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"
52 #include "cfglayout.h"
54 #include "langhooks.h"
58 #include "dwarf2out.h"
63 struct processor_costs {
67 /* Integer signed load */
70 /* Integer zeroed load */
76 /* fmov, fneg, fabs */
80 const int float_plusminus;
86 const int float_cmove;
92 const int float_div_sf;
95 const int float_div_df;
98 const int float_sqrt_sf;
101 const int float_sqrt_df;
109 /* integer multiply cost for each bit set past the most
110 significant 3, so the formula for multiply cost becomes:
113 highest_bit = highest_clear_bit(rs1);
115 highest_bit = highest_set_bit(rs1);
118 cost = int_mul{,X} + ((highest_bit - 3) / int_mul_bit_factor);
120 A value of zero indicates that the multiply costs is fixed,
122 const int int_mul_bit_factor;
133 /* penalty for shifts, due to scheduling rules etc. */
134 const int shift_penalty;
138 struct processor_costs cypress_costs = {
139 COSTS_N_INSNS (2), /* int load */
140 COSTS_N_INSNS (2), /* int signed load */
141 COSTS_N_INSNS (2), /* int zeroed load */
142 COSTS_N_INSNS (2), /* float load */
143 COSTS_N_INSNS (5), /* fmov, fneg, fabs */
144 COSTS_N_INSNS (5), /* fadd, fsub */
145 COSTS_N_INSNS (1), /* fcmp */
146 COSTS_N_INSNS (1), /* fmov, fmovr */
147 COSTS_N_INSNS (7), /* fmul */
148 COSTS_N_INSNS (37), /* fdivs */
149 COSTS_N_INSNS (37), /* fdivd */
150 COSTS_N_INSNS (63), /* fsqrts */
151 COSTS_N_INSNS (63), /* fsqrtd */
152 COSTS_N_INSNS (1), /* imul */
153 COSTS_N_INSNS (1), /* imulX */
154 0, /* imul bit factor */
155 COSTS_N_INSNS (1), /* idiv */
156 COSTS_N_INSNS (1), /* idivX */
157 COSTS_N_INSNS (1), /* movcc/movr */
158 0, /* shift penalty */
162 struct processor_costs supersparc_costs = {
163 COSTS_N_INSNS (1), /* int load */
164 COSTS_N_INSNS (1), /* int signed load */
165 COSTS_N_INSNS (1), /* int zeroed load */
166 COSTS_N_INSNS (0), /* float load */
167 COSTS_N_INSNS (3), /* fmov, fneg, fabs */
168 COSTS_N_INSNS (3), /* fadd, fsub */
169 COSTS_N_INSNS (3), /* fcmp */
170 COSTS_N_INSNS (1), /* fmov, fmovr */
171 COSTS_N_INSNS (3), /* fmul */
172 COSTS_N_INSNS (6), /* fdivs */
173 COSTS_N_INSNS (9), /* fdivd */
174 COSTS_N_INSNS (12), /* fsqrts */
175 COSTS_N_INSNS (12), /* fsqrtd */
176 COSTS_N_INSNS (4), /* imul */
177 COSTS_N_INSNS (4), /* imulX */
178 0, /* imul bit factor */
179 COSTS_N_INSNS (4), /* idiv */
180 COSTS_N_INSNS (4), /* idivX */
181 COSTS_N_INSNS (1), /* movcc/movr */
182 1, /* shift penalty */
186 struct processor_costs hypersparc_costs = {
187 COSTS_N_INSNS (1), /* int load */
188 COSTS_N_INSNS (1), /* int signed load */
189 COSTS_N_INSNS (1), /* int zeroed load */
190 COSTS_N_INSNS (1), /* float load */
191 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
192 COSTS_N_INSNS (1), /* fadd, fsub */
193 COSTS_N_INSNS (1), /* fcmp */
194 COSTS_N_INSNS (1), /* fmov, fmovr */
195 COSTS_N_INSNS (1), /* fmul */
196 COSTS_N_INSNS (8), /* fdivs */
197 COSTS_N_INSNS (12), /* fdivd */
198 COSTS_N_INSNS (17), /* fsqrts */
199 COSTS_N_INSNS (17), /* fsqrtd */
200 COSTS_N_INSNS (17), /* imul */
201 COSTS_N_INSNS (17), /* imulX */
202 0, /* imul bit factor */
203 COSTS_N_INSNS (17), /* idiv */
204 COSTS_N_INSNS (17), /* idivX */
205 COSTS_N_INSNS (1), /* movcc/movr */
206 0, /* shift penalty */
210 struct processor_costs leon_costs = {
211 COSTS_N_INSNS (1), /* int load */
212 COSTS_N_INSNS (1), /* int signed load */
213 COSTS_N_INSNS (1), /* int zeroed load */
214 COSTS_N_INSNS (1), /* float load */
215 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
216 COSTS_N_INSNS (1), /* fadd, fsub */
217 COSTS_N_INSNS (1), /* fcmp */
218 COSTS_N_INSNS (1), /* fmov, fmovr */
219 COSTS_N_INSNS (1), /* fmul */
220 COSTS_N_INSNS (15), /* fdivs */
221 COSTS_N_INSNS (15), /* fdivd */
222 COSTS_N_INSNS (23), /* fsqrts */
223 COSTS_N_INSNS (23), /* fsqrtd */
224 COSTS_N_INSNS (5), /* imul */
225 COSTS_N_INSNS (5), /* imulX */
226 0, /* imul bit factor */
227 COSTS_N_INSNS (5), /* idiv */
228 COSTS_N_INSNS (5), /* idivX */
229 COSTS_N_INSNS (1), /* movcc/movr */
230 0, /* shift penalty */
234 struct processor_costs sparclet_costs = {
235 COSTS_N_INSNS (3), /* int load */
236 COSTS_N_INSNS (3), /* int signed load */
237 COSTS_N_INSNS (1), /* int zeroed load */
238 COSTS_N_INSNS (1), /* float load */
239 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
240 COSTS_N_INSNS (1), /* fadd, fsub */
241 COSTS_N_INSNS (1), /* fcmp */
242 COSTS_N_INSNS (1), /* fmov, fmovr */
243 COSTS_N_INSNS (1), /* fmul */
244 COSTS_N_INSNS (1), /* fdivs */
245 COSTS_N_INSNS (1), /* fdivd */
246 COSTS_N_INSNS (1), /* fsqrts */
247 COSTS_N_INSNS (1), /* fsqrtd */
248 COSTS_N_INSNS (5), /* imul */
249 COSTS_N_INSNS (5), /* imulX */
250 0, /* imul bit factor */
251 COSTS_N_INSNS (5), /* idiv */
252 COSTS_N_INSNS (5), /* idivX */
253 COSTS_N_INSNS (1), /* movcc/movr */
254 0, /* shift penalty */
258 struct processor_costs ultrasparc_costs = {
259 COSTS_N_INSNS (2), /* int load */
260 COSTS_N_INSNS (3), /* int signed load */
261 COSTS_N_INSNS (2), /* int zeroed load */
262 COSTS_N_INSNS (2), /* float load */
263 COSTS_N_INSNS (1), /* fmov, fneg, fabs */
264 COSTS_N_INSNS (4), /* fadd, fsub */
265 COSTS_N_INSNS (1), /* fcmp */
266 COSTS_N_INSNS (2), /* fmov, fmovr */
267 COSTS_N_INSNS (4), /* fmul */
268 COSTS_N_INSNS (13), /* fdivs */
269 COSTS_N_INSNS (23), /* fdivd */
270 COSTS_N_INSNS (13), /* fsqrts */
271 COSTS_N_INSNS (23), /* fsqrtd */
272 COSTS_N_INSNS (4), /* imul */
273 COSTS_N_INSNS (4), /* imulX */
274 2, /* imul bit factor */
275 COSTS_N_INSNS (37), /* idiv */
276 COSTS_N_INSNS (68), /* idivX */
277 COSTS_N_INSNS (2), /* movcc/movr */
278 2, /* shift penalty */
282 struct processor_costs ultrasparc3_costs = {
283 COSTS_N_INSNS (2), /* int load */
284 COSTS_N_INSNS (3), /* int signed load */
285 COSTS_N_INSNS (3), /* int zeroed load */
286 COSTS_N_INSNS (2), /* float load */
287 COSTS_N_INSNS (3), /* fmov, fneg, fabs */
288 COSTS_N_INSNS (4), /* fadd, fsub */
289 COSTS_N_INSNS (5), /* fcmp */
290 COSTS_N_INSNS (3), /* fmov, fmovr */
291 COSTS_N_INSNS (4), /* fmul */
292 COSTS_N_INSNS (17), /* fdivs */
293 COSTS_N_INSNS (20), /* fdivd */
294 COSTS_N_INSNS (20), /* fsqrts */
295 COSTS_N_INSNS (29), /* fsqrtd */
296 COSTS_N_INSNS (6), /* imul */
297 COSTS_N_INSNS (6), /* imulX */
298 0, /* imul bit factor */
299 COSTS_N_INSNS (40), /* idiv */
300 COSTS_N_INSNS (71), /* idivX */
301 COSTS_N_INSNS (2), /* movcc/movr */
302 0, /* shift penalty */
306 struct processor_costs niagara_costs = {
307 COSTS_N_INSNS (3), /* int load */
308 COSTS_N_INSNS (3), /* int signed load */
309 COSTS_N_INSNS (3), /* int zeroed load */
310 COSTS_N_INSNS (9), /* float load */
311 COSTS_N_INSNS (8), /* fmov, fneg, fabs */
312 COSTS_N_INSNS (8), /* fadd, fsub */
313 COSTS_N_INSNS (26), /* fcmp */
314 COSTS_N_INSNS (8), /* fmov, fmovr */
315 COSTS_N_INSNS (29), /* fmul */
316 COSTS_N_INSNS (54), /* fdivs */
317 COSTS_N_INSNS (83), /* fdivd */
318 COSTS_N_INSNS (100), /* fsqrts - not implemented in hardware */
319 COSTS_N_INSNS (100), /* fsqrtd - not implemented in hardware */
320 COSTS_N_INSNS (11), /* imul */
321 COSTS_N_INSNS (11), /* imulX */
322 0, /* imul bit factor */
323 COSTS_N_INSNS (72), /* idiv */
324 COSTS_N_INSNS (72), /* idivX */
325 COSTS_N_INSNS (1), /* movcc/movr */
326 0, /* shift penalty */
330 struct processor_costs niagara2_costs = {
331 COSTS_N_INSNS (3), /* int load */
332 COSTS_N_INSNS (3), /* int signed load */
333 COSTS_N_INSNS (3), /* int zeroed load */
334 COSTS_N_INSNS (3), /* float load */
335 COSTS_N_INSNS (6), /* fmov, fneg, fabs */
336 COSTS_N_INSNS (6), /* fadd, fsub */
337 COSTS_N_INSNS (6), /* fcmp */
338 COSTS_N_INSNS (6), /* fmov, fmovr */
339 COSTS_N_INSNS (6), /* fmul */
340 COSTS_N_INSNS (19), /* fdivs */
341 COSTS_N_INSNS (33), /* fdivd */
342 COSTS_N_INSNS (19), /* fsqrts */
343 COSTS_N_INSNS (33), /* fsqrtd */
344 COSTS_N_INSNS (5), /* imul */
345 COSTS_N_INSNS (5), /* imulX */
346 0, /* imul bit factor */
347 COSTS_N_INSNS (26), /* idiv, average of 12 - 41 cycle range */
348 COSTS_N_INSNS (26), /* idivX, average of 12 - 41 cycle range */
349 COSTS_N_INSNS (1), /* movcc/movr */
350 0, /* shift penalty */
354 struct processor_costs niagara3_costs = {
355 COSTS_N_INSNS (3), /* int load */
356 COSTS_N_INSNS (3), /* int signed load */
357 COSTS_N_INSNS (3), /* int zeroed load */
358 COSTS_N_INSNS (3), /* float load */
359 COSTS_N_INSNS (9), /* fmov, fneg, fabs */
360 COSTS_N_INSNS (9), /* fadd, fsub */
361 COSTS_N_INSNS (9), /* fcmp */
362 COSTS_N_INSNS (9), /* fmov, fmovr */
363 COSTS_N_INSNS (9), /* fmul */
364 COSTS_N_INSNS (23), /* fdivs */
365 COSTS_N_INSNS (37), /* fdivd */
366 COSTS_N_INSNS (23), /* fsqrts */
367 COSTS_N_INSNS (37), /* fsqrtd */
368 COSTS_N_INSNS (9), /* imul */
369 COSTS_N_INSNS (9), /* imulX */
370 0, /* imul bit factor */
371 COSTS_N_INSNS (31), /* idiv, average of 17 - 45 cycle range */
372 COSTS_N_INSNS (30), /* idivX, average of 16 - 44 cycle range */
373 COSTS_N_INSNS (1), /* movcc/movr */
374 0, /* shift penalty */
378 struct processor_costs niagara4_costs = {
379 COSTS_N_INSNS (5), /* int load */
380 COSTS_N_INSNS (5), /* int signed load */
381 COSTS_N_INSNS (5), /* int zeroed load */
382 COSTS_N_INSNS (5), /* float load */
383 COSTS_N_INSNS (11), /* fmov, fneg, fabs */
384 COSTS_N_INSNS (11), /* fadd, fsub */
385 COSTS_N_INSNS (11), /* fcmp */
386 COSTS_N_INSNS (11), /* fmov, fmovr */
387 COSTS_N_INSNS (11), /* fmul */
388 COSTS_N_INSNS (24), /* fdivs */
389 COSTS_N_INSNS (37), /* fdivd */
390 COSTS_N_INSNS (24), /* fsqrts */
391 COSTS_N_INSNS (37), /* fsqrtd */
392 COSTS_N_INSNS (12), /* imul */
393 COSTS_N_INSNS (12), /* imulX */
394 0, /* imul bit factor */
395 COSTS_N_INSNS (50), /* idiv, average of 41 - 60 cycle range */
396 COSTS_N_INSNS (35), /* idivX, average of 26 - 44 cycle range */
397 COSTS_N_INSNS (1), /* movcc/movr */
398 0, /* shift penalty */
401 static const struct processor_costs *sparc_costs = &cypress_costs;
403 #ifdef HAVE_AS_RELAX_OPTION
404 /* If 'as' and 'ld' are relaxing tail call insns into branch always, use
405 "or %o7,%g0,X; call Y; or X,%g0,%o7" always, so that it can be optimized.
406 With sethi/jmp, neither 'as' nor 'ld' has an easy way how to find out if
407 somebody does not branch between the sethi and jmp. */
408 #define LEAF_SIBCALL_SLOT_RESERVED_P 1
410 #define LEAF_SIBCALL_SLOT_RESERVED_P \
411 ((TARGET_ARCH64 && !TARGET_CM_MEDLOW) || flag_pic)
414 /* Vector to say how input registers are mapped to output registers.
415 HARD_FRAME_POINTER_REGNUM cannot be remapped by this function to
416 eliminate it. You must use -fomit-frame-pointer to get that. */
417 char leaf_reg_remap[] =
418 { 0, 1, 2, 3, 4, 5, 6, 7,
419 -1, -1, -1, -1, -1, -1, 14, -1,
420 -1, -1, -1, -1, -1, -1, -1, -1,
421 8, 9, 10, 11, 12, 13, -1, 15,
423 32, 33, 34, 35, 36, 37, 38, 39,
424 40, 41, 42, 43, 44, 45, 46, 47,
425 48, 49, 50, 51, 52, 53, 54, 55,
426 56, 57, 58, 59, 60, 61, 62, 63,
427 64, 65, 66, 67, 68, 69, 70, 71,
428 72, 73, 74, 75, 76, 77, 78, 79,
429 80, 81, 82, 83, 84, 85, 86, 87,
430 88, 89, 90, 91, 92, 93, 94, 95,
431 96, 97, 98, 99, 100, 101, 102};
433 /* Vector, indexed by hard register number, which contains 1
434 for a register that is allowable in a candidate for leaf
435 function treatment. */
436 char sparc_leaf_regs[] =
437 { 1, 1, 1, 1, 1, 1, 1, 1,
438 0, 0, 0, 0, 0, 0, 1, 0,
439 0, 0, 0, 0, 0, 0, 0, 0,
440 1, 1, 1, 1, 1, 1, 0, 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, 1,
448 1, 1, 1, 1, 1, 1, 1, 1,
449 1, 1, 1, 1, 1, 1, 1};
451 struct GTY(()) machine_function
453 /* Size of the frame of the function. */
454 HOST_WIDE_INT frame_size;
456 /* Size of the frame of the function minus the register window save area
457 and the outgoing argument area. */
458 HOST_WIDE_INT apparent_frame_size;
460 /* Register we pretend the frame pointer is allocated to. Normally, this
461 is %fp, but if we are in a leaf procedure, this is (%sp + offset). We
462 record "offset" separately as it may be too big for (reg + disp). */
464 HOST_WIDE_INT frame_base_offset;
466 /* Some local-dynamic TLS symbol name. */
467 const char *some_ld_name;
469 /* Number of global or FP registers to be saved (as 4-byte quantities). */
470 int n_global_fp_regs;
472 /* True if the current function is leaf and uses only leaf regs,
473 so that the SPARC leaf function optimization can be applied.
474 Private version of current_function_uses_only_leaf_regs, see
475 sparc_expand_prologue for the rationale. */
478 /* True if the prologue saves local or in registers. */
479 bool save_local_in_regs_p;
481 /* True if the data calculated by sparc_expand_prologue are valid. */
482 bool prologue_data_valid_p;
485 #define sparc_frame_size cfun->machine->frame_size
486 #define sparc_apparent_frame_size cfun->machine->apparent_frame_size
487 #define sparc_frame_base_reg cfun->machine->frame_base_reg
488 #define sparc_frame_base_offset cfun->machine->frame_base_offset
489 #define sparc_n_global_fp_regs cfun->machine->n_global_fp_regs
490 #define sparc_leaf_function_p cfun->machine->leaf_function_p
491 #define sparc_save_local_in_regs_p cfun->machine->save_local_in_regs_p
492 #define sparc_prologue_data_valid_p cfun->machine->prologue_data_valid_p
494 /* 1 if the next opcode is to be specially indented. */
495 int sparc_indent_opcode = 0;
497 static void sparc_option_override (void);
498 static void sparc_init_modes (void);
499 static void scan_record_type (const_tree, int *, int *, int *);
500 static int function_arg_slotno (const CUMULATIVE_ARGS *, enum machine_mode,
501 const_tree, bool, bool, int *, int *);
503 static int supersparc_adjust_cost (rtx, rtx, rtx, int);
504 static int hypersparc_adjust_cost (rtx, rtx, rtx, int);
506 static void sparc_emit_set_const32 (rtx, rtx);
507 static void sparc_emit_set_const64 (rtx, rtx);
508 static void sparc_output_addr_vec (rtx);
509 static void sparc_output_addr_diff_vec (rtx);
510 static void sparc_output_deferred_case_vectors (void);
511 static bool sparc_legitimate_address_p (enum machine_mode, rtx, bool);
512 static bool sparc_legitimate_constant_p (enum machine_mode, rtx);
513 static rtx sparc_builtin_saveregs (void);
514 static int epilogue_renumber (rtx *, int);
515 static bool sparc_assemble_integer (rtx, unsigned int, int);
516 static int set_extends (rtx);
517 static void sparc_asm_function_prologue (FILE *, HOST_WIDE_INT);
518 static void sparc_asm_function_epilogue (FILE *, HOST_WIDE_INT);
519 #ifdef TARGET_SOLARIS
520 static void sparc_solaris_elf_asm_named_section (const char *, unsigned int,
521 tree) ATTRIBUTE_UNUSED;
523 static int sparc_adjust_cost (rtx, rtx, rtx, int);
524 static int sparc_issue_rate (void);
525 static void sparc_sched_init (FILE *, int, int);
526 static int sparc_use_sched_lookahead (void);
528 static void emit_soft_tfmode_libcall (const char *, int, rtx *);
529 static void emit_soft_tfmode_binop (enum rtx_code, rtx *);
530 static void emit_soft_tfmode_unop (enum rtx_code, rtx *);
531 static void emit_soft_tfmode_cvt (enum rtx_code, rtx *);
532 static void emit_hard_tfmode_operation (enum rtx_code, rtx *);
534 static bool sparc_function_ok_for_sibcall (tree, tree);
535 static void sparc_init_libfuncs (void);
536 static void sparc_init_builtins (void);
537 static void sparc_vis_init_builtins (void);
538 static rtx sparc_expand_builtin (tree, rtx, rtx, enum machine_mode, int);
539 static tree sparc_fold_builtin (tree, int, tree *, bool);
540 static int sparc_vis_mul8x16 (int, int);
541 static void sparc_handle_vis_mul8x16 (tree *, int, tree, tree, tree);
542 static void sparc_output_mi_thunk (FILE *, tree, HOST_WIDE_INT,
543 HOST_WIDE_INT, tree);
544 static bool sparc_can_output_mi_thunk (const_tree, HOST_WIDE_INT,
545 HOST_WIDE_INT, const_tree);
546 static void sparc_reorg (void);
547 static struct machine_function * sparc_init_machine_status (void);
548 static bool sparc_cannot_force_const_mem (enum machine_mode, rtx);
549 static rtx sparc_tls_get_addr (void);
550 static rtx sparc_tls_got (void);
551 static const char *get_some_local_dynamic_name (void);
552 static int get_some_local_dynamic_name_1 (rtx *, void *);
553 static int sparc_register_move_cost (enum machine_mode,
554 reg_class_t, reg_class_t);
555 static bool sparc_rtx_costs (rtx, int, int, int, int *, bool);
556 static rtx sparc_function_value (const_tree, const_tree, bool);
557 static rtx sparc_libcall_value (enum machine_mode, const_rtx);
558 static bool sparc_function_value_regno_p (const unsigned int);
559 static rtx sparc_struct_value_rtx (tree, int);
560 static enum machine_mode sparc_promote_function_mode (const_tree, enum machine_mode,
561 int *, const_tree, int);
562 static bool sparc_return_in_memory (const_tree, const_tree);
563 static bool sparc_strict_argument_naming (cumulative_args_t);
564 static void sparc_va_start (tree, rtx);
565 static tree sparc_gimplify_va_arg (tree, tree, gimple_seq *, gimple_seq *);
566 static bool sparc_vector_mode_supported_p (enum machine_mode);
567 static bool sparc_tls_referenced_p (rtx);
568 static rtx sparc_legitimize_tls_address (rtx);
569 static rtx sparc_legitimize_pic_address (rtx, rtx);
570 static rtx sparc_legitimize_address (rtx, rtx, enum machine_mode);
571 static rtx sparc_delegitimize_address (rtx);
572 static bool sparc_mode_dependent_address_p (const_rtx);
573 static bool sparc_pass_by_reference (cumulative_args_t,
574 enum machine_mode, const_tree, bool);
575 static void sparc_function_arg_advance (cumulative_args_t,
576 enum machine_mode, const_tree, bool);
577 static rtx sparc_function_arg_1 (cumulative_args_t,
578 enum machine_mode, const_tree, bool, bool);
579 static rtx sparc_function_arg (cumulative_args_t,
580 enum machine_mode, const_tree, bool);
581 static rtx sparc_function_incoming_arg (cumulative_args_t,
582 enum machine_mode, const_tree, bool);
583 static unsigned int sparc_function_arg_boundary (enum machine_mode,
585 static int sparc_arg_partial_bytes (cumulative_args_t,
586 enum machine_mode, tree, bool);
587 static void sparc_output_dwarf_dtprel (FILE *, int, rtx) ATTRIBUTE_UNUSED;
588 static void sparc_file_end (void);
589 static bool sparc_frame_pointer_required (void);
590 static bool sparc_can_eliminate (const int, const int);
591 static rtx sparc_builtin_setjmp_frame_value (void);
592 static void sparc_conditional_register_usage (void);
593 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
594 static const char *sparc_mangle_type (const_tree);
596 static void sparc_trampoline_init (rtx, tree, rtx);
597 static enum machine_mode sparc_preferred_simd_mode (enum machine_mode);
598 static reg_class_t sparc_preferred_reload_class (rtx x, reg_class_t rclass);
599 static bool sparc_print_operand_punct_valid_p (unsigned char);
600 static void sparc_print_operand (FILE *, rtx, int);
601 static void sparc_print_operand_address (FILE *, rtx);
602 static reg_class_t sparc_secondary_reload (bool, rtx, reg_class_t,
604 secondary_reload_info *);
606 #ifdef SUBTARGET_ATTRIBUTE_TABLE
607 /* Table of valid machine attributes. */
608 static const struct attribute_spec sparc_attribute_table[] =
610 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler,
612 SUBTARGET_ATTRIBUTE_TABLE,
613 { NULL, 0, 0, false, false, false, NULL, false }
617 /* Option handling. */
620 enum cmodel sparc_cmodel;
622 char sparc_hard_reg_printed[8];
624 /* Initialize the GCC target structure. */
626 /* The default is to use .half rather than .short for aligned HI objects. */
627 #undef TARGET_ASM_ALIGNED_HI_OP
628 #define TARGET_ASM_ALIGNED_HI_OP "\t.half\t"
630 #undef TARGET_ASM_UNALIGNED_HI_OP
631 #define TARGET_ASM_UNALIGNED_HI_OP "\t.uahalf\t"
632 #undef TARGET_ASM_UNALIGNED_SI_OP
633 #define TARGET_ASM_UNALIGNED_SI_OP "\t.uaword\t"
634 #undef TARGET_ASM_UNALIGNED_DI_OP
635 #define TARGET_ASM_UNALIGNED_DI_OP "\t.uaxword\t"
637 /* The target hook has to handle DI-mode values. */
638 #undef TARGET_ASM_INTEGER
639 #define TARGET_ASM_INTEGER sparc_assemble_integer
641 #undef TARGET_ASM_FUNCTION_PROLOGUE
642 #define TARGET_ASM_FUNCTION_PROLOGUE sparc_asm_function_prologue
643 #undef TARGET_ASM_FUNCTION_EPILOGUE
644 #define TARGET_ASM_FUNCTION_EPILOGUE sparc_asm_function_epilogue
646 #undef TARGET_SCHED_ADJUST_COST
647 #define TARGET_SCHED_ADJUST_COST sparc_adjust_cost
648 #undef TARGET_SCHED_ISSUE_RATE
649 #define TARGET_SCHED_ISSUE_RATE sparc_issue_rate
650 #undef TARGET_SCHED_INIT
651 #define TARGET_SCHED_INIT sparc_sched_init
652 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
653 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD sparc_use_sched_lookahead
655 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
656 #define TARGET_FUNCTION_OK_FOR_SIBCALL sparc_function_ok_for_sibcall
658 #undef TARGET_INIT_LIBFUNCS
659 #define TARGET_INIT_LIBFUNCS sparc_init_libfuncs
660 #undef TARGET_INIT_BUILTINS
661 #define TARGET_INIT_BUILTINS sparc_init_builtins
663 #undef TARGET_LEGITIMIZE_ADDRESS
664 #define TARGET_LEGITIMIZE_ADDRESS sparc_legitimize_address
665 #undef TARGET_DELEGITIMIZE_ADDRESS
666 #define TARGET_DELEGITIMIZE_ADDRESS sparc_delegitimize_address
667 #undef TARGET_MODE_DEPENDENT_ADDRESS_P
668 #define TARGET_MODE_DEPENDENT_ADDRESS_P sparc_mode_dependent_address_p
670 #undef TARGET_EXPAND_BUILTIN
671 #define TARGET_EXPAND_BUILTIN sparc_expand_builtin
672 #undef TARGET_FOLD_BUILTIN
673 #define TARGET_FOLD_BUILTIN sparc_fold_builtin
676 #undef TARGET_HAVE_TLS
677 #define TARGET_HAVE_TLS true
680 #undef TARGET_CANNOT_FORCE_CONST_MEM
681 #define TARGET_CANNOT_FORCE_CONST_MEM sparc_cannot_force_const_mem
683 #undef TARGET_ASM_OUTPUT_MI_THUNK
684 #define TARGET_ASM_OUTPUT_MI_THUNK sparc_output_mi_thunk
685 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
686 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK sparc_can_output_mi_thunk
688 #undef TARGET_MACHINE_DEPENDENT_REORG
689 #define TARGET_MACHINE_DEPENDENT_REORG sparc_reorg
691 #undef TARGET_RTX_COSTS
692 #define TARGET_RTX_COSTS sparc_rtx_costs
693 #undef TARGET_ADDRESS_COST
694 #define TARGET_ADDRESS_COST hook_int_rtx_bool_0
695 #undef TARGET_REGISTER_MOVE_COST
696 #define TARGET_REGISTER_MOVE_COST sparc_register_move_cost
698 #undef TARGET_PROMOTE_FUNCTION_MODE
699 #define TARGET_PROMOTE_FUNCTION_MODE sparc_promote_function_mode
701 #undef TARGET_FUNCTION_VALUE
702 #define TARGET_FUNCTION_VALUE sparc_function_value
703 #undef TARGET_LIBCALL_VALUE
704 #define TARGET_LIBCALL_VALUE sparc_libcall_value
705 #undef TARGET_FUNCTION_VALUE_REGNO_P
706 #define TARGET_FUNCTION_VALUE_REGNO_P sparc_function_value_regno_p
708 #undef TARGET_STRUCT_VALUE_RTX
709 #define TARGET_STRUCT_VALUE_RTX sparc_struct_value_rtx
710 #undef TARGET_RETURN_IN_MEMORY
711 #define TARGET_RETURN_IN_MEMORY sparc_return_in_memory
712 #undef TARGET_MUST_PASS_IN_STACK
713 #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size
714 #undef TARGET_PASS_BY_REFERENCE
715 #define TARGET_PASS_BY_REFERENCE sparc_pass_by_reference
716 #undef TARGET_ARG_PARTIAL_BYTES
717 #define TARGET_ARG_PARTIAL_BYTES sparc_arg_partial_bytes
718 #undef TARGET_FUNCTION_ARG_ADVANCE
719 #define TARGET_FUNCTION_ARG_ADVANCE sparc_function_arg_advance
720 #undef TARGET_FUNCTION_ARG
721 #define TARGET_FUNCTION_ARG sparc_function_arg
722 #undef TARGET_FUNCTION_INCOMING_ARG
723 #define TARGET_FUNCTION_INCOMING_ARG sparc_function_incoming_arg
724 #undef TARGET_FUNCTION_ARG_BOUNDARY
725 #define TARGET_FUNCTION_ARG_BOUNDARY sparc_function_arg_boundary
727 #undef TARGET_EXPAND_BUILTIN_SAVEREGS
728 #define TARGET_EXPAND_BUILTIN_SAVEREGS sparc_builtin_saveregs
729 #undef TARGET_STRICT_ARGUMENT_NAMING
730 #define TARGET_STRICT_ARGUMENT_NAMING sparc_strict_argument_naming
732 #undef TARGET_EXPAND_BUILTIN_VA_START
733 #define TARGET_EXPAND_BUILTIN_VA_START sparc_va_start
734 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
735 #define TARGET_GIMPLIFY_VA_ARG_EXPR sparc_gimplify_va_arg
737 #undef TARGET_VECTOR_MODE_SUPPORTED_P
738 #define TARGET_VECTOR_MODE_SUPPORTED_P sparc_vector_mode_supported_p
740 #undef TARGET_VECTORIZE_PREFERRED_SIMD_MODE
