1 /* Subroutines for insn-output.c for SPARC.
2 Copyright (C) 1987, 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998,
3 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010,
5 Free Software Foundation, Inc.
6 Contributed by Michael Tiemann (tiemann@cygnus.com)
7 64-bit SPARC-V9 support by Michael Tiemann, Jim Wilson, and Doug Evans,
10 This file is part of GCC.
12 GCC is free software; you can redistribute it and/or modify
13 it under the terms of the GNU General Public License as published by
14 the Free Software Foundation; either version 3, or (at your option)
17 GCC is distributed in the hope that it will be useful,
18 but WITHOUT ANY WARRANTY; without even the implied warranty of
19 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
20 GNU General Public License for more details.
22 You should have received a copy of the GNU General Public License
23 along with GCC; see the file COPYING3. If not see
24 <http://www.gnu.org/licenses/>. */
28 #include "coretypes.h"
33 #include "hard-reg-set.h"
34 #include "insn-config.h"
35 #include "insn-codes.h"
36 #include "conditions.h"
38 #include "insn-attr.h"
45 #include "diagnostic-core.h"
50 #include "target-def.h"
51 #include "common/common-target.h"
53 #include "langhooks.h"
61 struct processor_costs {
65 /* Integer signed load */
68 /* Integer zeroed load */
74 /* fmov, fneg, fabs */
78 const int float_plusminus;
84 const int float_cmove;
90 const int float_div_sf;
93 const int float_div_df;
96 const int float_sqrt_sf;
99 const int float_sqrt_df;
107 /* integer multiply cost for each bit set past the most
108 significant 3, so the formula for multiply cost becomes:
111 highest_bit = highest_clear_bit(rs1);
113 highest_bit = highest_set_bit(rs1);
116 cost = int_mul{,X} + ((highest_bit - 3) / int_mul_bit_factor);
118 A value of zero indicates that the multiply costs is fixed,
120 const int int_mul_bit_factor;
131 /* penalty for shifts, due to scheduling rules etc. */
132 const int shift_penalty;
136 struct processor_costs cypress_costs = {
137 COSTS_N_INSNS (2), /* int load */
138 COSTS_N_INSNS (2), /* int signed load */
139 COSTS_N_INSNS (2), /* int zeroed load */
140 COSTS_N_INSNS (2), /* float load */
141 COSTS_N_INSNS (5), /* fmov, fneg, fabs */
142 COSTS_N_INSNS (5), /* fadd, fsub */
143 COSTS_N_INSNS (1), /* fcmp */
144 COSTS_N_INSNS (1), /* fmov, fmovr */
145 COSTS_N_INSNS (7), /* fmul */
146 COSTS_N_INSNS (37), /* fdivs */
147 COSTS_N_INSNS (37), /* fdivd */
148 COSTS_N_INSNS (63), /* fsqrts */
149 COSTS_N_INSNS (63), /* fsqrtd */
150 COSTS_N_INSNS (1), /* imul */
151 COSTS_N_INSNS (1), /* imulX */
152 0, /* imul bit factor */
153 COSTS_N_INSNS (1), /* idiv */
154 COSTS_N_INSNS (1), /* idivX */
155 COSTS_N_INSNS (1), /* movcc/movr */
156 0, /* shift penalty */
160 struct processor_costs supersparc_costs = {
161 COSTS_N_INSNS (1), /* int load */
162 COSTS_N_INSNS (1), /* int signed load */
163 COSTS_N_INSNS (1), /* int zeroed load */
164 COSTS_N_INSNS (0), /* float load */
165 COSTS_N_INSNS (3), /* fmov, fneg, fabs */
166 COSTS_N_INSNS (3), /* fadd, fsub */
167 COSTS_N_INSNS (3), /* fcmp */
168 COSTS_N_INSNS (1), /* fmov, fmovr */
169 COSTS_N_INSNS (3), /* fmul */
170 COSTS_N_INSNS (6), /* fdivs */
171 COSTS_N_INSNS (9), /* fdivd */
172 COSTS_N_INSNS (12), /* fsqrts */
173 COSTS_N_INSNS (12), /* fsqrtd */
174 COSTS_N_INSNS (4), /* imul */
175 COSTS_N_INSNS (4), /* imulX */
176 0, /* imul bit factor */
177 COSTS_N_INSNS (4), /* idiv */
178 COSTS_N_INSNS (4), /* idivX */
179 COSTS_N_INSNS (1), /* movcc/movr */
180 1, /* shift penalty */
184 struct processor_costs hypersparc_costs = {
185 COSTS_N_INSNS (1), /* int load */
186 COSTS_N_INSNS (1), /* int signed load */
187 COSTS_N_INSNS (1), /* int zeroed load */
188 COSTS_N_INSNS (1), /* float load */
189 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
190 COSTS_N_INSNS (1), /* fadd, fsub */
191 COSTS_N_INSNS (1), /* fcmp */
192 COSTS_N_INSNS (1), /* fmov, fmovr */
193 COSTS_N_INSNS (1), /* fmul */
194 COSTS_N_INSNS (8), /* fdivs */
195 COSTS_N_INSNS (12), /* fdivd */
196 COSTS_N_INSNS (17), /* fsqrts */
197 COSTS_N_INSNS (17), /* fsqrtd */
198 COSTS_N_INSNS (17), /* imul */
199 COSTS_N_INSNS (17), /* imulX */
200 0, /* imul bit factor */
201 COSTS_N_INSNS (17), /* idiv */
202 COSTS_N_INSNS (17), /* idivX */
203 COSTS_N_INSNS (1), /* movcc/movr */
204 0, /* shift penalty */
208 struct processor_costs leon_costs = {
209 COSTS_N_INSNS (1), /* int load */
210 COSTS_N_INSNS (1), /* int signed load */
211 COSTS_N_INSNS (1), /* int zeroed load */
212 COSTS_N_INSNS (1), /* float load */
213 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
214 COSTS_N_INSNS (1), /* fadd, fsub */
215 COSTS_N_INSNS (1), /* fcmp */
216 COSTS_N_INSNS (1), /* fmov, fmovr */
217 COSTS_N_INSNS (1), /* fmul */
218 COSTS_N_INSNS (15), /* fdivs */
219 COSTS_N_INSNS (15), /* fdivd */
220 COSTS_N_INSNS (23), /* fsqrts */
221 COSTS_N_INSNS (23), /* fsqrtd */
222 COSTS_N_INSNS (5), /* imul */
223 COSTS_N_INSNS (5), /* imulX */
224 0, /* imul bit factor */
225 COSTS_N_INSNS (5), /* idiv */
226 COSTS_N_INSNS (5), /* idivX */
227 COSTS_N_INSNS (1), /* movcc/movr */
228 0, /* shift penalty */
232 struct processor_costs sparclet_costs = {
233 COSTS_N_INSNS (3), /* int load */
234 COSTS_N_INSNS (3), /* int signed load */
235 COSTS_N_INSNS (1), /* int zeroed load */
236 COSTS_N_INSNS (1), /* float load */
237 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
238 COSTS_N_INSNS (1), /* fadd, fsub */
239 COSTS_N_INSNS (1), /* fcmp */
240 COSTS_N_INSNS (1), /* fmov, fmovr */
241 COSTS_N_INSNS (1), /* fmul */
242 COSTS_N_INSNS (1), /* fdivs */
243 COSTS_N_INSNS (1), /* fdivd */
244 COSTS_N_INSNS (1), /* fsqrts */
245 COSTS_N_INSNS (1), /* fsqrtd */
246 COSTS_N_INSNS (5), /* imul */
247 COSTS_N_INSNS (5), /* imulX */
248 0, /* imul bit factor */
249 COSTS_N_INSNS (5), /* idiv */
250 COSTS_N_INSNS (5), /* idivX */
251 COSTS_N_INSNS (1), /* movcc/movr */
252 0, /* shift penalty */
256 struct processor_costs ultrasparc_costs = {
257 COSTS_N_INSNS (2), /* int load */
258 COSTS_N_INSNS (3), /* int signed load */
259 COSTS_N_INSNS (2), /* int zeroed load */
260 COSTS_N_INSNS (2), /* float load */
261 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
262 COSTS_N_INSNS (4), /* fadd, fsub */
263 COSTS_N_INSNS (1), /* fcmp */
264 COSTS_N_INSNS (2), /* fmov, fmovr */
265 COSTS_N_INSNS (4), /* fmul */
266 COSTS_N_INSNS (13), /* fdivs */
267 COSTS_N_INSNS (23), /* fdivd */
268 COSTS_N_INSNS (13), /* fsqrts */
269 COSTS_N_INSNS (23), /* fsqrtd */
270 COSTS_N_INSNS (4), /* imul */
271 COSTS_N_INSNS (4), /* imulX */
272 2, /* imul bit factor */
273 COSTS_N_INSNS (37), /* idiv */
274 COSTS_N_INSNS (68), /* idivX */
275 COSTS_N_INSNS (2), /* movcc/movr */
276 2, /* shift penalty */
280 struct processor_costs ultrasparc3_costs = {
281 COSTS_N_INSNS (2), /* int load */
282 COSTS_N_INSNS (3), /* int signed load */
283 COSTS_N_INSNS (3), /* int zeroed load */
284 COSTS_N_INSNS (2), /* float load */
285 COSTS_N_INSNS (3), /* fmov, fneg, fabs */
286 COSTS_N_INSNS (4), /* fadd, fsub */
287 COSTS_N_INSNS (5), /* fcmp */
288 COSTS_N_INSNS (3), /* fmov, fmovr */
289 COSTS_N_INSNS (4), /* fmul */
290 COSTS_N_INSNS (17), /* fdivs */
291 COSTS_N_INSNS (20), /* fdivd */
292 COSTS_N_INSNS (20), /* fsqrts */
293 COSTS_N_INSNS (29), /* fsqrtd */
294 COSTS_N_INSNS (6), /* imul */
295 COSTS_N_INSNS (6), /* imulX */
296 0, /* imul bit factor */
297 COSTS_N_INSNS (40), /* idiv */
298 COSTS_N_INSNS (71), /* idivX */
299 COSTS_N_INSNS (2), /* movcc/movr */
300 0, /* shift penalty */
304 struct processor_costs niagara_costs = {
305 COSTS_N_INSNS (3), /* int load */
306 COSTS_N_INSNS (3), /* int signed load */
307 COSTS_N_INSNS (3), /* int zeroed load */
308 COSTS_N_INSNS (9), /* float load */
309 COSTS_N_INSNS (8), /* fmov, fneg, fabs */
310 COSTS_N_INSNS (8), /* fadd, fsub */
311 COSTS_N_INSNS (26), /* fcmp */
312 COSTS_N_INSNS (8), /* fmov, fmovr */
313 COSTS_N_INSNS (29), /* fmul */
314 COSTS_N_INSNS (54), /* fdivs */
315 COSTS_N_INSNS (83), /* fdivd */
316 COSTS_N_INSNS (100), /* fsqrts - not implemented in hardware */
317 COSTS_N_INSNS (100), /* fsqrtd - not implemented in hardware */
318 COSTS_N_INSNS (11), /* imul */
319 COSTS_N_INSNS (11), /* imulX */
320 0, /* imul bit factor */
321 COSTS_N_INSNS (72), /* idiv */
322 COSTS_N_INSNS (72), /* idivX */
323 COSTS_N_INSNS (1), /* movcc/movr */
324 0, /* shift penalty */
328 struct processor_costs niagara2_costs = {
329 COSTS_N_INSNS (3), /* int load */
330 COSTS_N_INSNS (3), /* int signed load */
331 COSTS_N_INSNS (3), /* int zeroed load */
332 COSTS_N_INSNS (3), /* float load */
333 COSTS_N_INSNS (6), /* fmov, fneg, fabs */
334 COSTS_N_INSNS (6), /* fadd, fsub */
335 COSTS_N_INSNS (6), /* fcmp */
336 COSTS_N_INSNS (6), /* fmov, fmovr */
337 COSTS_N_INSNS (6), /* fmul */
338 COSTS_N_INSNS (19), /* fdivs */
339 COSTS_N_INSNS (33), /* fdivd */
340 COSTS_N_INSNS (19), /* fsqrts */
341 COSTS_N_INSNS (33), /* fsqrtd */
342 COSTS_N_INSNS (5), /* imul */
343 COSTS_N_INSNS (5), /* imulX */
344 0, /* imul bit factor */
345 COSTS_N_INSNS (26), /* idiv, average of 12 - 41 cycle range */
346 COSTS_N_INSNS (26), /* idivX, average of 12 - 41 cycle range */
347 COSTS_N_INSNS (1), /* movcc/movr */
348 0, /* shift penalty */
352 struct processor_costs niagara3_costs = {
353 COSTS_N_INSNS (3), /* int load */
354 COSTS_N_INSNS (3), /* int signed load */
355 COSTS_N_INSNS (3), /* int zeroed load */
356 COSTS_N_INSNS (3), /* float load */
357 COSTS_N_INSNS (9), /* fmov, fneg, fabs */
358 COSTS_N_INSNS (9), /* fadd, fsub */
359 COSTS_N_INSNS (9), /* fcmp */
360 COSTS_N_INSNS (9), /* fmov, fmovr */
361 COSTS_N_INSNS (9), /* fmul */
362 COSTS_N_INSNS (23), /* fdivs */
363 COSTS_N_INSNS (37), /* fdivd */
364 COSTS_N_INSNS (23), /* fsqrts */
365 COSTS_N_INSNS (37), /* fsqrtd */
366 COSTS_N_INSNS (9), /* imul */
367 COSTS_N_INSNS (9), /* imulX */
368 0, /* imul bit factor */
369 COSTS_N_INSNS (31), /* idiv, average of 17 - 45 cycle range */
370 COSTS_N_INSNS (30), /* idivX, average of 16 - 44 cycle range */
371 COSTS_N_INSNS (1), /* movcc/movr */
372 0, /* shift penalty */
376 struct processor_costs niagara4_costs = {
377 COSTS_N_INSNS (5), /* int load */
378 COSTS_N_INSNS (5), /* int signed load */
379 COSTS_N_INSNS (5), /* int zeroed load */
380 COSTS_N_INSNS (5), /* float load */
381 COSTS_N_INSNS (11), /* fmov, fneg, fabs */
382 COSTS_N_INSNS (11), /* fadd, fsub */
383 COSTS_N_INSNS (11), /* fcmp */
384 COSTS_N_INSNS (11), /* fmov, fmovr */
385 COSTS_N_INSNS (11), /* fmul */
386 COSTS_N_INSNS (24), /* fdivs */
387 COSTS_N_INSNS (37), /* fdivd */
388 COSTS_N_INSNS (24), /* fsqrts */
389 COSTS_N_INSNS (37), /* fsqrtd */
390 COSTS_N_INSNS (12), /* imul */
391 COSTS_N_INSNS (12), /* imulX */
392 0, /* imul bit factor */
393 COSTS_N_INSNS (50), /* idiv, average of 41 - 60 cycle range */
394 COSTS_N_INSNS (35), /* idivX, average of 26 - 44 cycle range */
395 COSTS_N_INSNS (1), /* movcc/movr */
396 0, /* shift penalty */
399 static const struct processor_costs *sparc_costs = &cypress_costs;
401 #ifdef HAVE_AS_RELAX_OPTION
402 /* If 'as' and 'ld' are relaxing tail call insns into branch always, use
403 "or %o7,%g0,X; call Y; or X,%g0,%o7" always, so that it can be optimized.
404 With sethi/jmp, neither 'as' nor 'ld' has an easy way how to find out if
405 somebody does not branch between the sethi and jmp. */
406 #define LEAF_SIBCALL_SLOT_RESERVED_P 1
408 #define LEAF_SIBCALL_SLOT_RESERVED_P \
409 ((TARGET_ARCH64 && !TARGET_CM_MEDLOW) || flag_pic)
412 /* Vector to say how input registers are mapped to output registers.
413 HARD_FRAME_POINTER_REGNUM cannot be remapped by this function to
414 eliminate it. You must use -fomit-frame-pointer to get that. */
415 char leaf_reg_remap[] =
416 { 0, 1, 2, 3, 4, 5, 6, 7,
417 -1, -1, -1, -1, -1, -1, 14, -1,
418 -1, -1, -1, -1, -1, -1, -1, -1,
419 8, 9, 10, 11, 12, 13, -1, 15,
421 32, 33, 34, 35, 36, 37, 38, 39,
422 40, 41, 42, 43, 44, 45, 46, 47,
423 48, 49, 50, 51, 52, 53, 54, 55,
424 56, 57, 58, 59, 60, 61, 62, 63,
425 64, 65, 66, 67, 68, 69, 70, 71,
426 72, 73, 74, 75, 76, 77, 78, 79,
427 80, 81, 82, 83, 84, 85, 86, 87,
428 88, 89, 90, 91, 92, 93, 94, 95,
429 96, 97, 98, 99, 100, 101, 102};
431 /* Vector, indexed by hard register number, which contains 1
432 for a register that is allowable in a candidate for leaf
433 function treatment. */
434 char sparc_leaf_regs[] =
435 { 1, 1, 1, 1, 1, 1, 1, 1,
436 0, 0, 0, 0, 0, 0, 1, 0,
437 0, 0, 0, 0, 0, 0, 0, 0,
438 1, 1, 1, 1, 1, 1, 0, 1,
439 1, 1, 1, 1, 1, 1, 1, 1,
440 1, 1, 1, 1, 1, 1, 1, 1,
441 1, 1, 1, 1, 1, 1, 1, 1,
442 1, 1, 1, 1, 1, 1, 1, 1,
443 1, 1, 1, 1, 1, 1, 1, 1,
444 1, 1, 1, 1, 1, 1, 1, 1,
445 1, 1, 1, 1, 1, 1, 1, 1,
446 1, 1, 1, 1, 1, 1, 1, 1,
447 1, 1, 1, 1, 1, 1, 1};
449 struct GTY(()) machine_function
451 /* Size of the frame of the function. */
452 HOST_WIDE_INT frame_size;
454 /* Size of the frame of the function minus the register window save area
455 and the outgoing argument area. */
456 HOST_WIDE_INT apparent_frame_size;
458 /* Register we pretend the frame pointer is allocated to. Normally, this
459 is %fp, but if we are in a leaf procedure, this is (%sp + offset). We
460 record "offset" separately as it may be too big for (reg + disp). */
462 HOST_WIDE_INT frame_base_offset;
464 /* Some local-dynamic TLS symbol name. */
465 const char *some_ld_name;
467 /* Number of global or FP registers to be saved (as 4-byte quantities). */
468 int n_global_fp_regs;
470 /* True if the current function is leaf and uses only leaf regs,
471 so that the SPARC leaf function optimization can be applied.
472 Private version of crtl->uses_only_leaf_regs, see
473 sparc_expand_prologue for the rationale. */
476 /* True if the prologue saves local or in registers. */
477 bool save_local_in_regs_p;
479 /* True if the data calculated by sparc_expand_prologue are valid. */
480 bool prologue_data_valid_p;
483 #define sparc_frame_size cfun->machine->frame_size
484 #define sparc_apparent_frame_size cfun->machine->apparent_frame_size
485 #define sparc_frame_base_reg cfun->machine->frame_base_reg
486 #define sparc_frame_base_offset cfun->machine->frame_base_offset
487 #define sparc_n_global_fp_regs cfun->machine->n_global_fp_regs
488 #define sparc_leaf_function_p cfun->machine->leaf_function_p
489 #define sparc_save_local_in_regs_p cfun->machine->save_local_in_regs_p
490 #define sparc_prologue_data_valid_p cfun->machine->prologue_data_valid_p
492 /* 1 if the next opcode is to be specially indented. */
493 int sparc_indent_opcode = 0;
495 static void sparc_option_override (void);
496 static void sparc_init_modes (void);
497 static void scan_record_type (const_tree, int *, int *, int *);
498 static int function_arg_slotno (const CUMULATIVE_ARGS *, enum machine_mode,
499 const_tree, bool, bool, int *, int *);
501 static int supersparc_adjust_cost (rtx, rtx, rtx, int);
502 static int hypersparc_adjust_cost (rtx, rtx, rtx, int);
504 static void sparc_emit_set_const32 (rtx, rtx);
505 static void sparc_emit_set_const64 (rtx, rtx);
506 static void sparc_output_addr_vec (rtx);
507 static void sparc_output_addr_diff_vec (rtx);
508 static void sparc_output_deferred_case_vectors (void);
509 static bool sparc_legitimate_address_p (enum machine_mode, rtx, bool);
510 static bool sparc_legitimate_constant_p (enum machine_mode, rtx);
511 static rtx sparc_builtin_saveregs (void);
512 static int epilogue_renumber (rtx *, int);
513 static bool sparc_assemble_integer (rtx, unsigned int, int);
514 static int set_extends (rtx);
515 static void sparc_asm_function_prologue (FILE *, HOST_WIDE_INT);
516 static void sparc_asm_function_epilogue (FILE *, HOST_WIDE_INT);
517 #ifdef TARGET_SOLARIS
518 static void sparc_solaris_elf_asm_named_section (const char *, unsigned int,
519 tree) ATTRIBUTE_UNUSED;
521 static int sparc_adjust_cost (rtx, rtx, rtx, int);
522 static int sparc_issue_rate (void);
523 static void sparc_sched_init (FILE *, int, int);
524 static int sparc_use_sched_lookahead (void);
526 static void emit_soft_tfmode_libcall (const char *, int, rtx *);
527 static void emit_soft_tfmode_binop (enum rtx_code, rtx *);
528 static void emit_soft_tfmode_unop (enum rtx_code, rtx *);
529 static void emit_soft_tfmode_cvt (enum rtx_code, rtx *);
530 static void emit_hard_tfmode_operation (enum rtx_code, rtx *);
532 static bool sparc_function_ok_for_sibcall (tree, tree);
533 static void sparc_init_libfuncs (void);
534 static void sparc_init_builtins (void);
535 static void sparc_vis_init_builtins (void);
536 static rtx sparc_expand_builtin (tree, rtx, rtx, enum machine_mode, int);
537 static tree sparc_fold_builtin (tree, int, tree *, bool);
538 static int sparc_vis_mul8x16 (int, int);
539 static void sparc_handle_vis_mul8x16 (tree *, int, tree, tree, tree);
540 static void sparc_output_mi_thunk (FILE *, tree, HOST_WIDE_INT,
541 HOST_WIDE_INT, tree);
542 static bool sparc_can_output_mi_thunk (const_tree, HOST_WIDE_INT,
543 HOST_WIDE_INT, const_tree);
544 static void sparc_reorg (void);
545 static struct machine_function * sparc_init_machine_status (void);
546 static bool sparc_cannot_force_const_mem (enum machine_mode, rtx);
547 static rtx sparc_tls_get_addr (void);
548 static rtx sparc_tls_got (void);
549 static const char *get_some_local_dynamic_name (void);
550 static int get_some_local_dynamic_name_1 (rtx *, void *);
551 static int sparc_register_move_cost (enum machine_mode,
552 reg_class_t, reg_class_t);
553 static bool sparc_rtx_costs (rtx, int, int, int, int *, bool);
554 static rtx sparc_function_value (const_tree, const_tree, bool);
555 static rtx sparc_libcall_value (enum machine_mode, const_rtx);
556 static bool sparc_function_value_regno_p (const unsigned int);
557 static rtx sparc_struct_value_rtx (tree, int);
558 static enum machine_mode sparc_promote_function_mode (const_tree, enum machine_mode,
559 int *, const_tree, int);
560 static bool sparc_return_in_memory (const_tree, const_tree);
561 static bool sparc_strict_argument_naming (cumulative_args_t);
562 static void sparc_va_start (tree, rtx);
563 static tree sparc_gimplify_va_arg (tree, tree, gimple_seq *, gimple_seq *);
564 static bool sparc_vector_mode_supported_p (enum machine_mode);
565 static bool sparc_tls_referenced_p (rtx);
566 static rtx sparc_legitimize_tls_address (rtx);
567 static rtx sparc_legitimize_pic_address (rtx, rtx);
568 static rtx sparc_legitimize_address (rtx, rtx, enum machine_mode);
569 static rtx sparc_delegitimize_address (rtx);
570 static bool sparc_mode_dependent_address_p (const_rtx);
571 static bool sparc_pass_by_reference (cumulative_args_t,
572 enum machine_mode, const_tree, bool);
573 static void sparc_function_arg_advance (cumulative_args_t,
574 enum machine_mode, const_tree, bool);
575 static rtx sparc_function_arg_1 (cumulative_args_t,
576 enum machine_mode, const_tree, bool, bool);
577 static rtx sparc_function_arg (cumulative_args_t,
578 enum machine_mode, const_tree, bool);
579 static rtx sparc_function_incoming_arg (cumulative_args_t,
580 enum machine_mode, const_tree, bool);
581 static unsigned int sparc_function_arg_boundary (enum machine_mode,
583 static int sparc_arg_partial_bytes (cumulative_args_t,
584 enum machine_mode, tree, bool);
585 static void sparc_output_dwarf_dtprel (FILE *, int, rtx) ATTRIBUTE_UNUSED;
586 static void sparc_file_end (void);
587 static bool sparc_frame_pointer_required (void);
588 static bool sparc_can_eliminate (const int, const int);
589 static rtx sparc_builtin_setjmp_frame_value (void);
590 static void sparc_conditional_register_usage (void);
591 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
592 static const char *sparc_mangle_type (const_tree);
594 static void sparc_trampoline_init (rtx, tree, rtx);
595 static enum machine_mode sparc_preferred_simd_mode (enum machine_mode);
596 static reg_class_t sparc_preferred_reload_class (rtx x, reg_class_t rclass);
597 static bool sparc_print_operand_punct_valid_p (unsigned char);
598 static void sparc_print_operand (FILE *, rtx, int);
599 static void sparc_print_operand_address (FILE *, rtx);
600 static reg_class_t sparc_secondary_reload (bool, rtx, reg_class_t,
602 secondary_reload_info *);
604 #ifdef SUBTARGET_ATTRIBUTE_TABLE
605 /* Table of valid machine attributes. */
606 static const struct attribute_spec sparc_attribute_table[] =
608 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler,
610 SUBTARGET_ATTRIBUTE_TABLE,
611 { NULL, 0, 0, false, false, false, NULL, false }
615 /* Option handling. */
618 enum cmodel sparc_cmodel;
620 char sparc_hard_reg_printed[8];
622 /* Initialize the GCC target structure. */
624 /* The default is to use .half rather than .short for aligned HI objects. */
625 #undef TARGET_ASM_ALIGNED_HI_OP
626 #define TARGET_ASM_ALIGNED_HI_OP "\t.half\t"
628 #undef TARGET_ASM_UNALIGNED_HI_OP
629 #define TARGET_ASM_UNALIGNED_HI_OP "\t.uahalf\t"
630 #undef TARGET_ASM_UNALIGNED_SI_OP
631 #define TARGET_ASM_UNALIGNED_SI_OP "\t.uaword\t"
632 #undef TARGET_ASM_UNALIGNED_DI_OP
633 #define TARGET_ASM_UNALIGNED_DI_OP "\t.uaxword\t"
635 /* The target hook has to handle DI-mode values. */
636 #undef TARGET_ASM_INTEGER
637 #define TARGET_ASM_INTEGER sparc_assemble_integer
639 #undef TARGET_ASM_FUNCTION_PROLOGUE
640 #define TARGET_ASM_FUNCTION_PROLOGUE sparc_asm_function_prologue
641 #undef TARGET_ASM_FUNCTION_EPILOGUE
642 #define TARGET_ASM_FUNCTION_EPILOGUE sparc_asm_function_epilogue
644 #undef TARGET_SCHED_ADJUST_COST
645 #define TARGET_SCHED_ADJUST_COST sparc_adjust_cost
646 #undef TARGET_SCHED_ISSUE_RATE
647 #define TARGET_SCHED_ISSUE_RATE sparc_issue_rate
648 #undef TARGET_SCHED_INIT
649 #define TARGET_SCHED_INIT sparc_sched_init
650 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
651 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD sparc_use_sched_lookahead
653 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
654 #define TARGET_FUNCTION_OK_FOR_SIBCALL sparc_function_ok_for_sibcall
656 #undef TARGET_INIT_LIBFUNCS
657 #define TARGET_INIT_LIBFUNCS sparc_init_libfuncs
658 #undef TARGET_INIT_BUILTINS
659 #define TARGET_INIT_BUILTINS sparc_init_builtins
661 #undef TARGET_LEGITIMIZE_ADDRESS
662 #define TARGET_LEGITIMIZE_ADDRESS sparc_legitimize_address
663 #undef TARGET_DELEGITIMIZE_ADDRESS
664 #define TARGET_DELEGITIMIZE_ADDRESS sparc_delegitimize_address
665 #undef TARGET_MODE_DEPENDENT_ADDRESS_P
666 #define TARGET_MODE_DEPENDENT_ADDRESS_P sparc_mode_dependent_address_p
668 #undef TARGET_EXPAND_BUILTIN
669 #define TARGET_EXPAND_BUILTIN sparc_expand_builtin
670 #undef TARGET_FOLD_BUILTIN
671 #define TARGET_FOLD_BUILTIN sparc_fold_builtin
674 #undef TARGET_HAVE_TLS
675 #define TARGET_HAVE_TLS true
678 #undef TARGET_CANNOT_FORCE_CONST_MEM
679 #define TARGET_CANNOT_FORCE_CONST_MEM sparc_cannot_force_const_mem
681 #undef TARGET_ASM_OUTPUT_MI_THUNK
682 #define TARGET_ASM_OUTPUT_MI_THUNK sparc_output_mi_thunk
683 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
684 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK sparc_can_output_mi_thunk
686 #undef TARGET_MACHINE_DEPENDENT_REORG
687 #define TARGET_MACHINE_DEPENDENT_REORG sparc_reorg
689 #undef TARGET_RTX_COSTS
690 #define TARGET_RTX_COSTS sparc_rtx_costs
691 #undef TARGET_ADDRESS_COST
692 #define TARGET_ADDRESS_COST hook_int_rtx_mode_as_bool_0
693 #undef TARGET_REGISTER_MOVE_COST
694 #define TARGET_REGISTER_MOVE_COST sparc_register_move_cost
696 #undef TARGET_PROMOTE_FUNCTION_MODE
697 #define TARGET_PROMOTE_FUNCTION_MODE sparc_promote_function_mode
699 #undef TARGET_FUNCTION_VALUE
700 #define TARGET_FUNCTION_VALUE sparc_function_value
701 #undef TARGET_LIBCALL_VALUE
702 #define TARGET_LIBCALL_VALUE sparc_libcall_value
703 #undef TARGET_FUNCTION_VALUE_REGNO_P
704 #define TARGET_FUNCTION_VALUE_REGNO_P sparc_function_value_regno_p
706 #undef TARGET_STRUCT_VALUE_RTX
707 #define TARGET_STRUCT_VALUE_RTX sparc_struct_value_rtx
708 #undef TARGET_RETURN_IN_MEMORY
709 #define TARGET_RETURN_IN_MEMORY sparc_return_in_memory
710 #undef TARGET_MUST_PASS_IN_STACK
711 #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size
712 #undef TARGET_PASS_BY_REFERENCE
713 #define TARGET_PASS_BY_REFERENCE sparc_pass_by_reference
714 #undef TARGET_ARG_PARTIAL_BYTES
715 #define TARGET_ARG_PARTIAL_BYTES sparc_arg_partial_bytes
716 #undef TARGET_FUNCTION_ARG_ADVANCE
717 #define TARGET_FUNCTION_ARG_ADVANCE sparc_function_arg_advance
718 #undef TARGET_FUNCTION_ARG
719 #define TARGET_FUNCTION_ARG sparc_function_arg
720 #undef TARGET_FUNCTION_INCOMING_ARG
721 #define TARGET_FUNCTION_INCOMING_ARG sparc_function_incoming_arg
722 #undef TARGET_FUNCTION_ARG_BOUNDARY
723 #define TARGET_FUNCTION_ARG_BOUNDARY sparc_function_arg_boundary
725 #undef TARGET_EXPAND_BUILTIN_SAVEREGS
726 #define TARGET_EXPAND_BUILTIN_SAVEREGS sparc_builtin_saveregs
727 #undef TARGET_STRICT_ARGUMENT_NAMING
728 #define TARGET_STRICT_ARGUMENT_NAMING sparc_strict_argument_naming
730 #undef TARGET_EXPAND_BUILTIN_VA_START
731 #define TARGET_EXPAND_BUILTIN_VA_START sparc_va_start
732 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
733 #define TARGET_GIMPLIFY_VA_ARG_EXPR sparc_gimplify_va_arg
735 #undef TARGET_VECTOR_MODE_SUPPORTED_P
736 #define TARGET_VECTOR_MODE_SUPPORTED_P sparc_vector_mode_supported_p
738 #undef TARGET_VECTORIZE_PREFERRED_SIMD_MODE
739 #define TARGET_VECTORIZE_PREFERRED_SIMD_MODE sparc_preferred_simd_mode
741 #ifdef SUBTARGET_INSERT_ATTRIBUTES
742 #undef TARGET_INSERT_ATTRIBUTES
743 #define TARGET_INSERT_ATTRIBUTES SUBTARGET_INSERT_ATTRIBUTES
746 #ifdef SUBTARGET_ATTRIBUTE_TABLE
747 #undef TARGET_ATTRIBUTE_TABLE
748 #define TARGET_ATTRIBUTE_TABLE sparc_attribute_table
751 #undef TARGET_RELAXED_ORDERING
752 #define TARGET_RELAXED_ORDERING SPARC_RELAXED_ORDERING
754 #undef TARGET_OPTION_OVERRIDE
755 #define TARGET_OPTION_OVERRIDE sparc_option_override
757 #if TARGET_GNU_TLS && defined(HAVE_AS_SPARC_UA_PCREL)
758 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
759 #define TARGET_ASM_OUTPUT_DWARF_DTPREL sparc_output_dwarf_dtprel
762 #undef TARGET_ASM_FILE_END
763 #define TARGET_ASM_FILE_END sparc_file_end
765 #undef TARGET_FRAME_POINTER_REQUIRED
766 #define TARGET_FRAME_POINTER_REQUIRED sparc_frame_pointer_required
768 #undef TARGET_BUILTIN_SETJMP_FRAME_VALUE
769 #define TARGET_BUILTIN_SETJMP_FRAME_VALUE sparc_builtin_setjmp_frame_value
771 #undef TARGET_CAN_ELIMINATE
772 #define TARGET_CAN_ELIMINATE sparc_can_eliminate
774 #undef TARGET_PREFERRED_RELOAD_CLASS
775 #define TARGET_PREFERRED_RELOAD_CLASS sparc_preferred_reload_class
777 #undef TARGET_SECONDARY_RELOAD
778 #define TARGET_SECONDARY_RELOAD sparc_secondary_reload
780 #undef TARGET_CONDITIONAL_REGISTER_USAGE
781 #define TARGET_CONDITIONAL_REGISTER_USAGE sparc_conditional_register_usage
783 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
784 #undef TARGET_MANGLE_TYPE
785 #define TARGET_MANGLE_TYPE sparc_mangle_type
788 #undef TARGET_LEGITIMATE_ADDRESS_P
789 #define TARGET_LEGITIMATE_ADDRESS_P sparc_legitimate_address_p
791 #undef TARGET_LEGITIMATE_CONSTANT_P
792 #define TARGET_LEGITIMATE_CONSTANT_P sparc_legitimate_constant_p
794 #undef TARGET_TRAMPOLINE_INIT
795 #define TARGET_TRAMPOLINE_INIT sparc_trampoline_init
797 #undef TARGET_PRINT_OPERAND_PUNCT_VALID_P
798 #define TARGET_PRINT_OPERAND_PUNCT_VALID_P sparc_print_operand_punct_valid_p
799 #undef TARGET_PRINT_OPERAND
800 #define TARGET_PRINT_OPERAND sparc_print_operand
801 #undef TARGET_PRINT_OPERAND_ADDRESS
802 #define TARGET_PRINT_OPERAND_ADDRESS sparc_print_operand_address
804 /* The value stored by LDSTUB. */
805 #undef TARGET_ATOMIC_TEST_AND_SET_TRUEVAL
806 #define TARGET_ATOMIC_TEST_AND_SET_TRUEVAL 0xff
808 struct gcc_target targetm = TARGET_INITIALIZER;
811 dump_target_flag_bits (const int flags)
813 if (flags & MASK_64BIT)
814 fprintf (stderr, "64BIT ");
815 if (flags & MASK_APP_REGS)
816 fprintf (stderr, "APP_REGS ");
817 if (flags & MASK_FASTER_STRUCTS)
818 fprintf (stderr, "FASTER_STRUCTS ");
819 if (flags & MASK_FLAT)
820 fprintf (stderr, "FLAT ");
821 if (flags & MASK_FMAF)
822 fprintf (stderr, "FMAF ");
823 if (flags & MASK_FPU)
824 fprintf (stderr, "FPU ");
825 if (flags & MASK_HARD_QUAD)
826 fprintf (stderr, "HARD_QUAD ");
827 if (flags & MASK_POPC)
828 fprintf (stderr, "POPC ");
829 if (flags & MASK_PTR64)
830 fprintf (stderr, "PTR64 ");
831 if (flags & MASK_STACK_BIAS)
832 fprintf (stderr, "STACK_BIAS ");
833 if (flags & MASK_UNALIGNED_DOUBLES)
834 fprintf (stderr, "UNALIGNED_DOUBLES ");
835 if (flags & MASK_V8PLUS)
836 fprintf (stderr, "V8PLUS ");
837 if (flags & MASK_VIS)
838 fprintf (stderr, "VIS ");
839 if (flags & MASK_VIS2)
840 fprintf (stderr, "VIS2 ");
841 if (flags & MASK_VIS3)
842 fprintf (stderr, "VIS3 ");
843 if (flags & MASK_DEPRECATED_V8_INSNS)
844 fprintf (stderr, "DEPRECATED_V8_INSNS ");
845 if (flags & MASK_SPARCLET)
846 fprintf (stderr, "SPARCLET ");
847 if (flags & MASK_SPARCLITE)
848 fprintf (stderr, "SPARCLITE ");
850 fprintf (stderr, "V8 ");
852 fprintf (stderr, "V9 ");
856 dump_target_flags (const char *prefix, const int flags)
858 fprintf (stderr, "%s: (%08x) [ ", prefix, flags);
859 dump_target_flag_bits (flags);
860 fprintf(stderr, "]\n");
863 /* Validate and override various options, and do some machine dependent
867 sparc_option_override (void)
869 static struct code_model {
870 const char *const name;
871 const enum cmodel value;
872 } const cmodels[] = {
874 { "medlow", CM_MEDLOW },
875 { "medmid", CM_MEDMID },
876 { "medany", CM_MEDANY },
877 { "embmedany", CM_EMBMEDANY },
878 { NULL, (enum cmodel) 0 }
880 const struct code_model *cmodel;
881 /* Map TARGET_CPU_DEFAULT to value for -m{cpu,tune}=. */
882 static struct cpu_default {
884 const enum processor_type processor;
885 } const cpu_default[] = {
886 /* There must be one entry here for each TARGET_CPU value. */
887 { TARGET_CPU_sparc, PROCESSOR_CYPRESS },
888 { TARGET_CPU_v8, PROCESSOR_V8 },
889 { TARGET_CPU_supersparc, PROCESSOR_SUPERSPARC },
890 { TARGET_CPU_hypersparc, PROCESSOR_HYPERSPARC },
891 { TARGET_CPU_leon, PROCESSOR_LEON },
892 { TARGET_CPU_sparclite, PROCESSOR_F930 },
893 { TARGET_CPU_sparclite86x, PROCESSOR_SPARCLITE86X },
894 { TARGET_CPU_sparclet, PROCESSOR_TSC701 },
895 { TARGET_CPU_v9, PROCESSOR_V9 },
896 { TARGET_CPU_ultrasparc, PROCESSOR_ULTRASPARC },
897 { TARGET_CPU_ultrasparc3, PROCESSOR_ULTRASPARC3 },
898 { TARGET_CPU_niagara, PROCESSOR_NIAGARA },
899 { TARGET_CPU_niagara2, PROCESSOR_NIAGARA2 },
900 { TARGET_CPU_niagara3, PROCESSOR_NIAGARA3 },
901 { TARGET_CPU_niagara4, PROCESSOR_NIAGARA4 },
904 const struct cpu_default *def;
905 /* Table of values for -m{cpu,tune}=. This must match the order of
906 the PROCESSOR_* enumeration. */
907 static struct cpu_table {
908 const char *const name;
911 } const cpu_table[] = {
912 { "v7", MASK_ISA, 0 },
913 { "cypress", MASK_ISA, 0 },
914 { "v8", MASK_ISA, MASK_V8 },
915 /* TI TMS390Z55 supersparc */
916 { "supersparc", MASK_ISA, MASK_V8 },
917 { "hypersparc", MASK_ISA, MASK_V8|MASK_FPU },
919 { "leon", MASK_ISA, MASK_V8|MASK_FPU },
920 { "sparclite", MASK_ISA, MASK_SPARCLITE },
921 /* The Fujitsu MB86930 is the original sparclite chip, with no FPU. */
922 { "f930", MASK_ISA|MASK_FPU, MASK_SPARCLITE },
923 /* The Fujitsu MB86934 is the recent sparclite chip, with an FPU. */
924 { "f934", MASK_ISA, MASK_SPARCLITE|MASK_FPU },
925 { "sparclite86x", MASK_ISA|MASK_FPU, MASK_SPARCLITE },
926 { "sparclet", MASK_ISA, MASK_SPARCLET },
928 { "tsc701", MASK_ISA, MASK_SPARCLET },
929 { "v9", MASK_ISA, MASK_V9 },
930 /* UltraSPARC I, II, IIi */
931 { "ultrasparc", MASK_ISA,
932 /* Although insns using %y are deprecated, it is a clear win. */
933 MASK_V9|MASK_DEPRECATED_V8_INSNS },
935 /* ??? Check if %y issue still holds true. */
936 { "ultrasparc3", MASK_ISA,
937 MASK_V9|MASK_DEPRECATED_V8_INSNS|MASK_VIS2 },
939 { "niagara", MASK_ISA,
940 MASK_V9|MASK_DEPRECATED_V8_INSNS },
942 { "niagara2", MASK_ISA,
943 MASK_V9|MASK_POPC|MASK_VIS2 },
945 { "niagara3", MASK_ISA,
946 MASK_V9|MASK_POPC|MASK_VIS2|MASK_VIS3|MASK_FMAF },
948 { "niagara4", MASK_ISA,
949 MASK_V9|MASK_POPC|MASK_VIS2|MASK_VIS3|MASK_FMAF },
951 const struct cpu_table *cpu;
955 if (sparc_debug_string != NULL)
960 p = ASTRDUP (sparc_debug_string);
961 while ((q = strtok (p, ",")) != NULL)
975 if (! strcmp (q, "all"))
976 mask = MASK_DEBUG_ALL;
977 else if (! strcmp (q, "options"))
978 mask = MASK_DEBUG_OPTIONS;
980 error ("unknown -mdebug-%s switch", q);
983 sparc_debug &= ~mask;
989 if (TARGET_DEBUG_OPTIONS)
991 dump_target_flags("Initial target_flags", target_flags);
992 dump_target_flags("target_flags_explicit", target_flags_explicit);
995 #ifdef SUBTARGET_OVERRIDE_OPTIONS
996 SUBTARGET_OVERRIDE_OPTIONS;
999 #ifndef SPARC_BI_ARCH
1000 /* Check for unsupported architecture size. */
1001 if (! TARGET_64BIT != DEFAULT_ARCH32_P)
1002 error ("%s is not supported by this configuration",
1003 DEFAULT_ARCH32_P ? "-m64" : "-m32");
1006 /* We force all 64bit archs to use 128 bit long double */
1007 if (TARGET_64BIT && ! TARGET_LONG_DOUBLE_128)
1009 error ("-mlong-double-64 not allowed with -m64");
1010 target_flags |= MASK_LONG_DOUBLE_128;
1013 /* Code model selection. */
1014 sparc_cmodel = SPARC_DEFAULT_CMODEL;
1016 #ifdef SPARC_BI_ARCH
1018 sparc_cmodel = CM_32;
1021 if (sparc_cmodel_string != NULL)
1025 for (cmodel = &cmodels[0]; cmodel->name; cmodel++)
1026 if (strcmp (sparc_cmodel_string, cmodel->name) == 0)
1028 if (cmodel->name == NULL)
1029 error ("bad value (%s) for -mcmodel= switch", sparc_cmodel_string);
1031 sparc_cmodel = cmodel->value;
1034 error ("-mcmodel= is not supported on 32 bit systems");
1037 /* Check that -fcall-saved-REG wasn't specified for out registers. */
1038 for (i = 8; i < 16; i++)
1039 if (!call_used_regs [i])
1041 error ("-fcall-saved-REG is not supported for out registers");
1042 call_used_regs [i] = 1;
1045 fpu = target_flags & MASK_FPU; /* save current -mfpu status */
1047 /* Set the default CPU. */
1048 if (!global_options_set.x_sparc_cpu_and_features)
1050 for (def = &cpu_default[0]; def->cpu != -1; ++def)
1051 if (def->cpu == TARGET_CPU_DEFAULT)
1053 gcc_assert (def->cpu != -1);
1054 sparc_cpu_and_features = def->processor;
1057 if (!global_options_set.x_sparc_cpu)
1058 sparc_cpu = sparc_cpu_and_features;
1060 cpu = &cpu_table[(int) sparc_cpu_and_features];
1062 if (TARGET_DEBUG_OPTIONS)
1064 fprintf (stderr, "sparc_cpu_and_features: %s\n", cpu->name);
1065 fprintf (stderr, "sparc_cpu: %s\n",
1066 cpu_table[(int) sparc_cpu].name);
1067 dump_target_flags ("cpu->disable", cpu->disable);
1068 dump_target_flags ("cpu->enable", cpu->enable);
1071 target_flags &= ~cpu->disable;
1072 target_flags |= (cpu->enable
1073 #ifndef HAVE_AS_FMAF_HPC_VIS3
1074 & ~(MASK_FMAF | MASK_VIS3)
1078 /* If -mfpu or -mno-fpu was explicitly used, don't override with
1079 the processor default. */
1080 if (target_flags_explicit & MASK_FPU)
1081 target_flags = (target_flags & ~MASK_FPU) | fpu;
1083 /* -mvis2 implies -mvis */
1085 target_flags |= MASK_VIS;
1087 /* -mvis3 implies -mvis2 and -mvis */
1089 target_flags |= MASK_VIS2 | MASK_VIS;
1091 /* Don't allow -mvis, -mvis2, -mvis3, or -mfmaf if FPU is disabled. */
1093 target_flags &= ~(MASK_VIS | MASK_VIS2 | MASK_VIS3 | MASK_FMAF);
1095 /* -mvis assumes UltraSPARC+, so we are sure v9 instructions
1097 -m64 also implies v9. */
1098 if (TARGET_VIS || TARGET_ARCH64)
1100 target_flags |= MASK_V9;
1101 target_flags &= ~(MASK_V8 | MASK_SPARCLET | MASK_SPARCLITE);
1104 /* -mvis also implies -mv8plus on 32-bit */
1105 if (TARGET_VIS && ! TARGET_ARCH64)
1106 target_flags |= MASK_V8PLUS;
1108 /* Use the deprecated v8 insns for sparc64 in 32 bit mode. */
1109 if (TARGET_V9 && TARGET_ARCH32)
1110 target_flags |= MASK_DEPRECATED_V8_INSNS;
1112 /* V8PLUS requires V9, makes no sense in 64 bit mode. */
1113 if (! TARGET_V9 || TARGET_ARCH64)
1114 target_flags &= ~MASK_V8PLUS;
1116 /* Don't use stack biasing in 32 bit mode. */
1118 target_flags &= ~MASK_STACK_BIAS;
1120 /* Supply a default value for align_functions. */
1121 if (align_functions == 0
1122 && (sparc_cpu == PROCESSOR_ULTRASPARC
1123 || sparc_cpu == PROCESSOR_ULTRASPARC3
1124 || sparc_cpu == PROCESSOR_NIAGARA
1125 || sparc_cpu == PROCESSOR_NIAGARA2
1126 || sparc_cpu == PROCESSOR_NIAGARA3
1127 || sparc_cpu == PROCESSOR_NIAGARA4))
1128 align_functions = 32;
1130 /* Validate PCC_STRUCT_RETURN. */
1131 if (flag_pcc_struct_return == DEFAULT_PCC_STRUCT_RETURN)
1132 flag_pcc_struct_return = (TARGET_ARCH64 ? 0 : 1);
1134 /* Only use .uaxword when compiling for a 64-bit target. */
1136 targetm.asm_out.unaligned_op.di = NULL;
1138 /* Do various machine dependent initializations. */
1139 sparc_init_modes ();
1141 /* Set up function hooks. */
1142 init_machine_status = sparc_init_machine_status;
1147 case PROCESSOR_CYPRESS:
1148 sparc_costs = &cypress_costs;
1151 case PROCESSOR_SPARCLITE:
1152 case PROCESSOR_SUPERSPARC:
1153 sparc_costs = &supersparc_costs;
1155 case PROCESSOR_F930:
1156 case PROCESSOR_F934:
1157 case PROCESSOR_HYPERSPARC:
1158 case PROCESSOR_SPARCLITE86X:
1159 sparc_costs = &hypersparc_costs;
1161 case PROCESSOR_LEON:
1162 sparc_costs = &leon_costs;
1164 case PROCESSOR_SPARCLET:
1165 case PROCESSOR_TSC701:
1166 sparc_costs = &sparclet_costs;
1169 case PROCESSOR_ULTRASPARC:
1170 sparc_costs = &ultrasparc_costs;
1172 case PROCESSOR_ULTRASPARC3:
1173 sparc_costs = &ultrasparc3_costs;
1175 case PROCESSOR_NIAGARA:
1176 sparc_costs = &niagara_costs;
1178 case PROCESSOR_NIAGARA2:
1179 sparc_costs = &niagara2_costs;
1181 case PROCESSOR_NIAGARA3:
1182 sparc_costs = &niagara3_costs;
1184 case PROCESSOR_NIAGARA4:
1185 sparc_costs = &niagara4_costs;
1187 case PROCESSOR_NATIVE:
1191 if (sparc_memory_model == SMM_DEFAULT)
1193 /* Choose the memory model for the operating system. */
1194 enum sparc_memory_model_type os_default = SUBTARGET_DEFAULT_MEMORY_MODEL;
1195 if (os_default != SMM_DEFAULT)
1196 sparc_memory_model = os_default;
1197 /* Choose the most relaxed model for the processor. */
1199 sparc_memory_model = SMM_RMO;
1201 sparc_memory_model = SMM_PSO;
1203 sparc_memory_model = SMM_SC;
1206 #ifdef TARGET_DEFAULT_LONG_DOUBLE_128
1207 if (!(target_flags_explicit & MASK_LONG_DOUBLE_128))
1208 target_flags |= MASK_LONG_DOUBLE_128;
1211 if (TARGET_DEBUG_OPTIONS)
1212 dump_target_flags ("Final target_flags", target_flags);
1214 maybe_set_param_value (PARAM_SIMULTANEOUS_PREFETCHES,
1215 ((sparc_cpu == PROCESSOR_ULTRASPARC
1216 || sparc_cpu == PROCESSOR_NIAGARA
1217 || sparc_cpu == PROCESSOR_NIAGARA2
1218 || sparc_cpu == PROCESSOR_NIAGARA3
1219 || sparc_cpu == PROCESSOR_NIAGARA4)
1221 : (sparc_cpu == PROCESSOR_ULTRASPARC3
1223 global_options.x_param_values,
1224 global_options_set.x_param_values);
1225 maybe_set_param_value (PARAM_L1_CACHE_LINE_SIZE,
1226 ((sparc_cpu == PROCESSOR_ULTRASPARC
1227 || sparc_cpu == PROCESSOR_ULTRASPARC3
1228 || sparc_cpu == PROCESSOR_NIAGARA
1229 || sparc_cpu == PROCESSOR_NIAGARA2
1230 || sparc_cpu == PROCESSOR_NIAGARA3
1231 || sparc_cpu == PROCESSOR_NIAGARA4)
1233 global_options.x_param_values,
1234 global_options_set.x_param_values);
1236 /* Disable save slot sharing for call-clobbered registers by default.
