoline (rtx m_tramp
, rtx fnaddr
, rtx cxt
)
9301 /* SPARC 64-bit trampoline:
9310 emit_move_insn (adjust_address (m_tramp
, SImode
, 0),
9311 GEN_INT (trunc_int_for_mode (0x83414000, SImode
)));
9312 emit_move_insn (adjust_address (m_tramp
, SImode
, 4),
9313 GEN_INT (trunc_int_for_mode (0xca586018, SImode
)));
9314 emit_move_insn (adjust_address (m_tramp
, SImode
, 8),
9315 GEN_INT (trunc_int_for_mode (0x81c14000, SImode
)));
9316 emit_move_insn (adjust_address (m_tramp
, SImode
, 12),
9317 GEN_INT (trunc_int_for_mode (0xca586010, SImode
)));
9318 emit_move_insn (adjust_address (m_tramp
, DImode
, 16), cxt
);
9319 emit_move_insn (adjust_address (m_tramp
, DImode
, 24), fnaddr
);
9320 emit_insn (gen_flushdi (validize_mem (adjust_address (m_tramp
, DImode
, 0))));
9322 if (sparc_cpu
!= PROCESSOR_ULTRASPARC
9323 && sparc_cpu
!= PROCESSOR_ULTRASPARC3
9324 && sparc_cpu
!= PROCESSOR_NIAGARA
9325 && sparc_cpu
!= PROCESSOR_NIAGARA2
9326 && sparc_cpu
!= PROCESSOR_NIAGARA3
9327 && sparc_cpu
!= PROCESSOR_NIAGARA4
9328 && sparc_cpu
!= PROCESSOR_NIAGARA7
)
9329 emit_insn (gen_flushdi (validize_mem (adjust_address (m_tramp
, DImode
, 8))));
9331 /* Call __enable_execute_stack after writing onto the stack to make sure
9332 the stack address is accessible. */
9333 #ifdef HAVE_ENABLE_EXECUTE_STACK
9334 emit_library_call (gen_rtx_SYMBOL_REF (Pmode
, "__enable_execute_stack"),
9335 LCT_NORMAL
, VOIDmode
, 1, XEXP (m_tramp
, 0), Pmode
);
9339 /* Worker for TARGET_TRAMPOLINE_INIT. */
9342 sparc_trampoline_init (rtx m_tramp
, tree fndecl
, rtx cxt
)
9344 rtx fnaddr
= force_reg (Pmode
, XEXP (DECL_RTL (fndecl
), 0));
9345 cxt
= force_reg (Pmode
, cxt
);
9347 sparc64_initialize_trampoline (m_tramp
, fnaddr
, cxt
);
9349 sparc32_initialize_trampoline (m_tramp
, fnaddr
, cxt
);
9352 /* Adjust the cost of a scheduling dependency. Return the new cost of
9353 a dependency LINK or INSN on DEP_INSN. COST is the current cost. */
9356 supersparc_adjust_cost (rtx_insn
*insn
, rtx link
, rtx_insn
*dep_insn
, int cost
)
9358 enum attr_type insn_type
;
9360 if (recog_memoized (insn
) < 0)
9363 insn_type
= get_attr_type (insn
);
9365 if (REG_NOTE_KIND (link
) == 0)
9367 /* Data dependency; DEP_INSN writes a register that INSN reads some
9370 /* if a load, then the dependence must be on the memory address;
9371 add an extra "cycle". Note that the cost could be two cycles
9372 if the reg was written late in an instruction group; we ca not tell
9374 if (insn_type
== TYPE_LOAD
|| insn_type
== TYPE_FPLOAD
)
9377 /* Get the delay only if the address of the store is the dependence. */
9378 if (insn_type
== TYPE_STORE
|| insn_type
== TYPE_FPSTORE
)
9380 rtx pat
= PATTERN(insn
);
9381 rtx dep_pat
= PATTERN (dep_insn
);
9383 if (GET_CODE (pat
) != SET
|| GET_CODE (dep_pat
) != SET
)
9384 return cost
; /* This should not happen! */
9386 /* The dependency between the two instructions was on the data that
9387 is being stored. Assume that this implies that the address of the
9388 store is not dependent. */
9389 if (rtx_equal_p (SET_DEST (dep_pat
), SET_SRC (pat
)))
9392 return cost
+ 3; /* An approximation. */
9395 /* A shift instruction cannot receive its data from an instruction
9396 in the same cycle; add a one cycle penalty. */
9397 if (insn_type
== TYPE_SHIFT
)
9398 return cost
+ 3; /* Split before cascade into shift. */
9402 /* Anti- or output- dependency; DEP_INSN reads/writes a register that
9403 INSN writes some cycles later. */
9405 /* These are only significant for the fpu unit; writing a fp reg before
9406 the fpu has finished with it stalls the processor. */
9408 /* Reusing an integer register causes no problems. */
9409 if (insn_type
== TYPE_IALU
|| insn_type
== TYPE_SHIFT
)
9417 hypersparc_adjust_cost (rtx_insn
*insn
, rtx link
, rtx_insn
*dep_insn
, int cost
)
9419 enum attr_type insn_type
, dep_type
;
9420 rtx pat
= PATTERN(insn
);
9421 rtx dep_pat
= PATTERN (dep_insn
);
9423 if (recog_memoized (insn
) < 0 || recog_memoized (dep_insn
) < 0)
9426 insn_type
= get_attr_type (insn
);
9427 dep_type
= get_attr_type (dep_insn
);
9429 switch (REG_NOTE_KIND (link
))
9432 /* Data dependency; DEP_INSN writes a register that INSN reads some
9439 /* Get the delay iff the address of the store is the dependence. */
9440 if (GET_CODE (pat
) != SET
|| GET_CODE (dep_pat
) != SET
)
9443 if (rtx_equal_p (SET_DEST (dep_pat
), SET_SRC (pat
)))
9450 /* If a load, then the dependence must be on the memory address. If
9451 the addresses aren't equal, then it might be a false dependency */
9452 if (dep_type
== TYPE_STORE
|| dep_type
== TYPE_FPSTORE
)
9454 if (GET_CODE (pat
) != SET
|| GET_CODE (dep_pat
) != SET
9455 || GET_CODE (SET_DEST (dep_pat
)) != MEM
9456 || GET_CODE (SET_SRC (pat
)) != MEM
9457 || ! rtx_equal_p (XEXP (SET_DEST (dep_pat
), 0),
9458 XEXP (SET_SRC (pat
), 0)))
9466 /* Compare to branch latency is 0. There is no benefit from
9467 separating compare and branch. */
9468 if (dep_type
== TYPE_COMPARE
)
9470 /* Floating point compare to branch latency is less than
9471 compare to conditional move. */
9472 if (dep_type
== TYPE_FPCMP
)
9481 /* Anti-dependencies only penalize the fpu unit. */
9482 if (insn_type
== TYPE_IALU
|| insn_type
== TYPE_SHIFT
)
9494 sparc_adjust_cost (rtx_insn
*insn
, rtx link
, rtx_insn
*dep
, int cost
)
9498 case PROCESSOR_SUPERSPARC
:
9499 cost
= supersparc_adjust_cost (insn
, link
, dep
, cost
);
9501 case PROCESSOR_HYPERSPARC
:
9502 case PROCESSOR_SPARCLITE86X
:
9503 cost
= hypersparc_adjust_cost (insn
, link
, dep
, cost
);
9512 sparc_sched_init (FILE *dump ATTRIBUTE_UNUSED
,
9513 int sched_verbose ATTRIBUTE_UNUSED
,
9514 int max_ready ATTRIBUTE_UNUSED
)
9518 sparc_use_sched_lookahead (void)
9520 if (sparc_cpu
== PROCESSOR_NIAGARA
9521 || sparc_cpu
== PROCESSOR_NIAGARA2
9522 || sparc_cpu
== PROCESSOR_NIAGARA3
)
9524 if (sparc_cpu
== PROCESSOR_NIAGARA4
9525 || sparc_cpu
== PROCESSOR_NIAGARA7
)
9527 if (sparc_cpu
== PROCESSOR_ULTRASPARC
9528 || sparc_cpu
== PROCESSOR_ULTRASPARC3
)
9530 if ((1 << sparc_cpu
) &
9531 ((1 << PROCESSOR_SUPERSPARC
) | (1 << PROCESSOR_HYPERSPARC
) |
9532 (1 << PROCESSOR_SPARCLITE86X
)))
9538 sparc_issue_rate (void)
9542 case PROCESSOR_NIAGARA
:
9543 case PROCESSOR_NIAGARA2
:
9544 case PROCESSOR_NIAGARA3
:
9547 case PROCESSOR_NIAGARA4
:
9548 case PROCESSOR_NIAGARA7
:
9550 /* Assume V9 processors are capable of at least dual-issue. */
9552 case PROCESSOR_SUPERSPARC
:
9554 case PROCESSOR_HYPERSPARC
:
9555 case PROCESSOR_SPARCLITE86X
:
9557 case PROCESSOR_ULTRASPARC
:
9558 case PROCESSOR_ULTRASPARC3
:
9564 set_extends (rtx_insn
*insn
)
9566 register rtx pat
= PATTERN (insn
);
9568 switch (GET_CODE (SET_SRC (pat
)))
9570 /* Load and some shift instructions zero extend. */
9573 /* sethi clears the high bits */
9575 /* LO_SUM is used with sethi. sethi cleared the high
9576 bits and the values used with lo_sum are positive */
9578 /* Store flag stores 0 or 1 */
9588 rtx op0
= XEXP (SET_SRC (pat
), 0);
9589 rtx op1
= XEXP (SET_SRC (pat
), 1);
9590 if (GET_CODE (op1
) == CONST_INT
)
9591 return INTVAL (op1
) >= 0;
9592 if (GET_CODE (op0
) != REG
)
9594 if (sparc_check_64 (op0
, insn
) == 1)
9596 return (GET_CODE (op1
) == REG
&& sparc_check_64 (op1
, insn
) == 1);
9601 rtx op0
= XEXP (SET_SRC (pat
), 0);
9602 rtx op1
= XEXP (SET_SRC (pat
), 1);
9603 if (GET_CODE (op0
) != REG
|| sparc_check_64 (op0
, insn
) <= 0)
9605 if (GET_CODE (op1
) == CONST_INT
)
9606 return INTVAL (op1
) >= 0;
9607 return (GET_CODE (op1
) == REG
&& sparc_check_64 (op1
, insn
) == 1);
9610 return GET_MODE (SET_SRC (pat
)) == SImode
;
9611 /* Positive integers leave the high bits zero. */
9613 return !(INTVAL (SET_SRC (pat
)) & 0x80000000);
9616 return - (GET_MODE (SET_SRC (pat
)) == SImode
);
9618 return sparc_check_64 (SET_SRC (pat
), insn
);
9624 /* We _ought_ to have only one kind per function, but... */
9625 static GTY(()) rtx sparc_addr_diff_list
;
9626 static GTY(()) rtx sparc_addr_list
;
9629 sparc_defer_case_vector (rtx lab
, rtx vec
, int diff
)
9631 vec
= gen_rtx_EXPR_LIST (VOIDmode
, lab
, vec
);
9633 sparc_addr_diff_list
9634 = gen_rtx_EXPR_LIST (VOIDmode
, vec
, sparc_addr_diff_list
);
9636 sparc_addr_list
= gen_rtx_EXPR_LIST (VOIDmode
, vec
, sparc_addr_list
);
9640 sparc_output_addr_vec (rtx vec
)
9642 rtx lab
= XEXP (vec
, 0), body
= XEXP (vec
, 1);
9643 int idx
, vlen
= XVECLEN (body
, 0);
9645 #ifdef ASM_OUTPUT_ADDR_VEC_START
9646 ASM_OUTPUT_ADDR_VEC_START (asm_out_file
);
9649 #ifdef ASM_OUTPUT_CASE_LABEL
9650 ASM_OUTPUT_CASE_LABEL (asm_out_file
, "L", CODE_LABEL_NUMBER (lab
),
9653 (*targetm
.asm_out
.internal_label
) (asm_out_file
, "L", CODE_LABEL_NUMBER (lab
));
9656 for (idx
= 0; idx
< vlen
; idx
++)
9658 ASM_OUTPUT_ADDR_VEC_ELT
9659 (asm_out_file
, CODE_LABEL_NUMBER (XEXP (XVECEXP (body
, 0, idx
), 0)));
9662 #ifdef ASM_OUTPUT_ADDR_VEC_END
9663 ASM_OUTPUT_ADDR_VEC_END (asm_out_file
);
9668 sparc_output_addr_diff_vec (rtx vec
)
9670 rtx lab
= XEXP (vec
, 0), body
= XEXP (vec
, 1);
9671 rtx base
= XEXP (XEXP (body
, 0), 0);
9672 int idx
, vlen
= XVECLEN (body
, 1);
9674 #ifdef ASM_OUTPUT_ADDR_VEC_START
9675 ASM_OUTPUT_ADDR_VEC_START (asm_out_file
);
9678 #ifdef ASM_OUTPUT_CASE_LABEL
9679 ASM_OUTPUT_CASE_LABEL (asm_out_file
, "L", CODE_LABEL_NUMBER (lab
),
9682 (*targetm
.asm_out
.internal_label
) (asm_out_file
, "L", CODE_LABEL_NUMBER (lab
));
9685 for (idx
= 0; idx
< vlen
; idx
++)
9687 ASM_OUTPUT_ADDR_DIFF_ELT
9690 CODE_LABEL_NUMBER (XEXP (XVECEXP (body
, 1, idx
), 0)),
9691 CODE_LABEL_NUMBER (base
));
9694 #ifdef ASM_OUTPUT_ADDR_VEC_END
9695 ASM_OUTPUT_ADDR_VEC_END (asm_out_file
);
9700 sparc_output_deferred_case_vectors (void)
9705 if (sparc_addr_list
== NULL_RTX
9706 && sparc_addr_diff_list
== NULL_RTX
)
9709 /* Align to cache line in the function's code section. */
9710 switch_to_section (current_function_section ());
9712 align
= floor_log2 (FUNCTION_BOUNDARY
/ BITS_PER_UNIT
);
9714 ASM_OUTPUT_ALIGN (asm_out_file
, align
);
9716 for (t
= sparc_addr_list
; t
; t
= XEXP (t
, 1))
9717 sparc_output_addr_vec (XEXP (t
, 0));
9718 for (t
= sparc_addr_diff_list
; t
; t
= XEXP (t
, 1))
9719 sparc_output_addr_diff_vec (XEXP (t
, 0));
9721 sparc_addr_list
= sparc_addr_diff_list
= NULL_RTX
;
9724 /* Return 0 if the high 32 bits of X (the low word of X, if DImode) are
9725 unknown. Return 1 if the high bits are zero, -1 if the register is
9728 sparc_check_64 (rtx x
, rtx_insn
*insn
)
9730 /* If a register is set only once it is safe to ignore insns this
9731 code does not know how to handle. The loop will either recognize
9732 the single set and return the correct value or fail to recognize
9737 gcc_assert (GET_CODE (x
) == REG
);
9739 if (GET_MODE (x
) == DImode
)
9740 y
= gen_rtx_REG (SImode
, REGNO (x
) + WORDS_BIG_ENDIAN
);
9742 if (flag_expensive_optimizations
9743 && df
&& DF_REG_DEF_COUNT (REGNO (y
)) == 1)
9749 insn
= get_last_insn_anywhere ();
9754 while ((insn
= PREV_INSN (insn
)))
9756 switch (GET_CODE (insn
))
9769 rtx pat
= PATTERN (insn
);
9770 if (GET_CODE (pat
) != SET
)
9772 if (rtx_equal_p (x
, SET_DEST (pat
)))
9773 return set_extends (insn
);
9774 if (y
&& rtx_equal_p (y
, SET_DEST (pat
)))
9775 return set_extends (insn
);
9776 if (reg_overlap_mentioned_p (SET_DEST (pat
), y
))
9784 /* Output a wide shift instruction in V8+ mode. INSN is the instruction,
9785 OPERANDS are its operands and OPCODE is the mnemonic to be used. */
9788 output_v8plus_shift (rtx_insn
*insn
, rtx
*operands
, const char *opcode
)
9790 static char asm_code
[60];
9792 /* The scratch register is only required when the destination
9793 register is not a 64-bit global or out register. */
9794 if (which_alternative
!= 2)
9795 operands
[3] = operands
[0];
9797 /* We can only shift by constants <= 63. */
9798 if (GET_CODE (operands
[2]) == CONST_INT
)
9799 operands
[2] = GEN_INT (INTVAL (operands
[2]) & 0x3f);
9801 if (GET_CODE (operands
[1]) == CONST_INT
)
9803 output_asm_insn ("mov\t%1, %3", operands
);
9807 output_asm_insn ("sllx\t%H1, 32, %3", operands
);
9808 if (sparc_check_64 (operands
[1], insn
) <= 0)
9809 output_asm_insn ("srl\t%L1, 0, %L1", operands
);
9810 output_asm_insn ("or\t%L1, %3, %3", operands
);
9813 strcpy (asm_code
, opcode
);
9815 if (which_alternative
!= 2)
9816 return strcat (asm_code
, "\t%0, %2, %L0\n\tsrlx\t%L0, 32, %H0");
9819 strcat (asm_code
, "\t%3, %2, %3\n\tsrlx\t%3, 32, %H0\n\tmov\t%3, %L0");
9822 /* Output rtl to increment the profiler label LABELNO
9823 for profiling a function entry. */
9826 sparc_profile_hook (int labelno
)
9831 fun
= gen_rtx_SYMBOL_REF (Pmode
, MCOUNT_FUNCTION
);
9832 if (NO_PROFILE_COUNTERS
)
9834 emit_library_call (fun
, LCT_NORMAL
, VOIDmode
, 0);
9838 ASM_GENERATE_INTERNAL_LABEL (buf
, "LP", labelno
);
9839 lab
= gen_rtx_SYMBOL_REF (Pmode
, ggc_strdup (buf
));
9840 emit_library_call (fun
, LCT_NORMAL
, VOIDmode
, 1, lab
, Pmode
);
9844 #ifdef TARGET_SOLARIS
9845 /* Solaris implementation of TARGET_ASM_NAMED_SECTION. */
9848 sparc_solaris_elf_asm_named_section (const char *name
, unsigned int flags
,
9849 tree decl ATTRIBUTE_UNUSED
)
9851 if (HAVE_COMDAT_GROUP
&& flags
& SECTION_LINKONCE
)
9853 solaris_elf_asm_comdat_section (name
, flags
, decl
);
9857 fprintf (asm_out_file
, "\t.section\t\"%s\"", name
);
9859 if (!(flags
& SECTION_DEBUG
))
9860 fputs (",#alloc", asm_out_file
);
9861 if (flags
& SECTION_WRITE
)
9862 fputs (",#write", asm_out_file
);
9863 if (flags
& SECTION_TLS
)
9864 fputs (",#tls", asm_out_file
);
9865 if (flags
& SECTION_CODE
)
9866 fputs (",#execinstr", asm_out_file
);
9868 if (flags
& SECTION_NOTYPE
)
9870 else if (flags
& SECTION_BSS
)
9871 fputs (",#nobits", asm_out_file
);
9873 fputs (",#progbits", asm_out_file
);
9875 fputc ('\n', asm_out_file
);
9877 #endif /* TARGET_SOLARIS */
9879 /* We do not allow indirect calls to be optimized into sibling calls.