741 #define TARGET_VECTORIZE_PREFERRED_SIMD_MODE sparc_preferred_simd_mode
743 #ifdef SUBTARGET_INSERT_ATTRIBUTES
744 #undef TARGET_INSERT_ATTRIBUTES
745 #define TARGET_INSERT_ATTRIBUTES SUBTARGET_INSERT_ATTRIBUTES
748 #ifdef SUBTARGET_ATTRIBUTE_TABLE
749 #undef TARGET_ATTRIBUTE_TABLE
750 #define TARGET_ATTRIBUTE_TABLE sparc_attribute_table
753 #undef TARGET_RELAXED_ORDERING
754 #define TARGET_RELAXED_ORDERING SPARC_RELAXED_ORDERING
756 #undef TARGET_OPTION_OVERRIDE
757 #define TARGET_OPTION_OVERRIDE sparc_option_override
759 #if TARGET_GNU_TLS && defined(HAVE_AS_SPARC_UA_PCREL)
760 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
761 #define TARGET_ASM_OUTPUT_DWARF_DTPREL sparc_output_dwarf_dtprel
764 #undef TARGET_ASM_FILE_END
765 #define TARGET_ASM_FILE_END sparc_file_end
767 #undef TARGET_FRAME_POINTER_REQUIRED
768 #define TARGET_FRAME_POINTER_REQUIRED sparc_frame_pointer_required
770 #undef TARGET_BUILTIN_SETJMP_FRAME_VALUE
771 #define TARGET_BUILTIN_SETJMP_FRAME_VALUE sparc_builtin_setjmp_frame_value
773 #undef TARGET_CAN_ELIMINATE
774 #define TARGET_CAN_ELIMINATE sparc_can_eliminate
776 #undef TARGET_PREFERRED_RELOAD_CLASS
777 #define TARGET_PREFERRED_RELOAD_CLASS sparc_preferred_reload_class
779 #undef TARGET_SECONDARY_RELOAD
780 #define TARGET_SECONDARY_RELOAD sparc_secondary_reload
782 #undef TARGET_CONDITIONAL_REGISTER_USAGE
783 #define TARGET_CONDITIONAL_REGISTER_USAGE sparc_conditional_register_usage
785 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
786 #undef TARGET_MANGLE_TYPE
787 #define TARGET_MANGLE_TYPE sparc_mangle_type
790 #undef TARGET_LEGITIMATE_ADDRESS_P
791 #define TARGET_LEGITIMATE_ADDRESS_P sparc_legitimate_address_p
793 #undef TARGET_LEGITIMATE_CONSTANT_P
794 #define TARGET_LEGITIMATE_CONSTANT_P sparc_legitimate_constant_p
796 #undef TARGET_TRAMPOLINE_INIT
797 #define TARGET_TRAMPOLINE_INIT sparc_trampoline_init
799 #undef TARGET_PRINT_OPERAND_PUNCT_VALID_P
800 #define TARGET_PRINT_OPERAND_PUNCT_VALID_P sparc_print_operand_punct_valid_p
801 #undef TARGET_PRINT_OPERAND
802 #define TARGET_PRINT_OPERAND sparc_print_operand
803 #undef TARGET_PRINT_OPERAND_ADDRESS
804 #define TARGET_PRINT_OPERAND_ADDRESS sparc_print_operand_address
806 /* The value stored by LDSTUB. */
807 #undef TARGET_ATOMIC_TEST_AND_SET_TRUEVAL
808 #define TARGET_ATOMIC_TEST_AND_SET_TRUEVAL 0xff
810 struct gcc_target targetm = TARGET_INITIALIZER;
813 dump_target_flag_bits (const int flags)
815 if (flags & MASK_64BIT)
816 fprintf (stderr, "64BIT ");
817 if (flags & MASK_APP_REGS)
818 fprintf (stderr, "APP_REGS ");
819 if (flags & MASK_FASTER_STRUCTS)
820 fprintf (stderr, "FASTER_STRUCTS ");
821 if (flags & MASK_FLAT)
822 fprintf (stderr, "FLAT ");
823 if (flags & MASK_FMAF)
824 fprintf (stderr, "FMAF ");
825 if (flags & MASK_FPU)
826 fprintf (stderr, "FPU ");
827 if (flags & MASK_HARD_QUAD)
828 fprintf (stderr, "HARD_QUAD ");
829 if (flags & MASK_POPC)
830 fprintf (stderr, "POPC ");
831 if (flags & MASK_PTR64)
832 fprintf (stderr, "PTR64 ");
833 if (flags & MASK_STACK_BIAS)
834 fprintf (stderr, "STACK_BIAS ");
835 if (flags & MASK_UNALIGNED_DOUBLES)
836 fprintf (stderr, "UNALIGNED_DOUBLES ");
837 if (flags & MASK_V8PLUS)
838 fprintf (stderr, "V8PLUS ");
839 if (flags & MASK_VIS)
840 fprintf (stderr, "VIS ");
841 if (flags & MASK_VIS2)
842 fprintf (stderr, "VIS2 ");
843 if (flags & MASK_VIS3)
844 fprintf (stderr, "VIS3 ");
845 if (flags & MASK_DEPRECATED_V8_INSNS)
846 fprintf (stderr, "DEPRECATED_V8_INSNS ");
847 if (flags & MASK_SPARCLET)
848 fprintf (stderr, "SPARCLET ");
849 if (flags & MASK_SPARCLITE)
850 fprintf (stderr, "SPARCLITE ");
852 fprintf (stderr, "V8 ");
854 fprintf (stderr, "V9 ");
858 dump_target_flags (const char *prefix, const int flags)
860 fprintf (stderr, "%s: (%08x) [ ", prefix, flags);
861 dump_target_flag_bits (flags);
862 fprintf(stderr, "]\n");
865 /* Validate and override various options, and do some machine dependent
869 sparc_option_override (void)
871 static struct code_model {
872 const char *const name;
873 const enum cmodel value;
874 } const cmodels[] = {
876 { "medlow", CM_MEDLOW },
877 { "medmid", CM_MEDMID },
878 { "medany", CM_MEDANY },
879 { "embmedany", CM_EMBMEDANY },
880 { NULL, (enum cmodel) 0 }
882 const struct code_model *cmodel;
883 /* Map TARGET_CPU_DEFAULT to value for -m{cpu,tune}=. */
884 static struct cpu_default {
886 const enum processor_type processor;
887 } const cpu_default[] = {
888 /* There must be one entry here for each TARGET_CPU value. */
889 { TARGET_CPU_sparc, PROCESSOR_CYPRESS },
890 { TARGET_CPU_v8, PROCESSOR_V8 },
891 { TARGET_CPU_supersparc, PROCESSOR_SUPERSPARC },
892 { TARGET_CPU_hypersparc, PROCESSOR_HYPERSPARC },
893 { TARGET_CPU_leon, PROCESSOR_LEON },
894 { TARGET_CPU_sparclite, PROCESSOR_F930 },
895 { TARGET_CPU_sparclite86x, PROCESSOR_SPARCLITE86X },
896 { TARGET_CPU_sparclet, PROCESSOR_TSC701 },
897 { TARGET_CPU_v9, PROCESSOR_V9 },
898 { TARGET_CPU_ultrasparc, PROCESSOR_ULTRASPARC },
899 { TARGET_CPU_ultrasparc3, PROCESSOR_ULTRASPARC3 },
900 { TARGET_CPU_niagara, PROCESSOR_NIAGARA },
901 { TARGET_CPU_niagara2, PROCESSOR_NIAGARA2 },
902 { TARGET_CPU_niagara3, PROCESSOR_NIAGARA3 },
903 { TARGET_CPU_niagara4, PROCESSOR_NIAGARA4 },
906 const struct cpu_default *def;
907 /* Table of values for -m{cpu,tune}=. This must match the order of
908 the PROCESSOR_* enumeration. */
909 static struct cpu_table {
910 const char *const name;
913 } const cpu_table[] = {
914 { "v7", MASK_ISA, 0 },
915 { "cypress", MASK_ISA, 0 },
916 { "v8", MASK_ISA, MASK_V8 },
917 /* TI TMS390Z55 supersparc */
918 { "supersparc", MASK_ISA, MASK_V8 },
919 { "hypersparc", MASK_ISA, MASK_V8|MASK_FPU },
921 { "leon", MASK_ISA, MASK_V8|MASK_FPU },
922 { "sparclite", MASK_ISA, MASK_SPARCLITE },
923 /* The Fujitsu MB86930 is the original sparclite chip, with no FPU. */
924 { "f930", MASK_ISA|MASK_FPU, MASK_SPARCLITE },
925 /* The Fujitsu MB86934 is the recent sparclite chip, with an FPU. */
926 { "f934", MASK_ISA, MASK_SPARCLITE|MASK_FPU },
927 { "sparclite86x", MASK_ISA|MASK_FPU, MASK_SPARCLITE },
928 { "sparclet", MASK_ISA, MASK_SPARCLET },
930 { "tsc701", MASK_ISA, MASK_SPARCLET },
931 { "v9", MASK_ISA, MASK_V9 },
932 /* UltraSPARC I, II, IIi */
933 { "ultrasparc", MASK_ISA,
934 /* Although insns using %y are deprecated, it is a clear win. */
935 MASK_V9|MASK_DEPRECATED_V8_INSNS },
937 /* ??? Check if %y issue still holds true. */
938 { "ultrasparc3", MASK_ISA,
939 MASK_V9|MASK_DEPRECATED_V8_INSNS|MASK_VIS2 },
941 { "niagara", MASK_ISA,
942 MASK_V9|MASK_DEPRECATED_V8_INSNS },
944 { "niagara2", MASK_ISA,
945 MASK_V9|MASK_POPC|MASK_VIS2 },
947 { "niagara3", MASK_ISA,
948 MASK_V9|MASK_POPC|MASK_VIS2|MASK_VIS3|MASK_FMAF },
950 { "niagara4", MASK_ISA,
951 MASK_V9|MASK_POPC|MASK_VIS2|MASK_VIS3|MASK_FMAF },
953 const struct cpu_table *cpu;
957 if (sparc_debug_string != NULL)
962 p = ASTRDUP (sparc_debug_string);
963 while ((q = strtok (p, ",")) != NULL)
977 if (! strcmp (q, "all"))
978 mask = MASK_DEBUG_ALL;
979 else if (! strcmp (q, "options"))
980 mask = MASK_DEBUG_OPTIONS;
982 error ("unknown -mdebug-%s switch", q);
985 sparc_debug &= ~mask;
991 if (TARGET_DEBUG_OPTIONS)
993 dump_target_flags("Initial target_flags", target_flags);
994 dump_target_flags("target_flags_explicit", target_flags_explicit);
997 #ifdef SUBTARGET_OVERRIDE_OPTIONS
998 SUBTARGET_OVERRIDE_OPTIONS;
1001 #ifndef SPARC_BI_ARCH
1002 /* Check for unsupported architecture size. */
1003 if (! TARGET_64BIT != DEFAULT_ARCH32_P)
1004 error ("%s is not supported by this configuration",
1005 DEFAULT_ARCH32_P ? "-m64" : "-m32");
1008 /* We force all 64bit archs to use 128 bit long double */
1009 if (TARGET_64BIT && ! TARGET_LONG_DOUBLE_128)
1011 error ("-mlong-double-64 not allowed with -m64");
1012 target_flags |= MASK_LONG_DOUBLE_128;
1015 /* Code model selection. */
1016 sparc_cmodel = SPARC_DEFAULT_CMODEL;
1018 #ifdef SPARC_BI_ARCH
1020 sparc_cmodel = CM_32;
1023 if (sparc_cmodel_string != NULL)
1027 for (cmodel = &cmodels[0]; cmodel->name; cmodel++)
1028 if (strcmp (sparc_cmodel_string, cmodel->name) == 0)
1030 if (cmodel->name == NULL)
1031 error ("bad value (%s) for -mcmodel= switch", sparc_cmodel_string);
1033 sparc_cmodel = cmodel->value;
1036 error ("-mcmodel= is not supported on 32 bit systems");
1039 /* Check that -fcall-saved-REG wasn't specified for out registers. */
1040 for (i = 8; i < 16; i++)
1041 if (!call_used_regs [i])
1043 error ("-fcall-saved-REG is not supported for out registers");
1044 call_used_regs [i] = 1;
1047 fpu = target_flags & MASK_FPU; /* save current -mfpu status */
1049 /* Set the default CPU. */
1050 if (!global_options_set.x_sparc_cpu_and_features)
1052 for (def = &cpu_default[0]; def->cpu != -1; ++def)
1053 if (def->cpu == TARGET_CPU_DEFAULT)
1055 gcc_assert (def->cpu != -1);
1056 sparc_cpu_and_features = def->processor;
1059 if (!global_options_set.x_sparc_cpu)
1060 sparc_cpu = sparc_cpu_and_features;
1062 cpu = &cpu_table[(int) sparc_cpu_and_features];
1064 if (TARGET_DEBUG_OPTIONS)
1066 fprintf (stderr, "sparc_cpu_and_features: %s\n", cpu->name);
1067 fprintf (stderr, "sparc_cpu: %s\n",
1068 cpu_table[(int) sparc_cpu].name);
1069 dump_target_flags ("cpu->disable", cpu->disable);
1070 dump_target_flags ("cpu->enable", cpu->enable);
1073 target_flags &= ~cpu->disable;
1074 target_flags |= (cpu->enable
1075 #ifndef HAVE_AS_FMAF_HPC_VIS3
1076 & ~(MASK_FMAF | MASK_VIS3)
1080 /* If -mfpu or -mno-fpu was explicitly used, don't override with
1081 the processor default. */
1082 if (target_flags_explicit & MASK_FPU)
1083 target_flags = (target_flags & ~MASK_FPU) | fpu;
1085 /* -mvis2 implies -mvis */
1087 target_flags |= MASK_VIS;
1089 /* -mvis3 implies -mvis2 and -mvis */
1091 target_flags |= MASK_VIS2 | MASK_VIS;
1093 /* Don't allow -mvis, -mvis2, -mvis3, or -mfmaf if FPU is disabled. */
1095 target_flags &= ~(MASK_VIS | MASK_VIS2 | MASK_VIS3 | MASK_FMAF);
1097 /* -mvis assumes UltraSPARC+, so we are sure v9 instructions
1099 -m64 also implies v9. */
1100 if (TARGET_VIS || TARGET_ARCH64)
1102 target_flags |= MASK_V9;
1103 target_flags &= ~(MASK_V8 | MASK_SPARCLET | MASK_SPARCLITE);
1106 /* -mvis also implies -mv8plus on 32-bit */
1107 if (TARGET_VIS && ! TARGET_ARCH64)
1108 target_flags |= MASK_V8PLUS;
1110 /* Use the deprecated v8 insns for sparc64 in 32 bit mode. */
1111 if (TARGET_V9 && TARGET_ARCH32)
1112 target_flags |= MASK_DEPRECATED_V8_INSNS;
1114 /* V8PLUS requires V9, makes no sense in 64 bit mode. */
1115 if (! TARGET_V9 || TARGET_ARCH64)
1116 target_flags &= ~MASK_V8PLUS;
1118 /* Don't use stack biasing in 32 bit mode. */
1120 target_flags &= ~MASK_STACK_BIAS;
1122 /* Supply a default value for align_functions. */
1123 if (align_functions == 0
1124 && (sparc_cpu == PROCESSOR_ULTRASPARC
1125 || sparc_cpu == PROCESSOR_ULTRASPARC3
1126 || sparc_cpu == PROCESSOR_NIAGARA
1127 || sparc_cpu == PROCESSOR_NIAGARA2
1128 || sparc_cpu == PROCESSOR_NIAGARA3
1129 || sparc_cpu == PROCESSOR_NIAGARA4))
1130 align_functions = 32;
1132 /* Validate PCC_STRUCT_RETURN. */
1133 if (flag_pcc_struct_return == DEFAULT_PCC_STRUCT_RETURN)
1134 flag_pcc_struct_return = (TARGET_ARCH64 ? 0 : 1);
1136 /* Only use .uaxword when compiling for a 64-bit target. */
1138 targetm.asm_out.unaligned_op.di = NULL;
1140 /* Do various machine dependent initializations. */
1141 sparc_init_modes ();
1143 /* Set up function hooks. */
1144 init_machine_status = sparc_init_machine_status;
1149 case PROCESSOR_CYPRESS:
1150 sparc_costs = &cypress_costs;
1153 case PROCESSOR_SPARCLITE:
1154 case PROCESSOR_SUPERSPARC:
1155 sparc_costs = &supersparc_costs;
1157 case PROCESSOR_F930:
1158 case PROCESSOR_F934:
1159 case PROCESSOR_HYPERSPARC:
1160 case PROCESSOR_SPARCLITE86X:
1161 sparc_costs = &hypersparc_costs;
1163 case PROCESSOR_LEON:
1164 sparc_costs = &leon_costs;
1166 case PROCESSOR_SPARCLET:
1167 case PROCESSOR_TSC701:
1168 sparc_costs = &sparclet_costs;
1171 case PROCESSOR_ULTRASPARC:
1172 sparc_costs = &ultrasparc_costs;
1174 case PROCESSOR_ULTRASPARC3:
1175 sparc_costs = &ultrasparc3_costs;
1177 case PROCESSOR_NIAGARA:
1178 sparc_costs = &niagara_costs;
1180 case PROCESSOR_NIAGARA2:
1181 sparc_costs = &niagara2_costs;
1183 case PROCESSOR_NIAGARA3:
1184 sparc_costs = &niagara3_costs;
1186 case PROCESSOR_NIAGARA4:
1187 sparc_costs = &niagara4_costs;
1189 case PROCESSOR_NATIVE:
1193 if (sparc_memory_model == SMM_DEFAULT)
1195 /* Choose the memory model for the operating system. */
1196 enum sparc_memory_model_type os_default = SUBTARGET_DEFAULT_MEMORY_MODEL;
1197 if (os_default != SMM_DEFAULT)
1198 sparc_memory_model = os_default;
1199 /* Choose the most relaxed model for the processor. */
1201 sparc_memory_model = SMM_RMO;
1203 sparc_memory_model = SMM_PSO;
1205 sparc_memory_model = SMM_SC;
1208 #ifdef TARGET_DEFAULT_LONG_DOUBLE_128
1209 if (!(target_flags_explicit & MASK_LONG_DOUBLE_128))
1210 target_flags |= MASK_LONG_DOUBLE_128;
1213 if (TARGET_DEBUG_OPTIONS)
1214 dump_target_flags ("Final target_flags", target_flags);
1216 maybe_set_param_value (PARAM_SIMULTANEOUS_PREFETCHES,
1217 ((sparc_cpu == PROCESSOR_ULTRASPARC
1218 || sparc_cpu == PROCESSOR_NIAGARA
1219 || sparc_cpu == PROCESSOR_NIAGARA2
1220 || sparc_cpu == PROCESSOR_NIAGARA3
1221 || sparc_cpu == PROCESSOR_NIAGARA4)
1223 : (sparc_cpu == PROCESSOR_ULTRASPARC3
1225 global_options.x_param_values,
1226 global_options_set.x_param_values);
1227 maybe_set_param_value (PARAM_L1_CACHE_LINE_SIZE,
1228 ((sparc_cpu == PROCESSOR_ULTRASPARC
1229 || sparc_cpu == PROCESSOR_ULTRASPARC3
1230 || sparc_cpu == PROCESSOR_NIAGARA
1231 || sparc_cpu == PROCESSOR_NIAGARA2
1232 || sparc_cpu == PROCESSOR_NIAGARA3
1233 || sparc_cpu == PROCESSOR_NIAGARA4)
1235 global_options.x_param_values,
1236 global_options_set.x_param_values);
1238 /* Disable save slot sharing for call-clobbered registers by default.
1239 The IRA sharing algorithm works on single registers only and this
1240 pessimizes for double floating-point registers. */
1241 if (!global_options_set.x_flag_ira_share_save_slots)
1242 flag_ira_share_save_slots = 0;
1245 /* Miscellaneous utilities. */
1247 /* Nonzero if CODE, a comparison, is suitable for use in v9 conditional move
1248 or branch on register contents instructions. */
1251 v9_regcmp_p (enum rtx_code code)
1253 return (code == EQ || code == NE || code == GE || code == LT
1254 || code == LE || code == GT);
1257 /* Nonzero if OP is a floating point constant which can
1258 be loaded into an integer register using a single
1259 sethi instruction. */
1264 if (GET_CODE (op) == CONST_DOUBLE)
1269 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
1270 REAL_VALUE_TO_TARGET_SINGLE (r, i);
1271 return !SPARC_SIMM13_P (i) && SPARC_SETHI_P (i);
1277 /* Nonzero if OP is a floating point constant which can
1278 be loaded into an integer register using a single
1284 if (GET_CODE (op) == CONST_DOUBLE)
1289 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
1290 REAL_VALUE_TO_TARGET_SINGLE (r, i);
1291 return SPARC_SIMM13_P (i);
1297 /* Nonzero if OP is a floating point constant which can
1298 be loaded into an integer register using a high/losum
1299 instruction sequence. */
1302 fp_high_losum_p (rtx op)
1304 /* The constraints calling this should only be in
1305 SFmode move insns, so any constant which cannot
1306 be moved using a single insn will do. */
1307 if (GET_CODE (op) == CONST_DOUBLE)
1312 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
1313 REAL_VALUE_TO_TARGET_SINGLE (r, i);
1314 return !SPARC_SIMM13_P (i) && !SPARC_SETHI_P (i);
1320 /* Return true if the address of LABEL can be loaded by means of the
1321 mov{si,di}_pic_label_ref patterns in PIC mode. */
1324 can_use_mov_pic_label_ref (rtx label)
1326 /* VxWorks does not impose a fixed gap between segments; the run-time
1327 gap can be different from the object-file gap. We therefore can't
1328 assume X - _GLOBAL_OFFSET_TABLE_ is a link-time constant unless we
1329 are absolutely sure that X is in the same segment as the GOT.
1330 Unfortunately, the flexibility of linker scripts means that we
1331 can't be sure of that in general, so assume that GOT-relative
1332 accesses are never valid on VxWorks. */
1333 if (TARGET_VXWORKS_RTP)
1336 /* Similarly, if the label is non-local, it might end up being placed
1337 in a different section than the current one; now mov_pic_label_ref
1338 requires the label and the code to be in the same section. */
1339 if (LABEL_REF_NONLOCAL_P (label))
1342 /* Finally, if we are reordering basic blocks and partition into hot
1343 and cold sections, this might happen for any label. */
1344 if (flag_reorder_blocks_and_partition)
1350 /* Expand a move instruction. Return true if all work is done. */
1353 sparc_expand_move (enum machine_mode mode, rtx *operands)
1355 /* Handle sets of MEM first. */
1356 if (GET_CODE (operands[0]) == MEM)
1358 /* 0 is a register (or a pair of registers) on SPARC. */
1359 if (register_or_zero_operand (operands[1], mode))
1362 if (!reload_in_progress)
1364 operands[0] = validize_mem (operands[0]);
1365 operands[1] = force_reg (mode, operands[1]);
1369 /* Fixup TLS cases. */
1371 && CONSTANT_P (operands[1])
1372 && sparc_tls_referenced_p (operands [1]))
1374 operands[1] = sparc_legitimize_tls_address (operands[1]);
1378 /* Fixup PIC cases. */
1379 if (flag_pic && CONSTANT_P (operands[1]))
1381 if (pic_address_needs_scratch (operands[1]))
1382 operands[1] = sparc_legitimize_pic_address (operands[1], NULL_RTX);
1384 /* We cannot use the mov{si,di}_pic_label_ref patterns in all cases. */
1385 if (GET_CODE (operands[1]) == LABEL_REF
1386 && can_use_mov_pic_label_ref (operands[1]))
1390 emit_insn (gen_movsi_pic_label_ref (operands[0], operands[1]));
1396 gcc_assert (TARGET_ARCH64);
1397 emit_insn (gen_movdi_pic_label_ref (operands[0], operands[1]));
1402 if (symbolic_operand (operands[1], mode))
1405 = sparc_legitimize_pic_address (operands[1],
1407 ? operands[0] : NULL_RTX);
1412 /* If we are trying to toss an integer constant into FP registers,
1413 or loading a FP or vector constant, force it into memory. */
1414 if (CONSTANT_P (operands[1])
1415 && REG_P (operands[0])
1416 && (SPARC_FP_REG_P (REGNO (operands[0]))
1417 || SCALAR_FLOAT_MODE_P (mode)
1418 || VECTOR_MODE_P (mode)))
1420 /* emit_group_store will send such bogosity to us when it is
1421 not storing directly into memory. So fix this up to avoid
1422 crashes in output_constant_pool. */
1423 if (operands [1] == const0_rtx)
1424 operands[1] = CONST0_RTX (mode);
1426 /* We can clear or set to all-ones FP registers if TARGET_VIS, and
1427 always other regs. */
1428 if ((TARGET_VIS || REGNO (operands[0]) < SPARC_FIRST_FP_REG)
1429 && (const_zero_operand (operands[1], mode)
1430 || const_all_ones_operand (operands[1], mode)))
1433 if (REGNO (operands[0]) < SPARC_FIRST_FP_REG
1434 /* We are able to build any SF constant in integer registers
1435 with at most 2 instructions. */
1437 /* And any DF constant in integer registers. */
1439 && ! can_create_pseudo_p ())))
1442 operands[1] = force_const_mem (mode, operands[1]);
1443 if (!reload_in_progress)
1444 operands[1] = validize_mem (operands[1]);
1448 /* Accept non-constants and valid constants unmodified. */
1449 if (!CONSTANT_P (operands[1])
1450 || GET_CODE (operands[1]) == HIGH
1451 || input_operand (operands[1], mode))
1457 /* All QImode constants require only one insn, so proceed. */
1462 sparc_emit_set_const32 (operands[0], operands[1]);
1466 /* input_operand should have filtered out 32-bit mode. */
1467 sparc_emit_set_const64 (operands[0], operands[1]);
1477 /* Load OP1, a 32-bit constant, into OP0, a register.
1478 We know it can't be done in one insn when we get
1479 here, the move expander guarantees this. */
1482 sparc_emit_set_const32 (rtx op0, rtx op1)
1484 enum machine_mode mode = GET_MODE (op0);
1487 if (can_create_pseudo_p ())
1488 temp = gen_reg_rtx (mode);
1490 if (GET_CODE (op1) == CONST_INT)
1492 gcc_assert (!small_int_operand (op1, mode)
1493 && !const_high_operand (op1, mode));
1495 /* Emit them as real moves instead of a HIGH/LO_SUM,
1496 this way CSE can see everything and reuse intermediate
1497 values if it wants. */
1498 emit_insn (gen_rtx_SET (VOIDmode, temp,
1499 GEN_INT (INTVAL (op1)
1500 & ~(HOST_WIDE_INT)0x3ff)));
1502 emit_insn (gen_rtx_SET (VOIDmode,
1504 gen_rtx_IOR (mode, temp,
1505 GEN_INT (INTVAL (op1) & 0x3ff))));
1509 /* A symbol, emit in the traditional way. */
1510 emit_insn (gen_rtx_SET (VOIDmode, temp,
1511 gen_rtx_HIGH (mode, op1)));
1512 emit_insn (gen_rtx_SET (VOIDmode,
1513 op0, gen_rtx_LO_SUM (mode, temp, op1)));
1517 /* Load OP1, a symbolic 64-bit constant, into OP0, a DImode register.
1518 If TEMP is nonzero, we are forbidden to use any other scratch
1519 registers. Otherwise, we are allowed to generate them as needed.
1521 Note that TEMP may have TImode if the code model is TARGET_CM_MEDANY
1522 or TARGET_CM_EMBMEDANY (see the reload_indi and reload_outdi patterns). */
1525 sparc_emit_set_symbolic_const64 (rtx op0, rtx op1, rtx temp)
1527 rtx temp1, temp2, temp3, temp4, temp5;
1530 if (temp && GET_MODE (temp) == TImode)
1533 temp = gen_rtx_REG (DImode, REGNO (temp));
1536 /* SPARC-V9 code-model support. */
1537 switch (sparc_cmodel)
1540 /* The range spanned by all instructions in the object is less
1541 than 2^31 bytes (2GB) and the distance from any instruction
1542 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1543 than 2^31 bytes (2GB).
1545 The executable must be in the low 4TB of the virtual address
1548 sethi %hi(symbol), %temp1
1549 or %temp1, %lo(symbol), %reg */
1551 temp1 = temp; /* op0 is allowed. */
1553 temp1 = gen_reg_rtx (DImode);
1555 emit_insn (gen_rtx_SET (VOIDmode, temp1, gen_rtx_HIGH (DImode, op1)));
1556 emit_insn (gen_rtx_SET (VOIDmode, op0, gen_rtx_LO_SUM (DImode, temp1, op1)));
1560 /* The range spanned by all instructions in the object is less
1561 than 2^31 bytes (2GB) and the distance from any instruction
1562 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1563 than 2^31 bytes (2GB).
1565 The executable must be in the low 16TB of the virtual address
1568 sethi %h44(symbol), %temp1
1569 or %temp1, %m44(symbol), %temp2
1570 sllx %temp2, 12, %temp3
1571 or %temp3, %l44(symbol), %reg */
1576 temp3 = temp; /* op0 is allowed. */
1580 temp1 = gen_reg_rtx (DImode);
1581 temp2 = gen_reg_rtx (DImode);
1582 temp3 = gen_reg_rtx (DImode);
1585 emit_insn (gen_seth44 (temp1, op1));
1586 emit_insn (gen_setm44 (temp2, temp1, op1));
1587 emit_insn (gen_rtx_SET (VOIDmode, temp3,
1588 gen_rtx_ASHIFT (DImode, temp2, GEN_INT (12))));
1589 emit_insn (gen_setl44 (op0, temp3, op1));
1593 /* The range spanned by all instructions in the object is less
1594 than 2^31 bytes (2GB) and the distance from any instruction
1595 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1596 than 2^31 bytes (2GB).
1598 The executable can be placed anywhere in the virtual address
1601 sethi %hh(symbol), %temp1
1602 sethi %lm(symbol), %temp2
1603 or %temp1, %hm(symbol), %temp3
1604 sllx %temp3, 32, %temp4
1605 or %temp4, %temp2, %temp5
1606 or %temp5, %lo(symbol), %reg */
1609 /* It is possible that one of the registers we got for operands[2]
1610 might coincide with that of operands[0] (which is why we made
1611 it TImode). Pick the other one to use as our scratch. */
1612 if (rtx_equal_p (temp, op0))
1614 gcc_assert (ti_temp);
1615 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1618 temp2 = temp; /* op0 is _not_ allowed, see above. */
1625 temp1 = gen_reg_rtx (DImode);
1626 temp2 = gen_reg_rtx (DImode);
1627 temp3 = gen_reg_rtx (DImode);
1628 temp4 = gen_reg_rtx (DImode);
1629 temp5 = gen_reg_rtx (DImode);
1632 emit_insn (gen_sethh (temp1, op1));
1633 emit_insn (gen_setlm (temp2, op1));
1634 emit_insn (gen_sethm (temp3, temp1, op1));
1635 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1636 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1637 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1638 gen_rtx_PLUS (DImode, temp4, temp2)));
1639 emit_insn (gen_setlo (op0, temp5, op1));
1643 /* Old old old backwards compatibility kruft here.