1237 The IRA sharing algorithm works on single registers only and this
1238 pessimizes for double floating-point registers. */
1239 if (!global_options_set.x_flag_ira_share_save_slots)
1240 flag_ira_share_save_slots = 0;
1243 /* Miscellaneous utilities. */
1245 /* Nonzero if CODE, a comparison, is suitable for use in v9 conditional move
1246 or branch on register contents instructions. */
1249 v9_regcmp_p (enum rtx_code code)
1251 return (code == EQ || code == NE || code == GE || code == LT
1252 || code == LE || code == GT);
1255 /* Nonzero if OP is a floating point constant which can
1256 be loaded into an integer register using a single
1257 sethi instruction. */
1262 if (GET_CODE (op) == CONST_DOUBLE)
1267 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
1268 REAL_VALUE_TO_TARGET_SINGLE (r, i);
1269 return !SPARC_SIMM13_P (i) && SPARC_SETHI_P (i);
1275 /* Nonzero if OP is a floating point constant which can
1276 be loaded into an integer register using a single
1282 if (GET_CODE (op) == CONST_DOUBLE)
1287 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
1288 REAL_VALUE_TO_TARGET_SINGLE (r, i);
1289 return SPARC_SIMM13_P (i);
1295 /* Nonzero if OP is a floating point constant which can
1296 be loaded into an integer register using a high/losum
1297 instruction sequence. */
1300 fp_high_losum_p (rtx op)
1302 /* The constraints calling this should only be in
1303 SFmode move insns, so any constant which cannot
1304 be moved using a single insn will do. */
1305 if (GET_CODE (op) == CONST_DOUBLE)
1310 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
1311 REAL_VALUE_TO_TARGET_SINGLE (r, i);
1312 return !SPARC_SIMM13_P (i) && !SPARC_SETHI_P (i);
1318 /* Return true if the address of LABEL can be loaded by means of the
1319 mov{si,di}_pic_label_ref patterns in PIC mode. */
1322 can_use_mov_pic_label_ref (rtx label)
1324 /* VxWorks does not impose a fixed gap between segments; the run-time
1325 gap can be different from the object-file gap. We therefore can't
1326 assume X - _GLOBAL_OFFSET_TABLE_ is a link-time constant unless we
1327 are absolutely sure that X is in the same segment as the GOT.
1328 Unfortunately, the flexibility of linker scripts means that we
1329 can't be sure of that in general, so assume that GOT-relative
1330 accesses are never valid on VxWorks. */
1331 if (TARGET_VXWORKS_RTP)
1334 /* Similarly, if the label is non-local, it might end up being placed
1335 in a different section than the current one; now mov_pic_label_ref
1336 requires the label and the code to be in the same section. */
1337 if (LABEL_REF_NONLOCAL_P (label))
1340 /* Finally, if we are reordering basic blocks and partition into hot
1341 and cold sections, this might happen for any label. */
1342 if (flag_reorder_blocks_and_partition)
1348 /* Expand a move instruction. Return true if all work is done. */
1351 sparc_expand_move (enum machine_mode mode, rtx *operands)
1353 /* Handle sets of MEM first. */
1354 if (GET_CODE (operands[0]) == MEM)
1356 /* 0 is a register (or a pair of registers) on SPARC. */
1357 if (register_or_zero_operand (operands[1], mode))
1360 if (!reload_in_progress)
1362 operands[0] = validize_mem (operands[0]);
1363 operands[1] = force_reg (mode, operands[1]);
1367 /* Fixup TLS cases. */
1369 && CONSTANT_P (operands[1])
1370 && sparc_tls_referenced_p (operands [1]))
1372 operands[1] = sparc_legitimize_tls_address (operands[1]);
1376 /* Fixup PIC cases. */
1377 if (flag_pic && CONSTANT_P (operands[1]))
1379 if (pic_address_needs_scratch (operands[1]))
1380 operands[1] = sparc_legitimize_pic_address (operands[1], NULL_RTX);
1382 /* We cannot use the mov{si,di}_pic_label_ref patterns in all cases. */
1383 if (GET_CODE (operands[1]) == LABEL_REF
1384 && can_use_mov_pic_label_ref (operands[1]))
1388 emit_insn (gen_movsi_pic_label_ref (operands[0], operands[1]));
1394 gcc_assert (TARGET_ARCH64);
1395 emit_insn (gen_movdi_pic_label_ref (operands[0], operands[1]));
1400 if (symbolic_operand (operands[1], mode))
1403 = sparc_legitimize_pic_address (operands[1],
1405 ? operands[0] : NULL_RTX);
1410 /* If we are trying to toss an integer constant into FP registers,
1411 or loading a FP or vector constant, force it into memory. */
1412 if (CONSTANT_P (operands[1])
1413 && REG_P (operands[0])
1414 && (SPARC_FP_REG_P (REGNO (operands[0]))
1415 || SCALAR_FLOAT_MODE_P (mode)
1416 || VECTOR_MODE_P (mode)))
1418 /* emit_group_store will send such bogosity to us when it is
1419 not storing directly into memory. So fix this up to avoid
1420 crashes in output_constant_pool. */
1421 if (operands [1] == const0_rtx)
1422 operands[1] = CONST0_RTX (mode);
1424 /* We can clear or set to all-ones FP registers if TARGET_VIS, and
1425 always other regs. */
1426 if ((TARGET_VIS || REGNO (operands[0]) < SPARC_FIRST_FP_REG)
1427 && (const_zero_operand (operands[1], mode)
1428 || const_all_ones_operand (operands[1], mode)))
1431 if (REGNO (operands[0]) < SPARC_FIRST_FP_REG
1432 /* We are able to build any SF constant in integer registers
1433 with at most 2 instructions. */
1435 /* And any DF constant in integer registers. */
1437 && ! can_create_pseudo_p ())))
1440 operands[1] = force_const_mem (mode, operands[1]);
1441 if (!reload_in_progress)
1442 operands[1] = validize_mem (operands[1]);
1446 /* Accept non-constants and valid constants unmodified. */
1447 if (!CONSTANT_P (operands[1])
1448 || GET_CODE (operands[1]) == HIGH
1449 || input_operand (operands[1], mode))
1455 /* All QImode constants require only one insn, so proceed. */
1460 sparc_emit_set_const32 (operands[0], operands[1]);
1464 /* input_operand should have filtered out 32-bit mode. */
1465 sparc_emit_set_const64 (operands[0], operands[1]);
1471 /* TImode isn't available in 32-bit mode. */
1472 split_double (operands[1], &high, &low);
1473 emit_insn (gen_movdi (operand_subword (operands[0], 0, 0, TImode),
1475 emit_insn (gen_movdi (operand_subword (operands[0], 1, 0, TImode),
1487 /* Load OP1, a 32-bit constant, into OP0, a register.
1488 We know it can't be done in one insn when we get
1489 here, the move expander guarantees this. */
1492 sparc_emit_set_const32 (rtx op0, rtx op1)
1494 enum machine_mode mode = GET_MODE (op0);
1497 if (can_create_pseudo_p ())
1498 temp = gen_reg_rtx (mode);
1500 if (GET_CODE (op1) == CONST_INT)
1502 gcc_assert (!small_int_operand (op1, mode)
1503 && !const_high_operand (op1, mode));
1505 /* Emit them as real moves instead of a HIGH/LO_SUM,
1506 this way CSE can see everything and reuse intermediate
1507 values if it wants. */
1508 emit_insn (gen_rtx_SET (VOIDmode, temp,
1509 GEN_INT (INTVAL (op1)
1510 & ~(HOST_WIDE_INT)0x3ff)));
1512 emit_insn (gen_rtx_SET (VOIDmode,
1514 gen_rtx_IOR (mode, temp,
1515 GEN_INT (INTVAL (op1) & 0x3ff))));
1519 /* A symbol, emit in the traditional way. */
1520 emit_insn (gen_rtx_SET (VOIDmode, temp,
1521 gen_rtx_HIGH (mode, op1)));
1522 emit_insn (gen_rtx_SET (VOIDmode,
1523 op0, gen_rtx_LO_SUM (mode, temp, op1)));
1527 /* Load OP1, a symbolic 64-bit constant, into OP0, a DImode register.
1528 If TEMP is nonzero, we are forbidden to use any other scratch
1529 registers. Otherwise, we are allowed to generate them as needed.
1531 Note that TEMP may have TImode if the code model is TARGET_CM_MEDANY
1532 or TARGET_CM_EMBMEDANY (see the reload_indi and reload_outdi patterns). */
1535 sparc_emit_set_symbolic_const64 (rtx op0, rtx op1, rtx temp)
1537 rtx temp1, temp2, temp3, temp4, temp5;
1540 if (temp && GET_MODE (temp) == TImode)
1543 temp = gen_rtx_REG (DImode, REGNO (temp));
1546 /* SPARC-V9 code-model support. */
1547 switch (sparc_cmodel)
1550 /* The range spanned by all instructions in the object is less
1551 than 2^31 bytes (2GB) and the distance from any instruction
1552 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1553 than 2^31 bytes (2GB).
1555 The executable must be in the low 4TB of the virtual address
1558 sethi %hi(symbol), %temp1
1559 or %temp1, %lo(symbol), %reg */
1561 temp1 = temp; /* op0 is allowed. */
1563 temp1 = gen_reg_rtx (DImode);
1565 emit_insn (gen_rtx_SET (VOIDmode, temp1, gen_rtx_HIGH (DImode, op1)));
1566 emit_insn (gen_rtx_SET (VOIDmode, op0, gen_rtx_LO_SUM (DImode, temp1, op1)));
1570 /* The range spanned by all instructions in the object is less
1571 than 2^31 bytes (2GB) and the distance from any instruction
1572 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1573 than 2^31 bytes (2GB).
1575 The executable must be in the low 16TB of the virtual address
1578 sethi %h44(symbol), %temp1
1579 or %temp1, %m44(symbol), %temp2
1580 sllx %temp2, 12, %temp3
1581 or %temp3, %l44(symbol), %reg */
1586 temp3 = temp; /* op0 is allowed. */
1590 temp1 = gen_reg_rtx (DImode);
1591 temp2 = gen_reg_rtx (DImode);
1592 temp3 = gen_reg_rtx (DImode);
1595 emit_insn (gen_seth44 (temp1, op1));
1596 emit_insn (gen_setm44 (temp2, temp1, op1));
1597 emit_insn (gen_rtx_SET (VOIDmode, temp3,
1598 gen_rtx_ASHIFT (DImode, temp2, GEN_INT (12))));
1599 emit_insn (gen_setl44 (op0, temp3, op1));
1603 /* The range spanned by all instructions in the object is less
1604 than 2^31 bytes (2GB) and the distance from any instruction
1605 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1606 than 2^31 bytes (2GB).
1608 The executable can be placed anywhere in the virtual address
1611 sethi %hh(symbol), %temp1
1612 sethi %lm(symbol), %temp2
1613 or %temp1, %hm(symbol), %temp3
1614 sllx %temp3, 32, %temp4
1615 or %temp4, %temp2, %temp5
1616 or %temp5, %lo(symbol), %reg */
1619 /* It is possible that one of the registers we got for operands[2]
1620 might coincide with that of operands[0] (which is why we made
1621 it TImode). Pick the other one to use as our scratch. */
1622 if (rtx_equal_p (temp, op0))
1624 gcc_assert (ti_temp);
1625 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1628 temp2 = temp; /* op0 is _not_ allowed, see above. */
1635 temp1 = gen_reg_rtx (DImode);
1636 temp2 = gen_reg_rtx (DImode);
1637 temp3 = gen_reg_rtx (DImode);
1638 temp4 = gen_reg_rtx (DImode);
1639 temp5 = gen_reg_rtx (DImode);
1642 emit_insn (gen_sethh (temp1, op1));
1643 emit_insn (gen_setlm (temp2, op1));
1644 emit_insn (gen_sethm (temp3, temp1, op1));
1645 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1646 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1647 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1648 gen_rtx_PLUS (DImode, temp4, temp2)));
1649 emit_insn (gen_setlo (op0, temp5, op1));
1653 /* Old old old backwards compatibility kruft here.
1654 Essentially it is MEDLOW with a fixed 64-bit
1655 virtual base added to all data segment addresses.
1656 Text-segment stuff is computed like MEDANY, we can't
1657 reuse the code above because the relocation knobs
1660 Data segment: sethi %hi(symbol), %temp1
1661 add %temp1, EMBMEDANY_BASE_REG, %temp2
1662 or %temp2, %lo(symbol), %reg */
1663 if (data_segment_operand (op1, GET_MODE (op1)))
1667 temp1 = temp; /* op0 is allowed. */
1672 temp1 = gen_reg_rtx (DImode);
1673 temp2 = gen_reg_rtx (DImode);
1676 emit_insn (gen_embmedany_sethi (temp1, op1));
1677 emit_insn (gen_embmedany_brsum (temp2, temp1));
1678 emit_insn (gen_embmedany_losum (op0, temp2, op1));
1681 /* Text segment: sethi %uhi(symbol), %temp1
1682 sethi %hi(symbol), %temp2
1683 or %temp1, %ulo(symbol), %temp3
1684 sllx %temp3, 32, %temp4
1685 or %temp4, %temp2, %temp5
1686 or %temp5, %lo(symbol), %reg */
1691 /* It is possible that one of the registers we got for operands[2]
1692 might coincide with that of operands[0] (which is why we made
1693 it TImode). Pick the other one to use as our scratch. */
1694 if (rtx_equal_p (temp, op0))
1696 gcc_assert (ti_temp);
1697 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1700 temp2 = temp; /* op0 is _not_ allowed, see above. */
1707 temp1 = gen_reg_rtx (DImode);
1708 temp2 = gen_reg_rtx (DImode);
1709 temp3 = gen_reg_rtx (DImode);
1710 temp4 = gen_reg_rtx (DImode);
1711 temp5 = gen_reg_rtx (DImode);
1714 emit_insn (gen_embmedany_textuhi (temp1, op1));
1715 emit_insn (gen_embmedany_texthi (temp2, op1));
1716 emit_insn (gen_embmedany_textulo (temp3, temp1, op1));
1717 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1718 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1719 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1720 gen_rtx_PLUS (DImode, temp4, temp2)));
1721 emit_insn (gen_embmedany_textlo (op0, temp5, op1));
1730 #if HOST_BITS_PER_WIDE_INT == 32
1732 sparc_emit_set_const64 (rtx op0 ATTRIBUTE_UNUSED, rtx op1 ATTRIBUTE_UNUSED)
1737 /* These avoid problems when cross compiling. If we do not
1738 go through all this hair then the optimizer will see
1739 invalid REG_EQUAL notes or in some cases none at all. */
1740 static rtx gen_safe_HIGH64 (rtx, HOST_WIDE_INT);
1741 static rtx gen_safe_SET64 (rtx, HOST_WIDE_INT);
1742 static rtx gen_safe_OR64 (rtx, HOST_WIDE_INT);
1743 static rtx gen_safe_XOR64 (rtx, HOST_WIDE_INT);
1745 /* The optimizer is not to assume anything about exactly
1746 which bits are set for a HIGH, they are unspecified.
1747 Unfortunately this leads to many missed optimizations
1748 during CSE. We mask out the non-HIGH bits, and matches
1749 a plain movdi, to alleviate this problem. */
1751 gen_safe_HIGH64 (rtx dest, HOST_WIDE_INT val)
1753 return gen_rtx_SET (VOIDmode, dest, GEN_INT (val & ~(HOST_WIDE_INT)0x3ff));
1757 gen_safe_SET64 (rtx dest, HOST_WIDE_INT val)
1759 return gen_rtx_SET (VOIDmode, dest, GEN_INT (val));
1763 gen_safe_OR64 (rtx src, HOST_WIDE_INT val)
1765 return gen_rtx_IOR (DImode, src, GEN_INT (val));
1769 gen_safe_XOR64 (rtx src, HOST_WIDE_INT val)
1771 return gen_rtx_XOR (DImode, src, GEN_INT (val));
1774 /* Worker routines for 64-bit constant formation on arch64.
1775 One of the key things to be doing in these emissions is
1776 to create as many temp REGs as possible. This makes it
1777 possible for half-built constants to be used later when
1778 such values are similar to something required later on.
1779 Without doing this, the optimizer cannot see such
1782 static void sparc_emit_set_const64_quick1 (rtx, rtx,
1783 unsigned HOST_WIDE_INT, int);
1786 sparc_emit_set_const64_quick1 (rtx op0, rtx temp,
1787 unsigned HOST_WIDE_INT low_bits, int is_neg)
1789 unsigned HOST_WIDE_INT high_bits;
1792 high_bits = (~low_bits) & 0xffffffff;
1794 high_bits = low_bits;
1796 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1799 emit_insn (gen_rtx_SET (VOIDmode, op0,
1800 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1804 /* If we are XOR'ing with -1, then we should emit a one's complement
1805 instead. This way the combiner will notice logical operations
1806 such as ANDN later on and substitute. */
1807 if ((low_bits & 0x3ff) == 0x3ff)
1809 emit_insn (gen_rtx_SET (VOIDmode, op0,
1810 gen_rtx_NOT (DImode, temp)));
1814 emit_insn (gen_rtx_SET (VOIDmode, op0,
1815 gen_safe_XOR64 (temp,
1816 (-(HOST_WIDE_INT)0x400
1817 | (low_bits & 0x3ff)))));
1822 static void sparc_emit_set_const64_quick2 (rtx, rtx, unsigned HOST_WIDE_INT,
1823 unsigned HOST_WIDE_INT, int);
1826 sparc_emit_set_const64_quick2 (rtx op0, rtx temp,
1827 unsigned HOST_WIDE_INT high_bits,
1828 unsigned HOST_WIDE_INT low_immediate,
1833 if ((high_bits & 0xfffffc00) != 0)
1835 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1836 if ((high_bits & ~0xfffffc00) != 0)
1837 emit_insn (gen_rtx_SET (VOIDmode, op0,
1838 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1844 emit_insn (gen_safe_SET64 (temp, high_bits));
1848 /* Now shift it up into place. */
1849 emit_insn (gen_rtx_SET (VOIDmode, op0,
1850 gen_rtx_ASHIFT (DImode, temp2,
1851 GEN_INT (shift_count))));
1853 /* If there is a low immediate part piece, finish up by
1854 putting that in as well. */
1855 if (low_immediate != 0)
1856 emit_insn (gen_rtx_SET (VOIDmode, op0,
1857 gen_safe_OR64 (op0, low_immediate)));
1860 static void sparc_emit_set_const64_longway (rtx, rtx, unsigned HOST_WIDE_INT,
1861 unsigned HOST_WIDE_INT);
1863 /* Full 64-bit constant decomposition. Even though this is the
1864 'worst' case, we still optimize a few things away. */
1866 sparc_emit_set_const64_longway (rtx op0, rtx temp,
1867 unsigned HOST_WIDE_INT high_bits,
1868 unsigned HOST_WIDE_INT low_bits)
1872 if (can_create_pseudo_p ())
1873 sub_temp = gen_reg_rtx (DImode);
1875 if ((high_bits & 0xfffffc00) != 0)
1877 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1878 if ((high_bits & ~0xfffffc00) != 0)
1879 emit_insn (gen_rtx_SET (VOIDmode,
1881 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1887 emit_insn (gen_safe_SET64 (temp, high_bits));
1891 if (can_create_pseudo_p ())
1893 rtx temp2 = gen_reg_rtx (DImode);
1894 rtx temp3 = gen_reg_rtx (DImode);
1895 rtx temp4 = gen_reg_rtx (DImode);
1897 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1898 gen_rtx_ASHIFT (DImode, sub_temp,
1901 emit_insn (gen_safe_HIGH64 (temp2, low_bits));
1902 if ((low_bits & ~0xfffffc00) != 0)
1904 emit_insn (gen_rtx_SET (VOIDmode, temp3,
1905 gen_safe_OR64 (temp2, (low_bits & 0x3ff))));
1906 emit_insn (gen_rtx_SET (VOIDmode, op0,
1907 gen_rtx_PLUS (DImode, temp4, temp3)));
1911 emit_insn (gen_rtx_SET (VOIDmode, op0,
1912 gen_rtx_PLUS (DImode, temp4, temp2)));
1917 rtx low1 = GEN_INT ((low_bits >> (32 - 12)) & 0xfff);
1918 rtx low2 = GEN_INT ((low_bits >> (32 - 12 - 12)) & 0xfff);
1919 rtx low3 = GEN_INT ((low_bits >> (32 - 12 - 12 - 8)) & 0x0ff);
1922 /* We are in the middle of reload, so this is really
1923 painful. However we do still make an attempt to
1924 avoid emitting truly stupid code. */
1925 if (low1 != const0_rtx)
1927 emit_insn (gen_rtx_SET (VOIDmode, op0,
1928 gen_rtx_ASHIFT (DImode, sub_temp,
1929 GEN_INT (to_shift))));
1930 emit_insn (gen_rtx_SET (VOIDmode, op0,
1931 gen_rtx_IOR (DImode, op0, low1)));
1939 if (low2 != const0_rtx)
1941 emit_insn (gen_rtx_SET (VOIDmode, op0,
1942 gen_rtx_ASHIFT (DImode, sub_temp,
1943 GEN_INT (to_shift))));
1944 emit_insn (gen_rtx_SET (VOIDmode, op0,
1945 gen_rtx_IOR (DImode, op0, low2)));
1953 emit_insn (gen_rtx_SET (VOIDmode, op0,
1954 gen_rtx_ASHIFT (DImode, sub_temp,
1955 GEN_INT (to_shift))));
1956 if (low3 != const0_rtx)
1957 emit_insn (gen_rtx_SET (VOIDmode, op0,
1958 gen_rtx_IOR (DImode, op0, low3)));
1963 /* Analyze a 64-bit constant for certain properties. */
1964 static void analyze_64bit_constant (unsigned HOST_WIDE_INT,
1965 unsigned HOST_WIDE_INT,
1966 int *, int *, int *);
1969 analyze_64bit_constant (unsigned HOST_WIDE_INT high_bits,
1970 unsigned HOST_WIDE_INT low_bits,
1971 int *hbsp, int *lbsp, int *abbasp)
1973 int lowest_bit_set, highest_bit_set, all_bits_between_are_set;
1976 lowest_bit_set = highest_bit_set = -1;
1980 if ((lowest_bit_set == -1)
1981 && ((low_bits >> i) & 1))
1983 if ((highest_bit_set == -1)
1984 && ((high_bits >> (32 - i - 1)) & 1))
1985 highest_bit_set = (64 - i - 1);
1988 && ((highest_bit_set == -1)
1989 || (lowest_bit_set == -1)));
1995 if ((lowest_bit_set == -1)
1996 && ((high_bits >> i) & 1))
1997 lowest_bit_set = i + 32;
1998 if ((highest_bit_set == -1)
1999 && ((low_bits >> (32 - i - 1)) & 1))
2000 highest_bit_set = 32 - i - 1;
2003 && ((highest_bit_set == -1)
2004 || (lowest_bit_set == -1)));
2006 /* If there are no bits set this should have gone out
2007 as one instruction! */
2008 gcc_assert (lowest_bit_set != -1 && highest_bit_set != -1);
2009 all_bits_between_are_set = 1;
2010 for (i = lowest_bit_set; i <= highest_bit_set; i++)
2014 if ((low_bits & (1 << i)) != 0)
2019 if ((high_bits & (1 << (i - 32))) != 0)
2022 all_bits_between_are_set = 0;
2025 *hbsp = highest_bit_set;
2026 *lbsp = lowest_bit_set;
2027 *abbasp = all_bits_between_are_set;
2030 static int const64_is_2insns (unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT);
2033 const64_is_2insns (unsigned HOST_WIDE_INT high_bits,
2034 unsigned HOST_WIDE_INT low_bits)
2036 int highest_bit_set, lowest_bit_set, all_bits_between_are_set;
2039 || high_bits == 0xffffffff)
2042 analyze_64bit_constant (high_bits, low_bits,
2043 &highest_bit_set, &lowest_bit_set,
2044 &all_bits_between_are_set);
2046 if ((highest_bit_set == 63
2047 || lowest_bit_set == 0)
2048 && all_bits_between_are_set != 0)
2051 if ((highest_bit_set - lowest_bit_set) < 21)
2057 static unsigned HOST_WIDE_INT create_simple_focus_bits (unsigned HOST_WIDE_INT,
2058 unsigned HOST_WIDE_INT,
2061 static unsigned HOST_WIDE_INT
2062 create_simple_focus_bits (unsigned HOST_WIDE_INT high_bits,
2063 unsigned HOST_WIDE_INT low_bits,
2064 int lowest_bit_set, int shift)
2066 HOST_WIDE_INT hi, lo;
2068 if (lowest_bit_set < 32)
2070 lo = (low_bits >> lowest_bit_set) << shift;
2071 hi = ((high_bits << (32 - lowest_bit_set)) << shift);
2076 hi = ((high_bits >> (lowest_bit_set - 32)) << shift);
2078 gcc_assert (! (hi & lo));
2082 /* Here we are sure to be arch64 and this is an integer constant
2083 being loaded into a register. Emit the most efficient
2084 insn sequence possible. Detection of all the 1-insn cases
2085 has been done already. */
2087 sparc_emit_set_const64 (rtx op0, rtx op1)
2089 unsigned HOST_WIDE_INT high_bits, low_bits;
2090 int lowest_bit_set, highest_bit_set;
2091 int all_bits_between_are_set;
2094 /* Sanity check that we know what we are working with. */
2095 gcc_assert (TARGET_ARCH64
2096 && (GET_CODE (op0) == SUBREG
2097 || (REG_P (op0) && ! SPARC_FP_REG_P (REGNO (op0)))));
2099 if (! can_create_pseudo_p ())
2102 if (GET_CODE (op1) != CONST_INT)
2104 sparc_emit_set_symbolic_const64 (op0, op1, temp);
2109 temp = gen_reg_rtx (DImode);
2111 high_bits = ((INTVAL (op1) >> 32) & 0xffffffff);
2112 low_bits = (INTVAL (op1) & 0xffffffff);
2114 /* low_bits bits 0 --> 31
2115 high_bits bits 32 --> 63 */
2117 analyze_64bit_constant (high_bits, low_bits,
2118 &highest_bit_set, &lowest_bit_set,
2119 &all_bits_between_are_set);
2121 /* First try for a 2-insn sequence. */
2123 /* These situations are preferred because the optimizer can
2124 * do more things with them:
2126 * sllx %reg, shift, %reg
2128 * srlx %reg, shift, %reg
2129 * 3) mov some_small_const, %reg
2130 * sllx %reg, shift, %reg
2132 if (((highest_bit_set == 63
2133 || lowest_bit_set == 0)
2134 && all_bits_between_are_set != 0)
2135 || ((highest_bit_set - lowest_bit_set) < 12))
2137 HOST_WIDE_INT the_const = -1;
2138 int shift = lowest_bit_set;
2140 if ((highest_bit_set != 63
2141 && lowest_bit_set != 0)
2142 || all_bits_between_are_set == 0)
2145 create_simple_focus_bits (high_bits, low_bits,
2148 else if (lowest_bit_set == 0)
2149 shift = -(63 - highest_bit_set);
2151 gcc_assert (SPARC_SIMM13_P (the_const));
2152 gcc_assert (shift != 0);
2154 emit_insn (gen_safe_SET64 (temp, the_const));
2156 emit_insn (gen_rtx_SET (VOIDmode,
2158 gen_rtx_ASHIFT (DImode,
2162 emit_insn (gen_rtx_SET (VOIDmode,
2164 gen_rtx_LSHIFTRT (DImode,
2166 GEN_INT (-shift))));
2170 /* Now a range of 22 or less bits set somewhere.
2171 * 1) sethi %hi(focus_bits), %reg
2172 * sllx %reg, shift, %reg
2173 * 2) sethi %hi(focus_bits), %reg
2174 * srlx %reg, shift, %reg
2176 if ((highest_bit_set - lowest_bit_set) < 21)
2178 unsigned HOST_WIDE_INT focus_bits =
2179 create_simple_focus_bits (high_bits, low_bits,
2180 lowest_bit_set, 10);
2182 gcc_assert (SPARC_SETHI_P (focus_bits));
2183 gcc_assert (lowest_bit_set != 10);
2185 emit_insn (gen_safe_HIGH64 (temp, focus_bits));
2187 /* If lowest_bit_set == 10 then a sethi alone could have done it. */
2188 if (lowest_bit_set < 10)
2189 emit_insn (gen_rtx_SET (VOIDmode,
2191 gen_rtx_LSHIFTRT (DImode, temp,
2192 GEN_INT (10 - lowest_bit_set))));
2193 else if (lowest_bit_set > 10)
2194 emit_insn (gen_rtx_SET (VOIDmode,
2196 gen_rtx_ASHIFT (DImode, temp,
2197 GEN_INT (lowest_bit_set - 10))));
2201 /* 1) sethi %hi(low_bits), %reg
2202 * or %reg, %lo(low_bits), %reg
2203 * 2) sethi %hi(~low_bits), %reg
2204 * xor %reg, %lo(-0x400 | (low_bits & 0x3ff)), %reg
2207 || high_bits == 0xffffffff)
2209 sparc_emit_set_const64_quick1 (op0, temp, low_bits,
2210 (high_bits == 0xffffffff));
2214 /* Now, try 3-insn sequences. */
2216 /* 1) sethi %hi(high_bits), %reg
2217 * or %reg, %lo(high_bits), %reg
2218 * sllx %reg, 32, %reg
2222 sparc_emit_set_const64_quick2 (op0, temp, high_bits, 0, 32);
2226 /* We may be able to do something quick
2227 when the constant is negated, so try that. */
2228 if (const64_is_2insns ((~high_bits) & 0xffffffff,
2229 (~low_bits) & 0xfffffc00))
2231 /* NOTE: The trailing bits get XOR'd so we need the
2232 non-negated bits, not the negated ones. */
2233 unsigned HOST_WIDE_INT trailing_bits = low_bits & 0x3ff;
2235 if ((((~high_bits) & 0xffffffff) == 0
2236 && ((~low_bits) & 0x80000000) == 0)
2237 || (((~high_bits) & 0xffffffff) == 0xffffffff
2238 && ((~low_bits) & 0x80000000) != 0))
2240 unsigned HOST_WIDE_INT fast_int = (~low_bits & 0xffffffff);
2242 if ((SPARC_SETHI_P (fast_int)
2243 && (~high_bits & 0xffffffff) == 0)
2244 || SPARC_SIMM13_P (fast_int))
2245 emit_insn (gen_safe_SET64 (temp, fast_int));
2247 sparc_emit_set_const64 (temp, GEN_INT (fast_int));
2252 negated_const = GEN_INT (((~low_bits) & 0xfffffc00) |
2253 (((HOST_WIDE_INT)((~high_bits) & 0xffffffff))<<32));
2254 sparc_emit_set_const64 (temp, negated_const);
2257 /* If we are XOR'ing with -1, then we should emit a one's complement
2258 instead. This way the combiner will notice logical operations
2259 such as ANDN later on and substitute. */
2260 if (trailing_bits == 0x3ff)
2262 emit_insn (gen_rtx_SET (VOIDmode, op0,
2263 gen_rtx_NOT (DImode, temp)));
2267 emit_insn (gen_rtx_SET (VOIDmode,
2269 gen_safe_XOR64 (temp,
2270 (-0x400 | trailing_bits))));
2275 /* 1) sethi %hi(xxx), %reg
2276 * or %reg, %lo(xxx), %reg
2277 * sllx %reg, yyy, %reg
2279 * ??? This is just a generalized version of the low_bits==0
2280 * thing above, FIXME...
2282 if ((highest_bit_set - lowest_bit_set) < 32)
2284 unsigned HOST_WIDE_INT focus_bits =
2285 create_simple_focus_bits (high_bits, low_bits,
2288 /* We can't get here in this state. */
2289 gcc_assert (highest_bit_set >= 32 && lowest_bit_set < 32);
2291 /* So what we know is that the set bits straddle the
2292 middle of the 64-bit word. */
2293 sparc_emit_set_const64_quick2 (op0, temp,
2299 /* 1) sethi %hi(high_bits), %reg
2300 * or %reg, %lo(high_bits), %reg
2301 * sllx %reg, 32, %reg
2302 * or %reg, low_bits, %reg
2304 if (SPARC_SIMM13_P(low_bits)
2305 && ((int)low_bits > 0))
2307 sparc_emit_set_const64_quick2 (op0, temp, high_bits, low_bits, 32);
2311 /* The easiest way when all else fails, is full decomposition. */
2312 sparc_emit_set_const64_longway (op0, temp, high_bits, low_bits);
2314 #endif /* HOST_BITS_PER_WIDE_INT == 32 */
2316 /* Given a comparison code (EQ, NE, etc.) and the first operand of a COMPARE,
2317 return the mode to be used for the comparison. For floating-point,
2318 CCFP[E]mode is used. CC_NOOVmode should be used when the first operand
2319 is a PLUS, MINUS, NEG, or ASHIFT. CCmode should be used when no special
2320 processing is needed. */
2323 select_cc_mode (enum rtx_code op, rtx x, rtx y ATTRIBUTE_UNUSED)
2325 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2351 else if (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS
2352 || GET_CODE (x) == NEG || GET_CODE (x) == ASHIFT)
2354 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
2355 return CCX_NOOVmode;
2361 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
2368 /* Emit the compare insn and return the CC reg for a CODE comparison
2369 with operands X and Y. */
2372 gen_compare_reg_1 (enum rtx_code code, rtx x, rtx y)
2374 enum machine_mode mode;
2377 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2380 mode = SELECT_CC_MODE (code, x, y);
2382 /* ??? We don't have movcc patterns so we cannot generate pseudo regs for the
2383 fcc regs (cse can't tell they're really call clobbered regs and will
2384 remove a duplicate comparison even if there is an intervening function
2385 call - it will then try to reload the cc reg via an int reg which is why
2386 we need the movcc patterns). It is possible to provide the movcc
2387 patterns by using the ldxfsr/stxfsr v9 insns. I tried it: you need two
2388 registers (say %g1,%g5) and it takes about 6 insns. A better fix would be
2389 to tell cse that CCFPE mode registers (even pseudos) are call
2392 /* ??? This is an experiment. Rather than making changes to cse which may
2393 or may not be easy/clean, we do our own cse. This is possible because
2394 we will generate hard registers. Cse knows they're call clobbered (it
2395 doesn't know the same thing about pseudos). If we guess wrong, no big
2396 deal, but if we win, great! */
2398 if (TARGET_V9 && GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2399 #if 1 /* experiment */
2402 /* We cycle through the registers to ensure they're all exercised. */
2403 static int next_fcc_reg = 0;
2404 /* Previous x,y for each fcc reg. */
2405 static rtx prev_args[4][2];
2407 /* Scan prev_args for x,y. */
2408 for (reg = 0; reg < 4; reg++)
2409 if (prev_args[reg][0] == x && prev_args[reg][1] == y)
2414 prev_args[reg][0] = x;
2415 prev_args[reg][1] = y;
2416 next_fcc_reg = (next_fcc_reg + 1) & 3;
2418 cc_reg = gen_rtx_REG (mode, reg + SPARC_FIRST_V9_FCC_REG);
2421 cc_reg = gen_reg_rtx (mode);
2422 #endif /* ! experiment */
2423 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2424 cc_reg = gen_rtx_REG (mode, SPARC_FCC_REG);
2426 cc_reg = gen_rtx_REG (mode, SPARC_ICC_REG);
2428 /* We shouldn't get there for TFmode if !TARGET_HARD_QUAD. If we do, this
2429 will only result in an unrecognizable insn so no point in asserting. */
2430 emit_insn (gen_rtx_SET (VOIDmode, cc_reg, gen_rtx_COMPARE (mode, x, y)));
2436 /* Emit the compare insn and return the CC reg for the comparison in CMP. */
2439 gen_compare_reg (rtx cmp)
2441 return gen_compare_reg_1 (GET_CODE (cmp), XEXP (cmp, 0), XEXP (cmp, 1));
2444 /* This function is used for v9 only.
2445 DEST is the target of the Scc insn.
2446 CODE is the code for an Scc's comparison.
2447 X and Y are the values we compare.
2449 This function is needed to turn
2452 (gt (reg:CCX 100 %icc)
2456 (gt:DI (reg:CCX 100 %icc)
2459 IE: The instruction recognizer needs to see the mode of the comparison to
2460 find the right instruction. We could use "gt:DI" right in the
2461 define_expand, but leaving it out allows us to handle DI, SI, etc. */
2464 gen_v9_scc (rtx dest, enum rtx_code compare_code, rtx x, rtx y)
2467 && (GET_MODE (x) == DImode
2468 || GET_MODE (dest) == DImode))
2471 /* Try to use the movrCC insns. */
2473 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
2475 && v9_regcmp_p (compare_code))
2480 /* Special case for op0 != 0. This can be done with one instruction if
2483 if (compare_code == NE
2484 && GET_MODE (dest) == DImode
2485 && rtx_equal_p (op0, dest))
2487 emit_insn (gen_rtx_SET (VOIDmode, dest,
2488 gen_rtx_IF_THEN_ELSE (DImode,
2489 gen_rtx_fmt_ee (compare_code, DImode,
2496 if (reg_overlap_mentioned_p (dest, op0))
2498 /* Handle the case where dest == x.
2499 We "early clobber" the result. */
2500 op0 = gen_reg_rtx (GET_MODE (x));
2501 emit_move_insn (op0, x);
2504 emit_insn (gen_rtx_SET (VOIDmode, dest, const0_rtx));
2505 if (GET_MODE (op0) != DImode)
2507 temp = gen_reg_rtx (DImode);
2508 convert_move (temp, op0, 0);
2512 emit_insn (gen_rtx_SET (VOIDmode, dest,
2513 gen_rtx_IF_THEN_ELSE (GET_MODE (dest),
2514 gen_rtx_fmt_ee (compare_code, DImode,
2522 x = gen_compare_reg_1 (compare_code, x, y);
2525 gcc_assert (GET_MODE (x) != CC_NOOVmode
2526 && GET_MODE (x) != CCX_NOOVmode);
2528 emit_insn (gen_rtx_SET (VOIDmode, dest, const0_rtx));
2529 emit_insn (gen_rtx_SET (VOIDmode, dest,
2530 gen_rtx_IF_THEN_ELSE (GET_MODE (dest),
2531 gen_rtx_fmt_ee (compare_code,
2532 GET_MODE (x), x, y),
2533 const1_rtx, dest)));
2539 /* Emit an scc insn. For seq, sne, sgeu, and sltu, we can do this
2540 without jumps using the addx/subx instructions. */
2543 emit_scc_insn (rtx operands[])
2550 /* The quad-word fp compare library routines all return nonzero to indicate
2551 true, which is different from the equivalent libgcc routines, so we must
2552 handle them specially here. */
2553 if (GET_MODE (operands[2]) == TFmode && ! TARGET_HARD_QUAD)
2555 operands[1] = sparc_emit_float_lib_cmp (operands[2], operands[3],
2556 GET_CODE (operands[1]));
2557 operands[2] = XEXP (operands[1], 0);
2558 operands[3] = XEXP (operands[1], 1);
2561 code = GET_CODE (operands[1]);
2565 /* For seq/sne on v9 we use the same code as v8 (the addx/subx method has
2566 more applications). The exception to this is "reg != 0" which can
2567 be done in one instruction on v9 (so we do it). */
2570 if (GET_MODE (x) == SImode)
2572 rtx pat = gen_seqsi_special (operands[0], x, y);
2576 else if (GET_MODE (x) == DImode)
2578 rtx pat = gen_seqdi_special (operands[0], x, y);
2586 if (GET_MODE (x) == SImode)
2588 rtx pat = gen_snesi_special (operands[0], x, y);
2592 else if (GET_MODE (x) == DImode)
2596 pat = gen_snedi_special_vis3 (operands[0], x, y);
2598 pat = gen_snedi_special (operands[0], x, y);
2606 && GET_MODE (x) == DImode
2608 && (code == GTU || code == LTU))
2609 && gen_v9_scc (operands[0], code, x, y))
2612 /* We can do LTU and GEU using the addx/subx instructions too. And
2613 for GTU/LEU, if both operands are registers swap them and fall
2614 back to the easy case. */
2615 if (code == GTU || code == LEU)
2617 if ((GET_CODE (x) == REG || GET_CODE (x) == SUBREG)
2618 && (GET_CODE (y) == REG || GET_CODE (y) == SUBREG))
2623 code = swap_condition (code);
2628 || (!TARGET_VIS3 && code == GEU))
2630 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
2631 gen_rtx_fmt_ee (code, SImode,
2632 gen_compare_reg_1 (code, x, y),
2637 /* All the posibilities to use addx/subx based sequences has been
2638 exhausted, try for a 3 instruction sequence using v9 conditional
2640 if (TARGET_V9 && gen_v9_scc (operands[0], code, x, y))
2643 /* Nope, do branches. */
2647 /* Emit a conditional jump insn for the v9 architecture using comparison code
2648 CODE and jump target LABEL.
2649 This function exists to take advantage of the v9 brxx insns. */
2652 emit_v9_brxx_insn (enum rtx_code code, rtx op0, rtx label)
2654 emit_jump_insn (gen_rtx_SET (VOIDmode,
2656 gen_rtx_IF_THEN_ELSE (VOIDmode,
2657 gen_rtx_fmt_ee (code, GET_MODE (op0),
2659 gen_rtx_LABEL_REF (VOIDmode, label),
2664 emit_conditional_branch_insn (rtx operands[])
2666 /* The quad-word fp compare library routines all return nonzero to indicate
2667 true, which is different from the equivalent libgcc routines, so we must
2668 handle them specially here. */
2669 if (GET_MODE (operands[1]) == TFmode && ! TARGET_HARD_QUAD)
2671 operands[0] = sparc_emit_float_lib_cmp (operands[1], operands[2],
2672 GET_CODE (operands[0]));
2673 operands[1] = XEXP (operands[0], 0);
2674 operands[2] = XEXP (operands[0], 1);
2677 if (TARGET_ARCH64 && operands[2] == const0_rtx
2678 && GET_CODE (operands[1]) == REG
2679 && GET_MODE (operands[1]) == DImode)
2681 emit_v9_brxx_insn (GET_CODE (operands[0]), operands[1], operands[3]);
2685 operands[1] = gen_compare_reg (operands[0]);
2686 operands[2] = const0_rtx;
2687 operands[0] = gen_rtx_fmt_ee (GET_CODE (operands[0]), VOIDmode,
2688 operands[1], operands[2]);
2689 emit_jump_insn (gen_cbranchcc4 (operands[0], operands[1], operands[2],
2694 /* Generate a DFmode part of a hard TFmode register.