9881 We cannot use sibling calls when delayed branches are disabled
9882 because they will likely require the call delay slot to be filled.
9884 Also, on SPARC 32-bit we cannot emit a sibling call when the
9885 current function returns a structure. This is because the "unimp
9886 after call" convention would cause the callee to return to the
9887 wrong place. The generic code already disallows cases where the
9888 function being called returns a structure.
9890 It may seem strange how this last case could occur. Usually there
9891 is code after the call which jumps to epilogue code which dumps the
9892 return value into the struct return area. That ought to invalidate
9893 the sibling call right? Well, in the C++ case we can end up passing
9894 the pointer to the struct return area to a constructor (which returns
9895 void) and then nothing else happens. Such a sibling call would look
9896 valid without the added check here.
9898 VxWorks PIC PLT entries require the global pointer to be initialized
9899 on entry. We therefore can't emit sibling calls to them. */
9901 sparc_function_ok_for_sibcall (tree decl
, tree exp ATTRIBUTE_UNUSED
)
9904 && flag_delayed_branch
9905 && (TARGET_ARCH64
|| ! cfun
->returns_struct
)
9906 && !(TARGET_VXWORKS_RTP
9908 && !targetm
.binds_local_p (decl
)));
9911 /* libfunc renaming. */
9914 sparc_init_libfuncs (void)
9918 /* Use the subroutines that Sun's library provides for integer
9919 multiply and divide. The `*' prevents an underscore from
9920 being prepended by the compiler. .umul is a little faster
9922 set_optab_libfunc (smul_optab
, SImode
, "*.umul");
9923 set_optab_libfunc (sdiv_optab
, SImode
, "*.div");
9924 set_optab_libfunc (udiv_optab
, SImode
, "*.udiv");
9925 set_optab_libfunc (smod_optab
, SImode
, "*.rem");
9926 set_optab_libfunc (umod_optab
, SImode
, "*.urem");
9928 /* TFmode arithmetic. These names are part of the SPARC 32bit ABI. */
9929 set_optab_libfunc (add_optab
, TFmode
, "_Q_add");
9930 set_optab_libfunc (sub_optab
, TFmode
, "_Q_sub");
9931 set_optab_libfunc (neg_optab
, TFmode
, "_Q_neg");
9932 set_optab_libfunc (smul_optab
, TFmode
, "_Q_mul");
9933 set_optab_libfunc (sdiv_optab
, TFmode
, "_Q_div");
9935 /* We can define the TFmode sqrt optab only if TARGET_FPU. This
9936 is because with soft-float, the SFmode and DFmode sqrt
9937 instructions will be absent, and the compiler will notice and
9938 try to use the TFmode sqrt instruction for calls to the
9939 builtin function sqrt, but this fails. */
9941 set_optab_libfunc (sqrt_optab
, TFmode
, "_Q_sqrt");
9943 set_optab_libfunc (eq_optab
, TFmode
, "_Q_feq");
9944 set_optab_libfunc (ne_optab
, TFmode
, "_Q_fne");
9945 set_optab_libfunc (gt_optab
, TFmode
, "_Q_fgt");
9946 set_optab_libfunc (ge_optab
, TFmode
, "_Q_fge");
9947 set_optab_libfunc (lt_optab
, TFmode
, "_Q_flt");
9948 set_optab_libfunc (le_optab
, TFmode
, "_Q_fle");
9950 set_conv_libfunc (sext_optab
, TFmode
, SFmode
, "_Q_stoq");
9951 set_conv_libfunc (sext_optab
, TFmode
, DFmode
, "_Q_dtoq");
9952 set_conv_libfunc (trunc_optab
, SFmode
, TFmode
, "_Q_qtos");
9953 set_conv_libfunc (trunc_optab
, DFmode
, TFmode
, "_Q_qtod");
9955 set_conv_libfunc (sfix_optab
, SImode
, TFmode
, "_Q_qtoi");
9956 set_conv_libfunc (ufix_optab
, SImode
, TFmode
, "_Q_qtou");
9957 set_conv_libfunc (sfloat_optab
, TFmode
, SImode
, "_Q_itoq");
9958 set_conv_libfunc (ufloat_optab
, TFmode
, SImode
, "_Q_utoq");
9960 if (DITF_CONVERSION_LIBFUNCS
)
9962 set_conv_libfunc (sfix_optab
, DImode
, TFmode
, "_Q_qtoll");
9963 set_conv_libfunc (ufix_optab
, DImode
, TFmode
, "_Q_qtoull");
9964 set_conv_libfunc (sfloat_optab
, TFmode
, DImode
, "_Q_lltoq");
9965 set_conv_libfunc (ufloat_optab
, TFmode
, DImode
, "_Q_ulltoq");
9968 if (SUN_CONVERSION_LIBFUNCS
)
9970 set_conv_libfunc (sfix_optab
, DImode
, SFmode
, "__ftoll");
9971 set_conv_libfunc (ufix_optab
, DImode
, SFmode
, "__ftoull");
9972 set_conv_libfunc (sfix_optab
, DImode
, DFmode
, "__dtoll");
9973 set_conv_libfunc (ufix_optab
, DImode
, DFmode
, "__dtoull");
9978 /* In the SPARC 64bit ABI, SImode multiply and divide functions
9979 do not exist in the library. Make sure the compiler does not
9980 emit calls to them by accident. (It should always use the
9981 hardware instructions.) */
9982 set_optab_libfunc (smul_optab
, SImode
, 0);
9983 set_optab_libfunc (sdiv_optab
, SImode
, 0);
9984 set_optab_libfunc (udiv_optab
, SImode
, 0);
9985 set_optab_libfunc (smod_optab
, SImode
, 0);
9986 set_optab_libfunc (umod_optab
, SImode
, 0);
9988 if (SUN_INTEGER_MULTIPLY_64
)
9990 set_optab_libfunc (smul_optab
, DImode
, "__mul64");
9991 set_optab_libfunc (sdiv_optab
, DImode
, "__div64");
9992 set_optab_libfunc (udiv_optab
, DImode
, "__udiv64");
9993 set_optab_libfunc (smod_optab
, DImode
, "__rem64");
9994 set_optab_libfunc (umod_optab
, DImode
, "__urem64");
9997 if (SUN_CONVERSION_LIBFUNCS
)
9999 set_conv_libfunc (sfix_optab
, DImode
, SFmode
, "__ftol");
10000 set_conv_libfunc (ufix_optab
, DImode
, SFmode
, "__ftoul");
10001 set_conv_libfunc (sfix_optab
, DImode
, DFmode
, "__dtol");
10002 set_conv_libfunc (ufix_optab
, DImode
, DFmode
, "__dtoul");
10007 /* SPARC builtins. */
10008 enum sparc_builtins
10010 /* FPU builtins. */
10011 SPARC_BUILTIN_LDFSR
,
10012 SPARC_BUILTIN_STFSR
,
10014 /* VIS 1.0 builtins. */
10015 SPARC_BUILTIN_FPACK16
,
10016 SPARC_BUILTIN_FPACK32
,
10017 SPARC_BUILTIN_FPACKFIX
,
10018 SPARC_BUILTIN_FEXPAND
,
10019 SPARC_BUILTIN_FPMERGE
,
10020 SPARC_BUILTIN_FMUL8X16
,
10021 SPARC_BUILTIN_FMUL8X16AU
,
10022 SPARC_BUILTIN_FMUL8X16AL
,
10023 SPARC_BUILTIN_FMUL8SUX16
,
10024 SPARC_BUILTIN_FMUL8ULX16
,
10025 SPARC_BUILTIN_FMULD8SUX16
,
10026 SPARC_BUILTIN_FMULD8ULX16
,
10027 SPARC_BUILTIN_FALIGNDATAV4HI
,
10028 SPARC_BUILTIN_FALIGNDATAV8QI
,
10029 SPARC_BUILTIN_FALIGNDATAV2SI
,
10030 SPARC_BUILTIN_FALIGNDATADI
,
10031 SPARC_BUILTIN_WRGSR
,
10032 SPARC_BUILTIN_RDGSR
,
10033 SPARC_BUILTIN_ALIGNADDR
,
10034 SPARC_BUILTIN_ALIGNADDRL
,
10035 SPARC_BUILTIN_PDIST
,
10036 SPARC_BUILTIN_EDGE8
,
10037 SPARC_BUILTIN_EDGE8L
,
10038 SPARC_BUILTIN_EDGE16
,
10039 SPARC_BUILTIN_EDGE16L
,
10040 SPARC_BUILTIN_EDGE32
,
10041 SPARC_BUILTIN_EDGE32L
,
10042 SPARC_BUILTIN_FCMPLE16
,
10043 SPARC_BUILTIN_FCMPLE32
,
10044 SPARC_BUILTIN_FCMPNE16
,
10045 SPARC_BUILTIN_FCMPNE32
,
10046 SPARC_BUILTIN_FCMPGT16
,
10047 SPARC_BUILTIN_FCMPGT32
,
10048 SPARC_BUILTIN_FCMPEQ16
,
10049 SPARC_BUILTIN_FCMPEQ32
,
10050 SPARC_BUILTIN_FPADD16
,
10051 SPARC_BUILTIN_FPADD16S
,
10052 SPARC_BUILTIN_FPADD32
,
10053 SPARC_BUILTIN_FPADD32S
,
10054 SPARC_BUILTIN_FPSUB16
,
10055 SPARC_BUILTIN_FPSUB16S
,
10056 SPARC_BUILTIN_FPSUB32
,
10057 SPARC_BUILTIN_FPSUB32S
,
10058 SPARC_BUILTIN_ARRAY8
,
10059 SPARC_BUILTIN_ARRAY16
,
10060 SPARC_BUILTIN_ARRAY32
,
10062 /* VIS 2.0 builtins. */
10063 SPARC_BUILTIN_EDGE8N
,
10064 SPARC_BUILTIN_EDGE8LN
,
10065 SPARC_BUILTIN_EDGE16N
,
10066 SPARC_BUILTIN_EDGE16LN
,
10067 SPARC_BUILTIN_EDGE32N
,
10068 SPARC_BUILTIN_EDGE32LN
,
10069 SPARC_BUILTIN_BMASK
,
10070 SPARC_BUILTIN_BSHUFFLEV4HI
,
10071 SPARC_BUILTIN_BSHUFFLEV8QI
,
10072 SPARC_BUILTIN_BSHUFFLEV2SI
,
10073 SPARC_BUILTIN_BSHUFFLEDI
,
10075 /* VIS 3.0 builtins. */
10076 SPARC_BUILTIN_CMASK8
,
10077 SPARC_BUILTIN_CMASK16
,
10078 SPARC_BUILTIN_CMASK32
,
10079 SPARC_BUILTIN_FCHKSM16
,
10080 SPARC_BUILTIN_FSLL16
,
10081 SPARC_BUILTIN_FSLAS16
,
10082 SPARC_BUILTIN_FSRL16
,
10083 SPARC_BUILTIN_FSRA16
,
10084 SPARC_BUILTIN_FSLL32
,
10085 SPARC_BUILTIN_FSLAS32
,
10086 SPARC_BUILTIN_FSRL32
,
10087 SPARC_BUILTIN_FSRA32
,
10088 SPARC_BUILTIN_PDISTN
,
10089 SPARC_BUILTIN_FMEAN16
,
10090 SPARC_BUILTIN_FPADD64
,
10091 SPARC_BUILTIN_FPSUB64
,
10092 SPARC_BUILTIN_FPADDS16
,
10093 SPARC_BUILTIN_FPADDS16S
,
10094 SPARC_BUILTIN_FPSUBS16
,
10095 SPARC_BUILTIN_FPSUBS16S
,
10096 SPARC_BUILTIN_FPADDS32
,
10097 SPARC_BUILTIN_FPADDS32S
,
10098 SPARC_BUILTIN_FPSUBS32
,
10099 SPARC_BUILTIN_FPSUBS32S
,
10100 SPARC_BUILTIN_FUCMPLE8
,
10101 SPARC_BUILTIN_FUCMPNE8
,
10102 SPARC_BUILTIN_FUCMPGT8
,
10103 SPARC_BUILTIN_FUCMPEQ8
,
10104 SPARC_BUILTIN_FHADDS
,
10105 SPARC_BUILTIN_FHADDD
,
10106 SPARC_BUILTIN_FHSUBS
,
10107 SPARC_BUILTIN_FHSUBD
,
10108 SPARC_BUILTIN_FNHADDS
,
10109 SPARC_BUILTIN_FNHADDD
,
10110 SPARC_BUILTIN_UMULXHI
,
10111 SPARC_BUILTIN_XMULX
,
10112 SPARC_BUILTIN_XMULXHI
,
10114 /* VIS 4.0 builtins. */
10115 SPARC_BUILTIN_FPADD8
,
10116 SPARC_BUILTIN_FPADDS8
,
10117 SPARC_BUILTIN_FPADDUS8
,
10118 SPARC_BUILTIN_FPADDUS16
,
10119 SPARC_BUILTIN_FPCMPLE8
,
10120 SPARC_BUILTIN_FPCMPGT8
,
10121 SPARC_BUILTIN_FPCMPULE16
,
10122 SPARC_BUILTIN_FPCMPUGT16
,
10123 SPARC_BUILTIN_FPCMPULE32
,
10124 SPARC_BUILTIN_FPCMPUGT32
,
10125 SPARC_BUILTIN_FPMAX8
,
10126 SPARC_BUILTIN_FPMAX16
,
10127 SPARC_BUILTIN_FPMAX32
,
10128 SPARC_BUILTIN_FPMAXU8
,
10129 SPARC_BUILTIN_FPMAXU16
,
10130 SPARC_BUILTIN_FPMAXU32
,
10131 SPARC_BUILTIN_FPMIN8
,
10132 SPARC_BUILTIN_FPMIN16
,
10133 SPARC_BUILTIN_FPMIN32
,
10134 SPARC_BUILTIN_FPMINU8
,
10135 SPARC_BUILTIN_FPMINU16
,
10136 SPARC_BUILTIN_FPMINU32
,
10137 SPARC_BUILTIN_FPSUB8
,
10138 SPARC_BUILTIN_FPSUBS8
,
10139 SPARC_BUILTIN_FPSUBUS8
,
10140 SPARC_BUILTIN_FPSUBUS16
,
10145 static GTY (()) tree sparc_builtins
[(int) SPARC_BUILTIN_MAX
];
10146 static enum insn_code sparc_builtins_icode
[(int) SPARC_BUILTIN_MAX
];
10148 /* Add a SPARC builtin function with NAME, ICODE, CODE and TYPE. Return the
10149 function decl or NULL_TREE if the builtin was not added. */
10152 def_builtin (const char *name
, enum insn_code icode
, enum sparc_builtins code
,
10156 = add_builtin_function (name
, type
, code
, BUILT_IN_MD
, NULL
, NULL_TREE
);
10160 sparc_builtins
[code
] = t
;
10161 sparc_builtins_icode
[code
] = icode
;
10167 /* Likewise, but also marks the function as "const". */
10170 def_builtin_const (const char *name
, enum insn_code icode
,
10171 enum sparc_builtins code
, tree type
)
10173 tree t
= def_builtin (name
, icode
, code
, type
);
10176 TREE_READONLY (t
) = 1;
10181 /* Implement the TARGET_INIT_BUILTINS target hook.