1644 Essentially it is MEDLOW with a fixed 64-bit
1645 virtual base added to all data segment addresses.
1646 Text-segment stuff is computed like MEDANY, we can't
1647 reuse the code above because the relocation knobs
1650 Data segment: sethi %hi(symbol), %temp1
1651 add %temp1, EMBMEDANY_BASE_REG, %temp2
1652 or %temp2, %lo(symbol), %reg */
1653 if (data_segment_operand (op1, GET_MODE (op1)))
1657 temp1 = temp; /* op0 is allowed. */
1662 temp1 = gen_reg_rtx (DImode);
1663 temp2 = gen_reg_rtx (DImode);
1666 emit_insn (gen_embmedany_sethi (temp1, op1));
1667 emit_insn (gen_embmedany_brsum (temp2, temp1));
1668 emit_insn (gen_embmedany_losum (op0, temp2, op1));
1671 /* Text segment: sethi %uhi(symbol), %temp1
1672 sethi %hi(symbol), %temp2
1673 or %temp1, %ulo(symbol), %temp3
1674 sllx %temp3, 32, %temp4
1675 or %temp4, %temp2, %temp5
1676 or %temp5, %lo(symbol), %reg */
1681 /* It is possible that one of the registers we got for operands[2]
1682 might coincide with that of operands[0] (which is why we made
1683 it TImode). Pick the other one to use as our scratch. */
1684 if (rtx_equal_p (temp, op0))
1686 gcc_assert (ti_temp);
1687 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1690 temp2 = temp; /* op0 is _not_ allowed, see above. */
1697 temp1 = gen_reg_rtx (DImode);
1698 temp2 = gen_reg_rtx (DImode);
1699 temp3 = gen_reg_rtx (DImode);
1700 temp4 = gen_reg_rtx (DImode);
1701 temp5 = gen_reg_rtx (DImode);
1704 emit_insn (gen_embmedany_textuhi (temp1, op1));
1705 emit_insn (gen_embmedany_texthi (temp2, op1));
1706 emit_insn (gen_embmedany_textulo (temp3, temp1, op1));
1707 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1708 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1709 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1710 gen_rtx_PLUS (DImode, temp4, temp2)));
1711 emit_insn (gen_embmedany_textlo (op0, temp5, op1));
1720 #if HOST_BITS_PER_WIDE_INT == 32
1722 sparc_emit_set_const64 (rtx op0 ATTRIBUTE_UNUSED, rtx op1 ATTRIBUTE_UNUSED)
1727 /* These avoid problems when cross compiling. If we do not
1728 go through all this hair then the optimizer will see
1729 invalid REG_EQUAL notes or in some cases none at all. */
1730 static rtx gen_safe_HIGH64 (rtx, HOST_WIDE_INT);
1731 static rtx gen_safe_SET64 (rtx, HOST_WIDE_INT);
1732 static rtx gen_safe_OR64 (rtx, HOST_WIDE_INT);
1733 static rtx gen_safe_XOR64 (rtx, HOST_WIDE_INT);
1735 /* The optimizer is not to assume anything about exactly
1736 which bits are set for a HIGH, they are unspecified.
1737 Unfortunately this leads to many missed optimizations
1738 during CSE. We mask out the non-HIGH bits, and matches
1739 a plain movdi, to alleviate this problem. */
1741 gen_safe_HIGH64 (rtx dest, HOST_WIDE_INT val)
1743 return gen_rtx_SET (VOIDmode, dest, GEN_INT (val & ~(HOST_WIDE_INT)0x3ff));
1747 gen_safe_SET64 (rtx dest, HOST_WIDE_INT val)
1749 return gen_rtx_SET (VOIDmode, dest, GEN_INT (val));
1753 gen_safe_OR64 (rtx src, HOST_WIDE_INT val)
1755 return gen_rtx_IOR (DImode, src, GEN_INT (val));
1759 gen_safe_XOR64 (rtx src, HOST_WIDE_INT val)
1761 return gen_rtx_XOR (DImode, src, GEN_INT (val));
1764 /* Worker routines for 64-bit constant formation on arch64.
1765 One of the key things to be doing in these emissions is
1766 to create as many temp REGs as possible. This makes it
1767 possible for half-built constants to be used later when
1768 such values are similar to something required later on.
1769 Without doing this, the optimizer cannot see such
1772 static void sparc_emit_set_const64_quick1 (rtx, rtx,
1773 unsigned HOST_WIDE_INT, int);
1776 sparc_emit_set_const64_quick1 (rtx op0, rtx temp,
1777 unsigned HOST_WIDE_INT low_bits, int is_neg)
1779 unsigned HOST_WIDE_INT high_bits;
1782 high_bits = (~low_bits) & 0xffffffff;
1784 high_bits = low_bits;
1786 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1789 emit_insn (gen_rtx_SET (VOIDmode, op0,
1790 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1794 /* If we are XOR'ing with -1, then we should emit a one's complement
1795 instead. This way the combiner will notice logical operations
1796 such as ANDN later on and substitute. */
1797 if ((low_bits & 0x3ff) == 0x3ff)
1799 emit_insn (gen_rtx_SET (VOIDmode, op0,
1800 gen_rtx_NOT (DImode, temp)));
1804 emit_insn (gen_rtx_SET (VOIDmode, op0,
1805 gen_safe_XOR64 (temp,
1806 (-(HOST_WIDE_INT)0x400
1807 | (low_bits & 0x3ff)))));
1812 static void sparc_emit_set_const64_quick2 (rtx, rtx, unsigned HOST_WIDE_INT,
1813 unsigned HOST_WIDE_INT, int);
1816 sparc_emit_set_const64_quick2 (rtx op0, rtx temp,
1817 unsigned HOST_WIDE_INT high_bits,
1818 unsigned HOST_WIDE_INT low_immediate,
1823 if ((high_bits & 0xfffffc00) != 0)
1825 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1826 if ((high_bits & ~0xfffffc00) != 0)
1827 emit_insn (gen_rtx_SET (VOIDmode, op0,
1828 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1834 emit_insn (gen_safe_SET64 (temp, high_bits));
1838 /* Now shift it up into place. */
1839 emit_insn (gen_rtx_SET (VOIDmode, op0,
1840 gen_rtx_ASHIFT (DImode, temp2,
1841 GEN_INT (shift_count))));
1843 /* If there is a low immediate part piece, finish up by
1844 putting that in as well. */
1845 if (low_immediate != 0)
1846 emit_insn (gen_rtx_SET (VOIDmode, op0,
1847 gen_safe_OR64 (op0, low_immediate)));
1850 static void sparc_emit_set_const64_longway (rtx, rtx, unsigned HOST_WIDE_INT,
1851 unsigned HOST_WIDE_INT);
1853 /* Full 64-bit constant decomposition. Even though this is the
1854 'worst' case, we still optimize a few things away. */
1856 sparc_emit_set_const64_longway (rtx op0, rtx temp,
1857 unsigned HOST_WIDE_INT high_bits,
1858 unsigned HOST_WIDE_INT low_bits)
1862 if (can_create_pseudo_p ())
1863 sub_temp = gen_reg_rtx (DImode);
1865 if ((high_bits & 0xfffffc00) != 0)
1867 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1868 if ((high_bits & ~0xfffffc00) != 0)
1869 emit_insn (gen_rtx_SET (VOIDmode,
1871 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1877 emit_insn (gen_safe_SET64 (temp, high_bits));
1881 if (can_create_pseudo_p ())
1883 rtx temp2 = gen_reg_rtx (DImode);
1884 rtx temp3 = gen_reg_rtx (DImode);
1885 rtx temp4 = gen_reg_rtx (DImode);
1887 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1888 gen_rtx_ASHIFT (DImode, sub_temp,
1891 emit_insn (gen_safe_HIGH64 (temp2, low_bits));
1892 if ((low_bits & ~0xfffffc00) != 0)
1894 emit_insn (gen_rtx_SET (VOIDmode, temp3,
1895 gen_safe_OR64 (temp2, (low_bits & 0x3ff))));
1896 emit_insn (gen_rtx_SET (VOIDmode, op0,
1897 gen_rtx_PLUS (DImode, temp4, temp3)));
1901 emit_insn (gen_rtx_SET (VOIDmode, op0,
1902 gen_rtx_PLUS (DImode, temp4, temp2)));
1907 rtx low1 = GEN_INT ((low_bits >> (32 - 12)) & 0xfff);
1908 rtx low2 = GEN_INT ((low_bits >> (32 - 12 - 12)) & 0xfff);
1909 rtx low3 = GEN_INT ((low_bits >> (32 - 12 - 12 - 8)) & 0x0ff);
1912 /* We are in the middle of reload, so this is really
1913 painful. However we do still make an attempt to
1914 avoid emitting truly stupid code. */
1915 if (low1 != const0_rtx)
1917 emit_insn (gen_rtx_SET (VOIDmode, op0,
1918 gen_rtx_ASHIFT (DImode, sub_temp,
1919 GEN_INT (to_shift))));
1920 emit_insn (gen_rtx_SET (VOIDmode, op0,
1921 gen_rtx_IOR (DImode, op0, low1)));
1929 if (low2 != const0_rtx)
1931 emit_insn (gen_rtx_SET (VOIDmode, op0,
1932 gen_rtx_ASHIFT (DImode, sub_temp,
1933 GEN_INT (to_shift))));
1934 emit_insn (gen_rtx_SET (VOIDmode, op0,
1935 gen_rtx_IOR (DImode, op0, low2)));
1943 emit_insn (gen_rtx_SET (VOIDmode, op0,
1944 gen_rtx_ASHIFT (DImode, sub_temp,
1945 GEN_INT (to_shift))));
1946 if (low3 != const0_rtx)
1947 emit_insn (gen_rtx_SET (VOIDmode, op0,
1948 gen_rtx_IOR (DImode, op0, low3)));
1953 /* Analyze a 64-bit constant for certain properties. */
1954 static void analyze_64bit_constant (unsigned HOST_WIDE_INT,
1955 unsigned HOST_WIDE_INT,
1956 int *, int *, int *);
1959 analyze_64bit_constant (unsigned HOST_WIDE_INT high_bits,
1960 unsigned HOST_WIDE_INT low_bits,
1961 int *hbsp, int *lbsp, int *abbasp)
1963 int lowest_bit_set, highest_bit_set, all_bits_between_are_set;
1966 lowest_bit_set = highest_bit_set = -1;
1970 if ((lowest_bit_set == -1)
1971 && ((low_bits >> i) & 1))
1973 if ((highest_bit_set == -1)
1974 && ((high_bits >> (32 - i - 1)) & 1))
1975 highest_bit_set = (64 - i - 1);
1978 && ((highest_bit_set == -1)
1979 || (lowest_bit_set == -1)));
1985 if ((lowest_bit_set == -1)
1986 && ((high_bits >> i) & 1))
1987 lowest_bit_set = i + 32;
1988 if ((highest_bit_set == -1)
1989 && ((low_bits >> (32 - i - 1)) & 1))
1990 highest_bit_set = 32 - i - 1;
1993 && ((highest_bit_set == -1)
1994 || (lowest_bit_set == -1)));
1996 /* If there are no bits set this should have gone out
1997 as one instruction! */
1998 gcc_assert (lowest_bit_set != -1 && highest_bit_set != -1);
1999 all_bits_between_are_set = 1;
2000 for (i = lowest_bit_set; i <= highest_bit_set; i++)
2004 if ((low_bits & (1 << i)) != 0)
2009 if ((high_bits & (1 << (i - 32))) != 0)
2012 all_bits_between_are_set = 0;
2015 *hbsp = highest_bit_set;
2016 *lbsp = lowest_bit_set;
2017 *abbasp = all_bits_between_are_set;
2020 static int const64_is_2insns (unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT);
2023 const64_is_2insns (unsigned HOST_WIDE_INT high_bits,
2024 unsigned HOST_WIDE_INT low_bits)
2026 int highest_bit_set, lowest_bit_set, all_bits_between_are_set;
2029 || high_bits == 0xffffffff)
2032 analyze_64bit_constant (high_bits, low_bits,
2033 &highest_bit_set, &lowest_bit_set,
2034 &all_bits_between_are_set);
2036 if ((highest_bit_set == 63
2037 || lowest_bit_set == 0)
2038 && all_bits_between_are_set != 0)
2041 if ((highest_bit_set - lowest_bit_set) < 21)
2047 static unsigned HOST_WIDE_INT create_simple_focus_bits (unsigned HOST_WIDE_INT,
2048 unsigned HOST_WIDE_INT,
2051 static unsigned HOST_WIDE_INT
2052 create_simple_focus_bits (unsigned HOST_WIDE_INT high_bits,
2053 unsigned HOST_WIDE_INT low_bits,
2054 int lowest_bit_set, int shift)
2056 HOST_WIDE_INT hi, lo;
2058 if (lowest_bit_set < 32)
2060 lo = (low_bits >> lowest_bit_set) << shift;
2061 hi = ((high_bits << (32 - lowest_bit_set)) << shift);
2066 hi = ((high_bits >> (lowest_bit_set - 32)) << shift);
2068 gcc_assert (! (hi & lo));
2072 /* Here we are sure to be arch64 and this is an integer constant
2073 being loaded into a register. Emit the most efficient
2074 insn sequence possible. Detection of all the 1-insn cases
2075 has been done already. */
2077 sparc_emit_set_const64 (rtx op0, rtx op1)
2079 unsigned HOST_WIDE_INT high_bits, low_bits;
2080 int lowest_bit_set, highest_bit_set;
2081 int all_bits_between_are_set;
2084 /* Sanity check that we know what we are working with. */
2085 gcc_assert (TARGET_ARCH64
2086 && (GET_CODE (op0) == SUBREG
2087 || (REG_P (op0) && ! SPARC_FP_REG_P (REGNO (op0)))));
2089 if (! can_create_pseudo_p ())
2092 if (GET_CODE (op1) != CONST_INT)
2094 sparc_emit_set_symbolic_const64 (op0, op1, temp);
2099 temp = gen_reg_rtx (DImode);
2101 high_bits = ((INTVAL (op1) >> 32) & 0xffffffff);
2102 low_bits = (INTVAL (op1) & 0xffffffff);
2104 /* low_bits bits 0 --> 31
2105 high_bits bits 32 --> 63 */
2107 analyze_64bit_constant (high_bits, low_bits,
2108 &highest_bit_set, &lowest_bit_set,
2109 &all_bits_between_are_set);
2111 /* First try for a 2-insn sequence. */
2113 /* These situations are preferred because the optimizer can
2114 * do more things with them:
2116 * sllx %reg, shift, %reg
2118 * srlx %reg, shift, %reg
2119 * 3) mov some_small_const, %reg
2120 * sllx %reg, shift, %reg
2122 if (((highest_bit_set == 63
2123 || lowest_bit_set == 0)
2124 && all_bits_between_are_set != 0)
2125 || ((highest_bit_set - lowest_bit_set) < 12))
2127 HOST_WIDE_INT the_const = -1;
2128 int shift = lowest_bit_set;
2130 if ((highest_bit_set != 63
2131 && lowest_bit_set != 0)
2132 || all_bits_between_are_set == 0)
2135 create_simple_focus_bits (high_bits, low_bits,
2138 else if (lowest_bit_set == 0)
2139 shift = -(63 - highest_bit_set);
2141 gcc_assert (SPARC_SIMM13_P (the_const));
2142 gcc_assert (shift != 0);
2144 emit_insn (gen_safe_SET64 (temp, the_const));
2146 emit_insn (gen_rtx_SET (VOIDmode,
2148 gen_rtx_ASHIFT (DImode,
2152 emit_insn (gen_rtx_SET (VOIDmode,
2154 gen_rtx_LSHIFTRT (DImode,
2156 GEN_INT (-shift))));
2160 /* Now a range of 22 or less bits set somewhere.
2161 * 1) sethi %hi(focus_bits), %reg
2162 * sllx %reg, shift, %reg
2163 * 2) sethi %hi(focus_bits), %reg
2164 * srlx %reg, shift, %reg
2166 if ((highest_bit_set - lowest_bit_set) < 21)
2168 unsigned HOST_WIDE_INT focus_bits =
2169 create_simple_focus_bits (high_bits, low_bits,
2170 lowest_bit_set, 10);
2172 gcc_assert (SPARC_SETHI_P (focus_bits));
2173 gcc_assert (lowest_bit_set != 10);
2175 emit_insn (gen_safe_HIGH64 (temp, focus_bits));
2177 /* If lowest_bit_set == 10 then a sethi alone could have done it. */
2178 if (lowest_bit_set < 10)
2179 emit_insn (gen_rtx_SET (VOIDmode,
2181 gen_rtx_LSHIFTRT (DImode, temp,
2182 GEN_INT (10 - lowest_bit_set))));
2183 else if (lowest_bit_set > 10)
2184 emit_insn (gen_rtx_SET (VOIDmode,
2186 gen_rtx_ASHIFT (DImode, temp,
2187 GEN_INT (lowest_bit_set - 10))));
2191 /* 1) sethi %hi(low_bits), %reg
2192 * or %reg, %lo(low_bits), %reg
2193 * 2) sethi %hi(~low_bits), %reg
2194 * xor %reg, %lo(-0x400 | (low_bits & 0x3ff)), %reg
2197 || high_bits == 0xffffffff)
2199 sparc_emit_set_const64_quick1 (op0, temp, low_bits,
2200 (high_bits == 0xffffffff));
2204 /* Now, try 3-insn sequences. */
2206 /* 1) sethi %hi(high_bits), %reg
2207 * or %reg, %lo(high_bits), %reg
2208 * sllx %reg, 32, %reg
2212 sparc_emit_set_const64_quick2 (op0, temp, high_bits, 0, 32);
2216 /* We may be able to do something quick
2217 when the constant is negated, so try that. */
2218 if (const64_is_2insns ((~high_bits) & 0xffffffff,
2219 (~low_bits) & 0xfffffc00))
2221 /* NOTE: The trailing bits get XOR'd so we need the
2222 non-negated bits, not the negated ones. */
2223 unsigned HOST_WIDE_INT trailing_bits = low_bits & 0x3ff;
2225 if ((((~high_bits) & 0xffffffff) == 0
2226 && ((~low_bits) & 0x80000000) == 0)
2227 || (((~high_bits) & 0xffffffff) == 0xffffffff
2228 && ((~low_bits) & 0x80000000) != 0))
2230 unsigned HOST_WIDE_INT fast_int = (~low_bits & 0xffffffff);
2232 if ((SPARC_SETHI_P (fast_int)
2233 && (~high_bits & 0xffffffff) == 0)
2234 || SPARC_SIMM13_P (fast_int))
2235 emit_insn (gen_safe_SET64 (temp, fast_int));
2237 sparc_emit_set_const64 (temp, GEN_INT (fast_int));
2242 negated_const = GEN_INT (((~low_bits) & 0xfffffc00) |
2243 (((HOST_WIDE_INT)((~high_bits) & 0xffffffff))<<32));
2244 sparc_emit_set_const64 (temp, negated_const);
2247 /* If we are XOR'ing with -1, then we should emit a one's complement
2248 instead. This way the combiner will notice logical operations
2249 such as ANDN later on and substitute. */
2250 if (trailing_bits == 0x3ff)
2252 emit_insn (gen_rtx_SET (VOIDmode, op0,
2253 gen_rtx_NOT (DImode, temp)));
2257 emit_insn (gen_rtx_SET (VOIDmode,
2259 gen_safe_XOR64 (temp,
2260 (-0x400 | trailing_bits))));
2265 /* 1) sethi %hi(xxx), %reg
2266 * or %reg, %lo(xxx), %reg
2267 * sllx %reg, yyy, %reg
2269 * ??? This is just a generalized version of the low_bits==0
2270 * thing above, FIXME...
2272 if ((highest_bit_set - lowest_bit_set) < 32)
2274 unsigned HOST_WIDE_INT focus_bits =
2275 create_simple_focus_bits (high_bits, low_bits,
2278 /* We can't get here in this state. */
2279 gcc_assert (highest_bit_set >= 32 && lowest_bit_set < 32);
2281 /* So what we know is that the set bits straddle the
2282 middle of the 64-bit word. */
2283 sparc_emit_set_const64_quick2 (op0, temp,
2289 /* 1) sethi %hi(high_bits), %reg
2290 * or %reg, %lo(high_bits), %reg
2291 * sllx %reg, 32, %reg
2292 * or %reg, low_bits, %reg
2294 if (SPARC_SIMM13_P(low_bits)
2295 && ((int)low_bits > 0))
2297 sparc_emit_set_const64_quick2 (op0, temp, high_bits, low_bits, 32);
2301 /* The easiest way when all else fails, is full decomposition. */
2302 sparc_emit_set_const64_longway (op0, temp, high_bits, low_bits);
2304 #endif /* HOST_BITS_PER_WIDE_INT == 32 */
2306 /* Given a comparison code (EQ, NE, etc.) and the first operand of a COMPARE,
2307 return the mode to be used for the comparison. For floating-point,
2308 CCFP[E]mode is used. CC_NOOVmode should be used when the first operand
2309 is a PLUS, MINUS, NEG, or ASHIFT. CCmode should be used when no special
2310 processing is needed. */
2313 select_cc_mode (enum rtx_code op, rtx x, rtx y ATTRIBUTE_UNUSED)
2315 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2341 else if (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS
2342 || GET_CODE (x) == NEG || GET_CODE (x) == ASHIFT)
2344 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
2345 return CCX_NOOVmode;
2351 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
2358 /* Emit the compare insn and return the CC reg for a CODE comparison
2359 with operands X and Y. */
2362 gen_compare_reg_1 (enum rtx_code code, rtx x, rtx y)
2364 enum machine_mode mode;
2367 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2370 mode = SELECT_CC_MODE (code, x, y);
2372 /* ??? We don't have movcc patterns so we cannot generate pseudo regs for the
2373 fcc regs (cse can't tell they're really call clobbered regs and will
2374 remove a duplicate comparison even if there is an intervening function
2375 call - it will then try to reload the cc reg via an int reg which is why
2376 we need the movcc patterns). It is possible to provide the movcc
2377 patterns by using the ldxfsr/stxfsr v9 insns. I tried it: you need two
2378 registers (say %g1,%g5) and it takes about 6 insns. A better fix would be
2379 to tell cse that CCFPE mode registers (even pseudos) are call
2382 /* ??? This is an experiment. Rather than making changes to cse which may
2383 or may not be easy/clean, we do our own cse. This is possible because
2384 we will generate hard registers. Cse knows they're call clobbered (it
2385 doesn't know the same thing about pseudos). If we guess wrong, no big
2386 deal, but if we win, great! */
2388 if (TARGET_V9 && GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2389 #if 1 /* experiment */
2392 /* We cycle through the registers to ensure they're all exercised. */
2393 static int next_fcc_reg = 0;
2394 /* Previous x,y for each fcc reg. */
2395 static rtx prev_args[4][2];
2397 /* Scan prev_args for x,y. */
2398 for (reg = 0; reg < 4; reg++)
2399 if (prev_args[reg][0] == x && prev_args[reg][1] == y)
2404 prev_args[reg][0] = x;
2405 prev_args[reg][1] = y;
2406 next_fcc_reg = (next_fcc_reg + 1) & 3;
2408 cc_reg = gen_rtx_REG (mode, reg + SPARC_FIRST_V9_FCC_REG);
2411 cc_reg = gen_reg_rtx (mode);
2412 #endif /* ! experiment */
2413 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2414 cc_reg = gen_rtx_REG (mode, SPARC_FCC_REG);
2416 cc_reg = gen_rtx_REG (mode, SPARC_ICC_REG);
2418 /* We shouldn't get there for TFmode if !TARGET_HARD_QUAD. If we do, this
2419 will only result in an unrecognizable insn so no point in asserting. */
2420 emit_insn (gen_rtx_SET (VOIDmode, cc_reg, gen_rtx_COMPARE (mode, x, y)));
2426 /* Emit the compare insn and return the CC reg for the comparison in CMP. */
2429 gen_compare_reg (rtx cmp)
2431 return gen_compare_reg_1 (GET_CODE (cmp), XEXP (cmp, 0), XEXP (cmp, 1));
2434 /* This function is used for v9 only.
2435 DEST is the target of the Scc insn.
2436 CODE is the code for an Scc's comparison.
2437 X and Y are the values we compare.
2439 This function is needed to turn
2442 (gt (reg:CCX 100 %icc)
2446 (gt:DI (reg:CCX 100 %icc)
2449 IE: The instruction recognizer needs to see the mode of the comparison to
2450 find the right instruction. We could use "gt:DI" right in the
2451 define_expand, but leaving it out allows us to handle DI, SI, etc. */
2454 gen_v9_scc (rtx dest, enum rtx_code compare_code, rtx x, rtx y)
2457 && (GET_MODE (x) == DImode
2458 || GET_MODE (dest) == DImode))
2461 /* Try to use the movrCC insns. */
2463 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
2465 && v9_regcmp_p (compare_code))
2470 /* Special case for op0 != 0. This can be done with one instruction if
2473 if (compare_code == NE
2474 && GET_MODE (dest) == DImode
2475 && rtx_equal_p (op0, dest))
2477 emit_insn (gen_rtx_SET (VOIDmode, dest,
2478 gen_rtx_IF_THEN_ELSE (DImode,
2479 gen_rtx_fmt_ee (compare_code, DImode,
2486 if (reg_overlap_mentioned_p (dest, op0))
2488 /* Handle the case where dest == x.
2489 We "early clobber" the result. */
2490 op0 = gen_reg_rtx (GET_MODE (x));
2491 emit_move_insn (op0, x);
2494 emit_insn (gen_rtx_SET (VOIDmode, dest, const0_rtx));
2495 if (GET_MODE (op0) != DImode)
2497 temp = gen_reg_rtx (DImode);
2498 convert_move (temp, op0, 0);
2502 emit_insn (gen_rtx_SET (VOIDmode, dest,
2503 gen_rtx_IF_THEN_ELSE (GET_MODE (dest),
2504 gen_rtx_fmt_ee (compare_code, DImode,
2512 x = gen_compare_reg_1 (compare_code, x, y);
2515 gcc_assert (GET_MODE (x) != CC_NOOVmode
2516 && GET_MODE (x) != CCX_NOOVmode);
2518 emit_insn (gen_rtx_SET (VOIDmode, dest, const0_rtx));
2519 emit_insn (gen_rtx_SET (VOIDmode, dest,
2520 gen_rtx_IF_THEN_ELSE (GET_MODE (dest),
2521 gen_rtx_fmt_ee (compare_code,
2522 GET_MODE (x), x, y),
2523 const1_rtx, dest)));
2529 /* Emit an scc insn. For seq, sne, sgeu, and sltu, we can do this
2530 without jumps using the addx/subx instructions. */
2533 emit_scc_insn (rtx operands[])
2540 /* The quad-word fp compare library routines all return nonzero to indicate
2541 true, which is different from the equivalent libgcc routines, so we must
2542 handle them specially here. */
2543 if (GET_MODE (operands[2]) == TFmode && ! TARGET_HARD_QUAD)
2545 operands[1] = sparc_emit_float_lib_cmp (operands[2], operands[3],
2546 GET_CODE (operands[1]));
2547 operands[2] = XEXP (operands[1], 0);
2548 operands[3] = XEXP (operands[1], 1);
2551 code = GET_CODE (operands[1]);
2555 /* For seq/sne on v9 we use the same code as v8 (the addx/subx method has
2556 more applications). The exception to this is "reg != 0" which can
2557 be done in one instruction on v9 (so we do it). */
2560 if (GET_MODE (x) == SImode)
2562 rtx pat = gen_seqsi_special (operands[0], x, y);
2566 else if (GET_MODE (x) == DImode)
2568 rtx pat = gen_seqdi_special (operands[0], x, y);
2576 if (GET_MODE (x) == SImode)
2578 rtx pat = gen_snesi_special (operands[0], x, y);
2582 else if (GET_MODE (x) == DImode)
2586 pat = gen_snedi_special_vis3 (operands[0], x, y);
2588 pat = gen_snedi_special (operands[0], x, y);
2596 && GET_MODE (x) == DImode
2598 && (code == GTU || code == LTU))
2599 && gen_v9_scc (operands[0], code, x, y))
2602 /* We can do LTU and GEU using the addx/subx instructions too. And
2603 for GTU/LEU, if both operands are registers swap them and fall
2604 back to the easy case. */
2605 if (code == GTU || code == LEU)
2607 if ((GET_CODE (x) == REG || GET_CODE (x) == SUBREG)
2608 && (GET_CODE (y) == REG || GET_CODE (y) == SUBREG))
2613 code = swap_condition (code);
2618 || (!TARGET_VIS3 && code == GEU))
2620 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
2621 gen_rtx_fmt_ee (code, SImode,
2622 gen_compare_reg_1 (code, x, y),
2627 /* All the posibilities to use addx/subx based sequences has been
2628 exhausted, try for a 3 instruction sequence using v9 conditional
2630 if (TARGET_V9 && gen_v9_scc (operands[0], code, x, y))
2633 /* Nope, do branches. */
2637 /* Emit a conditional jump insn for the v9 architecture using comparison code
2638 CODE and jump target LABEL.
2639 This function exists to take advantage of the v9 brxx insns. */
2642 emit_v9_brxx_insn (enum rtx_code code, rtx op0, rtx label)
2644 emit_jump_insn (gen_rtx_SET (VOIDmode,
2646 gen_rtx_IF_THEN_ELSE (VOIDmode,
2647 gen_rtx_fmt_ee (code, GET_MODE (op0),
2649 gen_rtx_LABEL_REF (VOIDmode, label),
2654 emit_conditional_branch_insn (rtx operands[])
2656 /* The quad-word fp compare library routines all return nonzero to indicate
2657 true, which is different from the equivalent libgcc routines, so we must
2658 handle them specially here. */
2659 if (GET_MODE (operands[1]) == TFmode && ! TARGET_HARD_QUAD)
2661 operands[0] = sparc_emit_float_lib_cmp (operands[1], operands[2],
2662 GET_CODE (operands[0]));
2663 operands[1] = XEXP (operands[0], 0);
2664 operands[2] = XEXP (operands[0], 1);
2667 if (TARGET_ARCH64 && operands[2] == const0_rtx
2668 && GET_CODE (operands[1]) == REG
2669 && GET_MODE (operands[1]) == DImode)
2671 emit_v9_brxx_insn (GET_CODE (operands[0]), operands[1], operands[3]);
2675 operands[1] = gen_compare_reg (operands[0]);
2676 operands[2] = const0_rtx;
2677 operands[0] = gen_rtx_fmt_ee (GET_CODE (operands[0]), VOIDmode,
2678 operands[1], operands[2]);
2679 emit_jump_insn (gen_cbranchcc4 (operands[0], operands[1], operands[2],
2684 /* Generate a DFmode part of a hard TFmode register.
2685 REG is the TFmode hard register, LOW is 1 for the
2686 low 64bit of the register and 0 otherwise.
2689 gen_df_reg (rtx reg, int low)
2691 int regno = REGNO (reg);
2693 if ((WORDS_BIG_ENDIAN == 0) ^ (low != 0))
2694 regno += (TARGET_ARCH64 && SPARC_INT_REG_P (regno)) ? 1 : 2;
2695 return gen_rtx_REG (DFmode, regno);
2698 /* Generate a call to FUNC with OPERANDS. Operand 0 is the return value.