2695 REG is the TFmode hard register, LOW is 1 for the
2696 low 64bit of the register and 0 otherwise.
2699 gen_df_reg (rtx reg, int low)
2701 int regno = REGNO (reg);
2703 if ((WORDS_BIG_ENDIAN == 0) ^ (low != 0))
2704 regno += (TARGET_ARCH64 && SPARC_INT_REG_P (regno)) ? 1 : 2;
2705 return gen_rtx_REG (DFmode, regno);
2708 /* Generate a call to FUNC with OPERANDS. Operand 0 is the return value.
2709 Unlike normal calls, TFmode operands are passed by reference. It is
2710 assumed that no more than 3 operands are required. */
2713 emit_soft_tfmode_libcall (const char *func_name, int nargs, rtx *operands)
2715 rtx ret_slot = NULL, arg[3], func_sym;
2718 /* We only expect to be called for conversions, unary, and binary ops. */
2719 gcc_assert (nargs == 2 || nargs == 3);
2721 for (i = 0; i < nargs; ++i)
2723 rtx this_arg = operands[i];
2726 /* TFmode arguments and return values are passed by reference. */
2727 if (GET_MODE (this_arg) == TFmode)
2729 int force_stack_temp;
2731 force_stack_temp = 0;
2732 if (TARGET_BUGGY_QP_LIB && i == 0)
2733 force_stack_temp = 1;
2735 if (GET_CODE (this_arg) == MEM
2736 && ! force_stack_temp)
2738 tree expr = MEM_EXPR (this_arg);
2740 mark_addressable (expr);
2741 this_arg = XEXP (this_arg, 0);
2743 else if (CONSTANT_P (this_arg)
2744 && ! force_stack_temp)
2746 this_slot = force_const_mem (TFmode, this_arg);
2747 this_arg = XEXP (this_slot, 0);
2751 this_slot = assign_stack_temp (TFmode, GET_MODE_SIZE (TFmode));
2753 /* Operand 0 is the return value. We'll copy it out later. */
2755 emit_move_insn (this_slot, this_arg);
2757 ret_slot = this_slot;
2759 this_arg = XEXP (this_slot, 0);
2766 func_sym = gen_rtx_SYMBOL_REF (Pmode, func_name);
2768 if (GET_MODE (operands[0]) == TFmode)
2771 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 2,
2772 arg[0], GET_MODE (arg[0]),
2773 arg[1], GET_MODE (arg[1]));
2775 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 3,
2776 arg[0], GET_MODE (arg[0]),
2777 arg[1], GET_MODE (arg[1]),
2778 arg[2], GET_MODE (arg[2]));
2781 emit_move_insn (operands[0], ret_slot);
2787 gcc_assert (nargs == 2);
2789 ret = emit_library_call_value (func_sym, operands[0], LCT_NORMAL,
2790 GET_MODE (operands[0]), 1,
2791 arg[1], GET_MODE (arg[1]));
2793 if (ret != operands[0])
2794 emit_move_insn (operands[0], ret);
2798 /* Expand soft-float TFmode calls to sparc abi routines. */
2801 emit_soft_tfmode_binop (enum rtx_code code, rtx *operands)
2823 emit_soft_tfmode_libcall (func, 3, operands);
2827 emit_soft_tfmode_unop (enum rtx_code code, rtx *operands)
2831 gcc_assert (code == SQRT);
2834 emit_soft_tfmode_libcall (func, 2, operands);
2838 emit_soft_tfmode_cvt (enum rtx_code code, rtx *operands)
2845 switch (GET_MODE (operands[1]))
2858 case FLOAT_TRUNCATE:
2859 switch (GET_MODE (operands[0]))
2873 switch (GET_MODE (operands[1]))
2878 operands[1] = gen_rtx_SIGN_EXTEND (DImode, operands[1]);
2888 case UNSIGNED_FLOAT:
2889 switch (GET_MODE (operands[1]))
2894 operands[1] = gen_rtx_ZERO_EXTEND (DImode, operands[1]);
2905 switch (GET_MODE (operands[0]))
2919 switch (GET_MODE (operands[0]))
2936 emit_soft_tfmode_libcall (func, 2, operands);
2939 /* Expand a hard-float tfmode operation. All arguments must be in
2943 emit_hard_tfmode_operation (enum rtx_code code, rtx *operands)
2947 if (GET_RTX_CLASS (code) == RTX_UNARY)
2949 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2950 op = gen_rtx_fmt_e (code, GET_MODE (operands[0]), operands[1]);
2954 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2955 operands[2] = force_reg (GET_MODE (operands[2]), operands[2]);
2956 op = gen_rtx_fmt_ee (code, GET_MODE (operands[0]),
2957 operands[1], operands[2]);
2960 if (register_operand (operands[0], VOIDmode))
2963 dest = gen_reg_rtx (GET_MODE (operands[0]));
2965 emit_insn (gen_rtx_SET (VOIDmode, dest, op));
2967 if (dest != operands[0])
2968 emit_move_insn (operands[0], dest);
2972 emit_tfmode_binop (enum rtx_code code, rtx *operands)
2974 if (TARGET_HARD_QUAD)
2975 emit_hard_tfmode_operation (code, operands);
2977 emit_soft_tfmode_binop (code, operands);
2981 emit_tfmode_unop (enum rtx_code code, rtx *operands)
2983 if (TARGET_HARD_QUAD)
2984 emit_hard_tfmode_operation (code, operands);
2986 emit_soft_tfmode_unop (code, operands);
2990 emit_tfmode_cvt (enum rtx_code code, rtx *operands)
2992 if (TARGET_HARD_QUAD)
2993 emit_hard_tfmode_operation (code, operands);
2995 emit_soft_tfmode_cvt (code, operands);
2998 /* Return nonzero if a branch/jump/call instruction will be emitting
2999 nop into its delay slot. */
3002 empty_delay_slot (rtx insn)
3006 /* If no previous instruction (should not happen), return true. */
3007 if (PREV_INSN (insn) == NULL)
3010 seq = NEXT_INSN (PREV_INSN (insn));
3011 if (GET_CODE (PATTERN (seq)) == SEQUENCE)
3017 /* Return nonzero if TRIAL can go into the call delay slot. */
3020 tls_call_delay (rtx trial)
3025 call __tls_get_addr, %tgd_call (foo)
3026 add %l7, %o0, %o0, %tgd_add (foo)
3027 while Sun as/ld does not. */
3028 if (TARGET_GNU_TLS || !TARGET_TLS)
3031 pat = PATTERN (trial);
3033 /* We must reject tgd_add{32|64}, i.e.
3034 (set (reg) (plus (reg) (unspec [(reg) (symbol_ref)] UNSPEC_TLSGD)))
3035 and tldm_add{32|64}, i.e.
3036 (set (reg) (plus (reg) (unspec [(reg) (symbol_ref)] UNSPEC_TLSLDM)))
3038 if (GET_CODE (pat) == SET
3039 && GET_CODE (SET_SRC (pat)) == PLUS)
3041 rtx unspec = XEXP (SET_SRC (pat), 1);
3043 if (GET_CODE (unspec) == UNSPEC
3044 && (XINT (unspec, 1) == UNSPEC_TLSGD
3045 || XINT (unspec, 1) == UNSPEC_TLSLDM))
3052 /* Return nonzero if TRIAL, an insn, can be combined with a 'restore'
3053 instruction. RETURN_P is true if the v9 variant 'return' is to be
3054 considered in the test too.
3056 TRIAL must be a SET whose destination is a REG appropriate for the
3057 'restore' instruction or, if RETURN_P is true, for the 'return'
3061 eligible_for_restore_insn (rtx trial, bool return_p)
3063 rtx pat = PATTERN (trial);
3064 rtx src = SET_SRC (pat);
3065 bool src_is_freg = false;
3068 /* Since we now can do moves between float and integer registers when
3069 VIS3 is enabled, we have to catch this case. We can allow such
3070 moves when doing a 'return' however. */
3072 if (GET_CODE (src_reg) == SUBREG)
3073 src_reg = SUBREG_REG (src_reg);
3074 if (GET_CODE (src_reg) == REG
3075 && SPARC_FP_REG_P (REGNO (src_reg)))
3078 /* The 'restore src,%g0,dest' pattern for word mode and below. */
3079 if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
3080 && arith_operand (src, GET_MODE (src))
3084 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
3086 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (SImode);
3089 /* The 'restore src,%g0,dest' pattern for double-word mode. */
3090 else if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
3091 && arith_double_operand (src, GET_MODE (src))
3093 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
3095 /* The 'restore src,%g0,dest' pattern for float if no FPU. */
3096 else if (! TARGET_FPU && register_operand (src, SFmode))
3099 /* The 'restore src,%g0,dest' pattern for double if no FPU. */
3100 else if (! TARGET_FPU && TARGET_ARCH64 && register_operand (src, DFmode))
3103 /* If we have the 'return' instruction, anything that does not use
3104 local or output registers and can go into a delay slot wins. */
3107 && !epilogue_renumber (&pat, 1)
3108 && get_attr_in_uncond_branch_delay (trial)
3109 == IN_UNCOND_BRANCH_DELAY_TRUE)
3112 /* The 'restore src1,src2,dest' pattern for SImode. */
3113 else if (GET_CODE (src) == PLUS
3114 && register_operand (XEXP (src, 0), SImode)
3115 && arith_operand (XEXP (src, 1), SImode))
3118 /* The 'restore src1,src2,dest' pattern for DImode. */
3119 else if (GET_CODE (src) == PLUS
3120 && register_operand (XEXP (src, 0), DImode)
3121 && arith_double_operand (XEXP (src, 1), DImode))
3124 /* The 'restore src1,%lo(src2),dest' pattern. */
3125 else if (GET_CODE (src) == LO_SUM
3126 && ! TARGET_CM_MEDMID
3127 && ((register_operand (XEXP (src, 0), SImode)
3128 && immediate_operand (XEXP (src, 1), SImode))
3130 && register_operand (XEXP (src, 0), DImode)
3131 && immediate_operand (XEXP (src, 1), DImode))))
3134 /* The 'restore src,src,dest' pattern. */
3135 else if (GET_CODE (src) == ASHIFT
3136 && (register_operand (XEXP (src, 0), SImode)
3137 || register_operand (XEXP (src, 0), DImode))
3138 && XEXP (src, 1) == const1_rtx)
3144 /* Return nonzero if TRIAL can go into the function return's delay slot. */
3147 eligible_for_return_delay (rtx trial)
3152 if (GET_CODE (trial) != INSN)
3155 if (get_attr_length (trial) != 1)
3158 /* If the function uses __builtin_eh_return, the eh_return machinery
3159 occupies the delay slot. */
3160 if (crtl->calls_eh_return)
3163 /* In the case of a leaf or flat function, anything can go into the slot. */
3164 if (sparc_leaf_function_p || TARGET_FLAT)
3166 get_attr_in_uncond_branch_delay (trial) == IN_UNCOND_BRANCH_DELAY_TRUE;
3168 pat = PATTERN (trial);
3169 if (GET_CODE (pat) == PARALLEL)
3175 for (i = XVECLEN (pat, 0) - 1; i >= 0; i--)
3177 rtx expr = XVECEXP (pat, 0, i);
3178 if (GET_CODE (expr) != SET)
3180 if (GET_CODE (SET_DEST (expr)) != REG)
3182 regno = REGNO (SET_DEST (expr));
3183 if (regno >= 8 && regno < 24)
3186 return !epilogue_renumber (&pat, 1)
3187 && (get_attr_in_uncond_branch_delay (trial)
3188 == IN_UNCOND_BRANCH_DELAY_TRUE);
3191 if (GET_CODE (pat) != SET)
3194 if (GET_CODE (SET_DEST (pat)) != REG)
3197 regno = REGNO (SET_DEST (pat));
3199 /* Otherwise, only operations which can be done in tandem with
3200 a `restore' or `return' insn can go into the delay slot. */
3201 if (regno >= 8 && regno < 24)
3204 /* If this instruction sets up floating point register and we have a return
3205 instruction, it can probably go in. But restore will not work
3207 if (! SPARC_INT_REG_P (regno))
3209 && !epilogue_renumber (&pat, 1)
3210 && get_attr_in_uncond_branch_delay (trial)
3211 == IN_UNCOND_BRANCH_DELAY_TRUE);
3213 return eligible_for_restore_insn (trial, true);
3216 /* Return nonzero if TRIAL can go into the sibling call's delay slot. */
3219 eligible_for_sibcall_delay (rtx trial)
3223 if (GET_CODE (trial) != INSN || GET_CODE (PATTERN (trial)) != SET)
3226 if (get_attr_length (trial) != 1)
3229 pat = PATTERN (trial);
3231 if (sparc_leaf_function_p || TARGET_FLAT)
3233 /* If the tail call is done using the call instruction,
3234 we have to restore %o7 in the delay slot. */
3235 if (LEAF_SIBCALL_SLOT_RESERVED_P)
3238 /* %g1 is used to build the function address */
3239 if (reg_mentioned_p (gen_rtx_REG (Pmode, 1), pat))
3245 /* Otherwise, only operations which can be done in tandem with
3246 a `restore' insn can go into the delay slot. */
3247 if (GET_CODE (SET_DEST (pat)) != REG
3248 || (REGNO (SET_DEST (pat)) >= 8 && REGNO (SET_DEST (pat)) < 24)
3249 || ! SPARC_INT_REG_P (REGNO (SET_DEST (pat))))
3252 /* If it mentions %o7, it can't go in, because sibcall will clobber it
3254 if (reg_mentioned_p (gen_rtx_REG (Pmode, 15), pat))
3257 return eligible_for_restore_insn (trial, false);
3260 /* Determine if it's legal to put X into the constant pool. This
3261 is not possible if X contains the address of a symbol that is
3262 not constant (TLS) or not known at final link time (PIC). */
3265 sparc_cannot_force_const_mem (enum machine_mode mode, rtx x)
3267 switch (GET_CODE (x))
3272 /* Accept all non-symbolic constants. */
3276 /* Labels are OK iff we are non-PIC. */
3277 return flag_pic != 0;
3280 /* 'Naked' TLS symbol references are never OK,
3281 non-TLS symbols are OK iff we are non-PIC. */
3282 if (SYMBOL_REF_TLS_MODEL (x))
3285 return flag_pic != 0;
3288 return sparc_cannot_force_const_mem (mode, XEXP (x, 0));
3291 return sparc_cannot_force_const_mem (mode, XEXP (x, 0))
3292 || sparc_cannot_force_const_mem (mode, XEXP (x, 1));
3300 /* Global Offset Table support. */
3301 static GTY(()) rtx got_helper_rtx = NULL_RTX;
3302 static GTY(()) rtx global_offset_table_rtx = NULL_RTX;
3304 /* Return the SYMBOL_REF for the Global Offset Table. */
3306 static GTY(()) rtx sparc_got_symbol = NULL_RTX;
3311 if (!sparc_got_symbol)
3312 sparc_got_symbol = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_");
3314 return sparc_got_symbol;
3317 /* Ensure that we are not using patterns that are not OK with PIC. */
3327 op = recog_data.operand[i];
3328 gcc_assert (GET_CODE (op) != SYMBOL_REF
3329 && (GET_CODE (op) != CONST
3330 || (GET_CODE (XEXP (op, 0)) == MINUS
3331 && XEXP (XEXP (op, 0), 0) == sparc_got ()
3332 && GET_CODE (XEXP (XEXP (op, 0), 1)) == CONST)));
3339 /* Return true if X is an address which needs a temporary register when
3340 reloaded while generating PIC code. */
3343 pic_address_needs_scratch (rtx x)
3345 /* An address which is a symbolic plus a non SMALL_INT needs a temp reg. */
3346 if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS
3347 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
3348 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
3349 && ! SMALL_INT (XEXP (XEXP (x, 0), 1)))
3355 /* Determine if a given RTX is a valid constant. We already know this
3356 satisfies CONSTANT_P. */
3359 sparc_legitimate_constant_p (enum machine_mode mode, rtx x)
3361 switch (GET_CODE (x))
3365 if (sparc_tls_referenced_p (x))
3370 if (GET_MODE (x) == VOIDmode)
3373 /* Floating point constants are generally not ok.
3374 The only exception is 0.0 and all-ones in VIS. */
3376 && SCALAR_FLOAT_MODE_P (mode)
3377 && (const_zero_operand (x, mode)
3378 || const_all_ones_operand (x, mode)))
3384 /* Vector constants are generally not ok.
3385 The only exception is 0 or -1 in VIS. */
3387 && (const_zero_operand (x, mode)
3388 || const_all_ones_operand (x, mode)))
3400 /* Determine if a given RTX is a valid constant address. */
3403 constant_address_p (rtx x)
3405 switch (GET_CODE (x))
3413 if (flag_pic && pic_address_needs_scratch (x))
3415 return sparc_legitimate_constant_p (Pmode, x);
3418 return !flag_pic && sparc_legitimate_constant_p (Pmode, x);
3425 /* Nonzero if the constant value X is a legitimate general operand
3426 when generating PIC code. It is given that flag_pic is on and
3427 that X satisfies CONSTANT_P or is a CONST_DOUBLE. */
3430 legitimate_pic_operand_p (rtx x)
3432 if (pic_address_needs_scratch (x))
3434 if (sparc_tls_referenced_p (x))
3439 #define RTX_OK_FOR_OFFSET_P(X, MODE) \
3441 && INTVAL (X) >= -0x1000 \
3442 && INTVAL (X) < (0x1000 - GET_MODE_SIZE (MODE)))
3444 #define RTX_OK_FOR_OLO10_P(X, MODE) \
3446 && INTVAL (X) >= -0x1000 \
3447 && INTVAL (X) < (0xc00 - GET_MODE_SIZE (MODE)))
3449 /* Handle the TARGET_LEGITIMATE_ADDRESS_P target hook.
3451 On SPARC, the actual legitimate addresses must be REG+REG or REG+SMALLINT
3452 ordinarily. This changes a bit when generating PIC. */
3455 sparc_legitimate_address_p (enum machine_mode mode, rtx addr, bool strict)
3457 rtx rs1 = NULL, rs2 = NULL, imm1 = NULL;
3459 if (REG_P (addr) || GET_CODE (addr) == SUBREG)
3461 else if (GET_CODE (addr) == PLUS)
3463 rs1 = XEXP (addr, 0);
3464 rs2 = XEXP (addr, 1);
3466 /* Canonicalize. REG comes first, if there are no regs,
3467 LO_SUM comes first. */
3469 && GET_CODE (rs1) != SUBREG
3471 || GET_CODE (rs2) == SUBREG
3472 || (GET_CODE (rs2) == LO_SUM && GET_CODE (rs1) != LO_SUM)))
3474 rs1 = XEXP (addr, 1);
3475 rs2 = XEXP (addr, 0);
3479 && rs1 == pic_offset_table_rtx
3481 && GET_CODE (rs2) != SUBREG
3482 && GET_CODE (rs2) != LO_SUM
3483 && GET_CODE (rs2) != MEM
3484 && !(GET_CODE (rs2) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs2))
3485 && (! symbolic_operand (rs2, VOIDmode) || mode == Pmode)
3486 && (GET_CODE (rs2) != CONST_INT || SMALL_INT (rs2)))
3488 || GET_CODE (rs1) == SUBREG)
3489 && RTX_OK_FOR_OFFSET_P (rs2, mode)))
3494 else if ((REG_P (rs1) || GET_CODE (rs1) == SUBREG)
3495 && (REG_P (rs2) || GET_CODE (rs2) == SUBREG))
3497 /* We prohibit REG + REG for TFmode when there are no quad move insns
3498 and we consequently need to split. We do this because REG+REG
3499 is not an offsettable address. If we get the situation in reload
3500 where source and destination of a movtf pattern are both MEMs with
3501 REG+REG address, then only one of them gets converted to an
3502 offsettable address. */
3504 && ! (TARGET_ARCH64 && TARGET_HARD_QUAD))
3507 /* Likewise for TImode, but in all cases. */
3511 /* We prohibit REG + REG on ARCH32 if not optimizing for
3512 DFmode/DImode because then mem_min_alignment is likely to be zero
3513 after reload and the forced split would lack a matching splitter
3515 if (TARGET_ARCH32 && !optimize
3516 && (mode == DFmode || mode == DImode))
3519 else if (USE_AS_OFFSETABLE_LO10
3520 && GET_CODE (rs1) == LO_SUM
3522 && ! TARGET_CM_MEDMID
3523 && RTX_OK_FOR_OLO10_P (rs2, mode))
3526 imm1 = XEXP (rs1, 1);
3527 rs1 = XEXP (rs1, 0);
3528 if (!CONSTANT_P (imm1)
3529 || (GET_CODE (rs1) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs1)))
3533 else if (GET_CODE (addr) == LO_SUM)
3535 rs1 = XEXP (addr, 0);
3536 imm1 = XEXP (addr, 1);
3538 if (!CONSTANT_P (imm1)
3539 || (GET_CODE (rs1) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs1)))
3542 /* We can't allow TFmode in 32-bit mode, because an offset greater
3543 than the alignment (8) may cause the LO_SUM to overflow. */
3544 if (mode == TFmode && TARGET_ARCH32)
3547 else if (GET_CODE (addr) == CONST_INT && SMALL_INT (addr))
3552 if (GET_CODE (rs1) == SUBREG)
3553 rs1 = SUBREG_REG (rs1);
3559 if (GET_CODE (rs2) == SUBREG)
3560 rs2 = SUBREG_REG (rs2);
3567 if (!REGNO_OK_FOR_BASE_P (REGNO (rs1))
3568 || (rs2 && !REGNO_OK_FOR_BASE_P (REGNO (rs2))))
3573 if ((! SPARC_INT_REG_P (REGNO (rs1))
3574 && REGNO (rs1) != FRAME_POINTER_REGNUM
3575 && REGNO (rs1) < FIRST_PSEUDO_REGISTER)
3577 && (! SPARC_INT_REG_P (REGNO (rs2))
3578 && REGNO (rs2) != FRAME_POINTER_REGNUM
3579 && REGNO (rs2) < FIRST_PSEUDO_REGISTER)))
3585 /* Return the SYMBOL_REF for the tls_get_addr function. */
3587 static GTY(()) rtx sparc_tls_symbol = NULL_RTX;
3590 sparc_tls_get_addr (void)
3592 if (!sparc_tls_symbol)
3593 sparc_tls_symbol = gen_rtx_SYMBOL_REF (Pmode, "__tls_get_addr");
3595 return sparc_tls_symbol;
3598 /* Return the Global Offset Table to be used in TLS mode. */
3601 sparc_tls_got (void)
3603 /* In PIC mode, this is just the PIC offset table. */
3606 crtl->uses_pic_offset_table = 1;
3607 return pic_offset_table_rtx;
3610 /* In non-PIC mode, Sun as (unlike GNU as) emits PC-relative relocations for
3611 the GOT symbol with the 32-bit ABI, so we reload the GOT register. */
3612 if (TARGET_SUN_TLS && TARGET_ARCH32)
3614 load_got_register ();
3615 return global_offset_table_rtx;
3618 /* In all other cases, we load a new pseudo with the GOT symbol. */
3619 return copy_to_reg (sparc_got ());
3622 /* Return true if X contains a thread-local symbol. */
3625 sparc_tls_referenced_p (rtx x)
3627 if (!TARGET_HAVE_TLS)
3630 if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS)
3631 x = XEXP (XEXP (x, 0), 0);
3633 if (GET_CODE (x) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (x))
3636 /* That's all we handle in sparc_legitimize_tls_address for now. */
3640 /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute
3641 this (thread-local) address. */
3644 sparc_legitimize_tls_address (rtx addr)
3646 rtx temp1, temp2, temp3, ret, o0, got, insn;
3648 gcc_assert (can_create_pseudo_p ());
3650 if (GET_CODE (addr) == SYMBOL_REF)
3651 switch (SYMBOL_REF_TLS_MODEL (addr))
3653 case TLS_MODEL_GLOBAL_DYNAMIC:
3655 temp1 = gen_reg_rtx (SImode);
3656 temp2 = gen_reg_rtx (SImode);
3657 ret = gen_reg_rtx (Pmode);
3658 o0 = gen_rtx_REG (Pmode, 8);
3659 got = sparc_tls_got ();
3660 emit_insn (gen_tgd_hi22 (temp1, addr));
3661 emit_insn (gen_tgd_lo10 (temp2, temp1, addr));
3664 emit_insn (gen_tgd_add32 (o0, got, temp2, addr));
3665 insn = emit_call_insn (gen_tgd_call32 (o0, sparc_tls_get_addr (),
3670 emit_insn (gen_tgd_add64 (o0, got, temp2, addr));
3671 insn = emit_call_insn (gen_tgd_call64 (o0, sparc_tls_get_addr (),
3674 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), o0);
3675 insn = get_insns ();
3677 emit_libcall_block (insn, ret, o0, addr);
3680 case TLS_MODEL_LOCAL_DYNAMIC:
3682 temp1 = gen_reg_rtx (SImode);
3683 temp2 = gen_reg_rtx (SImode);
3684 temp3 = gen_reg_rtx (Pmode);
3685 ret = gen_reg_rtx (Pmode);
3686 o0 = gen_rtx_REG (Pmode, 8);
3687 got = sparc_tls_got ();
3688 emit_insn (gen_tldm_hi22 (temp1));
3689 emit_insn (gen_tldm_lo10 (temp2, temp1));
3692 emit_insn (gen_tldm_add32 (o0, got, temp2));
3693 insn = emit_call_insn (gen_tldm_call32 (o0, sparc_tls_get_addr (),
3698 emit_insn (gen_tldm_add64 (o0, got, temp2));
3699 insn = emit_call_insn (gen_tldm_call64 (o0, sparc_tls_get_addr (),
3702 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), o0);
3703 insn = get_insns ();
3705 emit_libcall_block (insn, temp3, o0,
3706 gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
3707 UNSPEC_TLSLD_BASE));
3708 temp1 = gen_reg_rtx (SImode);
3709 temp2 = gen_reg_rtx (SImode);
3710 emit_insn (gen_tldo_hix22 (temp1, addr));
3711 emit_insn (gen_tldo_lox10 (temp2, temp1, addr));
3713 emit_insn (gen_tldo_add32 (ret, temp3, temp2, addr));
3715 emit_insn (gen_tldo_add64 (ret, temp3, temp2, addr));
3718 case TLS_MODEL_INITIAL_EXEC:
3719 temp1 = gen_reg_rtx (SImode);
3720 temp2 = gen_reg_rtx (SImode);
3721 temp3 = gen_reg_rtx (Pmode);
3722 got = sparc_tls_got ();
3723 emit_insn (gen_tie_hi22 (temp1, addr));
3724 emit_insn (gen_tie_lo10 (temp2, temp1, addr));
3726 emit_insn (gen_tie_ld32 (temp3, got, temp2, addr));
3728 emit_insn (gen_tie_ld64 (temp3, got, temp2, addr));
3731 ret = gen_reg_rtx (Pmode);
3733 emit_insn (gen_tie_add32 (ret, gen_rtx_REG (Pmode, 7),
3736 emit_insn (gen_tie_add64 (ret, gen_rtx_REG (Pmode, 7),
3740 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp3);
3743 case TLS_MODEL_LOCAL_EXEC:
3744 temp1 = gen_reg_rtx (Pmode);
3745 temp2 = gen_reg_rtx (Pmode);
3748 emit_insn (gen_tle_hix22_sp32 (temp1, addr));
3749 emit_insn (gen_tle_lox10_sp32 (temp2, temp1, addr));
3753 emit_insn (gen_tle_hix22_sp64 (temp1, addr));
3754 emit_insn (gen_tle_lox10_sp64 (temp2, temp1, addr));
3756 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp2);
3763 else if (GET_CODE (addr) == CONST)
3767 gcc_assert (GET_CODE (XEXP (addr, 0)) == PLUS);
3769 base = sparc_legitimize_tls_address (XEXP (XEXP (addr, 0), 0));
3770 offset = XEXP (XEXP (addr, 0), 1);
3772 base = force_operand (base, NULL_RTX);
3773 if (!(GET_CODE (offset) == CONST_INT && SMALL_INT (offset)))
3774 offset = force_reg (Pmode, offset);
3775 ret = gen_rtx_PLUS (Pmode, base, offset);
3779 gcc_unreachable (); /* for now ... */
3784 /* Legitimize PIC addresses. If the address is already position-independent,
3785 we return ORIG. Newly generated position-independent addresses go into a
3786 reg. This is REG if nonzero, otherwise we allocate register(s) as
3790 sparc_legitimize_pic_address (rtx orig, rtx reg)
3792 bool gotdata_op = false;
3794 if (GET_CODE (orig) == SYMBOL_REF
3795 /* See the comment in sparc_expand_move. */
3796 || (GET_CODE (orig) == LABEL_REF && !can_use_mov_pic_label_ref (orig)))
3798 rtx pic_ref, address;
3803 gcc_assert (can_create_pseudo_p ());
3804 reg = gen_reg_rtx (Pmode);
3809 /* If not during reload, allocate another temp reg here for loading
3810 in the address, so that these instructions can be optimized
3812 rtx temp_reg = (! can_create_pseudo_p ()
3813 ? reg : gen_reg_rtx (Pmode));
3815 /* Must put the SYMBOL_REF inside an UNSPEC here so that cse
3816 won't get confused into thinking that these two instructions
3817 are loading in the true address of the symbol. If in the
3818 future a PIC rtx exists, that should be used instead. */
3821 emit_insn (gen_movdi_high_pic (temp_reg, orig));
3822 emit_insn (gen_movdi_lo_sum_pic (temp_reg, temp_reg, orig));
3826 emit_insn (gen_movsi_high_pic (temp_reg, orig));
3827 emit_insn (gen_movsi_lo_sum_pic (temp_reg, temp_reg, orig));
3835 crtl->uses_pic_offset_table = 1;
3839 insn = emit_insn (gen_movdi_pic_gotdata_op (reg,
3840 pic_offset_table_rtx,
3843 insn = emit_insn (gen_movsi_pic_gotdata_op (reg,
3844 pic_offset_table_rtx,
3850 = gen_const_mem (Pmode,
3851 gen_rtx_PLUS (Pmode,
3852 pic_offset_table_rtx, address));
3853 insn = emit_move_insn (reg, pic_ref);
3856 /* Put a REG_EQUAL note on this insn, so that it can be optimized
3858 set_unique_reg_note (insn, REG_EQUAL, orig);
3861 else if (GET_CODE (orig) == CONST)
3865 if (GET_CODE (XEXP (orig, 0)) == PLUS
3866 && XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx)
3871 gcc_assert (can_create_pseudo_p ());
3872 reg = gen_reg_rtx (Pmode);
3875 gcc_assert (GET_CODE (XEXP (orig, 0)) == PLUS);
3876 base = sparc_legitimize_pic_address (XEXP (XEXP (orig, 0), 0), reg);
3877 offset = sparc_legitimize_pic_address (XEXP (XEXP (orig, 0), 1),
3878 base == reg ? NULL_RTX : reg);
3880 if (GET_CODE (offset) == CONST_INT)
3882 if (SMALL_INT (offset))
3883 return plus_constant (Pmode, base, INTVAL (offset));
3884 else if (can_create_pseudo_p ())
3885 offset = force_reg (Pmode, offset);
3887 /* If we reach here, then something is seriously wrong. */
3890 return gen_rtx_PLUS (Pmode, base, offset);
3892 else if (GET_CODE (orig) == LABEL_REF)
3893 /* ??? We ought to be checking that the register is live instead, in case
3894 it is eliminated. */
3895 crtl->uses_pic_offset_table = 1;
3900 /* Try machine-dependent ways of modifying an illegitimate address X
3901 to be legitimate. If we find one, return the new, valid address.
3903 OLDX is the address as it was before break_out_memory_refs was called.
3904 In some cases it is useful to look at this to decide what needs to be done.
3906 MODE is the mode of the operand pointed to by X.
3908 On SPARC, change REG+N into REG+REG, and REG+(X*Y) into REG+REG. */
3911 sparc_legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED,
3912 enum machine_mode mode)
3916 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == MULT)
3917 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3918 force_operand (XEXP (x, 0), NULL_RTX));
3919 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == MULT)
3920 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3921 force_operand (XEXP (x, 1), NULL_RTX));
3922 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == PLUS)
3923 x = gen_rtx_PLUS (Pmode, force_operand (XEXP (x, 0), NULL_RTX),
3925 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == PLUS)
3926 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3927 force_operand (XEXP (x, 1), NULL_RTX));
3929 if (x != orig_x && sparc_legitimate_address_p (mode, x, FALSE))
3932 if (sparc_tls_referenced_p (x))
3933 x = sparc_legitimize_tls_address (x);
3935 x = sparc_legitimize_pic_address (x, NULL_RTX);
3936 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 1)))
3937 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3938 copy_to_mode_reg (Pmode, XEXP (x, 1)));
3939 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 0)))
3940 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3941 copy_to_mode_reg (Pmode, XEXP (x, 0)));
3942 else if (GET_CODE (x) == SYMBOL_REF
3943 || GET_CODE (x) == CONST
3944 || GET_CODE (x) == LABEL_REF)
3945 x = copy_to_suggested_reg (x, NULL_RTX, Pmode);
3950 /* Delegitimize an address that was legitimized by the above function. */
3953 sparc_delegitimize_address (rtx x)
3955 x = delegitimize_mem_from_attrs (x);
3957 if (GET_CODE (x) == LO_SUM && GET_CODE (XEXP (x, 1)) == UNSPEC)
3958 switch (XINT (XEXP (x, 1), 1))
3960 case UNSPEC_MOVE_PIC:
3962 x = XVECEXP (XEXP (x, 1), 0, 0);
3963 gcc_assert (GET_CODE (x) == SYMBOL_REF);
3969 /* This is generated by mov{si,di}_pic_label_ref in PIC mode. */
3970 if (GET_CODE (x) == MINUS
3971 && REG_P (XEXP (x, 0))
3972 && REGNO (XEXP (x, 0)) == PIC_OFFSET_TABLE_REGNUM
3973 && GET_CODE (XEXP (x, 1)) == LO_SUM
3974 && GET_CODE (XEXP (XEXP (x, 1), 1)) == UNSPEC
3975 && XINT (XEXP (XEXP (x, 1), 1), 1) == UNSPEC_MOVE_PIC_LABEL)
3977 x = XVECEXP (XEXP (XEXP (x, 1), 1), 0, 0);
3978 gcc_assert (GET_CODE (x) == LABEL_REF);
3984 /* SPARC implementation of LEGITIMIZE_RELOAD_ADDRESS. Returns a value to
3985 replace the input X, or the original X if no replacement is called for.
3986 The output parameter *WIN is 1 if the calling macro should goto WIN,
3989 For SPARC, we wish to handle addresses by splitting them into
3990 HIGH+LO_SUM pairs, retaining the LO_SUM in the memory reference.
3991 This cuts the number of extra insns by one.
3993 Do nothing when generating PIC code and the address is a symbolic
3994 operand or requires a scratch register. */
3997 sparc_legitimize_reload_address (rtx x, enum machine_mode mode,
3998 int opnum, int type,
3999 int ind_levels ATTRIBUTE_UNUSED, int *win)
4001 /* Decompose SImode constants into HIGH+LO_SUM. */
4003 && (mode != TFmode || TARGET_ARCH64)
4004 && GET_MODE (x) == SImode
4005 && GET_CODE (x) != LO_SUM
4006 && GET_CODE (x) != HIGH
4007 && sparc_cmodel <= CM_MEDLOW
4009 && (symbolic_operand (x, Pmode) || pic_address_needs_scratch (x))))
4011 x = gen_rtx_LO_SUM (GET_MODE (x), gen_rtx_HIGH (GET_MODE (x), x), x);
4012 push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL,
4013 BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0,
4014 opnum, (enum reload_type)type);
4019 /* We have to recognize what we have already generated above. */
4020 if (GET_CODE (x) == LO_SUM && GET_CODE (XEXP (x, 0)) == HIGH)
4022 push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL,
4023 BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0,
4024 opnum, (enum reload_type)type);
4033 /* Return true if ADDR (a legitimate address expression)
4034 has an effect that depends on the machine mode it is used for.
4040 is not equivalent to
4042 (mem:QI [%l7+a]) (mem:QI [%l7+a+1])
4044 because [%l7+a+1] is interpreted as the address of (a+1). */
4048 sparc_mode_dependent_address_p (const_rtx addr)
4050 if (flag_pic && GET_CODE (addr) == PLUS)
4052 rtx op0 = XEXP (addr, 0);
4053 rtx op1 = XEXP (addr, 1);
4054 if (op0 == pic_offset_table_rtx
4055 && symbolic_operand (op1, VOIDmode))
4062 #ifdef HAVE_GAS_HIDDEN
4063 # define USE_HIDDEN_LINKONCE 1
4065 # define USE_HIDDEN_LINKONCE 0
4069 get_pc_thunk_name (char name[32], unsigned int regno)
4071 const char *reg_name = reg_names[regno];
4073 /* Skip the leading '%' as that cannot be used in a
4077 if (USE_HIDDEN_LINKONCE)
4078 sprintf (name, "__sparc_get_pc_thunk.%s", reg_name);
4080 ASM_GENERATE_INTERNAL_LABEL (name, "LADDPC", regno);
4083 /* Wrapper around the load_pcrel_sym{si,di} patterns. */
4086 gen_load_pcrel_sym (rtx op0, rtx op1, rtx op2, rtx op3)
4088 int orig_flag_pic = flag_pic;
4091 /* The load_pcrel_sym{si,di} patterns require absolute addressing. */
4094 insn = gen_load_pcrel_symdi (op0, op1, op2, op3);
4096 insn = gen_load_pcrel_symsi (op0, op1, op2, op3);
4097 flag_pic = orig_flag_pic;
4102 /* Emit code to load the GOT register. */
4105 load_got_register (void)
4107 /* In PIC mode, this will retrieve pic_offset_table_rtx. */
4108 if (!global_offset_table_rtx)
4109 global_offset_table_rtx = gen_rtx_REG (Pmode, GLOBAL_OFFSET_TABLE_REGNUM);
4111 if (TARGET_VXWORKS_RTP)
4112 emit_insn (gen_vxworks_load_got ());
4115 /* The GOT symbol is subject to a PC-relative relocation so we need a
4116 helper function to add the PC value and thus get the final value. */
4117 if (!got_helper_rtx)
4120 get_pc_thunk_name (name, GLOBAL_OFFSET_TABLE_REGNUM);
4121 got_helper_rtx = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (name));
4124 emit_insn (gen_load_pcrel_sym (global_offset_table_rtx, sparc_got (),
4126 GEN_INT (GLOBAL_OFFSET_TABLE_REGNUM)));
4129 /* Need to emit this whether or not we obey regdecls,
4130 since setjmp/longjmp can cause life info to screw up.
4131 ??? In the case where we don't obey regdecls, this is not sufficient
4132 since we may not fall out the bottom. */
4133 emit_use (global_offset_table_rtx);
4136 /* Emit a call instruction with the pattern given by PAT. ADDR is the
4137 address of the call target. */
4140 sparc_emit_call_insn (rtx pat, rtx addr)
4144 insn = emit_call_insn (pat);
4146 /* The PIC register is live on entry to VxWorks PIC PLT entries. */
4147 if (TARGET_VXWORKS_RTP
4149 && GET_CODE (addr) == SYMBOL_REF
4150 && (SYMBOL_REF_DECL (addr)
4151 ? !targetm.binds_local_p (SYMBOL_REF_DECL (addr))
4152 : !SYMBOL_REF_LOCAL_P (addr)))
4154 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), pic_offset_table_rtx);
4155 crtl->uses_pic_offset_table = 1;
4159 /* Return 1 if RTX is a MEM which is known to be aligned to at
4160 least a DESIRED byte boundary. */
4163 mem_min_alignment (rtx mem, int desired)
4165 rtx addr, base, offset;
4167 /* If it's not a MEM we can't accept it. */
4168 if (GET_CODE (mem) != MEM)
4172 if (!TARGET_UNALIGNED_DOUBLES
4173 && MEM_ALIGN (mem) / BITS_PER_UNIT >= (unsigned)desired)
4176 /* ??? The rest of the function predates MEM_ALIGN so
4177 there is probably a bit of redundancy. */
4178 addr = XEXP (mem, 0);
4179 base = offset = NULL_RTX;
4180 if (GET_CODE (addr) == PLUS)
4182 if (GET_CODE (XEXP (addr, 0)) == REG)
4184 base = XEXP (addr, 0);
4186 /* What we are saying here is that if the base
4187 REG is aligned properly, the compiler will make
4188 sure any REG based index upon it will be so
4190 if (GET_CODE (XEXP (addr, 1)) == CONST_INT)
4191 offset = XEXP (addr, 1);
4193 offset = const0_rtx;
4196 else if (GET_CODE (addr) == REG)
4199 offset = const0_rtx;
4202 if (base != NULL_RTX)
4204 int regno = REGNO (base);
4206 if (regno != HARD_FRAME_POINTER_REGNUM && regno != STACK_POINTER_REGNUM)
4208 /* Check if the compiler has recorded some information
4209 about the alignment of the base REG. If reload has
4210 completed, we already matched with proper alignments.
4211 If not running global_alloc, reload might give us
4212 unaligned pointer to local stack though. */
4214 && REGNO_POINTER_ALIGN (regno) >= desired * BITS_PER_UNIT)
4215 || (optimize && reload_completed))
4216 && (INTVAL (offset) & (desired - 1)) == 0)
4221 if (((INTVAL (offset) - SPARC_STACK_BIAS) & (desired - 1)) == 0)
4225 else if (! TARGET_UNALIGNED_DOUBLES
4226 || CONSTANT_P (addr)
4227 || GET_CODE (addr) == LO_SUM)
4229 /* Anything else we know is properly aligned unless TARGET_UNALIGNED_DOUBLES
4230 is true, in which case we can only assume that an access is aligned if
4231 it is to a constant address, or the address involves a LO_SUM. */
4235 /* An obviously unaligned address. */
4240 /* Vectors to keep interesting information about registers where it can easily
4241 be got. We used to use the actual mode value as the bit number, but there
4242 are more than 32 modes now. Instead we use two tables: one indexed by
4243 hard register number, and one indexed by mode. */
4245 /* The purpose of sparc_mode_class is to shrink the range of modes so that
4246 they all fit (as bit numbers) in a 32-bit word (again). Each real mode is
4247 mapped into one sparc_mode_class mode. */
4249 enum sparc_mode_class {
4250 S_MODE, D_MODE, T_MODE, O_MODE,
4251 SF_MODE, DF_MODE, TF_MODE, OF_MODE,
4255 /* Modes for single-word and smaller quantities. */
4256 #define S_MODES ((1 << (int) S_MODE) | (1 << (int) SF_MODE))
4258 /* Modes for double-word and smaller quantities. */
4259 #define D_MODES (S_MODES | (1 << (int) D_MODE) | (1 << DF_MODE))
4261 /* Modes for quad-word and smaller quantities. */
4262 #define T_MODES (D_MODES | (1 << (int) T_MODE) | (1 << (int) TF_MODE))
4264 /* Modes for 8-word and smaller quantities. */
4265 #define O_MODES (T_MODES | (1 << (int) O_MODE) | (1 << (int) OF_MODE))
4267 /* Modes for single-float quantities. We must allow any single word or
4268 smaller quantity. This is because the fix/float conversion instructions
4269 take integer inputs/outputs from the float registers. */
4270 #define SF_MODES (S_MODES)
4272 /* Modes for double-float and smaller quantities. */
4273 #define DF_MODES (D_MODES)
4275 /* Modes for quad-float and smaller quantities. */
4276 #define TF_MODES (DF_MODES | (1 << (int) TF_MODE))
4278 /* Modes for quad-float pairs and smaller quantities. */
4279 #define OF_MODES (TF_MODES | (1 << (int) OF_MODE))
4281 /* Modes for double-float only quantities. */
4282 #define DF_MODES_NO_S ((1 << (int) D_MODE) | (1 << (int) DF_MODE))
4284 /* Modes for quad-float and double-float only quantities. */
4285 #define TF_MODES_NO_S (DF_MODES_NO_S | (1 << (int) TF_MODE))
4287 /* Modes for quad-float pairs and double-float only quantities. */
4288 #define OF_MODES_NO_S (TF_MODES_NO_S | (1 << (int) OF_MODE))
4290 /* Modes for condition codes. */
4291 #define CC_MODES (1 << (int) CC_MODE)
4292 #define CCFP_MODES (1 << (int) CCFP_MODE)
4294 /* Value is 1 if register/mode pair is acceptable on sparc.