10182 Create builtin functions for special SPARC instructions. */
10185 sparc_init_builtins (void)
10188 sparc_fpu_init_builtins ();
10191 sparc_vis_init_builtins ();
10194 /* Create builtin functions for FPU instructions. */
10197 sparc_fpu_init_builtins (void)
10200 = build_function_type_list (void_type_node
,
10201 build_pointer_type (unsigned_type_node
), 0);
10202 def_builtin ("__builtin_load_fsr", CODE_FOR_ldfsr
,
10203 SPARC_BUILTIN_LDFSR
, ftype
);
10204 def_builtin ("__builtin_store_fsr", CODE_FOR_stfsr
,
10205 SPARC_BUILTIN_STFSR
, ftype
);
10208 /* Create builtin functions for VIS instructions. */
10211 sparc_vis_init_builtins (void)
10213 tree v4qi
= build_vector_type (unsigned_intQI_type_node
, 4);
10214 tree v8qi
= build_vector_type (unsigned_intQI_type_node
, 8);
10215 tree v4hi
= build_vector_type (intHI_type_node
, 4);
10216 tree v2hi
= build_vector_type (intHI_type_node
, 2);
10217 tree v2si
= build_vector_type (intSI_type_node
, 2);
10218 tree v1si
= build_vector_type (intSI_type_node
, 1);
10220 tree v4qi_ftype_v4hi
= build_function_type_list (v4qi
, v4hi
, 0);
10221 tree v8qi_ftype_v2si_v8qi
= build_function_type_list (v8qi
, v2si
, v8qi
, 0);
10222 tree v2hi_ftype_v2si
= build_function_type_list (v2hi
, v2si
, 0);
10223 tree v4hi_ftype_v4qi
= build_function_type_list (v4hi
, v4qi
, 0);
10224 tree v8qi_ftype_v4qi_v4qi
= build_function_type_list (v8qi
, v4qi
, v4qi
, 0);
10225 tree v4hi_ftype_v4qi_v4hi
= build_function_type_list (v4hi
, v4qi
, v4hi
, 0);
10226 tree v4hi_ftype_v4qi_v2hi
= build_function_type_list (v4hi
, v4qi
, v2hi
, 0);
10227 tree v2si_ftype_v4qi_v2hi
= build_function_type_list (v2si
, v4qi
, v2hi
, 0);
10228 tree v4hi_ftype_v8qi_v4hi
= build_function_type_list (v4hi
, v8qi
, v4hi
, 0);
10229 tree v4hi_ftype_v4hi_v4hi
= build_function_type_list (v4hi
, v4hi
, v4hi
, 0);
10230 tree v2si_ftype_v2si_v2si
= build_function_type_list (v2si
, v2si
, v2si
, 0);
10231 tree v8qi_ftype_v8qi_v8qi
= build_function_type_list (v8qi
, v8qi
, v8qi
, 0);
10232 tree v2hi_ftype_v2hi_v2hi
= build_function_type_list (v2hi
, v2hi
, v2hi
, 0);
10233 tree v1si_ftype_v1si_v1si
= build_function_type_list (v1si
, v1si
, v1si
, 0);
10234 tree di_ftype_v8qi_v8qi_di
= build_function_type_list (intDI_type_node
,
10236 intDI_type_node
, 0);
10237 tree di_ftype_v8qi_v8qi
= build_function_type_list (intDI_type_node
,
10239 tree si_ftype_v8qi_v8qi
= build_function_type_list (intSI_type_node
,
10241 tree di_ftype_di_di
= build_function_type_list (intDI_type_node
,
10243 intDI_type_node
, 0);
10244 tree si_ftype_si_si
= build_function_type_list (intSI_type_node
,
10246 intSI_type_node
, 0);
10247 tree ptr_ftype_ptr_si
= build_function_type_list (ptr_type_node
,
10249 intSI_type_node
, 0);
10250 tree ptr_ftype_ptr_di
= build_function_type_list (ptr_type_node
,
10252 intDI_type_node
, 0);
10253 tree si_ftype_ptr_ptr
= build_function_type_list (intSI_type_node
,
10256 tree di_ftype_ptr_ptr
= build_function_type_list (intDI_type_node
,
10259 tree si_ftype_v4hi_v4hi
= build_function_type_list (intSI_type_node
,
10261 tree si_ftype_v2si_v2si
= build_function_type_list (intSI_type_node
,
10263 tree di_ftype_v4hi_v4hi
= build_function_type_list (intDI_type_node
,
10265 tree di_ftype_v2si_v2si
= build_function_type_list (intDI_type_node
,
10267 tree void_ftype_di
= build_function_type_list (void_type_node
,
10268 intDI_type_node
, 0);
10269 tree di_ftype_void
= build_function_type_list (intDI_type_node
,
10270 void_type_node
, 0);
10271 tree void_ftype_si
= build_function_type_list (void_type_node
,
10272 intSI_type_node
, 0);
10273 tree sf_ftype_sf_sf
= build_function_type_list (float_type_node
,
10275 float_type_node
, 0);
10276 tree df_ftype_df_df
= build_function_type_list (double_type_node
,
10278 double_type_node
, 0);
10280 /* Packing and expanding vectors. */
10281 def_builtin ("__builtin_vis_fpack16", CODE_FOR_fpack16_vis
,
10282 SPARC_BUILTIN_FPACK16
, v4qi_ftype_v4hi
);
10283 def_builtin ("__builtin_vis_fpack32", CODE_FOR_fpack32_vis
,
10284 SPARC_BUILTIN_FPACK32
, v8qi_ftype_v2si_v8qi
);
10285 def_builtin ("__builtin_vis_fpackfix", CODE_FOR_fpackfix_vis
,
10286 SPARC_BUILTIN_FPACKFIX
, v2hi_ftype_v2si
);
10287 def_builtin_const ("__builtin_vis_fexpand", CODE_FOR_fexpand_vis
,
10288 SPARC_BUILTIN_FEXPAND
, v4hi_ftype_v4qi
);
10289 def_builtin_const ("__builtin_vis_fpmerge", CODE_FOR_fpmerge_vis
,
10290 SPARC_BUILTIN_FPMERGE
, v8qi_ftype_v4qi_v4qi
);
10292 /* Multiplications. */
10293 def_builtin_const ("__builtin_vis_fmul8x16", CODE_FOR_fmul8x16_vis
,
10294 SPARC_BUILTIN_FMUL8X16
, v4hi_ftype_v4qi_v4hi
);
10295 def_builtin_const ("__builtin_vis_fmul8x16au", CODE_FOR_fmul8x16au_vis
,
10296 SPARC_BUILTIN_FMUL8X16AU
, v4hi_ftype_v4qi_v2hi
);
10297 def_builtin_const ("__builtin_vis_fmul8x16al", CODE_FOR_fmul8x16al_vis
,
10298 SPARC_BUILTIN_FMUL8X16AL
, v4hi_ftype_v4qi_v2hi
);
10299 def_builtin_const ("__builtin_vis_fmul8sux16", CODE_FOR_fmul8sux16_vis
,
10300 SPARC_BUILTIN_FMUL8SUX16
, v4hi_ftype_v8qi_v4hi
);
10301 def_builtin_const ("__builtin_vis_fmul8ulx16", CODE_FOR_fmul8ulx16_vis
,
10302 SPARC_BUILTIN_FMUL8ULX16
, v4hi_ftype_v8qi_v4hi
);
10303 def_builtin_const ("__builtin_vis_fmuld8sux16", CODE_FOR_fmuld8sux16_vis
,
10304 SPARC_BUILTIN_FMULD8SUX16
, v2si_ftype_v4qi_v2hi
);
10305 def_builtin_const ("__builtin_vis_fmuld8ulx16", CODE_FOR_fmuld8ulx16_vis
,
10306 SPARC_BUILTIN_FMULD8ULX16
, v2si_ftype_v4qi_v2hi
);
10308 /* Data aligning. */
10309 def_builtin ("__builtin_vis_faligndatav4hi", CODE_FOR_faligndatav4hi_vis
,
10310 SPARC_BUILTIN_FALIGNDATAV4HI
, v4hi_ftype_v4hi_v4hi
);
10311 def_builtin ("__builtin_vis_faligndatav8qi", CODE_FOR_faligndatav8qi_vis
,
10312 SPARC_BUILTIN_FALIGNDATAV8QI
, v8qi_ftype_v8qi_v8qi
);
10313 def_builtin ("__builtin_vis_faligndatav2si", CODE_FOR_faligndatav2si_vis
,
10314 SPARC_BUILTIN_FALIGNDATAV2SI
, v2si_ftype_v2si_v2si
);
10315 def_builtin ("__builtin_vis_faligndatadi", CODE_FOR_faligndatav1di_vis
,
10316 SPARC_BUILTIN_FALIGNDATADI
, di_ftype_di_di
);
10318 def_builtin ("__builtin_vis_write_gsr", CODE_FOR_wrgsr_vis
,
10319 SPARC_BUILTIN_WRGSR
, void_ftype_di
);
10320 def_builtin ("__builtin_vis_read_gsr", CODE_FOR_rdgsr_vis
,
10321 SPARC_BUILTIN_RDGSR
, di_ftype_void
);
10325 def_builtin ("__builtin_vis_alignaddr", CODE_FOR_alignaddrdi_vis
,
10326 SPARC_BUILTIN_ALIGNADDR
, ptr_ftype_ptr_di
);
10327 def_builtin ("__builtin_vis_alignaddrl", CODE_FOR_alignaddrldi_vis
,
10328 SPARC_BUILTIN_ALIGNADDRL
, ptr_ftype_ptr_di
);
10332 def_builtin ("__builtin_vis_alignaddr", CODE_FOR_alignaddrsi_vis
,
10333 SPARC_BUILTIN_ALIGNADDR
, ptr_ftype_ptr_si
);
10334 def_builtin ("__builtin_vis_alignaddrl", CODE_FOR_alignaddrlsi_vis
,
10335 SPARC_BUILTIN_ALIGNADDRL
, ptr_ftype_ptr_si
);
10338 /* Pixel distance. */
10339 def_builtin_const ("__builtin_vis_pdist", CODE_FOR_pdist_vis
,
10340 SPARC_BUILTIN_PDIST
, di_ftype_v8qi_v8qi_di
);
10342 /* Edge handling. */
10345 def_builtin_const ("__builtin_vis_edge8", CODE_FOR_edge8di_vis
,
10346 SPARC_BUILTIN_EDGE8
, di_ftype_ptr_ptr
);
10347 def_builtin_const ("__builtin_vis_edge8l", CODE_FOR_edge8ldi_vis
,
10348 SPARC_BUILTIN_EDGE8L
, di_ftype_ptr_ptr
);
10349 def_builtin_const ("__builtin_vis_edge16", CODE_FOR_edge16di_vis
,
10350 SPARC_BUILTIN_EDGE16
, di_ftype_ptr_ptr
);
10351 def_builtin_const ("__builtin_vis_edge16l", CODE_FOR_edge16ldi_vis
,
10352 SPARC_BUILTIN_EDGE16L
, di_ftype_ptr_ptr
);
10353 def_builtin_const ("__builtin_vis_edge32", CODE_FOR_edge32di_vis
,
10354 SPARC_BUILTIN_EDGE32
, di_ftype_ptr_ptr
);
10355 def_builtin_const ("__builtin_vis_edge32l", CODE_FOR_edge32ldi_vis
,
10356 SPARC_BUILTIN_EDGE32L
, di_ftype_ptr_ptr
);
10360 def_builtin_const ("__builtin_vis_edge8", CODE_FOR_edge8si_vis
,
10361 SPARC_BUILTIN_EDGE8
, si_ftype_ptr_ptr
);
10362 def_builtin_const ("__builtin_vis_edge8l", CODE_FOR_edge8lsi_vis
,
10363 SPARC_BUILTIN_EDGE8L
, si_ftype_ptr_ptr
);
10364 def_builtin_const ("__builtin_vis_edge16", CODE_FOR_edge16si_vis
,
10365 SPARC_BUILTIN_EDGE16
, si_ftype_ptr_ptr
);
10366 def_builtin_const ("__builtin_vis_edge16l", CODE_FOR_edge16lsi_vis
,
10367 SPARC_BUILTIN_EDGE16L
, si_ftype_ptr_ptr
);
10368 def_builtin_const ("__builtin_vis_edge32", CODE_FOR_edge32si_vis
,
10369 SPARC_BUILTIN_EDGE32
, si_ftype_ptr_ptr
);
10370 def_builtin_const ("__builtin_vis_edge32l", CODE_FOR_edge32lsi_vis
,
10371 SPARC_BUILTIN_EDGE32L
, si_ftype_ptr_ptr
);
10374 /* Pixel compare. */
10377 def_builtin_const ("__builtin_vis_fcmple16", CODE_FOR_fcmple16di_vis
,
10378 SPARC_BUILTIN_FCMPLE16
, di_ftype_v4hi_v4hi
);
10379 def_builtin_const ("__builtin_vis_fcmple32", CODE_FOR_fcmple32di_vis
,
10380 SPARC_BUILTIN_FCMPLE32
, di_ftype_v2si_v2si
);
10381 def_builtin_const ("__builtin_vis_fcmpne16", CODE_FOR_fcmpne16di_vis
,
10382 SPARC_BUILTIN_FCMPNE16
, di_ftype_v4hi_v4hi
);
10383 def_builtin_const ("__builtin_vis_fcmpne32", CODE_FOR_fcmpne32di_vis
,
10384 SPARC_BUILTIN_FCMPNE32
, di_ftype_v2si_v2si
);
10385 def_builtin_const ("__builtin_vis_fcmpgt16", CODE_FOR_fcmpgt16di_vis
,
10386 SPARC_BUILTIN_FCMPGT16
, di_ftype_v4hi_v4hi
);
10387 def_builtin_const ("__builtin_vis_fcmpgt32", CODE_FOR_fcmpgt32di_vis
,
10388 SPARC_BUILTIN_FCMPGT32
, di_ftype_v2si_v2si
);
10389 def_builtin_const ("__builtin_vis_fcmpeq16", CODE_FOR_fcmpeq16di_vis
,
10390 SPARC_BUILTIN_FCMPEQ16
, di_ftype_v4hi_v4hi
);
10391 def_builtin_const ("__builtin_vis_fcmpeq32", CODE_FOR_fcmpeq32di_vis
,
10392 SPARC_BUILTIN_FCMPEQ32
, di_ftype_v2si_v2si
);
10396 def_builtin_const ("__builtin_vis_fcmple16", CODE_FOR_fcmple16si_vis
,
10397 SPARC_BUILTIN_FCMPLE16
, si_ftype_v4hi_v4hi
);
10398 def_builtin_const ("__builtin_vis_fcmple32", CODE_FOR_fcmple32si_vis
,
10399 SPARC_BUILTIN_FCMPLE32
, si_ftype_v2si_v2si
);
10400 def_builtin_const ("__builtin_vis_fcmpne16", CODE_FOR_fcmpne16si_vis
,
10401 SPARC_BUILTIN_FCMPNE16
, si_ftype_v4hi_v4hi
);
10402 def_builtin_const ("__builtin_vis_fcmpne32", CODE_FOR_fcmpne32si_vis
,
10403 SPARC_BUILTIN_FCMPNE32
, si_ftype_v2si_v2si
);
10404 def_builtin_const ("__builtin_vis_fcmpgt16", CODE_FOR_fcmpgt16si_vis
,
10405 SPARC_BUILTIN_FCMPGT16
, si_ftype_v4hi_v4hi
);
10406 def_builtin_const ("__builtin_vis_fcmpgt32", CODE_FOR_fcmpgt32si_vis
,
10407 SPARC_BUILTIN_FCMPGT32
, si_ftype_v2si_v2si
);
10408 def_builtin_const ("__builtin_vis_fcmpeq16", CODE_FOR_fcmpeq16si_vis
,
10409 SPARC_BUILTIN_FCMPEQ16
, si_ftype_v4hi_v4hi
);
10410 def_builtin_const ("__builtin_vis_fcmpeq32", CODE_FOR_fcmpeq32si_vis
,
10411 SPARC_BUILTIN_FCMPEQ32
, si_ftype_v2si_v2si
);
10414 /* Addition and subtraction. */
10415 def_builtin_const ("__builtin_vis_fpadd16", CODE_FOR_addv4hi3
,
10416 SPARC_BUILTIN_FPADD16
, v4hi_ftype_v4hi_v4hi
);
10417 def_builtin_const ("__builtin_vis_fpadd16s", CODE_FOR_addv2hi3
,
10418 SPARC_BUILTIN_FPADD16S
, v2hi_ftype_v2hi_v2hi
);
10419 def_builtin_const ("__builtin_vis_fpadd32", CODE_FOR_addv2si3
,
10420 SPARC_BUILTIN_FPADD32
, v2si_ftype_v2si_v2si
);
10421 def_builtin_const ("__builtin_vis_fpadd32s", CODE_FOR_addv1si3
,
10422 SPARC_BUILTIN_FPADD32S
, v1si_ftype_v1si_v1si
);
10423 def_builtin_const ("__builtin_vis_fpsub16", CODE_FOR_subv4hi3
,
10424 SPARC_BUILTIN_FPSUB16
, v4hi_ftype_v4hi_v4hi
);
10425 def_builtin_const ("__builtin_vis_fpsub16s", CODE_FOR_subv2hi3
,
10426 SPARC_BUILTIN_FPSUB16S
, v2hi_ftype_v2hi_v2hi
);
10427 def_builtin_const ("__builtin_vis_fpsub32", CODE_FOR_subv2si3
,
10428 SPARC_BUILTIN_FPSUB32
, v2si_ftype_v2si_v2si
);
10429 def_builtin_const ("__builtin_vis_fpsub32s", CODE_FOR_subv1si3
,
10430 SPARC_BUILTIN_FPSUB32S
, v1si_ftype_v1si_v1si
);
10432 /* Three-dimensional array addressing. */
10435 def_builtin_const ("__builtin_vis_array8", CODE_FOR_array8di_vis
,
10436 SPARC_BUILTIN_ARRAY8