2699 Unlike normal calls, TFmode operands are passed by reference. It is
2700 assumed that no more than 3 operands are required. */
2703 emit_soft_tfmode_libcall (const char *func_name, int nargs, rtx *operands)
2705 rtx ret_slot = NULL, arg[3], func_sym;
2708 /* We only expect to be called for conversions, unary, and binary ops. */
2709 gcc_assert (nargs == 2 || nargs == 3);
2711 for (i = 0; i < nargs; ++i)
2713 rtx this_arg = operands[i];
2716 /* TFmode arguments and return values are passed by reference. */
2717 if (GET_MODE (this_arg) == TFmode)
2719 int force_stack_temp;
2721 force_stack_temp = 0;
2722 if (TARGET_BUGGY_QP_LIB && i == 0)
2723 force_stack_temp = 1;
2725 if (GET_CODE (this_arg) == MEM
2726 && ! force_stack_temp)
2728 tree expr = MEM_EXPR (this_arg);
2730 mark_addressable (expr);
2731 this_arg = XEXP (this_arg, 0);
2733 else if (CONSTANT_P (this_arg)
2734 && ! force_stack_temp)
2736 this_slot = force_const_mem (TFmode, this_arg);
2737 this_arg = XEXP (this_slot, 0);
2741 this_slot = assign_stack_temp (TFmode, GET_MODE_SIZE (TFmode), 0);
2743 /* Operand 0 is the return value. We'll copy it out later. */
2745 emit_move_insn (this_slot, this_arg);
2747 ret_slot = this_slot;
2749 this_arg = XEXP (this_slot, 0);
2756 func_sym = gen_rtx_SYMBOL_REF (Pmode, func_name);
2758 if (GET_MODE (operands[0]) == TFmode)
2761 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 2,
2762 arg[0], GET_MODE (arg[0]),
2763 arg[1], GET_MODE (arg[1]));
2765 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 3,
2766 arg[0], GET_MODE (arg[0]),
2767 arg[1], GET_MODE (arg[1]),
2768 arg[2], GET_MODE (arg[2]));
2771 emit_move_insn (operands[0], ret_slot);
2777 gcc_assert (nargs == 2);
2779 ret = emit_library_call_value (func_sym, operands[0], LCT_NORMAL,
2780 GET_MODE (operands[0]), 1,
2781 arg[1], GET_MODE (arg[1]));
2783 if (ret != operands[0])
2784 emit_move_insn (operands[0], ret);
2788 /* Expand soft-float TFmode calls to sparc abi routines. */
2791 emit_soft_tfmode_binop (enum rtx_code code, rtx *operands)
2813 emit_soft_tfmode_libcall (func, 3, operands);
2817 emit_soft_tfmode_unop (enum rtx_code code, rtx *operands)
2821 gcc_assert (code == SQRT);
2824 emit_soft_tfmode_libcall (func, 2, operands);
2828 emit_soft_tfmode_cvt (enum rtx_code code, rtx *operands)
2835 switch (GET_MODE (operands[1]))
2848 case FLOAT_TRUNCATE:
2849 switch (GET_MODE (operands[0]))
2863 switch (GET_MODE (operands[1]))
2868 operands[1] = gen_rtx_SIGN_EXTEND (DImode, operands[1]);
2878 case UNSIGNED_FLOAT:
2879 switch (GET_MODE (operands[1]))
2884 operands[1] = gen_rtx_ZERO_EXTEND (DImode, operands[1]);
2895 switch (GET_MODE (operands[0]))
2909 switch (GET_MODE (operands[0]))
2926 emit_soft_tfmode_libcall (func, 2, operands);
2929 /* Expand a hard-float tfmode operation. All arguments must be in
2933 emit_hard_tfmode_operation (enum rtx_code code, rtx *operands)
2937 if (GET_RTX_CLASS (code) == RTX_UNARY)
2939 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2940 op = gen_rtx_fmt_e (code, GET_MODE (operands[0]), operands[1]);
2944 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2945 operands[2] = force_reg (GET_MODE (operands[2]), operands[2]);
2946 op = gen_rtx_fmt_ee (code, GET_MODE (operands[0]),
2947 operands[1], operands[2]);
2950 if (register_operand (operands[0], VOIDmode))
2953 dest = gen_reg_rtx (GET_MODE (operands[0]));
2955 emit_insn (gen_rtx_SET (VOIDmode, dest, op));
2957 if (dest != operands[0])
2958 emit_move_insn (operands[0], dest);
2962 emit_tfmode_binop (enum rtx_code code, rtx *operands)
2964 if (TARGET_HARD_QUAD)
2965 emit_hard_tfmode_operation (code, operands);
2967 emit_soft_tfmode_binop (code, operands);
2971 emit_tfmode_unop (enum rtx_code code, rtx *operands)
2973 if (TARGET_HARD_QUAD)
2974 emit_hard_tfmode_operation (code, operands);
2976 emit_soft_tfmode_unop (code, operands);
2980 emit_tfmode_cvt (enum rtx_code code, rtx *operands)
2982 if (TARGET_HARD_QUAD)
2983 emit_hard_tfmode_operation (code, operands);
2985 emit_soft_tfmode_cvt (code, operands);
2988 /* Return nonzero if a branch/jump/call instruction will be emitting
2989 nop into its delay slot. */
2992 empty_delay_slot (rtx insn)
2996 /* If no previous instruction (should not happen), return true. */
2997 if (PREV_INSN (insn) == NULL)
3000 seq = NEXT_INSN (PREV_INSN (insn));
3001 if (GET_CODE (PATTERN (seq)) == SEQUENCE)
3007 /* Return nonzero if TRIAL can go into the call delay slot. */
3010 tls_call_delay (rtx trial)
3015 call __tls_get_addr, %tgd_call (foo)
3016 add %l7, %o0, %o0, %tgd_add (foo)
3017 while Sun as/ld does not. */
3018 if (TARGET_GNU_TLS || !TARGET_TLS)
3021 pat = PATTERN (trial);
3023 /* We must reject tgd_add{32|64}, i.e.
3024 (set (reg) (plus (reg) (unspec [(reg) (symbol_ref)] UNSPEC_TLSGD)))
3025 and tldm_add{32|64}, i.e.
3026 (set (reg) (plus (reg) (unspec [(reg) (symbol_ref)] UNSPEC_TLSLDM)))
3028 if (GET_CODE (pat) == SET
3029 && GET_CODE (SET_SRC (pat)) == PLUS)
3031 rtx unspec = XEXP (SET_SRC (pat), 1);
3033 if (GET_CODE (unspec) == UNSPEC
3034 && (XINT (unspec, 1) == UNSPEC_TLSGD
3035 || XINT (unspec, 1) == UNSPEC_TLSLDM))
3042 /* Return nonzero if TRIAL, an insn, can be combined with a 'restore'
3043 instruction. RETURN_P is true if the v9 variant 'return' is to be
3044 considered in the test too.
3046 TRIAL must be a SET whose destination is a REG appropriate for the
3047 'restore' instruction or, if RETURN_P is true, for the 'return'
3051 eligible_for_restore_insn (rtx trial, bool return_p)
3053 rtx pat = PATTERN (trial);
3054 rtx src = SET_SRC (pat);
3055 bool src_is_freg = false;
3058 /* Since we now can do moves between float and integer registers when
3059 VIS3 is enabled, we have to catch this case. We can allow such
3060 moves when doing a 'return' however. */
3062 if (GET_CODE (src_reg) == SUBREG)
3063 src_reg = SUBREG_REG (src_reg);
3064 if (GET_CODE (src_reg) == REG
3065 && SPARC_FP_REG_P (REGNO (src_reg)))
3068 /* The 'restore src,%g0,dest' pattern for word mode and below. */
3069 if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
3070 && arith_operand (src, GET_MODE (src))
3074 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
3076 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (SImode);
3079 /* The 'restore src,%g0,dest' pattern for double-word mode. */
3080 else if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
3081 && arith_double_operand (src, GET_MODE (src))
3083 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
3085 /* The 'restore src,%g0,dest' pattern for float if no FPU. */
3086 else if (! TARGET_FPU && register_operand (src, SFmode))
3089 /* The 'restore src,%g0,dest' pattern for double if no FPU. */
3090 else if (! TARGET_FPU && TARGET_ARCH64 && register_operand (src, DFmode))
3093 /* If we have the 'return' instruction, anything that does not use
3094 local or output registers and can go into a delay slot wins. */
3097 && !epilogue_renumber (&pat, 1)
3098 && get_attr_in_uncond_branch_delay (trial)
3099 == IN_UNCOND_BRANCH_DELAY_TRUE)
3102 /* The 'restore src1,src2,dest' pattern for SImode. */
3103 else if (GET_CODE (src) == PLUS
3104 && register_operand (XEXP (src, 0), SImode)
3105 && arith_operand (XEXP (src, 1), SImode))
3108 /* The 'restore src1,src2,dest' pattern for DImode. */
3109 else if (GET_CODE (src) == PLUS
3110 && register_operand (XEXP (src, 0), DImode)
3111 && arith_double_operand (XEXP (src, 1), DImode))
3114 /* The 'restore src1,%lo(src2),dest' pattern. */
3115 else if (GET_CODE (src) == LO_SUM
3116 && ! TARGET_CM_MEDMID
3117 && ((register_operand (XEXP (src, 0), SImode)
3118 && immediate_operand (XEXP (src, 1), SImode))
3120 && register_operand (XEXP (src, 0), DImode)
3121 && immediate_operand (XEXP (src, 1), DImode))))
3124 /* The 'restore src,src,dest' pattern. */
3125 else if (GET_CODE (src) == ASHIFT
3126 && (register_operand (XEXP (src, 0), SImode)
3127 || register_operand (XEXP (src, 0), DImode))
3128 && XEXP (src, 1) == const1_rtx)
3134 /* Return nonzero if TRIAL can go into the function return's delay slot. */
3137 eligible_for_return_delay (rtx trial)
3142 if (GET_CODE (trial) != INSN)
3145 if (get_attr_length (trial) != 1)
3148 /* If the function uses __builtin_eh_return, the eh_return machinery
3149 occupies the delay slot. */
3150 if (crtl->calls_eh_return)
3153 /* In the case of a leaf or flat function, anything can go into the slot. */
3154 if (sparc_leaf_function_p || TARGET_FLAT)
3156 get_attr_in_uncond_branch_delay (trial) == IN_UNCOND_BRANCH_DELAY_TRUE;
3158 pat = PATTERN (trial);
3159 if (GET_CODE (pat) == PARALLEL)
3165 for (i = XVECLEN (pat, 0) - 1; i >= 0; i--)
3167 rtx expr = XVECEXP (pat, 0, i);
3168 if (GET_CODE (expr) != SET)
3170 if (GET_CODE (SET_DEST (expr)) != REG)
3172 regno = REGNO (SET_DEST (expr));
3173 if (regno >= 8 && regno < 24)
3176 return !epilogue_renumber (&pat, 1)
3177 && (get_attr_in_uncond_branch_delay (trial)
3178 == IN_UNCOND_BRANCH_DELAY_TRUE);
3181 if (GET_CODE (pat) != SET)
3184 if (GET_CODE (SET_DEST (pat)) != REG)
3187 regno = REGNO (SET_DEST (pat));
3189 /* Otherwise, only operations which can be done in tandem with
3190 a `restore' or `return' insn can go into the delay slot. */
3191 if (regno >= 8 && regno < 24)
3194 /* If this instruction sets up floating point register and we have a return
3195 instruction, it can probably go in. But restore will not work
3197 if (! SPARC_INT_REG_P (regno))
3199 && !epilogue_renumber (&pat, 1)
3200 && get_attr_in_uncond_branch_delay (trial)
3201 == IN_UNCOND_BRANCH_DELAY_TRUE);
3203 return eligible_for_restore_insn (trial, true);
3206 /* Return nonzero if TRIAL can go into the sibling call's delay slot. */
3209 eligible_for_sibcall_delay (rtx trial)
3213 if (GET_CODE (trial) != INSN || GET_CODE (PATTERN (trial)) != SET)
3216 if (get_attr_length (trial) != 1)
3219 pat = PATTERN (trial);
3221 if (sparc_leaf_function_p || TARGET_FLAT)
3223 /* If the tail call is done using the call instruction,
3224 we have to restore %o7 in the delay slot. */
3225 if (LEAF_SIBCALL_SLOT_RESERVED_P)
3228 /* %g1 is used to build the function address */
3229 if (reg_mentioned_p (gen_rtx_REG (Pmode, 1), pat))
3235 /* Otherwise, only operations which can be done in tandem with
3236 a `restore' insn can go into the delay slot. */
3237 if (GET_CODE (SET_DEST (pat)) != REG
3238 || (REGNO (SET_DEST (pat)) >= 8 && REGNO (SET_DEST (pat)) < 24)
3239 || ! SPARC_INT_REG_P (REGNO (SET_DEST (pat))))
3242 /* If it mentions %o7, it can't go in, because sibcall will clobber it
3244 if (reg_mentioned_p (gen_rtx_REG (Pmode, 15), pat))
3247 return eligible_for_restore_insn (trial, false);
3250 /* Determine if it's legal to put X into the constant pool. This
3251 is not possible if X contains the address of a symbol that is
3252 not constant (TLS) or not known at final link time (PIC). */
3255 sparc_cannot_force_const_mem (enum machine_mode mode, rtx x)
3257 switch (GET_CODE (x))
3262 /* Accept all non-symbolic constants. */
3266 /* Labels are OK iff we are non-PIC. */
3267 return flag_pic != 0;
3270 /* 'Naked' TLS symbol references are never OK,
3271 non-TLS symbols are OK iff we are non-PIC. */
3272 if (SYMBOL_REF_TLS_MODEL (x))
3275 return flag_pic != 0;
3278 return sparc_cannot_force_const_mem (mode, XEXP (x, 0));
3281 return sparc_cannot_force_const_mem (mode, XEXP (x, 0))
3282 || sparc_cannot_force_const_mem (mode, XEXP (x, 1));
3290 /* Global Offset Table support. */
3291 static GTY(()) rtx got_helper_rtx = NULL_RTX;
3292 static GTY(()) rtx global_offset_table_rtx = NULL_RTX;
3294 /* Return the SYMBOL_REF for the Global Offset Table. */
3296 static GTY(()) rtx sparc_got_symbol = NULL_RTX;
3301 if (!sparc_got_symbol)
3302 sparc_got_symbol = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_");
3304 return sparc_got_symbol;
3307 /* Ensure that we are not using patterns that are not OK with PIC. */
3317 op = recog_data.operand[i];
3318 gcc_assert (GET_CODE (op) != SYMBOL_REF
3319 && (GET_CODE (op) != CONST
3320 || (GET_CODE (XEXP (op, 0)) == MINUS
3321 && XEXP (XEXP (op, 0), 0) == sparc_got ()
3322 && GET_CODE (XEXP (XEXP (op, 0), 1)) == CONST)));
3329 /* Return true if X is an address which needs a temporary register when
3330 reloaded while generating PIC code. */
3333 pic_address_needs_scratch (rtx x)
3335 /* An address which is a symbolic plus a non SMALL_INT needs a temp reg. */
3336 if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS
3337 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
3338 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
3339 && ! SMALL_INT (XEXP (XEXP (x, 0), 1)))
3345 /* Determine if a given RTX is a valid constant. We already know this
3346 satisfies CONSTANT_P. */
3349 sparc_legitimate_constant_p (enum machine_mode mode, rtx x)
3351 switch (GET_CODE (x))
3355 if (sparc_tls_referenced_p (x))
3360 if (GET_MODE (x) == VOIDmode)
3363 /* Floating point constants are generally not ok.
3364 The only exception is 0.0 and all-ones in VIS. */
3366 && SCALAR_FLOAT_MODE_P (mode)
3367 && (const_zero_operand (x, mode)
3368 || const_all_ones_operand (x, mode)))
3374 /* Vector constants are generally not ok.
3375 The only exception is 0 or -1 in VIS. */
3377 && (const_zero_operand (x, mode)
3378 || const_all_ones_operand (x, mode)))
3390 /* Determine if a given RTX is a valid constant address. */
3393 constant_address_p (rtx x)
3395 switch (GET_CODE (x))
3403 if (flag_pic && pic_address_needs_scratch (x))
3405 return sparc_legitimate_constant_p (Pmode, x);
3408 return !flag_pic && sparc_legitimate_constant_p (Pmode, x);
3415 /* Nonzero if the constant value X is a legitimate general operand
3416 when generating PIC code. It is given that flag_pic is on and
3417 that X satisfies CONSTANT_P or is a CONST_DOUBLE. */
3420 legitimate_pic_operand_p (rtx x)
3422 if (pic_address_needs_scratch (x))
3424 if (sparc_tls_referenced_p (x))
3429 #define RTX_OK_FOR_OFFSET_P(X, MODE) \
3431 && INTVAL (X) >= -0x1000 \
3432 && INTVAL (X) < (0x1000 - GET_MODE_SIZE (MODE)))
3434 #define RTX_OK_FOR_OLO10_P(X, MODE) \
3436 && INTVAL (X) >= -0x1000 \
3437 && INTVAL (X) < (0xc00 - GET_MODE_SIZE (MODE)))
3439 /* Handle the TARGET_LEGITIMATE_ADDRESS_P target hook.
3441 On SPARC, the actual legitimate addresses must be REG+REG or REG+SMALLINT
3442 ordinarily. This changes a bit when generating PIC. */
3445 sparc_legitimate_address_p (enum machine_mode mode, rtx addr, bool strict)
3447 rtx rs1 = NULL, rs2 = NULL, imm1 = NULL;
3449 if (REG_P (addr) || GET_CODE (addr) == SUBREG)
3451 else if (GET_CODE (addr) == PLUS)
3453 rs1 = XEXP (addr, 0);
3454 rs2 = XEXP (addr, 1);
3456 /* Canonicalize. REG comes first, if there are no regs,
3457 LO_SUM comes first. */
3459 && GET_CODE (rs1) != SUBREG
3461 || GET_CODE (rs2) == SUBREG
3462 || (GET_CODE (rs2) == LO_SUM && GET_CODE (rs1) != LO_SUM)))
3464 rs1 = XEXP (addr, 1);
3465 rs2 = XEXP (addr, 0);
3469 && rs1 == pic_offset_table_rtx
3471 && GET_CODE (rs2) != SUBREG
3472 && GET_CODE (rs2) != LO_SUM
3473 && GET_CODE (rs2) != MEM
3474 && !(GET_CODE (rs2) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs2))
3475 && (! symbolic_operand (rs2, VOIDmode) || mode == Pmode)
3476 && (GET_CODE (rs2) != CONST_INT || SMALL_INT (rs2)))
3478 || GET_CODE (rs1) == SUBREG)
3479 && RTX_OK_FOR_OFFSET_P (rs2, mode)))
3484 else if ((REG_P (rs1) || GET_CODE (rs1) == SUBREG)
3485 && (REG_P (rs2) || GET_CODE (rs2) == SUBREG))
3487 /* We prohibit REG + REG for TFmode when there are no quad move insns
3488 and we consequently need to split. We do this because REG+REG
3489 is not an offsettable address. If we get the situation in reload
3490 where source and destination of a movtf pattern are both MEMs with
3491 REG+REG address, then only one of them gets converted to an
3492 offsettable address. */
3494 && ! (TARGET_ARCH64 && TARGET_HARD_QUAD))
3497 /* We prohibit REG + REG on ARCH32 if not optimizing for
3498 DFmode/DImode because then mem_min_alignment is likely to be zero
3499 after reload and the forced split would lack a matching splitter
3501 if (TARGET_ARCH32 && !optimize
3502 && (mode == DFmode || mode == DImode))
3505 else if (USE_AS_OFFSETABLE_LO10
3506 && GET_CODE (rs1) == LO_SUM
3508 && ! TARGET_CM_MEDMID
3509 && RTX_OK_FOR_OLO10_P (rs2, mode))
3512 imm1 = XEXP (rs1, 1);
3513 rs1 = XEXP (rs1, 0);
3514 if (!CONSTANT_P (imm1)
3515 || (GET_CODE (rs1) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs1)))
3519 else if (GET_CODE (addr) == LO_SUM)
3521 rs1 = XEXP (addr, 0);
3522 imm1 = XEXP (addr, 1);
3524 if (!CONSTANT_P (imm1)
3525 || (GET_CODE (rs1) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs1)))
3528 /* We can't allow TFmode in 32-bit mode, because an offset greater
3529 than the alignment (8) may cause the LO_SUM to overflow. */
3530 if (mode == TFmode && TARGET_ARCH32)
3533 else if (GET_CODE (addr) == CONST_INT && SMALL_INT (addr))
3538 if (GET_CODE (rs1) == SUBREG)
3539 rs1 = SUBREG_REG (rs1);
3545 if (GET_CODE (rs2) == SUBREG)
3546 rs2 = SUBREG_REG (rs2);
3553 if (!REGNO_OK_FOR_BASE_P (REGNO (rs1))
3554 || (rs2 && !REGNO_OK_FOR_BASE_P (REGNO (rs2))))
3559 if ((! SPARC_INT_REG_P (REGNO (rs1))
3560 && REGNO (rs1) != FRAME_POINTER_REGNUM
3561 && REGNO (rs1) < FIRST_PSEUDO_REGISTER)
3563 && (! SPARC_INT_REG_P (REGNO (rs2))
3564 && REGNO (rs2) != FRAME_POINTER_REGNUM
3565 && REGNO (rs2) < FIRST_PSEUDO_REGISTER)))
3571 /* Return the SYMBOL_REF for the tls_get_addr function. */
3573 static GTY(()) rtx sparc_tls_symbol = NULL_RTX;
3576 sparc_tls_get_addr (void)
3578 if (!sparc_tls_symbol)
3579 sparc_tls_symbol = gen_rtx_SYMBOL_REF (Pmode, "__tls_get_addr");
3581 return sparc_tls_symbol;
3584 /* Return the Global Offset Table to be used in TLS mode. */
3587 sparc_tls_got (void)
3589 /* In PIC mode, this is just the PIC offset table. */
3592 crtl->uses_pic_offset_table = 1;
3593 return pic_offset_table_rtx;
3596 /* In non-PIC mode, Sun as (unlike GNU as) emits PC-relative relocations for
3597 the GOT symbol with the 32-bit ABI, so we reload the GOT register. */
3598 if (TARGET_SUN_TLS && TARGET_ARCH32)
3600 load_got_register ();
3601 return global_offset_table_rtx;
3604 /* In all other cases, we load a new pseudo with the GOT symbol. */
3605 return copy_to_reg (sparc_got ());
3608 /* Return true if X contains a thread-local symbol. */
3611 sparc_tls_referenced_p (rtx x)
3613 if (!TARGET_HAVE_TLS)
3616 if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS)
3617 x = XEXP (XEXP (x, 0), 0);
3619 if (GET_CODE (x) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (x))
3622 /* That's all we handle in sparc_legitimize_tls_address for now. */
3626 /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute
3627 this (thread-local) address. */
3630 sparc_legitimize_tls_address (rtx addr)
3632 rtx temp1, temp2, temp3, ret, o0, got, insn;
3634 gcc_assert (can_create_pseudo_p ());
3636 if (GET_CODE (addr) == SYMBOL_REF)
3637 switch (SYMBOL_REF_TLS_MODEL (addr))
3639 case TLS_MODEL_GLOBAL_DYNAMIC:
3641 temp1 = gen_reg_rtx (SImode);
3642 temp2 = gen_reg_rtx (SImode);
3643 ret = gen_reg_rtx (Pmode);
3644 o0 = gen_rtx_REG (Pmode, 8);
3645 got = sparc_tls_got ();
3646 emit_insn (gen_tgd_hi22 (temp1, addr));
3647 emit_insn (gen_tgd_lo10 (temp2, temp1, addr));
3650 emit_insn (gen_tgd_add32 (o0, got, temp2, addr));
3651 insn = emit_call_insn (gen_tgd_call32 (o0, sparc_tls_get_addr (),
3656 emit_insn (gen_tgd_add64 (o0, got, temp2, addr));
3657 insn = emit_call_insn (gen_tgd_call64 (o0, sparc_tls_get_addr (),
3660 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), o0);
3661 insn = get_insns ();
3663 emit_libcall_block (insn, ret, o0, addr);
3666 case TLS_MODEL_LOCAL_DYNAMIC:
3668 temp1 = gen_reg_rtx (SImode);
3669 temp2 = gen_reg_rtx (SImode);
3670 temp3 = gen_reg_rtx (Pmode);
3671 ret = gen_reg_rtx (Pmode);
3672 o0 = gen_rtx_REG (Pmode, 8);
3673 got = sparc_tls_got ();
3674 emit_insn (gen_tldm_hi22 (temp1));
3675 emit_insn (gen_tldm_lo10 (temp2, temp1));
3678 emit_insn (gen_tldm_add32 (o0, got, temp2));
3679 insn = emit_call_insn (gen_tldm_call32 (o0, sparc_tls_get_addr (),
3684 emit_insn (gen_tldm_add64 (o0, got, temp2));
3685 insn = emit_call_insn (gen_tldm_call64 (o0, sparc_tls_get_addr (),
3688 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), o0);
3689 insn = get_insns ();
3691 emit_libcall_block (insn, temp3, o0,
3692 gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
3693 UNSPEC_TLSLD_BASE));
3694 temp1 = gen_reg_rtx (SImode);
3695 temp2 = gen_reg_rtx (SImode);
3696 emit_insn (gen_tldo_hix22 (temp1, addr));
3697 emit_insn (gen_tldo_lox10 (temp2, temp1, addr));
3699 emit_insn (gen_tldo_add32 (ret, temp3, temp2, addr));
3701 emit_insn (gen_tldo_add64 (ret, temp3, temp2, addr));
3704 case TLS_MODEL_INITIAL_EXEC:
3705 temp1 = gen_reg_rtx (SImode);
3706 temp2 = gen_reg_rtx (SImode);
3707 temp3 = gen_reg_rtx (Pmode);
3708 got = sparc_tls_got ();
3709 emit_insn (gen_tie_hi22 (temp1, addr));
3710 emit_insn (gen_tie_lo10 (temp2, temp1, addr));
3712 emit_insn (gen_tie_ld32 (temp3, got, temp2, addr));
3714 emit_insn (gen_tie_ld64 (temp3, got, temp2, addr));
3717 ret = gen_reg_rtx (Pmode);
3719 emit_insn (gen_tie_add32 (ret, gen_rtx_REG (Pmode, 7),
3722 emit_insn (gen_tie_add64 (ret, gen_rtx_REG (Pmode, 7),
3726 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp3);
3729 case TLS_MODEL_LOCAL_EXEC:
3730 temp1 = gen_reg_rtx (Pmode);
3731 temp2 = gen_reg_rtx (Pmode);
3734 emit_insn (gen_tle_hix22_sp32 (temp1, addr));
3735 emit_insn (gen_tle_lox10_sp32 (temp2, temp1, addr));
3739 emit_insn (gen_tle_hix22_sp64 (temp1, addr));
3740 emit_insn (gen_tle_lox10_sp64 (temp2, temp1, addr));
3742 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp2);
3749 else if (GET_CODE (addr) == CONST)
3753 gcc_assert (GET_CODE (XEXP (addr, 0)) == PLUS);
3755 base = sparc_legitimize_tls_address (XEXP (XEXP (addr, 0), 0));
3756 offset = XEXP (XEXP (addr, 0), 1);
3758 base = force_operand (base, NULL_RTX);
3759 if (!(GET_CODE (offset) == CONST_INT && SMALL_INT (offset)))
3760 offset = force_reg (Pmode, offset);
3761 ret = gen_rtx_PLUS (Pmode, base, offset);
3765 gcc_unreachable (); /* for now ... */
3770 /* Legitimize PIC addresses. If the address is already position-independent,
3771 we return ORIG. Newly generated position-independent addresses go into a
3772 reg. This is REG if nonzero, otherwise we allocate register(s) as
3776 sparc_legitimize_pic_address (rtx orig, rtx reg)
3778 bool gotdata_op = false;
3780 if (GET_CODE (orig) == SYMBOL_REF
3781 /* See the comment in sparc_expand_move. */
3782 || (GET_CODE (orig) == LABEL_REF && !can_use_mov_pic_label_ref (orig)))
3784 rtx pic_ref, address;
3789 gcc_assert (can_create_pseudo_p ());
3790 reg = gen_reg_rtx (Pmode);
3795 /* If not during reload, allocate another temp reg here for loading
3796 in the address, so that these instructions can be optimized
3798 rtx temp_reg = (! can_create_pseudo_p ()
3799 ? reg : gen_reg_rtx (Pmode));
3801 /* Must put the SYMBOL_REF inside an UNSPEC here so that cse
3802 won't get confused into thinking that these two instructions
3803 are loading in the true address of the symbol. If in the
3804 future a PIC rtx exists, that should be used instead. */
3807 emit_insn (gen_movdi_high_pic (temp_reg, orig));
3808 emit_insn (gen_movdi_lo_sum_pic (temp_reg, temp_reg, orig));
3812 emit_insn (gen_movsi_high_pic (temp_reg, orig));
3813 emit_insn (gen_movsi_lo_sum_pic (temp_reg, temp_reg, orig));
3821 crtl->uses_pic_offset_table = 1;
3825 insn = emit_insn (gen_movdi_pic_gotdata_op (reg,
3826 pic_offset_table_rtx,
3829 insn = emit_insn (gen_movsi_pic_gotdata_op (reg,
3830 pic_offset_table_rtx,
3836 = gen_const_mem (Pmode,
3837 gen_rtx_PLUS (Pmode,
3838 pic_offset_table_rtx, address));
3839 insn = emit_move_insn (reg, pic_ref);
3842 /* Put a REG_EQUAL note on this insn, so that it can be optimized
3844 set_unique_reg_note (insn, REG_EQUAL, orig);
3847 else if (GET_CODE (orig) == CONST)
3851 if (GET_CODE (XEXP (orig, 0)) == PLUS
3852 && XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx)
3857 gcc_assert (can_create_pseudo_p ());
3858 reg = gen_reg_rtx (Pmode);
3861 gcc_assert (GET_CODE (XEXP (orig, 0)) == PLUS);
3862 base = sparc_legitimize_pic_address (XEXP (XEXP (orig, 0), 0), reg);
3863 offset = sparc_legitimize_pic_address (XEXP (XEXP (orig, 0), 1),
3864 base == reg ? NULL_RTX : reg);
3866 if (GET_CODE (offset) == CONST_INT)
3868 if (SMALL_INT (offset))
3869 return plus_constant (Pmode, base, INTVAL (offset));
3870 else if (can_create_pseudo_p ())
3871 offset = force_reg (Pmode, offset);
3873 /* If we reach here, then something is seriously wrong. */
3876 return gen_rtx_PLUS (Pmode, base, offset);
3878 else if (GET_CODE (orig) == LABEL_REF)
3879 /* ??? We ought to be checking that the register is live instead, in case
3880 it is eliminated. */
3881 crtl->uses_pic_offset_table = 1;
3886 /* Try machine-dependent ways of modifying an illegitimate address X
3887 to be legitimate. If we find one, return the new, valid address.