4295 The funny mixture of D and T modes is because integer operations
4296 do not specially operate on tetra quantities, so non-quad-aligned
4297 registers can hold quadword quantities (except %o4 and %i4 because
4298 they cross fixed registers). */
4300 /* This points to either the 32 bit or the 64 bit version. */
4301 const int *hard_regno_mode_classes;
4303 static const int hard_32bit_mode_classes[] = {
4304 S_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
4305 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
4306 T_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
4307 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
4309 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4310 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4311 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4312 OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
4314 /* FP regs f32 to f63. Only the even numbered registers actually exist,
4315 and none can hold SFmode/SImode values. */
4316 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4317 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4318 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4319 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4322 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
4324 /* %icc, %sfp, %gsr */
4325 CC_MODES, 0, D_MODES
4328 static const int hard_64bit_mode_classes[] = {
4329 D_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4330 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4331 T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4332 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4334 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4335 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4336 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4337 OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
4339 /* FP regs f32 to f63. Only the even numbered registers actually exist,
4340 and none can hold SFmode/SImode values. */
4341 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4342 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4343 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4344 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4347 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
4349 /* %icc, %sfp, %gsr */
4350 CC_MODES, 0, D_MODES
4353 int sparc_mode_class [NUM_MACHINE_MODES];
4355 enum reg_class sparc_regno_reg_class[FIRST_PSEUDO_REGISTER];
4358 sparc_init_modes (void)
4362 for (i = 0; i < NUM_MACHINE_MODES; i++)
4364 switch (GET_MODE_CLASS (i))
4367 case MODE_PARTIAL_INT:
4368 case MODE_COMPLEX_INT:
4369 if (GET_MODE_SIZE (i) <= 4)
4370 sparc_mode_class[i] = 1 << (int) S_MODE;
4371 else if (GET_MODE_SIZE (i) == 8)
4372 sparc_mode_class[i] = 1 << (int) D_MODE;
4373 else if (GET_MODE_SIZE (i) == 16)
4374 sparc_mode_class[i] = 1 << (int) T_MODE;
4375 else if (GET_MODE_SIZE (i) == 32)
4376 sparc_mode_class[i] = 1 << (int) O_MODE;
4378 sparc_mode_class[i] = 0;
4380 case MODE_VECTOR_INT:
4381 if (GET_MODE_SIZE (i) <= 4)
4382 sparc_mode_class[i] = 1 << (int)SF_MODE;
4383 else if (GET_MODE_SIZE (i) == 8)
4384 sparc_mode_class[i] = 1 << (int)DF_MODE;
4387 case MODE_COMPLEX_FLOAT:
4388 if (GET_MODE_SIZE (i) <= 4)
4389 sparc_mode_class[i] = 1 << (int) SF_MODE;
4390 else if (GET_MODE_SIZE (i) == 8)
4391 sparc_mode_class[i] = 1 << (int) DF_MODE;
4392 else if (GET_MODE_SIZE (i) == 16)
4393 sparc_mode_class[i] = 1 << (int) TF_MODE;
4394 else if (GET_MODE_SIZE (i) == 32)
4395 sparc_mode_class[i] = 1 << (int) OF_MODE;
4397 sparc_mode_class[i] = 0;
4400 if (i == (int) CCFPmode || i == (int) CCFPEmode)
4401 sparc_mode_class[i] = 1 << (int) CCFP_MODE;
4403 sparc_mode_class[i] = 1 << (int) CC_MODE;
4406 sparc_mode_class[i] = 0;
4412 hard_regno_mode_classes = hard_64bit_mode_classes;
4414 hard_regno_mode_classes = hard_32bit_mode_classes;
4416 /* Initialize the array used by REGNO_REG_CLASS. */
4417 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
4419 if (i < 16 && TARGET_V8PLUS)
4420 sparc_regno_reg_class[i] = I64_REGS;
4421 else if (i < 32 || i == FRAME_POINTER_REGNUM)
4422 sparc_regno_reg_class[i] = GENERAL_REGS;
4424 sparc_regno_reg_class[i] = FP_REGS;
4426 sparc_regno_reg_class[i] = EXTRA_FP_REGS;
4428 sparc_regno_reg_class[i] = FPCC_REGS;
4430 sparc_regno_reg_class[i] = NO_REGS;
4434 /* Return whether REGNO, a global or FP register, must be saved/restored. */
4437 save_global_or_fp_reg_p (unsigned int regno,
4438 int leaf_function ATTRIBUTE_UNUSED)
4440 return !call_used_regs[regno] && df_regs_ever_live_p (regno);
4443 /* Return whether the return address register (%i7) is needed. */
4446 return_addr_reg_needed_p (int leaf_function)
4448 /* If it is live, for example because of __builtin_return_address (0). */
4449 if (df_regs_ever_live_p (RETURN_ADDR_REGNUM))
4452 /* Otherwise, it is needed as save register if %o7 is clobbered. */
4454 /* Loading the GOT register clobbers %o7. */
4455 || crtl->uses_pic_offset_table
4456 || df_regs_ever_live_p (INCOMING_RETURN_ADDR_REGNUM))
4462 /* Return whether REGNO, a local or in register, must be saved/restored. */
4465 save_local_or_in_reg_p (unsigned int regno, int leaf_function)
4467 /* General case: call-saved registers live at some point. */
4468 if (!call_used_regs[regno] && df_regs_ever_live_p (regno))
4471 /* Frame pointer register (%fp) if needed. */
4472 if (regno == HARD_FRAME_POINTER_REGNUM && frame_pointer_needed)
4475 /* Return address register (%i7) if needed. */
4476 if (regno == RETURN_ADDR_REGNUM && return_addr_reg_needed_p (leaf_function))
4479 /* GOT register (%l7) if needed. */
4480 if (regno == PIC_OFFSET_TABLE_REGNUM && crtl->uses_pic_offset_table)
4483 /* If the function accesses prior frames, the frame pointer and the return
4484 address of the previous frame must be saved on the stack. */
4485 if (crtl->accesses_prior_frames
4486 && (regno == HARD_FRAME_POINTER_REGNUM || regno == RETURN_ADDR_REGNUM))
4492 /* Compute the frame size required by the function. This function is called
4493 during the reload pass and also by sparc_expand_prologue. */
4496 sparc_compute_frame_size (HOST_WIDE_INT size, int leaf_function)
4498 HOST_WIDE_INT frame_size, apparent_frame_size;
4499 int args_size, n_global_fp_regs = 0;
4500 bool save_local_in_regs_p = false;
4503 /* If the function allocates dynamic stack space, the dynamic offset is
4504 computed early and contains REG_PARM_STACK_SPACE, so we need to cope. */
4505 if (leaf_function && !cfun->calls_alloca)
4508 args_size = crtl->outgoing_args_size + REG_PARM_STACK_SPACE (cfun->decl);
4510 /* Calculate space needed for global registers. */
4512 for (i = 0; i < 8; i++)
4513 if (save_global_or_fp_reg_p (i, 0))
4514 n_global_fp_regs += 2;
4516 for (i = 0; i < 8; i += 2)
4517 if (save_global_or_fp_reg_p (i, 0) || save_global_or_fp_reg_p (i + 1, 0))
4518 n_global_fp_regs += 2;
4520 /* In the flat window model, find out which local and in registers need to
4521 be saved. We don't reserve space in the current frame for them as they
4522 will be spilled into the register window save area of the caller's frame.
4523 However, as soon as we use this register window save area, we must create
4524 that of the current frame to make it the live one. */
4526 for (i = 16; i < 32; i++)
4527 if (save_local_or_in_reg_p (i, leaf_function))
4529 save_local_in_regs_p = true;
4533 /* Calculate space needed for FP registers. */
4534 for (i = 32; i < (TARGET_V9 ? 96 : 64); i += 2)
4535 if (save_global_or_fp_reg_p (i, 0) || save_global_or_fp_reg_p (i + 1, 0))
4536 n_global_fp_regs += 2;
4539 && n_global_fp_regs == 0
4541 && !save_local_in_regs_p)
4542 frame_size = apparent_frame_size = 0;
4545 /* We subtract STARTING_FRAME_OFFSET, remember it's negative. */
4546 apparent_frame_size = (size - STARTING_FRAME_OFFSET + 7) & -8;
4547 apparent_frame_size += n_global_fp_regs * 4;
4549 /* We need to add the size of the outgoing argument area. */
4550 frame_size = apparent_frame_size + ((args_size + 7) & -8);
4552 /* And that of the register window save area. */
4553 frame_size += FIRST_PARM_OFFSET (cfun->decl);
4555 /* Finally, bump to the appropriate alignment. */
4556 frame_size = SPARC_STACK_ALIGN (frame_size);
4559 /* Set up values for use in prologue and epilogue. */
4560 sparc_frame_size = frame_size;
4561 sparc_apparent_frame_size = apparent_frame_size;
4562 sparc_n_global_fp_regs = n_global_fp_regs;
4563 sparc_save_local_in_regs_p = save_local_in_regs_p;
4568 /* Implement the macro INITIAL_ELIMINATION_OFFSET, return the OFFSET. */
4571 sparc_initial_elimination_offset (int to)
4575 if (to == STACK_POINTER_REGNUM)
4576 offset = sparc_compute_frame_size (get_frame_size (), crtl->is_leaf);
4580 offset += SPARC_STACK_BIAS;
4584 /* Output any necessary .register pseudo-ops. */
4587 sparc_output_scratch_registers (FILE *file ATTRIBUTE_UNUSED)
4589 #ifdef HAVE_AS_REGISTER_PSEUDO_OP
4595 /* Check if %g[2367] were used without
4596 .register being printed for them already. */
4597 for (i = 2; i < 8; i++)
4599 if (df_regs_ever_live_p (i)
4600 && ! sparc_hard_reg_printed [i])
4602 sparc_hard_reg_printed [i] = 1;
4603 /* %g7 is used as TLS base register, use #ignore
4604 for it instead of #scratch. */
4605 fprintf (file, "\t.register\t%%g%d, #%s\n", i,
4606 i == 7 ? "ignore" : "scratch");
4613 #define PROBE_INTERVAL (1 << STACK_CHECK_PROBE_INTERVAL_EXP)
4615 #if PROBE_INTERVAL > 4096
4616 #error Cannot use indexed addressing mode for stack probing
4619 /* Emit code to probe a range of stack addresses from FIRST to FIRST+SIZE,
4620 inclusive. These are offsets from the current stack pointer.
4622 Note that we don't use the REG+REG addressing mode for the probes because
4623 of the stack bias in 64-bit mode. And it doesn't really buy us anything
4624 so the advantages of having a single code win here. */
4627 sparc_emit_probe_stack_range (HOST_WIDE_INT first, HOST_WIDE_INT size)
4629 rtx g1 = gen_rtx_REG (Pmode, 1);
4631 /* See if we have a constant small number of probes to generate. If so,
4632 that's the easy case. */
4633 if (size <= PROBE_INTERVAL)
4635 emit_move_insn (g1, GEN_INT (first));
4636 emit_insn (gen_rtx_SET (VOIDmode, g1,
4637 gen_rtx_MINUS (Pmode, stack_pointer_rtx, g1)));
4638 emit_stack_probe (plus_constant (Pmode, g1, -size));
4641 /* The run-time loop is made up of 10 insns in the generic case while the
4642 compile-time loop is made up of 4+2*(n-2) insns for n # of intervals. */
4643 else if (size <= 5 * PROBE_INTERVAL)
4647 emit_move_insn (g1, GEN_INT (first + PROBE_INTERVAL));
4648 emit_insn (gen_rtx_SET (VOIDmode, g1,
4649 gen_rtx_MINUS (Pmode, stack_pointer_rtx, g1)));
4650 emit_stack_probe (g1);
4652 /* Probe at FIRST + N * PROBE_INTERVAL for values of N from 2 until
4653 it exceeds SIZE. If only two probes are needed, this will not
4654 generate any code. Then probe at FIRST + SIZE. */
4655 for (i = 2 * PROBE_INTERVAL; i < size; i += PROBE_INTERVAL)
4657 emit_insn (gen_rtx_SET (VOIDmode, g1,
4658 plus_constant (Pmode, g1, -PROBE_INTERVAL)));
4659 emit_stack_probe (g1);
4662 emit_stack_probe (plus_constant (Pmode, g1,
4663 (i - PROBE_INTERVAL) - size));
4666 /* Otherwise, do the same as above, but in a loop. Note that we must be
4667 extra careful with variables wrapping around because we might be at
4668 the very top (or the very bottom) of the address space and we have
4669 to be able to handle this case properly; in particular, we use an
4670 equality test for the loop condition. */
4673 HOST_WIDE_INT rounded_size;
4674 rtx g4 = gen_rtx_REG (Pmode, 4);
4676 emit_move_insn (g1, GEN_INT (first));
4679 /* Step 1: round SIZE to the previous multiple of the interval. */
4681 rounded_size = size & -PROBE_INTERVAL;
4682 emit_move_insn (g4, GEN_INT (rounded_size));
4685 /* Step 2: compute initial and final value of the loop counter. */
4687 /* TEST_ADDR = SP + FIRST. */
4688 emit_insn (gen_rtx_SET (VOIDmode, g1,
4689 gen_rtx_MINUS (Pmode, stack_pointer_rtx, g1)));
4691 /* LAST_ADDR = SP + FIRST + ROUNDED_SIZE. */
4692 emit_insn (gen_rtx_SET (VOIDmode, g4, gen_rtx_MINUS (Pmode, g1, g4)));
4697 while (TEST_ADDR != LAST_ADDR)
4699 TEST_ADDR = TEST_ADDR + PROBE_INTERVAL
4703 probes at FIRST + N * PROBE_INTERVAL for values of N from 1
4704 until it is equal to ROUNDED_SIZE. */
4707 emit_insn (gen_probe_stack_rangedi (g1, g1, g4));
4709 emit_insn (gen_probe_stack_rangesi (g1, g1, g4));
4712 /* Step 4: probe at FIRST + SIZE if we cannot assert at compile-time
4713 that SIZE is equal to ROUNDED_SIZE. */
4715 if (size != rounded_size)
4716 emit_stack_probe (plus_constant (Pmode, g4, rounded_size - size));
4719 /* Make sure nothing is scheduled before we are done. */
4720 emit_insn (gen_blockage ());
4723 /* Probe a range of stack addresses from REG1 to REG2 inclusive. These are
4724 absolute addresses. */
4727 output_probe_stack_range (rtx reg1, rtx reg2)
4729 static int labelno = 0;
4730 char loop_lab[32], end_lab[32];
4733 ASM_GENERATE_INTERNAL_LABEL (loop_lab, "LPSRL", labelno);
4734 ASM_GENERATE_INTERNAL_LABEL (end_lab, "LPSRE", labelno++);
4736 ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, loop_lab);
4738 /* Jump to END_LAB if TEST_ADDR == LAST_ADDR. */
4741 output_asm_insn ("cmp\t%0, %1", xops);
4743 fputs ("\tbe,pn\t%xcc,", asm_out_file);
4745 fputs ("\tbe\t", asm_out_file);
4746 assemble_name_raw (asm_out_file, end_lab);
4747 fputc ('\n', asm_out_file);
4749 /* TEST_ADDR = TEST_ADDR + PROBE_INTERVAL. */
4750 xops[1] = GEN_INT (-PROBE_INTERVAL);
4751 output_asm_insn (" add\t%0, %1, %0", xops);
4753 /* Probe at TEST_ADDR and branch. */
4755 fputs ("\tba,pt\t%xcc,", asm_out_file);
4757 fputs ("\tba\t", asm_out_file);
4758 assemble_name_raw (asm_out_file, loop_lab);
4759 fputc ('\n', asm_out_file);
4760 xops[1] = GEN_INT (SPARC_STACK_BIAS);
4761 output_asm_insn (" st\t%%g0, [%0+%1]", xops);
4763 ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, end_lab);
4768 /* Emit code to save/restore registers from LOW to HIGH at BASE+OFFSET as
4769 needed. LOW is supposed to be double-word aligned for 32-bit registers.
4770 SAVE_P decides whether a register must be saved/restored. ACTION_TRUE
4771 is the action to be performed if SAVE_P returns true and ACTION_FALSE
4772 the action to be performed if it returns false. Return the new offset. */
4774 typedef bool (*sorr_pred_t) (unsigned int, int);
4775 typedef enum { SORR_NONE, SORR_ADVANCE, SORR_SAVE, SORR_RESTORE } sorr_act_t;
4778 emit_save_or_restore_regs (unsigned int low, unsigned int high, rtx base,
4779 int offset, int leaf_function, sorr_pred_t save_p,
4780 sorr_act_t action_true, sorr_act_t action_false)
4785 if (TARGET_ARCH64 && high <= 32)
4789 for (i = low; i < high; i++)
4791 if (save_p (i, leaf_function))
4793 mem = gen_frame_mem (DImode, plus_constant (Pmode,
4795 if (action_true == SORR_SAVE)
4797 insn = emit_move_insn (mem, gen_rtx_REG (DImode, i));
4798 RTX_FRAME_RELATED_P (insn) = 1;
4800 else /* action_true == SORR_RESTORE */
4802 /* The frame pointer must be restored last since its old
4803 value may be used as base address for the frame. This
4804 is problematic in 64-bit mode only because of the lack
4805 of double-word load instruction. */
4806 if (i == HARD_FRAME_POINTER_REGNUM)
4809 emit_move_insn (gen_rtx_REG (DImode, i), mem);
4813 else if (action_false == SORR_ADVANCE)
4819 mem = gen_frame_mem (DImode, plus_constant (Pmode, base, fp_offset));
4820 emit_move_insn (hard_frame_pointer_rtx, mem);
4825 for (i = low; i < high; i += 2)
4827 bool reg0 = save_p (i, leaf_function);
4828 bool reg1 = save_p (i + 1, leaf_function);
4829 enum machine_mode mode;
4834 mode = SPARC_INT_REG_P (i) ? DImode : DFmode;
4839 mode = SPARC_INT_REG_P (i) ? SImode : SFmode;
4844 mode = SPARC_INT_REG_P (i) ? SImode : SFmode;
4850 if (action_false == SORR_ADVANCE)
4855 mem = gen_frame_mem (mode, plus_constant (Pmode, base, offset));
4856 if (action_true == SORR_SAVE)
4858 insn = emit_move_insn (mem, gen_rtx_REG (mode, regno));
4859 RTX_FRAME_RELATED_P (insn) = 1;
4863 mem = gen_frame_mem (SImode, plus_constant (Pmode, base,
4865 set1 = gen_rtx_SET (VOIDmode, mem,
4866 gen_rtx_REG (SImode, regno));
4867 RTX_FRAME_RELATED_P (set1) = 1;
4869 = gen_frame_mem (SImode, plus_constant (Pmode, base,
4871 set2 = gen_rtx_SET (VOIDmode, mem,
4872 gen_rtx_REG (SImode, regno + 1));
4873 RTX_FRAME_RELATED_P (set2) = 1;
4874 add_reg_note (insn, REG_FRAME_RELATED_EXPR,
4875 gen_rtx_PARALLEL (VOIDmode,
4876 gen_rtvec (2, set1, set2)));
4879 else /* action_true == SORR_RESTORE */
4880 emit_move_insn (gen_rtx_REG (mode, regno), mem);
4882 /* Always preserve double-word alignment. */
4883 offset = (offset + 8) & -8;
4890 /* Emit code to adjust BASE to OFFSET. Return the new base. */
4893 emit_adjust_base_to_offset (rtx base, int offset)
4895 /* ??? This might be optimized a little as %g1 might already have a
4896 value close enough that a single add insn will do. */
4897 /* ??? Although, all of this is probably only a temporary fix because
4898 if %g1 can hold a function result, then sparc_expand_epilogue will
4899 lose (the result will be clobbered). */
4900 rtx new_base = gen_rtx_REG (Pmode, 1);
4901 emit_move_insn (new_base, GEN_INT (offset));
4902 emit_insn (gen_rtx_SET (VOIDmode,
4903 new_base, gen_rtx_PLUS (Pmode, base, new_base)));
4907 /* Emit code to save/restore call-saved global and FP registers. */
4910 emit_save_or_restore_global_fp_regs (rtx base, int offset, sorr_act_t action)
4912 if (offset < -4096 || offset + sparc_n_global_fp_regs * 4 > 4095)
4914 base = emit_adjust_base_to_offset (base, offset);
4919 = emit_save_or_restore_regs (0, 8, base, offset, 0,
4920 save_global_or_fp_reg_p, action, SORR_NONE);
4921 emit_save_or_restore_regs (32, TARGET_V9 ? 96 : 64, base, offset, 0,
4922 save_global_or_fp_reg_p, action, SORR_NONE);
4925 /* Emit code to save/restore call-saved local and in registers. */
4928 emit_save_or_restore_local_in_regs (rtx base, int offset, sorr_act_t action)
4930 if (offset < -4096 || offset + 16 * UNITS_PER_WORD > 4095)
4932 base = emit_adjust_base_to_offset (base, offset);
4936 emit_save_or_restore_regs (16, 32, base, offset, sparc_leaf_function_p,
4937 save_local_or_in_reg_p, action, SORR_ADVANCE);
4940 /* Emit a window_save insn. */
4943 emit_window_save (rtx increment)
4945 rtx insn = emit_insn (gen_window_save (increment));
4946 RTX_FRAME_RELATED_P (insn) = 1;
4948 /* The incoming return address (%o7) is saved in %i7. */
4949 add_reg_note (insn, REG_CFA_REGISTER,
4950 gen_rtx_SET (VOIDmode,
4951 gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM),
4953 INCOMING_RETURN_ADDR_REGNUM)));
4955 /* The window save event. */
4956 add_reg_note (insn, REG_CFA_WINDOW_SAVE, const0_rtx);
4958 /* The CFA is %fp, the hard frame pointer. */
4959 add_reg_note (insn, REG_CFA_DEF_CFA,
4960 plus_constant (Pmode, hard_frame_pointer_rtx,
4961 INCOMING_FRAME_SP_OFFSET));
4966 /* Generate an increment for the stack pointer. */
4969 gen_stack_pointer_inc (rtx increment)
4971 return gen_rtx_SET (VOIDmode,
4973 gen_rtx_PLUS (Pmode,
4978 /* Generate a decrement for the stack pointer. */
4981 gen_stack_pointer_dec (rtx decrement)
4983 return gen_rtx_SET (VOIDmode,
4985 gen_rtx_MINUS (Pmode,
4990 /* Expand the function prologue. The prologue is responsible for reserving
4991 storage for the frame, saving the call-saved registers and loading the
4992 GOT register if needed. */
4995 sparc_expand_prologue (void)
5000 /* Compute a snapshot of crtl->uses_only_leaf_regs. Relying
5001 on the final value of the flag means deferring the prologue/epilogue
5002 expansion until just before the second scheduling pass, which is too
5003 late to emit multiple epilogues or return insns.
5005 Of course we are making the assumption that the value of the flag
5006 will not change between now and its final value. Of the three parts
5007 of the formula, only the last one can reasonably vary. Let's take a
5008 closer look, after assuming that the first two ones are set to true
5009 (otherwise the last value is effectively silenced).
5011 If only_leaf_regs_used returns false, the global predicate will also
5012 be false so the actual frame size calculated below will be positive.
5013 As a consequence, the save_register_window insn will be emitted in
5014 the instruction stream; now this insn explicitly references %fp
5015 which is not a leaf register so only_leaf_regs_used will always
5016 return false subsequently.
5018 If only_leaf_regs_used returns true, we hope that the subsequent
5019 optimization passes won't cause non-leaf registers to pop up. For
5020 example, the regrename pass has special provisions to not rename to
5021 non-leaf registers in a leaf function. */
5022 sparc_leaf_function_p
5023 = optimize > 0 && crtl->is_leaf && only_leaf_regs_used ();
5025 size = sparc_compute_frame_size (get_frame_size(), sparc_leaf_function_p);
5027 if (flag_stack_usage_info)
5028 current_function_static_stack_size = size;
5030 if (flag_stack_check == STATIC_BUILTIN_STACK_CHECK && size)
5031 sparc_emit_probe_stack_range (STACK_CHECK_PROTECT, size);
5035 else if (sparc_leaf_function_p)
5037 rtx size_int_rtx = GEN_INT (-size);
5040 insn = emit_insn (gen_stack_pointer_inc (size_int_rtx));
5041 else if (size <= 8192)
5043 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (-4096)));
5044 RTX_FRAME_RELATED_P (insn) = 1;
5046 /* %sp is still the CFA register. */
5047 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (4096 - size)));
5051 rtx size_rtx = gen_rtx_REG (Pmode, 1);
5052 emit_move_insn (size_rtx, size_int_rtx);
5053 insn = emit_insn (gen_stack_pointer_inc (size_rtx));
5054 add_reg_note (insn, REG_FRAME_RELATED_EXPR,
5055 gen_stack_pointer_inc (size_int_rtx));
5058 RTX_FRAME_RELATED_P (insn) = 1;
5062 rtx size_int_rtx = GEN_INT (-size);
5065 emit_window_save (size_int_rtx);
5066 else if (size <= 8192)
5068 emit_window_save (GEN_INT (-4096));
5070 /* %sp is not the CFA register anymore. */
5071 emit_insn (gen_stack_pointer_inc (GEN_INT (4096 - size)));
5073 /* Make sure no %fp-based store is issued until after the frame is
5074 established. The offset between the frame pointer and the stack
5075 pointer is calculated relative to the value of the stack pointer
5076 at the end of the function prologue, and moving instructions that
5077 access the stack via the frame pointer between the instructions
5078 that decrement the stack pointer could result in accessing the
5079 register window save area, which is volatile. */
5080 emit_insn (gen_frame_blockage ());
5084 rtx size_rtx = gen_rtx_REG (Pmode, 1);
5085 emit_move_insn (size_rtx, size_int_rtx);
5086 emit_window_save (size_rtx);
5090 if (sparc_leaf_function_p)
5092 sparc_frame_base_reg = stack_pointer_rtx;
5093 sparc_frame_base_offset = size + SPARC_STACK_BIAS;
5097 sparc_frame_base_reg = hard_frame_pointer_rtx;
5098 sparc_frame_base_offset = SPARC_STACK_BIAS;
5101 if (sparc_n_global_fp_regs > 0)
5102 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg,
5103 sparc_frame_base_offset
5104 - sparc_apparent_frame_size,
5107 /* Load the GOT register if needed. */
5108 if (crtl->uses_pic_offset_table)
5109 load_got_register ();
5111 /* Advertise that the data calculated just above are now valid. */
5112 sparc_prologue_data_valid_p = true;
5115 /* Expand the function prologue. The prologue is responsible for reserving
5116 storage for the frame, saving the call-saved registers and loading the
5117 GOT register if needed. */
5120 sparc_flat_expand_prologue (void)
5125 sparc_leaf_function_p = optimize > 0 && crtl->is_leaf;
5127 size = sparc_compute_frame_size (get_frame_size(), sparc_leaf_function_p);
5129 if (flag_stack_usage_info)
5130 current_function_static_stack_size = size;
5132 if (flag_stack_check == STATIC_BUILTIN_STACK_CHECK && size)
5133 sparc_emit_probe_stack_range (STACK_CHECK_PROTECT, size);
5135 if (sparc_save_local_in_regs_p)
5136 emit_save_or_restore_local_in_regs (stack_pointer_rtx, SPARC_STACK_BIAS,
5143 rtx size_int_rtx, size_rtx;
5145 size_rtx = size_int_rtx = GEN_INT (-size);
5147 /* We establish the frame (i.e. decrement the stack pointer) first, even
5148 if we use a frame pointer, because we cannot clobber any call-saved
5149 registers, including the frame pointer, if we haven't created a new
5150 register save area, for the sake of compatibility with the ABI. */
5152 insn = emit_insn (gen_stack_pointer_inc (size_int_rtx));
5153 else if (size <= 8192 && !frame_pointer_needed)
5155 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (-4096)));
5156 RTX_FRAME_RELATED_P (insn) = 1;
5157 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (4096 - size)));
5161 size_rtx = gen_rtx_REG (Pmode, 1);
5162 emit_move_insn (size_rtx, size_int_rtx);
5163 insn = emit_insn (gen_stack_pointer_inc (size_rtx));
5164 add_reg_note (insn, REG_CFA_ADJUST_CFA,
5165 gen_stack_pointer_inc (size_int_rtx));
5167 RTX_FRAME_RELATED_P (insn) = 1;
5169 /* Ensure nothing is scheduled until after the frame is established. */
5170 emit_insn (gen_blockage ());
5172 if (frame_pointer_needed)
5174 insn = emit_insn (gen_rtx_SET (VOIDmode, hard_frame_pointer_rtx,
5175 gen_rtx_MINUS (Pmode,
5178 RTX_FRAME_RELATED_P (insn) = 1;
5180 add_reg_note (insn, REG_CFA_ADJUST_CFA,
5181 gen_rtx_SET (VOIDmode, hard_frame_pointer_rtx,
5182 plus_constant (Pmode, stack_pointer_rtx,
5186 if (return_addr_reg_needed_p (sparc_leaf_function_p))
5188 rtx o7 = gen_rtx_REG (Pmode, INCOMING_RETURN_ADDR_REGNUM);
5189 rtx i7 = gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM);
5191 insn = emit_move_insn (i7, o7);
5192 RTX_FRAME_RELATED_P (insn) = 1;
5194 add_reg_note (insn, REG_CFA_REGISTER,
5195 gen_rtx_SET (VOIDmode, i7, o7));
5197 /* Prevent this instruction from ever being considered dead,
5198 even if this function has no epilogue. */
5203 if (frame_pointer_needed)
5205 sparc_frame_base_reg = hard_frame_pointer_rtx;
5206 sparc_frame_base_offset = SPARC_STACK_BIAS;
5210 sparc_frame_base_reg = stack_pointer_rtx;
5211 sparc_frame_base_offset = size + SPARC_STACK_BIAS;
5214 if (sparc_n_global_fp_regs > 0)
5215 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg,
5216 sparc_frame_base_offset
5217 - sparc_apparent_frame_size,
5220 /* Load the GOT register if needed. */
5221 if (crtl->uses_pic_offset_table)
5222 load_got_register ();
5224 /* Advertise that the data calculated just above are now valid. */
5225 sparc_prologue_data_valid_p = true;
5228 /* This function generates the assembly code for function entry, which boils
5229 down to emitting the necessary .register directives. */
5232 sparc_asm_function_prologue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
5234 /* Check that the assumption we made in sparc_expand_prologue is valid. */
5236 gcc_assert (sparc_leaf_function_p == crtl->uses_only_leaf_regs);
5238 sparc_output_scratch_registers (file);
5241 /* Expand the function epilogue, either normal or part of a sibcall.
5242 We emit all the instructions except the return or the call. */
5245 sparc_expand_epilogue (bool for_eh)
5247 HOST_WIDE_INT size = sparc_frame_size;
5249 if (sparc_n_global_fp_regs > 0)
5250 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg,
5251 sparc_frame_base_offset
5252 - sparc_apparent_frame_size,
5255 if (size == 0 || for_eh)
5257 else if (sparc_leaf_function_p)
5260 emit_insn (gen_stack_pointer_dec (GEN_INT (-size)));
5261 else if (size <= 8192)
5263 emit_insn (gen_stack_pointer_dec (GEN_INT (-4096)));
5264 emit_insn (gen_stack_pointer_dec (GEN_INT (4096 - size)));
5268 rtx reg = gen_rtx_REG (Pmode, 1);
5269 emit_move_insn (reg, GEN_INT (-size));
5270 emit_insn (gen_stack_pointer_dec (reg));
5275 /* Expand the function epilogue, either normal or part of a sibcall.
5276 We emit all the instructions except the return or the call. */
5279 sparc_flat_expand_epilogue (bool for_eh)
5281 HOST_WIDE_INT size = sparc_frame_size;
5283 if (sparc_n_global_fp_regs > 0)
5284 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg,
5285 sparc_frame_base_offset
5286 - sparc_apparent_frame_size,
5289 /* If we have a frame pointer, we'll need both to restore it before the
5290 frame is destroyed and use its current value in destroying the frame.
5291 Since we don't have an atomic way to do that in the flat window model,
5292 we save the current value into a temporary register (%g1). */
5293 if (frame_pointer_needed && !for_eh)
5294 emit_move_insn (gen_rtx_REG (Pmode, 1), hard_frame_pointer_rtx);
5296 if (return_addr_reg_needed_p (sparc_leaf_function_p))
5297 emit_move_insn (gen_rtx_REG (Pmode, INCOMING_RETURN_ADDR_REGNUM),
5298 gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM));
5300 if (sparc_save_local_in_regs_p)
5301 emit_save_or_restore_local_in_regs (sparc_frame_base_reg,
5302 sparc_frame_base_offset,
5305 if (size == 0 || for_eh)
5307 else if (frame_pointer_needed)
5309 /* Make sure the frame is destroyed after everything else is done. */
5310 emit_insn (gen_blockage ());
5312 emit_move_insn (stack_pointer_rtx, gen_rtx_REG (Pmode, 1));
5317 emit_insn (gen_blockage ());
5320 emit_insn (gen_stack_pointer_dec (GEN_INT (-size)));
5321 else if (size <= 8192)
5323 emit_insn (gen_stack_pointer_dec (GEN_INT (-4096)));
5324 emit_insn (gen_stack_pointer_dec (GEN_INT (4096 - size)));
5328 rtx reg = gen_rtx_REG (Pmode, 1);
5329 emit_move_insn (reg, GEN_INT (-size));
5330 emit_insn (gen_stack_pointer_dec (reg));
5335 /* Return true if it is appropriate to emit `return' instructions in the
5336 body of a function. */
5339 sparc_can_use_return_insn_p (void)
5341 return sparc_prologue_data_valid_p
5342 && sparc_n_global_fp_regs == 0
5344 ? (sparc_frame_size == 0 && !sparc_save_local_in_regs_p)
5345 : (sparc_frame_size == 0 || !sparc_leaf_function_p);
5348 /* This function generates the assembly code for function exit. */
5351 sparc_asm_function_epilogue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
5353 /* If the last two instructions of a function are "call foo; dslot;"
5354 the return address might point to the first instruction in the next
5355 function and we have to output a dummy nop for the sake of sane
5356 backtraces in such cases. This is pointless for sibling calls since
5357 the return address is explicitly adjusted. */
5359 rtx insn, last_real_insn;
5361 insn = get_last_insn ();
5363 last_real_insn = prev_real_insn (insn);
5365 && GET_CODE (last_real_insn) == INSN
5366 && GET_CODE (PATTERN (last_real_insn)) == SEQUENCE)
5367 last_real_insn = XVECEXP (PATTERN (last_real_insn), 0, 0);
5370 && CALL_P (last_real_insn)
5371 && !SIBLING_CALL_P (last_real_insn))
5372 fputs("\tnop\n", file);
5374 sparc_output_deferred_case_vectors ();
5377 /* Output a 'restore' instruction. */
5380 output_restore (rtx pat)
5386 fputs ("\t restore\n", asm_out_file);
5390 gcc_assert (GET_CODE (pat) == SET);
5392 operands[0] = SET_DEST (pat);
5393 pat = SET_SRC (pat);
5395 switch (GET_CODE (pat))
5398 operands[1] = XEXP (pat, 0);
5399 operands[2] = XEXP (pat, 1);
5400 output_asm_insn (" restore %r1, %2, %Y0", operands);
5403 operands[1] = XEXP (pat, 0);
5404 operands[2] = XEXP (pat, 1);
5405 output_asm_insn (" restore %r1, %%lo(%a2), %Y0", operands);
5408 operands[1] = XEXP (pat, 0);
5409 gcc_assert (XEXP (pat, 1) == const1_rtx);
5410 output_asm_insn (" restore %r1, %r1, %Y0", operands);
5414 output_asm_insn (" restore %%g0, %1, %Y0", operands);
5419 /* Output a return. */
5422 output_return (rtx insn)
5424 if (crtl->calls_eh_return)
5426 /* If the function uses __builtin_eh_return, the eh_return
5427 machinery occupies the delay slot. */
5428 gcc_assert (!final_sequence);
5430 if (flag_delayed_branch)
5432 if (!TARGET_FLAT && TARGET_V9)
5433 fputs ("\treturn\t%i7+8\n", asm_out_file);
5437 fputs ("\trestore\n", asm_out_file);
5439 fputs ("\tjmp\t%o7+8\n", asm_out_file);
5442 fputs ("\t add\t%sp, %g1, %sp\n", asm_out_file);
5447 fputs ("\trestore\n", asm_out_file);
5449 fputs ("\tadd\t%sp, %g1, %sp\n", asm_out_file);
5450 fputs ("\tjmp\t%o7+8\n\t nop\n", asm_out_file);
5453 else if (sparc_leaf_function_p || TARGET_FLAT)
5455 /* This is a leaf or flat function so we don't have to bother restoring
5456 the register window, which frees us from dealing with the convoluted
5457 semantics of restore/return. We simply output the jump to the
5458 return address and the insn in the delay slot (if any). */
5460 return "jmp\t%%o7+%)%#";
5464 /* This is a regular function so we have to restore the register window.
5465 We may have a pending insn for the delay slot, which will be either
5466 combined with the 'restore' instruction or put in the delay slot of
5467 the 'return' instruction. */
5473 delay = NEXT_INSN (insn);
5476 pat = PATTERN (delay);
5478 if (TARGET_V9 && ! epilogue_renumber (&pat, 1))
5480 epilogue_renumber (&pat, 0);
5481 return "return\t%%i7+%)%#";
5485 output_asm_insn ("jmp\t%%i7+%)", NULL);
5486 output_restore (pat);
5487 PATTERN (delay) = gen_blockage ();
5488 INSN_CODE (delay) = -1;
5493 /* The delay slot is empty. */
5495 return "return\t%%i7+%)\n\t nop";
5496 else if (flag_delayed_branch)
5497 return "jmp\t%%i7+%)\n\t restore";
5499 return "restore\n\tjmp\t%%o7+%)\n\t nop";
5506 /* Output a sibling call. */
5509 output_sibcall (rtx insn, rtx call_operand)
5513 gcc_assert (flag_delayed_branch);
5515 operands[0] = call_operand;
5517 if (sparc_leaf_function_p || TARGET_FLAT)
5519 /* This is a leaf or flat function so we don't have to bother restoring
5520 the register window. We simply output the jump to the function and
5521 the insn in the delay slot (if any). */
5523 gcc_assert (!(LEAF_SIBCALL_SLOT_RESERVED_P && final_sequence));
5526 output_asm_insn ("sethi\t%%hi(%a0), %%g1\n\tjmp\t%%g1 + %%lo(%a0)%#",
5529 /* Use or with rs2 %%g0 instead of mov, so that as/ld can optimize
5530 it into branch if possible. */
5531 output_asm_insn ("or\t%%o7, %%g0, %%g1\n\tcall\t%a0, 0\n\t or\t%%g1, %%g0, %%o7",
5536 /* This is a regular function so we have to restore the register window.
5537 We may have a pending insn for the delay slot, which will be combined
5538 with the 'restore' instruction. */
5540 output_asm_insn ("call\t%a0, 0", operands);
5544 rtx delay = NEXT_INSN (insn);
5547 output_restore (PATTERN (delay));
5549 PATTERN (delay) = gen_blockage ();
5550 INSN_CODE (delay) = -1;
5553 output_restore (NULL_RTX);
5559 /* Functions for handling argument passing.
5561 For 32-bit, the first 6 args are normally in registers and the rest are
5562 pushed. Any arg that starts within the first 6 words is at least
5563 partially passed in a register unless its data type forbids.
5565 For 64-bit, the argument registers are laid out as an array of 16 elements
5566 and arguments are added sequentially. The first 6 int args and up to the
5567 first 16 fp args (depending on size) are passed in regs.
5569 Slot Stack Integral Float Float in structure Double Long Double
5570 ---- ----- -------- ----- ------------------ ------ -----------
5571 15 [SP+248] %f31 %f30,%f31 %d30
5572 14 [SP+240] %f29 %f28,%f29 %d28 %q28
5573 13 [SP+232] %f27 %f26,%f27 %d26
5574 12 [SP+224] %f25 %f24,%f25 %d24 %q24
5575 11 [SP+216] %f23 %f22,%f23 %d22
5576 10 [SP+208] %f21 %f20,%f21 %d20 %q20
5577 9 [SP+200] %f19 %f18,%f19 %d18
5578 8 [SP+192] %f17 %f16,%f17 %d16 %q16
5579 7 [SP+184] %f15 %f14,%f15 %d14
5580 6 [SP+176] %f13 %f12,%f13 %d12 %q12
5581 5 [SP+168] %o5 %f11 %f10,%f11 %d10
5582 4 [SP+160] %o4 %f9 %f8,%f9 %d8 %q8
5583 3 [SP+152] %o3 %f7 %f6,%f7 %d6
5584 2 [SP+144] %o2 %f5 %f4,%f5 %d4 %q4
5585 1 [SP+136] %o1 %f3 %f2,%f3 %d2
5586 0 [SP+128] %o0 %f1 %f0,%f1 %d0 %q0
5588 Here SP = %sp if -mno-stack-bias or %sp+stack_bias otherwise.
5590 Integral arguments are always passed as 64-bit quantities appropriately
5593 Passing of floating point values is handled as follows.
5594 If a prototype is in scope:
5595 If the value is in a named argument (i.e. not a stdarg function or a
5596 value not part of the `...') then the value is passed in the appropriate
5598 If the value is part of the `...' and is passed in one of the first 6
5599 slots then the value is passed in the appropriate int reg.
5600 If the value is part of the `...' and is not passed in one of the first 6
5601 slots then the value is passed in memory.
5602 If a prototype is not in scope:
5603 If the value is one of the first 6 arguments the value is passed in the
5604 appropriate integer reg and the appropriate fp reg.
5605 If the value is not one of the first 6 arguments the value is passed in
5606 the appropriate fp reg and in memory.
5609 Summary of the calling conventions implemented by GCC on the SPARC:
5612 size argument return value
5614 small integer <4 int. reg. int. reg.
5615 word 4 int. reg. int. reg.
5616 double word 8 int. reg. int. reg.
5618 _Complex small integer <8 int. reg. int. reg.
5619 _Complex word 8 int. reg. int. reg.
5620 _Complex double word 16 memory int. reg.
5622 vector integer <=8 int. reg. FP reg.
5623 vector integer >8 memory memory
5625 float 4 int. reg. FP reg.
5626 double 8 int. reg. FP reg.
5627 long double 16 memory memory
5629 _Complex float 8 memory FP reg.
5630 _Complex double 16 memory FP reg.
5631 _Complex long double 32 memory FP reg.
5633 vector float any memory memory
5635 aggregate any memory memory
5640 size argument return value
5642 small integer <8 int. reg. int. reg.
5643 word 8 int. reg. int. reg.
5644 double word 16 int. reg. int. reg.
5646 _Complex small integer <16 int. reg. int. reg.
5647 _Complex word 16 int. reg. int. reg.
5648 _Complex double word 32 memory int. reg.
5650 vector integer <=16 FP reg. FP reg.
5651 vector integer 16<s<=32 memory FP reg.
5652 vector integer >32 memory memory
5654 float 4 FP reg. FP reg.
5655 double 8 FP reg. FP reg.
5656 long double 16 FP reg. FP reg.
5658 _Complex float 8 FP reg. FP reg.
5659 _Complex double 16 FP reg. FP reg.
5660 _Complex long double 32 memory FP reg.
5662 vector float <=16 FP reg. FP reg.
5663 vector float 16<s<=32 memory FP reg.
5664 vector float >32 memory memory
5666 aggregate <=16 reg. reg.
5667 aggregate 16<s<=32 memory reg.
5668 aggregate >32 memory memory
5672 Note #1: complex floating-point types follow the extended SPARC ABIs as
5673 implemented by the Sun compiler.
5675 Note #2: integral vector types follow the scalar floating-point types
5676 conventions to match what is implemented by the Sun VIS SDK.
5678 Note #3: floating-point vector types follow the aggregate types
5682 /* Maximum number of int regs for args. */
5683 #define SPARC_INT_ARG_MAX 6
5684 /* Maximum number of fp regs for args. */
5685 #define SPARC_FP_ARG_MAX 16
5687 #define ROUND_ADVANCE(SIZE) (((SIZE) + UNITS_PER_WORD - 1) / UNITS_PER_WORD)
5689 /* Handle the INIT_CUMULATIVE_ARGS macro.
5690 Initialize a variable CUM of type CUMULATIVE_ARGS
5691 for a call to a function whose data type is FNTYPE.
5692 For a library call, FNTYPE is 0. */
5695 init_cumulative_args (struct sparc_args *cum, tree fntype,
5696 rtx libname ATTRIBUTE_UNUSED,
5697 tree fndecl ATTRIBUTE_UNUSED)
5700 cum->prototype_p = fntype && prototype_p (fntype);
5701 cum->libcall_p = fntype == 0;
5704 /* Handle promotion of pointer and integer arguments. */
5706 static enum machine_mode
5707 sparc_promote_function_mode (const_tree type,
5708 enum machine_mode mode,
5710 const_tree fntype ATTRIBUTE_UNUSED,
5711 int for_return ATTRIBUTE_UNUSED)
5713 if (type != NULL_TREE && POINTER_TYPE_P (type))
5715 *punsignedp = POINTERS_EXTEND_UNSIGNED;
5719 /* Integral arguments are passed as full words, as per the ABI. */
5720 if (GET_MODE_CLASS (mode) == MODE_INT
5721 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5727 /* Handle the TARGET_STRICT_ARGUMENT_NAMING target hook. */
5730 sparc_strict_argument_naming (cumulative_args_t ca ATTRIBUTE_UNUSED)
5732 return TARGET_ARCH64 ? true : false;
5735 /* Scan the record type TYPE and return the following predicates:
5736 - INTREGS_P: the record contains at least one field or sub-field
5737 that is eligible for promotion in integer registers.
5738 - FP_REGS_P: the record contains at least one field or sub-field
5739 that is eligible for promotion in floating-point registers.
5740 - PACKED_P: the record contains at least one field that is packed.
5742 Sub-fields are not taken into account for the PACKED_P predicate. */
5745 scan_record_type (const_tree type, int *intregs_p, int *fpregs_p,
5750 for (field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field))
5752 if (TREE_CODE (field) == FIELD_DECL)
5754 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
5755 scan_record_type (TREE_TYPE (field), intregs_p, fpregs_p, 0);
5756 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
5757 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
5763 if (packed_p && DECL_PACKED (field))
5769 /* Compute the slot number to pass an argument in.
5770 Return the slot number or -1 if passing on the stack.
5772 CUM is a variable of type CUMULATIVE_ARGS which gives info about
5773 the preceding args and about the function being called.
5774 MODE is the argument's machine mode.
5775 TYPE is the data type of the argument (as a tree).
5776 This is null for libcalls where that information may
5778 NAMED is nonzero if this argument is a named parameter
5779 (otherwise it is an extra parameter matching an ellipsis).
5780 INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG.
5781 *PREGNO records the register number to use if scalar type.