, di_ftype_di_di
);
10437 def_builtin_const ("__builtin_vis_array16", CODE_FOR_array16di_vis
,
10438 SPARC_BUILTIN_ARRAY16
, di_ftype_di_di
);
10439 def_builtin_const ("__builtin_vis_array32", CODE_FOR_array32di_vis
,
10440 SPARC_BUILTIN_ARRAY32
, di_ftype_di_di
);
10444 def_builtin_const ("__builtin_vis_array8", CODE_FOR_array8si_vis
,
10445 SPARC_BUILTIN_ARRAY8
, si_ftype_si_si
);
10446 def_builtin_const ("__builtin_vis_array16", CODE_FOR_array16si_vis
,
10447 SPARC_BUILTIN_ARRAY16
, si_ftype_si_si
);
10448 def_builtin_const ("__builtin_vis_array32", CODE_FOR_array32si_vis
,
10449 SPARC_BUILTIN_ARRAY32
, si_ftype_si_si
);
10454 /* Edge handling. */
10457 def_builtin_const ("__builtin_vis_edge8n", CODE_FOR_edge8ndi_vis
,
10458 SPARC_BUILTIN_EDGE8N
, di_ftype_ptr_ptr
);
10459 def_builtin_const ("__builtin_vis_edge8ln", CODE_FOR_edge8lndi_vis
,
10460 SPARC_BUILTIN_EDGE8LN
, di_ftype_ptr_ptr
);
10461 def_builtin_const ("__builtin_vis_edge16n", CODE_FOR_edge16ndi_vis
,
10462 SPARC_BUILTIN_EDGE16N
, di_ftype_ptr_ptr
);
10463 def_builtin_const ("__builtin_vis_edge16ln", CODE_FOR_edge16lndi_vis
,
10464 SPARC_BUILTIN_EDGE16LN
, di_ftype_ptr_ptr
);
10465 def_builtin_const ("__builtin_vis_edge32n", CODE_FOR_edge32ndi_vis
,
10466 SPARC_BUILTIN_EDGE32N
, di_ftype_ptr_ptr
);
10467 def_builtin_const ("__builtin_vis_edge32ln", CODE_FOR_edge32lndi_vis
,
10468 SPARC_BUILTIN_EDGE32LN
, di_ftype_ptr_ptr
);
10472 def_builtin_const ("__builtin_vis_edge8n", CODE_FOR_edge8nsi_vis
,
10473 SPARC_BUILTIN_EDGE8N
, si_ftype_ptr_ptr
);
10474 def_builtin_const ("__builtin_vis_edge8ln", CODE_FOR_edge8lnsi_vis
,
10475 SPARC_BUILTIN_EDGE8LN
, si_ftype_ptr_ptr
);
10476 def_builtin_const ("__builtin_vis_edge16n", CODE_FOR_edge16nsi_vis
,
10477 SPARC_BUILTIN_EDGE16N
, si_ftype_ptr_ptr
);
10478 def_builtin_const ("__builtin_vis_edge16ln", CODE_FOR_edge16lnsi_vis
,
10479 SPARC_BUILTIN_EDGE16LN
, si_ftype_ptr_ptr
);
10480 def_builtin_const ("__builtin_vis_edge32n", CODE_FOR_edge32nsi_vis
,
10481 SPARC_BUILTIN_EDGE32N
, si_ftype_ptr_ptr
);
10482 def_builtin_const ("__builtin_vis_edge32ln", CODE_FOR_edge32lnsi_vis
,
10483 SPARC_BUILTIN_EDGE32LN
, si_ftype_ptr_ptr
);
10486 /* Byte mask and shuffle. */
10488 def_builtin ("__builtin_vis_bmask", CODE_FOR_bmaskdi_vis
,
10489 SPARC_BUILTIN_BMASK
, di_ftype_di_di
);
10491 def_builtin ("__builtin_vis_bmask", CODE_FOR_bmasksi_vis
,
10492 SPARC_BUILTIN_BMASK
, si_ftype_si_si
);
10493 def_builtin ("__builtin_vis_bshufflev4hi", CODE_FOR_bshufflev4hi_vis
,
10494 SPARC_BUILTIN_BSHUFFLEV4HI
, v4hi_ftype_v4hi_v4hi
);
10495 def_builtin ("__builtin_vis_bshufflev8qi", CODE_FOR_bshufflev8qi_vis
,
10496 SPARC_BUILTIN_BSHUFFLEV8QI
, v8qi_ftype_v8qi_v8qi
);
10497 def_builtin ("__builtin_vis_bshufflev2si", CODE_FOR_bshufflev2si_vis
,
10498 SPARC_BUILTIN_BSHUFFLEV2SI
, v2si_ftype_v2si_v2si
);
10499 def_builtin ("__builtin_vis_bshuffledi", CODE_FOR_bshufflev1di_vis
,
10500 SPARC_BUILTIN_BSHUFFLEDI
, di_ftype_di_di
);
10507 def_builtin ("__builtin_vis_cmask8", CODE_FOR_cmask8di_vis
,
10508 SPARC_BUILTIN_CMASK8
, void_ftype_di
);
10509 def_builtin ("__builtin_vis_cmask16", CODE_FOR_cmask16di_vis
,
10510 SPARC_BUILTIN_CMASK16
, void_ftype_di
);
10511 def_builtin ("__builtin_vis_cmask32", CODE_FOR_cmask32di_vis
,
10512 SPARC_BUILTIN_CMASK32
, void_ftype_di
);
10516 def_builtin ("__builtin_vis_cmask8", CODE_FOR_cmask8si_vis
,
10517 SPARC_BUILTIN_CMASK8
, void_ftype_si
);
10518 def_builtin ("__builtin_vis_cmask16", CODE_FOR_cmask16si_vis
,
10519 SPARC_BUILTIN_CMASK16
, void_ftype_si
);
10520 def_builtin ("__builtin_vis_cmask32", CODE_FOR_cmask32si_vis
,
10521 SPARC_BUILTIN_CMASK32
, void_ftype_si
);
10524 def_builtin_const ("__builtin_vis_fchksm16", CODE_FOR_fchksm16_vis
,
10525 SPARC_BUILTIN_FCHKSM16
, v4hi_ftype_v4hi_v4hi
);
10527 def_builtin_const ("__builtin_vis_fsll16", CODE_FOR_vashlv4hi3
,
10528 SPARC_BUILTIN_FSLL16
, v4hi_ftype_v4hi_v4hi
);
10529 def_builtin_const ("__builtin_vis_fslas16", CODE_FOR_vssashlv4hi3
,
10530 SPARC_BUILTIN_FSLAS16
, v4hi_ftype_v4hi_v4hi
);
10531 def_builtin_const ("__builtin_vis_fsrl16", CODE_FOR_vlshrv4hi3
,
10532 SPARC_BUILTIN_FSRL16
, v4hi_ftype_v4hi_v4hi
);
10533 def_builtin_const ("__builtin_vis_fsra16", CODE_FOR_vashrv4hi3
,
10534 SPARC_BUILTIN_FSRA16
, v4hi_ftype_v4hi_v4hi
);
10535 def_builtin_const ("__builtin_vis_fsll32", CODE_FOR_vashlv2si3
,
10536 SPARC_BUILTIN_FSLL32
, v2si_ftype_v2si_v2si
);
10537 def_builtin_const ("__builtin_vis_fslas32", CODE_FOR_vssashlv2si3
,
10538 SPARC_BUILTIN_FSLAS32
, v2si_ftype_v2si_v2si
);
10539 def_builtin_const ("__builtin_vis_fsrl32", CODE_FOR_vlshrv2si3
,
10540 SPARC_BUILTIN_FSRL32
, v2si_ftype_v2si_v2si
);
10541 def_builtin_const ("__builtin_vis_fsra32", CODE_FOR_vashrv2si3
,
10542 SPARC_BUILTIN_FSRA32
, v2si_ftype_v2si_v2si
);
10545 def_builtin_const ("__builtin_vis_pdistn", CODE_FOR_pdistndi_vis
,
10546 SPARC_BUILTIN_PDISTN
, di_ftype_v8qi_v8qi
);
10548 def_builtin_const ("__builtin_vis_pdistn", CODE_FOR_pdistnsi_vis
,
10549 SPARC_BUILTIN_PDISTN
, si_ftype_v8qi_v8qi
);
10551 def_builtin_const ("__builtin_vis_fmean16", CODE_FOR_fmean16_vis
,
10552 SPARC_BUILTIN_FMEAN16
, v4hi_ftype_v4hi_v4hi
);
10553 def_builtin_const ("__builtin_vis_fpadd64", CODE_FOR_fpadd64_vis
,
10554 SPARC_BUILTIN_FPADD64
, di_ftype_di_di
);
10555 def_builtin_const ("__builtin_vis_fpsub64", CODE_FOR_fpsub64_vis
,
10556 SPARC_BUILTIN_FPSUB64
, di_ftype_di_di
);
10558 def_builtin_const ("__builtin_vis_fpadds16", CODE_FOR_ssaddv4hi3
,
10559 SPARC_BUILTIN_FPADDS16
, v4hi_ftype_v4hi_v4hi
);
10560 def_builtin_const ("__builtin_vis_fpadds16s", CODE_FOR_ssaddv2hi3
,
10561 SPARC_BUILTIN_FPADDS16S
, v2hi_ftype_v2hi_v2hi
);
10562 def_builtin_const ("__builtin_vis_fpsubs16", CODE_FOR_sssubv4hi3
,
10563 SPARC_BUILTIN_FPSUBS16
, v4hi_ftype_v4hi_v4hi
);
10564 def_builtin_const ("__builtin_vis_fpsubs16s", CODE_FOR_sssubv2hi3
,
10565 SPARC_BUILTIN_FPSUBS16S
, v2hi_ftype_v2hi_v2hi
);
10566 def_builtin_const ("__builtin_vis_fpadds32", CODE_FOR_ssaddv2si3
,
10567 SPARC_BUILTIN_FPADDS32
, v2si_ftype_v2si_v2si
);
10568 def_builtin_const ("__builtin_vis_fpadds32s", CODE_FOR_ssaddv1si3
,
10569 SPARC_BUILTIN_FPADDS32S
, v1si_ftype_v1si_v1si
);
10570 def_builtin_const ("__builtin_vis_fpsubs32", CODE_FOR_sssubv2si3
,
10571 SPARC_BUILTIN_FPSUBS32
, v2si_ftype_v2si_v2si
);
10572 def_builtin_const ("__builtin_vis_fpsubs32s", CODE_FOR_sssubv1si3
,
10573 SPARC_BUILTIN_FPSUBS32S
, v1si_ftype_v1si_v1si
);
10577 def_builtin_const ("__builtin_vis_fucmple8", CODE_FOR_fucmple8di_vis
,
10578 SPARC_BUILTIN_FUCMPLE8
, di_ftype_v8qi_v8qi
);
10579 def_builtin_const ("__builtin_vis_fucmpne8", CODE_FOR_fucmpne8di_vis
,
10580 SPARC_BUILTIN_FUCMPNE8
, di_ftype_v8qi_v8qi
);
10581 def_builtin_const ("__builtin_vis_fucmpgt8", CODE_FOR_fucmpgt8di_vis
,
10582 SPARC_BUILTIN_FUCMPGT8
, di_ftype_v8qi_v8qi
);
10583 def_builtin_const ("__builtin_vis_fucmpeq8", CODE_FOR_fucmpeq8di_vis
,
10584 SPARC_BUILTIN_FUCMPEQ8
, di_ftype_v8qi_v8qi
);
10588 def_builtin_const ("__builtin_vis_fucmple8", CODE_FOR_fucmple8si_vis
,
10589 SPARC_BUILTIN_FUCMPLE8
, si_ftype_v8qi_v8qi
);
10590 def_builtin_const ("__builtin_vis_fucmpne8", CODE_FOR_fucmpne8si_vis
,
10591 SPARC_BUILTIN_FUCMPNE8
, si_ftype_v8qi_v8qi
);
10592 def_builtin_const ("__builtin_vis_fucmpgt8", CODE_FOR_fucmpgt8si_vis
,
10593 SPARC_BUILTIN_FUCMPGT8
, si_ftype_v8qi_v8qi
);
10594 def_builtin_const ("__builtin_vis_fucmpeq8", CODE_FOR_fucmpeq8si_vis
,
10595 SPARC_BUILTIN_FUCMPEQ8
, si_ftype_v8qi_v8qi
);
10598 def_builtin_const ("__builtin_vis_fhadds", CODE_FOR_fhaddsf_vis
,
10599 SPARC_BUILTIN_FHADDS
, sf_ftype_sf_sf
);
10600 def_builtin_const ("__builtin_vis_fhaddd", CODE_FOR_fhadddf_vis
,
10601 SPARC_BUILTIN_FHADDD
, df_ftype_df_df
);
10602 def_builtin_const ("__builtin_vis_fhsubs", CODE_FOR_fhsubsf_vis
,
10603 SPARC_BUILTIN_FHSUBS
, sf_ftype_sf_sf
);
10604 def_builtin_const ("__builtin_vis_fhsubd", CODE_FOR_fhsubdf_vis
,
10605 SPARC_BUILTIN_FHSUBD
, df_ftype_df_df
);
10606 def_builtin_const ("__builtin_vis_fnhadds", CODE_FOR_fnhaddsf_vis
,
10607 SPARC_BUILTIN_FNHADDS
, sf_ftype_sf_sf
);
10608 def_builtin_const ("__builtin_vis_fnhaddd", CODE_FOR_fnhadddf_vis
,
10609 SPARC_BUILTIN_FNHADDD
, df_ftype_df_df
);
10611 def_builtin_const ("__builtin_vis_umulxhi", CODE_FOR_umulxhi_vis
,
10612 SPARC_BUILTIN_UMULXHI
, di_ftype_di_di
);
10613 def_builtin_const ("__builtin_vis_xmulx", CODE_FOR_xmulx_vis
,
10614 SPARC_BUILTIN_XMULX
, di_ftype_di_di
);
10615 def_builtin_const ("__builtin_vis_xmulxhi", CODE_FOR_xmulxhi_vis
,
10616 SPARC_BUILTIN_XMULXHI
, di_ftype_di_di
);
10621 def_builtin_const ("__builtin_vis_fpadd8", CODE_FOR_addv8qi3
,
10622 SPARC_BUILTIN_FPADD8
, v8qi_ftype_v8qi_v8qi
);
10623 def_builtin_const ("__builtin_vis_fpadds8", CODE_FOR_ssaddv8qi3
,
10624 SPARC_BUILTIN_FPADDS8
, v8qi_ftype_v8qi_v8qi
);
10625 def_builtin_const ("__builtin_vis_fpaddus8", CODE_FOR_usaddv8qi3
,
10626 SPARC_BUILTIN_FPADDUS8
, v8qi_ftype_v8qi_v8qi
);
10627 def_builtin_const ("__builtin_vis_fpaddus16", CODE_FOR_usaddv4hi3
,
10628 SPARC_BUILTIN_FPADDUS16
, v4hi_ftype_v4hi_v4hi
);
10633 def_builtin_const ("__builtin_vis_fpcmple8", CODE_FOR_fpcmple8di_vis
,
10634 SPARC_BUILTIN_FPCMPLE8
, di_ftype_v8qi_v8qi
);
10635 def_builtin_const ("__builtin_vis_fpcmpgt8", CODE_FOR_fpcmpgt8di_vis
,
10636 SPARC_BUILTIN_FPCMPGT8
, di_ftype_v8qi_v8qi
);
10637 def_builtin_const ("__builtin_vis_fpcmpule16", CODE_FOR_fpcmpule16di_vis
,
10638 SPARC_BUILTIN_FPCMPULE16
, di_ftype_v4hi_v4hi
);
10639 def_builtin_const ("__builtin_vis_fpcmpugt16", CODE_FOR_fpcmpugt16di_vis
,
10640 SPARC_BUILTIN_FPCMPUGT16
, di_ftype_v4hi_v4hi
);
10641 def_builtin_const ("__builtin_vis_fpcmpule32", CODE_FOR_fpcmpule32di_vis
,
10642 SPARC_BUILTIN_FPCMPULE32
, di_ftype_v2si_v2si
);
10643 def_builtin_const ("__builtin_vis_fpcmpugt32", CODE_FOR_fpcmpugt32di_vis
,
10644 SPARC_BUILTIN_FPCMPUGT32
, di_ftype_v2si_v2si
);
10648 def_builtin_const ("__builtin_vis_fpcmple8", CODE_FOR_fpcmple8si_vis
,
10649 SPARC_BUILTIN_FPCMPLE8
, si_ftype_v8qi_v8qi
);
10650 def_builtin_const ("__builtin_vis_fpcmpgt8", CODE_FOR_fpcmpgt8si_vis
,
10651 SPARC_BUILTIN_FPCMPGT8
, si_ftype_v8qi_v8qi
);
10652 def_builtin_const ("__builtin_vis_fpcmpule16", CODE_FOR_fpcmpule16si_vis
,
10653 SPARC_BUILTIN_FPCMPULE16
, si_ftype_v4hi_v4hi
);
10654 def_builtin_const ("__builtin_vis_fpcmpugt16", CODE_FOR_fpcmpugt16si_vis
,
10655 SPARC_BUILTIN_FPCMPUGT16
, si_ftype_v4hi_v4hi
);
10656 def_builtin_const ("__builtin_vis_fpcmpule32", CODE_FOR_fpcmpule32si_vis
,
10657 SPARC_BUILTIN_FPCMPULE32
, di_ftype_v2si_v2si
);
10658 def_builtin_const ("__builtin_vis_fpcmpugt32", CODE_FOR_fpcmpugt32si_vis
,
10659 SPARC_BUILTIN_FPCMPUGT32
, di_ftype_v2si_v2si
);
10662 def_builtin_const ("__builtin_vis_fpmax8", CODE_FOR_maxv8qi3
,
10663 SPARC_BUILTIN_FPMAX8
, v8qi_ftype_v8qi_v8qi
);
10664 def_builtin_const ("__builtin_vis_fpmax16", CODE_FOR_maxv4hi3
,
10665 SPARC_BUILTIN_FPMAX16
, v4hi_ftype_v4hi_v4hi
);
10666 def_builtin_const ("__builtin_vis_fpmax32", CODE_FOR_maxv2si3
,
10667 SPARC_BUILTIN_FPMAX32
, v2si_ftype_v2si_v2si
);
10668 def_builtin_const ("__builtin_vis_fpmaxu8", CODE_FOR_maxuv8qi3
,
10669 SPARC_BUILTIN_FPMAXU8
, v8qi_ftype_v8qi_v8qi
);
10670 def_builtin_const ("__builtin_vis_fpmaxu16", CODE_FOR_maxuv4hi3
,
10671 SPARC_BUILTIN_FPMAXU16
, v4hi_ftype_v4hi_v4hi
);
10672 def_builtin_const ("__builtin_vis_fpmaxu32", CODE_FOR_maxuv2si3
,
10673 SPARC_BUILTIN_FPMAXU32
, v2si_ftype_v2si_v2si
);
10674 def_builtin_const ("__builtin_vis_fpmin8", CODE_FOR_minv8qi3
,
10675 SPARC_BUILTIN_FPMIN8
, v8qi_ftype_v8qi_v8qi
);
10676 def_builtin_const ("__builtin_vis_fpmin16", CODE_FOR_minv4hi3
,
10677 SPARC_BUILTIN_FPMIN16
, v4hi_ftype_v4hi_v4hi
);