3889 OLDX is the address as it was before break_out_memory_refs was called.
3890 In some cases it is useful to look at this to decide what needs to be done.
3892 MODE is the mode of the operand pointed to by X.
3894 On SPARC, change REG+N into REG+REG, and REG+(X*Y) into REG+REG. */
3897 sparc_legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED,
3898 enum machine_mode mode)
3902 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == MULT)
3903 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3904 force_operand (XEXP (x, 0), NULL_RTX));
3905 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == MULT)
3906 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3907 force_operand (XEXP (x, 1), NULL_RTX));
3908 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == PLUS)
3909 x = gen_rtx_PLUS (Pmode, force_operand (XEXP (x, 0), NULL_RTX),
3911 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == PLUS)
3912 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3913 force_operand (XEXP (x, 1), NULL_RTX));
3915 if (x != orig_x && sparc_legitimate_address_p (mode, x, FALSE))
3918 if (sparc_tls_referenced_p (x))
3919 x = sparc_legitimize_tls_address (x);
3921 x = sparc_legitimize_pic_address (x, NULL_RTX);
3922 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 1)))
3923 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3924 copy_to_mode_reg (Pmode, XEXP (x, 1)));
3925 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 0)))
3926 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3927 copy_to_mode_reg (Pmode, XEXP (x, 0)));
3928 else if (GET_CODE (x) == SYMBOL_REF
3929 || GET_CODE (x) == CONST
3930 || GET_CODE (x) == LABEL_REF)
3931 x = copy_to_suggested_reg (x, NULL_RTX, Pmode);
3936 /* Delegitimize an address that was legitimized by the above function. */
3939 sparc_delegitimize_address (rtx x)
3941 x = delegitimize_mem_from_attrs (x);
3943 if (GET_CODE (x) == LO_SUM && GET_CODE (XEXP (x, 1)) == UNSPEC)
3944 switch (XINT (XEXP (x, 1), 1))
3946 case UNSPEC_MOVE_PIC:
3948 x = XVECEXP (XEXP (x, 1), 0, 0);
3949 gcc_assert (GET_CODE (x) == SYMBOL_REF);
3955 /* This is generated by mov{si,di}_pic_label_ref in PIC mode. */
3956 if (GET_CODE (x) == MINUS
3957 && REG_P (XEXP (x, 0))
3958 && REGNO (XEXP (x, 0)) == PIC_OFFSET_TABLE_REGNUM
3959 && GET_CODE (XEXP (x, 1)) == LO_SUM
3960 && GET_CODE (XEXP (XEXP (x, 1), 1)) == UNSPEC
3961 && XINT (XEXP (XEXP (x, 1), 1), 1) == UNSPEC_MOVE_PIC_LABEL)
3963 x = XVECEXP (XEXP (XEXP (x, 1), 1), 0, 0);
3964 gcc_assert (GET_CODE (x) == LABEL_REF);
3970 /* SPARC implementation of LEGITIMIZE_RELOAD_ADDRESS. Returns a value to
3971 replace the input X, or the original X if no replacement is called for.
3972 The output parameter *WIN is 1 if the calling macro should goto WIN,
3975 For SPARC, we wish to handle addresses by splitting them into
3976 HIGH+LO_SUM pairs, retaining the LO_SUM in the memory reference.
3977 This cuts the number of extra insns by one.
3979 Do nothing when generating PIC code and the address is a symbolic
3980 operand or requires a scratch register. */
3983 sparc_legitimize_reload_address (rtx x, enum machine_mode mode,
3984 int opnum, int type,
3985 int ind_levels ATTRIBUTE_UNUSED, int *win)
3987 /* Decompose SImode constants into HIGH+LO_SUM. */
3989 && (mode != TFmode || TARGET_ARCH64)
3990 && GET_MODE (x) == SImode
3991 && GET_CODE (x) != LO_SUM
3992 && GET_CODE (x) != HIGH
3993 && sparc_cmodel <= CM_MEDLOW
3995 && (symbolic_operand (x, Pmode) || pic_address_needs_scratch (x))))
3997 x = gen_rtx_LO_SUM (GET_MODE (x), gen_rtx_HIGH (GET_MODE (x), x), x);
3998 push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL,
3999 BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0,
4000 opnum, (enum reload_type)type);
4005 /* We have to recognize what we have already generated above. */
4006 if (GET_CODE (x) == LO_SUM && GET_CODE (XEXP (x, 0)) == HIGH)
4008 push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL,
4009 BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0,
4010 opnum, (enum reload_type)type);
4019 /* Return true if ADDR (a legitimate address expression)
4020 has an effect that depends on the machine mode it is used for.
4026 is not equivalent to
4028 (mem:QI [%l7+a]) (mem:QI [%l7+a+1])
4030 because [%l7+a+1] is interpreted as the address of (a+1). */
4034 sparc_mode_dependent_address_p (const_rtx addr)
4036 if (flag_pic && GET_CODE (addr) == PLUS)
4038 rtx op0 = XEXP (addr, 0);
4039 rtx op1 = XEXP (addr, 1);
4040 if (op0 == pic_offset_table_rtx
4041 && symbolic_operand (op1, VOIDmode))
4048 #ifdef HAVE_GAS_HIDDEN
4049 # define USE_HIDDEN_LINKONCE 1
4051 # define USE_HIDDEN_LINKONCE 0
4055 get_pc_thunk_name (char name[32], unsigned int regno)
4057 const char *reg_name = reg_names[regno];
4059 /* Skip the leading '%' as that cannot be used in a
4063 if (USE_HIDDEN_LINKONCE)
4064 sprintf (name, "__sparc_get_pc_thunk.%s", reg_name);
4066 ASM_GENERATE_INTERNAL_LABEL (name, "LADDPC", regno);
4069 /* Wrapper around the load_pcrel_sym{si,di} patterns. */
4072 gen_load_pcrel_sym (rtx op0, rtx op1, rtx op2, rtx op3)
4074 int orig_flag_pic = flag_pic;
4077 /* The load_pcrel_sym{si,di} patterns require absolute addressing. */
4080 insn = gen_load_pcrel_symdi (op0, op1, op2, op3);
4082 insn = gen_load_pcrel_symsi (op0, op1, op2, op3);
4083 flag_pic = orig_flag_pic;
4088 /* Emit code to load the GOT register. */
4091 load_got_register (void)
4093 /* In PIC mode, this will retrieve pic_offset_table_rtx. */
4094 if (!global_offset_table_rtx)
4095 global_offset_table_rtx = gen_rtx_REG (Pmode, GLOBAL_OFFSET_TABLE_REGNUM);
4097 if (TARGET_VXWORKS_RTP)
4098 emit_insn (gen_vxworks_load_got ());
4101 /* The GOT symbol is subject to a PC-relative relocation so we need a
4102 helper function to add the PC value and thus get the final value. */
4103 if (!got_helper_rtx)
4106 get_pc_thunk_name (name, GLOBAL_OFFSET_TABLE_REGNUM);
4107 got_helper_rtx = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (name));
4110 emit_insn (gen_load_pcrel_sym (global_offset_table_rtx, sparc_got (),
4112 GEN_INT (GLOBAL_OFFSET_TABLE_REGNUM)));
4115 /* Need to emit this whether or not we obey regdecls,
4116 since setjmp/longjmp can cause life info to screw up.
4117 ??? In the case where we don't obey regdecls, this is not sufficient
4118 since we may not fall out the bottom. */
4119 emit_use (global_offset_table_rtx);
4122 /* Emit a call instruction with the pattern given by PAT. ADDR is the
4123 address of the call target. */
4126 sparc_emit_call_insn (rtx pat, rtx addr)
4130 insn = emit_call_insn (pat);
4132 /* The PIC register is live on entry to VxWorks PIC PLT entries. */
4133 if (TARGET_VXWORKS_RTP
4135 && GET_CODE (addr) == SYMBOL_REF
4136 && (SYMBOL_REF_DECL (addr)
4137 ? !targetm.binds_local_p (SYMBOL_REF_DECL (addr))
4138 : !SYMBOL_REF_LOCAL_P (addr)))
4140 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), pic_offset_table_rtx);
4141 crtl->uses_pic_offset_table = 1;
4145 /* Return 1 if RTX is a MEM which is known to be aligned to at
4146 least a DESIRED byte boundary. */
4149 mem_min_alignment (rtx mem, int desired)
4151 rtx addr, base, offset;
4153 /* If it's not a MEM we can't accept it. */
4154 if (GET_CODE (mem) != MEM)
4158 if (!TARGET_UNALIGNED_DOUBLES
4159 && MEM_ALIGN (mem) / BITS_PER_UNIT >= (unsigned)desired)
4162 /* ??? The rest of the function predates MEM_ALIGN so
4163 there is probably a bit of redundancy. */
4164 addr = XEXP (mem, 0);
4165 base = offset = NULL_RTX;
4166 if (GET_CODE (addr) == PLUS)
4168 if (GET_CODE (XEXP (addr, 0)) == REG)
4170 base = XEXP (addr, 0);
4172 /* What we are saying here is that if the base
4173 REG is aligned properly, the compiler will make
4174 sure any REG based index upon it will be so
4176 if (GET_CODE (XEXP (addr, 1)) == CONST_INT)
4177 offset = XEXP (addr, 1);
4179 offset = const0_rtx;
4182 else if (GET_CODE (addr) == REG)
4185 offset = const0_rtx;
4188 if (base != NULL_RTX)
4190 int regno = REGNO (base);
4192 if (regno != HARD_FRAME_POINTER_REGNUM && regno != STACK_POINTER_REGNUM)
4194 /* Check if the compiler has recorded some information
4195 about the alignment of the base REG. If reload has
4196 completed, we already matched with proper alignments.
4197 If not running global_alloc, reload might give us
4198 unaligned pointer to local stack though. */
4200 && REGNO_POINTER_ALIGN (regno) >= desired * BITS_PER_UNIT)
4201 || (optimize && reload_completed))
4202 && (INTVAL (offset) & (desired - 1)) == 0)
4207 if (((INTVAL (offset) - SPARC_STACK_BIAS) & (desired - 1)) == 0)
4211 else if (! TARGET_UNALIGNED_DOUBLES
4212 || CONSTANT_P (addr)
4213 || GET_CODE (addr) == LO_SUM)
4215 /* Anything else we know is properly aligned unless TARGET_UNALIGNED_DOUBLES
4216 is true, in which case we can only assume that an access is aligned if
4217 it is to a constant address, or the address involves a LO_SUM. */
4221 /* An obviously unaligned address. */
4226 /* Vectors to keep interesting information about registers where it can easily
4227 be got. We used to use the actual mode value as the bit number, but there
4228 are more than 32 modes now. Instead we use two tables: one indexed by
4229 hard register number, and one indexed by mode. */
4231 /* The purpose of sparc_mode_class is to shrink the range of modes so that
4232 they all fit (as bit numbers) in a 32-bit word (again). Each real mode is
4233 mapped into one sparc_mode_class mode. */
4235 enum sparc_mode_class {
4236 S_MODE, D_MODE, T_MODE, O_MODE,
4237 SF_MODE, DF_MODE, TF_MODE, OF_MODE,
4241 /* Modes for single-word and smaller quantities. */
4242 #define S_MODES ((1 << (int) S_MODE) | (1 << (int) SF_MODE))
4244 /* Modes for double-word and smaller quantities. */
4245 #define D_MODES (S_MODES | (1 << (int) D_MODE) | (1 << DF_MODE))
4247 /* Modes for quad-word and smaller quantities. */
4248 #define T_MODES (D_MODES | (1 << (int) T_MODE) | (1 << (int) TF_MODE))
4250 /* Modes for 8-word and smaller quantities. */
4251 #define O_MODES (T_MODES | (1 << (int) O_MODE) | (1 << (int) OF_MODE))
4253 /* Modes for single-float quantities. We must allow any single word or
4254 smaller quantity. This is because the fix/float conversion instructions
4255 take integer inputs/outputs from the float registers. */
4256 #define SF_MODES (S_MODES)
4258 /* Modes for double-float and smaller quantities. */
4259 #define DF_MODES (D_MODES)
4261 /* Modes for quad-float and smaller quantities. */
4262 #define TF_MODES (DF_MODES | (1 << (int) TF_MODE))
4264 /* Modes for quad-float pairs and smaller quantities. */
4265 #define OF_MODES (TF_MODES | (1 << (int) OF_MODE))
4267 /* Modes for double-float only quantities. */
4268 #define DF_MODES_NO_S ((1 << (int) D_MODE) | (1 << (int) DF_MODE))
4270 /* Modes for quad-float and double-float only quantities. */
4271 #define TF_MODES_NO_S (DF_MODES_NO_S | (1 << (int) TF_MODE))
4273 /* Modes for quad-float pairs and double-float only quantities. */
4274 #define OF_MODES_NO_S (TF_MODES_NO_S | (1 << (int) OF_MODE))
4276 /* Modes for condition codes. */
4277 #define CC_MODES (1 << (int) CC_MODE)
4278 #define CCFP_MODES (1 << (int) CCFP_MODE)
4280 /* Value is 1 if register/mode pair is acceptable on sparc.
4281 The funny mixture of D and T modes is because integer operations
4282 do not specially operate on tetra quantities, so non-quad-aligned
4283 registers can hold quadword quantities (except %o4 and %i4 because
4284 they cross fixed registers). */
4286 /* This points to either the 32 bit or the 64 bit version. */
4287 const int *hard_regno_mode_classes;
4289 static const int hard_32bit_mode_classes[] = {
4290 S_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
4291 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
4292 T_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
4293 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
4295 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4296 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4297 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4298 OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
4300 /* FP regs f32 to f63. Only the even numbered registers actually exist,
4301 and none can hold SFmode/SImode values. */
4302 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4303 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4304 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4305 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4308 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
4310 /* %icc, %sfp, %gsr */
4311 CC_MODES, 0, D_MODES
4314 static const int hard_64bit_mode_classes[] = {
4315 D_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4316 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4317 T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4318 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4320 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4321 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4322 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4323 OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
4325 /* FP regs f32 to f63. Only the even numbered registers actually exist,
4326 and none can hold SFmode/SImode values. */
4327 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4328 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4329 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4330 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4333 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
4335 /* %icc, %sfp, %gsr */
4336 CC_MODES, 0, D_MODES
4339 int sparc_mode_class [NUM_MACHINE_MODES];
4341 enum reg_class sparc_regno_reg_class[FIRST_PSEUDO_REGISTER];
4344 sparc_init_modes (void)
4348 for (i = 0; i < NUM_MACHINE_MODES; i++)
4350 switch (GET_MODE_CLASS (i))
4353 case MODE_PARTIAL_INT:
4354 case MODE_COMPLEX_INT:
4355 if (GET_MODE_SIZE (i) <= 4)
4356 sparc_mode_class[i] = 1 << (int) S_MODE;
4357 else if (GET_MODE_SIZE (i) == 8)
4358 sparc_mode_class[i] = 1 << (int) D_MODE;
4359 else if (GET_MODE_SIZE (i) == 16)
4360 sparc_mode_class[i] = 1 << (int) T_MODE;
4361 else if (GET_MODE_SIZE (i) == 32)
4362 sparc_mode_class[i] = 1 << (int) O_MODE;
4364 sparc_mode_class[i] = 0;
4366 case MODE_VECTOR_INT:
4367 if (GET_MODE_SIZE (i) <= 4)
4368 sparc_mode_class[i] = 1 << (int)SF_MODE;
4369 else if (GET_MODE_SIZE (i) == 8)
4370 sparc_mode_class[i] = 1 << (int)DF_MODE;
4373 case MODE_COMPLEX_FLOAT:
4374 if (GET_MODE_SIZE (i) <= 4)
4375 sparc_mode_class[i] = 1 << (int) SF_MODE;
4376 else if (GET_MODE_SIZE (i) == 8)
4377 sparc_mode_class[i] = 1 << (int) DF_MODE;
4378 else if (GET_MODE_SIZE (i) == 16)
4379 sparc_mode_class[i] = 1 << (int) TF_MODE;
4380 else if (GET_MODE_SIZE (i) == 32)
4381 sparc_mode_class[i] = 1 << (int) OF_MODE;
4383 sparc_mode_class[i] = 0;
4386 if (i == (int) CCFPmode || i == (int) CCFPEmode)
4387 sparc_mode_class[i] = 1 << (int) CCFP_MODE;
4389 sparc_mode_class[i] = 1 << (int) CC_MODE;
4392 sparc_mode_class[i] = 0;
4398 hard_regno_mode_classes = hard_64bit_mode_classes;
4400 hard_regno_mode_classes = hard_32bit_mode_classes;
4402 /* Initialize the array used by REGNO_REG_CLASS. */
4403 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
4405 if (i < 16 && TARGET_V8PLUS)
4406 sparc_regno_reg_class[i] = I64_REGS;
4407 else if (i < 32 || i == FRAME_POINTER_REGNUM)
4408 sparc_regno_reg_class[i] = GENERAL_REGS;
4410 sparc_regno_reg_class[i] = FP_REGS;
4412 sparc_regno_reg_class[i] = EXTRA_FP_REGS;
4414 sparc_regno_reg_class[i] = FPCC_REGS;
4416 sparc_regno_reg_class[i] = NO_REGS;
4420 /* Return whether REGNO, a global or FP register, must be saved/restored. */
4423 save_global_or_fp_reg_p (unsigned int regno,
4424 int leaf_function ATTRIBUTE_UNUSED)
4426 return !call_used_regs[regno] && df_regs_ever_live_p (regno);
4429 /* Return whether the return address register (%i7) is needed. */
4432 return_addr_reg_needed_p (int leaf_function)
4434 /* If it is live, for example because of __builtin_return_address (0). */
4435 if (df_regs_ever_live_p (RETURN_ADDR_REGNUM))
4438 /* Otherwise, it is needed as save register if %o7 is clobbered. */
4440 /* Loading the GOT register clobbers %o7. */
4441 || crtl->uses_pic_offset_table
4442 || df_regs_ever_live_p (INCOMING_RETURN_ADDR_REGNUM))
4448 /* Return whether REGNO, a local or in register, must be saved/restored. */
4451 save_local_or_in_reg_p (unsigned int regno, int leaf_function)
4453 /* General case: call-saved registers live at some point. */
4454 if (!call_used_regs[regno] && df_regs_ever_live_p (regno))
4457 /* Frame pointer register (%fp) if needed. */
4458 if (regno == HARD_FRAME_POINTER_REGNUM && frame_pointer_needed)
4461 /* Return address register (%i7) if needed. */
4462 if (regno == RETURN_ADDR_REGNUM && return_addr_reg_needed_p (leaf_function))
4465 /* GOT register (%l7) if needed. */
4466 if (regno == PIC_OFFSET_TABLE_REGNUM && crtl->uses_pic_offset_table)
4469 /* If the function accesses prior frames, the frame pointer and the return
4470 address of the previous frame must be saved on the stack. */
4471 if (crtl->accesses_prior_frames
4472 && (regno == HARD_FRAME_POINTER_REGNUM || regno == RETURN_ADDR_REGNUM))
4478 /* Compute the frame size required by the function. This function is called
4479 during the reload pass and also by sparc_expand_prologue. */
4482 sparc_compute_frame_size (HOST_WIDE_INT size, int leaf_function)
4484 HOST_WIDE_INT frame_size, apparent_frame_size;
4485 int args_size, n_global_fp_regs = 0;
4486 bool save_local_in_regs_p = false;
4489 /* If the function allocates dynamic stack space, the dynamic offset is
4490 computed early and contains REG_PARM_STACK_SPACE, so we need to cope. */
4491 if (leaf_function && !cfun->calls_alloca)
4494 args_size = crtl->outgoing_args_size + REG_PARM_STACK_SPACE (cfun->decl);
4496 /* Calculate space needed for global registers. */
4498 for (i = 0; i < 8; i++)
4499 if (save_global_or_fp_reg_p (i, 0))
4500 n_global_fp_regs += 2;
4502 for (i = 0; i < 8; i += 2)
4503 if (save_global_or_fp_reg_p (i, 0) || save_global_or_fp_reg_p (i + 1, 0))
4504 n_global_fp_regs += 2;
4506 /* In the flat window model, find out which local and in registers need to
4507 be saved. We don't reserve space in the current frame for them as they
4508 will be spilled into the register window save area of the caller's frame.
4509 However, as soon as we use this register window save area, we must create
4510 that of the current frame to make it the live one. */
4512 for (i = 16; i < 32; i++)
4513 if (save_local_or_in_reg_p (i, leaf_function))
4515 save_local_in_regs_p = true;
4519 /* Calculate space needed for FP registers. */
4520 for (i = 32; i < (TARGET_V9 ? 96 : 64); i += 2)
4521 if (save_global_or_fp_reg_p (i, 0) || save_global_or_fp_reg_p (i + 1, 0))
4522 n_global_fp_regs += 2;
4525 && n_global_fp_regs == 0
4527 && !save_local_in_regs_p)
4528 frame_size = apparent_frame_size = 0;
4531 /* We subtract STARTING_FRAME_OFFSET, remember it's negative. */
4532 apparent_frame_size = (size - STARTING_FRAME_OFFSET + 7) & -8;
4533 apparent_frame_size += n_global_fp_regs * 4;
4535 /* We need to add the size of the outgoing argument area. */
4536 frame_size = apparent_frame_size + ((args_size + 7) & -8);
4538 /* And that of the register window save area. */
4539 frame_size += FIRST_PARM_OFFSET (cfun->decl);
4541 /* Finally, bump to the appropriate alignment. */
4542 frame_size = SPARC_STACK_ALIGN (frame_size);
4545 /* Set up values for use in prologue and epilogue. */
4546 sparc_frame_size = frame_size;
4547 sparc_apparent_frame_size = apparent_frame_size;
4548 sparc_n_global_fp_regs = n_global_fp_regs;
4549 sparc_save_local_in_regs_p = save_local_in_regs_p;
4554 /* Output any necessary .register pseudo-ops. */
4557 sparc_output_scratch_registers (FILE *file ATTRIBUTE_UNUSED)
4559 #ifdef HAVE_AS_REGISTER_PSEUDO_OP
4565 /* Check if %g[2367] were used without
4566 .register being printed for them already. */
4567 for (i = 2; i < 8; i++)
4569 if (df_regs_ever_live_p (i)
4570 && ! sparc_hard_reg_printed [i])
4572 sparc_hard_reg_printed [i] = 1;
4573 /* %g7 is used as TLS base register, use #ignore
4574 for it instead of #scratch. */
4575 fprintf (file, "\t.register\t%%g%d, #%s\n", i,
4576 i == 7 ? "ignore" : "scratch");
4583 #define PROBE_INTERVAL (1 << STACK_CHECK_PROBE_INTERVAL_EXP)
4585 #if PROBE_INTERVAL > 4096
4586 #error Cannot use indexed addressing mode for stack probing
4589 /* Emit code to probe a range of stack addresses from FIRST to FIRST+SIZE,
4590 inclusive. These are offsets from the current stack pointer.
4592 Note that we don't use the REG+REG addressing mode for the probes because
4593 of the stack bias in 64-bit mode. And it doesn't really buy us anything
4594 so the advantages of having a single code win here. */
4597 sparc_emit_probe_stack_range (HOST_WIDE_INT first, HOST_WIDE_INT size)
4599 rtx g1 = gen_rtx_REG (Pmode, 1);
4601 /* See if we have a constant small number of probes to generate. If so,
4602 that's the easy case. */
4603 if (size <= PROBE_INTERVAL)
4605 emit_move_insn (g1, GEN_INT (first));
4606 emit_insn (gen_rtx_SET (VOIDmode, g1,
4607 gen_rtx_MINUS (Pmode, stack_pointer_rtx, g1)));
4608 emit_stack_probe (plus_constant (Pmode, g1, -size));
4611 /* The run-time loop is made up of 10 insns in the generic case while the
4612 compile-time loop is made up of 4+2*(n-2) insns for n # of intervals. */
4613 else if (size <= 5 * PROBE_INTERVAL)
4617 emit_move_insn (g1, GEN_INT (first + PROBE_INTERVAL));
4618 emit_insn (gen_rtx_SET (VOIDmode, g1,
4619 gen_rtx_MINUS (Pmode, stack_pointer_rtx, g1)));
4620 emit_stack_probe (g1);
4622 /* Probe at FIRST + N * PROBE_INTERVAL for values of N from 2 until
4623 it exceeds SIZE. If only two probes are needed, this will not
4624 generate any code. Then probe at FIRST + SIZE. */
4625 for (i = 2 * PROBE_INTERVAL; i < size; i += PROBE_INTERVAL)
4627 emit_insn (gen_rtx_SET (VOIDmode, g1,
4628 plus_constant (Pmode, g1, -PROBE_INTERVAL)));
4629 emit_stack_probe (g1);
4632 emit_stack_probe (plus_constant (Pmode, g1,
4633 (i - PROBE_INTERVAL) - size));
4636 /* Otherwise, do the same as above, but in a loop. Note that we must be
4637 extra careful with variables wrapping around because we might be at
4638 the very top (or the very bottom) of the address space and we have
4639 to be able to handle this case properly; in particular, we use an
4640 equality test for the loop condition. */
4643 HOST_WIDE_INT rounded_size;
4644 rtx g4 = gen_rtx_REG (Pmode, 4);
4646 emit_move_insn (g1, GEN_INT (first));
4649 /* Step 1: round SIZE to the previous multiple of the interval. */
4651 rounded_size = size & -PROBE_INTERVAL;
4652 emit_move_insn (g4, GEN_INT (rounded_size));
4655 /* Step 2: compute initial and final value of the loop counter. */
4657 /* TEST_ADDR = SP + FIRST. */
4658 emit_insn (gen_rtx_SET (VOIDmode, g1,
4659 gen_rtx_MINUS (Pmode, stack_pointer_rtx, g1)));
4661 /* LAST_ADDR = SP + FIRST + ROUNDED_SIZE. */
4662 emit_insn (gen_rtx_SET (VOIDmode, g4, gen_rtx_MINUS (Pmode, g1, g4)));
4667 while (TEST_ADDR != LAST_ADDR)
4669 TEST_ADDR = TEST_ADDR + PROBE_INTERVAL
4673 probes at FIRST + N * PROBE_INTERVAL for values of N from 1
4674 until it is equal to ROUNDED_SIZE. */
4677 emit_insn (gen_probe_stack_rangedi (g1, g1, g4));
4679 emit_insn (gen_probe_stack_rangesi (g1, g1, g4));
4682 /* Step 4: probe at FIRST + SIZE if we cannot assert at compile-time
4683 that SIZE is equal to ROUNDED_SIZE. */
4685 if (size != rounded_size)
4686 emit_stack_probe (plus_constant (Pmode, g4, rounded_size - size));
4689 /* Make sure nothing is scheduled before we are done. */
4690 emit_insn (gen_blockage ());
4693 /* Probe a range of stack addresses from REG1 to REG2 inclusive. These are
4694 absolute addresses. */
4697 output_probe_stack_range (rtx reg1, rtx reg2)
4699 static int labelno = 0;
4700 char loop_lab[32], end_lab[32];
4703 ASM_GENERATE_INTERNAL_LABEL (loop_lab, "LPSRL", labelno);
4704 ASM_GENERATE_INTERNAL_LABEL (end_lab, "LPSRE", labelno++);
4706 ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, loop_lab);
4708 /* Jump to END_LAB if TEST_ADDR == LAST_ADDR. */
4711 output_asm_insn ("cmp\t%0, %1", xops);
4713 fputs ("\tbe,pn\t%xcc,", asm_out_file);
4715 fputs ("\tbe\t", asm_out_file);
4716 assemble_name_raw (asm_out_file, end_lab);
4717 fputc ('\n', asm_out_file);
4719 /* TEST_ADDR = TEST_ADDR + PROBE_INTERVAL. */
4720 xops[1] = GEN_INT (-PROBE_INTERVAL);
4721 output_asm_insn (" add\t%0, %1, %0", xops);
4723 /* Probe at TEST_ADDR and branch. */
4725 fputs ("\tba,pt\t%xcc,", asm_out_file);
4727 fputs ("\tba\t", asm_out_file);
4728 assemble_name_raw (asm_out_file, loop_lab);
4729 fputc ('\n', asm_out_file);
4730 xops[1] = GEN_INT (SPARC_STACK_BIAS);
4731 output_asm_insn (" st\t%%g0, [%0+%1]", xops);
4733 ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, end_lab);
4738 /* Emit code to save/restore registers from LOW to HIGH at BASE+OFFSET as
4739 needed. LOW is supposed to be double-word aligned for 32-bit registers.