5782 *PPADDING records the amount of padding needed in words. */
5785 function_arg_slotno (const struct sparc_args *cum, enum machine_mode mode,
5786 const_tree type, bool named, bool incoming_p,
5787 int *pregno, int *ppadding)
5789 int regbase = (incoming_p
5790 ? SPARC_INCOMING_INT_ARG_FIRST
5791 : SPARC_OUTGOING_INT_ARG_FIRST);
5792 int slotno = cum->words;
5793 enum mode_class mclass;
5798 if (type && TREE_ADDRESSABLE (type))
5804 && TYPE_ALIGN (type) % PARM_BOUNDARY != 0)
5807 /* For SPARC64, objects requiring 16-byte alignment get it. */
5809 && (type ? TYPE_ALIGN (type) : GET_MODE_ALIGNMENT (mode)) >= 128
5810 && (slotno & 1) != 0)
5811 slotno++, *ppadding = 1;
5813 mclass = GET_MODE_CLASS (mode);
5814 if (type && TREE_CODE (type) == VECTOR_TYPE)
5816 /* Vector types deserve special treatment because they are
5817 polymorphic wrt their mode, depending upon whether VIS
5818 instructions are enabled. */
5819 if (TREE_CODE (TREE_TYPE (type)) == REAL_TYPE)
5821 /* The SPARC port defines no floating-point vector modes. */
5822 gcc_assert (mode == BLKmode);
5826 /* Integral vector types should either have a vector
5827 mode or an integral mode, because we are guaranteed
5828 by pass_by_reference that their size is not greater
5829 than 16 bytes and TImode is 16-byte wide. */
5830 gcc_assert (mode != BLKmode);
5832 /* Vector integers are handled like floats according to
5834 mclass = MODE_FLOAT;
5841 case MODE_COMPLEX_FLOAT:
5842 case MODE_VECTOR_INT:
5843 if (TARGET_ARCH64 && TARGET_FPU && named)
5845 if (slotno >= SPARC_FP_ARG_MAX)
5847 regno = SPARC_FP_ARG_FIRST + slotno * 2;
5848 /* Arguments filling only one single FP register are
5849 right-justified in the outer double FP register. */
5850 if (GET_MODE_SIZE (mode) <= 4)
5857 case MODE_COMPLEX_INT:
5858 if (slotno >= SPARC_INT_ARG_MAX)
5860 regno = regbase + slotno;
5864 if (mode == VOIDmode)
5865 /* MODE is VOIDmode when generating the actual call. */
5868 gcc_assert (mode == BLKmode);
5872 || (TREE_CODE (type) != VECTOR_TYPE
5873 && TREE_CODE (type) != RECORD_TYPE))
5875 if (slotno >= SPARC_INT_ARG_MAX)
5877 regno = regbase + slotno;
5879 else /* TARGET_ARCH64 && type */
5881 int intregs_p = 0, fpregs_p = 0, packed_p = 0;
5883 /* First see what kinds of registers we would need. */
5884 if (TREE_CODE (type) == VECTOR_TYPE)
5887 scan_record_type (type, &intregs_p, &fpregs_p, &packed_p);
5889 /* The ABI obviously doesn't specify how packed structures
5890 are passed. These are defined to be passed in int regs
5891 if possible, otherwise memory. */
5892 if (packed_p || !named)
5893 fpregs_p = 0, intregs_p = 1;
5895 /* If all arg slots are filled, then must pass on stack. */
5896 if (fpregs_p && slotno >= SPARC_FP_ARG_MAX)
5899 /* If there are only int args and all int arg slots are filled,
5900 then must pass on stack. */
5901 if (!fpregs_p && intregs_p && slotno >= SPARC_INT_ARG_MAX)
5904 /* Note that even if all int arg slots are filled, fp members may
5905 still be passed in regs if such regs are available.
5906 *PREGNO isn't set because there may be more than one, it's up
5907 to the caller to compute them. */
5920 /* Handle recursive register counting for structure field layout. */
5922 struct function_arg_record_value_parms
5924 rtx ret; /* return expression being built. */
5925 int slotno; /* slot number of the argument. */
5926 int named; /* whether the argument is named. */
5927 int regbase; /* regno of the base register. */
5928 int stack; /* 1 if part of the argument is on the stack. */
5929 int intoffset; /* offset of the first pending integer field. */
5930 unsigned int nregs; /* number of words passed in registers. */
5933 static void function_arg_record_value_3
5934 (HOST_WIDE_INT, struct function_arg_record_value_parms *);
5935 static void function_arg_record_value_2
5936 (const_tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
5937 static void function_arg_record_value_1
5938 (const_tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
5939 static rtx function_arg_record_value (const_tree, enum machine_mode, int, int, int);
5940 static rtx function_arg_union_value (int, enum machine_mode, int, int);
5942 /* A subroutine of function_arg_record_value. Traverse the structure
5943 recursively and determine how many registers will be required. */
5946 function_arg_record_value_1 (const_tree type, HOST_WIDE_INT startbitpos,
5947 struct function_arg_record_value_parms *parms,
5952 /* We need to compute how many registers are needed so we can
5953 allocate the PARALLEL but before we can do that we need to know
5954 whether there are any packed fields. The ABI obviously doesn't
5955 specify how structures are passed in this case, so they are
5956 defined to be passed in int regs if possible, otherwise memory,
5957 regardless of whether there are fp values present. */
5960 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5962 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
5969 /* Compute how many registers we need. */
5970 for (field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field))
5972 if (TREE_CODE (field) == FIELD_DECL)
5974 HOST_WIDE_INT bitpos = startbitpos;
5976 if (DECL_SIZE (field) != 0)
5978 if (integer_zerop (DECL_SIZE (field)))
5981 if (host_integerp (bit_position (field), 1))
5982 bitpos += int_bit_position (field);
5985 /* ??? FIXME: else assume zero offset. */
5987 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
5988 function_arg_record_value_1 (TREE_TYPE (field),
5992 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
5993 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
5998 if (parms->intoffset != -1)
6000 unsigned int startbit, endbit;
6001 int intslots, this_slotno;
6003 startbit = parms->intoffset & -BITS_PER_WORD;
6004 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
6006 intslots = (endbit - startbit) / BITS_PER_WORD;
6007 this_slotno = parms->slotno + parms->intoffset
6010 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
6012 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
6013 /* We need to pass this field on the stack. */
6017 parms->nregs += intslots;
6018 parms->intoffset = -1;
6021 /* There's no need to check this_slotno < SPARC_FP_ARG MAX.
6022 If it wasn't true we wouldn't be here. */
6023 if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE
6024 && DECL_MODE (field) == BLKmode)
6025 parms->nregs += TYPE_VECTOR_SUBPARTS (TREE_TYPE (field));
6026 else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
6033 if (parms->intoffset == -1)
6034 parms->intoffset = bitpos;
6040 /* A subroutine of function_arg_record_value. Assign the bits of the
6041 structure between parms->intoffset and bitpos to integer registers. */
6044 function_arg_record_value_3 (HOST_WIDE_INT bitpos,
6045 struct function_arg_record_value_parms *parms)
6047 enum machine_mode mode;
6049 unsigned int startbit, endbit;
6050 int this_slotno, intslots, intoffset;
6053 if (parms->intoffset == -1)
6056 intoffset = parms->intoffset;
6057 parms->intoffset = -1;
6059 startbit = intoffset & -BITS_PER_WORD;
6060 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
6061 intslots = (endbit - startbit) / BITS_PER_WORD;
6062 this_slotno = parms->slotno + intoffset / BITS_PER_WORD;
6064 intslots = MIN (intslots, SPARC_INT_ARG_MAX - this_slotno);
6068 /* If this is the trailing part of a word, only load that much into
6069 the register. Otherwise load the whole register. Note that in
6070 the latter case we may pick up unwanted bits. It's not a problem
6071 at the moment but may wish to revisit. */
6073 if (intoffset % BITS_PER_WORD != 0)
6074 mode = smallest_mode_for_size (BITS_PER_WORD - intoffset % BITS_PER_WORD,
6079 intoffset /= BITS_PER_UNIT;
6082 regno = parms->regbase + this_slotno;
6083 reg = gen_rtx_REG (mode, regno);
6084 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
6085 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (intoffset));
6088 intoffset = (intoffset | (UNITS_PER_WORD-1)) + 1;
6093 while (intslots > 0);
6096 /* A subroutine of function_arg_record_value. Traverse the structure
6097 recursively and assign bits to floating point registers. Track which
6098 bits in between need integer registers; invoke function_arg_record_value_3
6099 to make that happen. */
6102 function_arg_record_value_2 (const_tree type, HOST_WIDE_INT startbitpos,
6103 struct function_arg_record_value_parms *parms,
6109 for (field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field))
6111 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
6118 for (field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field))
6120 if (TREE_CODE (field) == FIELD_DECL)
6122 HOST_WIDE_INT bitpos = startbitpos;
6124 if (DECL_SIZE (field) != 0)
6126 if (integer_zerop (DECL_SIZE (field)))
6129 if (host_integerp (bit_position (field), 1))
6130 bitpos += int_bit_position (field);
6133 /* ??? FIXME: else assume zero offset. */
6135 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
6136 function_arg_record_value_2 (TREE_TYPE (field),
6140 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
6141 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
6146 int this_slotno = parms->slotno + bitpos / BITS_PER_WORD;
6147 int regno, nregs, pos;
6148 enum machine_mode mode = DECL_MODE (field);
6151 function_arg_record_value_3 (bitpos, parms);
6153 if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE
6156 mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field)));
6157 nregs = TYPE_VECTOR_SUBPARTS (TREE_TYPE (field));
6159 else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
6161 mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field)));
6167 regno = SPARC_FP_ARG_FIRST + this_slotno * 2;
6168 if (GET_MODE_SIZE (mode) <= 4 && (bitpos & 32) != 0)
6170 reg = gen_rtx_REG (mode, regno);
6171 pos = bitpos / BITS_PER_UNIT;
6172 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
6173 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos));
6177 regno += GET_MODE_SIZE (mode) / 4;
6178 reg = gen_rtx_REG (mode, regno);
6179 pos += GET_MODE_SIZE (mode);
6180 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
6181 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos));
6187 if (parms->intoffset == -1)
6188 parms->intoffset = bitpos;
6194 /* Used by function_arg and sparc_function_value_1 to implement the complex
6195 conventions of the 64-bit ABI for passing and returning structures.
6196 Return an expression valid as a return value for the FUNCTION_ARG
6197 and TARGET_FUNCTION_VALUE.
6199 TYPE is the data type of the argument (as a tree).
6200 This is null for libcalls where that information may
6202 MODE is the argument's machine mode.
6203 SLOTNO is the index number of the argument's slot in the parameter array.
6204 NAMED is nonzero if this argument is a named parameter
6205 (otherwise it is an extra parameter matching an ellipsis).
6206 REGBASE is the regno of the base register for the parameter array. */
6209 function_arg_record_value (const_tree type, enum machine_mode mode,
6210 int slotno, int named, int regbase)
6212 HOST_WIDE_INT typesize = int_size_in_bytes (type);
6213 struct function_arg_record_value_parms parms;
6216 parms.ret = NULL_RTX;
6217 parms.slotno = slotno;
6218 parms.named = named;
6219 parms.regbase = regbase;
6222 /* Compute how many registers we need. */
6224 parms.intoffset = 0;
6225 function_arg_record_value_1 (type, 0, &parms, false);
6227 /* Take into account pending integer fields. */
6228 if (parms.intoffset != -1)
6230 unsigned int startbit, endbit;
6231 int intslots, this_slotno;
6233 startbit = parms.intoffset & -BITS_PER_WORD;
6234 endbit = (typesize*BITS_PER_UNIT + BITS_PER_WORD - 1) & -BITS_PER_WORD;
6235 intslots = (endbit - startbit) / BITS_PER_WORD;
6236 this_slotno = slotno + parms.intoffset / BITS_PER_WORD;
6238 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
6240 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
6241 /* We need to pass this field on the stack. */
6245 parms.nregs += intslots;
6247 nregs = parms.nregs;
6249 /* Allocate the vector and handle some annoying special cases. */
6252 /* ??? Empty structure has no value? Duh? */
6255 /* Though there's nothing really to store, return a word register
6256 anyway so the rest of gcc doesn't go nuts. Returning a PARALLEL
6257 leads to breakage due to the fact that there are zero bytes to
6259 return gen_rtx_REG (mode, regbase);
6263 /* ??? C++ has structures with no fields, and yet a size. Give up
6264 for now and pass everything back in integer registers. */
6265 nregs = (typesize + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
6267 if (nregs + slotno > SPARC_INT_ARG_MAX)
6268 nregs = SPARC_INT_ARG_MAX - slotno;
6270 gcc_assert (nregs != 0);
6272 parms.ret = gen_rtx_PARALLEL (mode, rtvec_alloc (parms.stack + nregs));
6274 /* If at least one field must be passed on the stack, generate
6275 (parallel [(expr_list (nil) ...) ...]) so that all fields will
6276 also be passed on the stack. We can't do much better because the
6277 semantics of TARGET_ARG_PARTIAL_BYTES doesn't handle the case
6278 of structures for which the fields passed exclusively in registers
6279 are not at the beginning of the structure. */
6281 XVECEXP (parms.ret, 0, 0)
6282 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
6284 /* Fill in the entries. */
6286 parms.intoffset = 0;
6287 function_arg_record_value_2 (type, 0, &parms, false);
6288 function_arg_record_value_3 (typesize * BITS_PER_UNIT, &parms);
6290 gcc_assert (parms.nregs == nregs);
6295 /* Used by function_arg and sparc_function_value_1 to implement the conventions
6296 of the 64-bit ABI for passing and returning unions.
6297 Return an expression valid as a return value for the FUNCTION_ARG
6298 and TARGET_FUNCTION_VALUE.
6300 SIZE is the size in bytes of the union.
6301 MODE is the argument's machine mode.
6302 REGNO is the hard register the union will be passed in. */
6305 function_arg_union_value (int size, enum machine_mode mode, int slotno,
6308 int nwords = ROUND_ADVANCE (size), i;
6311 /* See comment in previous function for empty structures. */
6313 return gen_rtx_REG (mode, regno);
6315 if (slotno == SPARC_INT_ARG_MAX - 1)
6318 regs = gen_rtx_PARALLEL (mode, rtvec_alloc (nwords));
6320 for (i = 0; i < nwords; i++)
6322 /* Unions are passed left-justified. */
6323 XVECEXP (regs, 0, i)
6324 = gen_rtx_EXPR_LIST (VOIDmode,
6325 gen_rtx_REG (word_mode, regno),
6326 GEN_INT (UNITS_PER_WORD * i));
6333 /* Used by function_arg and sparc_function_value_1 to implement the conventions
6334 for passing and returning large (BLKmode) vectors.
6335 Return an expression valid as a return value for the FUNCTION_ARG
6336 and TARGET_FUNCTION_VALUE.
6338 SIZE is the size in bytes of the vector (at least 8 bytes).
6339 REGNO is the FP hard register the vector will be passed in. */
6342 function_arg_vector_value (int size, int regno)
6344 int i, nregs = size / 8;
6347 regs = gen_rtx_PARALLEL (BLKmode, rtvec_alloc (nregs));
6349 for (i = 0; i < nregs; i++)
6351 XVECEXP (regs, 0, i)
6352 = gen_rtx_EXPR_LIST (VOIDmode,
6353 gen_rtx_REG (DImode, regno + 2*i),
6360 /* Determine where to put an argument to a function.
6361 Value is zero to push the argument on the stack,
6362 or a hard register in which to store the argument.
6364 CUM is a variable of type CUMULATIVE_ARGS which gives info about
6365 the preceding args and about the function being called.
6366 MODE is the argument's machine mode.
6367 TYPE is the data type of the argument (as a tree).
6368 This is null for libcalls where that information may
6370 NAMED is true if this argument is a named parameter
6371 (otherwise it is an extra parameter matching an ellipsis).
6372 INCOMING_P is false for TARGET_FUNCTION_ARG, true for
6373 TARGET_FUNCTION_INCOMING_ARG. */
6376 sparc_function_arg_1 (cumulative_args_t cum_v, enum machine_mode mode,
6377 const_tree type, bool named, bool incoming_p)
6379 const CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v);
6381 int regbase = (incoming_p
6382 ? SPARC_INCOMING_INT_ARG_FIRST
6383 : SPARC_OUTGOING_INT_ARG_FIRST);
6384 int slotno, regno, padding;
6385 enum mode_class mclass = GET_MODE_CLASS (mode);
6387 slotno = function_arg_slotno (cum, mode, type, named, incoming_p,
6392 /* Vector types deserve special treatment because they are polymorphic wrt
6393 their mode, depending upon whether VIS instructions are enabled. */
6394 if (type && TREE_CODE (type) == VECTOR_TYPE)
6396 HOST_WIDE_INT size = int_size_in_bytes (type);
6397 gcc_assert ((TARGET_ARCH32 && size <= 8)
6398 || (TARGET_ARCH64 && size <= 16));
6400 if (mode == BLKmode)
6401 return function_arg_vector_value (size,
6402 SPARC_FP_ARG_FIRST + 2*slotno);
6404 mclass = MODE_FLOAT;
6408 return gen_rtx_REG (mode, regno);
6410 /* Structures up to 16 bytes in size are passed in arg slots on the stack
6411 and are promoted to registers if possible. */
6412 if (type && TREE_CODE (type) == RECORD_TYPE)
6414 HOST_WIDE_INT size = int_size_in_bytes (type);
6415 gcc_assert (size <= 16);
6417 return function_arg_record_value (type, mode, slotno, named, regbase);
6420 /* Unions up to 16 bytes in size are passed in integer registers. */
6421 else if (type && TREE_CODE (type) == UNION_TYPE)
6423 HOST_WIDE_INT size = int_size_in_bytes (type);
6424 gcc_assert (size <= 16);
6426 return function_arg_union_value (size, mode, slotno, regno);
6429 /* v9 fp args in reg slots beyond the int reg slots get passed in regs
6430 but also have the slot allocated for them.
6431 If no prototype is in scope fp values in register slots get passed
6432 in two places, either fp regs and int regs or fp regs and memory. */
6433 else if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT)
6434 && SPARC_FP_REG_P (regno))
6436 rtx reg = gen_rtx_REG (mode, regno);
6437 if (cum->prototype_p || cum->libcall_p)
6439 /* "* 2" because fp reg numbers are recorded in 4 byte
6442 /* ??? This will cause the value to be passed in the fp reg and
6443 in the stack. When a prototype exists we want to pass the
6444 value in the reg but reserve space on the stack. That's an
6445 optimization, and is deferred [for a bit]. */
6446 if ((regno - SPARC_FP_ARG_FIRST) >= SPARC_INT_ARG_MAX * 2)
6447 return gen_rtx_PARALLEL (mode,
6449 gen_rtx_EXPR_LIST (VOIDmode,
6450 NULL_RTX, const0_rtx),
6451 gen_rtx_EXPR_LIST (VOIDmode,
6455 /* ??? It seems that passing back a register even when past
6456 the area declared by REG_PARM_STACK_SPACE will allocate
6457 space appropriately, and will not copy the data onto the
6458 stack, exactly as we desire.
6460 This is due to locate_and_pad_parm being called in
6461 expand_call whenever reg_parm_stack_space > 0, which
6462 while beneficial to our example here, would seem to be
6463 in error from what had been intended. Ho hum... -- r~ */
6471 if ((regno - SPARC_FP_ARG_FIRST) < SPARC_INT_ARG_MAX * 2)
6475 /* On incoming, we don't need to know that the value
6476 is passed in %f0 and %i0, and it confuses other parts
6477 causing needless spillage even on the simplest cases. */
6481 intreg = (SPARC_OUTGOING_INT_ARG_FIRST
6482 + (regno - SPARC_FP_ARG_FIRST) / 2);
6484 v0 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
6485 v1 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_REG (mode, intreg),
6487 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
6491 v0 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
6492 v1 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
6493 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
6498 /* All other aggregate types are passed in an integer register in a mode
6499 corresponding to the size of the type. */
6500 else if (type && AGGREGATE_TYPE_P (type))
6502 HOST_WIDE_INT size = int_size_in_bytes (type);
6503 gcc_assert (size <= 16);
6505 mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0);
6508 return gen_rtx_REG (mode, regno);
6511 /* Handle the TARGET_FUNCTION_ARG target hook. */
6514 sparc_function_arg (cumulative_args_t cum, enum machine_mode mode,
6515 const_tree type, bool named)
6517 return sparc_function_arg_1 (cum, mode, type, named, false);
6520 /* Handle the TARGET_FUNCTION_INCOMING_ARG target hook. */
6523 sparc_function_incoming_arg (cumulative_args_t cum, enum machine_mode mode,
6524 const_tree type, bool named)
6526 return sparc_function_arg_1 (cum, mode, type, named, true);
6529 /* For sparc64, objects requiring 16 byte alignment are passed that way. */
6532 sparc_function_arg_boundary (enum machine_mode mode, const_tree type)
6534 return ((TARGET_ARCH64
6535 && (GET_MODE_ALIGNMENT (mode) == 128
6536 || (type && TYPE_ALIGN (type) == 128)))
6541 /* For an arg passed partly in registers and partly in memory,
6542 this is the number of bytes of registers used.
6543 For args passed entirely in registers or entirely in memory, zero.
6545 Any arg that starts in the first 6 regs but won't entirely fit in them
6546 needs partial registers on v8. On v9, structures with integer
6547 values in arg slots 5,6 will be passed in %o5 and SP+176, and complex fp
6548 values that begin in the last fp reg [where "last fp reg" varies with the
6549 mode] will be split between that reg and memory. */
6552 sparc_arg_partial_bytes (cumulative_args_t cum, enum machine_mode mode,
6553 tree type, bool named)
6555 int slotno, regno, padding;
6557 /* We pass false for incoming_p here, it doesn't matter. */
6558 slotno = function_arg_slotno (get_cumulative_args (cum), mode, type, named,
6559 false, ®no, &padding);
6566 if ((slotno + (mode == BLKmode
6567 ? ROUND_ADVANCE (int_size_in_bytes (type))
6568 : ROUND_ADVANCE (GET_MODE_SIZE (mode))))
6569 > SPARC_INT_ARG_MAX)
6570 return (SPARC_INT_ARG_MAX - slotno) * UNITS_PER_WORD;
6574 /* We are guaranteed by pass_by_reference that the size of the
6575 argument is not greater than 16 bytes, so we only need to return
6576 one word if the argument is partially passed in registers. */
6578 if (type && AGGREGATE_TYPE_P (type))
6580 int size = int_size_in_bytes (type);
6582 if (size > UNITS_PER_WORD
6583 && slotno == SPARC_INT_ARG_MAX - 1)
6584 return UNITS_PER_WORD;
6586 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_INT
6587 || (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT
6588 && ! (TARGET_FPU && named)))
6590 /* The complex types are passed as packed types. */
6591 if (GET_MODE_SIZE (mode) > UNITS_PER_WORD
6592 && slotno == SPARC_INT_ARG_MAX - 1)
6593 return UNITS_PER_WORD;
6595 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT)
6597 if ((slotno + GET_MODE_SIZE (mode) / UNITS_PER_WORD)
6599 return UNITS_PER_WORD;
6606 /* Handle the TARGET_PASS_BY_REFERENCE target hook.
6607 Specify whether to pass the argument by reference. */
6610 sparc_pass_by_reference (cumulative_args_t cum ATTRIBUTE_UNUSED,
6611 enum machine_mode mode, const_tree type,
6612 bool named ATTRIBUTE_UNUSED)
6615 /* Original SPARC 32-bit ABI says that structures and unions,
6616 and quad-precision floats are passed by reference. For Pascal,
6617 also pass arrays by reference. All other base types are passed
6620 Extended ABI (as implemented by the Sun compiler) says that all
6621 complex floats are passed by reference. Pass complex integers
6622 in registers up to 8 bytes. More generally, enforce the 2-word
6623 cap for passing arguments in registers.
6625 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6626 integers are passed like floats of the same size, that is in
6627 registers up to 8 bytes. Pass all vector floats by reference
6628 like structure and unions. */
6629 return ((type && (AGGREGATE_TYPE_P (type) || VECTOR_FLOAT_TYPE_P (type)))
6631 /* Catch CDImode, TFmode, DCmode and TCmode. */
6632 || GET_MODE_SIZE (mode) > 8
6634 && TREE_CODE (type) == VECTOR_TYPE
6635 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8));
6637 /* Original SPARC 64-bit ABI says that structures and unions
6638 smaller than 16 bytes are passed in registers, as well as
6639 all other base types.
6641 Extended ABI (as implemented by the Sun compiler) says that
6642 complex floats are passed in registers up to 16 bytes. Pass
6643 all complex integers in registers up to 16 bytes. More generally,
6644 enforce the 2-word cap for passing arguments in registers.
6646 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6647 integers are passed like floats of the same size, that is in
6648 registers (up to 16 bytes). Pass all vector floats like structure
6651 && (AGGREGATE_TYPE_P (type) || TREE_CODE (type) == VECTOR_TYPE)
6652 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 16)
6653 /* Catch CTImode and TCmode. */
6654 || GET_MODE_SIZE (mode) > 16);
6657 /* Handle the TARGET_FUNCTION_ARG_ADVANCE hook.
6658 Update the data in CUM to advance over an argument
6659 of mode MODE and data type TYPE.
6660 TYPE is null for libcalls where that information may not be available. */
6663 sparc_function_arg_advance (cumulative_args_t cum_v, enum machine_mode mode,
6664 const_tree type, bool named)
6666 CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v);
6669 /* We pass false for incoming_p here, it doesn't matter. */
6670 function_arg_slotno (cum, mode, type, named, false, ®no, &padding);
6672 /* If argument requires leading padding, add it. */
6673 cum->words += padding;
6677 cum->words += (mode != BLKmode
6678 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
6679 : ROUND_ADVANCE (int_size_in_bytes (type)));
6683 if (type && AGGREGATE_TYPE_P (type))
6685 int size = int_size_in_bytes (type);
6689 else if (size <= 16)
6691 else /* passed by reference */
6696 cum->words += (mode != BLKmode
6697 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
6698 : ROUND_ADVANCE (int_size_in_bytes (type)));
6703 /* Handle the FUNCTION_ARG_PADDING macro.
6704 For the 64 bit ABI structs are always stored left shifted in their
6708 function_arg_padding (enum machine_mode mode, const_tree type)
6710 if (TARGET_ARCH64 && type != 0 && AGGREGATE_TYPE_P (type))
6713 /* Fall back to the default. */
6714 return DEFAULT_FUNCTION_ARG_PADDING (mode, type);
6717 /* Handle the TARGET_RETURN_IN_MEMORY target hook.
6718 Specify whether to return the return value in memory. */
6721 sparc_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
6724 /* Original SPARC 32-bit ABI says that structures and unions,
6725 and quad-precision floats are returned in memory. All other
6726 base types are returned in registers.
6728 Extended ABI (as implemented by the Sun compiler) says that
6729 all complex floats are returned in registers (8 FP registers
6730 at most for '_Complex long double'). Return all complex integers
6731 in registers (4 at most for '_Complex long long').
6733 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6734 integers are returned like floats of the same size, that is in
6735 registers up to 8 bytes and in memory otherwise. Return all
6736 vector floats in memory like structure and unions; note that
6737 they always have BLKmode like the latter. */
6738 return (TYPE_MODE (type) == BLKmode
6739 || TYPE_MODE (type) == TFmode
6740 || (TREE_CODE (type) == VECTOR_TYPE
6741 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8));
6743 /* Original SPARC 64-bit ABI says that structures and unions
6744 smaller than 32 bytes are returned in registers, as well as
6745 all other base types.
6747 Extended ABI (as implemented by the Sun compiler) says that all
6748 complex floats are returned in registers (8 FP registers at most
6749 for '_Complex long double'). Return all complex integers in
6750 registers (4 at most for '_Complex TItype').
6752 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6753 integers are returned like floats of the same size, that is in
6754 registers. Return all vector floats like structure and unions;
6755 note that they always have BLKmode like the latter. */
6756 return (TYPE_MODE (type) == BLKmode
6757 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 32);
6760 /* Handle the TARGET_STRUCT_VALUE target hook.
6761 Return where to find the structure return value address. */
6764 sparc_struct_value_rtx (tree fndecl, int incoming)
6773 mem = gen_frame_mem (Pmode, plus_constant (Pmode, frame_pointer_rtx,
6774 STRUCT_VALUE_OFFSET));
6776 mem = gen_frame_mem (Pmode, plus_constant (Pmode, stack_pointer_rtx,
6777 STRUCT_VALUE_OFFSET));
6779 /* Only follow the SPARC ABI for fixed-size structure returns.
6780 Variable size structure returns are handled per the normal
6781 procedures in GCC. This is enabled by -mstd-struct-return */
6783 && sparc_std_struct_return
6784 && TYPE_SIZE_UNIT (TREE_TYPE (fndecl))
6785 && TREE_CODE (TYPE_SIZE_UNIT (TREE_TYPE (fndecl))) == INTEGER_CST)
6787 /* We must check and adjust the return address, as it is
6788 optional as to whether the return object is really
6790 rtx ret_reg = gen_rtx_REG (Pmode, 31);
6791 rtx scratch = gen_reg_rtx (SImode);
6792 rtx endlab = gen_label_rtx ();
6794 /* Calculate the return object size */
6795 tree size = TYPE_SIZE_UNIT (TREE_TYPE (fndecl));
6796 rtx size_rtx = GEN_INT (TREE_INT_CST_LOW (size) & 0xfff);
6797 /* Construct a temporary return value */
6799 = assign_stack_local (Pmode, TREE_INT_CST_LOW (size), 0);
6801 /* Implement SPARC 32-bit psABI callee return struct checking:
6803 Fetch the instruction where we will return to and see if
6804 it's an unimp instruction (the most significant 10 bits
6806 emit_move_insn (scratch, gen_rtx_MEM (SImode,
6807 plus_constant (Pmode,
6809 /* Assume the size is valid and pre-adjust */
6810 emit_insn (gen_add3_insn (ret_reg, ret_reg, GEN_INT (4)));
6811 emit_cmp_and_jump_insns (scratch, size_rtx, EQ, const0_rtx, SImode,
6813 emit_insn (gen_sub3_insn (ret_reg, ret_reg, GEN_INT (4)));
6814 /* Write the address of the memory pointed to by temp_val into
6815 the memory pointed to by mem */
6816 emit_move_insn (mem, XEXP (temp_val, 0));
6817 emit_label (endlab);
6824 /* Handle TARGET_FUNCTION_VALUE, and TARGET_LIBCALL_VALUE target hook.
6825 For v9, function return values are subject to the same rules as arguments,
6826 except that up to 32 bytes may be returned in registers. */
6829 sparc_function_value_1 (const_tree type, enum machine_mode mode,
6832 /* Beware that the two values are swapped here wrt function_arg. */
6833 int regbase = (outgoing
6834 ? SPARC_INCOMING_INT_ARG_FIRST
6835 : SPARC_OUTGOING_INT_ARG_FIRST);
6836 enum mode_class mclass = GET_MODE_CLASS (mode);
6839 /* Vector types deserve special treatment because they are polymorphic wrt
6840 their mode, depending upon whether VIS instructions are enabled. */
6841 if (type && TREE_CODE (type) == VECTOR_TYPE)
6843 HOST_WIDE_INT size = int_size_in_bytes (type);
6844 gcc_assert ((TARGET_ARCH32 && size <= 8)
6845 || (TARGET_ARCH64 && size <= 32));
6847 if (mode == BLKmode)
6848 return function_arg_vector_value (size,
6849 SPARC_FP_ARG_FIRST);
6851 mclass = MODE_FLOAT;
6854 if (TARGET_ARCH64 && type)
6856 /* Structures up to 32 bytes in size are returned in registers. */
6857 if (TREE_CODE (type) == RECORD_TYPE)
6859 HOST_WIDE_INT size = int_size_in_bytes (type);
6860 gcc_assert (size <= 32);
6862 return function_arg_record_value (type, mode, 0, 1, regbase);
6865 /* Unions up to 32 bytes in size are returned in integer registers. */
6866 else if (TREE_CODE (type) == UNION_TYPE)
6868 HOST_WIDE_INT size = int_size_in_bytes (type);
6869 gcc_assert (size <= 32);
6871 return function_arg_union_value (size, mode, 0, regbase);
6874 /* Objects that require it are returned in FP registers. */
6875 else if (mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT)
6878 /* All other aggregate types are returned in an integer register in a
6879 mode corresponding to the size of the type. */
6880 else if (AGGREGATE_TYPE_P (type))
6882 /* All other aggregate types are passed in an integer register
6883 in a mode corresponding to the size of the type. */
6884 HOST_WIDE_INT size = int_size_in_bytes (type);
6885 gcc_assert (size <= 32);
6887 mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0);
6889 /* ??? We probably should have made the same ABI change in
6890 3.4.0 as the one we made for unions. The latter was
6891 required by the SCD though, while the former is not
6892 specified, so we favored compatibility and efficiency.
6894 Now we're stuck for aggregates larger than 16 bytes,
6895 because OImode vanished in the meantime. Let's not
6896 try to be unduly clever, and simply follow the ABI
6897 for unions in that case. */
6898 if (mode == BLKmode)
6899 return function_arg_union_value (size, mode, 0, regbase);
6904 /* We should only have pointer and integer types at this point. This
6905 must match sparc_promote_function_mode. */
6906 else if (mclass == MODE_INT && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
6910 /* We should only have pointer and integer types at this point. This must
6911 match sparc_promote_function_mode. */
6912 else if (TARGET_ARCH32
6913 && mclass == MODE_INT
6914 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
6917 if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT) && TARGET_FPU)
6918 regno = SPARC_FP_ARG_FIRST;
6922 return gen_rtx_REG (mode, regno);
6925 /* Handle TARGET_FUNCTION_VALUE.
6926 On the SPARC, the value is found in the first "output" register, but the
6927 called function leaves it in the first "input" register. */
6930 sparc_function_value (const_tree valtype,
6931 const_tree fn_decl_or_type ATTRIBUTE_UNUSED,
6934 return sparc_function_value_1 (valtype, TYPE_MODE (valtype), outgoing);
6937 /* Handle TARGET_LIBCALL_VALUE. */
6940 sparc_libcall_value (enum machine_mode mode,
6941 const_rtx fun ATTRIBUTE_UNUSED)
6943 return sparc_function_value_1 (NULL_TREE, mode, false);
6946 /* Handle FUNCTION_VALUE_REGNO_P.
6947 On the SPARC, the first "output" reg is used for integer values, and the
6948 first floating point register is used for floating point values. */
6951 sparc_function_value_regno_p (const unsigned int regno)
6953 return (regno == 8 || regno == 32);
6956 /* Do what is necessary for `va_start'. We look at the current function
6957 to determine if stdarg or varargs is used and return the address of
6958 the first unnamed parameter. */
6961 sparc_builtin_saveregs (void)
6963 int first_reg = crtl->args.info.words;
6967 for (regno = first_reg; regno < SPARC_INT_ARG_MAX; regno++)
6968 emit_move_insn (gen_rtx_MEM (word_mode,
6969 gen_rtx_PLUS (Pmode,
6971 GEN_INT (FIRST_PARM_OFFSET (0)
6974 gen_rtx_REG (word_mode,
6975 SPARC_INCOMING_INT_ARG_FIRST + regno));
6977 address = gen_rtx_PLUS (Pmode,
6979 GEN_INT (FIRST_PARM_OFFSET (0)
6980 + UNITS_PER_WORD * first_reg));
6985 /* Implement `va_start' for stdarg. */
6988 sparc_va_start (tree valist, rtx nextarg)
6990 nextarg = expand_builtin_saveregs ();
6991 std_expand_builtin_va_start (valist, nextarg);
6994 /* Implement `va_arg' for stdarg. */
6997 sparc_gimplify_va_arg (tree valist, tree type, gimple_seq *pre_p,
7000 HOST_WIDE_INT size, rsize, align;
7003 tree ptrtype = build_pointer_type (type);
7005 if (pass_by_reference (NULL, TYPE_MODE (type), type, false))
7008 size = rsize = UNITS_PER_WORD;
7014 size = int_size_in_bytes (type);
7015 rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD;
7020 /* For SPARC64, objects requiring 16-byte alignment get it. */
7021 if (TYPE_ALIGN (type) >= 2 * (unsigned) BITS_PER_WORD)
7022 align = 2 * UNITS_PER_WORD;
7024 /* SPARC-V9 ABI states that structures up to 16 bytes in size
7025 are left-justified in their slots. */
7026 if (AGGREGATE_TYPE_P (type))
7029 size = rsize = UNITS_PER_WORD;
7039 incr = fold_build_pointer_plus_hwi (incr, align - 1);
7040 incr = fold_convert (sizetype, incr);
7041 incr = fold_build2 (BIT_AND_EXPR, sizetype, incr,
7043 incr = fold_convert (ptr_type_node, incr);
7046 gimplify_expr (&incr, pre_p, post_p, is_gimple_val, fb_rvalue);
7049 if (BYTES_BIG_ENDIAN && size < rsize)
7050 addr = fold_build_pointer_plus_hwi (incr, rsize - size);
7054 addr = fold_convert (build_pointer_type (ptrtype), addr);
7055 addr = build_va_arg_indirect_ref (addr);
7058 /* If the address isn't aligned properly for the type, we need a temporary.
7059 FIXME: This is inefficient, usually we can do this in registers. */
7060 else if (align == 0 && TYPE_ALIGN (type) > BITS_PER_WORD)
7062 tree tmp = create_tmp_var (type, "va_arg_tmp");
7063 tree dest_addr = build_fold_addr_expr (tmp);
7064 tree copy = build_call_expr (builtin_decl_implicit (BUILT_IN_MEMCPY),
7065 3, dest_addr, addr, size_int (rsize));
7066 TREE_ADDRESSABLE (tmp) = 1;
7067 gimplify_and_add (copy, pre_p);
7072 addr = fold_convert (ptrtype, addr);
7074 incr = fold_build_pointer_plus_hwi (incr, rsize);
7075 gimplify_assign (valist, incr, post_p);
7077 return build_va_arg_indirect_ref (addr);
7080 /* Implement the TARGET_VECTOR_MODE_SUPPORTED_P target hook.
7081 Specify whether the vector mode is supported by the hardware. */
7084 sparc_vector_mode_supported_p (enum machine_mode mode)
7086 return TARGET_VIS && VECTOR_MODE_P (mode) ? true : false;
7089 /* Implement the TARGET_VECTORIZE_PREFERRED_SIMD_MODE target hook. */
7091 static enum machine_mode
7092 sparc_preferred_simd_mode (enum machine_mode mode)
7110 /* Return the string to output an unconditional branch to LABEL, which is
7111 the operand number of the label.
7113 DEST is the destination insn (i.e. the label), INSN is the source. */
7116 output_ubranch (rtx dest, int label, rtx insn)
7118 static char string[64];
7119 bool v9_form = false;
7122 if (TARGET_V9 && INSN_ADDRESSES_SET_P ())
7124 int delta = (INSN_ADDRESSES (INSN_UID (dest))
7125 - INSN_ADDRESSES (INSN_UID (insn)));
7126 /* Leave some instructions for "slop". */
7127 if (delta >= -260000 && delta < 260000)
7132 strcpy (string, "ba%*,pt\t%%xcc, ");
7134 strcpy (string, "b%*\t");
7136 p = strchr (string, '\0');
7147 /* Return the string to output a conditional branch to LABEL, which is
7148 the operand number of the label. OP is the conditional expression.
7149 XEXP (OP, 0) is assumed to be a condition code register (integer or
7150 floating point) and its mode specifies what kind of comparison we made.
7152 DEST is the destination insn (i.e. the label), INSN is the source.
7154 REVERSED is nonzero if we should reverse the sense of the comparison.
7156 ANNUL is nonzero if we should generate an annulling branch. */
7159 output_cbranch (rtx op, rtx dest, int label, int reversed, int annul,
7162 static char string[64];
7163 enum rtx_code code = GET_CODE (op);
7164 rtx cc_reg = XEXP (op, 0);
7165 enum machine_mode mode = GET_MODE (cc_reg);
7166 const char *labelno, *branch;
7167 int spaces = 8, far;
7170 /* v9 branches are limited to +-1MB. If it is too far away,
7183 fbne,a,pn %fcc2, .LC29
7191 far = TARGET_V9 && (get_attr_length (insn) >= 3);
7194 /* Reversal of FP compares takes care -- an ordered compare
7195 becomes an unordered compare and vice versa. */
7196 if (mode == CCFPmode || mode == CCFPEmode)
7197 code = reverse_condition_maybe_unordered (code);
7199 code = reverse_condition (code);
7202 /* Start by writing the branch condition. */
7203 if (mode == CCFPmode || mode == CCFPEmode)
7254 /* ??? !v9: FP branches cannot be preceded by another floating point
7255 insn. Because there is currently no concept of pre-delay slots,
7256 we can fix this only by always emitting a nop before a floating
7261 strcpy (string, "nop\n\t");
7262 strcat (string, branch);
7275 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
7287 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
7308 strcpy (string, branch);
7310 spaces -= strlen (branch);
7311 p = strchr (string, '\0');
7313 /* Now add the annulling, the label, and a possible noop. */
7326 if (! far && insn && INSN_ADDRESSES_SET_P ())
7328 int delta = (INSN_ADDRESSES (INSN_UID (dest))
7329 - INSN_ADDRESSES (INSN_UID (insn)));
7330 /* Leave some instructions for "slop". */
7331 if (delta < -260000 || delta >= 260000)
7335 if (mode == CCFPmode || mode == CCFPEmode)
7337 static char v9_fcc_labelno[] = "%%fccX, ";
7338 /* Set the char indicating the number of the fcc reg to use. */
7339 v9_fcc_labelno[5] = REGNO (cc_reg) - SPARC_FIRST_V9_FCC_REG + '0';
7340 labelno = v9_fcc_labelno;
7343 gcc_assert (REGNO (cc_reg) == SPARC_FCC_REG);
7347 else if (mode == CCXmode || mode == CCX_NOOVmode)
7349 labelno = "%%xcc, ";
7354 labelno = "%%icc, ";
7359 if (*labelno && insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
7362 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
7375 strcpy (p, labelno);
7376 p = strchr (p, '\0');
7379 strcpy (p, ".+12\n\t nop\n\tb\t");
7380 /* Skip the next insn if requested or
7381 if we know that it will be a nop. */
7382 if (annul || ! final_sequence)
7396 /* Emit a library call comparison between floating point X and Y.
7397 COMPARISON is the operator to compare with (EQ, NE, GT, etc).
7398 Return the new operator to be used in the comparison sequence.
7400 TARGET_ARCH64 uses _Qp_* functions, which use pointers to TFmode
7401 values as arguments instead of the TFmode registers themselves,
7402 that's why we cannot call emit_float_lib_cmp. */
7405 sparc_emit_float_lib_cmp (rtx x, rtx y, enum rtx_code comparison)
7408 rtx slot0, slot1, result, tem, tem2, libfunc;
7409 enum machine_mode mode;
7410 enum rtx_code new_comparison;
7415 qpfunc = (TARGET_ARCH64 ? "_Qp_feq" : "_Q_feq");
7419 qpfunc = (TARGET_ARCH64 ? "_Qp_fne" : "_Q_fne");
7423 qpfunc = (TARGET_ARCH64 ? "_Qp_fgt" : "_Q_fgt");
7427 qpfunc = (TARGET_ARCH64 ? "_Qp_fge" : "_Q_fge");
7431 qpfunc = (TARGET_ARCH64 ? "_Qp_flt" : "_Q_flt");
7435 qpfunc = (TARGET_ARCH64 ? "_Qp_fle" : "_Q_fle");
7446 qpfunc = (TARGET_ARCH64 ? "_Qp_cmp" : "_Q_cmp");
7457 tree expr = MEM_EXPR (x);
7459 mark_addressable (expr);
7464 slot0 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode));
7465 emit_move_insn (slot0, x);
7470 tree expr = MEM_EXPR (y);
7472 mark_addressable (expr);
7477 slot1 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode));
7478 emit_move_insn (slot1, y);
7481 libfunc = gen_rtx_SYMBOL_REF (Pmode, qpfunc);
7482 emit_library_call (libfunc, LCT_NORMAL,
7484 XEXP (slot0, 0), Pmode,
7485 XEXP (slot1, 0), Pmode);
7490 libfunc = gen_rtx_SYMBOL_REF (Pmode, qpfunc);
7491 emit_library_call (libfunc, LCT_NORMAL,
7493 x, TFmode, y, TFmode);
7498 /* Immediately move the result of the libcall into a pseudo
7499 register so reload doesn't clobber the value if it needs
7500 the return register for a spill reg. */
7501 result = gen_reg_rtx (mode);
7502 emit_move_insn (result, hard_libcall_value (mode, libfunc));
7507 return gen_rtx_NE (VOIDmode, result, const0_rtx);
7510 new_comparison = (comparison == UNORDERED ? EQ : NE);
7511 return gen_rtx_fmt_ee (new_comparison, VOIDmode, result, GEN_INT(3));
7514 new_comparison = (comparison == UNGT ? GT : NE);
7515 return gen_rtx_fmt_ee (new_comparison, VOIDmode, result, const1_rtx);
7517 return gen_rtx_NE (VOIDmode, result, const2_rtx);
7519 tem = gen_reg_rtx (mode);
7521 emit_insn (gen_andsi3 (tem, result, const1_rtx));
7523 emit_insn (gen_anddi3 (tem, result, const1_rtx));
7524 return gen_rtx_NE (VOIDmode, tem, const0_rtx);
7527 tem = gen_reg_rtx (mode);
7529 emit_insn (gen_addsi3 (tem, result, const1_rtx));
7531 emit_insn (gen_adddi3 (tem, result, const1_rtx));
7532 tem2 = gen_reg_rtx (mode);
7534 emit_insn (gen_andsi3 (tem2, tem, const2_rtx));
7536 emit_insn (gen_anddi3 (tem2, tem, const2_rtx));
7537 new_comparison = (comparison == UNEQ ? EQ : NE);
7538 return gen_rtx_fmt_ee (new_comparison, VOIDmode, tem2, const0_rtx);
7544 /* Generate an unsigned DImode to FP conversion. This is the same code
7545 optabs would emit if we didn't have TFmode patterns. */
7548 sparc_emit_floatunsdi (rtx *operands, enum machine_mode mode)
7550 rtx neglab, donelab, i0, i1, f0, in, out;
7553 in = force_reg (DImode, operands[1]);
7554 neglab = gen_label_rtx ();
7555 donelab = gen_label_rtx ();
7556 i0 = gen_reg_rtx (DImode);
7557 i1 = gen_reg_rtx (DImode);
7558 f0 = gen_reg_rtx (mode);
7560 emit_cmp_and_jump_insns (in, const0_rtx, LT, const0_rtx, DImode, 0, neglab);
7562 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_FLOAT (mode, in)));
7563 emit_jump_insn (gen_jump (donelab));
7566 emit_label (neglab);
7568 emit_insn (gen_lshrdi3 (i0, in, const1_rtx));
7569 emit_insn (gen_anddi3 (i1, in, const1_rtx));
7570 emit_insn (gen_iordi3 (i0, i0, i1));
7571 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_FLOAT (mode, i0)));
7572 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_PLUS (mode, f0, f0)));
7574 emit_label (donelab);
7577 /* Generate an FP to unsigned DImode conversion. This is the same code
7578 optabs would emit if we didn't have TFmode patterns. */
7581 sparc_emit_fixunsdi (rtx *operands, enum machine_mode mode)
7583 rtx neglab, donelab, i0, i1, f0, in, out, limit;
7586 in = force_reg (mode, operands[1]);
7587 neglab = gen_label_rtx ();
7588 donelab = gen_label_rtx ();
7589 i0 = gen_reg_rtx (DImode);
7590 i1 = gen_reg_rtx (DImode);
7591 limit = gen_reg_rtx (mode);
7592 f0 = gen_reg_rtx (mode);
7594 emit_move_insn (limit,
7595 CONST_DOUBLE_FROM_REAL_VALUE (
7596 REAL_VALUE_ATOF ("9223372036854775808.0", mode), mode));
7597 emit_cmp_and_jump_insns (in, limit, GE, NULL_RTX, mode, 0, neglab);
7599 emit_insn (gen_rtx_SET (VOIDmode,
7601 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, in))));
7602 emit_jump_insn (gen_jump (donelab));
7605 emit_label (neglab);
7607 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_MINUS (mode, in, limit)));
7608 emit_insn (gen_rtx_SET (VOIDmode,
7610 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, f0))));
7611 emit_insn (gen_movdi (i1, const1_rtx));
7612 emit_insn (gen_ashldi3 (i1, i1, GEN_INT (63)));
7613 emit_insn (gen_xordi3 (out, i0, i1));
7615 emit_label (donelab);
7618 /* Return the string to output a conditional branch to LABEL, testing
7619 register REG. LABEL is the operand number of the label; REG is the
7620 operand number of the reg. OP is the conditional expression. The mode
7621 of REG says what kind of comparison we made.