10678 def_builtin_const ("__builtin_vis_fpmin32", CODE_FOR_minv2si3
,
10679 SPARC_BUILTIN_FPMIN32
, v2si_ftype_v2si_v2si
);
10680 def_builtin_const ("__builtin_vis_fpminu8", CODE_FOR_minuv8qi3
,
10681 SPARC_BUILTIN_FPMINU8
, v8qi_ftype_v8qi_v8qi
);
10682 def_builtin_const ("__builtin_vis_fpminu16", CODE_FOR_minuv4hi3
,
10683 SPARC_BUILTIN_FPMINU16
, v4hi_ftype_v4hi_v4hi
);
10684 def_builtin_const ("__builtin_vis_fpminu32", CODE_FOR_minuv2si3
,
10685 SPARC_BUILTIN_FPMINU32
, v2si_ftype_v2si_v2si
);
10686 def_builtin_const ("__builtin_vis_fpsub8", CODE_FOR_subv8qi3
,
10687 SPARC_BUILTIN_FPSUB8
, v8qi_ftype_v8qi_v8qi
);
10688 def_builtin_const ("__builtin_vis_fpsubs8", CODE_FOR_sssubv8qi3
,
10689 SPARC_BUILTIN_FPSUBS8
, v8qi_ftype_v8qi_v8qi
);
10690 def_builtin_const ("__builtin_vis_fpsubus8", CODE_FOR_ussubv8qi3
,
10691 SPARC_BUILTIN_FPSUBUS8
, v8qi_ftype_v8qi_v8qi
);
10692 def_builtin_const ("__builtin_vis_fpsubus16", CODE_FOR_ussubv4hi3
,
10693 SPARC_BUILTIN_FPSUBUS16
, v4hi_ftype_v4hi_v4hi
);
10697 /* Implement TARGET_BUILTIN_DECL hook. */
10700 sparc_builtin_decl (unsigned code
, bool initialize_p ATTRIBUTE_UNUSED
)
10702 if (code
>= SPARC_BUILTIN_MAX
)
10703 return error_mark_node
;
10705 return sparc_builtins
[code
];
10708 /* Implemented TARGET_EXPAND_BUILTIN hook. */
10711 sparc_expand_builtin (tree exp
, rtx target
,
10712 rtx subtarget ATTRIBUTE_UNUSED
,
10713 machine_mode tmode ATTRIBUTE_UNUSED
,
10714 int ignore ATTRIBUTE_UNUSED
)
10716 tree fndecl
= TREE_OPERAND (CALL_EXPR_FN (exp
), 0);
10717 enum sparc_builtins code
= (enum sparc_builtins
) DECL_FUNCTION_CODE (fndecl
);
10718 enum insn_code icode
= sparc_builtins_icode
[code
];
10719 bool nonvoid
= TREE_TYPE (TREE_TYPE (fndecl
)) != void_type_node
;
10720 call_expr_arg_iterator iter
;
10727 machine_mode tmode
= insn_data
[icode
].operand
[0].mode
;
10729 || GET_MODE (target
) != tmode
10730 || ! (*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
10731 op
[0] = gen_reg_rtx (tmode
);
10736 FOR_EACH_CALL_EXPR_ARG (arg
, iter
, exp
)
10738 const struct insn_operand_data
*insn_op
;
10741 if (arg
== error_mark_node
)
10745 idx
= arg_count
- !nonvoid
;
10746 insn_op
= &insn_data
[icode
].operand
[idx
];
10747 op
[arg_count
] = expand_normal (arg
);
10749 if (code
== SPARC_BUILTIN_LDFSR
|| code
== SPARC_BUILTIN_STFSR
)
10751 if (!address_operand (op
[arg_count
], SImode
))
10753 op
[arg_count
] = convert_memory_address (Pmode
, op
[arg_count
]);
10754 op
[arg_count
] = copy_addr_to_reg (op
[arg_count
]);
10756 op
[arg_count
] = gen_rtx_MEM (SImode
, op
[arg_count
]);
10759 else if (insn_op
->mode
== V1DImode
10760 && GET_MODE (op
[arg_count
]) == DImode
)
10761 op
[arg_count
] = gen_lowpart (V1DImode
, op
[arg_count
]);
10763 else if (insn_op
->mode
== V1SImode
10764 && GET_MODE (op
[arg_count
]) == SImode
)
10765 op
[arg_count
] = gen_lowpart (V1SImode
, op
[arg_count
]);
10767 if (! (*insn_data
[icode
].operand
[idx
].predicate
) (op
[arg_count
],
10769 op
[arg_count
] = copy_to_mode_reg (insn_op
->mode
, op
[arg_count
]);
10775 pat
= GEN_FCN (icode
) (op
[0]);
10779 pat
= GEN_FCN (icode
) (op
[0], op
[1]);
10781 pat
= GEN_FCN (icode
) (op
[1]);
10784 pat
= GEN_FCN (icode
) (op
[0], op
[1], op
[2]);
10787 pat
= GEN_FCN (icode
) (op
[0], op
[1], op
[2], op
[3]);
10790 gcc_unreachable ();
10798 return (nonvoid
? op
[0] : const0_rtx
);
10801 /* Return the upper 16 bits of the 8x16 multiplication. */
10804 sparc_vis_mul8x16 (int e8
, int e16
)
10806 return (e8
* e16
+ 128) / 256;
10809 /* Multiply the VECTOR_CSTs CST0 and CST1 as specified by FNCODE and put
10810 the result into the array N_ELTS, whose elements are of INNER_TYPE. */
10813 sparc_handle_vis_mul8x16 (tree
*n_elts
, enum sparc_builtins fncode
,
10814 tree inner_type
, tree cst0
, tree cst1
)
10816 unsigned i
, num
= VECTOR_CST_NELTS (cst0
);
10821 case SPARC_BUILTIN_FMUL8X16
:
10822 for (i
= 0; i
< num
; ++i
)
10825 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (VECTOR_CST_ELT (cst0
, i
)),
10826 TREE_INT_CST_LOW (VECTOR_CST_ELT (cst1
, i
)));
10827 n_elts
[i
] = build_int_cst (inner_type
, val
);
10831 case SPARC_BUILTIN_FMUL8X16AU
:
10832 scale
= TREE_INT_CST_LOW (VECTOR_CST_ELT (cst1
, 0));
10834 for (i
= 0; i
< num
; ++i
)
10837 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (VECTOR_CST_ELT (cst0
, i
)),
10839 n_elts
[i
] = build_int_cst (inner_type
, val
);
10843 case SPARC_BUILTIN_FMUL8X16AL
:
10844 scale
= TREE_INT_CST_LOW (VECTOR_CST_ELT (cst1
, 1));
10846 for (i
= 0; i
< num
; ++i
)
10849 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (VECTOR_CST_ELT (cst0
, i
)),
10851 n_elts
[i
] = build_int_cst (inner_type
, val
);
10856 gcc_unreachable ();
10860 /* Implement TARGET_FOLD_BUILTIN hook.
10862 Fold builtin functions for SPARC intrinsics. If IGNORE is true the
10863 result of the function call is ignored. NULL_TREE is returned if the
10864 function could not be folded. */
10867 sparc_fold_builtin (tree fndecl
, int n_args ATTRIBUTE_UNUSED
,
10868 tree
*args
, bool ignore
)
10870 enum sparc_builtins code
= (enum sparc_builtins
) DECL_FUNCTION_CODE (fndecl
);
10871 tree rtype
= TREE_TYPE (TREE_TYPE (fndecl
));
10872 tree arg0
, arg1
, arg2
;
10877 case SPARC_BUILTIN_LDFSR
:
10878 case SPARC_BUILTIN_STFSR
:
10879 case SPARC_BUILTIN_ALIGNADDR
:
10880 case SPARC_BUILTIN_WRGSR
:
10881 case SPARC_BUILTIN_BMASK
:
10882 case SPARC_BUILTIN_CMASK8
:
10883 case SPARC_BUILTIN_CMASK16
:
10884 case SPARC_BUILTIN_CMASK32
:
10888 return build_zero_cst (rtype
);
10893 case SPARC_BUILTIN_FEXPAND
:
10897 if (TREE_CODE (arg0
) == VECTOR_CST
)
10899 tree inner_type
= TREE_TYPE (rtype
);
10903 n_elts
= XALLOCAVEC (tree
, VECTOR_CST_NELTS (arg0
));
10904 for (i
= 0; i
< VECTOR_CST_NELTS (arg0
); ++i
)
10905 n_elts
[i
] = build_int_cst (inner_type
,
10907 (VECTOR_CST_ELT (arg0
, i
)) << 4);
10908 return build_vector (rtype
, n_elts
);
10912 case SPARC_BUILTIN_FMUL8X16
:
10913 case SPARC_BUILTIN_FMUL8X16AU
:
10914 case SPARC_BUILTIN_FMUL8X16AL
:
10920 if (TREE_CODE (arg0
) == VECTOR_CST
&& TREE_CODE (arg1
) == VECTOR_CST
)
10922 tree inner_type
= TREE_TYPE (rtype
);
10923 tree
*n_elts
= XALLOCAVEC (tree
, VECTOR_CST_NELTS (arg0
));
10924 sparc_handle_vis_mul8x16 (n_elts
, code
, inner_type
, arg0
, arg1
);
10925 return build_vector (rtype
, n_elts
);
10929 case SPARC_BUILTIN_FPMERGE
:
10935 if (TREE_CODE (arg0
) == VECTOR_CST
&& TREE_CODE (arg1
) == VECTOR_CST
)
10937 tree
*n_elts
= XALLOCAVEC (tree
, 2 * VECTOR_CST_NELTS (arg0
));
10939 for (i
= 0; i
< VECTOR_CST_NELTS (arg0
); ++i
)
10941 n_elts
[2*i
] = VECTOR_CST_ELT (arg0
, i
);
10942 n_elts
[2*i
+1] = VECTOR_CST_ELT (arg1
, i
);
10945 return build_vector (rtype
, n_elts
);
10949 case SPARC_BUILTIN_PDIST
:
10950 case SPARC_BUILTIN_PDISTN
:
10955 if (code
== SPARC_BUILTIN_PDIST
)
10961 arg2
= integer_zero_node
;
10963 if (TREE_CODE (arg0
) == VECTOR_CST
10964 && TREE_CODE (arg1
) == VECTOR_CST
10965 && TREE_CODE (arg2
) == INTEGER_CST
)
10967 bool overflow
= false;
10968 widest_int result
= wi::to_widest (arg2
);
10972 for (i
= 0; i
< VECTOR_CST_NELTS (arg0
); ++i
)
10974 tree e0
= VECTOR_CST_ELT (arg0
, i
);
10975 tree e1
= VECTOR_CST_ELT (arg1
, i
);
10977 bool neg1_ovf
, neg2_ovf
, add1_ovf
, add2_ovf
;
10979 tmp
= wi::neg (wi::to_widest (e1
), &neg1_ovf
);
10980 tmp
= wi::add (wi::to_widest (e0
), tmp
, SIGNED
, &add1_ovf
);
10981 if (wi::neg_p (tmp
))
10982 tmp
= wi::neg (tmp
, &neg2_ovf
);
10985 result
= wi::add (result
, tmp
, SIGNED
, &add2_ovf
);
10986 overflow
|= neg1_ovf
| neg2_ovf
| add1_ovf
| add2_ovf
;
10989 gcc_assert (!overflow
);
10991 return wide_int_to_tree (rtype
, result
);
11001 /* ??? This duplicates information provided to the compiler by the
11002 ??? scheduler description. Some day, teach genautomata to output
11003 ??? the latencies and then CSE will just use that. */
11006 sparc_rtx_costs (rtx x
, machine_mode mode
, int outer_code
,
11007 int opno ATTRIBUTE_UNUSED
,
11008 int *total
, bool speed ATTRIBUTE_UNUSED
)
11010 int code
= GET_CODE (x
);
11011 bool float_mode_p
= FLOAT_MODE_P (mode
);
11022 case CONST_WIDE_INT
:
11024 if (!SPARC_SIMM13_P (CONST_WIDE_INT_ELT (x
, 0)))
11026 if (!SPARC_SIMM13_P (CONST_WIDE_INT_ELT (x
, 1)))
11045 /* If outer-code was a sign or zero extension, a cost
11046 of COSTS_N_INSNS (1) was already added in. This is
11047 why we are subtracting it back out. */
11048 if (outer_code
== ZERO_EXTEND
)
11050 *total
= sparc_costs
->int_zload
- COSTS_N_INSNS (1);
11052 else if (outer_code
== SIGN_EXTEND
)
11054 *total
= sparc_costs
->int_sload
- COSTS_N_INSNS (1);
11056 else if (float_mode_p
)
11058 *total
= sparc_costs
->float_load
;
11062 *total
= sparc_costs
->int_load
;
11070 *total
= sparc_costs
->float_plusminus
;
11072 *total
= COSTS_N_INSNS (1);
11079 gcc_assert (float_mode_p
);
11080 *total
= sparc_costs
->float_mul
;
11083 if (GET_CODE (sub
) == NEG
)
11084 sub
= XEXP (sub
, 0);
11085 *total
+= rtx_cost (sub
, mode
, FMA
, 0, speed
);
11088 if (GET_CODE (sub
) == NEG
)
11089 sub
= XEXP (sub
, 0);
11090 *total
+= rtx_cost (sub
, mode
, FMA
, 2, speed
);
11096 *total
= sparc_costs
->float_mul
;
11097 else if (TARGET_ARCH32
&& !TARGET_HARD_MUL
)
11098 *total
= COSTS_N_INSNS (25);
11104 if (sparc_costs
->int_mul_bit_factor
)
11108 if (GET_CODE (XEXP (x
, 1)) == CONST_INT
)
11110 unsigned HOST_WIDE_INT value
= INTVAL (XEXP (x
, 1));
11111 for (nbits
= 0; value
!= 0; value
&= value
- 1)
11119 bit_cost
= (nbits
- 3) / sparc_costs
->int_mul_bit_factor
;
11120 bit_cost
= COSTS_N_INSNS (bit_cost
);
11123 if (mode
== DImode
|| !TARGET_HARD_MUL
)
11124 *total
= sparc_costs
->int_mulX
+ bit_cost
;
11126 *total
= sparc_costs
->int_mul
+ bit_cost
;
11133 *total
= COSTS_N_INSNS (1) + sparc_costs
->shift_penalty
;
11142 if (mode
== DFmode
)
11143 *total
= sparc_costs
->float_div_df
;
11145 *total
= sparc_costs
->float_div_sf
;
11149 if (mode
== DImode
)
11150 *total
= sparc_costs
->int_divX
;
11152 *total
= sparc_costs
->int_div
;
11157 if (! float_mode_p
)
11159 *total
= COSTS_N_INSNS (1);
11166 case UNSIGNED_FLOAT
:
11170 case FLOAT_TRUNCATE
:
11171 *total
= sparc_costs
->float_move
;
11175 if (mode
== DFmode
)
11176 *total
= sparc_costs
->float_sqrt_df
;
11178 *total
= sparc_costs
->float_sqrt_sf
;
11183 *total
= sparc_costs
->float_cmp
;
11185 *total
= COSTS_N_INSNS (1);
11190 *total
= sparc_costs
->float_cmove
;
11192 *total
= sparc_costs
->int_cmove
;
11196 /* Handle the NAND vector patterns. */
11197 if (sparc_vector_mode_supported_p (mode
)
11198 && GET_CODE (XEXP (x
, 0)) == NOT
11199 && GET_CODE (XEXP (x
, 1)) == NOT
)
11201 *total
= COSTS_N_INSNS (1);
11212 /* Return true if CLASS is either GENERAL_REGS or I64_REGS. */
11215 general_or_i64_p (reg_class_t rclass
)
11217 return (rclass
== GENERAL_REGS
|| rclass
== I64_REGS
);
11220 /* Implement TARGET_REGISTER_MOVE_COST. */
11223 sparc_register_move_cost (machine_mode mode ATTRIBUTE_UNUSED
,
11224 reg_class_t from
, reg_class_t to
)
11226 bool need_memory
= false;
11228 if (from
== FPCC_REGS
|| to
== FPCC_REGS
)
11229 need_memory
= true;
11230 else if ((FP_REG_CLASS_P (from
) && general_or_i64_p (to
))
11231 || (general_or_i64_p (from
) && FP_REG_CLASS_P (to
)))
11235 int size
= GET_MODE_SIZE (mode
);
11236 if (size
== 8 || size
== 4)
11238 if (! TARGET_ARCH32
|| size
== 4)
11244 need_memory
= true;
11249 if (sparc_cpu
== PROCESSOR_ULTRASPARC
11250 || sparc_cpu
== PROCESSOR_ULTRASPARC3
11251 || sparc_cpu
== PROCESSOR_NIAGARA
11252 || sparc_cpu
== PROCESSOR_NIAGARA2
11253 || sparc_cpu
== PROCESSOR_NIAGARA3
11254 || sparc_cpu
== PROCESSOR_NIAGARA4
11255 || sparc_cpu
== PROCESSOR_NIAGARA7
)
11264 /* Emit the sequence of insns SEQ while preserving the registers REG and REG2.