4740 SAVE_P decides whether a register must be saved/restored. ACTION_TRUE
4741 is the action to be performed if SAVE_P returns true and ACTION_FALSE
4742 the action to be performed if it returns false. Return the new offset. */
4744 typedef bool (*sorr_pred_t) (unsigned int, int);
4745 typedef enum { SORR_NONE, SORR_ADVANCE, SORR_SAVE, SORR_RESTORE } sorr_act_t;
4748 emit_save_or_restore_regs (unsigned int low, unsigned int high, rtx base,
4749 int offset, int leaf_function, sorr_pred_t save_p,
4750 sorr_act_t action_true, sorr_act_t action_false)
4755 if (TARGET_ARCH64 && high <= 32)
4759 for (i = low; i < high; i++)
4761 if (save_p (i, leaf_function))
4763 mem = gen_frame_mem (DImode, plus_constant (Pmode,
4765 if (action_true == SORR_SAVE)
4767 insn = emit_move_insn (mem, gen_rtx_REG (DImode, i));
4768 RTX_FRAME_RELATED_P (insn) = 1;
4770 else /* action_true == SORR_RESTORE */
4772 /* The frame pointer must be restored last since its old
4773 value may be used as base address for the frame. This
4774 is problematic in 64-bit mode only because of the lack
4775 of double-word load instruction. */
4776 if (i == HARD_FRAME_POINTER_REGNUM)
4779 emit_move_insn (gen_rtx_REG (DImode, i), mem);
4783 else if (action_false == SORR_ADVANCE)
4789 mem = gen_frame_mem (DImode, plus_constant (Pmode, base, fp_offset));
4790 emit_move_insn (hard_frame_pointer_rtx, mem);
4795 for (i = low; i < high; i += 2)
4797 bool reg0 = save_p (i, leaf_function);
4798 bool reg1 = save_p (i + 1, leaf_function);
4799 enum machine_mode mode;
4804 mode = SPARC_INT_REG_P (i) ? DImode : DFmode;
4809 mode = SPARC_INT_REG_P (i) ? SImode : SFmode;
4814 mode = SPARC_INT_REG_P (i) ? SImode : SFmode;
4820 if (action_false == SORR_ADVANCE)
4825 mem = gen_frame_mem (mode, plus_constant (Pmode, base, offset));
4826 if (action_true == SORR_SAVE)
4828 insn = emit_move_insn (mem, gen_rtx_REG (mode, regno));
4829 RTX_FRAME_RELATED_P (insn) = 1;
4833 mem = gen_frame_mem (SImode, plus_constant (Pmode, base,
4835 set1 = gen_rtx_SET (VOIDmode, mem,
4836 gen_rtx_REG (SImode, regno));
4837 RTX_FRAME_RELATED_P (set1) = 1;
4839 = gen_frame_mem (SImode, plus_constant (Pmode, base,
4841 set2 = gen_rtx_SET (VOIDmode, mem,
4842 gen_rtx_REG (SImode, regno + 1));
4843 RTX_FRAME_RELATED_P (set2) = 1;
4844 add_reg_note (insn, REG_FRAME_RELATED_EXPR,
4845 gen_rtx_PARALLEL (VOIDmode,
4846 gen_rtvec (2, set1, set2)));
4849 else /* action_true == SORR_RESTORE */
4850 emit_move_insn (gen_rtx_REG (mode, regno), mem);
4852 /* Always preserve double-word alignment. */
4853 offset = (offset + 8) & -8;
4860 /* Emit code to adjust BASE to OFFSET. Return the new base. */
4863 emit_adjust_base_to_offset (rtx base, int offset)
4865 /* ??? This might be optimized a little as %g1 might already have a
4866 value close enough that a single add insn will do. */
4867 /* ??? Although, all of this is probably only a temporary fix because
4868 if %g1 can hold a function result, then sparc_expand_epilogue will
4869 lose (the result will be clobbered). */
4870 rtx new_base = gen_rtx_REG (Pmode, 1);
4871 emit_move_insn (new_base, GEN_INT (offset));
4872 emit_insn (gen_rtx_SET (VOIDmode,
4873 new_base, gen_rtx_PLUS (Pmode, base, new_base)));
4877 /* Emit code to save/restore call-saved global and FP registers. */
4880 emit_save_or_restore_global_fp_regs (rtx base, int offset, sorr_act_t action)
4882 if (offset < -4096 || offset + sparc_n_global_fp_regs * 4 > 4095)
4884 base = emit_adjust_base_to_offset (base, offset);
4889 = emit_save_or_restore_regs (0, 8, base, offset, 0,
4890 save_global_or_fp_reg_p, action, SORR_NONE);
4891 emit_save_or_restore_regs (32, TARGET_V9 ? 96 : 64, base, offset, 0,
4892 save_global_or_fp_reg_p, action, SORR_NONE);
4895 /* Emit code to save/restore call-saved local and in registers. */
4898 emit_save_or_restore_local_in_regs (rtx base, int offset, sorr_act_t action)
4900 if (offset < -4096 || offset + 16 * UNITS_PER_WORD > 4095)
4902 base = emit_adjust_base_to_offset (base, offset);
4906 emit_save_or_restore_regs (16, 32, base, offset, sparc_leaf_function_p,
4907 save_local_or_in_reg_p, action, SORR_ADVANCE);
4910 /* Emit a window_save insn. */
4913 emit_window_save (rtx increment)
4915 rtx insn = emit_insn (gen_window_save (increment));
4916 RTX_FRAME_RELATED_P (insn) = 1;
4918 /* The incoming return address (%o7) is saved in %i7. */
4919 add_reg_note (insn, REG_CFA_REGISTER,
4920 gen_rtx_SET (VOIDmode,
4921 gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM),
4923 INCOMING_RETURN_ADDR_REGNUM)));
4925 /* The window save event. */
4926 add_reg_note (insn, REG_CFA_WINDOW_SAVE, const0_rtx);
4928 /* The CFA is %fp, the hard frame pointer. */
4929 add_reg_note (insn, REG_CFA_DEF_CFA,
4930 plus_constant (Pmode, hard_frame_pointer_rtx,
4931 INCOMING_FRAME_SP_OFFSET));
4936 /* Generate an increment for the stack pointer. */
4939 gen_stack_pointer_inc (rtx increment)
4941 return gen_rtx_SET (VOIDmode,
4943 gen_rtx_PLUS (Pmode,
4948 /* Generate a decrement for the stack pointer. */
4951 gen_stack_pointer_dec (rtx decrement)
4953 return gen_rtx_SET (VOIDmode,
4955 gen_rtx_MINUS (Pmode,
4960 /* Expand the function prologue. The prologue is responsible for reserving
4961 storage for the frame, saving the call-saved registers and loading the
4962 GOT register if needed. */
4965 sparc_expand_prologue (void)
4970 /* Compute a snapshot of current_function_uses_only_leaf_regs. Relying
4971 on the final value of the flag means deferring the prologue/epilogue
4972 expansion until just before the second scheduling pass, which is too
4973 late to emit multiple epilogues or return insns.
4975 Of course we are making the assumption that the value of the flag
4976 will not change between now and its final value. Of the three parts
4977 of the formula, only the last one can reasonably vary. Let's take a
4978 closer look, after assuming that the first two ones are set to true
4979 (otherwise the last value is effectively silenced).
4981 If only_leaf_regs_used returns false, the global predicate will also
4982 be false so the actual frame size calculated below will be positive.
4983 As a consequence, the save_register_window insn will be emitted in
4984 the instruction stream; now this insn explicitly references %fp
4985 which is not a leaf register so only_leaf_regs_used will always
4986 return false subsequently.
4988 If only_leaf_regs_used returns true, we hope that the subsequent
4989 optimization passes won't cause non-leaf registers to pop up. For
4990 example, the regrename pass has special provisions to not rename to
4991 non-leaf registers in a leaf function. */
4992 sparc_leaf_function_p
4993 = optimize > 0 && current_function_is_leaf && only_leaf_regs_used ();
4995 size = sparc_compute_frame_size (get_frame_size(), sparc_leaf_function_p);
4997 if (flag_stack_usage_info)
4998 current_function_static_stack_size = size;
5000 if (flag_stack_check == STATIC_BUILTIN_STACK_CHECK && size)
5001 sparc_emit_probe_stack_range (STACK_CHECK_PROTECT, size);
5005 else if (sparc_leaf_function_p)
5007 rtx size_int_rtx = GEN_INT (-size);
5010 insn = emit_insn (gen_stack_pointer_inc (size_int_rtx));
5011 else if (size <= 8192)
5013 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (-4096)));
5014 RTX_FRAME_RELATED_P (insn) = 1;
5016 /* %sp is still the CFA register. */
5017 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (4096 - size)));
5021 rtx size_rtx = gen_rtx_REG (Pmode, 1);
5022 emit_move_insn (size_rtx, size_int_rtx);
5023 insn = emit_insn (gen_stack_pointer_inc (size_rtx));
5024 add_reg_note (insn, REG_FRAME_RELATED_EXPR,
5025 gen_stack_pointer_inc (size_int_rtx));
5028 RTX_FRAME_RELATED_P (insn) = 1;
5032 rtx size_int_rtx = GEN_INT (-size);
5035 emit_window_save (size_int_rtx);
5036 else if (size <= 8192)
5038 emit_window_save (GEN_INT (-4096));
5040 /* %sp is not the CFA register anymore. */
5041 emit_insn (gen_stack_pointer_inc (GEN_INT (4096 - size)));
5043 /* Make sure no %fp-based store is issued until after the frame is
5044 established. The offset between the frame pointer and the stack
5045 pointer is calculated relative to the value of the stack pointer
5046 at the end of the function prologue, and moving instructions that
5047 access the stack via the frame pointer between the instructions
5048 that decrement the stack pointer could result in accessing the
5049 register window save area, which is volatile. */
5050 emit_insn (gen_frame_blockage ());
5054 rtx size_rtx = gen_rtx_REG (Pmode, 1);
5055 emit_move_insn (size_rtx, size_int_rtx);
5056 emit_window_save (size_rtx);
5060 if (sparc_leaf_function_p)
5062 sparc_frame_base_reg = stack_pointer_rtx;
5063 sparc_frame_base_offset = size + SPARC_STACK_BIAS;
5067 sparc_frame_base_reg = hard_frame_pointer_rtx;
5068 sparc_frame_base_offset = SPARC_STACK_BIAS;
5071 if (sparc_n_global_fp_regs > 0)
5072 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg,
5073 sparc_frame_base_offset
5074 - sparc_apparent_frame_size,
5077 /* Load the GOT register if needed. */
5078 if (crtl->uses_pic_offset_table)
5079 load_got_register ();
5081 /* Advertise that the data calculated just above are now valid. */
5082 sparc_prologue_data_valid_p = true;
5085 /* Expand the function prologue. The prologue is responsible for reserving
5086 storage for the frame, saving the call-saved registers and loading the
5087 GOT register if needed. */
5090 sparc_flat_expand_prologue (void)
5095 sparc_leaf_function_p = optimize > 0 && current_function_is_leaf;
5097 size = sparc_compute_frame_size (get_frame_size(), sparc_leaf_function_p);
5099 if (flag_stack_usage_info)
5100 current_function_static_stack_size = size;
5102 if (flag_stack_check == STATIC_BUILTIN_STACK_CHECK && size)
5103 sparc_emit_probe_stack_range (STACK_CHECK_PROTECT, size);
5105 if (sparc_save_local_in_regs_p)
5106 emit_save_or_restore_local_in_regs (stack_pointer_rtx, SPARC_STACK_BIAS,
5113 rtx size_int_rtx, size_rtx;
5115 size_rtx = size_int_rtx = GEN_INT (-size);
5117 /* We establish the frame (i.e. decrement the stack pointer) first, even
5118 if we use a frame pointer, because we cannot clobber any call-saved
5119 registers, including the frame pointer, if we haven't created a new
5120 register save area, for the sake of compatibility with the ABI. */
5122 insn = emit_insn (gen_stack_pointer_inc (size_int_rtx));
5123 else if (size <= 8192 && !frame_pointer_needed)
5125 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (-4096)));
5126 RTX_FRAME_RELATED_P (insn) = 1;
5127 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (4096 - size)));
5131 size_rtx = gen_rtx_REG (Pmode, 1);
5132 emit_move_insn (size_rtx, size_int_rtx);
5133 insn = emit_insn (gen_stack_pointer_inc (size_rtx));
5134 add_reg_note (insn, REG_CFA_ADJUST_CFA,
5135 gen_stack_pointer_inc (size_int_rtx));
5137 RTX_FRAME_RELATED_P (insn) = 1;
5139 /* Ensure nothing is scheduled until after the frame is established. */
5140 emit_insn (gen_blockage ());
5142 if (frame_pointer_needed)
5144 insn = emit_insn (gen_rtx_SET (VOIDmode, hard_frame_pointer_rtx,
5145 gen_rtx_MINUS (Pmode,
5148 RTX_FRAME_RELATED_P (insn) = 1;
5150 add_reg_note (insn, REG_CFA_ADJUST_CFA,
5151 gen_rtx_SET (VOIDmode, hard_frame_pointer_rtx,
5152 plus_constant (Pmode, stack_pointer_rtx,
5156 if (return_addr_reg_needed_p (sparc_leaf_function_p))
5158 rtx o7 = gen_rtx_REG (Pmode, INCOMING_RETURN_ADDR_REGNUM);
5159 rtx i7 = gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM);
5161 insn = emit_move_insn (i7, o7);
5162 RTX_FRAME_RELATED_P (insn) = 1;
5164 add_reg_note (insn, REG_CFA_REGISTER,
5165 gen_rtx_SET (VOIDmode, i7, o7));
5167 /* Prevent this instruction from ever being considered dead,
5168 even if this function has no epilogue. */
5173 if (frame_pointer_needed)
5175 sparc_frame_base_reg = hard_frame_pointer_rtx;
5176 sparc_frame_base_offset = SPARC_STACK_BIAS;
5180 sparc_frame_base_reg = stack_pointer_rtx;
5181 sparc_frame_base_offset = size + SPARC_STACK_BIAS;
5184 if (sparc_n_global_fp_regs > 0)
5185 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg,
5186 sparc_frame_base_offset
5187 - sparc_apparent_frame_size,
5190 /* Load the GOT register if needed. */
5191 if (crtl->uses_pic_offset_table)
5192 load_got_register ();
5194 /* Advertise that the data calculated just above are now valid. */
5195 sparc_prologue_data_valid_p = true;
5198 /* This function generates the assembly code for function entry, which boils
5199 down to emitting the necessary .register directives. */
5202 sparc_asm_function_prologue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
5204 /* Check that the assumption we made in sparc_expand_prologue is valid. */
5206 gcc_assert (sparc_leaf_function_p == current_function_uses_only_leaf_regs);
5208 sparc_output_scratch_registers (file);
5211 /* Expand the function epilogue, either normal or part of a sibcall.
5212 We emit all the instructions except the return or the call. */
5215 sparc_expand_epilogue (bool for_eh)
5217 HOST_WIDE_INT size = sparc_frame_size;
5219 if (sparc_n_global_fp_regs > 0)
5220 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg,
5221 sparc_frame_base_offset
5222 - sparc_apparent_frame_size,
5225 if (size == 0 || for_eh)
5227 else if (sparc_leaf_function_p)
5230 emit_insn (gen_stack_pointer_dec (GEN_INT (-size)));
5231 else if (size <= 8192)
5233 emit_insn (gen_stack_pointer_dec (GEN_INT (-4096)));
5234 emit_insn (gen_stack_pointer_dec (GEN_INT (4096 - size)));
5238 rtx reg = gen_rtx_REG (Pmode, 1);
5239 emit_move_insn (reg, GEN_INT (-size));
5240 emit_insn (gen_stack_pointer_dec (reg));
5245 /* Expand the function epilogue, either normal or part of a sibcall.
5246 We emit all the instructions except the return or the call. */
5249 sparc_flat_expand_epilogue (bool for_eh)
5251 HOST_WIDE_INT size = sparc_frame_size;
5253 if (sparc_n_global_fp_regs > 0)
5254 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg,
5255 sparc_frame_base_offset
5256 - sparc_apparent_frame_size,
5259 /* If we have a frame pointer, we'll need both to restore it before the
5260 frame is destroyed and use its current value in destroying the frame.
5261 Since we don't have an atomic way to do that in the flat window model,
5262 we save the current value into a temporary register (%g1). */
5263 if (frame_pointer_needed && !for_eh)
5264 emit_move_insn (gen_rtx_REG (Pmode, 1), hard_frame_pointer_rtx);
5266 if (return_addr_reg_needed_p (sparc_leaf_function_p))
5267 emit_move_insn (gen_rtx_REG (Pmode, INCOMING_RETURN_ADDR_REGNUM),
5268 gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM));
5270 if (sparc_save_local_in_regs_p)
5271 emit_save_or_restore_local_in_regs (sparc_frame_base_reg,
5272 sparc_frame_base_offset,
5275 if (size == 0 || for_eh)
5277 else if (frame_pointer_needed)
5279 /* Make sure the frame is destroyed after everything else is done. */
5280 emit_insn (gen_blockage ());
5282 emit_move_insn (stack_pointer_rtx, gen_rtx_REG (Pmode, 1));
5287 emit_insn (gen_blockage ());
5290 emit_insn (gen_stack_pointer_dec (GEN_INT (-size)));
5291 else if (size <= 8192)
5293 emit_insn (gen_stack_pointer_dec (GEN_INT (-4096)));
5294 emit_insn (gen_stack_pointer_dec (GEN_INT (4096 - size)));
5298 rtx reg = gen_rtx_REG (Pmode, 1);
5299 emit_move_insn (reg, GEN_INT (-size));
5300 emit_insn (gen_stack_pointer_dec (reg));
5305 /* Return true if it is appropriate to emit `return' instructions in the
5306 body of a function. */
5309 sparc_can_use_return_insn_p (void)
5311 return sparc_prologue_data_valid_p
5312 && sparc_n_global_fp_regs == 0
5314 ? (sparc_frame_size == 0 && !sparc_save_local_in_regs_p)
5315 : (sparc_frame_size == 0 || !sparc_leaf_function_p);
5318 /* This function generates the assembly code for function exit. */
5321 sparc_asm_function_epilogue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
5323 /* If the last two instructions of a function are "call foo; dslot;"
5324 the return address might point to the first instruction in the next
5325 function and we have to output a dummy nop for the sake of sane
5326 backtraces in such cases. This is pointless for sibling calls since
5327 the return address is explicitly adjusted. */
5329 rtx insn, last_real_insn;
5331 insn = get_last_insn ();
5333 last_real_insn = prev_real_insn (insn);
5335 && GET_CODE (last_real_insn) == INSN
5336 && GET_CODE (PATTERN (last_real_insn)) == SEQUENCE)
5337 last_real_insn = XVECEXP (PATTERN (last_real_insn), 0, 0);
5340 && CALL_P (last_real_insn)
5341 && !SIBLING_CALL_P (last_real_insn))
5342 fputs("\tnop\n", file);
5344 sparc_output_deferred_case_vectors ();
5347 /* Output a 'restore' instruction. */
5350 output_restore (rtx pat)
5356 fputs ("\t restore\n", asm_out_file);
5360 gcc_assert (GET_CODE (pat) == SET);
5362 operands[0] = SET_DEST (pat);
5363 pat = SET_SRC (pat);
5365 switch (GET_CODE (pat))
5368 operands[1] = XEXP (pat, 0);
5369 operands[2] = XEXP (pat, 1);
5370 output_asm_insn (" restore %r1, %2, %Y0", operands);
5373 operands[1] = XEXP (pat, 0);
5374 operands[2] = XEXP (pat, 1);
5375 output_asm_insn (" restore %r1, %%lo(%a2), %Y0", operands);
5378 operands[1] = XEXP (pat, 0);
5379 gcc_assert (XEXP (pat, 1) == const1_rtx);
5380 output_asm_insn (" restore %r1, %r1, %Y0", operands);
5384 output_asm_insn (" restore %%g0, %1, %Y0", operands);
5389 /* Output a return. */
5392 output_return (rtx insn)
5394 if (crtl->calls_eh_return)
5396 /* If the function uses __builtin_eh_return, the eh_return
5397 machinery occupies the delay slot. */
5398 gcc_assert (!final_sequence);
5400 if (flag_delayed_branch)
5402 if (!TARGET_FLAT && TARGET_V9)
5403 fputs ("\treturn\t%i7+8\n", asm_out_file);
5407 fputs ("\trestore\n", asm_out_file);
5409 fputs ("\tjmp\t%o7+8\n", asm_out_file);
5412 fputs ("\t add\t%sp, %g1, %sp\n", asm_out_file);
5417 fputs ("\trestore\n", asm_out_file);
5419 fputs ("\tadd\t%sp, %g1, %sp\n", asm_out_file);
5420 fputs ("\tjmp\t%o7+8\n\t nop\n", asm_out_file);
5423 else if (sparc_leaf_function_p || TARGET_FLAT)
5425 /* This is a leaf or flat function so we don't have to bother restoring
5426 the register window, which frees us from dealing with the convoluted
5427 semantics of restore/return. We simply output the jump to the
5428 return address and the insn in the delay slot (if any). */
5430 return "jmp\t%%o7+%)%#";
5434 /* This is a regular function so we have to restore the register window.
5435 We may have a pending insn for the delay slot, which will be either
5436 combined with the 'restore' instruction or put in the delay slot of
5437 the 'return' instruction. */
5443 delay = NEXT_INSN (insn);
5446 pat = PATTERN (delay);
5448 if (TARGET_V9 && ! epilogue_renumber (&pat, 1))
5450 epilogue_renumber (&pat, 0);
5451 return "return\t%%i7+%)%#";
5455 output_asm_insn ("jmp\t%%i7+%)", NULL);
5456 output_restore (pat);
5457 PATTERN (delay) = gen_blockage ();
5458 INSN_CODE (delay) = -1;
5463 /* The delay slot is empty. */
5465 return "return\t%%i7+%)\n\t nop";
5466 else if (flag_delayed_branch)
5467 return "jmp\t%%i7+%)\n\t restore";
5469 return "restore\n\tjmp\t%%o7+%)\n\t nop";
5476 /* Output a sibling call. */
5479 output_sibcall (rtx insn, rtx call_operand)
5483 gcc_assert (flag_delayed_branch);
5485 operands[0] = call_operand;
5487 if (sparc_leaf_function_p || TARGET_FLAT)
5489 /* This is a leaf or flat function so we don't have to bother restoring
5490 the register window. We simply output the jump to the function and
5491 the insn in the delay slot (if any). */
5493 gcc_assert (!(LEAF_SIBCALL_SLOT_RESERVED_P && final_sequence));
5496 output_asm_insn ("sethi\t%%hi(%a0), %%g1\n\tjmp\t%%g1 + %%lo(%a0)%#",
5499 /* Use or with rs2 %%g0 instead of mov, so that as/ld can optimize
5500 it into branch if possible. */
5501 output_asm_insn ("or\t%%o7, %%g0, %%g1\n\tcall\t%a0, 0\n\t or\t%%g1, %%g0, %%o7",
5506 /* This is a regular function so we have to restore the register window.
5507 We may have a pending insn for the delay slot, which will be combined
5508 with the 'restore' instruction. */
5510 output_asm_insn ("call\t%a0, 0", operands);
5514 rtx delay = NEXT_INSN (insn);
5517 output_restore (PATTERN (delay));
5519 PATTERN (delay) = gen_blockage ();
5520 INSN_CODE (delay) = -1;
5523 output_restore (NULL_RTX);
5529 /* Functions for handling argument passing.
5531 For 32-bit, the first 6 args are normally in registers and the rest are
5532 pushed. Any arg that starts within the first 6 words is at least
5533 partially passed in a register unless its data type forbids.
5535 For 64-bit, the argument registers are laid out as an array of 16 elements
5536 and arguments are added sequentially. The first 6 int args and up to the
5537 first 16 fp args (depending on size) are passed in regs.
5539 Slot Stack Integral Float Float in structure Double Long Double
5540 ---- ----- -------- ----- ------------------ ------ -----------
5541 15 [SP+248] %f31 %f30,%f31 %d30
5542 14 [SP+240] %f29 %f28,%f29 %d28 %q28
5543 13 [SP+232] %f27 %f26,%f27 %d26
5544 12 [SP+224] %f25 %f24,%f25 %d24 %q24
5545 11 [SP+216] %f23 %f22,%f23 %d22
5546 10 [SP+208] %f21 %f20,%f21 %d20 %q20
5547 9 [SP+200] %f19 %f18,%f19 %d18
5548 8 [SP+192] %f17 %f16,%f17 %d16 %q16
5549 7 [SP+184] %f15 %f14,%f15 %d14
5550 6 [SP+176] %f13 %f12,%f13 %d12 %q12
5551 5 [SP+168] %o5 %f11 %f10,%f11 %d10
5552 4 [SP+160] %o4 %f9 %f8,%f9 %d8 %q8
5553 3 [SP+152] %o3 %f7 %f6,%f7 %d6
5554 2 [SP+144] %o2 %f5 %f4,%f5 %d4 %q4
5555 1 [SP+136] %o1 %f3 %f2,%f3 %d2
5556 0 [SP+128] %o0 %f1 %f0,%f1 %d0 %q0
5558 Here SP = %sp if -mno-stack-bias or %sp+stack_bias otherwise.
5560 Integral arguments are always passed as 64-bit quantities appropriately
5563 Passing of floating point values is handled as follows.
5564 If a prototype is in scope:
5565 If the value is in a named argument (i.e. not a stdarg function or a
5566 value not part of the `...') then the value is passed in the appropriate
5568 If the value is part of the `...' and is passed in one of the first 6
5569 slots then the value is passed in the appropriate int reg.
5570 If the value is part of the `...' and is not passed in one of the first 6
5571 slots then the value is passed in memory.
5572 If a prototype is not in scope:
5573 If the value is one of the first 6 arguments the value is passed in the
5574 appropriate integer reg and the appropriate fp reg.
5575 If the value is not one of the first 6 arguments the value is passed in
5576 the appropriate fp reg and in memory.
5579 Summary of the calling conventions implemented by GCC on the SPARC:
5582 size argument return value
5584 small integer <4 int. reg. int. reg.
5585 word 4 int. reg. int. reg.
5586 double word 8 int. reg. int. reg.
5588 _Complex small integer <8 int. reg. int. reg.
5589 _Complex word 8 int. reg. int. reg.
5590 _Complex double word 16 memory int. reg.
5592 vector integer <=8 int. reg. FP reg.
5593 vector integer >8 memory memory
5595 float 4 int. reg. FP reg.
5596 double 8 int. reg. FP reg.
5597 long double 16 memory memory
5599 _Complex float 8 memory FP reg.
5600 _Complex double 16 memory FP reg.
5601 _Complex long double 32 memory FP reg.
5603 vector float any memory memory
5605 aggregate any memory memory
5610 size argument return value
5612 small integer <8 int. reg. int. reg.
5613 word 8 int. reg. int. reg.
5614 double word 16 int. reg. int. reg.
5616 _Complex small integer <16 int. reg. int. reg.
5617 _Complex word 16 int. reg. int. reg.
5618 _Complex double word 32 memory int. reg.
5620 vector integer <=16 FP reg. FP reg.
5621 vector integer 16<s<=32 memory FP reg.
5622 vector integer >32 memory memory
5624 float 4 FP reg. FP reg.
5625 double 8 FP reg. FP reg.
5626 long double 16 FP reg. FP reg.
5628 _Complex float 8 FP reg. FP reg.
5629 _Complex double 16 FP reg. FP reg.
5630 _Complex long double 32 memory FP reg.
5632 vector float <=16 FP reg. FP reg.
5633 vector float 16<s<=32 memory FP reg.
5634 vector float >32 memory memory
5636 aggregate <=16 reg. reg.
5637 aggregate 16<s<=32 memory reg.
5638 aggregate >32 memory memory
5642 Note #1: complex floating-point types follow the extended SPARC ABIs as
5643 implemented by the Sun compiler.
5645 Note #2: integral vector types follow the scalar floating-point types
5646 conventions to match what is implemented by the Sun VIS SDK.
5648 Note #3: floating-point vector types follow the aggregate types
5652 /* Maximum number of int regs for args. */
5653 #define SPARC_INT_ARG_MAX 6
5654 /* Maximum number of fp regs for args. */
5655 #define SPARC_FP_ARG_MAX 16
5657 #define ROUND_ADVANCE(SIZE) (((SIZE) + UNITS_PER_WORD - 1) / UNITS_PER_WORD)
5659 /* Handle the INIT_CUMULATIVE_ARGS macro.
5660 Initialize a variable CUM of type CUMULATIVE_ARGS
5661 for a call to a function whose data type is FNTYPE.
5662 For a library call, FNTYPE is 0. */
5665 init_cumulative_args (struct sparc_args *cum, tree fntype,
5666 rtx libname ATTRIBUTE_UNUSED,
5667 tree fndecl ATTRIBUTE_UNUSED)
5670 cum->prototype_p = fntype && prototype_p (fntype);
5671 cum->libcall_p = fntype == 0;
5674 /* Handle promotion of pointer and integer arguments. */
5676 static enum machine_mode
5677 sparc_promote_function_mode (const_tree type,
5678 enum machine_mode mode,
5680 const_tree fntype ATTRIBUTE_UNUSED,
5681 int for_return ATTRIBUTE_UNUSED)
5683 if (type != NULL_TREE && POINTER_TYPE_P (type))
5685 *punsignedp = POINTERS_EXTEND_UNSIGNED;
5689 /* Integral arguments are passed as full words, as per the ABI. */
5690 if (GET_MODE_CLASS (mode) == MODE_INT
5691 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5697 /* Handle the TARGET_STRICT_ARGUMENT_NAMING target hook. */
5700 sparc_strict_argument_naming (cumulative_args_t ca ATTRIBUTE_UNUSED)
5702 return TARGET_ARCH64 ? true : false;
5705 /* Scan the record type TYPE and return the following predicates:
5706 - INTREGS_P: the record contains at least one field or sub-field
5707 that is eligible for promotion in integer registers.
5708 - FP_REGS_P: the record contains at least one field or sub-field
5709 that is eligible for promotion in floating-point registers.
5710 - PACKED_P: the record contains at least one field that is packed.
5712 Sub-fields are not taken into account for the PACKED_P predicate. */
5715 scan_record_type (const_tree type, int *intregs_p, int *fpregs_p,
5720 for (field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field))
5722 if (TREE_CODE (field) == FIELD_DECL)
5724 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
5725 scan_record_type (TREE_TYPE (field), intregs_p, fpregs_p, 0);
5726 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
5727 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
5733 if (packed_p && DECL_PACKED (field))
5739 /* Compute the slot number to pass an argument in.
5740 Return the slot number or -1 if passing on the stack.
5742 CUM is a variable of type CUMULATIVE_ARGS which gives info about
5743 the preceding args and about the function being called.
5744 MODE is the argument's machine mode.
5745 TYPE is the data type of the argument (as a tree).
5746 This is null for libcalls where that information may
5748 NAMED is nonzero if this argument is a named parameter
5749 (otherwise it is an extra parameter matching an ellipsis).
5750 INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG.
5751 *PREGNO records the register number to use if scalar type.
5752 *PPADDING records the amount of padding needed in words. */
5755 function_arg_slotno (const struct sparc_args *cum, enum machine_mode mode,
5756 const_tree type, bool named, bool incoming_p,
5757 int *pregno, int *ppadding)
5759 int regbase = (incoming_p
5760 ? SPARC_INCOMING_INT_ARG_FIRST
5761 : SPARC_OUTGOING_INT_ARG_FIRST);
5762 int slotno = cum->words;
5763 enum mode_class mclass;
5768 if (type && TREE_ADDRESSABLE (type))
5774 && TYPE_ALIGN (type) % PARM_BOUNDARY != 0)
5777 /* For SPARC64, objects requiring 16-byte alignment get it. */
5779 && (type ? TYPE_ALIGN (type) : GET_MODE_ALIGNMENT (mode)) >= 128
5780 && (slotno & 1) != 0)
5781 slotno++, *ppadding = 1;
5783 mclass = GET_MODE_CLASS (mode);
5784 if (type && TREE_CODE (type) == VECTOR_TYPE)
5786 /* Vector types deserve special treatment because they are
5787 polymorphic wrt their mode, depending upon whether VIS
5788 instructions are enabled. */
5789 if (TREE_CODE (TREE_TYPE (type)) == REAL_TYPE)
5791 /* The SPARC port defines no floating-point vector modes. */
5792 gcc_assert (mode == BLKmode);
5796 /* Integral vector types should either have a vector
5797 mode or an integral mode, because we are guaranteed
5798 by pass_by_reference that their size is not greater
5799 than 16 bytes and TImode is 16-byte wide. */
5800 gcc_assert (mode != BLKmode);
5802 /* Vector integers are handled like floats according to
5804 mclass = MODE_FLOAT;
5811 case MODE_COMPLEX_FLOAT:
5812 case MODE_VECTOR_INT:
5813 if (TARGET_ARCH64 && TARGET_FPU && named)
5815 if (slotno >= SPARC_FP_ARG_MAX)
5817 regno = SPARC_FP_ARG_FIRST + slotno * 2;
5818 /* Arguments filling only one single FP register are
5819 right-justified in the outer double FP register. */
5820 if (GET_MODE_SIZE (mode) <= 4)
5827 case MODE_COMPLEX_INT:
5828 if (slotno >= SPARC_INT_ARG_MAX)
5830 regno = regbase + slotno;
5834 if (mode == VOIDmode)
5835 /* MODE is VOIDmode when generating the actual call. */
5838 gcc_assert (mode == BLKmode);
5842 || (TREE_CODE (type) != VECTOR_TYPE
5843 && TREE_CODE (type) != RECORD_TYPE))
5845 if (slotno >= SPARC_INT_ARG_MAX)
5847 regno = regbase + slotno;
5849 else /* TARGET_ARCH64 && type */
5851 int intregs_p = 0, fpregs_p = 0, packed_p = 0;
5853 /* First see what kinds of registers we would need. */
5854 if (TREE_CODE (type) == VECTOR_TYPE)
5857 scan_record_type (type, &intregs_p, &fpregs_p, &packed_p);
5859 /* The ABI obviously doesn't specify how packed structures
5860 are passed. These are defined to be passed in int regs
5861 if possible, otherwise memory. */
5862 if (packed_p || !named)
5863 fpregs_p = 0, intregs_p = 1;
5865 /* If all arg slots are filled, then must pass on stack. */
5866 if (fpregs_p && slotno >= SPARC_FP_ARG_MAX)
5869 /* If there are only int args and all int arg slots are filled,
5870 then must pass on stack. */
5871 if (!fpregs_p && intregs_p && slotno >= SPARC_INT_ARG_MAX)
5874 /* Note that even if all int arg slots are filled, fp members may
5875 still be passed in regs if such regs are available.