7623 DEST is the destination insn (i.e. the label), INSN is the source.
7625 REVERSED is nonzero if we should reverse the sense of the comparison.
7627 ANNUL is nonzero if we should generate an annulling branch. */
7630 output_v9branch (rtx op, rtx dest, int reg, int label, int reversed,
7631 int annul, rtx insn)
7633 static char string[64];
7634 enum rtx_code code = GET_CODE (op);
7635 enum machine_mode mode = GET_MODE (XEXP (op, 0));
7640 /* branch on register are limited to +-128KB. If it is too far away,
7653 brgez,a,pn %o1, .LC29
7659 ba,pt %xcc, .LC29 */
7661 far = get_attr_length (insn) >= 3;
7663 /* If not floating-point or if EQ or NE, we can just reverse the code. */
7665 code = reverse_condition (code);
7667 /* Only 64 bit versions of these instructions exist. */
7668 gcc_assert (mode == DImode);
7670 /* Start by writing the branch condition. */
7675 strcpy (string, "brnz");
7679 strcpy (string, "brz");
7683 strcpy (string, "brgez");
7687 strcpy (string, "brlz");
7691 strcpy (string, "brlez");
7695 strcpy (string, "brgz");
7702 p = strchr (string, '\0');
7704 /* Now add the annulling, reg, label, and nop. */
7711 if (insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
7714 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
7719 *p = p < string + 8 ? '\t' : ' ';
7727 int veryfar = 1, delta;
7729 if (INSN_ADDRESSES_SET_P ())
7731 delta = (INSN_ADDRESSES (INSN_UID (dest))
7732 - INSN_ADDRESSES (INSN_UID (insn)));
7733 /* Leave some instructions for "slop". */
7734 if (delta >= -260000 && delta < 260000)
7738 strcpy (p, ".+12\n\t nop\n\t");
7739 /* Skip the next insn if requested or
7740 if we know that it will be a nop. */
7741 if (annul || ! final_sequence)
7751 strcpy (p, "ba,pt\t%%xcc, ");
7765 /* Return 1, if any of the registers of the instruction are %l[0-7] or %o[0-7].
7766 Such instructions cannot be used in the delay slot of return insn on v9.
7767 If TEST is 0, also rename all %i[0-7] registers to their %o[0-7] counterparts.
7771 epilogue_renumber (register rtx *where, int test)
7773 register const char *fmt;
7775 register enum rtx_code code;
7780 code = GET_CODE (*where);
7785 if (REGNO (*where) >= 8 && REGNO (*where) < 24) /* oX or lX */
7787 if (! test && REGNO (*where) >= 24 && REGNO (*where) < 32)
7788 *where = gen_rtx_REG (GET_MODE (*where), OUTGOING_REGNO (REGNO(*where)));
7796 /* Do not replace the frame pointer with the stack pointer because
7797 it can cause the delayed instruction to load below the stack.
7798 This occurs when instructions like:
7800 (set (reg/i:SI 24 %i0)
7801 (mem/f:SI (plus:SI (reg/f:SI 30 %fp)
7802 (const_int -20 [0xffffffec])) 0))
7804 are in the return delayed slot. */
7806 if (GET_CODE (XEXP (*where, 0)) == REG
7807 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM
7808 && (GET_CODE (XEXP (*where, 1)) != CONST_INT
7809 || INTVAL (XEXP (*where, 1)) < SPARC_STACK_BIAS))
7814 if (SPARC_STACK_BIAS
7815 && GET_CODE (XEXP (*where, 0)) == REG
7816 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM)
7824 fmt = GET_RTX_FORMAT (code);
7826 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7831 for (j = XVECLEN (*where, i) - 1; j >= 0; j--)
7832 if (epilogue_renumber (&(XVECEXP (*where, i, j)), test))
7835 else if (fmt[i] == 'e'
7836 && epilogue_renumber (&(XEXP (*where, i)), test))
7842 /* Leaf functions and non-leaf functions have different needs. */
7845 reg_leaf_alloc_order[] = REG_LEAF_ALLOC_ORDER;
7848 reg_nonleaf_alloc_order[] = REG_ALLOC_ORDER;
7850 static const int *const reg_alloc_orders[] = {
7851 reg_leaf_alloc_order,
7852 reg_nonleaf_alloc_order};
7855 order_regs_for_local_alloc (void)
7857 static int last_order_nonleaf = 1;
7859 if (df_regs_ever_live_p (15) != last_order_nonleaf)
7861 last_order_nonleaf = !last_order_nonleaf;
7862 memcpy ((char *) reg_alloc_order,
7863 (const char *) reg_alloc_orders[last_order_nonleaf],
7864 FIRST_PSEUDO_REGISTER * sizeof (int));
7868 /* Return 1 if REG and MEM are legitimate enough to allow the various
7869 mem<-->reg splits to be run. */
7872 sparc_splitdi_legitimate (rtx reg, rtx mem)
7874 /* Punt if we are here by mistake. */
7875 gcc_assert (reload_completed);
7877 /* We must have an offsettable memory reference. */
7878 if (! offsettable_memref_p (mem))
7881 /* If we have legitimate args for ldd/std, we do not want
7882 the split to happen. */
7883 if ((REGNO (reg) % 2) == 0
7884 && mem_min_alignment (mem, 8))
7891 /* Like sparc_splitdi_legitimate but for REG <--> REG moves. */
7894 sparc_split_regreg_legitimate (rtx reg1, rtx reg2)
7898 if (GET_CODE (reg1) == SUBREG)
7899 reg1 = SUBREG_REG (reg1);
7900 if (GET_CODE (reg1) != REG)
7902 regno1 = REGNO (reg1);
7904 if (GET_CODE (reg2) == SUBREG)
7905 reg2 = SUBREG_REG (reg2);
7906 if (GET_CODE (reg2) != REG)
7908 regno2 = REGNO (reg2);
7910 if (SPARC_INT_REG_P (regno1) && SPARC_INT_REG_P (regno2))
7915 if ((SPARC_INT_REG_P (regno1) && SPARC_FP_REG_P (regno2))
7916 || (SPARC_FP_REG_P (regno1) && SPARC_INT_REG_P (regno2)))
7923 /* Return 1 if x and y are some kind of REG and they refer to
7924 different hard registers. This test is guaranteed to be
7925 run after reload. */
7928 sparc_absnegfloat_split_legitimate (rtx x, rtx y)
7930 if (GET_CODE (x) != REG)
7932 if (GET_CODE (y) != REG)
7934 if (REGNO (x) == REGNO (y))
7939 /* Return 1 if REGNO (reg1) is even and REGNO (reg1) == REGNO (reg2) - 1.
7940 This makes them candidates for using ldd and std insns.
7942 Note reg1 and reg2 *must* be hard registers. */
7945 registers_ok_for_ldd_peep (rtx reg1, rtx reg2)
7947 /* We might have been passed a SUBREG. */
7948 if (GET_CODE (reg1) != REG || GET_CODE (reg2) != REG)
7951 if (REGNO (reg1) % 2 != 0)
7954 /* Integer ldd is deprecated in SPARC V9 */
7955 if (TARGET_V9 && SPARC_INT_REG_P (REGNO (reg1)))
7958 return (REGNO (reg1) == REGNO (reg2) - 1);
7961 /* Return 1 if the addresses in mem1 and mem2 are suitable for use in
7964 This can only happen when addr1 and addr2, the addresses in mem1
7965 and mem2, are consecutive memory locations (addr1 + 4 == addr2).
7966 addr1 must also be aligned on a 64-bit boundary.
7968 Also iff dependent_reg_rtx is not null it should not be used to
7969 compute the address for mem1, i.e. we cannot optimize a sequence
7981 But, note that the transformation from:
7986 is perfectly fine. Thus, the peephole2 patterns always pass us
7987 the destination register of the first load, never the second one.
7989 For stores we don't have a similar problem, so dependent_reg_rtx is
7993 mems_ok_for_ldd_peep (rtx mem1, rtx mem2, rtx dependent_reg_rtx)
7997 HOST_WIDE_INT offset1;
7999 /* The mems cannot be volatile. */
8000 if (MEM_VOLATILE_P (mem1) || MEM_VOLATILE_P (mem2))
8003 /* MEM1 should be aligned on a 64-bit boundary. */
8004 if (MEM_ALIGN (mem1) < 64)
8007 addr1 = XEXP (mem1, 0);
8008 addr2 = XEXP (mem2, 0);
8010 /* Extract a register number and offset (if used) from the first addr. */
8011 if (GET_CODE (addr1) == PLUS)
8013 /* If not a REG, return zero. */
8014 if (GET_CODE (XEXP (addr1, 0)) != REG)
8018 reg1 = REGNO (XEXP (addr1, 0));
8019 /* The offset must be constant! */
8020 if (GET_CODE (XEXP (addr1, 1)) != CONST_INT)
8022 offset1 = INTVAL (XEXP (addr1, 1));
8025 else if (GET_CODE (addr1) != REG)
8029 reg1 = REGNO (addr1);
8030 /* This was a simple (mem (reg)) expression. Offset is 0. */
8034 /* Make sure the second address is a (mem (plus (reg) (const_int). */
8035 if (GET_CODE (addr2) != PLUS)
8038 if (GET_CODE (XEXP (addr2, 0)) != REG
8039 || GET_CODE (XEXP (addr2, 1)) != CONST_INT)
8042 if (reg1 != REGNO (XEXP (addr2, 0)))
8045 if (dependent_reg_rtx != NULL_RTX && reg1 == REGNO (dependent_reg_rtx))
8048 /* The first offset must be evenly divisible by 8 to ensure the
8049 address is 64 bit aligned. */
8050 if (offset1 % 8 != 0)
8053 /* The offset for the second addr must be 4 more than the first addr. */
8054 if (INTVAL (XEXP (addr2, 1)) != offset1 + 4)
8057 /* All the tests passed. addr1 and addr2 are valid for ldd and std
8062 /* Return 1 if reg is a pseudo, or is the first register in
8063 a hard register pair. This makes it suitable for use in
8064 ldd and std insns. */
8067 register_ok_for_ldd (rtx reg)
8069 /* We might have been passed a SUBREG. */
8073 if (REGNO (reg) < FIRST_PSEUDO_REGISTER)
8074 return (REGNO (reg) % 2 == 0);
8079 /* Return 1 if OP is a memory whose address is known to be
8080 aligned to 8-byte boundary, or a pseudo during reload.
8081 This makes it suitable for use in ldd and std insns. */
8084 memory_ok_for_ldd (rtx op)
8088 /* In 64-bit mode, we assume that the address is word-aligned. */
8089 if (TARGET_ARCH32 && !mem_min_alignment (op, 8))
8092 if (! can_create_pseudo_p ()
8093 && !strict_memory_address_p (Pmode, XEXP (op, 0)))
8096 else if (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER)
8098 if (!(reload_in_progress && reg_renumber [REGNO (op)] < 0))
8107 /* Implement TARGET_PRINT_OPERAND_PUNCT_VALID_P. */
8110 sparc_print_operand_punct_valid_p (unsigned char code)
8123 /* Implement TARGET_PRINT_OPERAND.
8124 Print operand X (an rtx) in assembler syntax to file FILE.
8125 CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified.
8126 For `%' followed by punctuation, CODE is the punctuation and X is null. */
8129 sparc_print_operand (FILE *file, rtx x, int code)
8134 /* Output an insn in a delay slot. */
8136 sparc_indent_opcode = 1;
8138 fputs ("\n\t nop", file);
8141 /* Output an annul flag if there's nothing for the delay slot and we
8142 are optimizing. This is always used with '(' below.
8143 Sun OS 4.1.1 dbx can't handle an annulled unconditional branch;
8144 this is a dbx bug. So, we only do this when optimizing.
8145 On UltraSPARC, a branch in a delay slot causes a pipeline flush.
8146 Always emit a nop in case the next instruction is a branch. */
8147 if (! final_sequence && (optimize && (int)sparc_cpu < PROCESSOR_V9))
8151 /* Output a 'nop' if there's nothing for the delay slot and we are
8152 not optimizing. This is always used with '*' above. */
8153 if (! final_sequence && ! (optimize && (int)sparc_cpu < PROCESSOR_V9))
8154 fputs ("\n\t nop", file);
8155 else if (final_sequence)
8156 sparc_indent_opcode = 1;
8159 /* Output the right displacement from the saved PC on function return.
8160 The caller may have placed an "unimp" insn immediately after the call
8161 so we have to account for it. This insn is used in the 32-bit ABI
8162 when calling a function that returns a non zero-sized structure. The
8163 64-bit ABI doesn't have it. Be careful to have this test be the same
8164 as that for the call. The exception is when sparc_std_struct_return
8165 is enabled, the psABI is followed exactly and the adjustment is made
8166 by the code in sparc_struct_value_rtx. The call emitted is the same
8167 when sparc_std_struct_return is enabled. */
8169 && cfun->returns_struct
8170 && !sparc_std_struct_return
8171 && DECL_SIZE (DECL_RESULT (current_function_decl))
8172 && TREE_CODE (DECL_SIZE (DECL_RESULT (current_function_decl)))
8174 && !integer_zerop (DECL_SIZE (DECL_RESULT (current_function_decl))))
8180 /* Output the Embedded Medium/Anywhere code model base register. */
8181 fputs (EMBMEDANY_BASE_REG, file);
8184 /* Print some local dynamic TLS name. */
8185 assemble_name (file, get_some_local_dynamic_name ());
8189 /* Adjust the operand to take into account a RESTORE operation. */
8190 if (GET_CODE (x) == CONST_INT)
8192 else if (GET_CODE (x) != REG)
8193 output_operand_lossage ("invalid %%Y operand");
8194 else if (REGNO (x) < 8)
8195 fputs (reg_names[REGNO (x)], file);
8196 else if (REGNO (x) >= 24 && REGNO (x) < 32)
8197 fputs (reg_names[REGNO (x)-16], file);
8199 output_operand_lossage ("invalid %%Y operand");
8202 /* Print out the low order register name of a register pair. */
8203 if (WORDS_BIG_ENDIAN)
8204 fputs (reg_names[REGNO (x)+1], file);
8206 fputs (reg_names[REGNO (x)], file);
8209 /* Print out the high order register name of a register pair. */
8210 if (WORDS_BIG_ENDIAN)
8211 fputs (reg_names[REGNO (x)], file);
8213 fputs (reg_names[REGNO (x)+1], file);
8216 /* Print out the second register name of a register pair or quad.
8217 I.e., R (%o0) => %o1. */
8218 fputs (reg_names[REGNO (x)+1], file);
8221 /* Print out the third register name of a register quad.
8222 I.e., S (%o0) => %o2. */
8223 fputs (reg_names[REGNO (x)+2], file);
8226 /* Print out the fourth register name of a register quad.
8227 I.e., T (%o0) => %o3. */
8228 fputs (reg_names[REGNO (x)+3], file);
8231 /* Print a condition code register. */
8232 if (REGNO (x) == SPARC_ICC_REG)
8234 /* We don't handle CC[X]_NOOVmode because they're not supposed
8236 if (GET_MODE (x) == CCmode)
8237 fputs ("%icc", file);
8238 else if (GET_MODE (x) == CCXmode)
8239 fputs ("%xcc", file);
8244 /* %fccN register */
8245 fputs (reg_names[REGNO (x)], file);
8248 /* Print the operand's address only. */
8249 output_address (XEXP (x, 0));
8252 /* In this case we need a register. Use %g0 if the
8253 operand is const0_rtx. */
8255 || (GET_MODE (x) != VOIDmode && x == CONST0_RTX (GET_MODE (x))))
8257 fputs ("%g0", file);
8264 switch (GET_CODE (x))
8266 case IOR: fputs ("or", file); break;
8267 case AND: fputs ("and", file); break;
8268 case XOR: fputs ("xor", file); break;
8269 default: output_operand_lossage ("invalid %%A operand");
8274 switch (GET_CODE (x))
8276 case IOR: fputs ("orn", file); break;
8277 case AND: fputs ("andn", file); break;
8278 case XOR: fputs ("xnor", file); break;
8279 default: output_operand_lossage ("invalid %%B operand");
8283 /* This is used by the conditional move instructions. */
8286 enum rtx_code rc = GET_CODE (x);
8290 case NE: fputs ("ne", file); break;
8291 case EQ: fputs ("e", file); break;
8292 case GE: fputs ("ge", file); break;
8293 case GT: fputs ("g", file); break;
8294 case LE: fputs ("le", file); break;
8295 case LT: fputs ("l", file); break;
8296 case GEU: fputs ("geu", file); break;
8297 case GTU: fputs ("gu", file); break;
8298 case LEU: fputs ("leu", file); break;
8299 case LTU: fputs ("lu", file); break;
8300 case LTGT: fputs ("lg", file); break;
8301 case UNORDERED: fputs ("u", file); break;
8302 case ORDERED: fputs ("o", file); break;
8303 case UNLT: fputs ("ul", file); break;
8304 case UNLE: fputs ("ule", file); break;
8305 case UNGT: fputs ("ug", file); break;
8306 case UNGE: fputs ("uge", file); break;
8307 case UNEQ: fputs ("ue", file); break;
8308 default: output_operand_lossage ("invalid %%C operand");
8313 /* This are used by the movr instruction pattern. */
8316 enum rtx_code rc = GET_CODE (x);
8319 case NE: fputs ("ne", file); break;
8320 case EQ: fputs ("e", file); break;
8321 case GE: fputs ("gez", file); break;
8322 case LT: fputs ("lz", file); break;
8323 case LE: fputs ("lez", file); break;
8324 case GT: fputs ("gz", file); break;
8325 default: output_operand_lossage ("invalid %%D operand");
8332 /* Print a sign-extended character. */
8333 int i = trunc_int_for_mode (INTVAL (x), QImode);
8334 fprintf (file, "%d", i);
8339 /* Operand must be a MEM; write its address. */
8340 if (GET_CODE (x) != MEM)
8341 output_operand_lossage ("invalid %%f operand");
8342 output_address (XEXP (x, 0));
8347 /* Print a sign-extended 32-bit value. */
8349 if (GET_CODE(x) == CONST_INT)
8351 else if (GET_CODE(x) == CONST_DOUBLE)
8352 i = CONST_DOUBLE_LOW (x);
8355 output_operand_lossage ("invalid %%s operand");
8358 i = trunc_int_for_mode (i, SImode);
8359 fprintf (file, HOST_WIDE_INT_PRINT_DEC, i);
8364 /* Do nothing special. */
8368 /* Undocumented flag. */
8369 output_operand_lossage ("invalid operand output code");
8372 if (GET_CODE (x) == REG)
8373 fputs (reg_names[REGNO (x)], file);
8374 else if (GET_CODE (x) == MEM)
8377 /* Poor Sun assembler doesn't understand absolute addressing. */
8378 if (CONSTANT_P (XEXP (x, 0)))
8379 fputs ("%g0+", file);
8380 output_address (XEXP (x, 0));
8383 else if (GET_CODE (x) == HIGH)
8385 fputs ("%hi(", file);
8386 output_addr_const (file, XEXP (x, 0));
8389 else if (GET_CODE (x) == LO_SUM)
8391 sparc_print_operand (file, XEXP (x, 0), 0);
8392 if (TARGET_CM_MEDMID)
8393 fputs ("+%l44(", file);
8395 fputs ("+%lo(", file);
8396 output_addr_const (file, XEXP (x, 1));
8399 else if (GET_CODE (x) == CONST_DOUBLE
8400 && (GET_MODE (x) == VOIDmode
8401 || GET_MODE_CLASS (GET_MODE (x)) == MODE_INT))
8403 if (CONST_DOUBLE_HIGH (x) == 0)
8404 fprintf (file, "%u", (unsigned int) CONST_DOUBLE_LOW (x));
8405 else if (CONST_DOUBLE_HIGH (x) == -1
8406 && CONST_DOUBLE_LOW (x) < 0)
8407 fprintf (file, "%d", (int) CONST_DOUBLE_LOW (x));
8409 output_operand_lossage ("long long constant not a valid immediate operand");
8411 else if (GET_CODE (x) == CONST_DOUBLE)
8412 output_operand_lossage ("floating point constant not a valid immediate operand");
8413 else { output_addr_const (file, x); }
8416 /* Implement TARGET_PRINT_OPERAND_ADDRESS. */
8419 sparc_print_operand_address (FILE *file, rtx x)
8421 register rtx base, index = 0;
8423 register rtx addr = x;
8426 fputs (reg_names[REGNO (addr)], file);
8427 else if (GET_CODE (addr) == PLUS)
8429 if (CONST_INT_P (XEXP (addr, 0)))
8430 offset = INTVAL (XEXP (addr, 0)), base = XEXP (addr, 1);
8431 else if (CONST_INT_P (XEXP (addr, 1)))
8432 offset = INTVAL (XEXP (addr, 1)), base = XEXP (addr, 0);
8434 base = XEXP (addr, 0), index = XEXP (addr, 1);
8435 if (GET_CODE (base) == LO_SUM)
8437 gcc_assert (USE_AS_OFFSETABLE_LO10
8439 && ! TARGET_CM_MEDMID);
8440 output_operand (XEXP (base, 0), 0);
8441 fputs ("+%lo(", file);
8442 output_address (XEXP (base, 1));
8443 fprintf (file, ")+%d", offset);
8447 fputs (reg_names[REGNO (base)], file);
8449 fprintf (file, "%+d", offset);
8450 else if (REG_P (index))
8451 fprintf (file, "+%s", reg_names[REGNO (index)]);
8452 else if (GET_CODE (index) == SYMBOL_REF
8453 || GET_CODE (index) == LABEL_REF
8454 || GET_CODE (index) == CONST)
8455 fputc ('+', file), output_addr_const (file, index);
8456 else gcc_unreachable ();
8459 else if (GET_CODE (addr) == MINUS
8460 && GET_CODE (XEXP (addr, 1)) == LABEL_REF)
8462 output_addr_const (file, XEXP (addr, 0));
8464 output_addr_const (file, XEXP (addr, 1));
8465 fputs ("-.)", file);
8467 else if (GET_CODE (addr) == LO_SUM)
8469 output_operand (XEXP (addr, 0), 0);
8470 if (TARGET_CM_MEDMID)
8471 fputs ("+%l44(", file);
8473 fputs ("+%lo(", file);
8474 output_address (XEXP (addr, 1));
8478 && GET_CODE (addr) == CONST
8479 && GET_CODE (XEXP (addr, 0)) == MINUS
8480 && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST
8481 && GET_CODE (XEXP (XEXP (XEXP (addr, 0), 1), 0)) == MINUS
8482 && XEXP (XEXP (XEXP (XEXP (addr, 0), 1), 0), 1) == pc_rtx)
8484 addr = XEXP (addr, 0);
8485 output_addr_const (file, XEXP (addr, 0));
8486 /* Group the args of the second CONST in parenthesis. */
8488 /* Skip past the second CONST--it does nothing for us. */
8489 output_addr_const (file, XEXP (XEXP (addr, 1), 0));
8490 /* Close the parenthesis. */
8495 output_addr_const (file, addr);
8499 /* Target hook for assembling integer objects. The sparc version has
8500 special handling for aligned DI-mode objects. */
8503 sparc_assemble_integer (rtx x, unsigned int size, int aligned_p)
8505 /* ??? We only output .xword's for symbols and only then in environments
8506 where the assembler can handle them. */
8507 if (aligned_p && size == 8
8508 && (GET_CODE (x) != CONST_INT && GET_CODE (x) != CONST_DOUBLE))
8512 assemble_integer_with_op ("\t.xword\t", x);
8517 assemble_aligned_integer (4, const0_rtx);
8518 assemble_aligned_integer (4, x);
8522 return default_assemble_integer (x, size, aligned_p);
8525 /* Return the value of a code used in the .proc pseudo-op that says
8526 what kind of result this function returns. For non-C types, we pick
8527 the closest C type. */
8529 #ifndef SHORT_TYPE_SIZE
8530 #define SHORT_TYPE_SIZE (BITS_PER_UNIT * 2)
8533 #ifndef INT_TYPE_SIZE
8534 #define INT_TYPE_SIZE BITS_PER_WORD
8537 #ifndef LONG_TYPE_SIZE
8538 #define LONG_TYPE_SIZE BITS_PER_WORD
8541 #ifndef LONG_LONG_TYPE_SIZE
8542 #define LONG_LONG_TYPE_SIZE (BITS_PER_WORD * 2)
8545 #ifndef FLOAT_TYPE_SIZE
8546 #define FLOAT_TYPE_SIZE BITS_PER_WORD
8549 #ifndef DOUBLE_TYPE_SIZE
8550 #define DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
8553 #ifndef LONG_DOUBLE_TYPE_SIZE
8554 #define LONG_DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
8558 sparc_type_code (register tree type)
8560 register unsigned long qualifiers = 0;
8561 register unsigned shift;
8563 /* Only the first 30 bits of the qualifier are valid. We must refrain from
8564 setting more, since some assemblers will give an error for this. Also,
8565 we must be careful to avoid shifts of 32 bits or more to avoid getting
8566 unpredictable results. */
8568 for (shift = 6; shift < 30; shift += 2, type = TREE_TYPE (type))
8570 switch (TREE_CODE (type))
8576 qualifiers |= (3 << shift);
8581 qualifiers |= (2 << shift);
8585 case REFERENCE_TYPE:
8587 qualifiers |= (1 << shift);
8591 return (qualifiers | 8);
8594 case QUAL_UNION_TYPE:
8595 return (qualifiers | 9);
8598 return (qualifiers | 10);
8601 return (qualifiers | 16);
8604 /* If this is a range type, consider it to be the underlying
8606 if (TREE_TYPE (type) != 0)
8609 /* Carefully distinguish all the standard types of C,
8610 without messing up if the language is not C. We do this by
8611 testing TYPE_PRECISION and TYPE_UNSIGNED. The old code used to
8612 look at both the names and the above fields, but that's redundant.
8613 Any type whose size is between two C types will be considered
8614 to be the wider of the two types. Also, we do not have a
8615 special code to use for "long long", so anything wider than
8616 long is treated the same. Note that we can't distinguish
8617 between "int" and "long" in this code if they are the same
8618 size, but that's fine, since neither can the assembler. */
8620 if (TYPE_PRECISION (type) <= CHAR_TYPE_SIZE)
8621 return (qualifiers | (TYPE_UNSIGNED (type) ? 12 : 2));
8623 else if (TYPE_PRECISION (type) <= SHORT_TYPE_SIZE)
8624 return (qualifiers | (TYPE_UNSIGNED (type) ? 13 : 3));
8626 else if (TYPE_PRECISION (type) <= INT_TYPE_SIZE)
8627 return (qualifiers | (TYPE_UNSIGNED (type) ? 14 : 4));
8630 return (qualifiers | (TYPE_UNSIGNED (type) ? 15 : 5));
8633 /* If this is a range type, consider it to be the underlying
8635 if (TREE_TYPE (type) != 0)
8638 /* Carefully distinguish all the standard types of C,
8639 without messing up if the language is not C. */
8641 if (TYPE_PRECISION (type) == FLOAT_TYPE_SIZE)
8642 return (qualifiers | 6);
8645 return (qualifiers | 7);
8647 case COMPLEX_TYPE: /* GNU Fortran COMPLEX type. */
8648 /* ??? We need to distinguish between double and float complex types,
8649 but I don't know how yet because I can't reach this code from
8650 existing front-ends. */
8651 return (qualifiers | 7); /* Who knows? */
8654 case BOOLEAN_TYPE: /* Boolean truth value type. */
8660 gcc_unreachable (); /* Not a type! */
8667 /* Nested function support. */
8669 /* Emit RTL insns to initialize the variable parts of a trampoline.
8670 FNADDR is an RTX for the address of the function's pure code.
8671 CXT is an RTX for the static chain value for the function.
8673 This takes 16 insns: 2 shifts & 2 ands (to split up addresses), 4 sethi
8674 (to load in opcodes), 4 iors (to merge address and opcodes), and 4 writes
8675 (to store insns). This is a bit excessive. Perhaps a different
8676 mechanism would be better here.
8678 Emit enough FLUSH insns to synchronize the data and instruction caches. */
8681 sparc32_initialize_trampoline (rtx m_tramp, rtx fnaddr, rtx cxt)
8683 /* SPARC 32-bit trampoline:
8686 sethi %hi(static), %g2
8688 or %g2, %lo(static), %g2
8690 SETHI i,r = 00rr rrr1 00ii iiii iiii iiii iiii iiii
8691 JMPL r+i,d = 10dd ddd1 1100 0rrr rr1i iiii iiii iiii
8695 (adjust_address (m_tramp, SImode, 0),
8696 expand_binop (SImode, ior_optab,
8697 expand_shift (RSHIFT_EXPR, SImode, fnaddr, 10, 0, 1),
8698 GEN_INT (trunc_int_for_mode (0x03000000, SImode)),
8699 NULL_RTX, 1, OPTAB_DIRECT));
8702 (adjust_address (m_tramp, SImode, 4),
8703 expand_binop (SImode, ior_optab,
8704 expand_shift (RSHIFT_EXPR, SImode, cxt, 10, 0, 1),
8705 GEN_INT (trunc_int_for_mode (0x05000000, SImode)),
8706 NULL_RTX, 1, OPTAB_DIRECT));
8709 (adjust_address (m_tramp, SImode, 8),
8710 expand_binop (SImode, ior_optab,
8711 expand_and (SImode, fnaddr, GEN_INT (0x3ff), NULL_RTX),
8712 GEN_INT (trunc_int_for_mode (0x81c06000, SImode)),
8713 NULL_RTX, 1, OPTAB_DIRECT));
8716 (adjust_address (m_tramp, SImode, 12),
8717 expand_binop (SImode, ior_optab,
8718 expand_and (SImode, cxt, GEN_INT (0x3ff), NULL_RTX),
8719 GEN_INT (trunc_int_for_mode (0x8410a000, SImode)),
8720 NULL_RTX, 1, OPTAB_DIRECT));
8722 /* On UltraSPARC a flush flushes an entire cache line. The trampoline is
8723 aligned on a 16 byte boundary so one flush clears it all. */
8724 emit_insn (gen_flush (validize_mem (adjust_address (m_tramp, SImode, 0))));
8725 if (sparc_cpu != PROCESSOR_ULTRASPARC
8726 && sparc_cpu != PROCESSOR_ULTRASPARC3
8727 && sparc_cpu != PROCESSOR_NIAGARA
8728 && sparc_cpu != PROCESSOR_NIAGARA2
8729 && sparc_cpu != PROCESSOR_NIAGARA3
8730 && sparc_cpu != PROCESSOR_NIAGARA4)
8731 emit_insn (gen_flush (validize_mem (adjust_address (m_tramp, SImode, 8))));
8733 /* Call __enable_execute_stack after writing onto the stack to make sure
8734 the stack address is accessible. */
8735 #ifdef HAVE_ENABLE_EXECUTE_STACK
8736 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
8737 LCT_NORMAL, VOIDmode, 1, XEXP (m_tramp, 0), Pmode);
8742 /* The 64-bit version is simpler because it makes more sense to load the
8743 values as "immediate" data out of the trampoline. It's also easier since
8744 we can read the PC without clobbering a register. */
8747 sparc64_initialize_trampoline (rtx m_tramp, rtx fnaddr, rtx cxt)
8749 /* SPARC 64-bit trampoline:
8758 emit_move_insn (adjust_address (m_tramp, SImode, 0),
8759 GEN_INT (trunc_int_for_mode (0x83414000, SImode)));
8760 emit_move_insn (adjust_address (m_tramp, SImode, 4),
8761 GEN_INT (trunc_int_for_mode (0xca586018, SImode)));
8762 emit_move_insn (adjust_address (m_tramp, SImode, 8),
8763 GEN_INT (trunc_int_for_mode (0x81c14000, SImode)));
8764 emit_move_insn (adjust_address (m_tramp, SImode, 12),
8765 GEN_INT (trunc_int_for_mode (0xca586010, SImode)));
8766 emit_move_insn (adjust_address (m_tramp, DImode, 16), cxt);
8767 emit_move_insn (adjust_address (m_tramp, DImode, 24), fnaddr);
8768 emit_insn (gen_flushdi (validize_mem (adjust_address (m_tramp, DImode, 0))));
8770 if (sparc_cpu != PROCESSOR_ULTRASPARC
8771 && sparc_cpu != PROCESSOR_ULTRASPARC3
8772 && sparc_cpu != PROCESSOR_NIAGARA
8773 && sparc_cpu != PROCESSOR_NIAGARA2
8774 && sparc_cpu != PROCESSOR_NIAGARA3
8775 && sparc_cpu != PROCESSOR_NIAGARA4)
8776 emit_insn (gen_flushdi (validize_mem (adjust_address (m_tramp, DImode, 8))));
8778 /* Call __enable_execute_stack after writing onto the stack to make sure
8779 the stack address is accessible. */
8780 #ifdef HAVE_ENABLE_EXECUTE_STACK
8781 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
8782 LCT_NORMAL, VOIDmode, 1, XEXP (m_tramp, 0), Pmode);
8786 /* Worker for TARGET_TRAMPOLINE_INIT. */
8789 sparc_trampoline_init (rtx m_tramp, tree fndecl, rtx cxt)
8791 rtx fnaddr = force_reg (Pmode, XEXP (DECL_RTL (fndecl), 0));
8792 cxt = force_reg (Pmode, cxt);
8794 sparc64_initialize_trampoline (m_tramp, fnaddr, cxt);
8796 sparc32_initialize_trampoline (m_tramp, fnaddr, cxt);
8799 /* Adjust the cost of a scheduling dependency. Return the new cost of
8800 a dependency LINK or INSN on DEP_INSN. COST is the current cost. */
8803 supersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
8805 enum attr_type insn_type;
8807 if (! recog_memoized (insn))
8810 insn_type = get_attr_type (insn);
8812 if (REG_NOTE_KIND (link) == 0)
8814 /* Data dependency; DEP_INSN writes a register that INSN reads some
8817 /* if a load, then the dependence must be on the memory address;
8818 add an extra "cycle". Note that the cost could be two cycles
8819 if the reg was written late in an instruction group; we ca not tell
8821 if (insn_type == TYPE_LOAD || insn_type == TYPE_FPLOAD)
8824 /* Get the delay only if the address of the store is the dependence. */
8825 if (insn_type == TYPE_STORE || insn_type == TYPE_FPSTORE)
8827 rtx pat = PATTERN(insn);
8828 rtx dep_pat = PATTERN (dep_insn);
8830 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
8831 return cost; /* This should not happen! */
8833 /* The dependency between the two instructions was on the data that
8834 is being stored. Assume that this implies that the address of the
8835 store is not dependent. */
8836 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
8839 return cost + 3; /* An approximation. */
8842 /* A shift instruction cannot receive its data from an instruction
8843 in the same cycle; add a one cycle penalty. */
8844 if (insn_type == TYPE_SHIFT)
8845 return cost + 3; /* Split before cascade into shift. */
8849 /* Anti- or output- dependency; DEP_INSN reads/writes a register that
8850 INSN writes some cycles later. */
8852 /* These are only significant for the fpu unit; writing a fp reg before
8853 the fpu has finished with it stalls the processor. */
8855 /* Reusing an integer register causes no problems. */
8856 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
8864 hypersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
8866 enum attr_type insn_type, dep_type;
8867 rtx pat = PATTERN(insn);
8868 rtx dep_pat = PATTERN (dep_insn);
8870 if (recog_memoized (insn) < 0 || recog_memoized (dep_insn) < 0)
8873 insn_type = get_attr_type (insn);
8874 dep_type = get_attr_type (dep_insn);
8876 switch (REG_NOTE_KIND (link))
8879 /* Data dependency; DEP_INSN writes a register that INSN reads some
8886 /* Get the delay iff the address of the store is the dependence. */
8887 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
8890 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
8897 /* If a load, then the dependence must be on the memory address. If
8898 the addresses aren't equal, then it might be a false dependency */
8899 if (dep_type == TYPE_STORE || dep_type == TYPE_FPSTORE)
8901 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET
8902 || GET_CODE (SET_DEST (dep_pat)) != MEM
8903 || GET_CODE (SET_SRC (pat)) != MEM
8904 || ! rtx_equal_p (XEXP (SET_DEST (dep_pat), 0),
8905 XEXP (SET_SRC (pat), 0)))
8913 /* Compare to branch latency is 0. There is no benefit from
8914 separating compare and branch. */
8915 if (dep_type == TYPE_COMPARE)
8917 /* Floating point compare to branch latency is less than
8918 compare to conditional move. */
8919 if (dep_type == TYPE_FPCMP)
8928 /* Anti-dependencies only penalize the fpu unit. */
8929 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
8941 sparc_adjust_cost(rtx insn, rtx link, rtx dep, int cost)
8945 case PROCESSOR_SUPERSPARC:
8946 cost = supersparc_adjust_cost (insn, link, dep, cost);
8948 case PROCESSOR_HYPERSPARC:
8949 case PROCESSOR_SPARCLITE86X:
8950 cost = hypersparc_adjust_cost (insn, link, dep, cost);
8959 sparc_sched_init (FILE *dump ATTRIBUTE_UNUSED,
8960 int sched_verbose ATTRIBUTE_UNUSED,
8961 int max_ready ATTRIBUTE_UNUSED)
8965 sparc_use_sched_lookahead (void)
8967 if (sparc_cpu == PROCESSOR_NIAGARA
8968 || sparc_cpu == PROCESSOR_NIAGARA2
8969 || sparc_cpu == PROCESSOR_NIAGARA3)
8971 if (sparc_cpu == PROCESSOR_NIAGARA4)
8973 if (sparc_cpu == PROCESSOR_ULTRASPARC
8974 || sparc_cpu == PROCESSOR_ULTRASPARC3)
8976 if ((1 << sparc_cpu) &
8977 ((1 << PROCESSOR_SUPERSPARC) | (1 << PROCESSOR_HYPERSPARC) |
8978 (1 << PROCESSOR_SPARCLITE86X)))
8984 sparc_issue_rate (void)
8988 case PROCESSOR_NIAGARA:
8989 case PROCESSOR_NIAGARA2:
8990 case PROCESSOR_NIAGARA3:
8993 case PROCESSOR_NIAGARA4:
8995 /* Assume V9 processors are capable of at least dual-issue. */
8997 case PROCESSOR_SUPERSPARC:
8999 case PROCESSOR_HYPERSPARC:
9000 case PROCESSOR_SPARCLITE86X:
9002 case PROCESSOR_ULTRASPARC:
9003 case PROCESSOR_ULTRASPARC3:
9009 set_extends (rtx insn)
9011 register rtx pat = PATTERN (insn);
9013 switch (GET_CODE (SET_SRC (pat)))
9015 /* Load and some shift instructions zero extend. */
9018 /* sethi clears the high bits */
9020 /* LO_SUM is used with sethi. sethi cleared the high
9021 bits and the values used with lo_sum are positive */
9023 /* Store flag stores 0 or 1 */
9033 rtx op0 = XEXP (SET_SRC (pat), 0);
9034 rtx op1 = XEXP (SET_SRC (pat), 1);
9035 if (GET_CODE (op1) == CONST_INT)
9036 return INTVAL (op1) >= 0;
9037 if (GET_CODE (op0) != REG)
9039 if (sparc_check_64 (op0, insn) == 1)
9041 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
9046 rtx op0 = XEXP (SET_SRC (pat), 0);
9047 rtx op1 = XEXP (SET_SRC (pat), 1);
9048 if (GET_CODE (op0) != REG || sparc_check_64 (op0, insn) <= 0)
9050 if (GET_CODE (op1) == CONST_INT)
9051 return INTVAL (op1) >= 0;
9052 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
9055 return GET_MODE (SET_SRC (pat)) == SImode;
9056 /* Positive integers leave the high bits zero. */
9058 return ! (CONST_DOUBLE_LOW (SET_SRC (pat)) & 0x80000000);
9060 return ! (INTVAL (SET_SRC (pat)) & 0x80000000);
9063 return - (GET_MODE (SET_SRC (pat)) == SImode);
9065 return sparc_check_64 (SET_SRC (pat), insn);
9071 /* We _ought_ to have only one kind per function, but... */
9072 static GTY(()) rtx sparc_addr_diff_list;
9073 static GTY(()) rtx sparc_addr_list;
9076 sparc_defer_case_vector (rtx lab, rtx vec, int diff)
9078 vec = gen_rtx_EXPR_LIST (VOIDmode, lab, vec);
9080 sparc_addr_diff_list
9081 = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_diff_list);
9083 sparc_addr_list = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_list);
9087 sparc_output_addr_vec (rtx vec)
9089 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
9090 int idx, vlen = XVECLEN (body, 0);
9092 #ifdef ASM_OUTPUT_ADDR_VEC_START
9093 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
9096 #ifdef ASM_OUTPUT_CASE_LABEL
9097 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
9100 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
9103 for (idx = 0; idx < vlen; idx++)
9105 ASM_OUTPUT_ADDR_VEC_ELT
9106 (asm_out_file, CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 0, idx), 0)));
9109 #ifdef ASM_OUTPUT_ADDR_VEC_END
9110 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
9115 sparc_output_addr_diff_vec (rtx vec)
9117 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
9118 rtx base = XEXP (XEXP (body, 0), 0);
9119 int idx, vlen = XVECLEN (body, 1);
9121 #ifdef ASM_OUTPUT_ADDR_VEC_START
9122 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
9125 #ifdef ASM_OUTPUT_CASE_LABEL
9126 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
9129 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
9132 for (idx = 0; idx < vlen; idx++)
9134 ASM_OUTPUT_ADDR_DIFF_ELT
9137 CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 1, idx), 0)),
9138 CODE_LABEL_NUMBER (base));
9141 #ifdef ASM_OUTPUT_ADDR_VEC_END
9142 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
9147 sparc_output_deferred_case_vectors (void)
9152 if (sparc_addr_list == NULL_RTX
9153 && sparc_addr_diff_list == NULL_RTX)
9156 /* Align to cache line in the function's code section. */
9157 switch_to_section (current_function_section ());
9159 align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
9161 ASM_OUTPUT_ALIGN (asm_out_file, align);
9163 for (t = sparc_addr_list; t ; t = XEXP (t, 1))
9164 sparc_output_addr_vec (XEXP (t, 0));
9165 for (t = sparc_addr_diff_list; t ; t = XEXP (t, 1))
9166 sparc_output_addr_diff_vec (XEXP (t, 0));
9168 sparc_addr_list = sparc_addr_diff_list = NULL_RTX;
9171 /* Return 0 if the high 32 bits of X (the low word of X, if DImode) are
9172 unknown. Return 1 if the high bits are zero, -1 if the register is
9175 sparc_check_64 (rtx x, rtx insn)
9177 /* If a register is set only once it is safe to ignore insns this
9178 code does not know how to handle. The loop will either recognize
9179 the single set and return the correct value or fail to recognize
9184 gcc_assert (GET_CODE (x) == REG);
9186 if (GET_MODE (x) == DImode)
9187 y = gen_rtx_REG (SImode, REGNO (x) + WORDS_BIG_ENDIAN);
9189 if (flag_expensive_optimizations
9190 && df && DF_REG_DEF_COUNT (REGNO (y)) == 1)
9196 insn = get_last_insn_anywhere ();
9201 while ((insn = PREV_INSN (insn)))
9203 switch (GET_CODE (insn))
9216 rtx pat = PATTERN (insn);
9217 if (GET_CODE (pat) != SET)
9219 if (rtx_equal_p (x, SET_DEST (pat)))
9220 return set_extends (insn);
9221 if (y && rtx_equal_p (y, SET_DEST (pat)))
9222 return set_extends (insn);
9223 if (reg_overlap_mentioned_p (SET_DEST (pat), y))
9231 /* Output a wide shift instruction in V8+ mode. INSN is the instruction,
9232 OPERANDS are its operands and OPCODE is the mnemonic to be used. */
9235 output_v8plus_shift (rtx insn, rtx *operands, const char *opcode)
9237 static char asm_code[60];
9239 /* The scratch register is only required when the destination
9240 register is not a 64-bit global or out register. */
9241 if (which_alternative != 2)
9242 operands[3] = operands[0];
9244 /* We can only shift by constants <= 63. */
9245 if (GET_CODE (operands[2]) == CONST_INT)
9246 operands[2] = GEN_INT (INTVAL (operands[2]) & 0x3f);
9248 if (GET_CODE (operands[1]) == CONST_INT)
9250 output_asm_insn ("mov\t%1, %3", operands);
9254 output_asm_insn ("sllx\t%H1, 32, %3", operands);
9255 if (sparc_check_64 (operands[1], insn) <= 0)
9256 output_asm_insn ("srl\t%L1, 0, %L1", operands);
9257 output_asm_insn ("or\t%L1, %3, %3", operands);
9260 strcpy (asm_code, opcode);
9262 if (which_alternative != 2)
9263 return strcat (asm_code, "\t%0, %2, %L0\n\tsrlx\t%L0, 32, %H0");
9266 strcat (asm_code, "\t%3, %2, %3\n\tsrlx\t%3, 32, %H0\n\tmov\t%3, %L0");
9269 /* Output rtl to increment the profiler label LABELNO
9270 for profiling a function entry. */
9273 sparc_profile_hook (int labelno)
9278 fun = gen_rtx_SYMBOL_REF (Pmode, MCOUNT_FUNCTION);
9279 if (NO_PROFILE_COUNTERS)
9281 emit_library_call (fun, LCT_NORMAL, VOIDmode, 0);
9285 ASM_GENERATE_INTERNAL_LABEL (buf, "LP", labelno);
9286 lab = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (buf));
9287 emit_library_call (fun, LCT_NORMAL, VOIDmode, 1, lab, Pmode);
9291 #ifdef TARGET_SOLARIS
9292 /* Solaris implementation of TARGET_ASM_NAMED_SECTION. */
9295 sparc_solaris_elf_asm_named_section (const char *name, unsigned int flags,
9296 tree decl ATTRIBUTE_UNUSED)
9298 if (HAVE_COMDAT_GROUP && flags & SECTION_LINKONCE)
9300 solaris_elf_asm_comdat_section (name, flags, decl);
9304 fprintf (asm_out_file, "\t.section\t\"%s\"", name);
9306 if (!(flags & SECTION_DEBUG))
9307 fputs (",#alloc", asm_out_file);
9308 if (flags & SECTION_WRITE)
9309 fputs (",#write", asm_out_file);
9310 if (flags & SECTION_TLS)
9311 fputs (",#tls", asm_out_file);
9312 if (flags & SECTION_CODE)
9313 fputs (",#execinstr", asm_out_file);
9315 /* ??? Handle SECTION_BSS. */
9317 fputc ('\n', asm_out_file);
9319 #endif /* TARGET_SOLARIS */
9321 /* We do not allow indirect calls to be optimized into sibling calls.