11265 This is achieved by means of a manual dynamic stack space allocation in
11266 the current frame. We make the assumption that SEQ doesn't contain any
11267 function calls, with the possible exception of calls to the GOT helper. */
11270 emit_and_preserve (rtx seq
, rtx reg
, rtx reg2
)
11272 /* We must preserve the lowest 16 words for the register save area. */
11273 HOST_WIDE_INT offset
= 16*UNITS_PER_WORD
;
11274 /* We really need only 2 words of fresh stack space. */
11275 HOST_WIDE_INT size
= SPARC_STACK_ALIGN (offset
+ 2*UNITS_PER_WORD
);
11278 = gen_rtx_MEM (word_mode
, plus_constant (Pmode
, stack_pointer_rtx
,
11279 SPARC_STACK_BIAS
+ offset
));
11281 emit_insn (gen_stack_pointer_inc (GEN_INT (-size
)));
11282 emit_insn (gen_rtx_SET (slot
, reg
));
11284 emit_insn (gen_rtx_SET (adjust_address (slot
, word_mode
, UNITS_PER_WORD
),
11288 emit_insn (gen_rtx_SET (reg2
,
11289 adjust_address (slot
, word_mode
, UNITS_PER_WORD
)));
11290 emit_insn (gen_rtx_SET (reg
, slot
));
11291 emit_insn (gen_stack_pointer_inc (GEN_INT (size
)));
11294 /* Output the assembler code for a thunk function. THUNK_DECL is the
11295 declaration for the thunk function itself, FUNCTION is the decl for
11296 the target function. DELTA is an immediate constant offset to be
11297 added to THIS. If VCALL_OFFSET is nonzero, the word at address
11298 (*THIS + VCALL_OFFSET) should be additionally added to THIS. */
11301 sparc_output_mi_thunk (FILE *file
, tree thunk_fndecl ATTRIBUTE_UNUSED
,
11302 HOST_WIDE_INT delta
, HOST_WIDE_INT vcall_offset
,
11305 rtx this_rtx
, funexp
;
11307 unsigned int int_arg_first
;
11309 reload_completed
= 1;
11310 epilogue_completed
= 1;
11312 emit_note (NOTE_INSN_PROLOGUE_END
);
11316 sparc_leaf_function_p
= 1;
11318 int_arg_first
= SPARC_OUTGOING_INT_ARG_FIRST
;
11320 else if (flag_delayed_branch
)
11322 /* We will emit a regular sibcall below, so we need to instruct
11323 output_sibcall that we are in a leaf function. */
11324 sparc_leaf_function_p
= crtl
->uses_only_leaf_regs
= 1;
11326 /* This will cause final.c to invoke leaf_renumber_regs so we
11327 must behave as if we were in a not-yet-leafified function. */
11328 int_arg_first
= SPARC_INCOMING_INT_ARG_FIRST
;
11332 /* We will emit the sibcall manually below, so we will need to
11333 manually spill non-leaf registers. */
11334 sparc_leaf_function_p
= crtl
->uses_only_leaf_regs
= 0;
11336 /* We really are in a leaf function. */
11337 int_arg_first
= SPARC_OUTGOING_INT_ARG_FIRST
;
11340 /* Find the "this" pointer. Normally in %o0, but in ARCH64 if the function
11341 returns a structure, the structure return pointer is there instead. */
11343 && aggregate_value_p (TREE_TYPE (TREE_TYPE (function
)), function
))
11344 this_rtx
= gen_rtx_REG (Pmode
, int_arg_first
+ 1);
11346 this_rtx
= gen_rtx_REG (Pmode
, int_arg_first
);
11348 /* Add DELTA. When possible use a plain add, otherwise load it into
11349 a register first. */
11352 rtx delta_rtx
= GEN_INT (delta
);
11354 if (! SPARC_SIMM13_P (delta
))
11356 rtx scratch
= gen_rtx_REG (Pmode
, 1);
11357 emit_move_insn (scratch
, delta_rtx
);
11358 delta_rtx
= scratch
;
11361 /* THIS_RTX += DELTA. */
11362 emit_insn (gen_add2_insn (this_rtx
, delta_rtx
));
11365 /* Add the word at address (*THIS_RTX + VCALL_OFFSET). */
11368 rtx vcall_offset_rtx
= GEN_INT (vcall_offset
);
11369 rtx scratch
= gen_rtx_REG (Pmode
, 1);
11371 gcc_assert (vcall_offset
< 0);
11373 /* SCRATCH = *THIS_RTX. */
11374 emit_move_insn (scratch
, gen_rtx_MEM (Pmode
, this_rtx
));
11376 /* Prepare for adding VCALL_OFFSET. The difficulty is that we
11377 may not have any available scratch register at this point. */
11378 if (SPARC_SIMM13_P (vcall_offset
))
11380 /* This is the case if ARCH64 (unless -ffixed-g5 is passed). */
11381 else if (! fixed_regs
[5]
11382 /* The below sequence is made up of at least 2 insns,
11383 while the default method may need only one. */
11384 && vcall_offset
< -8192)
11386 rtx scratch2
= gen_rtx_REG (Pmode
, 5);
11387 emit_move_insn (scratch2
, vcall_offset_rtx
);
11388 vcall_offset_rtx
= scratch2
;
11392 rtx increment
= GEN_INT (-4096);
11394 /* VCALL_OFFSET is a negative number whose typical range can be
11395 estimated as -32768..0 in 32-bit mode. In almost all cases
11396 it is therefore cheaper to emit multiple add insns than
11397 spilling and loading the constant into a register (at least
11399 while (! SPARC_SIMM13_P (vcall_offset
))
11401 emit_insn (gen_add2_insn (scratch
, increment
));
11402 vcall_offset
+= 4096;
11404 vcall_offset_rtx
= GEN_INT (vcall_offset
); /* cannot be 0 */
11407 /* SCRATCH = *(*THIS_RTX + VCALL_OFFSET). */
11408 emit_move_insn (scratch
, gen_rtx_MEM (Pmode
,
11409 gen_rtx_PLUS (Pmode
,
11411 vcall_offset_rtx
)));
11413 /* THIS_RTX += *(*THIS_RTX + VCALL_OFFSET). */
11414 emit_insn (gen_add2_insn (this_rtx
, scratch
));
11417 /* Generate a tail call to the target function. */
11418 if (! TREE_USED (function
))
11420 assemble_external (function
);
11421 TREE_USED (function
) = 1;
11423 funexp
= XEXP (DECL_RTL (function
), 0);
11425 if (flag_delayed_branch
)
11427 funexp
= gen_rtx_MEM (FUNCTION_MODE
, funexp
);
11428 insn
= emit_call_insn (gen_sibcall (funexp
));
11429 SIBLING_CALL_P (insn
) = 1;
11433 /* The hoops we have to jump through in order to generate a sibcall
11434 without using delay slots... */
11435 rtx spill_reg
, seq
, scratch
= gen_rtx_REG (Pmode
, 1);
11439 spill_reg
= gen_rtx_REG (word_mode
, 15); /* %o7 */
11441 load_got_register (); /* clobbers %o7 */
11442 scratch
= sparc_legitimize_pic_address (funexp
, scratch
);
11443 seq
= get_insns ();
11445 emit_and_preserve (seq
, spill_reg
, pic_offset_table_rtx
);
11447 else if (TARGET_ARCH32
)
11449 emit_insn (gen_rtx_SET (scratch
,
11450 gen_rtx_HIGH (SImode
, funexp
)));
11451 emit_insn (gen_rtx_SET (scratch
,
11452 gen_rtx_LO_SUM (SImode
, scratch
, funexp
)));
11454 else /* TARGET_ARCH64 */
11456 switch (sparc_cmodel
)
11460 /* The destination can serve as a temporary. */
11461 sparc_emit_set_symbolic_const64 (scratch
, funexp
, scratch
);
11466 /* The destination cannot serve as a temporary. */
11467 spill_reg
= gen_rtx_REG (DImode
, 15); /* %o7 */
11469 sparc_emit_set_symbolic_const64 (scratch
, funexp
, spill_reg
);
11470 seq
= get_insns ();
11472 emit_and_preserve (seq
, spill_reg
, 0);
11476 gcc_unreachable ();
11480 emit_jump_insn (gen_indirect_jump (scratch
));
11485 /* Run just enough of rest_of_compilation to get the insns emitted.
11486 There's not really enough bulk here to make other passes such as
11487 instruction scheduling worth while. Note that use_thunk calls
11488 assemble_start_function and assemble_end_function. */
11489 insn
= get_insns ();
11490 shorten_branches (insn
);
11491 final_start_function (insn
, file
, 1);
11492 final (insn
, file
, 1);
11493 final_end_function ();
11495 reload_completed
= 0;
11496 epilogue_completed
= 0;
11499 /* Return true if sparc_output_mi_thunk would be able to output the
11500 assembler code for the thunk function specified by the arguments
11501 it is passed, and false otherwise. */
11503 sparc_can_output_mi_thunk (const_tree thunk_fndecl ATTRIBUTE_UNUSED
,
11504 HOST_WIDE_INT delta ATTRIBUTE_UNUSED
,
11505 HOST_WIDE_INT vcall_offset
,
11506 const_tree function ATTRIBUTE_UNUSED
)
11508 /* Bound the loop used in the default method above. */
11509 return (vcall_offset
>= -32768 || ! fixed_regs
[5]);
11512 /* How to allocate a 'struct machine_function'. */
11514 static struct machine_function
*
11515 sparc_init_machine_status (void)
11517 return ggc_cleared_alloc
<machine_function
> ();
11520 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
11521 We need to emit DTP-relative relocations. */
11524 sparc_output_dwarf_dtprel (FILE *file
, int size
, rtx x
)
11529 fputs ("\t.word\t%r_tls_dtpoff32(", file
);
11532 fputs ("\t.xword\t%r_tls_dtpoff64(", file
);
11535 gcc_unreachable ();
11537 output_addr_const (file
, x
);
11541 /* Do whatever processing is required at the end of a file. */
11544 sparc_file_end (void)
11546 /* If we need to emit the special GOT helper function, do so now. */
11547 if (got_helper_rtx
)
11549 const char *name
= XSTR (got_helper_rtx
, 0);
11550 const char *reg_name
= reg_names
[GLOBAL_OFFSET_TABLE_REGNUM
];
11551 #ifdef DWARF2_UNWIND_INFO
11555 if (USE_HIDDEN_LINKONCE
)
11557 tree decl
= build_decl (BUILTINS_LOCATION
, FUNCTION_DECL
,
11558 get_identifier (name
),
11559 build_function_type_list (void_type_node
,
11561 DECL_RESULT (decl
) = build_decl (BUILTINS_LOCATION
, RESULT_DECL
,
11562 NULL_TREE
, void_type_node
);
11563 TREE_PUBLIC (decl
) = 1;
11564 TREE_STATIC (decl
) = 1;
11565 make_decl_one_only (decl
, DECL_ASSEMBLER_NAME (decl
));
11566 DECL_VISIBILITY (decl
) = VISIBILITY_HIDDEN
;
11567 DECL_VISIBILITY_SPECIFIED (decl
) = 1;
11568 resolve_unique_section (decl
, 0, flag_function_sections
);
11569 allocate_struct_function (decl
, true);
11570 cfun
->is_thunk
= 1;
11571 current_function_decl
= decl
;
11572 init_varasm_status ();
11573 assemble_start_function (decl
, name
);
11577 const int align
= floor_log2 (FUNCTION_BOUNDARY
/ BITS_PER_UNIT
);
11578 switch_to_section (text_section
);
11580 ASM_OUTPUT_ALIGN (asm_out_file
, align
);
11581 ASM_OUTPUT_LABEL (asm_out_file
, name
);
11584 #ifdef DWARF2_UNWIND_INFO
11585 do_cfi
= dwarf2out_do_cfi_asm ();
11587 fprintf (asm_out_file
, "\t.cfi_startproc\n");
11589 if (flag_delayed_branch
)
11590 fprintf (asm_out_file
, "\tjmp\t%%o7+8\n\t add\t%%o7, %s, %s\n",
11591 reg_name
, reg_name
);
11593 fprintf (asm_out_file
, "\tadd\t%%o7, %s, %s\n\tjmp\t%%o7+8\n\t nop\n",
11594 reg_name
, reg_name
);
11595 #ifdef DWARF2_UNWIND_INFO
11597 fprintf (asm_out_file
, "\t.cfi_endproc\n");
11601 if (NEED_INDICATE_EXEC_STACK
)
11602 file_end_indicate_exec_stack ();
11604 #ifdef TARGET_SOLARIS
11605 solaris_file_end ();
11609 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
11610 /* Implement TARGET_MANGLE_TYPE. */
11612 static const char *
11613 sparc_mangle_type (const_tree type
)
11616 && TYPE_MAIN_VARIANT (type
) == long_double_type_node
11617 && TARGET_LONG_DOUBLE_128
)
11620 /* For all other types, use normal C++ mangling. */
11625 /* Expand a membar instruction for various use cases. Both the LOAD_STORE
11626 and BEFORE_AFTER arguments of the form X_Y. They are two-bit masks where
11627 bit 0 indicates that X is true, and bit 1 indicates Y is true. */
11630 sparc_emit_membar_for_model (enum memmodel model
,
11631 int load_store
, int before_after
)
11633 /* Bits for the MEMBAR mmask field. */
11634 const int LoadLoad
= 1;
11635 const int StoreLoad
= 2;
11636 const int LoadStore
= 4;
11637 const int StoreStore
= 8;
11639 int mm
= 0, implied
= 0;
11641 switch (sparc_memory_model
)
11644 /* Sequential Consistency. All memory transactions are immediately
11645 visible in sequential execution order. No barriers needed. */
11646 implied
= LoadLoad
| StoreLoad
| LoadStore
| StoreStore
;
11650 /* Total Store Ordering: all memory transactions with store semantics
11651 are followed by an implied StoreStore. */
11652 implied
|= StoreStore
;
11654 /* If we're not looking for a raw barrer (before+after), then atomic
11655 operations get the benefit of being both load and store. */
11656 if (load_store
== 3 && before_after
== 1)
11657 implied
|= StoreLoad
;
11661 /* Partial Store Ordering: all memory transactions with load semantics
11662 are followed by an implied LoadLoad | LoadStore. */
11663 implied
|= LoadLoad
| LoadStore
;
11665 /* If we're not looking for a raw barrer (before+after), then atomic
11666 operations get the benefit of being both load and store. */
11667 if (load_store
== 3 && before_after
== 2)
11668 implied
|= StoreLoad
| StoreStore
;
11672 /* Relaxed Memory Ordering: no implicit bits. */
11676 gcc_unreachable ();
11679 if (before_after
& 1)
11681 if (is_mm_release (model
) || is_mm_acq_rel (model
)
11682 || is_mm_seq_cst (model
))
11684 if (load_store
& 1)
11685 mm
|= LoadLoad
| StoreLoad
;
11686 if (load_store
& 2)
11687 mm
|= LoadStore
| StoreStore
;
11690 if (before_after
& 2)
11692 if (is_mm_acquire (model
) || is_mm_acq_rel (model
)
11693 || is_mm_seq_cst (model
))
11695 if (load_store
& 1)
11696 mm
|= LoadLoad
| LoadStore
;
11697 if (load_store
& 2)
11698 mm
|= StoreLoad
| StoreStore
;
11702 /* Remove the bits implied by the system memory model. */
11705 /* For raw barriers (before+after), always emit a barrier.