5876 *PREGNO isn't set because there may be more than one, it's up
5877 to the caller to compute them. */
5890 /* Handle recursive register counting for structure field layout. */
5892 struct function_arg_record_value_parms
5894 rtx ret; /* return expression being built. */
5895 int slotno; /* slot number of the argument. */
5896 int named; /* whether the argument is named. */
5897 int regbase; /* regno of the base register. */
5898 int stack; /* 1 if part of the argument is on the stack. */
5899 int intoffset; /* offset of the first pending integer field. */
5900 unsigned int nregs; /* number of words passed in registers. */
5903 static void function_arg_record_value_3
5904 (HOST_WIDE_INT, struct function_arg_record_value_parms *);
5905 static void function_arg_record_value_2
5906 (const_tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
5907 static void function_arg_record_value_1
5908 (const_tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
5909 static rtx function_arg_record_value (const_tree, enum machine_mode, int, int, int);
5910 static rtx function_arg_union_value (int, enum machine_mode, int, int);
5912 /* A subroutine of function_arg_record_value. Traverse the structure
5913 recursively and determine how many registers will be required. */
5916 function_arg_record_value_1 (const_tree type, HOST_WIDE_INT startbitpos,
5917 struct function_arg_record_value_parms *parms,
5922 /* We need to compute how many registers are needed so we can
5923 allocate the PARALLEL but before we can do that we need to know
5924 whether there are any packed fields. The ABI obviously doesn't
5925 specify how structures are passed in this case, so they are
5926 defined to be passed in int regs if possible, otherwise memory,
5927 regardless of whether there are fp values present. */
5930 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5932 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
5939 /* Compute how many registers we need. */
5940 for (field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field))
5942 if (TREE_CODE (field) == FIELD_DECL)
5944 HOST_WIDE_INT bitpos = startbitpos;
5946 if (DECL_SIZE (field) != 0)
5948 if (integer_zerop (DECL_SIZE (field)))
5951 if (host_integerp (bit_position (field), 1))
5952 bitpos += int_bit_position (field);
5955 /* ??? FIXME: else assume zero offset. */
5957 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
5958 function_arg_record_value_1 (TREE_TYPE (field),
5962 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
5963 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
5968 if (parms->intoffset != -1)
5970 unsigned int startbit, endbit;
5971 int intslots, this_slotno;
5973 startbit = parms->intoffset & -BITS_PER_WORD;
5974 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
5976 intslots = (endbit - startbit) / BITS_PER_WORD;
5977 this_slotno = parms->slotno + parms->intoffset
5980 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
5982 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
5983 /* We need to pass this field on the stack. */
5987 parms->nregs += intslots;
5988 parms->intoffset = -1;
5991 /* There's no need to check this_slotno < SPARC_FP_ARG MAX.
5992 If it wasn't true we wouldn't be here. */
5993 if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE
5994 && DECL_MODE (field) == BLKmode)
5995 parms->nregs += TYPE_VECTOR_SUBPARTS (TREE_TYPE (field));
5996 else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
6003 if (parms->intoffset == -1)
6004 parms->intoffset = bitpos;
6010 /* A subroutine of function_arg_record_value. Assign the bits of the
6011 structure between parms->intoffset and bitpos to integer registers. */
6014 function_arg_record_value_3 (HOST_WIDE_INT bitpos,
6015 struct function_arg_record_value_parms *parms)
6017 enum machine_mode mode;
6019 unsigned int startbit, endbit;
6020 int this_slotno, intslots, intoffset;
6023 if (parms->intoffset == -1)
6026 intoffset = parms->intoffset;
6027 parms->intoffset = -1;
6029 startbit = intoffset & -BITS_PER_WORD;
6030 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
6031 intslots = (endbit - startbit) / BITS_PER_WORD;
6032 this_slotno = parms->slotno + intoffset / BITS_PER_WORD;
6034 intslots = MIN (intslots, SPARC_INT_ARG_MAX - this_slotno);
6038 /* If this is the trailing part of a word, only load that much into
6039 the register. Otherwise load the whole register. Note that in
6040 the latter case we may pick up unwanted bits. It's not a problem
6041 at the moment but may wish to revisit. */
6043 if (intoffset % BITS_PER_WORD != 0)
6044 mode = smallest_mode_for_size (BITS_PER_WORD - intoffset % BITS_PER_WORD,
6049 intoffset /= BITS_PER_UNIT;
6052 regno = parms->regbase + this_slotno;
6053 reg = gen_rtx_REG (mode, regno);
6054 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
6055 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (intoffset));
6058 intoffset = (intoffset | (UNITS_PER_WORD-1)) + 1;
6063 while (intslots > 0);
6066 /* A subroutine of function_arg_record_value. Traverse the structure
6067 recursively and assign bits to floating point registers. Track which
6068 bits in between need integer registers; invoke function_arg_record_value_3
6069 to make that happen. */
6072 function_arg_record_value_2 (const_tree type, HOST_WIDE_INT startbitpos,
6073 struct function_arg_record_value_parms *parms,
6079 for (field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field))
6081 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
6088 for (field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field))
6090 if (TREE_CODE (field) == FIELD_DECL)
6092 HOST_WIDE_INT bitpos = startbitpos;
6094 if (DECL_SIZE (field) != 0)
6096 if (integer_zerop (DECL_SIZE (field)))
6099 if (host_integerp (bit_position (field), 1))
6100 bitpos += int_bit_position (field);
6103 /* ??? FIXME: else assume zero offset. */
6105 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
6106 function_arg_record_value_2 (TREE_TYPE (field),
6110 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
6111 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
6116 int this_slotno = parms->slotno + bitpos / BITS_PER_WORD;
6117 int regno, nregs, pos;
6118 enum machine_mode mode = DECL_MODE (field);
6121 function_arg_record_value_3 (bitpos, parms);
6123 if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE
6126 mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field)));
6127 nregs = TYPE_VECTOR_SUBPARTS (TREE_TYPE (field));
6129 else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
6131 mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field)));
6137 regno = SPARC_FP_ARG_FIRST + this_slotno * 2;
6138 if (GET_MODE_SIZE (mode) <= 4 && (bitpos & 32) != 0)
6140 reg = gen_rtx_REG (mode, regno);
6141 pos = bitpos / BITS_PER_UNIT;
6142 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
6143 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos));
6147 regno += GET_MODE_SIZE (mode) / 4;
6148 reg = gen_rtx_REG (mode, regno);
6149 pos += GET_MODE_SIZE (mode);
6150 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
6151 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos));
6157 if (parms->intoffset == -1)
6158 parms->intoffset = bitpos;
6164 /* Used by function_arg and sparc_function_value_1 to implement the complex
6165 conventions of the 64-bit ABI for passing and returning structures.
6166 Return an expression valid as a return value for the FUNCTION_ARG
6167 and TARGET_FUNCTION_VALUE.
6169 TYPE is the data type of the argument (as a tree).
6170 This is null for libcalls where that information may
6172 MODE is the argument's machine mode.
6173 SLOTNO is the index number of the argument's slot in the parameter array.
6174 NAMED is nonzero if this argument is a named parameter
6175 (otherwise it is an extra parameter matching an ellipsis).
6176 REGBASE is the regno of the base register for the parameter array. */
6179 function_arg_record_value (const_tree type, enum machine_mode mode,
6180 int slotno, int named, int regbase)
6182 HOST_WIDE_INT typesize = int_size_in_bytes (type);
6183 struct function_arg_record_value_parms parms;
6186 parms.ret = NULL_RTX;
6187 parms.slotno = slotno;
6188 parms.named = named;
6189 parms.regbase = regbase;
6192 /* Compute how many registers we need. */
6194 parms.intoffset = 0;
6195 function_arg_record_value_1 (type, 0, &parms, false);
6197 /* Take into account pending integer fields. */
6198 if (parms.intoffset != -1)
6200 unsigned int startbit, endbit;
6201 int intslots, this_slotno;
6203 startbit = parms.intoffset & -BITS_PER_WORD;
6204 endbit = (typesize*BITS_PER_UNIT + BITS_PER_WORD - 1) & -BITS_PER_WORD;
6205 intslots = (endbit - startbit) / BITS_PER_WORD;
6206 this_slotno = slotno + parms.intoffset / BITS_PER_WORD;
6208 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
6210 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
6211 /* We need to pass this field on the stack. */
6215 parms.nregs += intslots;
6217 nregs = parms.nregs;
6219 /* Allocate the vector and handle some annoying special cases. */
6222 /* ??? Empty structure has no value? Duh? */
6225 /* Though there's nothing really to store, return a word register
6226 anyway so the rest of gcc doesn't go nuts. Returning a PARALLEL
6227 leads to breakage due to the fact that there are zero bytes to
6229 return gen_rtx_REG (mode, regbase);
6233 /* ??? C++ has structures with no fields, and yet a size. Give up
6234 for now and pass everything back in integer registers. */
6235 nregs = (typesize + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
6237 if (nregs + slotno > SPARC_INT_ARG_MAX)
6238 nregs = SPARC_INT_ARG_MAX - slotno;
6240 gcc_assert (nregs != 0);
6242 parms.ret = gen_rtx_PARALLEL (mode, rtvec_alloc (parms.stack + nregs));
6244 /* If at least one field must be passed on the stack, generate
6245 (parallel [(expr_list (nil) ...) ...]) so that all fields will
6246 also be passed on the stack. We can't do much better because the
6247 semantics of TARGET_ARG_PARTIAL_BYTES doesn't handle the case
6248 of structures for which the fields passed exclusively in registers
6249 are not at the beginning of the structure. */
6251 XVECEXP (parms.ret, 0, 0)
6252 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
6254 /* Fill in the entries. */
6256 parms.intoffset = 0;
6257 function_arg_record_value_2 (type, 0, &parms, false);
6258 function_arg_record_value_3 (typesize * BITS_PER_UNIT, &parms);
6260 gcc_assert (parms.nregs == nregs);
6265 /* Used by function_arg and sparc_function_value_1 to implement the conventions
6266 of the 64-bit ABI for passing and returning unions.
6267 Return an expression valid as a return value for the FUNCTION_ARG
6268 and TARGET_FUNCTION_VALUE.
6270 SIZE is the size in bytes of the union.
6271 MODE is the argument's machine mode.
6272 REGNO is the hard register the union will be passed in. */
6275 function_arg_union_value (int size, enum machine_mode mode, int slotno,
6278 int nwords = ROUND_ADVANCE (size), i;
6281 /* See comment in previous function for empty structures. */
6283 return gen_rtx_REG (mode, regno);
6285 if (slotno == SPARC_INT_ARG_MAX - 1)
6288 regs = gen_rtx_PARALLEL (mode, rtvec_alloc (nwords));
6290 for (i = 0; i < nwords; i++)
6292 /* Unions are passed left-justified. */
6293 XVECEXP (regs, 0, i)
6294 = gen_rtx_EXPR_LIST (VOIDmode,
6295 gen_rtx_REG (word_mode, regno),
6296 GEN_INT (UNITS_PER_WORD * i));
6303 /* Used by function_arg and sparc_function_value_1 to implement the conventions
6304 for passing and returning large (BLKmode) vectors.
6305 Return an expression valid as a return value for the FUNCTION_ARG
6306 and TARGET_FUNCTION_VALUE.
6308 SIZE is the size in bytes of the vector (at least 8 bytes).
6309 REGNO is the FP hard register the vector will be passed in. */
6312 function_arg_vector_value (int size, int regno)
6314 int i, nregs = size / 8;
6317 regs = gen_rtx_PARALLEL (BLKmode, rtvec_alloc (nregs));
6319 for (i = 0; i < nregs; i++)
6321 XVECEXP (regs, 0, i)
6322 = gen_rtx_EXPR_LIST (VOIDmode,
6323 gen_rtx_REG (DImode, regno + 2*i),
6330 /* Determine where to put an argument to a function.
6331 Value is zero to push the argument on the stack,
6332 or a hard register in which to store the argument.
6334 CUM is a variable of type CUMULATIVE_ARGS which gives info about
6335 the preceding args and about the function being called.
6336 MODE is the argument's machine mode.
6337 TYPE is the data type of the argument (as a tree).
6338 This is null for libcalls where that information may
6340 NAMED is true if this argument is a named parameter
6341 (otherwise it is an extra parameter matching an ellipsis).
6342 INCOMING_P is false for TARGET_FUNCTION_ARG, true for
6343 TARGET_FUNCTION_INCOMING_ARG. */
6346 sparc_function_arg_1 (cumulative_args_t cum_v, enum machine_mode mode,
6347 const_tree type, bool named, bool incoming_p)
6349 const CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v);
6351 int regbase = (incoming_p
6352 ? SPARC_INCOMING_INT_ARG_FIRST
6353 : SPARC_OUTGOING_INT_ARG_FIRST);
6354 int slotno, regno, padding;
6355 enum mode_class mclass = GET_MODE_CLASS (mode);
6357 slotno = function_arg_slotno (cum, mode, type, named, incoming_p,
6362 /* Vector types deserve special treatment because they are polymorphic wrt
6363 their mode, depending upon whether VIS instructions are enabled. */
6364 if (type && TREE_CODE (type) == VECTOR_TYPE)
6366 HOST_WIDE_INT size = int_size_in_bytes (type);
6367 gcc_assert ((TARGET_ARCH32 && size <= 8)
6368 || (TARGET_ARCH64 && size <= 16));
6370 if (mode == BLKmode)
6371 return function_arg_vector_value (size,
6372 SPARC_FP_ARG_FIRST + 2*slotno);
6374 mclass = MODE_FLOAT;
6378 return gen_rtx_REG (mode, regno);
6380 /* Structures up to 16 bytes in size are passed in arg slots on the stack
6381 and are promoted to registers if possible. */
6382 if (type && TREE_CODE (type) == RECORD_TYPE)
6384 HOST_WIDE_INT size = int_size_in_bytes (type);
6385 gcc_assert (size <= 16);
6387 return function_arg_record_value (type, mode, slotno, named, regbase);
6390 /* Unions up to 16 bytes in size are passed in integer registers. */
6391 else if (type && TREE_CODE (type) == UNION_TYPE)
6393 HOST_WIDE_INT size = int_size_in_bytes (type);
6394 gcc_assert (size <= 16);
6396 return function_arg_union_value (size, mode, slotno, regno);
6399 /* v9 fp args in reg slots beyond the int reg slots get passed in regs
6400 but also have the slot allocated for them.
6401 If no prototype is in scope fp values in register slots get passed
6402 in two places, either fp regs and int regs or fp regs and memory. */
6403 else if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT)
6404 && SPARC_FP_REG_P (regno))
6406 rtx reg = gen_rtx_REG (mode, regno);
6407 if (cum->prototype_p || cum->libcall_p)
6409 /* "* 2" because fp reg numbers are recorded in 4 byte
6412 /* ??? This will cause the value to be passed in the fp reg and
6413 in the stack. When a prototype exists we want to pass the
6414 value in the reg but reserve space on the stack. That's an
6415 optimization, and is deferred [for a bit]. */
6416 if ((regno - SPARC_FP_ARG_FIRST) >= SPARC_INT_ARG_MAX * 2)
6417 return gen_rtx_PARALLEL (mode,
6419 gen_rtx_EXPR_LIST (VOIDmode,
6420 NULL_RTX, const0_rtx),
6421 gen_rtx_EXPR_LIST (VOIDmode,
6425 /* ??? It seems that passing back a register even when past
6426 the area declared by REG_PARM_STACK_SPACE will allocate
6427 space appropriately, and will not copy the data onto the
6428 stack, exactly as we desire.
6430 This is due to locate_and_pad_parm being called in
6431 expand_call whenever reg_parm_stack_space > 0, which
6432 while beneficial to our example here, would seem to be
6433 in error from what had been intended. Ho hum... -- r~ */
6441 if ((regno - SPARC_FP_ARG_FIRST) < SPARC_INT_ARG_MAX * 2)
6445 /* On incoming, we don't need to know that the value
6446 is passed in %f0 and %i0, and it confuses other parts
6447 causing needless spillage even on the simplest cases. */
6451 intreg = (SPARC_OUTGOING_INT_ARG_FIRST
6452 + (regno - SPARC_FP_ARG_FIRST) / 2);
6454 v0 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
6455 v1 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_REG (mode, intreg),
6457 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
6461 v0 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
6462 v1 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
6463 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
6468 /* All other aggregate types are passed in an integer register in a mode
6469 corresponding to the size of the type. */
6470 else if (type && AGGREGATE_TYPE_P (type))
6472 HOST_WIDE_INT size = int_size_in_bytes (type);
6473 gcc_assert (size <= 16);
6475 mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0);
6478 return gen_rtx_REG (mode, regno);
6481 /* Handle the TARGET_FUNCTION_ARG target hook. */
6484 sparc_function_arg (cumulative_args_t cum, enum machine_mode mode,
6485 const_tree type, bool named)
6487 return sparc_function_arg_1 (cum, mode, type, named, false);
6490 /* Handle the TARGET_FUNCTION_INCOMING_ARG target hook. */
6493 sparc_function_incoming_arg (cumulative_args_t cum, enum machine_mode mode,
6494 const_tree type, bool named)
6496 return sparc_function_arg_1 (cum, mode, type, named, true);
6499 /* For sparc64, objects requiring 16 byte alignment are passed that way. */
6502 sparc_function_arg_boundary (enum machine_mode mode, const_tree type)
6504 return ((TARGET_ARCH64
6505 && (GET_MODE_ALIGNMENT (mode) == 128
6506 || (type && TYPE_ALIGN (type) == 128)))
6511 /* For an arg passed partly in registers and partly in memory,
6512 this is the number of bytes of registers used.
6513 For args passed entirely in registers or entirely in memory, zero.
6515 Any arg that starts in the first 6 regs but won't entirely fit in them
6516 needs partial registers on v8. On v9, structures with integer
6517 values in arg slots 5,6 will be passed in %o5 and SP+176, and complex fp
6518 values that begin in the last fp reg [where "last fp reg" varies with the
6519 mode] will be split between that reg and memory. */
6522 sparc_arg_partial_bytes (cumulative_args_t cum, enum machine_mode mode,
6523 tree type, bool named)
6525 int slotno, regno, padding;
6527 /* We pass false for incoming_p here, it doesn't matter. */
6528 slotno = function_arg_slotno (get_cumulative_args (cum), mode, type, named,
6529 false, ®no, &padding);
6536 if ((slotno + (mode == BLKmode
6537 ? ROUND_ADVANCE (int_size_in_bytes (type))
6538 : ROUND_ADVANCE (GET_MODE_SIZE (mode))))
6539 > SPARC_INT_ARG_MAX)
6540 return (SPARC_INT_ARG_MAX - slotno) * UNITS_PER_WORD;
6544 /* We are guaranteed by pass_by_reference that the size of the
6545 argument is not greater than 16 bytes, so we only need to return
6546 one word if the argument is partially passed in registers. */
6548 if (type && AGGREGATE_TYPE_P (type))
6550 int size = int_size_in_bytes (type);
6552 if (size > UNITS_PER_WORD
6553 && slotno == SPARC_INT_ARG_MAX - 1)
6554 return UNITS_PER_WORD;
6556 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_INT
6557 || (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT
6558 && ! (TARGET_FPU && named)))
6560 /* The complex types are passed as packed types. */
6561 if (GET_MODE_SIZE (mode) > UNITS_PER_WORD
6562 && slotno == SPARC_INT_ARG_MAX - 1)
6563 return UNITS_PER_WORD;
6565 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT)
6567 if ((slotno + GET_MODE_SIZE (mode) / UNITS_PER_WORD)
6569 return UNITS_PER_WORD;
6576 /* Handle the TARGET_PASS_BY_REFERENCE target hook.
6577 Specify whether to pass the argument by reference. */
6580 sparc_pass_by_reference (cumulative_args_t cum ATTRIBUTE_UNUSED,
6581 enum machine_mode mode, const_tree type,
6582 bool named ATTRIBUTE_UNUSED)
6585 /* Original SPARC 32-bit ABI says that structures and unions,
6586 and quad-precision floats are passed by reference. For Pascal,
6587 also pass arrays by reference. All other base types are passed
6590 Extended ABI (as implemented by the Sun compiler) says that all
6591 complex floats are passed by reference. Pass complex integers
6592 in registers up to 8 bytes. More generally, enforce the 2-word
6593 cap for passing arguments in registers.
6595 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6596 integers are passed like floats of the same size, that is in
6597 registers up to 8 bytes. Pass all vector floats by reference
6598 like structure and unions. */
6599 return ((type && (AGGREGATE_TYPE_P (type) || VECTOR_FLOAT_TYPE_P (type)))
6601 /* Catch CDImode, TFmode, DCmode and TCmode. */
6602 || GET_MODE_SIZE (mode) > 8
6604 && TREE_CODE (type) == VECTOR_TYPE
6605 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8));
6607 /* Original SPARC 64-bit ABI says that structures and unions
6608 smaller than 16 bytes are passed in registers, as well as
6609 all other base types.
6611 Extended ABI (as implemented by the Sun compiler) says that
6612 complex floats are passed in registers up to 16 bytes. Pass
6613 all complex integers in registers up to 16 bytes. More generally,
6614 enforce the 2-word cap for passing arguments in registers.
6616 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6617 integers are passed like floats of the same size, that is in
6618 registers (up to 16 bytes). Pass all vector floats like structure
6621 && (AGGREGATE_TYPE_P (type) || TREE_CODE (type) == VECTOR_TYPE)
6622 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 16)
6623 /* Catch CTImode and TCmode. */
6624 || GET_MODE_SIZE (mode) > 16);
6627 /* Handle the TARGET_FUNCTION_ARG_ADVANCE hook.
6628 Update the data in CUM to advance over an argument
6629 of mode MODE and data type TYPE.
6630 TYPE is null for libcalls where that information may not be available. */
6633 sparc_function_arg_advance (cumulative_args_t cum_v, enum machine_mode mode,
6634 const_tree type, bool named)
6636 CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v);
6639 /* We pass false for incoming_p here, it doesn't matter. */
6640 function_arg_slotno (cum, mode, type, named, false, ®no, &padding);
6642 /* If argument requires leading padding, add it. */
6643 cum->words += padding;
6647 cum->words += (mode != BLKmode
6648 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
6649 : ROUND_ADVANCE (int_size_in_bytes (type)));
6653 if (type && AGGREGATE_TYPE_P (type))
6655 int size = int_size_in_bytes (type);
6659 else if (size <= 16)
6661 else /* passed by reference */
6666 cum->words += (mode != BLKmode
6667 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
6668 : ROUND_ADVANCE (int_size_in_bytes (type)));
6673 /* Handle the FUNCTION_ARG_PADDING macro.
6674 For the 64 bit ABI structs are always stored left shifted in their
6678 function_arg_padding (enum machine_mode mode, const_tree type)
6680 if (TARGET_ARCH64 && type != 0 && AGGREGATE_TYPE_P (type))
6683 /* Fall back to the default. */
6684 return DEFAULT_FUNCTION_ARG_PADDING (mode, type);
6687 /* Handle the TARGET_RETURN_IN_MEMORY target hook.
6688 Specify whether to return the return value in memory. */
6691 sparc_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
6694 /* Original SPARC 32-bit ABI says that structures and unions,
6695 and quad-precision floats are returned in memory. All other
6696 base types are returned in registers.
6698 Extended ABI (as implemented by the Sun compiler) says that
6699 all complex floats are returned in registers (8 FP registers
6700 at most for '_Complex long double'). Return all complex integers
6701 in registers (4 at most for '_Complex long long').
6703 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6704 integers are returned like floats of the same size, that is in
6705 registers up to 8 bytes and in memory otherwise. Return all
6706 vector floats in memory like structure and unions; note that
6707 they always have BLKmode like the latter. */
6708 return (TYPE_MODE (type) == BLKmode
6709 || TYPE_MODE (type) == TFmode
6710 || (TREE_CODE (type) == VECTOR_TYPE
6711 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8));
6713 /* Original SPARC 64-bit ABI says that structures and unions
6714 smaller than 32 bytes are returned in registers, as well as
6715 all other base types.
6717 Extended ABI (as implemented by the Sun compiler) says that all
6718 complex floats are returned in registers (8 FP registers at most
6719 for '_Complex long double'). Return all complex integers in
6720 registers (4 at most for '_Complex TItype').
6722 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6723 integers are returned like floats of the same size, that is in
6724 registers. Return all vector floats like structure and unions;
6725 note that they always have BLKmode like the latter. */
6726 return (TYPE_MODE (type) == BLKmode
6727 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 32);
6730 /* Handle the TARGET_STRUCT_VALUE target hook.
6731 Return where to find the structure return value address. */
6734 sparc_struct_value_rtx (tree fndecl, int incoming)
6743 mem = gen_frame_mem (Pmode, plus_constant (Pmode, frame_pointer_rtx,
6744 STRUCT_VALUE_OFFSET));
6746 mem = gen_frame_mem (Pmode, plus_constant (Pmode, stack_pointer_rtx,
6747 STRUCT_VALUE_OFFSET));
6749 /* Only follow the SPARC ABI for fixed-size structure returns.
6750 Variable size structure returns are handled per the normal
6751 procedures in GCC. This is enabled by -mstd-struct-return */
6753 && sparc_std_struct_return
6754 && TYPE_SIZE_UNIT (TREE_TYPE (fndecl))
6755 && TREE_CODE (TYPE_SIZE_UNIT (TREE_TYPE (fndecl))) == INTEGER_CST)
6757 /* We must check and adjust the return address, as it is
6758 optional as to whether the return object is really
6760 rtx ret_reg = gen_rtx_REG (Pmode, 31);
6761 rtx scratch = gen_reg_rtx (SImode);
6762 rtx endlab = gen_label_rtx ();
6764 /* Calculate the return object size */
6765 tree size = TYPE_SIZE_UNIT (TREE_TYPE (fndecl));
6766 rtx size_rtx = GEN_INT (TREE_INT_CST_LOW (size) & 0xfff);
6767 /* Construct a temporary return value */
6769 = assign_stack_local (Pmode, TREE_INT_CST_LOW (size), 0);
6771 /* Implement SPARC 32-bit psABI callee return struct checking:
6773 Fetch the instruction where we will return to and see if
6774 it's an unimp instruction (the most significant 10 bits
6776 emit_move_insn (scratch, gen_rtx_MEM (SImode,
6777 plus_constant (Pmode,
6779 /* Assume the size is valid and pre-adjust */
6780 emit_insn (gen_add3_insn (ret_reg, ret_reg, GEN_INT (4)));
6781 emit_cmp_and_jump_insns (scratch, size_rtx, EQ, const0_rtx, SImode,
6783 emit_insn (gen_sub3_insn (ret_reg, ret_reg, GEN_INT (4)));
6784 /* Write the address of the memory pointed to by temp_val into
6785 the memory pointed to by mem */
6786 emit_move_insn (mem, XEXP (temp_val, 0));
6787 emit_label (endlab);
6794 /* Handle TARGET_FUNCTION_VALUE, and TARGET_LIBCALL_VALUE target hook.
6795 For v9, function return values are subject to the same rules as arguments,
6796 except that up to 32 bytes may be returned in registers. */
6799 sparc_function_value_1 (const_tree type, enum machine_mode mode,
6802 /* Beware that the two values are swapped here wrt function_arg. */
6803 int regbase = (outgoing
6804 ? SPARC_INCOMING_INT_ARG_FIRST
6805 : SPARC_OUTGOING_INT_ARG_FIRST);
6806 enum mode_class mclass = GET_MODE_CLASS (mode);
6809 /* Vector types deserve special treatment because they are polymorphic wrt
6810 their mode, depending upon whether VIS instructions are enabled. */
6811 if (type && TREE_CODE (type) == VECTOR_TYPE)
6813 HOST_WIDE_INT size = int_size_in_bytes (type);
6814 gcc_assert ((TARGET_ARCH32 && size <= 8)
6815 || (TARGET_ARCH64 && size <= 32));
6817 if (mode == BLKmode)
6818 return function_arg_vector_value (size,
6819 SPARC_FP_ARG_FIRST);
6821 mclass = MODE_FLOAT;
6824 if (TARGET_ARCH64 && type)
6826 /* Structures up to 32 bytes in size are returned in registers. */
6827 if (TREE_CODE (type) == RECORD_TYPE)
6829 HOST_WIDE_INT size = int_size_in_bytes (type);
6830 gcc_assert (size <= 32);
6832 return function_arg_record_value (type, mode, 0, 1, regbase);
6835 /* Unions up to 32 bytes in size are returned in integer registers. */
6836 else if (TREE_CODE (type) == UNION_TYPE)
6838 HOST_WIDE_INT size = int_size_in_bytes (type);
6839 gcc_assert (size <= 32);
6841 return function_arg_union_value (size, mode, 0, regbase);
6844 /* Objects that require it are returned in FP registers. */
6845 else if (mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT)
6848 /* All other aggregate types are returned in an integer register in a
6849 mode corresponding to the size of the type. */
6850 else if (AGGREGATE_TYPE_P (type))
6852 /* All other aggregate types are passed in an integer register
6853 in a mode corresponding to the size of the type. */
6854 HOST_WIDE_INT size = int_size_in_bytes (type);
6855 gcc_assert (size <= 32);
6857 mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0);
6859 /* ??? We probably should have made the same ABI change in
6860 3.4.0 as the one we made for unions. The latter was
6861 required by the SCD though, while the former is not
6862 specified, so we favored compatibility and efficiency.
6864 Now we're stuck for aggregates larger than 16 bytes,
6865 because OImode vanished in the meantime. Let's not
6866 try to be unduly clever, and simply follow the ABI
6867 for unions in that case. */
6868 if (mode == BLKmode)
6869 return function_arg_union_value (size, mode, 0, regbase);
6874 /* We should only have pointer and integer types at this point. This
6875 must match sparc_promote_function_mode. */
6876 else if (mclass == MODE_INT && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
6880 /* We should only have pointer and integer types at this point. This must
6881 match sparc_promote_function_mode. */
6882 else if (TARGET_ARCH32
6883 && mclass == MODE_INT
6884 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
6887 if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT) && TARGET_FPU)
6888 regno = SPARC_FP_ARG_FIRST;
6892 return gen_rtx_REG (mode, regno);
6895 /* Handle TARGET_FUNCTION_VALUE.
6896 On the SPARC, the value is found in the first "output" register, but the
6897 called function leaves it in the first "input" register. */
6900 sparc_function_value (const_tree valtype,
6901 const_tree fn_decl_or_type ATTRIBUTE_UNUSED,
6904 return sparc_function_value_1 (valtype, TYPE_MODE (valtype), outgoing);
6907 /* Handle TARGET_LIBCALL_VALUE. */
6910 sparc_libcall_value (enum machine_mode mode,
6911 const_rtx fun ATTRIBUTE_UNUSED)
6913 return sparc_function_value_1 (NULL_TREE, mode, false);
6916 /* Handle FUNCTION_VALUE_REGNO_P.