9323 We cannot use sibling calls when delayed branches are disabled
9324 because they will likely require the call delay slot to be filled.
9326 Also, on SPARC 32-bit we cannot emit a sibling call when the
9327 current function returns a structure. This is because the "unimp
9328 after call" convention would cause the callee to return to the
9329 wrong place. The generic code already disallows cases where the
9330 function being called returns a structure.
9332 It may seem strange how this last case could occur. Usually there
9333 is code after the call which jumps to epilogue code which dumps the
9334 return value into the struct return area. That ought to invalidate
9335 the sibling call right? Well, in the C++ case we can end up passing
9336 the pointer to the struct return area to a constructor (which returns
9337 void) and then nothing else happens. Such a sibling call would look
9338 valid without the added check here.
9340 VxWorks PIC PLT entries require the global pointer to be initialized
9341 on entry. We therefore can't emit sibling calls to them. */
9343 sparc_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
9346 && flag_delayed_branch
9347 && (TARGET_ARCH64 || ! cfun->returns_struct)
9348 && !(TARGET_VXWORKS_RTP
9350 && !targetm.binds_local_p (decl)));
9353 /* libfunc renaming. */
9356 sparc_init_libfuncs (void)
9360 /* Use the subroutines that Sun's library provides for integer
9361 multiply and divide. The `*' prevents an underscore from
9362 being prepended by the compiler. .umul is a little faster
9364 set_optab_libfunc (smul_optab, SImode, "*.umul");
9365 set_optab_libfunc (sdiv_optab, SImode, "*.div");
9366 set_optab_libfunc (udiv_optab, SImode, "*.udiv");
9367 set_optab_libfunc (smod_optab, SImode, "*.rem");
9368 set_optab_libfunc (umod_optab, SImode, "*.urem");
9370 /* TFmode arithmetic. These names are part of the SPARC 32bit ABI. */
9371 set_optab_libfunc (add_optab, TFmode, "_Q_add");
9372 set_optab_libfunc (sub_optab, TFmode, "_Q_sub");
9373 set_optab_libfunc (neg_optab, TFmode, "_Q_neg");
9374 set_optab_libfunc (smul_optab, TFmode, "_Q_mul");
9375 set_optab_libfunc (sdiv_optab, TFmode, "_Q_div");
9377 /* We can define the TFmode sqrt optab only if TARGET_FPU. This
9378 is because with soft-float, the SFmode and DFmode sqrt
9379 instructions will be absent, and the compiler will notice and
9380 try to use the TFmode sqrt instruction for calls to the
9381 builtin function sqrt, but this fails. */
9383 set_optab_libfunc (sqrt_optab, TFmode, "_Q_sqrt");
9385 set_optab_libfunc (eq_optab, TFmode, "_Q_feq");
9386 set_optab_libfunc (ne_optab, TFmode, "_Q_fne");
9387 set_optab_libfunc (gt_optab, TFmode, "_Q_fgt");
9388 set_optab_libfunc (ge_optab, TFmode, "_Q_fge");
9389 set_optab_libfunc (lt_optab, TFmode, "_Q_flt");
9390 set_optab_libfunc (le_optab, TFmode, "_Q_fle");
9392 set_conv_libfunc (sext_optab, TFmode, SFmode, "_Q_stoq");
9393 set_conv_libfunc (sext_optab, TFmode, DFmode, "_Q_dtoq");
9394 set_conv_libfunc (trunc_optab, SFmode, TFmode, "_Q_qtos");
9395 set_conv_libfunc (trunc_optab, DFmode, TFmode, "_Q_qtod");
9397 set_conv_libfunc (sfix_optab, SImode, TFmode, "_Q_qtoi");
9398 set_conv_libfunc (ufix_optab, SImode, TFmode, "_Q_qtou");
9399 set_conv_libfunc (sfloat_optab, TFmode, SImode, "_Q_itoq");
9400 set_conv_libfunc (ufloat_optab, TFmode, SImode, "_Q_utoq");
9402 if (DITF_CONVERSION_LIBFUNCS)
9404 set_conv_libfunc (sfix_optab, DImode, TFmode, "_Q_qtoll");
9405 set_conv_libfunc (ufix_optab, DImode, TFmode, "_Q_qtoull");
9406 set_conv_libfunc (sfloat_optab, TFmode, DImode, "_Q_lltoq");
9407 set_conv_libfunc (ufloat_optab, TFmode, DImode, "_Q_ulltoq");
9410 if (SUN_CONVERSION_LIBFUNCS)
9412 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftoll");
9413 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoull");
9414 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtoll");
9415 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoull");
9420 /* In the SPARC 64bit ABI, SImode multiply and divide functions
9421 do not exist in the library. Make sure the compiler does not
9422 emit calls to them by accident. (It should always use the
9423 hardware instructions.) */
9424 set_optab_libfunc (smul_optab, SImode, 0);
9425 set_optab_libfunc (sdiv_optab, SImode, 0);
9426 set_optab_libfunc (udiv_optab, SImode, 0);
9427 set_optab_libfunc (smod_optab, SImode, 0);
9428 set_optab_libfunc (umod_optab, SImode, 0);
9430 if (SUN_INTEGER_MULTIPLY_64)
9432 set_optab_libfunc (smul_optab, DImode, "__mul64");
9433 set_optab_libfunc (sdiv_optab, DImode, "__div64");
9434 set_optab_libfunc (udiv_optab, DImode, "__udiv64");
9435 set_optab_libfunc (smod_optab, DImode, "__rem64");
9436 set_optab_libfunc (umod_optab, DImode, "__urem64");
9439 if (SUN_CONVERSION_LIBFUNCS)
9441 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftol");
9442 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoul");
9443 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtol");
9444 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoul");
9449 static tree def_builtin(const char *name, int code, tree type)
9451 return add_builtin_function(name, type, code, BUILT_IN_MD, NULL,
9455 static tree def_builtin_const(const char *name, int code, tree type)
9457 tree t = def_builtin(name, code, type);
9460 TREE_READONLY (t) = 1;
9465 /* Implement the TARGET_INIT_BUILTINS target hook.
9466 Create builtin functions for special SPARC instructions. */
9469 sparc_init_builtins (void)
9472 sparc_vis_init_builtins ();
9475 /* Create builtin functions for VIS 1.0 instructions. */
9478 sparc_vis_init_builtins (void)
9480 tree v4qi = build_vector_type (unsigned_intQI_type_node, 4);
9481 tree v8qi = build_vector_type (unsigned_intQI_type_node, 8);
9482 tree v4hi = build_vector_type (intHI_type_node, 4);
9483 tree v2hi = build_vector_type (intHI_type_node, 2);
9484 tree v2si = build_vector_type (intSI_type_node, 2);
9485 tree v1si = build_vector_type (intSI_type_node, 1);
9487 tree v4qi_ftype_v4hi = build_function_type_list (v4qi, v4hi, 0);
9488 tree v8qi_ftype_v2si_v8qi = build_function_type_list (v8qi, v2si, v8qi, 0);
9489 tree v2hi_ftype_v2si = build_function_type_list (v2hi, v2si, 0);
9490 tree v4hi_ftype_v4qi = build_function_type_list (v4hi, v4qi, 0);
9491 tree v8qi_ftype_v4qi_v4qi = build_function_type_list (v8qi, v4qi, v4qi, 0);
9492 tree v4hi_ftype_v4qi_v4hi = build_function_type_list (v4hi, v4qi, v4hi, 0);
9493 tree v4hi_ftype_v4qi_v2hi = build_function_type_list (v4hi, v4qi, v2hi, 0);
9494 tree v2si_ftype_v4qi_v2hi = build_function_type_list (v2si, v4qi, v2hi, 0);
9495 tree v4hi_ftype_v8qi_v4hi = build_function_type_list (v4hi, v8qi, v4hi, 0);
9496 tree v4hi_ftype_v4hi_v4hi = build_function_type_list (v4hi, v4hi, v4hi, 0);
9497 tree v2si_ftype_v2si_v2si = build_function_type_list (v2si, v2si, v2si, 0);
9498 tree v8qi_ftype_v8qi_v8qi = build_function_type_list (v8qi, v8qi, v8qi, 0);
9499 tree v2hi_ftype_v2hi_v2hi = build_function_type_list (v2hi, v2hi, v2hi, 0);
9500 tree v1si_ftype_v1si_v1si = build_function_type_list (v1si, v1si, v1si, 0);
9501 tree di_ftype_v8qi_v8qi_di = build_function_type_list (intDI_type_node,
9503 intDI_type_node, 0);
9504 tree di_ftype_v8qi_v8qi = build_function_type_list (intDI_type_node,
9506 tree si_ftype_v8qi_v8qi = build_function_type_list (intSI_type_node,
9508 tree di_ftype_di_di = build_function_type_list (intDI_type_node,
9510 intDI_type_node, 0);
9511 tree si_ftype_si_si = build_function_type_list (intSI_type_node,
9513 intSI_type_node, 0);
9514 tree ptr_ftype_ptr_si = build_function_type_list (ptr_type_node,
9516 intSI_type_node, 0);
9517 tree ptr_ftype_ptr_di = build_function_type_list (ptr_type_node,
9519 intDI_type_node, 0);
9520 tree si_ftype_ptr_ptr = build_function_type_list (intSI_type_node,
9523 tree di_ftype_ptr_ptr = build_function_type_list (intDI_type_node,
9526 tree si_ftype_v4hi_v4hi = build_function_type_list (intSI_type_node,
9528 tree si_ftype_v2si_v2si = build_function_type_list (intSI_type_node,
9530 tree di_ftype_v4hi_v4hi = build_function_type_list (intDI_type_node,
9532 tree di_ftype_v2si_v2si = build_function_type_list (intDI_type_node,
9534 tree void_ftype_di = build_function_type_list (void_type_node,
9535 intDI_type_node, 0);
9536 tree di_ftype_void = build_function_type_list (intDI_type_node,
9538 tree void_ftype_si = build_function_type_list (void_type_node,
9539 intSI_type_node, 0);
9540 tree sf_ftype_sf_sf = build_function_type_list (float_type_node,
9542 float_type_node, 0);
9543 tree df_ftype_df_df = build_function_type_list (double_type_node,
9545 double_type_node, 0);
9547 /* Packing and expanding vectors. */
9548 def_builtin ("__builtin_vis_fpack16", CODE_FOR_fpack16_vis,
9550 def_builtin ("__builtin_vis_fpack32", CODE_FOR_fpack32_vis,
9551 v8qi_ftype_v2si_v8qi);
9552 def_builtin ("__builtin_vis_fpackfix", CODE_FOR_fpackfix_vis,
9554 def_builtin_const ("__builtin_vis_fexpand", CODE_FOR_fexpand_vis,
9556 def_builtin_const ("__builtin_vis_fpmerge", CODE_FOR_fpmerge_vis,
9557 v8qi_ftype_v4qi_v4qi);
9559 /* Multiplications. */
9560 def_builtin_const ("__builtin_vis_fmul8x16", CODE_FOR_fmul8x16_vis,
9561 v4hi_ftype_v4qi_v4hi);
9562 def_builtin_const ("__builtin_vis_fmul8x16au", CODE_FOR_fmul8x16au_vis,
9563 v4hi_ftype_v4qi_v2hi);
9564 def_builtin_const ("__builtin_vis_fmul8x16al", CODE_FOR_fmul8x16al_vis,
9565 v4hi_ftype_v4qi_v2hi);
9566 def_builtin_const ("__builtin_vis_fmul8sux16", CODE_FOR_fmul8sux16_vis,
9567 v4hi_ftype_v8qi_v4hi);
9568 def_builtin_const ("__builtin_vis_fmul8ulx16", CODE_FOR_fmul8ulx16_vis,
9569 v4hi_ftype_v8qi_v4hi);
9570 def_builtin_const ("__builtin_vis_fmuld8sux16", CODE_FOR_fmuld8sux16_vis,
9571 v2si_ftype_v4qi_v2hi);
9572 def_builtin_const ("__builtin_vis_fmuld8ulx16", CODE_FOR_fmuld8ulx16_vis,
9573 v2si_ftype_v4qi_v2hi);
9575 /* Data aligning. */
9576 def_builtin ("__builtin_vis_faligndatav4hi", CODE_FOR_faligndatav4hi_vis,
9577 v4hi_ftype_v4hi_v4hi);
9578 def_builtin ("__builtin_vis_faligndatav8qi", CODE_FOR_faligndatav8qi_vis,
9579 v8qi_ftype_v8qi_v8qi);
9580 def_builtin ("__builtin_vis_faligndatav2si", CODE_FOR_faligndatav2si_vis,
9581 v2si_ftype_v2si_v2si);
9582 def_builtin ("__builtin_vis_faligndatadi", CODE_FOR_faligndatav1di_vis,
9585 def_builtin ("__builtin_vis_write_gsr", CODE_FOR_wrgsr_vis,
9587 def_builtin ("__builtin_vis_read_gsr", CODE_FOR_rdgsr_vis,
9592 def_builtin ("__builtin_vis_alignaddr", CODE_FOR_alignaddrdi_vis,
9594 def_builtin ("__builtin_vis_alignaddrl", CODE_FOR_alignaddrldi_vis,
9599 def_builtin ("__builtin_vis_alignaddr", CODE_FOR_alignaddrsi_vis,
9601 def_builtin ("__builtin_vis_alignaddrl", CODE_FOR_alignaddrlsi_vis,
9605 /* Pixel distance. */
9606 def_builtin_const ("__builtin_vis_pdist", CODE_FOR_pdist_vis,
9607 di_ftype_v8qi_v8qi_di);
9609 /* Edge handling. */
9612 def_builtin_const ("__builtin_vis_edge8", CODE_FOR_edge8di_vis,
9614 def_builtin_const ("__builtin_vis_edge8l", CODE_FOR_edge8ldi_vis,
9616 def_builtin_const ("__builtin_vis_edge16", CODE_FOR_edge16di_vis,
9618 def_builtin_const ("__builtin_vis_edge16l", CODE_FOR_edge16ldi_vis,
9620 def_builtin_const ("__builtin_vis_edge32", CODE_FOR_edge32di_vis,
9622 def_builtin_const ("__builtin_vis_edge32l", CODE_FOR_edge32ldi_vis,
9626 def_builtin_const ("__builtin_vis_edge8n", CODE_FOR_edge8ndi_vis,
9628 def_builtin_const ("__builtin_vis_edge8ln", CODE_FOR_edge8lndi_vis,
9630 def_builtin_const ("__builtin_vis_edge16n", CODE_FOR_edge16ndi_vis,
9632 def_builtin_const ("__builtin_vis_edge16ln", CODE_FOR_edge16lndi_vis,
9634 def_builtin_const ("__builtin_vis_edge32n", CODE_FOR_edge32ndi_vis,
9636 def_builtin_const ("__builtin_vis_edge32ln", CODE_FOR_edge32lndi_vis,
9642 def_builtin_const ("__builtin_vis_edge8", CODE_FOR_edge8si_vis,
9644 def_builtin_const ("__builtin_vis_edge8l", CODE_FOR_edge8lsi_vis,
9646 def_builtin_const ("__builtin_vis_edge16", CODE_FOR_edge16si_vis,
9648 def_builtin_const ("__builtin_vis_edge16l", CODE_FOR_edge16lsi_vis,
9650 def_builtin_const ("__builtin_vis_edge32", CODE_FOR_edge32si_vis,
9652 def_builtin_const ("__builtin_vis_edge32l", CODE_FOR_edge32lsi_vis,
9656 def_builtin_const ("__builtin_vis_edge8n", CODE_FOR_edge8nsi_vis,
9658 def_builtin_const ("__builtin_vis_edge8ln", CODE_FOR_edge8lnsi_vis,
9660 def_builtin_const ("__builtin_vis_edge16n", CODE_FOR_edge16nsi_vis,
9662 def_builtin_const ("__builtin_vis_edge16ln", CODE_FOR_edge16lnsi_vis,
9664 def_builtin_const ("__builtin_vis_edge32n", CODE_FOR_edge32nsi_vis,
9666 def_builtin_const ("__builtin_vis_edge32ln", CODE_FOR_edge32lnsi_vis,
9671 /* Pixel compare. */
9674 def_builtin_const ("__builtin_vis_fcmple16", CODE_FOR_fcmple16di_vis,
9675 di_ftype_v4hi_v4hi);
9676 def_builtin_const ("__builtin_vis_fcmple32", CODE_FOR_fcmple32di_vis,
9677 di_ftype_v2si_v2si);
9678 def_builtin_const ("__builtin_vis_fcmpne16", CODE_FOR_fcmpne16di_vis,
9679 di_ftype_v4hi_v4hi);
9680 def_builtin_const ("__builtin_vis_fcmpne32", CODE_FOR_fcmpne32di_vis,
9681 di_ftype_v2si_v2si);
9682 def_builtin_const ("__builtin_vis_fcmpgt16", CODE_FOR_fcmpgt16di_vis,
9683 di_ftype_v4hi_v4hi);
9684 def_builtin_const ("__builtin_vis_fcmpgt32", CODE_FOR_fcmpgt32di_vis,
9685 di_ftype_v2si_v2si);
9686 def_builtin_const ("__builtin_vis_fcmpeq16", CODE_FOR_fcmpeq16di_vis,
9687 di_ftype_v4hi_v4hi);
9688 def_builtin_const ("__builtin_vis_fcmpeq32", CODE_FOR_fcmpeq32di_vis,
9689 di_ftype_v2si_v2si);
9693 def_builtin_const ("__builtin_vis_fcmple16", CODE_FOR_fcmple16si_vis,
9694 si_ftype_v4hi_v4hi);
9695 def_builtin_const ("__builtin_vis_fcmple32", CODE_FOR_fcmple32si_vis,
9696 si_ftype_v2si_v2si);
9697 def_builtin_const ("__builtin_vis_fcmpne16", CODE_FOR_fcmpne16si_vis,
9698 si_ftype_v4hi_v4hi);
9699 def_builtin_const ("__builtin_vis_fcmpne32", CODE_FOR_fcmpne32si_vis,
9700 si_ftype_v2si_v2si);
9701 def_builtin_const ("__builtin_vis_fcmpgt16", CODE_FOR_fcmpgt16si_vis,
9702 si_ftype_v4hi_v4hi);
9703 def_builtin_const ("__builtin_vis_fcmpgt32", CODE_FOR_fcmpgt32si_vis,
9704 si_ftype_v2si_v2si);
9705 def_builtin_const ("__builtin_vis_fcmpeq16", CODE_FOR_fcmpeq16si_vis,
9706 si_ftype_v4hi_v4hi);
9707 def_builtin_const ("__builtin_vis_fcmpeq32", CODE_FOR_fcmpeq32si_vis,
9708 si_ftype_v2si_v2si);
9711 /* Addition and subtraction. */
9712 def_builtin_const ("__builtin_vis_fpadd16", CODE_FOR_addv4hi3,
9713 v4hi_ftype_v4hi_v4hi);
9714 def_builtin_const ("__builtin_vis_fpadd16s", CODE_FOR_addv2hi3,
9715 v2hi_ftype_v2hi_v2hi);
9716 def_builtin_const ("__builtin_vis_fpadd32", CODE_FOR_addv2si3,
9717 v2si_ftype_v2si_v2si);
9718 def_builtin_const ("__builtin_vis_fpadd32s", CODE_FOR_addv1si3,
9719 v1si_ftype_v1si_v1si);
9720 def_builtin_const ("__builtin_vis_fpsub16", CODE_FOR_subv4hi3,
9721 v4hi_ftype_v4hi_v4hi);
9722 def_builtin_const ("__builtin_vis_fpsub16s", CODE_FOR_subv2hi3,
9723 v2hi_ftype_v2hi_v2hi);
9724 def_builtin_const ("__builtin_vis_fpsub32", CODE_FOR_subv2si3,
9725 v2si_ftype_v2si_v2si);
9726 def_builtin_const ("__builtin_vis_fpsub32s", CODE_FOR_subv1si3,
9727 v1si_ftype_v1si_v1si);
9729 /* Three-dimensional array addressing. */
9732 def_builtin_const ("__builtin_vis_array8", CODE_FOR_array8di_vis,
9734 def_builtin_const ("__builtin_vis_array16", CODE_FOR_array16di_vis,
9736 def_builtin_const ("__builtin_vis_array32", CODE_FOR_array32di_vis,
9741 def_builtin_const ("__builtin_vis_array8", CODE_FOR_array8si_vis,
9743 def_builtin_const ("__builtin_vis_array16", CODE_FOR_array16si_vis,
9745 def_builtin_const ("__builtin_vis_array32", CODE_FOR_array32si_vis,
9751 /* Byte mask and shuffle */
9753 def_builtin ("__builtin_vis_bmask", CODE_FOR_bmaskdi_vis,
9756 def_builtin ("__builtin_vis_bmask", CODE_FOR_bmasksi_vis,
9758 def_builtin ("__builtin_vis_bshufflev4hi", CODE_FOR_bshufflev4hi_vis,
9759 v4hi_ftype_v4hi_v4hi);
9760 def_builtin ("__builtin_vis_bshufflev8qi", CODE_FOR_bshufflev8qi_vis,
9761 v8qi_ftype_v8qi_v8qi);
9762 def_builtin ("__builtin_vis_bshufflev2si", CODE_FOR_bshufflev2si_vis,
9763 v2si_ftype_v2si_v2si);
9764 def_builtin ("__builtin_vis_bshuffledi", CODE_FOR_bshufflev1di_vis,
9772 def_builtin ("__builtin_vis_cmask8", CODE_FOR_cmask8di_vis,
9774 def_builtin ("__builtin_vis_cmask16", CODE_FOR_cmask16di_vis,
9776 def_builtin ("__builtin_vis_cmask32", CODE_FOR_cmask32di_vis,
9781 def_builtin ("__builtin_vis_cmask8", CODE_FOR_cmask8si_vis,
9783 def_builtin ("__builtin_vis_cmask16", CODE_FOR_cmask16si_vis,
9785 def_builtin ("__builtin_vis_cmask32", CODE_FOR_cmask32si_vis,
9789 def_builtin_const ("__builtin_vis_fchksm16", CODE_FOR_fchksm16_vis,
9790 v4hi_ftype_v4hi_v4hi);
9792 def_builtin_const ("__builtin_vis_fsll16", CODE_FOR_vashlv4hi3,
9793 v4hi_ftype_v4hi_v4hi);
9794 def_builtin_const ("__builtin_vis_fslas16", CODE_FOR_vssashlv4hi3,
9795 v4hi_ftype_v4hi_v4hi);
9796 def_builtin_const ("__builtin_vis_fsrl16", CODE_FOR_vlshrv4hi3,
9797 v4hi_ftype_v4hi_v4hi);
9798 def_builtin_const ("__builtin_vis_fsra16", CODE_FOR_vashrv4hi3,
9799 v4hi_ftype_v4hi_v4hi);
9800 def_builtin_const ("__builtin_vis_fsll32", CODE_FOR_vashlv2si3,
9801 v2si_ftype_v2si_v2si);
9802 def_builtin_const ("__builtin_vis_fslas32", CODE_FOR_vssashlv2si3,
9803 v2si_ftype_v2si_v2si);
9804 def_builtin_const ("__builtin_vis_fsrl32", CODE_FOR_vlshrv2si3,
9805 v2si_ftype_v2si_v2si);
9806 def_builtin_const ("__builtin_vis_fsra32", CODE_FOR_vashrv2si3,
9807 v2si_ftype_v2si_v2si);
9810 def_builtin_const ("__builtin_vis_pdistn", CODE_FOR_pdistndi_vis,
9811 di_ftype_v8qi_v8qi);
9813 def_builtin_const ("__builtin_vis_pdistn", CODE_FOR_pdistnsi_vis,
9814 si_ftype_v8qi_v8qi);
9816 def_builtin_const ("__builtin_vis_fmean16", CODE_FOR_fmean16_vis,
9817 v4hi_ftype_v4hi_v4hi);
9818 def_builtin_const ("__builtin_vis_fpadd64", CODE_FOR_fpadd64_vis,
9820 def_builtin_const ("__builtin_vis_fpsub64", CODE_FOR_fpsub64_vis,
9823 def_builtin_const ("__builtin_vis_fpadds16", CODE_FOR_ssaddv4hi3,
9824 v4hi_ftype_v4hi_v4hi);
9825 def_builtin_const ("__builtin_vis_fpadds16s", CODE_FOR_ssaddv2hi3,
9826 v2hi_ftype_v2hi_v2hi);
9827 def_builtin_const ("__builtin_vis_fpsubs16", CODE_FOR_sssubv4hi3,
9828 v4hi_ftype_v4hi_v4hi);
9829 def_builtin_const ("__builtin_vis_fpsubs16s", CODE_FOR_sssubv2hi3,
9830 v2hi_ftype_v2hi_v2hi);
9831 def_builtin_const ("__builtin_vis_fpadds32", CODE_FOR_ssaddv2si3,
9832 v2si_ftype_v2si_v2si);
9833 def_builtin_const ("__builtin_vis_fpadds32s", CODE_FOR_ssaddv1si3,
9834 v1si_ftype_v1si_v1si);
9835 def_builtin_const ("__builtin_vis_fpsubs32", CODE_FOR_sssubv2si3,
9836 v2si_ftype_v2si_v2si);
9837 def_builtin_const ("__builtin_vis_fpsubs32s", CODE_FOR_sssubv1si3,
9838 v1si_ftype_v1si_v1si);
9842 def_builtin_const ("__builtin_vis_fucmple8", CODE_FOR_fucmple8di_vis,
9843 di_ftype_v8qi_v8qi);
9844 def_builtin_const ("__builtin_vis_fucmpne8", CODE_FOR_fucmpne8di_vis,
9845 di_ftype_v8qi_v8qi);
9846 def_builtin_const ("__builtin_vis_fucmpgt8", CODE_FOR_fucmpgt8di_vis,
9847 di_ftype_v8qi_v8qi);
9848 def_builtin_const ("__builtin_vis_fucmpeq8", CODE_FOR_fucmpeq8di_vis,
9849 di_ftype_v8qi_v8qi);
9853 def_builtin_const ("__builtin_vis_fucmple8", CODE_FOR_fucmple8si_vis,
9854 si_ftype_v8qi_v8qi);
9855 def_builtin_const ("__builtin_vis_fucmpne8", CODE_FOR_fucmpne8si_vis,
9856 si_ftype_v8qi_v8qi);
9857 def_builtin_const ("__builtin_vis_fucmpgt8", CODE_FOR_fucmpgt8si_vis,
9858 si_ftype_v8qi_v8qi);
9859 def_builtin_const ("__builtin_vis_fucmpeq8", CODE_FOR_fucmpeq8si_vis,
9860 si_ftype_v8qi_v8qi);
9863 def_builtin_const ("__builtin_vis_fhadds", CODE_FOR_fhaddsf_vis,
9865 def_builtin_const ("__builtin_vis_fhaddd", CODE_FOR_fhadddf_vis,
9867 def_builtin_const ("__builtin_vis_fhsubs", CODE_FOR_fhsubsf_vis,
9869 def_builtin_const ("__builtin_vis_fhsubd", CODE_FOR_fhsubdf_vis,
9871 def_builtin_const ("__builtin_vis_fnhadds", CODE_FOR_fnhaddsf_vis,
9873 def_builtin_const ("__builtin_vis_fnhaddd", CODE_FOR_fnhadddf_vis,
9876 def_builtin_const ("__builtin_vis_umulxhi", CODE_FOR_umulxhi_vis,
9878 def_builtin_const ("__builtin_vis_xmulx", CODE_FOR_xmulx_vis,
9880 def_builtin_const ("__builtin_vis_xmulxhi", CODE_FOR_xmulxhi_vis,
9885 /* Handle TARGET_EXPAND_BUILTIN target hook.
9886 Expand builtin functions for sparc intrinsics. */
9889 sparc_expand_builtin (tree exp, rtx target,
9890 rtx subtarget ATTRIBUTE_UNUSED,
9891 enum machine_mode tmode ATTRIBUTE_UNUSED,
9892 int ignore ATTRIBUTE_UNUSED)
9895 call_expr_arg_iterator iter;
9896 tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0);
9897 unsigned int icode = DECL_FUNCTION_CODE (fndecl);
9902 nonvoid = TREE_TYPE (TREE_TYPE (fndecl)) != void_type_node;
9906 enum machine_mode tmode = insn_data[icode].operand[0].mode;
9908 || GET_MODE (target) != tmode
9909 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
9910 op[0] = gen_reg_rtx (tmode);
9914 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
9916 const struct insn_operand_data *insn_op;
9919 if (arg == error_mark_node)
9923 idx = arg_count - !nonvoid;
9924 insn_op = &insn_data[icode].operand[idx];
9925 op[arg_count] = expand_normal (arg);
9927 if (insn_op->mode == V1DImode
9928 && GET_MODE (op[arg_count]) == DImode)
9929 op[arg_count] = gen_lowpart (V1DImode, op[arg_count]);
9930 else if (insn_op->mode == V1SImode
9931 && GET_MODE (op[arg_count]) == SImode)
9932 op[arg_count] = gen_lowpart (V1SImode, op[arg_count]);
9934 if (! (*insn_data[icode].operand[idx].predicate) (op[arg_count],
9936 op[arg_count] = copy_to_mode_reg (insn_op->mode, op[arg_count]);
9942 pat = GEN_FCN (icode) (op[0]);
9946 pat = GEN_FCN (icode) (op[0], op[1]);
9948 pat = GEN_FCN (icode) (op[1]);
9951 pat = GEN_FCN (icode) (op[0], op[1], op[2]);
9954 pat = GEN_FCN (icode) (op[0], op[1], op[2], op[3]);
9972 sparc_vis_mul8x16 (int e8, int e16)
9974 return (e8 * e16 + 128) / 256;
9977 /* Multiply the VECTOR_CSTs CST0 and CST1 as specified by FNCODE and put
9978 the result into the array N_ELTS, whose elements are of INNER_TYPE. */
9981 sparc_handle_vis_mul8x16 (tree *n_elts, int fncode, tree inner_type,
9982 tree cst0, tree cst1)
9984 unsigned i, num = VECTOR_CST_NELTS (cst0);
9989 case CODE_FOR_fmul8x16_vis:
9990 for (i = 0; i < num; ++i)
9993 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (VECTOR_CST_ELT (cst0, i)),
9994 TREE_INT_CST_LOW (VECTOR_CST_ELT (cst1, i)));
9995 n_elts[i] = build_int_cst (inner_type, val);
9999 case CODE_FOR_fmul8x16au_vis:
10000 scale = TREE_INT_CST_LOW (VECTOR_CST_ELT (cst1, 0));
10002 for (i = 0; i < num; ++i)
10005 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (VECTOR_CST_ELT (cst0, i)),
10007 n_elts[i] = build_int_cst (inner_type, val);
10011 case CODE_FOR_fmul8x16al_vis:
10012 scale = TREE_INT_CST_LOW (VECTOR_CST_ELT (cst1, 1));
10014 for (i = 0; i < num; ++i)
10017 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (VECTOR_CST_ELT (cst0, i)),
10019 n_elts[i] = build_int_cst (inner_type, val);
10024 gcc_unreachable ();
10028 /* Handle TARGET_FOLD_BUILTIN target hook.
10029 Fold builtin functions for SPARC intrinsics. If IGNORE is true the
10030 result of the function call is ignored. NULL_TREE is returned if the
10031 function could not be folded. */
10034 sparc_fold_builtin (tree fndecl, int n_args ATTRIBUTE_UNUSED,
10035 tree *args, bool ignore)
10037 tree arg0, arg1, arg2;
10038 tree rtype = TREE_TYPE (TREE_TYPE (fndecl));
10039 enum insn_code icode = (enum insn_code) DECL_FUNCTION_CODE (fndecl);
10043 /* Note that a switch statement instead of the sequence of tests would
10044 be incorrect as many of the CODE_FOR values could be CODE_FOR_nothing
10045 and that would yield multiple alternatives with identical values. */
10046 if (icode == CODE_FOR_alignaddrsi_vis
10047 || icode == CODE_FOR_alignaddrdi_vis
10048 || icode == CODE_FOR_wrgsr_vis
10049 || icode == CODE_FOR_bmasksi_vis
10050 || icode == CODE_FOR_bmaskdi_vis
10051 || icode == CODE_FOR_cmask8si_vis
10052 || icode == CODE_FOR_cmask8di_vis
10053 || icode == CODE_FOR_cmask16si_vis
10054 || icode == CODE_FOR_cmask16di_vis
10055 || icode == CODE_FOR_cmask32si_vis
10056 || icode == CODE_FOR_cmask32di_vis)
10059 return build_zero_cst (rtype);
10064 case CODE_FOR_fexpand_vis:
10068 if (TREE_CODE (arg0) == VECTOR_CST)
10070 tree inner_type = TREE_TYPE (rtype);
10074 n_elts = XALLOCAVEC (tree, VECTOR_CST_NELTS (arg0));
10075 for (i = 0; i < VECTOR_CST_NELTS (arg0); ++i)
10076 n_elts[i] = build_int_cst (inner_type,
10078 (VECTOR_CST_ELT (arg0, i)) << 4);
10079 return build_vector (rtype, n_elts);
10083 case CODE_FOR_fmul8x16_vis:
10084 case CODE_FOR_fmul8x16au_vis:
10085 case CODE_FOR_fmul8x16al_vis:
10091 if (TREE_CODE (arg0) == VECTOR_CST && TREE_CODE (arg1) == VECTOR_CST)
10093 tree inner_type = TREE_TYPE (rtype);
10094 tree *n_elts = XALLOCAVEC (tree, VECTOR_CST_NELTS (arg0));
10095 sparc_handle_vis_mul8x16 (n_elts, icode, inner_type, arg0, arg1);
10096 return build_vector (rtype, n_elts);
10100 case CODE_FOR_fpmerge_vis:
10106 if (TREE_CODE (arg0) == VECTOR_CST && TREE_CODE (arg1) == VECTOR_CST)
10108 tree *n_elts = XALLOCAVEC (tree, 2 * VECTOR_CST_NELTS (arg0));
10110 for (i = 0; i < VECTOR_CST_NELTS (arg0); ++i)
10112 n_elts[2*i] = VECTOR_CST_ELT (arg0, i);
10113 n_elts[2*i+1] = VECTOR_CST_ELT (arg1, i);
10116 return build_vector (rtype, n_elts);
10120 case CODE_FOR_pdist_vis:
10128 if (TREE_CODE (arg0) == VECTOR_CST
10129 && TREE_CODE (arg1) == VECTOR_CST
10130 && TREE_CODE (arg2) == INTEGER_CST)
10133 unsigned HOST_WIDE_INT low = TREE_INT_CST_LOW (arg2);
10134 HOST_WIDE_INT high = TREE_INT_CST_HIGH (arg2);
10137 for (i = 0; i < VECTOR_CST_NELTS (arg0); ++i)
10139 unsigned HOST_WIDE_INT
10140 low0 = TREE_INT_CST_LOW (VECTOR_CST_ELT (arg0, i)),
10141 low1 = TREE_INT_CST_LOW (VECTOR_CST_ELT (arg1, i));
10143 high0 = TREE_INT_CST_HIGH (VECTOR_CST_ELT (arg0, i));
10145 high1 = TREE_INT_CST_HIGH (VECTOR_CST_ELT (arg1, i));
10147 unsigned HOST_WIDE_INT l;
10150 overflow |= neg_double (low1, high1, &l, &h);
10151 overflow |= add_double (low0, high0, l, h, &l, &h);
10153 overflow |= neg_double (l, h, &l, &h);
10155 overflow |= add_double (low, high, l, h, &low, &high);
10158 gcc_assert (overflow == 0);
10160 return build_int_cst_wide (rtype, low, high);
10170 /* ??? This duplicates information provided to the compiler by the
10171 ??? scheduler description. Some day, teach genautomata to output
10172 ??? the latencies and then CSE will just use that. */
10175 sparc_rtx_costs (rtx x, int code, int outer_code, int opno ATTRIBUTE_UNUSED,
10176 int *total, bool speed ATTRIBUTE_UNUSED)
10178 enum machine_mode mode = GET_MODE (x);
10179 bool float_mode_p = FLOAT_MODE_P (mode);
10184 if (INTVAL (x) < 0x1000 && INTVAL (x) >= -0x1000)
10202 if (GET_MODE (x) == VOIDmode
10203 && ((CONST_DOUBLE_HIGH (x) == 0
10204 && CONST_DOUBLE_LOW (x) < 0x1000)
10205 || (CONST_DOUBLE_HIGH (x) == -1
10206 && CONST_DOUBLE_LOW (x) < 0
10207 && CONST_DOUBLE_LOW (x) >= -0x1000)))
10214 /* If outer-code was a sign or zero extension, a cost
10215 of COSTS_N_INSNS (1) was already added in. This is
10216 why we are subtracting it back out. */
10217 if (outer_code == ZERO_EXTEND)
10219 *total = sparc_costs->int_zload - COSTS_N_INSNS (1);
10221 else if (outer_code == SIGN_EXTEND)
10223 *total = sparc_costs->int_sload - COSTS_N_INSNS (1);
10225 else if (float_mode_p)
10227 *total = sparc_costs->float_load;
10231 *total = sparc_costs->int_load;
10239 *total = sparc_costs->float_plusminus;
10241 *total = COSTS_N_INSNS (1);
10248 gcc_assert (float_mode_p);
10249 *total = sparc_costs->float_mul;
10252 if (GET_CODE (sub) == NEG)
10253 sub = XEXP (sub, 0);
10254 *total += rtx_cost (sub, FMA, 0, speed);
10257 if (GET_CODE (sub) == NEG)
10258 sub = XEXP (sub, 0);
10259 *total += rtx_cost (sub, FMA, 2, speed);
10265 *total = sparc_costs->float_mul;
10266 else if (! TARGET_HARD_MUL)
10267 *total = COSTS_N_INSNS (25);
10273 if (sparc_costs->int_mul_bit_factor)
10277 if (GET_CODE (XEXP (x, 1)) == CONST_INT)
10279 unsigned HOST_WIDE_INT value = INTVAL (XEXP (x, 1));
10280 for (nbits = 0; value != 0; value &= value - 1)
10283 else if (GET_CODE (XEXP (x, 1)) == CONST_DOUBLE
10284 && GET_MODE (XEXP (x, 1)) == VOIDmode)
10286 rtx x1 = XEXP (x, 1);
10287 unsigned HOST_WIDE_INT value1 = CONST_DOUBLE_LOW (x1);
10288 unsigned HOST_WIDE_INT value2 = CONST_DOUBLE_HIGH (x1);
10290 for (nbits = 0; value1 != 0; value1 &= value1 - 1)
10292 for (; value2 != 0; value2 &= value2 - 1)
10300 bit_cost = (nbits - 3) / sparc_costs->int_mul_bit_factor;
10301 bit_cost = COSTS_N_INSNS (bit_cost);
10304 if (mode == DImode)
10305 *total = sparc_costs->int_mulX + bit_cost;
10307 *total = sparc_costs->int_mul + bit_cost;
10314 *total = COSTS_N_INSNS (1) + sparc_costs->shift_penalty;
10323 if (mode == DFmode)
10324 *total = sparc_costs->float_div_df;
10326 *total = sparc_costs->float_div_sf;
10330 if (mode == DImode)
10331 *total = sparc_costs->int_divX;
10333 *total = sparc_costs->int_div;
10338 if (! float_mode_p)
10340 *total = COSTS_N_INSNS (1);
10347 case UNSIGNED_FLOAT:
10351 case FLOAT_TRUNCATE:
10352 *total = sparc_costs->float_move;
10356 if (mode == DFmode)
10357 *total = sparc_costs->float_sqrt_df;
10359 *total = sparc_costs->float_sqrt_sf;
10364 *total = sparc_costs->float_cmp;
10366 *total = COSTS_N_INSNS (1);
10371 *total = sparc_costs->float_cmove;
10373 *total = sparc_costs->int_cmove;
10377 /* Handle the NAND vector patterns. */
10378 if (sparc_vector_mode_supported_p (GET_MODE (x))
10379 && GET_CODE (XEXP (x, 0)) == NOT
10380 && GET_CODE (XEXP (x, 1)) == NOT)
10382 *total = COSTS_N_INSNS (1);
10393 /* Return true if CLASS is either GENERAL_REGS or I64_REGS. */
10396 general_or_i64_p (reg_class_t rclass)
10398 return (rclass == GENERAL_REGS || rclass == I64_REGS);
10401 /* Implement TARGET_REGISTER_MOVE_COST. */
10404 sparc_register_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED,
10405 reg_class_t from, reg_class_t to)
10407 bool need_memory = false;
10409 if (from == FPCC_REGS || to == FPCC_REGS)
10410 need_memory = true;
10411 else if ((FP_REG_CLASS_P (from) && general_or_i64_p (to))
10412 || (general_or_i64_p (from) && FP_REG_CLASS_P (to)))
10416 int size = GET_MODE_SIZE (mode);
10417 if (size == 8 || size == 4)
10419 if (! TARGET_ARCH32 || size == 4)
10425 need_memory = true;
10430 if (sparc_cpu == PROCESSOR_ULTRASPARC
10431 || sparc_cpu == PROCESSOR_ULTRASPARC3
10432 || sparc_cpu == PROCESSOR_NIAGARA
10433 || sparc_cpu == PROCESSOR_NIAGARA2
10434 || sparc_cpu == PROCESSOR_NIAGARA3
10435 || sparc_cpu == PROCESSOR_NIAGARA4)
10444 /* Emit the sequence of insns SEQ while preserving the registers REG and REG2.