11706 This will become a compile-time barrier if needed. */
11707 if (mm
|| before_after
== 3)
11708 emit_insn (gen_membar (GEN_INT (mm
)));
11711 /* Expand code to perform a 8 or 16-bit compare and swap by doing 32-bit
11712 compare and swap on the word containing the byte or half-word. */
11715 sparc_expand_compare_and_swap_12 (rtx bool_result
, rtx result
, rtx mem
,
11716 rtx oldval
, rtx newval
)
11718 rtx addr1
= force_reg (Pmode
, XEXP (mem
, 0));
11719 rtx addr
= gen_reg_rtx (Pmode
);
11720 rtx off
= gen_reg_rtx (SImode
);
11721 rtx oldv
= gen_reg_rtx (SImode
);
11722 rtx newv
= gen_reg_rtx (SImode
);
11723 rtx oldvalue
= gen_reg_rtx (SImode
);
11724 rtx newvalue
= gen_reg_rtx (SImode
);
11725 rtx res
= gen_reg_rtx (SImode
);
11726 rtx resv
= gen_reg_rtx (SImode
);
11727 rtx memsi
, val
, mask
, cc
;
11729 emit_insn (gen_rtx_SET (addr
, gen_rtx_AND (Pmode
, addr1
, GEN_INT (-4))));
11731 if (Pmode
!= SImode
)
11732 addr1
= gen_lowpart (SImode
, addr1
);
11733 emit_insn (gen_rtx_SET (off
, gen_rtx_AND (SImode
, addr1
, GEN_INT (3))));
11735 memsi
= gen_rtx_MEM (SImode
, addr
);
11736 set_mem_alias_set (memsi
, ALIAS_SET_MEMORY_BARRIER
);
11737 MEM_VOLATILE_P (memsi
) = MEM_VOLATILE_P (mem
);
11739 val
= copy_to_reg (memsi
);
11741 emit_insn (gen_rtx_SET (off
,
11742 gen_rtx_XOR (SImode
, off
,
11743 GEN_INT (GET_MODE (mem
) == QImode
11746 emit_insn (gen_rtx_SET (off
, gen_rtx_ASHIFT (SImode
, off
, GEN_INT (3))));
11748 if (GET_MODE (mem
) == QImode
)
11749 mask
= force_reg (SImode
, GEN_INT (0xff));
11751 mask
= force_reg (SImode
, GEN_INT (0xffff));
11753 emit_insn (gen_rtx_SET (mask
, gen_rtx_ASHIFT (SImode
, mask
, off
)));
11755 emit_insn (gen_rtx_SET (val
,
11756 gen_rtx_AND (SImode
, gen_rtx_NOT (SImode
, mask
),
11759 oldval
= gen_lowpart (SImode
, oldval
);
11760 emit_insn (gen_rtx_SET (oldv
, gen_rtx_ASHIFT (SImode
, oldval
, off
)));
11762 newval
= gen_lowpart_common (SImode
, newval
);
11763 emit_insn (gen_rtx_SET (newv
, gen_rtx_ASHIFT (SImode
, newval
, off
)));
11765 emit_insn (gen_rtx_SET (oldv
, gen_rtx_AND (SImode
, oldv
, mask
)));
11767 emit_insn (gen_rtx_SET (newv
, gen_rtx_AND (SImode
, newv
, mask
)));
11769 rtx_code_label
*end_label
= gen_label_rtx ();
11770 rtx_code_label
*loop_label
= gen_label_rtx ();
11771 emit_label (loop_label
);
11773 emit_insn (gen_rtx_SET (oldvalue
, gen_rtx_IOR (SImode
, oldv
, val
)));
11775 emit_insn (gen_rtx_SET (newvalue
, gen_rtx_IOR (SImode
, newv
, val
)));
11777 emit_move_insn (bool_result
, const1_rtx
);
11779 emit_insn (gen_atomic_compare_and_swapsi_1 (res
, memsi
, oldvalue
, newvalue
));
11781 emit_cmp_and_jump_insns (res
, oldvalue
, EQ
, NULL
, SImode
, 0, end_label
);
11783 emit_insn (gen_rtx_SET (resv
,
11784 gen_rtx_AND (SImode
, gen_rtx_NOT (SImode
, mask
),
11787 emit_move_insn (bool_result
, const0_rtx
);
11789 cc
= gen_compare_reg_1 (NE
, resv
, val
);
11790 emit_insn (gen_rtx_SET (val
, resv
));
11792 /* Use cbranchcc4 to separate the compare and branch! */
11793 emit_jump_insn (gen_cbranchcc4 (gen_rtx_NE (VOIDmode
, cc
, const0_rtx
),
11794 cc
, const0_rtx
, loop_label
));
11796 emit_label (end_label
);
11798 emit_insn (gen_rtx_SET (res
, gen_rtx_AND (SImode
, res
, mask
)));
11800 emit_insn (gen_rtx_SET (res
, gen_rtx_LSHIFTRT (SImode
, res
, off
)));
11802 emit_move_insn (result
, gen_lowpart (GET_MODE (result
), res
));
11805 /* Expand code to perform a compare-and-swap. */
11808 sparc_expand_compare_and_swap (rtx operands
[])
11810 rtx bval
, retval
, mem
, oldval
, newval
;
11812 enum memmodel model
;
11814 bval
= operands
[0];
11815 retval
= operands
[1];
11817 oldval
= operands
[3];
11818 newval
= operands
[4];
11819 model
= (enum memmodel
) INTVAL (operands
[6]);
11820 mode
= GET_MODE (mem
);
11822 sparc_emit_membar_for_model (model
, 3, 1);
11824 if (reg_overlap_mentioned_p (retval
, oldval
))
11825 oldval
= copy_to_reg (oldval
);
11827 if (mode
== QImode
|| mode
== HImode
)
11828 sparc_expand_compare_and_swap_12 (bval
, retval
, mem
, oldval
, newval
);
11831 rtx (*gen
) (rtx
, rtx
, rtx
, rtx
);
11834 if (mode
== SImode
)
11835 gen
= gen_atomic_compare_and_swapsi_1
;
11837 gen
= gen_atomic_compare_and_swapdi_1
;
11838 emit_insn (gen (retval
, mem
, oldval
, newval
));
11840 x
= emit_store_flag (bval
, EQ
, retval
, oldval
, mode
, 1, 1);
11842 convert_move (bval
, x
, 1);
11845 sparc_emit_membar_for_model (model
, 3, 2);
11849 sparc_expand_vec_perm_bmask (machine_mode vmode
, rtx sel
)
11853 sel
= gen_lowpart (DImode
, sel
);
11857 /* inp = xxxxxxxAxxxxxxxB */
11858 t_1
= expand_simple_binop (DImode
, LSHIFTRT
, sel
, GEN_INT (16),
11859 NULL_RTX
, 1, OPTAB_DIRECT
);
11860 /* t_1 = ....xxxxxxxAxxx. */
11861 sel
= expand_simple_binop (SImode
, AND
, gen_lowpart (SImode
, sel
),
11862 GEN_INT (3), NULL_RTX
, 1, OPTAB_DIRECT
);
11863 t_1
= expand_simple_binop (SImode
, AND
, gen_lowpart (SImode
, t_1
),
11864 GEN_INT (0x30000), NULL_RTX
, 1, OPTAB_DIRECT
);
11865 /* sel = .......B */
11866 /* t_1 = ...A.... */
11867 sel
= expand_simple_binop (SImode
, IOR
, sel
, t_1
, sel
, 1, OPTAB_DIRECT
);
11868 /* sel = ...A...B */
11869 sel
= expand_mult (SImode
, sel
, GEN_INT (0x4444), sel
, 1);
11870 /* sel = AAAABBBB * 4 */
11871 t_1
= force_reg (SImode
, GEN_INT (0x01230123));
11872 /* sel = { A*4, A*4+1, A*4+2, ... } */
11876 /* inp = xxxAxxxBxxxCxxxD */
11877 t_1
= expand_simple_binop (DImode
, LSHIFTRT
, sel
, GEN_INT (8),
11878 NULL_RTX
, 1, OPTAB_DIRECT
);
11879 t_2
= expand_simple_binop (DImode
, LSHIFTRT
, sel
, GEN_INT (16),
11880 NULL_RTX
, 1, OPTAB_DIRECT
);
11881 t_3
= expand_simple_binop (DImode
, LSHIFTRT
, sel
, GEN_INT (24),
11882 NULL_RTX
, 1, OPTAB_DIRECT
);
11883 /* t_1 = ..xxxAxxxBxxxCxx */
11884 /* t_2 = ....xxxAxxxBxxxC */
11885 /* t_3 = ......xxxAxxxBxx */
11886 sel
= expand_simple_binop (SImode
, AND
, gen_lowpart (SImode
, sel
),
11888 NULL_RTX
, 1, OPTAB_DIRECT
);
11889 t_1
= expand_simple_binop (SImode
, AND
, gen_lowpart (SImode
, t_1
),
11891 NULL_RTX
, 1, OPTAB_DIRECT
);
11892 t_2
= expand_simple_binop (SImode
, AND
, gen_lowpart (SImode
, t_2
),
11893 GEN_INT (0x070000),
11894 NULL_RTX
, 1, OPTAB_DIRECT
);
11895 t_3
= expand_simple_binop (SImode
, AND
, gen_lowpart (SImode
, t_3
),
11896 GEN_INT (0x07000000),
11897 NULL_RTX
, 1, OPTAB_DIRECT
);
11898 /* sel = .......D */
11899 /* t_1 = .....C.. */
11900 /* t_2 = ...B.... */
11901 /* t_3 = .A...... */
11902 sel
= expand_simple_binop (SImode
, IOR
, sel
, t_1
, sel
, 1, OPTAB_DIRECT
);
11903 t_2
= expand_simple_binop (SImode
, IOR
, t_2
, t_3
, t_2
, 1, OPTAB_DIRECT
);
11904 sel
= expand_simple_binop (SImode
, IOR
, sel
, t_2
, sel
, 1, OPTAB_DIRECT
);
11905 /* sel = .A.B.C.D */
11906 sel
= expand_mult (SImode
, sel
, GEN_INT (0x22), sel
, 1);
11907 /* sel = AABBCCDD * 2 */
11908 t_1
= force_reg (SImode
, GEN_INT (0x01010101));
11909 /* sel = { A*2, A*2+1, B*2, B*2+1, ... } */
11913 /* input = xAxBxCxDxExFxGxH */
11914 sel
= expand_simple_binop (DImode
, AND
, sel
,
11915 GEN_INT ((HOST_WIDE_INT
)0x0f0f0f0f << 32
11917 NULL_RTX
, 1, OPTAB_DIRECT
);
11918 /* sel = .A.B.C.D.E.F.G.H */
11919 t_1
= expand_simple_binop (DImode
, LSHIFTRT
, sel
, GEN_INT (4),
11920 NULL_RTX
, 1, OPTAB_DIRECT
);
11921 /* t_1 = ..A.B.C.D.E.F.G. */
11922 sel
= expand_simple_binop (DImode
, IOR
, sel
, t_1
,
11923 NULL_RTX
, 1, OPTAB_DIRECT
);
11924 /* sel = .AABBCCDDEEFFGGH */
11925 sel
= expand_simple_binop (DImode
, AND
, sel
,
11926 GEN_INT ((HOST_WIDE_INT
)0xff00ff << 32
11928 NULL_RTX
, 1, OPTAB_DIRECT
);
11929 /* sel = ..AB..CD..EF..GH */
11930 t_1
= expand_simple_binop (DImode
, LSHIFTRT
, sel
, GEN_INT (8),
11931 NULL_RTX
, 1, OPTAB_DIRECT
);
11932 /* t_1 = ....AB..CD..EF.. */
11933 sel
= expand_simple_binop (DImode
, IOR
, sel
, t_1
,
11934 NULL_RTX
, 1, OPTAB_DIRECT
);
11935 /* sel = ..ABABCDCDEFEFGH */
11936 sel
= expand_simple_binop (DImode
, AND
, sel
,
11937 GEN_INT ((HOST_WIDE_INT
)0xffff << 32 | 0xffff),
11938 NULL_RTX
, 1, OPTAB_DIRECT
);
11939 /* sel = ....ABCD....EFGH */
11940 t_1
= expand_simple_binop (DImode
, LSHIFTRT
, sel
, GEN_INT (16),
11941 NULL_RTX
, 1, OPTAB_DIRECT
);
11942 /* t_1 = ........ABCD.... */
11943 sel
= gen_lowpart (SImode
, sel
);
11944 t_1
= gen_lowpart (SImode
, t_1
);
11948 gcc_unreachable ();
11951 /* Always perform the final addition/merge within the bmask insn. */
11952 emit_insn (gen_bmasksi_vis (gen_rtx_REG (SImode
, 0), sel
, t_1
));
11955 /* Implement TARGET_FRAME_POINTER_REQUIRED. */
11958 sparc_frame_pointer_required (void)
11960 /* If the stack pointer is dynamically modified in the function, it cannot
11961 serve as the frame pointer. */
11962 if (cfun
->calls_alloca
)
11965 /* If the function receives nonlocal gotos, it needs to save the frame
11966 pointer in the nonlocal_goto_save_area object. */
11967 if (cfun
->has_nonlocal_label
)
11970 /* In flat mode, that's it. */
11974 /* Otherwise, the frame pointer is required if the function isn't leaf. */
11975 return !(crtl
->is_leaf
&& only_leaf_regs_used ());
11978 /* The way this is structured, we can't eliminate SFP in favor of SP
11979 if the frame pointer is required: we want to use the SFP->HFP elimination
11980 in that case. But the test in update_eliminables doesn't know we are
11981 assuming below that we only do the former elimination. */
11984 sparc_can_eliminate (const int from ATTRIBUTE_UNUSED
, const int to
)
11986 return to
== HARD_FRAME_POINTER_REGNUM
|| !sparc_frame_pointer_required ();
11989 /* Return the hard frame pointer directly to bypass the stack bias. */
11992 sparc_builtin_setjmp_frame_value (void)
11994 return hard_frame_pointer_rtx
;
11997 /* If !TARGET_FPU, then make the fp registers and fp cc regs fixed so that
11998 they won't be allocated. */
12001 sparc_conditional_register_usage (void)
12003 if (PIC_OFFSET_TABLE_REGNUM
!= INVALID_REGNUM
)
12005 fixed_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
12006 call_used_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
12008 /* If the user has passed -f{fixed,call-{used,saved}}-g5 */
12009 /* then honor it. */
12010 if (TARGET_ARCH32
&& fixed_regs
[5])
12012 else if (TARGET_ARCH64
&& fixed_regs
[5] == 2)
12017 for (regno
= SPARC_FIRST_V9_FP_REG
;
12018 regno
<= SPARC_LAST_V9_FP_REG
;
12020 fixed_regs
[regno
] = 1;
12021 /* %fcc0 is used by v8 and v9. */
12022 for (regno
= SPARC_FIRST_V9_FCC_REG
+ 1;
12023 regno
<= SPARC_LAST_V9_FCC_REG
;
12025 fixed_regs
[regno
] = 1;
12030 for (regno
= 32; regno
< SPARC_LAST_V9_FCC_REG
; regno
++)
12031 fixed_regs
[regno
] = 1;
12033 /* If the user has passed -f{fixed,call-{used,saved}}-g2 */
12034 /* then honor it. Likewise with g3 and g4. */
12035 if (fixed_regs
[2] == 2)
12036 fixed_regs
[2] = ! TARGET_APP_REGS
;
12037 if (fixed_regs
[3] == 2)
12038 fixed_regs
[3] = ! TARGET_APP_REGS
;
12039 if (TARGET_ARCH32
&& fixed_regs
[4] == 2)
12040 fixed_regs
[4] = ! TARGET_APP_REGS
;
12041 else if (TARGET_CM_EMBMEDANY
)
12043 else if (fixed_regs
[4] == 2)
12048 /* Disable leaf functions. */
12049 memset (sparc_leaf_regs
, 0, FIRST_PSEUDO_REGISTER
);
12050 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
12051 leaf_reg_remap
[regno
] = regno
;
12054 global_regs
[SPARC_GSR_REG
] = 1;
12057 /* Implement TARGET_PREFERRED_RELOAD_CLASS:
12059 - We can't load constants into FP registers.
12060 - We can't load FP constants into integer registers when soft-float,
12061 because there is no soft-float pattern with a r/F constraint.
12062 - We can't load FP constants into integer registers for TFmode unless
12063 it is 0.0L, because there is no movtf pattern with a r/F constraint.
12064 - Try and reload integer constants (symbolic or otherwise) back into
12065 registers directly, rather than having them dumped to memory. */
12068 sparc_preferred_reload_class (rtx x
, reg_class_t rclass
)
12070 machine_mode mode
= GET_MODE (x
);
12071 if (CONSTANT_P (x
))
12073 if (FP_REG_CLASS_P (rclass
)
12074 || rclass
== GENERAL_OR_FP_REGS
12075 || rclass
== GENERAL_OR_EXTRA_FP_REGS
12076 || (GET_MODE_CLASS (mode
) == MODE_FLOAT
&& ! TARGET_FPU
)
12077 || (mode
== TFmode
&& ! const_zero_operand (x
, mode
)))
12080 if (GET_MODE_CLASS (mode
) == MODE_INT
)
12081 return GENERAL_REGS
;
12083 if (GET_MODE_CLASS (mode
) == MODE_VECTOR_INT
)
12085 if (! FP_REG_CLASS_P (rclass
)
12086 || !(const_zero_operand (x
, mode
)
12087 || const_all_ones_operand (x
, mode
)))
12094 && (rclass
== EXTRA_FP_REGS
12095 || rclass
== GENERAL_OR_EXTRA_FP_REGS
))
12097 int regno
= true_regnum (x
);
12099 if (SPARC_INT_REG_P (regno
))
12100 return (rclass
== EXTRA_FP_REGS
12101 ? FP_REGS
: GENERAL_OR_FP_REGS
);
12107 /* Output a wide multiply instruction in V8+ mode. INSN is the instruction,
12108 OPERANDS are its operands and OPCODE is the mnemonic to be used. */
12111 output_v8plus_mult (rtx_insn
*insn
, rtx
*operands
, const char *opcode
)
12115 gcc_assert (! TARGET_ARCH64
);
12117 if (sparc_check_64 (operands
[1], insn
) <= 0)
12118 output_asm_insn ("srl\t%L1, 0, %L1", operands
);
12119 if (which_alternative
== 1)
12120 output_asm_insn ("sllx\t%H1, 32, %H1", operands
);
12121 if (GET_CODE (operands
[2]) == CONST_INT
)
12123 if (which_alternative
== 1)
12125 output_asm_insn ("or\t%L1, %H1, %H1", operands
);
12126 sprintf (mulstr
, "%s\t%%H1, %%2, %%L0", opcode
);
12127 output_asm_insn (mulstr
, operands
);
12128 return "srlx\t%L0, 32, %H0";
12132 output_asm_insn ("sllx\t%H1, 32, %3", operands
);
12133 output_asm_insn ("or\t%L1, %3, %3", operands
);
12134 sprintf (mulstr
, "%s\t%%3, %%2, %%3", opcode
);
12135 output_asm_insn (mulstr
, operands
);
12136 output_asm_insn ("srlx\t%3, 32, %H0", operands
);
12137 return "mov\t%3, %L0";
12140 else if (rtx_equal_p (operands
[1], operands
[2]))
12142 if (which_alternative
== 1)
12144 output_asm_insn ("or\t%L1, %H1, %H1", operands
);
12145 sprintf (mulstr
, "%s\t%%H1, %%H1, %%L0", opcode
);
12146 output_asm_insn (mulstr
, operands
);
12147 return "srlx\t%L0, 32, %H0";
12151 output_asm_insn ("sllx\t%H1, 32, %3", operands
);
12152 output_asm_insn ("or\t%L1, %3, %3", operands
);
12153 sprintf (mulstr
, "%s\t%%3, %%3, %%3", opcode
);
12154 output_asm_insn (mulstr
, operands
);
12155 output_asm_insn ("srlx\t%3, 32, %H0", operands
);
12156 return "mov\t%3, %L0";
12159 if (sparc_check_64 (operands
[2], insn
) <= 0)
12160 output_asm_insn ("srl\t%L2, 0, %L2", operands
);
12161 if (which_alternative
== 1)
12163 output_asm_insn ("or\t%L1, %H1, %H1", operands
);
12164 output_asm_insn ("sllx\t%H2, 32, %L1", operands
);
12165 output_asm_insn ("or\t%L2, %L1, %L1", operands
);
12166 sprintf (mulstr
, "%s\t%%H1, %%L1, %%L0", opcode
);
12167 output_asm_insn (mulstr
, operands
);
12168 return "srlx\t%L0, 32, %H0";
12172 output_asm_insn ("sllx\t%H1, 32, %3", operands
);
12173 output_asm_insn ("sllx\t%H2, 32, %4", operands
);
12174 output_asm_insn ("or\t%L1, %3, %3", operands
);
12175 output_asm_insn ("or\t%L2, %4, %4", operands
);
12176 sprintf (mulstr
, "%s\t%%3, %%4, %%3", opcode
);
12177 output_asm_insn (mulstr
, operands
);
12178 output_asm_insn ("srlx\t%3, 32, %H0", operands
);
12179 return "mov\t%3, %L0";
12183 /* Subroutine of sparc_expand_vector_init. Emit code to initialize
12184 all fields of TARGET to ELT by means of VIS2 BSHUFFLE insn. MODE
12185 and INNER_MODE are the modes describing TARGET. */
12188 vector_init_bshuffle (rtx target
, rtx elt
, machine_mode mode
,
12189 machine_mode inner_mode
)
12191 rtx t1
, final_insn
, sel
;
12194 t1
= gen_reg_rtx (mode
);
12196 elt
= convert_modes (SImode
, inner_mode
, elt
, true);
12197 emit_move_insn (gen_lowpart(SImode
, t1
), elt
);
12202 final_insn
= gen_bshufflev2si_vis (target
, t1
, t1
);
12203 bmask
= 0x45674567;
12206 final_insn
= gen_bshufflev4hi_vis (target
, t1
, t1
);
12207 bmask
= 0x67676767;
12210 final_insn
= gen_bshufflev8qi_vis (target
, t1
, t1
);
12211 bmask
= 0x77777777;
12214 gcc_unreachable ();
12217 sel
= force_reg (SImode
, GEN_INT (bmask
));
12218 emit_insn (gen_bmasksi_vis (gen_rtx_REG (SImode
, 0), sel
, const0_rtx
));
12219 emit_insn (final_insn
);
12222 /* Subroutine of sparc_expand_vector_init. Emit code to initialize
12223 all fields of TARGET to ELT in V8QI by means of VIS FPMERGE insn. */
12226 vector_init_fpmerge (rtx target
, rtx elt
)
12228 rtx t1
, t2
, t2_low
, t3
, t3_low
;
12230 t1
= gen_reg_rtx (V4QImode
);
12231 elt
= convert_modes (SImode
, QImode
, elt
, true);
12232 emit_move_insn (gen_lowpart (SImode
, t1
), elt
);
12234 t2
= gen_reg_rtx (V8QImode
);
12235 t2_low
= gen_lowpart (V4QImode
, t2
);
12236 emit_insn (gen_fpmerge_vis (t2
, t1
, t1
));
12238 t3
= gen_reg_rtx (V8QImode
);
12239 t3_low
= gen_lowpart (V4QImode
, t3
);
12240 emit_insn (gen_fpmerge_vis (t3
, t2_low
, t2_low
));
12242 emit_insn (gen_fpmerge_vis (target
, t3_low
, t3_low
));
12245 /* Subroutine of sparc_expand_vector_init. Emit code to initialize
12246 all fields of TARGET to ELT in V4HI by means of VIS FALIGNDATA insn. */
12249 vector_init_faligndata (rtx target
, rtx elt
)
12251 rtx t1
= gen_reg_rtx (V4HImode
);
12254 elt
= convert_modes (SImode
, HImode
, elt
, true);
12255 emit_move_insn (gen_lowpart (SImode
, t1
), elt
);
12257 emit_insn (gen_alignaddrsi_vis (gen_reg_rtx (SImode
),
12258 force_reg (SImode
, GEN_INT (6)),
12261 for (i
= 0; i
< 4; i
++)
12262 emit_insn (gen_faligndatav4hi_vis (target
, t1
, target
));
12265 /* Emit code to initialize TARGET to values for individual fields VALS. */
12268 sparc_expand_vector_init (rtx target
, rtx vals
)
12270 const machine_mode mode
= GET_MODE (target
);
12271 const machine_mode inner_mode
= GET_MODE_INNER (mode
);
12272 const int n_elts
= GET_MODE_NUNITS (mode
);
12278 for (i
= 0; i
< n_elts
; i
++)
12280 rtx x
= XVECEXP (vals
, 0, i
);
12281 if (!CONSTANT_P (x
))
12284 if (i
> 0 && !rtx_equal_p (x
, XVECEXP (vals
, 0, 0)))
12290 emit_move_insn (target
, gen_rtx_CONST_VECTOR (mode
, XVEC (vals
, 0)));
12294 if (GET_MODE_SIZE (inner_mode
) == GET_MODE_SIZE (mode
))
12296 if (GET_MODE_SIZE (inner_mode
) == 4)
12298 emit_move_insn (gen_lowpart (SImode
, target
),
12299 gen_lowpart (SImode
, XVECEXP (vals
, 0, 0)));
12302 else if (GET_MODE_SIZE (inner_mode
) == 8)
12304 emit_move_insn (gen_lowpart (DImode
, target
),
12305 gen_lowpart (DImode
, XVECEXP (vals
, 0, 0)));
12309 else if (GET_MODE_SIZE (inner_mode
) == GET_MODE_SIZE (word_mode
)
12310 && GET_MODE_SIZE (mode
) == 2 * GET_MODE_SIZE (word_mode
))
12312 emit_move_insn (gen_highpart (word_mode
, target
),
12313 gen_lowpart (word_mode
, XVECEXP (vals
, 0, 0)));
12314 emit_move_insn (gen_lowpart (word_mode
, target
),
12315 gen_lowpart (word_mode
, XVECEXP (vals
, 0, 1)));
12319 if (all_same
&& GET_MODE_SIZE (mode
) == 8)
12323 vector_init_bshuffle (target
, XVECEXP (vals
, 0, 0), mode
, inner_mode
);
12326 if (mode
== V8QImode
)
12328 vector_init_fpmerge (target
, XVECEXP (vals
, 0, 0));
12331 if (mode
== V4HImode
)
12333 vector_init_faligndata (target
, XVECEXP (vals
, 0, 0));
12338 mem
= assign_stack_temp (mode
, GET_MODE_SIZE (mode
));
12339 for (i
= 0; i
< n_elts
; i
++)
12340 emit_move_insn (adjust_address_nv (mem
, inner_mode
,
12341 i
* GET_MODE_SIZE (inner_mode
)),
12342 XVECEXP (vals
, 0, i
));
12343 emit_move_insn (target
, mem
);
12346 /* Implement TARGET_SECONDARY_RELOAD. */
12349 sparc_secondary_reload (bool in_p
, rtx x
, reg_class_t rclass_i
,
12350 machine_mode mode
, secondary_reload_info
*sri
)
12352 enum reg_class rclass
= (enum reg_class
) rclass_i
;
12354 sri
->icode
= CODE_FOR_nothing
;
12355 sri
->extra_cost
= 0;
12357 /* We need a temporary when loading/storing a HImode/QImode value
12358 between memory and the FPU registers. This can happen when combine puts
12359 a paradoxical subreg in a float/fix conversion insn. */
12360 if (FP_REG_CLASS_P (rclass
)
12361 && (mode
== HImode
|| mode
== QImode
)
12362 && (GET_CODE (x
) == MEM
12363 || ((GET_CODE (x
) == REG
|| GET_CODE (x
) == SUBREG
)
12364 && true_regnum (x
) == -1)))
12365 return GENERAL_REGS
;
12367 /* On 32-bit we need a temporary when loading/storing a DFmode value
12368 between unaligned memory and the upper FPU registers. */
12370 && rclass
== EXTRA_FP_REGS
12372 && GET_CODE (x
) == MEM
12373 && ! mem_min_alignment (x
, 8))
12376 if (((TARGET_CM_MEDANY
12377 && symbolic_operand (x
, mode
))
12378 || (TARGET_CM_EMBMEDANY
12379 && text_segment_operand (x
, mode
)))
12383 sri
->icode
= direct_optab_handler (reload_in_optab
, mode
);
12385 sri
->icode
= direct_optab_handler (reload_out_optab
, mode
);
12389 if (TARGET_VIS3
&& TARGET_ARCH32
)
12391 int regno
= true_regnum (x
);
12393 /* When using VIS3 fp<-->int register moves, on 32-bit we have
12394 to move 8-byte values in 4-byte pieces. This only works via
12395 FP_REGS, and not via EXTRA_FP_REGS. Therefore if we try to
12396 move between EXTRA_FP_REGS and GENERAL_REGS, we will need
12397 an FP_REGS intermediate move. */
12398 if ((rclass
== EXTRA_FP_REGS
&& SPARC_INT_REG_P (regno
))
12399 || ((general_or_i64_p (rclass
)
12400 || rclass
== GENERAL_OR_FP_REGS
)
12401 && SPARC_FP_REG_P (regno
)))
12403 sri
->extra_cost
= 2;
12411 /* Emit code to conditionally move either OPERANDS[2] or OPERANDS[3] into
12412 OPERANDS[0] in MODE. OPERANDS[1] is the operator of the condition. */
12415 sparc_expand_conditional_move (machine_mode mode
, rtx
*operands
)
12417 enum rtx_code rc
= GET_CODE (operands
[1]);
12418 machine_mode cmp_mode
;
12419 rtx cc_reg
, dst
, cmp
;
12422 if (GET_MODE (XEXP (cmp
, 0)) == DImode
&& !TARGET_ARCH64
)
12425 if (GET_MODE (XEXP (cmp
, 0)) == TFmode
&& !TARGET_HARD_QUAD
)
12426 cmp
= sparc_emit_float_lib_cmp (XEXP (cmp
, 0), XEXP (cmp
, 1), rc
);
12428 cmp_mode
= GET_MODE (XEXP (cmp
, 0));
12429 rc
= GET_CODE (cmp
);
12432 if (! rtx_equal_p (operands
[2], dst
)
12433 && ! rtx_equal_p (operands
[3], dst
))
12435 if (reg_overlap_mentioned_p (dst
, cmp
))
12436 dst
= gen_reg_rtx (mode
);
12438 emit_move_insn (dst
, operands
[3]);
12440 else if (operands
[2] == dst
)
12442 operands
[2] = operands
[3];
12444 if (GET_MODE_CLASS (cmp_mode
) == MODE_FLOAT
)
12445 rc
= reverse_condition_maybe_unordered (rc
);
12447 rc
= reverse_condition (rc
);
12450 if (XEXP (cmp
, 1) == const0_rtx
12451 && GET_CODE (XEXP (cmp
, 0)) == REG
12452 && cmp_mode
== DImode
12453 && v9_regcmp_p (rc
))
12454 cc_reg
= XEXP (cmp
, 0);
12456 cc_reg
= gen_compare_reg_1 (rc
, XEXP (cmp
, 0), XEXP (cmp
, 1));
12458 cmp
= gen_rtx_fmt_ee (rc
, GET_MODE (cc_reg
), cc_reg
, const0_rtx
);
12460 emit_insn (gen_rtx_SET (dst
,
12461 gen_rtx_IF_THEN_ELSE (mode
, cmp
, operands
[2], dst
)));
12463 if (dst
!= operands
[0])
12464 emit_move_insn (operands
[0], dst
);
12469 /* Emit code to conditionally move a combination of OPERANDS[1] and OPERANDS[2]
12470 into OPERANDS[0] in MODE, depending on the outcome of the comparison of
12471 OPERANDS[4] and OPERANDS[5]. OPERANDS[3] is the operator of the condition.
12472 FCODE is the machine code to be used for OPERANDS[3] and CCODE the machine
12473 code to be used for the condition mask. */
12476 sparc_expand_vcond (machine_mode mode
, rtx
*operands
, int ccode
, int fcode
)
12478 rtx mask
, cop0
, cop1
, fcmp
, cmask
, bshuf
, gsr
;
12479 enum rtx_code code
= GET_CODE (operands
[3]);
12481 mask
= gen_reg_rtx (Pmode
);
12482 cop0
= operands
[4];
12483 cop1
= operands
[5];
12484 if (code
== LT
|| code
== GE
)
12488 code
= swap_condition (code
);
12489 t
= cop0
; cop0
= cop1
; cop1
= t
;
12492 gsr
= gen_rtx_REG (DImode
, SPARC_GSR_REG
);
12494 fcmp
= gen_rtx_UNSPEC (Pmode
,
12495 gen_rtvec (1, gen_rtx_fmt_ee (code
, mode
, cop0
, cop1
)),
12498 cmask
= gen_rtx_UNSPEC (DImode
,
12499 gen_rtvec (2, mask
, gsr
),
12502 bshuf
= gen_rtx_UNSPEC (mode
,
12503 gen_rtvec (3, operands
[1], operands
[2], gsr
),
12506 emit_insn (gen_rtx_SET (mask
, fcmp
));
12507 emit_insn (gen_rtx_SET (gsr
, cmask
));
12509 emit_insn (gen_rtx_SET (operands
[0], bshuf
));
12512 /* On sparc, any mode which naturally allocates into the float
12513 registers should return 4 here. */
12516 sparc_regmode_natural_size (machine_mode mode
)
12518 int size
= UNITS_PER_WORD
;
12522 enum mode_class mclass
= GET_MODE_CLASS (mode
);
12524 if (mclass
== MODE_FLOAT
|| mclass
== MODE_VECTOR_INT
)
12531 /* Return TRUE if it is a good idea to tie two pseudo registers
12532 when one has mode MODE1 and one has mode MODE2.
12533 If HARD_REGNO_MODE_OK could produce different values for MODE1 and MODE2,
12534 for any hard reg, then this must be FALSE for correct output.
12536 For V9 we have to deal with the fact that only the lower 32 floating
12537 point registers are 32-bit addressable. */
12540 sparc_modes_tieable_p (machine_mode mode1
, machine_mode mode2
)
12542 enum mode_class mclass1
, mclass2
;
12543 unsigned short size1
, size2
;
12545 if (mode1
== mode2
)
12548 mclass1
= GET_MODE_CLASS (mode1
);
12549 mclass2
= GET_MODE_CLASS (mode2
);
12550 if (mclass1
!= mclass2
)
12556 /* Classes are the same and we are V9 so we have to deal with upper
12557 vs. lower floating point registers. If one of the modes is a
12558 4-byte mode, and the other is not, we have to mark them as not
12559 tieable because only the lower 32 floating point register are
12560 addressable 32-bits at a time.
12562 We can't just test explicitly for SFmode, otherwise we won't
12563 cover the vector mode cases properly. */
12565 if (mclass1
!= MODE_FLOAT
&& mclass1
!= MODE_VECTOR_INT
)
12568 size1
= GET_MODE_SIZE (mode1
);
12569 size2
= GET_MODE_SIZE (mode2
);
12570 if ((size1
> 4 && size2
== 4)
12571 || (size2
> 4 && size1
== 4))
12577 /* Implement TARGET_CSTORE_MODE. */
12579 static machine_mode
12580 sparc_cstore_mode (enum insn_code icode ATTRIBUTE_UNUSED
)
12582 return (TARGET_ARCH64
? DImode
: SImode
);
12585 /* Return the compound expression made of T1 and T2. */
12588 compound_expr (tree t1
, tree t2
)
12590 return build2 (COMPOUND_EXPR
, void_type_node
, t1
, t2
);
12593 /* Implement TARGET_ATOMIC_ASSIGN_EXPAND_FENV hook. */
12596 sparc_atomic_assign_expand_fenv (tree
*hold
, tree
*clear
, tree
*update
)
12601 const unsigned HOST_WIDE_INT accrued_exception_mask
= 0x1f << 5;
12602 const unsigned HOST_WIDE_INT trap_enable_mask
= 0x1f << 23;
12604 /* We generate the equivalent of feholdexcept (&fenv_var):
12606 unsigned int fenv_var;
12607 __builtin_store_fsr (&fenv_var);
12609 unsigned int tmp1_var;
12610 tmp1_var = fenv_var & ~(accrued_exception_mask | trap_enable_mask);
12612 __builtin_load_fsr (&tmp1_var); */
12614 tree fenv_var
= create_tmp_var_raw (unsigned_type_node
);
12615 TREE_ADDRESSABLE (fenv_var
) = 1;
12616 tree fenv_addr
= build_fold_addr_expr (fenv_var
);
12617 tree stfsr
= sparc_builtins
[SPARC_BUILTIN_STFSR
];
12619 = build4 (TARGET_EXPR
, unsigned_type_node
, fenv_var
,
12620 build_call_expr (stfsr
, 1, fenv_addr
), NULL_TREE
, NULL_TREE
);
12622 tree tmp1_var
= create_tmp_var_raw (unsigned_type_node
);
12623 TREE_ADDRESSABLE (tmp1_var
) = 1;
12624 tree masked_fenv_var
12625 = build2 (BIT_AND_EXPR
, unsigned_type_node
, fenv_var
,
12626 build_int_cst (unsigned_type_node
,
12627 ~(accrued_exception_mask
| trap_enable_mask
)));
12629 = build4 (TARGET_EXPR
, unsigned_type_node
, tmp1_var
, masked_fenv_var
,
12630 NULL_TREE
, NULL_TREE
);
12632 tree tmp1_addr
= build_fold_addr_expr (tmp1_var
);
12633 tree ldfsr
= sparc_builtins
[SPARC_BUILTIN_LDFSR
];
12634 tree hold_ldfsr
= build_call_expr (ldfsr
, 1, tmp1_addr
);
12636 *hold
= compound_expr (compound_expr (hold_stfsr
, hold_mask
), hold_ldfsr
);
12638 /* We reload the value of tmp1_var to clear the exceptions:
12640 __builtin_load_fsr (&tmp1_var); */
12642 *clear
= build_call_expr (ldfsr
, 1, tmp1_addr
);
12644 /* We generate the equivalent of feupdateenv (&fenv_var):
12646 unsigned int tmp2_var;
12647 __builtin_store_fsr (&tmp2_var);
12649 __builtin_load_fsr (&fenv_var);
12651 if (SPARC_LOW_FE_EXCEPT_VALUES)
12653 __atomic_feraiseexcept ((int) tmp2_var); */
12655 tree tmp2_var
= create_tmp_var_raw (unsigned_type_node
);
12656 TREE_ADDRESSABLE (tmp2_var
) = 1;
12657 tree tmp2_addr
= build_fold_addr_expr (tmp2_var
);
12659 = build4 (TARGET_EXPR
, unsigned_type_node
, tmp2_var
,
12660 build_call_expr (stfsr
, 1, tmp2_addr
), NULL_TREE
, NULL_TREE
);
12662 tree update_ldfsr
= build_call_expr (ldfsr
, 1, fenv_addr
);
12664 tree atomic_feraiseexcept
12665 = builtin_decl_implicit (BUILT_IN_ATOMIC_FERAISEEXCEPT
);
12667 = build_call_expr (atomic_feraiseexcept
, 1,
12668 fold_convert (integer_type_node
, tmp2_var
));
12670 if (SPARC_LOW_FE_EXCEPT_VALUES
)
12672 tree shifted_tmp2_var
12673 = build2 (RSHIFT_EXPR
, unsigned_type_node
, tmp2_var
,
12674 build_int_cst (unsigned_type_node
, 5));
12676 = build2 (MODIFY_EXPR
, void_type_node
, tmp2_var
, shifted_tmp2_var
);
12677 update_call
= compound_expr (update_shift
, update_call
);
12681 = compound_expr (compound_expr (update_stfsr
, update_ldfsr
), update_call
);
12684 #include "gt-sparc.h"