10445 This is achieved by means of a manual dynamic stack space allocation in
10446 the current frame. We make the assumption that SEQ doesn't contain any
10447 function calls, with the possible exception of calls to the GOT helper. */
10450 emit_and_preserve (rtx seq, rtx reg, rtx reg2)
10452 /* We must preserve the lowest 16 words for the register save area. */
10453 HOST_WIDE_INT offset = 16*UNITS_PER_WORD;
10454 /* We really need only 2 words of fresh stack space. */
10455 HOST_WIDE_INT size = SPARC_STACK_ALIGN (offset + 2*UNITS_PER_WORD);
10458 = gen_rtx_MEM (word_mode, plus_constant (Pmode, stack_pointer_rtx,
10459 SPARC_STACK_BIAS + offset));
10461 emit_insn (gen_stack_pointer_dec (GEN_INT (size)));
10462 emit_insn (gen_rtx_SET (VOIDmode, slot, reg));
10464 emit_insn (gen_rtx_SET (VOIDmode,
10465 adjust_address (slot, word_mode, UNITS_PER_WORD),
10469 emit_insn (gen_rtx_SET (VOIDmode,
10471 adjust_address (slot, word_mode, UNITS_PER_WORD)));
10472 emit_insn (gen_rtx_SET (VOIDmode, reg, slot));
10473 emit_insn (gen_stack_pointer_inc (GEN_INT (size)));
10476 /* Output the assembler code for a thunk function. THUNK_DECL is the
10477 declaration for the thunk function itself, FUNCTION is the decl for
10478 the target function. DELTA is an immediate constant offset to be
10479 added to THIS. If VCALL_OFFSET is nonzero, the word at address
10480 (*THIS + VCALL_OFFSET) should be additionally added to THIS. */
10483 sparc_output_mi_thunk (FILE *file, tree thunk_fndecl ATTRIBUTE_UNUSED,
10484 HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset,
10487 rtx this_rtx, insn, funexp;
10488 unsigned int int_arg_first;
10490 reload_completed = 1;
10491 epilogue_completed = 1;
10493 emit_note (NOTE_INSN_PROLOGUE_END);
10497 sparc_leaf_function_p = 1;
10499 int_arg_first = SPARC_OUTGOING_INT_ARG_FIRST;
10501 else if (flag_delayed_branch)
10503 /* We will emit a regular sibcall below, so we need to instruct
10504 output_sibcall that we are in a leaf function. */
10505 sparc_leaf_function_p = crtl->uses_only_leaf_regs = 1;
10507 /* This will cause final.c to invoke leaf_renumber_regs so we
10508 must behave as if we were in a not-yet-leafified function. */
10509 int_arg_first = SPARC_INCOMING_INT_ARG_FIRST;
10513 /* We will emit the sibcall manually below, so we will need to
10514 manually spill non-leaf registers. */
10515 sparc_leaf_function_p = crtl->uses_only_leaf_regs = 0;
10517 /* We really are in a leaf function. */
10518 int_arg_first = SPARC_OUTGOING_INT_ARG_FIRST;
10521 /* Find the "this" pointer. Normally in %o0, but in ARCH64 if the function
10522 returns a structure, the structure return pointer is there instead. */
10524 && aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function))
10525 this_rtx = gen_rtx_REG (Pmode, int_arg_first + 1);
10527 this_rtx = gen_rtx_REG (Pmode, int_arg_first);
10529 /* Add DELTA. When possible use a plain add, otherwise load it into
10530 a register first. */
10533 rtx delta_rtx = GEN_INT (delta);
10535 if (! SPARC_SIMM13_P (delta))
10537 rtx scratch = gen_rtx_REG (Pmode, 1);
10538 emit_move_insn (scratch, delta_rtx);
10539 delta_rtx = scratch;
10542 /* THIS_RTX += DELTA. */
10543 emit_insn (gen_add2_insn (this_rtx, delta_rtx));
10546 /* Add the word at address (*THIS_RTX + VCALL_OFFSET). */
10549 rtx vcall_offset_rtx = GEN_INT (vcall_offset);
10550 rtx scratch = gen_rtx_REG (Pmode, 1);
10552 gcc_assert (vcall_offset < 0);
10554 /* SCRATCH = *THIS_RTX. */
10555 emit_move_insn (scratch, gen_rtx_MEM (Pmode, this_rtx));
10557 /* Prepare for adding VCALL_OFFSET. The difficulty is that we
10558 may not have any available scratch register at this point. */
10559 if (SPARC_SIMM13_P (vcall_offset))
10561 /* This is the case if ARCH64 (unless -ffixed-g5 is passed). */
10562 else if (! fixed_regs[5]
10563 /* The below sequence is made up of at least 2 insns,
10564 while the default method may need only one. */
10565 && vcall_offset < -8192)
10567 rtx scratch2 = gen_rtx_REG (Pmode, 5);
10568 emit_move_insn (scratch2, vcall_offset_rtx);
10569 vcall_offset_rtx = scratch2;
10573 rtx increment = GEN_INT (-4096);
10575 /* VCALL_OFFSET is a negative number whose typical range can be
10576 estimated as -32768..0 in 32-bit mode. In almost all cases
10577 it is therefore cheaper to emit multiple add insns than
10578 spilling and loading the constant into a register (at least
10580 while (! SPARC_SIMM13_P (vcall_offset))
10582 emit_insn (gen_add2_insn (scratch, increment));
10583 vcall_offset += 4096;
10585 vcall_offset_rtx = GEN_INT (vcall_offset); /* cannot be 0 */
10588 /* SCRATCH = *(*THIS_RTX + VCALL_OFFSET). */
10589 emit_move_insn (scratch, gen_rtx_MEM (Pmode,
10590 gen_rtx_PLUS (Pmode,
10592 vcall_offset_rtx)));
10594 /* THIS_RTX += *(*THIS_RTX + VCALL_OFFSET). */
10595 emit_insn (gen_add2_insn (this_rtx, scratch));
10598 /* Generate a tail call to the target function. */
10599 if (! TREE_USED (function))
10601 assemble_external (function);
10602 TREE_USED (function) = 1;
10604 funexp = XEXP (DECL_RTL (function), 0);
10606 if (flag_delayed_branch)
10608 funexp = gen_rtx_MEM (FUNCTION_MODE, funexp);
10609 insn = emit_call_insn (gen_sibcall (funexp));
10610 SIBLING_CALL_P (insn) = 1;
10614 /* The hoops we have to jump through in order to generate a sibcall
10615 without using delay slots... */
10616 rtx spill_reg, seq, scratch = gen_rtx_REG (Pmode, 1);
10620 spill_reg = gen_rtx_REG (word_mode, 15); /* %o7 */
10622 load_got_register (); /* clobbers %o7 */
10623 scratch = sparc_legitimize_pic_address (funexp, scratch);
10624 seq = get_insns ();
10626 emit_and_preserve (seq, spill_reg, pic_offset_table_rtx);
10628 else if (TARGET_ARCH32)
10630 emit_insn (gen_rtx_SET (VOIDmode,
10632 gen_rtx_HIGH (SImode, funexp)));
10633 emit_insn (gen_rtx_SET (VOIDmode,
10635 gen_rtx_LO_SUM (SImode, scratch, funexp)));
10637 else /* TARGET_ARCH64 */
10639 switch (sparc_cmodel)
10643 /* The destination can serve as a temporary. */
10644 sparc_emit_set_symbolic_const64 (scratch, funexp, scratch);
10649 /* The destination cannot serve as a temporary. */
10650 spill_reg = gen_rtx_REG (DImode, 15); /* %o7 */
10652 sparc_emit_set_symbolic_const64 (scratch, funexp, spill_reg);
10653 seq = get_insns ();
10655 emit_and_preserve (seq, spill_reg, 0);
10659 gcc_unreachable ();
10663 emit_jump_insn (gen_indirect_jump (scratch));
10668 /* Run just enough of rest_of_compilation to get the insns emitted.
10669 There's not really enough bulk here to make other passes such as
10670 instruction scheduling worth while. Note that use_thunk calls
10671 assemble_start_function and assemble_end_function. */
10672 insn = get_insns ();
10673 insn_locators_alloc ();
10674 shorten_branches (insn);
10675 final_start_function (insn, file, 1);
10676 final (insn, file, 1);
10677 final_end_function ();
10679 reload_completed = 0;
10680 epilogue_completed = 0;
10683 /* Return true if sparc_output_mi_thunk would be able to output the
10684 assembler code for the thunk function specified by the arguments
10685 it is passed, and false otherwise. */
10687 sparc_can_output_mi_thunk (const_tree thunk_fndecl ATTRIBUTE_UNUSED,
10688 HOST_WIDE_INT delta ATTRIBUTE_UNUSED,
10689 HOST_WIDE_INT vcall_offset,
10690 const_tree function ATTRIBUTE_UNUSED)
10692 /* Bound the loop used in the default method above. */
10693 return (vcall_offset >= -32768 || ! fixed_regs[5]);
10696 /* We use the machine specific reorg pass to enable workarounds for errata. */
10703 /* The only erratum we handle for now is that of the AT697F processor. */
10704 if (!sparc_fix_at697f)
10707 /* We need to have the (essentially) final form of the insn stream in order
10708 to properly detect the various hazards. Run delay slot scheduling. */
10709 if (optimize > 0 && flag_delayed_branch)
10711 cleanup_barriers ();
10712 dbr_schedule (get_insns ());
10715 /* Now look for specific patterns in the insn stream. */
10716 for (insn = get_insns (); insn; insn = next)
10718 bool insert_nop = false;
10721 /* Look for a single-word load into an odd-numbered FP register. */
10722 if (NONJUMP_INSN_P (insn)
10723 && (set = single_set (insn)) != NULL_RTX
10724 && GET_MODE_SIZE (GET_MODE (SET_SRC (set))) == 4
10725 && MEM_P (SET_SRC (set))
10726 && REG_P (SET_DEST (set))
10727 && REGNO (SET_DEST (set)) > 31
10728 && REGNO (SET_DEST (set)) % 2 != 0)
10730 /* The wrong dependency is on the enclosing double register. */
10731 unsigned int x = REGNO (SET_DEST (set)) - 1;
10732 unsigned int src1, src2, dest;
10735 /* If the insn has a delay slot, then it cannot be problematic. */
10736 next = next_active_insn (insn);
10737 if (NONJUMP_INSN_P (next) && GET_CODE (PATTERN (next)) == SEQUENCE)
10741 extract_insn (next);
10742 code = INSN_CODE (next);
10747 case CODE_FOR_adddf3:
10748 case CODE_FOR_subdf3:
10749 case CODE_FOR_muldf3:
10750 case CODE_FOR_divdf3:
10751 dest = REGNO (recog_data.operand[0]);
10752 src1 = REGNO (recog_data.operand[1]);
10753 src2 = REGNO (recog_data.operand[2]);
10757 ld [address], %fx+1
10758 FPOPd %f{x,y}, %f{y,x}, %f{x,y} */
10759 if ((src1 == x || src2 == x)
10760 && (dest == src1 || dest == src2))
10766 ld [address], %fx+1
10767 FPOPd %fx, %fx, %fx */
10770 && (code == CODE_FOR_adddf3 || code == CODE_FOR_muldf3))
10775 case CODE_FOR_sqrtdf2:
10776 dest = REGNO (recog_data.operand[0]);
10777 src1 = REGNO (recog_data.operand[1]);
10779 ld [address], %fx+1
10781 if (src1 == x && dest == src1)
10790 next = NEXT_INSN (insn);
10793 emit_insn_after (gen_nop (), insn);
10797 /* How to allocate a 'struct machine_function'. */
10799 static struct machine_function *
10800 sparc_init_machine_status (void)
10802 return ggc_alloc_cleared_machine_function ();
10805 /* Locate some local-dynamic symbol still in use by this function
10806 so that we can print its name in local-dynamic base patterns. */
10808 static const char *
10809 get_some_local_dynamic_name (void)
10813 if (cfun->machine->some_ld_name)
10814 return cfun->machine->some_ld_name;
10816 for (insn = get_insns (); insn ; insn = NEXT_INSN (insn))
10818 && for_each_rtx (&PATTERN (insn), get_some_local_dynamic_name_1, 0))
10819 return cfun->machine->some_ld_name;
10821 gcc_unreachable ();
10825 get_some_local_dynamic_name_1 (rtx *px, void *data ATTRIBUTE_UNUSED)
10830 && GET_CODE (x) == SYMBOL_REF
10831 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_DYNAMIC)
10833 cfun->machine->some_ld_name = XSTR (x, 0);
10840 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
10841 We need to emit DTP-relative relocations. */
10844 sparc_output_dwarf_dtprel (FILE *file, int size, rtx x)
10849 fputs ("\t.word\t%r_tls_dtpoff32(", file);
10852 fputs ("\t.xword\t%r_tls_dtpoff64(", file);
10855 gcc_unreachable ();
10857 output_addr_const (file, x);
10861 /* Do whatever processing is required at the end of a file. */
10864 sparc_file_end (void)
10866 /* If we need to emit the special GOT helper function, do so now. */
10867 if (got_helper_rtx)
10869 const char *name = XSTR (got_helper_rtx, 0);
10870 const char *reg_name = reg_names[GLOBAL_OFFSET_TABLE_REGNUM];
10871 #ifdef DWARF2_UNWIND_INFO
10875 if (USE_HIDDEN_LINKONCE)
10877 tree decl = build_decl (BUILTINS_LOCATION, FUNCTION_DECL,
10878 get_identifier (name),
10879 build_function_type_list (void_type_node,
10881 DECL_RESULT (decl) = build_decl (BUILTINS_LOCATION, RESULT_DECL,
10882 NULL_TREE, void_type_node);
10883 TREE_PUBLIC (decl) = 1;
10884 TREE_STATIC (decl) = 1;
10885 make_decl_one_only (decl, DECL_ASSEMBLER_NAME (decl));
10886 DECL_VISIBILITY (decl) = VISIBILITY_HIDDEN;
10887 DECL_VISIBILITY_SPECIFIED (decl) = 1;
10888 resolve_unique_section (decl, 0, flag_function_sections);
10889 allocate_struct_function (decl, true);
10890 cfun->is_thunk = 1;
10891 current_function_decl = decl;
10892 init_varasm_status ();
10893 assemble_start_function (decl, name);
10897 const int align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
10898 switch_to_section (text_section);
10900 ASM_OUTPUT_ALIGN (asm_out_file, align);
10901 ASM_OUTPUT_LABEL (asm_out_file, name);
10904 #ifdef DWARF2_UNWIND_INFO
10905 do_cfi = dwarf2out_do_cfi_asm ();
10907 fprintf (asm_out_file, "\t.cfi_startproc\n");
10909 if (flag_delayed_branch)
10910 fprintf (asm_out_file, "\tjmp\t%%o7+8\n\t add\t%%o7, %s, %s\n",
10911 reg_name, reg_name);
10913 fprintf (asm_out_file, "\tadd\t%%o7, %s, %s\n\tjmp\t%%o7+8\n\t nop\n",
10914 reg_name, reg_name);
10915 #ifdef DWARF2_UNWIND_INFO
10917 fprintf (asm_out_file, "\t.cfi_endproc\n");
10921 if (NEED_INDICATE_EXEC_STACK)
10922 file_end_indicate_exec_stack ();
10924 #ifdef TARGET_SOLARIS
10925 solaris_file_end ();
10929 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
10930 /* Implement TARGET_MANGLE_TYPE. */
10932 static const char *
10933 sparc_mangle_type (const_tree type)
10936 && TYPE_MAIN_VARIANT (type) == long_double_type_node
10937 && TARGET_LONG_DOUBLE_128)
10940 /* For all other types, use normal C++ mangling. */
10945 /* Expand a membar instruction for various use cases. Both the LOAD_STORE
10946 and BEFORE_AFTER arguments of the form X_Y. They are two-bit masks where
10947 bit 0 indicates that X is true, and bit 1 indicates Y is true. */
10950 sparc_emit_membar_for_model (enum memmodel model,
10951 int load_store, int before_after)
10953 /* Bits for the MEMBAR mmask field. */
10954 const int LoadLoad = 1;
10955 const int StoreLoad = 2;
10956 const int LoadStore = 4;
10957 const int StoreStore = 8;
10959 int mm = 0, implied = 0;
10961 switch (sparc_memory_model)
10964 /* Sequential Consistency. All memory transactions are immediately
10965 visible in sequential execution order. No barriers needed. */
10966 implied = LoadLoad | StoreLoad | LoadStore | StoreStore;
10970 /* Total Store Ordering: all memory transactions with store semantics
10971 are followed by an implied StoreStore. */
10972 implied |= StoreStore;
10976 /* Partial Store Ordering: all memory transactions with load semantics
10977 are followed by an implied LoadLoad | LoadStore. */
10978 implied |= LoadLoad | LoadStore;
10980 /* If we're not looking for a raw barrer (before+after), then atomic
10981 operations get the benefit of being both load and store. */
10982 if (load_store == 3 && before_after == 2)
10983 implied |= StoreLoad | StoreStore;
10987 /* Relaxed Memory Ordering: no implicit bits. */
10991 gcc_unreachable ();
10994 if (before_after & 1)
10996 if (model == MEMMODEL_ACQUIRE
10997 || model == MEMMODEL_ACQ_REL
10998 || model == MEMMODEL_SEQ_CST)
11000 if (load_store & 1)
11001 mm |= LoadLoad | LoadStore;
11002 if (load_store & 2)
11003 mm |= StoreLoad | StoreStore;
11006 if (before_after & 2)
11008 if (model == MEMMODEL_RELEASE
11009 || model == MEMMODEL_ACQ_REL
11010 || model == MEMMODEL_SEQ_CST)
11012 if (load_store & 1)
11013 mm |= LoadLoad | StoreLoad;
11014 if (load_store & 2)
11015 mm |= LoadStore | StoreStore;
11019 /* Remove the bits implied by the system memory model. */
11022 /* For raw barriers (before+after), always emit a barrier.
11023 This will become a compile-time barrier if needed. */
11024 if (mm || before_after == 3)
11025 emit_insn (gen_membar (GEN_INT (mm)));
11028 /* Expand code to perform a 8 or 16-bit compare and swap by doing 32-bit
11029 compare and swap on the word containing the byte or half-word. */
11032 sparc_expand_compare_and_swap_12 (rtx bool_result, rtx result, rtx mem,
11033 rtx oldval, rtx newval)
11035 rtx addr1 = force_reg (Pmode, XEXP (mem, 0));
11036 rtx addr = gen_reg_rtx (Pmode);
11037 rtx off = gen_reg_rtx (SImode);
11038 rtx oldv = gen_reg_rtx (SImode);
11039 rtx newv = gen_reg_rtx (SImode);
11040 rtx oldvalue = gen_reg_rtx (SImode);
11041 rtx newvalue = gen_reg_rtx (SImode);
11042 rtx res = gen_reg_rtx (SImode);
11043 rtx resv = gen_reg_rtx (SImode);
11044 rtx memsi, val, mask, end_label, loop_label, cc;
11046 emit_insn (gen_rtx_SET (VOIDmode, addr,
11047 gen_rtx_AND (Pmode, addr1, GEN_INT (-4))));
11049 if (Pmode != SImode)
11050 addr1 = gen_lowpart (SImode, addr1);
11051 emit_insn (gen_rtx_SET (VOIDmode, off,
11052 gen_rtx_AND (SImode, addr1, GEN_INT (3))));
11054 memsi = gen_rtx_MEM (SImode, addr);
11055 set_mem_alias_set (memsi, ALIAS_SET_MEMORY_BARRIER);
11056 MEM_VOLATILE_P (memsi) = MEM_VOLATILE_P (mem);
11058 val = copy_to_reg (memsi);
11060 emit_insn (gen_rtx_SET (VOIDmode, off,
11061 gen_rtx_XOR (SImode, off,
11062 GEN_INT (GET_MODE (mem) == QImode
11065 emit_insn (gen_rtx_SET (VOIDmode, off,
11066 gen_rtx_ASHIFT (SImode, off, GEN_INT (3))));
11068 if (GET_MODE (mem) == QImode)
11069 mask = force_reg (SImode, GEN_INT (0xff));
11071 mask = force_reg (SImode, GEN_INT (0xffff));
11073 emit_insn (gen_rtx_SET (VOIDmode, mask,
11074 gen_rtx_ASHIFT (SImode, mask, off)));
11076 emit_insn (gen_rtx_SET (VOIDmode, val,
11077 gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask),
11080 oldval = gen_lowpart (SImode, oldval);
11081 emit_insn (gen_rtx_SET (VOIDmode, oldv,
11082 gen_rtx_ASHIFT (SImode, oldval, off)));
11084 newval = gen_lowpart_common (SImode, newval);
11085 emit_insn (gen_rtx_SET (VOIDmode, newv,
11086 gen_rtx_ASHIFT (SImode, newval, off)));
11088 emit_insn (gen_rtx_SET (VOIDmode, oldv,
11089 gen_rtx_AND (SImode, oldv, mask)));
11091 emit_insn (gen_rtx_SET (VOIDmode, newv,
11092 gen_rtx_AND (SImode, newv, mask)));
11094 end_label = gen_label_rtx ();
11095 loop_label = gen_label_rtx ();
11096 emit_label (loop_label);
11098 emit_insn (gen_rtx_SET (VOIDmode, oldvalue,
11099 gen_rtx_IOR (SImode, oldv, val)));
11101 emit_insn (gen_rtx_SET (VOIDmode, newvalue,
11102 gen_rtx_IOR (SImode, newv, val)));
11104 emit_move_insn (bool_result, const1_rtx);
11106 emit_insn (gen_atomic_compare_and_swapsi_1 (res, memsi, oldvalue, newvalue));
11108 emit_cmp_and_jump_insns (res, oldvalue, EQ, NULL, SImode, 0, end_label);
11110 emit_insn (gen_rtx_SET (VOIDmode, resv,
11111 gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask),
11114 emit_move_insn (bool_result, const0_rtx);
11116 cc = gen_compare_reg_1 (NE, resv, val);
11117 emit_insn (gen_rtx_SET (VOIDmode, val, resv));
11119 /* Use cbranchcc4 to separate the compare and branch! */
11120 emit_jump_insn (gen_cbranchcc4 (gen_rtx_NE (VOIDmode, cc, const0_rtx),
11121 cc, const0_rtx, loop_label));
11123 emit_label (end_label);
11125 emit_insn (gen_rtx_SET (VOIDmode, res,
11126 gen_rtx_AND (SImode, res, mask)));
11128 emit_insn (gen_rtx_SET (VOIDmode, res,
11129 gen_rtx_LSHIFTRT (SImode, res, off)));
11131 emit_move_insn (result, gen_lowpart (GET_MODE (result), res));
11134 /* Expand code to perform a compare-and-swap. */
11137 sparc_expand_compare_and_swap (rtx operands[])
11139 rtx bval, retval, mem, oldval, newval;
11140 enum machine_mode mode;
11141 enum memmodel model;
11143 bval = operands[0];
11144 retval = operands[1];
11146 oldval = operands[3];
11147 newval = operands[4];
11148 model = (enum memmodel) INTVAL (operands[6]);
11149 mode = GET_MODE (mem);
11151 sparc_emit_membar_for_model (model, 3, 1);
11153 if (reg_overlap_mentioned_p (retval, oldval))
11154 oldval = copy_to_reg (oldval);
11156 if (mode == QImode || mode == HImode)
11157 sparc_expand_compare_and_swap_12 (bval, retval, mem, oldval, newval);
11160 rtx (*gen) (rtx, rtx, rtx, rtx);
11163 if (mode == SImode)
11164 gen = gen_atomic_compare_and_swapsi_1;
11166 gen = gen_atomic_compare_and_swapdi_1;
11167 emit_insn (gen (retval, mem, oldval, newval));
11169 x = emit_store_flag (bval, EQ, retval, oldval, mode, 1, 1);
11171 convert_move (bval, x, 1);
11174 sparc_emit_membar_for_model (model, 3, 2);
11178 sparc_expand_vec_perm_bmask (enum machine_mode vmode, rtx sel)
11182 sel = gen_lowpart (DImode, sel);
11186 /* inp = xxxxxxxAxxxxxxxB */
11187 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (16),
11188 NULL_RTX, 1, OPTAB_DIRECT);
11189 /* t_1 = ....xxxxxxxAxxx. */
11190 sel = expand_simple_binop (SImode, AND, gen_lowpart (SImode, sel),
11191 GEN_INT (3), NULL_RTX, 1, OPTAB_DIRECT);
11192 t_1 = expand_simple_binop (SImode, AND, gen_lowpart (SImode, t_1),
11193 GEN_INT (0x30000), NULL_RTX, 1, OPTAB_DIRECT);
11194 /* sel = .......B */
11195 /* t_1 = ...A.... */
11196 sel = expand_simple_binop (SImode, IOR, sel, t_1, sel, 1, OPTAB_DIRECT);
11197 /* sel = ...A...B */
11198 sel = expand_mult (SImode, sel, GEN_INT (0x4444), sel, 1);
11199 /* sel = AAAABBBB * 4 */
11200 t_1 = force_reg (SImode, GEN_INT (0x01230123));
11201 /* sel = { A*4, A*4+1, A*4+2, ... } */
11205 /* inp = xxxAxxxBxxxCxxxD */
11206 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (8),
11207 NULL_RTX, 1, OPTAB_DIRECT);
11208 t_2 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (16),
11209 NULL_RTX, 1, OPTAB_DIRECT);
11210 t_3 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (24),
11211 NULL_RTX, 1, OPTAB_DIRECT);
11212 /* t_1 = ..xxxAxxxBxxxCxx */
11213 /* t_2 = ....xxxAxxxBxxxC */
11214 /* t_3 = ......xxxAxxxBxx */
11215 sel = expand_simple_binop (SImode, AND, gen_lowpart (SImode, sel),
11217 NULL_RTX, 1, OPTAB_DIRECT);
11218 t_1 = expand_simple_binop (SImode, AND, gen_lowpart (SImode, t_1),
11220 NULL_RTX, 1, OPTAB_DIRECT);
11221 t_2 = expand_simple_binop (SImode, AND, gen_lowpart (SImode, t_2),
11222 GEN_INT (0x070000),
11223 NULL_RTX, 1, OPTAB_DIRECT);
11224 t_3 = expand_simple_binop (SImode, AND, gen_lowpart (SImode, t_3),
11225 GEN_INT (0x07000000),
11226 NULL_RTX, 1, OPTAB_DIRECT);
11227 /* sel = .......D */
11228 /* t_1 = .....C.. */
11229 /* t_2 = ...B.... */
11230 /* t_3 = .A...... */
11231 sel = expand_simple_binop (SImode, IOR, sel, t_1, sel, 1, OPTAB_DIRECT);
11232 t_2 = expand_simple_binop (SImode, IOR, t_2, t_3, t_2, 1, OPTAB_DIRECT);
11233 sel = expand_simple_binop (SImode, IOR, sel, t_2, sel, 1, OPTAB_DIRECT);
11234 /* sel = .A.B.C.D */
11235 sel = expand_mult (SImode, sel, GEN_INT (0x22), sel, 1);
11236 /* sel = AABBCCDD * 2 */
11237 t_1 = force_reg (SImode, GEN_INT (0x01010101));
11238 /* sel = { A*2, A*2+1, B*2, B*2+1, ... } */
11242 /* input = xAxBxCxDxExFxGxH */
11243 sel = expand_simple_binop (DImode, AND, sel,
11244 GEN_INT ((HOST_WIDE_INT)0x0f0f0f0f << 32
11246 NULL_RTX, 1, OPTAB_DIRECT);
11247 /* sel = .A.B.C.D.E.F.G.H */
11248 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (4),
11249 NULL_RTX, 1, OPTAB_DIRECT);
11250 /* t_1 = ..A.B.C.D.E.F.G. */
11251 sel = expand_simple_binop (DImode, IOR, sel, t_1,
11252 NULL_RTX, 1, OPTAB_DIRECT);
11253 /* sel = .AABBCCDDEEFFGGH */
11254 sel = expand_simple_binop (DImode, AND, sel,
11255 GEN_INT ((HOST_WIDE_INT)0xff00ff << 32
11257 NULL_RTX, 1, OPTAB_DIRECT);
11258 /* sel = ..AB..CD..EF..GH */
11259 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (8),
11260 NULL_RTX, 1, OPTAB_DIRECT);
11261 /* t_1 = ....AB..CD..EF.. */
11262 sel = expand_simple_binop (DImode, IOR, sel, t_1,
11263 NULL_RTX, 1, OPTAB_DIRECT);
11264 /* sel = ..ABABCDCDEFEFGH */
11265 sel = expand_simple_binop (DImode, AND, sel,
11266 GEN_INT ((HOST_WIDE_INT)0xffff << 32 | 0xffff),
11267 NULL_RTX, 1, OPTAB_DIRECT);
11268 /* sel = ....ABCD....EFGH */
11269 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (16),
11270 NULL_RTX, 1, OPTAB_DIRECT);
11271 /* t_1 = ........ABCD.... */
11272 sel = gen_lowpart (SImode, sel);
11273 t_1 = gen_lowpart (SImode, t_1);
11277 gcc_unreachable ();
11280 /* Always perform the final addition/merge within the bmask insn. */
11281 emit_insn (gen_bmasksi_vis (gen_reg_rtx (SImode), sel, t_1));
11284 /* Implement TARGET_FRAME_POINTER_REQUIRED. */
11287 sparc_frame_pointer_required (void)
11289 /* If the stack pointer is dynamically modified in the function, it cannot
11290 serve as the frame pointer. */
11291 if (cfun->calls_alloca)
11294 /* If the function receives nonlocal gotos, it needs to save the frame
11295 pointer in the nonlocal_goto_save_area object. */
11296 if (cfun->has_nonlocal_label)
11299 /* In flat mode, that's it. */
11303 /* Otherwise, the frame pointer is required if the function isn't leaf. */
11304 return !(crtl->is_leaf && only_leaf_regs_used ());
11307 /* The way this is structured, we can't eliminate SFP in favor of SP
11308 if the frame pointer is required: we want to use the SFP->HFP elimination
11309 in that case. But the test in update_eliminables doesn't know we are
11310 assuming below that we only do the former elimination. */
11313 sparc_can_eliminate (const int from ATTRIBUTE_UNUSED, const int to)
11315 return to == HARD_FRAME_POINTER_REGNUM || !sparc_frame_pointer_required ();
11318 /* Return the hard frame pointer directly to bypass the stack bias. */
11321 sparc_builtin_setjmp_frame_value (void)
11323 return hard_frame_pointer_rtx;
11326 /* If !TARGET_FPU, then make the fp registers and fp cc regs fixed so that
11327 they won't be allocated. */
11330 sparc_conditional_register_usage (void)
11332 if (PIC_OFFSET_TABLE_REGNUM != INVALID_REGNUM)
11334 fixed_regs[PIC_OFFSET_TABLE_REGNUM] = 1;
11335 call_used_regs[PIC_OFFSET_TABLE_REGNUM] = 1;
11337 /* If the user has passed -f{fixed,call-{used,saved}}-g5 */
11338 /* then honor it. */
11339 if (TARGET_ARCH32 && fixed_regs[5])
11341 else if (TARGET_ARCH64 && fixed_regs[5] == 2)
11346 for (regno = SPARC_FIRST_V9_FP_REG;
11347 regno <= SPARC_LAST_V9_FP_REG;
11349 fixed_regs[regno] = 1;
11350 /* %fcc0 is used by v8 and v9. */
11351 for (regno = SPARC_FIRST_V9_FCC_REG + 1;
11352 regno <= SPARC_LAST_V9_FCC_REG;
11354 fixed_regs[regno] = 1;
11359 for (regno = 32; regno < SPARC_LAST_V9_FCC_REG; regno++)
11360 fixed_regs[regno] = 1;
11362 /* If the user has passed -f{fixed,call-{used,saved}}-g2 */
11363 /* then honor it. Likewise with g3 and g4. */
11364 if (fixed_regs[2] == 2)
11365 fixed_regs[2] = ! TARGET_APP_REGS;
11366 if (fixed_regs[3] == 2)
11367 fixed_regs[3] = ! TARGET_APP_REGS;
11368 if (TARGET_ARCH32 && fixed_regs[4] == 2)
11369 fixed_regs[4] = ! TARGET_APP_REGS;
11370 else if (TARGET_CM_EMBMEDANY)
11372 else if (fixed_regs[4] == 2)
11377 /* Disable leaf functions. */
11378 memset (sparc_leaf_regs, 0, FIRST_PSEUDO_REGISTER);
11379 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
11380 leaf_reg_remap [regno] = regno;
11383 global_regs[SPARC_GSR_REG] = 1;
11386 /* Implement TARGET_PREFERRED_RELOAD_CLASS:
11388 - We can't load constants into FP registers.
11389 - We can't load FP constants into integer registers when soft-float,
11390 because there is no soft-float pattern with a r/F constraint.
11391 - We can't load FP constants into integer registers for TFmode unless
11392 it is 0.0L, because there is no movtf pattern with a r/F constraint.
11393 - Try and reload integer constants (symbolic or otherwise) back into
11394 registers directly, rather than having them dumped to memory. */
11397 sparc_preferred_reload_class (rtx x, reg_class_t rclass)
11399 enum machine_mode mode = GET_MODE (x);
11400 if (CONSTANT_P (x))
11402 if (FP_REG_CLASS_P (rclass)
11403 || rclass == GENERAL_OR_FP_REGS
11404 || rclass == GENERAL_OR_EXTRA_FP_REGS
11405 || (GET_MODE_CLASS (mode) == MODE_FLOAT && ! TARGET_FPU)
11406 || (mode == TFmode && ! const_zero_operand (x, mode)))
11409 if (GET_MODE_CLASS (mode) == MODE_INT)
11410 return GENERAL_REGS;
11412 if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT)
11414 if (! FP_REG_CLASS_P (rclass)
11415 || !(const_zero_operand (x, mode)
11416 || const_all_ones_operand (x, mode)))
11423 && (rclass == EXTRA_FP_REGS
11424 || rclass == GENERAL_OR_EXTRA_FP_REGS))
11426 int regno = true_regnum (x);
11428 if (SPARC_INT_REG_P (regno))
11429 return (rclass == EXTRA_FP_REGS
11430 ? FP_REGS : GENERAL_OR_FP_REGS);
11436 /* Output a wide multiply instruction in V8+ mode. INSN is the instruction,
11437 OPERANDS are its operands and OPCODE is the mnemonic to be used. */
11440 output_v8plus_mult (rtx insn, rtx *operands, const char *opcode)
11444 gcc_assert (! TARGET_ARCH64);
11446 if (sparc_check_64 (operands[1], insn) <= 0)
11447 output_asm_insn ("srl\t%L1, 0, %L1", operands);
11448 if (which_alternative == 1)
11449 output_asm_insn ("sllx\t%H1, 32, %H1", operands);
11450 if (GET_CODE (operands[2]) == CONST_INT)
11452 if (which_alternative == 1)
11454 output_asm_insn ("or\t%L1, %H1, %H1", operands);
11455 sprintf (mulstr, "%s\t%%H1, %%2, %%L0", opcode);
11456 output_asm_insn (mulstr, operands);
11457 return "srlx\t%L0, 32, %H0";
11461 output_asm_insn ("sllx\t%H1, 32, %3", operands);
11462 output_asm_insn ("or\t%L1, %3, %3", operands);
11463 sprintf (mulstr, "%s\t%%3, %%2, %%3", opcode);
11464 output_asm_insn (mulstr, operands);
11465 output_asm_insn ("srlx\t%3, 32, %H0", operands);
11466 return "mov\t%3, %L0";
11469 else if (rtx_equal_p (operands[1], operands[2]))
11471 if (which_alternative == 1)
11473 output_asm_insn ("or\t%L1, %H1, %H1", operands);
11474 sprintf (mulstr, "%s\t%%H1, %%H1, %%L0", opcode);
11475 output_asm_insn (mulstr, operands);
11476 return "srlx\t%L0, 32, %H0";
11480 output_asm_insn ("sllx\t%H1, 32, %3", operands);
11481 output_asm_insn ("or\t%L1, %3, %3", operands);
11482 sprintf (mulstr, "%s\t%%3, %%3, %%3", opcode);
11483 output_asm_insn (mulstr, operands);
11484 output_asm_insn ("srlx\t%3, 32, %H0", operands);
11485 return "mov\t%3, %L0";
11488 if (sparc_check_64 (operands[2], insn) <= 0)
11489 output_asm_insn ("srl\t%L2, 0, %L2", operands);
11490 if (which_alternative == 1)
11492 output_asm_insn ("or\t%L1, %H1, %H1", operands);
11493 output_asm_insn ("sllx\t%H2, 32, %L1", operands);
11494 output_asm_insn ("or\t%L2, %L1, %L1", operands);
11495 sprintf (mulstr, "%s\t%%H1, %%L1, %%L0", opcode);
11496 output_asm_insn (mulstr, operands);
11497 return "srlx\t%L0, 32, %H0";
11501 output_asm_insn ("sllx\t%H1, 32, %3", operands);
11502 output_asm_insn ("sllx\t%H2, 32, %4", operands);
11503 output_asm_insn ("or\t%L1, %3, %3", operands);
11504 output_asm_insn ("or\t%L2, %4, %4", operands);
11505 sprintf (mulstr, "%s\t%%3, %%4, %%3", opcode);
11506 output_asm_insn (mulstr, operands);
11507 output_asm_insn ("srlx\t%3, 32, %H0", operands);
11508 return "mov\t%3, %L0";
11512 /* Subroutine of sparc_expand_vector_init. Emit code to initialize
11513 all fields of TARGET to ELT by means of VIS2 BSHUFFLE insn. MODE
11514 and INNER_MODE are the modes describing TARGET. */
11517 vector_init_bshuffle (rtx target, rtx elt, enum machine_mode mode,
11518 enum machine_mode inner_mode)
11520 rtx t1, final_insn;
11523 t1 = gen_reg_rtx (mode);
11525 elt = convert_modes (SImode, inner_mode, elt, true);
11526 emit_move_insn (gen_lowpart(SImode, t1), elt);
11531 final_insn = gen_bshufflev2si_vis (target, t1, t1);
11532 bmask = 0x45674567;
11535 final_insn = gen_bshufflev4hi_vis (target, t1, t1);
11536 bmask = 0x67676767;
11539 final_insn = gen_bshufflev8qi_vis (target, t1, t1);
11540 bmask = 0x77777777;
11543 gcc_unreachable ();
11546 emit_insn (gen_bmasksi_vis (gen_reg_rtx (SImode), CONST0_RTX (SImode),
11547 force_reg (SImode, GEN_INT (bmask))));
11548 emit_insn (final_insn);
11551 /* Subroutine of sparc_expand_vector_init. Emit code to initialize
11552 all fields of TARGET to ELT in V8QI by means of VIS FPMERGE insn. */
11555 vector_init_fpmerge (rtx target, rtx elt)
11557 rtx t1, t2, t2_low, t3, t3_low;
11559 t1 = gen_reg_rtx (V4QImode);
11560 elt = convert_modes (SImode, QImode, elt, true);
11561 emit_move_insn (gen_lowpart (SImode, t1), elt);
11563 t2 = gen_reg_rtx (V8QImode);
11564 t2_low = gen_lowpart (V4QImode, t2);
11565 emit_insn (gen_fpmerge_vis (t2, t1, t1));
11567 t3 = gen_reg_rtx (V8QImode);
11568 t3_low = gen_lowpart (V4QImode, t3);
11569 emit_insn (gen_fpmerge_vis (t3, t2_low, t2_low));
11571 emit_insn (gen_fpmerge_vis (target, t3_low, t3_low));
11574 /* Subroutine of sparc_expand_vector_init. Emit code to initialize
11575 all fields of TARGET to ELT in V4HI by means of VIS FALIGNDATA insn. */
11578 vector_init_faligndata (rtx target, rtx elt)
11580 rtx t1 = gen_reg_rtx (V4HImode);
11583 elt = convert_modes (SImode, HImode, elt, true);
11584 emit_move_insn (gen_lowpart (SImode, t1), elt);
11586 emit_insn (gen_alignaddrsi_vis (gen_reg_rtx (SImode),
11587 force_reg (SImode, GEN_INT (6)),
11590 for (i = 0; i < 4; i++)
11591 emit_insn (gen_faligndatav4hi_vis (target, t1, target));
11594 /* Emit code to initialize TARGET to values for individual fields VALS. */
11597 sparc_expand_vector_init (rtx target, rtx vals)
11599 const enum machine_mode mode = GET_MODE (target);
11600 const enum machine_mode inner_mode = GET_MODE_INNER (mode);
11601 const int n_elts = GET_MODE_NUNITS (mode);
11607 for (i = 0; i < n_elts; i++)
11609 rtx x = XVECEXP (vals, 0, i);
11610 if (!CONSTANT_P (x))
11613 if (i > 0 && !rtx_equal_p (x, XVECEXP (vals, 0, 0)))
11619 emit_move_insn (target, gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0)));
11623 if (GET_MODE_SIZE (inner_mode) == GET_MODE_SIZE (mode))
11625 if (GET_MODE_SIZE (inner_mode) == 4)
11627 emit_move_insn (gen_lowpart (SImode, target),
11628 gen_lowpart (SImode, XVECEXP (vals, 0, 0)));
11631 else if (GET_MODE_SIZE (inner_mode) == 8)
11633 emit_move_insn (gen_lowpart (DImode, target),
11634 gen_lowpart (DImode, XVECEXP (vals, 0, 0)));
11638 else if (GET_MODE_SIZE (inner_mode) == GET_MODE_SIZE (word_mode)
11639 && GET_MODE_SIZE (mode) == 2 * GET_MODE_SIZE (word_mode))
11641 emit_move_insn (gen_highpart (word_mode, target),
11642 gen_lowpart (word_mode, XVECEXP (vals, 0, 0)));
11643 emit_move_insn (gen_lowpart (word_mode, target),
11644 gen_lowpart (word_mode, XVECEXP (vals, 0, 1)));
11648 if (all_same && GET_MODE_SIZE (mode) == 8)
11652 vector_init_bshuffle (target, XVECEXP (vals, 0, 0), mode, inner_mode);
11655 if (mode == V8QImode)
11657 vector_init_fpmerge (target, XVECEXP (vals, 0, 0));
11660 if (mode == V4HImode)
11662 vector_init_faligndata (target, XVECEXP (vals, 0, 0));
11667 mem = assign_stack_temp (mode, GET_MODE_SIZE (mode));
11668 for (i = 0; i < n_elts; i++)
11669 emit_move_insn (adjust_address_nv (mem, inner_mode,
11670 i * GET_MODE_SIZE (inner_mode)),
11671 XVECEXP (vals, 0, i));
11672 emit_move_insn (target, mem);
11675 /* Implement TARGET_SECONDARY_RELOAD. */
11678 sparc_secondary_reload (bool in_p, rtx x, reg_class_t rclass_i,
11679 enum machine_mode mode, secondary_reload_info *sri)
11681 enum reg_class rclass = (enum reg_class) rclass_i;
11683 sri->icode = CODE_FOR_nothing;
11684 sri->extra_cost = 0;
11686 /* We need a temporary when loading/storing a HImode/QImode value
11687 between memory and the FPU registers. This can happen when combine puts
11688 a paradoxical subreg in a float/fix conversion insn. */
11689 if (FP_REG_CLASS_P (rclass)
11690 && (mode == HImode || mode == QImode)
11691 && (GET_CODE (x) == MEM
11692 || ((GET_CODE (x) == REG || GET_CODE (x) == SUBREG)
11693 && true_regnum (x) == -1)))
11694 return GENERAL_REGS;
11696 /* On 32-bit we need a temporary when loading/storing a DFmode value
11697 between unaligned memory and the upper FPU registers. */
11699 && rclass == EXTRA_FP_REGS
11701 && GET_CODE (x) == MEM
11702 && ! mem_min_alignment (x, 8))
11705 if (((TARGET_CM_MEDANY
11706 && symbolic_operand (x, mode))
11707 || (TARGET_CM_EMBMEDANY
11708 && text_segment_operand (x, mode)))
11712 sri->icode = direct_optab_handler (reload_in_optab, mode);
11714 sri->icode = direct_optab_handler (reload_out_optab, mode);
11718 if (TARGET_VIS3 && TARGET_ARCH32)
11720 int regno = true_regnum (x);
11722 /* When using VIS3 fp<-->int register moves, on 32-bit we have
11723 to move 8-byte values in 4-byte pieces. This only works via
11724 FP_REGS, and not via EXTRA_FP_REGS. Therefore if we try to
11725 move between EXTRA_FP_REGS and GENERAL_REGS, we will need
11726 an FP_REGS intermediate move. */
11727 if ((rclass == EXTRA_FP_REGS && SPARC_INT_REG_P (regno))
11728 || ((general_or_i64_p (rclass)
11729 || rclass == GENERAL_OR_FP_REGS)
11730 && SPARC_FP_REG_P (regno)))
11732 sri->extra_cost = 2;
11740 /* Emit code to conditionally move either OPERANDS[2] or OPERANDS[3] into
11741 OPERANDS[0] in MODE. OPERANDS[1] is the operator of the condition. */
11744 sparc_expand_conditional_move (enum machine_mode mode, rtx *operands)
11746 enum rtx_code rc = GET_CODE (operands[1]);
11747 enum machine_mode cmp_mode;
11748 rtx cc_reg, dst, cmp;
11751 if (GET_MODE (XEXP (cmp, 0)) == DImode && !TARGET_ARCH64)
11754 if (GET_MODE (XEXP (cmp, 0)) == TFmode && !TARGET_HARD_QUAD)
11755 cmp = sparc_emit_float_lib_cmp (XEXP (cmp, 0), XEXP (cmp, 1), rc);
11757 cmp_mode = GET_MODE (XEXP (cmp, 0));
11758 rc = GET_CODE (cmp);
11761 if (! rtx_equal_p (operands[2], dst)
11762 && ! rtx_equal_p (operands[3], dst))
11764 if (reg_overlap_mentioned_p (dst, cmp))
11765 dst = gen_reg_rtx (mode);
11767 emit_move_insn (dst, operands[3]);
11769 else if (operands[2] == dst)
11771 operands[2] = operands[3];
11773 if (GET_MODE_CLASS (cmp_mode) == MODE_FLOAT)
11774 rc = reverse_condition_maybe_unordered (rc);
11776 rc = reverse_condition (rc);
11779 if (XEXP (cmp, 1) == const0_rtx
11780 && GET_CODE (XEXP (cmp, 0)) == REG
11781 && cmp_mode == DImode
11782 && v9_regcmp_p (rc))
11783 cc_reg = XEXP (cmp, 0);
11785 cc_reg = gen_compare_reg_1 (rc, XEXP (cmp, 0), XEXP (cmp, 1));
11787 cmp = gen_rtx_fmt_ee (rc, GET_MODE (cc_reg), cc_reg, const0_rtx);
11789 emit_insn (gen_rtx_SET (VOIDmode, dst,
11790 gen_rtx_IF_THEN_ELSE (mode, cmp, operands[2], dst)));
11792 if (dst != operands[0])
11793 emit_move_insn (operands[0], dst);
11798 /* Emit code to conditionally move a combination of OPERANDS[1] and OPERANDS[2]
11799 into OPERANDS[0] in MODE, depending on the outcome of the comparison of
11800 OPERANDS[4] and OPERANDS[5]. OPERANDS[3] is the operator of the condition.
11801 FCODE is the machine code to be used for OPERANDS[3] and CCODE the machine
11802 code to be used for the condition mask. */
11805 sparc_expand_vcond (enum machine_mode mode, rtx *operands, int ccode, int fcode)
11807 rtx mask, cop0, cop1, fcmp, cmask, bshuf, gsr;
11808 enum rtx_code code = GET_CODE (operands[3]);
11810 mask = gen_reg_rtx (Pmode);
11811 cop0 = operands[4];
11812 cop1 = operands[5];
11813 if (code == LT || code == GE)
11817 code = swap_condition (code);
11818 t = cop0; cop0 = cop1; cop1 = t;
11821 gsr = gen_rtx_REG (DImode, SPARC_GSR_REG);
11823 fcmp = gen_rtx_UNSPEC (Pmode,
11824 gen_rtvec (1, gen_rtx_fmt_ee (code, mode, cop0, cop1)),
11827 cmask = gen_rtx_UNSPEC (DImode,
11828 gen_rtvec (2, mask, gsr),
11831 bshuf = gen_rtx_UNSPEC (mode,
11832 gen_rtvec (3, operands[1], operands[2], gsr),
11835 emit_insn (gen_rtx_SET (VOIDmode, mask, fcmp));
11836 emit_insn (gen_rtx_SET (VOIDmode, gsr, cmask));
11838 emit_insn (gen_rtx_SET (VOIDmode, operands[0], bshuf));
11841 /* On sparc, any mode which naturally allocates into the float
11842 registers should return 4 here. */
11845 sparc_regmode_natural_size (enum machine_mode mode)
11847 int size = UNITS_PER_WORD;
11851 enum mode_class mclass = GET_MODE_CLASS (mode);
11853 if (mclass == MODE_FLOAT || mclass == MODE_VECTOR_INT)
11860 /* Return TRUE if it is a good idea to tie two pseudo registers
11861 when one has mode MODE1 and one has mode MODE2.
11862 If HARD_REGNO_MODE_OK could produce different values for MODE1 and MODE2,
11863 for any hard reg, then this must be FALSE for correct output.
11865 For V9 we have to deal with the fact that only the lower 32 floating
11866 point registers are 32-bit addressable. */
11869 sparc_modes_tieable_p (enum machine_mode mode1, enum machine_mode mode2)
11871 enum mode_class mclass1, mclass2;
11872 unsigned short size1, size2;
11874 if (mode1 == mode2)
11877 mclass1 = GET_MODE_CLASS (mode1);
11878 mclass2 = GET_MODE_CLASS (mode2);
11879 if (mclass1 != mclass2)
11885 /* Classes are the same and we are V9 so we have to deal with upper
11886 vs. lower floating point registers. If one of the modes is a
11887 4-byte mode, and the other is not, we have to mark them as not
11888 tieable because only the lower 32 floating point register are
11889 addressable 32-bits at a time.
11891 We can't just test explicitly for SFmode, otherwise we won't
11892 cover the vector mode cases properly. */
11894 if (mclass1 != MODE_FLOAT && mclass1 != MODE_VECTOR_INT)
11897 size1 = GET_MODE_SIZE (mode1);
11898 size2 = GET_MODE_SIZE (mode2);
11899 if ((size1 > 4 && size2 == 4)
11900 || (size2 > 4 && size1 == 4))
11906 #include "gt-sparc.h"