6917 On the SPARC, the first "output" reg is used for integer values, and the
6918 first floating point register is used for floating point values. */
6921 sparc_function_value_regno_p (const unsigned int regno)
6923 return (regno == 8 || regno == 32);
6926 /* Do what is necessary for `va_start'. We look at the current function
6927 to determine if stdarg or varargs is used and return the address of
6928 the first unnamed parameter. */
6931 sparc_builtin_saveregs (void)
6933 int first_reg = crtl->args.info.words;
6937 for (regno = first_reg; regno < SPARC_INT_ARG_MAX; regno++)
6938 emit_move_insn (gen_rtx_MEM (word_mode,
6939 gen_rtx_PLUS (Pmode,
6941 GEN_INT (FIRST_PARM_OFFSET (0)
6944 gen_rtx_REG (word_mode,
6945 SPARC_INCOMING_INT_ARG_FIRST + regno));
6947 address = gen_rtx_PLUS (Pmode,
6949 GEN_INT (FIRST_PARM_OFFSET (0)
6950 + UNITS_PER_WORD * first_reg));
6955 /* Implement `va_start' for stdarg. */
6958 sparc_va_start (tree valist, rtx nextarg)
6960 nextarg = expand_builtin_saveregs ();
6961 std_expand_builtin_va_start (valist, nextarg);
6964 /* Implement `va_arg' for stdarg. */
6967 sparc_gimplify_va_arg (tree valist, tree type, gimple_seq *pre_p,
6970 HOST_WIDE_INT size, rsize, align;
6973 tree ptrtype = build_pointer_type (type);
6975 if (pass_by_reference (NULL, TYPE_MODE (type), type, false))
6978 size = rsize = UNITS_PER_WORD;
6984 size = int_size_in_bytes (type);
6985 rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD;
6990 /* For SPARC64, objects requiring 16-byte alignment get it. */
6991 if (TYPE_ALIGN (type) >= 2 * (unsigned) BITS_PER_WORD)
6992 align = 2 * UNITS_PER_WORD;
6994 /* SPARC-V9 ABI states that structures up to 16 bytes in size
6995 are left-justified in their slots. */
6996 if (AGGREGATE_TYPE_P (type))
6999 size = rsize = UNITS_PER_WORD;
7009 incr = fold_build_pointer_plus_hwi (incr, align - 1);
7010 incr = fold_convert (sizetype, incr);
7011 incr = fold_build2 (BIT_AND_EXPR, sizetype, incr,
7013 incr = fold_convert (ptr_type_node, incr);
7016 gimplify_expr (&incr, pre_p, post_p, is_gimple_val, fb_rvalue);
7019 if (BYTES_BIG_ENDIAN && size < rsize)
7020 addr = fold_build_pointer_plus_hwi (incr, rsize - size);
7024 addr = fold_convert (build_pointer_type (ptrtype), addr);
7025 addr = build_va_arg_indirect_ref (addr);
7028 /* If the address isn't aligned properly for the type, we need a temporary.
7029 FIXME: This is inefficient, usually we can do this in registers. */
7030 else if (align == 0 && TYPE_ALIGN (type) > BITS_PER_WORD)
7032 tree tmp = create_tmp_var (type, "va_arg_tmp");
7033 tree dest_addr = build_fold_addr_expr (tmp);
7034 tree copy = build_call_expr (builtin_decl_implicit (BUILT_IN_MEMCPY),
7035 3, dest_addr, addr, size_int (rsize));
7036 TREE_ADDRESSABLE (tmp) = 1;
7037 gimplify_and_add (copy, pre_p);
7042 addr = fold_convert (ptrtype, addr);
7044 incr = fold_build_pointer_plus_hwi (incr, rsize);
7045 gimplify_assign (valist, incr, post_p);
7047 return build_va_arg_indirect_ref (addr);
7050 /* Implement the TARGET_VECTOR_MODE_SUPPORTED_P target hook.
7051 Specify whether the vector mode is supported by the hardware. */
7054 sparc_vector_mode_supported_p (enum machine_mode mode)
7056 return TARGET_VIS && VECTOR_MODE_P (mode) ? true : false;
7059 /* Implement the TARGET_VECTORIZE_PREFERRED_SIMD_MODE target hook. */
7061 static enum machine_mode
7062 sparc_preferred_simd_mode (enum machine_mode mode)
7080 /* Return the string to output an unconditional branch to LABEL, which is
7081 the operand number of the label.
7083 DEST is the destination insn (i.e. the label), INSN is the source. */
7086 output_ubranch (rtx dest, int label, rtx insn)
7088 static char string[64];
7089 bool v9_form = false;
7092 if (TARGET_V9 && INSN_ADDRESSES_SET_P ())
7094 int delta = (INSN_ADDRESSES (INSN_UID (dest))
7095 - INSN_ADDRESSES (INSN_UID (insn)));
7096 /* Leave some instructions for "slop". */
7097 if (delta >= -260000 && delta < 260000)
7102 strcpy (string, "ba%*,pt\t%%xcc, ");
7104 strcpy (string, "b%*\t");
7106 p = strchr (string, '\0');
7117 /* Return the string to output a conditional branch to LABEL, which is
7118 the operand number of the label. OP is the conditional expression.
7119 XEXP (OP, 0) is assumed to be a condition code register (integer or
7120 floating point) and its mode specifies what kind of comparison we made.
7122 DEST is the destination insn (i.e. the label), INSN is the source.
7124 REVERSED is nonzero if we should reverse the sense of the comparison.
7126 ANNUL is nonzero if we should generate an annulling branch. */
7129 output_cbranch (rtx op, rtx dest, int label, int reversed, int annul,
7132 static char string[64];
7133 enum rtx_code code = GET_CODE (op);
7134 rtx cc_reg = XEXP (op, 0);
7135 enum machine_mode mode = GET_MODE (cc_reg);
7136 const char *labelno, *branch;
7137 int spaces = 8, far;
7140 /* v9 branches are limited to +-1MB. If it is too far away,
7153 fbne,a,pn %fcc2, .LC29
7161 far = TARGET_V9 && (get_attr_length (insn) >= 3);
7164 /* Reversal of FP compares takes care -- an ordered compare
7165 becomes an unordered compare and vice versa. */
7166 if (mode == CCFPmode || mode == CCFPEmode)
7167 code = reverse_condition_maybe_unordered (code);
7169 code = reverse_condition (code);
7172 /* Start by writing the branch condition. */
7173 if (mode == CCFPmode || mode == CCFPEmode)
7224 /* ??? !v9: FP branches cannot be preceded by another floating point
7225 insn. Because there is currently no concept of pre-delay slots,
7226 we can fix this only by always emitting a nop before a floating
7231 strcpy (string, "nop\n\t");
7232 strcat (string, branch);
7245 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
7257 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
7278 strcpy (string, branch);
7280 spaces -= strlen (branch);
7281 p = strchr (string, '\0');
7283 /* Now add the annulling, the label, and a possible noop. */
7296 if (! far && insn && INSN_ADDRESSES_SET_P ())
7298 int delta = (INSN_ADDRESSES (INSN_UID (dest))
7299 - INSN_ADDRESSES (INSN_UID (insn)));
7300 /* Leave some instructions for "slop". */
7301 if (delta < -260000 || delta >= 260000)
7305 if (mode == CCFPmode || mode == CCFPEmode)
7307 static char v9_fcc_labelno[] = "%%fccX, ";
7308 /* Set the char indicating the number of the fcc reg to use. */
7309 v9_fcc_labelno[5] = REGNO (cc_reg) - SPARC_FIRST_V9_FCC_REG + '0';
7310 labelno = v9_fcc_labelno;
7313 gcc_assert (REGNO (cc_reg) == SPARC_FCC_REG);
7317 else if (mode == CCXmode || mode == CCX_NOOVmode)
7319 labelno = "%%xcc, ";
7324 labelno = "%%icc, ";
7329 if (*labelno && insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
7332 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
7345 strcpy (p, labelno);
7346 p = strchr (p, '\0');
7349 strcpy (p, ".+12\n\t nop\n\tb\t");
7350 /* Skip the next insn if requested or
7351 if we know that it will be a nop. */
7352 if (annul || ! final_sequence)
7366 /* Emit a library call comparison between floating point X and Y.
7367 COMPARISON is the operator to compare with (EQ, NE, GT, etc).
7368 Return the new operator to be used in the comparison sequence.
7370 TARGET_ARCH64 uses _Qp_* functions, which use pointers to TFmode
7371 values as arguments instead of the TFmode registers themselves,
7372 that's why we cannot call emit_float_lib_cmp. */
7375 sparc_emit_float_lib_cmp (rtx x, rtx y, enum rtx_code comparison)
7378 rtx slot0, slot1, result, tem, tem2, libfunc;
7379 enum machine_mode mode;
7380 enum rtx_code new_comparison;
7385 qpfunc = (TARGET_ARCH64 ? "_Qp_feq" : "_Q_feq");
7389 qpfunc = (TARGET_ARCH64 ? "_Qp_fne" : "_Q_fne");
7393 qpfunc = (TARGET_ARCH64 ? "_Qp_fgt" : "_Q_fgt");
7397 qpfunc = (TARGET_ARCH64 ? "_Qp_fge" : "_Q_fge");
7401 qpfunc = (TARGET_ARCH64 ? "_Qp_flt" : "_Q_flt");
7405 qpfunc = (TARGET_ARCH64 ? "_Qp_fle" : "_Q_fle");
7416 qpfunc = (TARGET_ARCH64 ? "_Qp_cmp" : "_Q_cmp");
7427 tree expr = MEM_EXPR (x);
7429 mark_addressable (expr);
7434 slot0 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0);
7435 emit_move_insn (slot0, x);
7440 tree expr = MEM_EXPR (y);
7442 mark_addressable (expr);
7447 slot1 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0);
7448 emit_move_insn (slot1, y);
7451 libfunc = gen_rtx_SYMBOL_REF (Pmode, qpfunc);
7452 emit_library_call (libfunc, LCT_NORMAL,
7454 XEXP (slot0, 0), Pmode,
7455 XEXP (slot1, 0), Pmode);
7460 libfunc = gen_rtx_SYMBOL_REF (Pmode, qpfunc);
7461 emit_library_call (libfunc, LCT_NORMAL,
7463 x, TFmode, y, TFmode);
7468 /* Immediately move the result of the libcall into a pseudo
7469 register so reload doesn't clobber the value if it needs
7470 the return register for a spill reg. */
7471 result = gen_reg_rtx (mode);
7472 emit_move_insn (result, hard_libcall_value (mode, libfunc));
7477 return gen_rtx_NE (VOIDmode, result, const0_rtx);
7480 new_comparison = (comparison == UNORDERED ? EQ : NE);
7481 return gen_rtx_fmt_ee (new_comparison, VOIDmode, result, GEN_INT(3));
7484 new_comparison = (comparison == UNGT ? GT : NE);
7485 return gen_rtx_fmt_ee (new_comparison, VOIDmode, result, const1_rtx);
7487 return gen_rtx_NE (VOIDmode, result, const2_rtx);
7489 tem = gen_reg_rtx (mode);
7491 emit_insn (gen_andsi3 (tem, result, const1_rtx));
7493 emit_insn (gen_anddi3 (tem, result, const1_rtx));
7494 return gen_rtx_NE (VOIDmode, tem, const0_rtx);
7497 tem = gen_reg_rtx (mode);
7499 emit_insn (gen_addsi3 (tem, result, const1_rtx));
7501 emit_insn (gen_adddi3 (tem, result, const1_rtx));
7502 tem2 = gen_reg_rtx (mode);
7504 emit_insn (gen_andsi3 (tem2, tem, const2_rtx));
7506 emit_insn (gen_anddi3 (tem2, tem, const2_rtx));
7507 new_comparison = (comparison == UNEQ ? EQ : NE);
7508 return gen_rtx_fmt_ee (new_comparison, VOIDmode, tem2, const0_rtx);
7514 /* Generate an unsigned DImode to FP conversion. This is the same code
7515 optabs would emit if we didn't have TFmode patterns. */
7518 sparc_emit_floatunsdi (rtx *operands, enum machine_mode mode)
7520 rtx neglab, donelab, i0, i1, f0, in, out;
7523 in = force_reg (DImode, operands[1]);
7524 neglab = gen_label_rtx ();
7525 donelab = gen_label_rtx ();
7526 i0 = gen_reg_rtx (DImode);
7527 i1 = gen_reg_rtx (DImode);
7528 f0 = gen_reg_rtx (mode);
7530 emit_cmp_and_jump_insns (in, const0_rtx, LT, const0_rtx, DImode, 0, neglab);
7532 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_FLOAT (mode, in)));
7533 emit_jump_insn (gen_jump (donelab));
7536 emit_label (neglab);
7538 emit_insn (gen_lshrdi3 (i0, in, const1_rtx));
7539 emit_insn (gen_anddi3 (i1, in, const1_rtx));
7540 emit_insn (gen_iordi3 (i0, i0, i1));
7541 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_FLOAT (mode, i0)));
7542 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_PLUS (mode, f0, f0)));
7544 emit_label (donelab);
7547 /* Generate an FP to unsigned DImode conversion. This is the same code
7548 optabs would emit if we didn't have TFmode patterns. */
7551 sparc_emit_fixunsdi (rtx *operands, enum machine_mode mode)
7553 rtx neglab, donelab, i0, i1, f0, in, out, limit;
7556 in = force_reg (mode, operands[1]);
7557 neglab = gen_label_rtx ();
7558 donelab = gen_label_rtx ();
7559 i0 = gen_reg_rtx (DImode);
7560 i1 = gen_reg_rtx (DImode);
7561 limit = gen_reg_rtx (mode);
7562 f0 = gen_reg_rtx (mode);
7564 emit_move_insn (limit,
7565 CONST_DOUBLE_FROM_REAL_VALUE (
7566 REAL_VALUE_ATOF ("9223372036854775808.0", mode), mode));
7567 emit_cmp_and_jump_insns (in, limit, GE, NULL_RTX, mode, 0, neglab);
7569 emit_insn (gen_rtx_SET (VOIDmode,
7571 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, in))));
7572 emit_jump_insn (gen_jump (donelab));
7575 emit_label (neglab);
7577 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_MINUS (mode, in, limit)));
7578 emit_insn (gen_rtx_SET (VOIDmode,
7580 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, f0))));
7581 emit_insn (gen_movdi (i1, const1_rtx));
7582 emit_insn (gen_ashldi3 (i1, i1, GEN_INT (63)));
7583 emit_insn (gen_xordi3 (out, i0, i1));
7585 emit_label (donelab);
7588 /* Return the string to output a conditional branch to LABEL, testing
7589 register REG. LABEL is the operand number of the label; REG is the
7590 operand number of the reg. OP is the conditional expression. The mode
7591 of REG says what kind of comparison we made.
7593 DEST is the destination insn (i.e. the label), INSN is the source.
7595 REVERSED is nonzero if we should reverse the sense of the comparison.
7597 ANNUL is nonzero if we should generate an annulling branch. */
7600 output_v9branch (rtx op, rtx dest, int reg, int label, int reversed,
7601 int annul, rtx insn)
7603 static char string[64];
7604 enum rtx_code code = GET_CODE (op);
7605 enum machine_mode mode = GET_MODE (XEXP (op, 0));
7610 /* branch on register are limited to +-128KB. If it is too far away,
7623 brgez,a,pn %o1, .LC29
7629 ba,pt %xcc, .LC29 */
7631 far = get_attr_length (insn) >= 3;
7633 /* If not floating-point or if EQ or NE, we can just reverse the code. */
7635 code = reverse_condition (code);
7637 /* Only 64 bit versions of these instructions exist. */
7638 gcc_assert (mode == DImode);
7640 /* Start by writing the branch condition. */
7645 strcpy (string, "brnz");
7649 strcpy (string, "brz");
7653 strcpy (string, "brgez");
7657 strcpy (string, "brlz");
7661 strcpy (string, "brlez");
7665 strcpy (string, "brgz");
7672 p = strchr (string, '\0');
7674 /* Now add the annulling, reg, label, and nop. */
7681 if (insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
7684 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
7689 *p = p < string + 8 ? '\t' : ' ';
7697 int veryfar = 1, delta;
7699 if (INSN_ADDRESSES_SET_P ())
7701 delta = (INSN_ADDRESSES (INSN_UID (dest))
7702 - INSN_ADDRESSES (INSN_UID (insn)));
7703 /* Leave some instructions for "slop". */
7704 if (delta >= -260000 && delta < 260000)
7708 strcpy (p, ".+12\n\t nop\n\t");
7709 /* Skip the next insn if requested or
7710 if we know that it will be a nop. */
7711 if (annul || ! final_sequence)
7721 strcpy (p, "ba,pt\t%%xcc, ");
7735 /* Return 1, if any of the registers of the instruction are %l[0-7] or %o[0-7].
7736 Such instructions cannot be used in the delay slot of return insn on v9.
7737 If TEST is 0, also rename all %i[0-7] registers to their %o[0-7] counterparts.
7741 epilogue_renumber (register rtx *where, int test)
7743 register const char *fmt;
7745 register enum rtx_code code;
7750 code = GET_CODE (*where);
7755 if (REGNO (*where) >= 8 && REGNO (*where) < 24) /* oX or lX */
7757 if (! test && REGNO (*where) >= 24 && REGNO (*where) < 32)
7758 *where = gen_rtx_REG (GET_MODE (*where), OUTGOING_REGNO (REGNO(*where)));
7766 /* Do not replace the frame pointer with the stack pointer because
7767 it can cause the delayed instruction to load below the stack.
7768 This occurs when instructions like:
7770 (set (reg/i:SI 24 %i0)
7771 (mem/f:SI (plus:SI (reg/f:SI 30 %fp)
7772 (const_int -20 [0xffffffec])) 0))
7774 are in the return delayed slot. */
7776 if (GET_CODE (XEXP (*where, 0)) == REG
7777 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM
7778 && (GET_CODE (XEXP (*where, 1)) != CONST_INT
7779 || INTVAL (XEXP (*where, 1)) < SPARC_STACK_BIAS))
7784 if (SPARC_STACK_BIAS
7785 && GET_CODE (XEXP (*where, 0)) == REG
7786 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM)
7794 fmt = GET_RTX_FORMAT (code);
7796 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7801 for (j = XVECLEN (*where, i) - 1; j >= 0; j--)
7802 if (epilogue_renumber (&(XVECEXP (*where, i, j)), test))
7805 else if (fmt[i] == 'e'
7806 && epilogue_renumber (&(XEXP (*where, i)), test))
7812 /* Leaf functions and non-leaf functions have different needs. */
7815 reg_leaf_alloc_order[] = REG_LEAF_ALLOC_ORDER;
7818 reg_nonleaf_alloc_order[] = REG_ALLOC_ORDER;
7820 static const int *const reg_alloc_orders[] = {
7821 reg_leaf_alloc_order,
7822 reg_nonleaf_alloc_order};
7825 order_regs_for_local_alloc (void)
7827 static int last_order_nonleaf = 1;
7829 if (df_regs_ever_live_p (15) != last_order_nonleaf)
7831 last_order_nonleaf = !last_order_nonleaf;
7832 memcpy ((char *) reg_alloc_order,
7833 (const char *) reg_alloc_orders[last_order_nonleaf],
7834 FIRST_PSEUDO_REGISTER * sizeof (int));
7838 /* Return 1 if REG and MEM are legitimate enough to allow the various
7839 mem<-->reg splits to be run. */
7842 sparc_splitdi_legitimate (rtx reg, rtx mem)
7844 /* Punt if we are here by mistake. */
7845 gcc_assert (reload_completed);
7847 /* We must have an offsettable memory reference. */
7848 if (! offsettable_memref_p (mem))
7851 /* If we have legitimate args for ldd/std, we do not want
7852 the split to happen. */
7853 if ((REGNO (reg) % 2) == 0
7854 && mem_min_alignment (mem, 8))
7861 /* Like sparc_splitdi_legitimate but for REG <--> REG moves. */
7864 sparc_split_regreg_legitimate (rtx reg1, rtx reg2)
7868 if (GET_CODE (reg1) == SUBREG)
7869 reg1 = SUBREG_REG (reg1);
7870 if (GET_CODE (reg1) != REG)
7872 regno1 = REGNO (reg1);
7874 if (GET_CODE (reg2) == SUBREG)
7875 reg2 = SUBREG_REG (reg2);
7876 if (GET_CODE (reg2) != REG)
7878 regno2 = REGNO (reg2);
7880 if (SPARC_INT_REG_P (regno1) && SPARC_INT_REG_P (regno2))
7885 if ((SPARC_INT_REG_P (regno1) && SPARC_FP_REG_P (regno2))
7886 || (SPARC_FP_REG_P (regno1) && SPARC_INT_REG_P (regno2)))
7893 /* Return 1 if x and y are some kind of REG and they refer to
7894 different hard registers. This test is guaranteed to be
7895 run after reload. */
7898 sparc_absnegfloat_split_legitimate (rtx x, rtx y)
7900 if (GET_CODE (x) != REG)
7902 if (GET_CODE (y) != REG)
7904 if (REGNO (x) == REGNO (y))
7909 /* Return 1 if REGNO (reg1) is even and REGNO (reg1) == REGNO (reg2) - 1.
7910 This makes them candidates for using ldd and std insns.
7912 Note reg1 and reg2 *must* be hard registers. */
7915 registers_ok_for_ldd_peep (rtx reg1, rtx reg2)
7917 /* We might have been passed a SUBREG. */
7918 if (GET_CODE (reg1) != REG || GET_CODE (reg2) != REG)
7921 if (REGNO (reg1) % 2 != 0)
7924 /* Integer ldd is deprecated in SPARC V9 */
7925 if (TARGET_V9 && SPARC_INT_REG_P (REGNO (reg1)))
7928 return (REGNO (reg1) == REGNO (reg2) - 1);
7931 /* Return 1 if the addresses in mem1 and mem2 are suitable for use in
7934 This can only happen when addr1 and addr2, the addresses in mem1
7935 and mem2, are consecutive memory locations (addr1 + 4 == addr2).
7936 addr1 must also be aligned on a 64-bit boundary.
7938 Also iff dependent_reg_rtx is not null it should not be used to
7939 compute the address for mem1, i.e. we cannot optimize a sequence
7951 But, note that the transformation from:
7956 is perfectly fine. Thus, the peephole2 patterns always pass us
7957 the destination register of the first load, never the second one.
7959 For stores we don't have a similar problem, so dependent_reg_rtx is
7963 mems_ok_for_ldd_peep (rtx mem1, rtx mem2, rtx dependent_reg_rtx)
7967 HOST_WIDE_INT offset1;
7969 /* The mems cannot be volatile. */
7970 if (MEM_VOLATILE_P (mem1) || MEM_VOLATILE_P (mem2))
7973 /* MEM1 should be aligned on a 64-bit boundary. */
7974 if (MEM_ALIGN (mem1) < 64)
7977 addr1 = XEXP (mem1, 0);
7978 addr2 = XEXP (mem2, 0);
7980 /* Extract a register number and offset (if used) from the first addr. */
7981 if (GET_CODE (addr1) == PLUS)
7983 /* If not a REG, return zero. */
7984 if (GET_CODE (XEXP (addr1, 0)) != REG)
7988 reg1 = REGNO (XEXP (addr1, 0));
7989 /* The offset must be constant! */
7990 if (GET_CODE (XEXP (addr1, 1)) != CONST_INT)
7992 offset1 = INTVAL (XEXP (addr1, 1));
7995 else if (GET_CODE (addr1) != REG)
7999 reg1 = REGNO (addr1);
8000 /* This was a simple (mem (reg)) expression. Offset is 0. */
8004 /* Make sure the second address is a (mem (plus (reg) (const_int). */
8005 if (GET_CODE (addr2) != PLUS)
8008 if (GET_CODE (XEXP (addr2, 0)) != REG
8009 || GET_CODE (XEXP (addr2, 1)) != CONST_INT)
8012 if (reg1 != REGNO (XEXP (addr2, 0)))
8015 if (dependent_reg_rtx != NULL_RTX && reg1 == REGNO (dependent_reg_rtx))
8018 /* The first offset must be evenly divisible by 8 to ensure the
8019 address is 64 bit aligned. */
8020 if (offset1 % 8 != 0)
8023 /* The offset for the second addr must be 4 more than the first addr. */
8024 if (INTVAL (XEXP (addr2, 1)) != offset1 + 4)
8027 /* All the tests passed. addr1 and addr2 are valid for ldd and std
8032 /* Return 1 if reg is a pseudo, or is the first register in
8033 a hard register pair. This makes it suitable for use in
8034 ldd and std insns. */
8037 register_ok_for_ldd (rtx reg)
8039 /* We might have been passed a SUBREG. */
8043 if (REGNO (reg) < FIRST_PSEUDO_REGISTER)
8044 return (REGNO (reg) % 2 == 0);
8049 /* Return 1 if OP is a memory whose address is known to be
8050 aligned to 8-byte boundary, or a pseudo during reload.
8051 This makes it suitable for use in ldd and std insns. */
8054 memory_ok_for_ldd (rtx op)
8058 /* In 64-bit mode, we assume that the address is word-aligned. */
8059 if (TARGET_ARCH32 && !mem_min_alignment (op, 8))
8062 if (! can_create_pseudo_p ()
8063 && !strict_memory_address_p (Pmode, XEXP (op, 0)))
8066 else if (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER)
8068 if (!(reload_in_progress && reg_renumber [REGNO (op)] < 0))
8077 /* Implement TARGET_PRINT_OPERAND_PUNCT_VALID_P. */
8080 sparc_print_operand_punct_valid_p (unsigned char code)
8093 /* Implement TARGET_PRINT_OPERAND.
8094 Print operand X (an rtx) in assembler syntax to file FILE.
8095 CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified.
8096 For `%' followed by punctuation, CODE is the punctuation and X is null. */
8099 sparc_print_operand (FILE *file, rtx x, int code)
8104 /* Output an insn in a delay slot. */
8106 sparc_indent_opcode = 1;
8108 fputs ("\n\t nop", file);
8111 /* Output an annul flag if there's nothing for the delay slot and we
8112 are optimizing. This is always used with '(' below.
8113 Sun OS 4.1.1 dbx can't handle an annulled unconditional branch;
8114 this is a dbx bug. So, we only do this when optimizing.
8115 On UltraSPARC, a branch in a delay slot causes a pipeline flush.
8116 Always emit a nop in case the next instruction is a branch. */
8117 if (! final_sequence && (optimize && (int)sparc_cpu < PROCESSOR_V9))
8121 /* Output a 'nop' if there's nothing for the delay slot and we are
8122 not optimizing. This is always used with '*' above. */
8123 if (! final_sequence && ! (optimize && (int)sparc_cpu < PROCESSOR_V9))
8124 fputs ("\n\t nop", file);
8125 else if (final_sequence)
8126 sparc_indent_opcode = 1;
8129 /* Output the right displacement from the saved PC on function return.
8130 The caller may have placed an "unimp" insn immediately after the call
8131 so we have to account for it. This insn is used in the 32-bit ABI
8132 when calling a function that returns a non zero-sized structure. The
8133 64-bit ABI doesn't have it. Be careful to have this test be the same
8134 as that for the call. The exception is when sparc_std_struct_return
8135 is enabled, the psABI is followed exactly and the adjustment is made
8136 by the code in sparc_struct_value_rtx. The call emitted is the same
8137 when sparc_std_struct_return is enabled. */
8139 && cfun->returns_struct
8140 && !sparc_std_struct_return
8141 && DECL_SIZE (DECL_RESULT (current_function_decl))
8142 && TREE_CODE (DECL_SIZE (DECL_RESULT (current_function_decl)))
8144 && !integer_zerop (DECL_SIZE (DECL_RESULT (current_function_decl))))
8150 /* Output the Embedded Medium/Anywhere code model base register. */
8151 fputs (EMBMEDANY_BASE_REG, file);
8154 /* Print some local dynamic TLS name. */
8155 assemble_name (file, get_some_local_dynamic_name ());
8159 /* Adjust the operand to take into account a RESTORE operation. */
8160 if (GET_CODE (x) == CONST_INT)
8162 else if (GET_CODE (x) != REG)
8163 output_operand_lossage ("invalid %%Y operand");
8164 else if (REGNO (x) < 8)
8165 fputs (reg_names[REGNO (x)], file);
8166 else if (REGNO (x) >= 24 && REGNO (x) < 32)
8167 fputs (reg_names[REGNO (x)-16], file);
8169 output_operand_lossage ("invalid %%Y operand");
8172 /* Print out the low order register name of a register pair. */
8173 if (WORDS_BIG_ENDIAN)
8174 fputs (reg_names[REGNO (x)+1], file);
8176 fputs (reg_names[REGNO (x)], file);
8179 /* Print out the high order register name of a register pair. */
8180 if (WORDS_BIG_ENDIAN)
8181 fputs (reg_names[REGNO (x)], file);
8183 fputs (reg_names[REGNO (x)+1], file);
8186 /* Print out the second register name of a register pair or quad.
8187 I.e., R (%o0) => %o1. */
8188 fputs (reg_names[REGNO (x)+1], file);
8191 /* Print out the third register name of a register quad.
8192 I.e., S (%o0) => %o2. */
8193 fputs (reg_names[REGNO (x)+2], file);
8196 /* Print out the fourth register name of a register quad.
8197 I.e., T (%o0) => %o3. */
8198 fputs (reg_names[REGNO (x)+3], file);
8201 /* Print a condition code register. */
8202 if (REGNO (x) == SPARC_ICC_REG)
8204 /* We don't handle CC[X]_NOOVmode because they're not supposed
8206 if (GET_MODE (x) == CCmode)
8207 fputs ("%icc", file);
8208 else if (GET_MODE (x) == CCXmode)
8209 fputs ("%xcc", file);
8214 /* %fccN register */
8215 fputs (reg_names[REGNO (x)], file);
8218 /* Print the operand's address only. */
8219 output_address (XEXP (x, 0));
8222 /* In this case we need a register. Use %g0 if the
8223 operand is const0_rtx. */
8225 || (GET_MODE (x) != VOIDmode && x == CONST0_RTX (GET_MODE (x))))
8227 fputs ("%g0", file);
8234 switch (GET_CODE (x))
8236 case IOR: fputs ("or", file); break;
8237 case AND: fputs ("and", file); break;
8238 case XOR: fputs ("xor", file); break;
8239 default: output_operand_lossage ("invalid %%A operand");
8244 switch (GET_CODE (x))
8246 case IOR: fputs ("orn", file); break;
8247 case AND: fputs ("andn", file); break;
8248 case XOR: fputs ("xnor", file); break;
8249 default: output_operand_lossage ("invalid %%B operand");
8253 /* This is used by the conditional move instructions. */
8256 enum rtx_code rc = GET_CODE (x);
8260 case NE: fputs ("ne", file); break;
8261 case EQ: fputs ("e", file); break;
8262 case GE: fputs ("ge", file); break;
8263 case GT: fputs ("g", file); break;
8264 case LE: fputs ("le", file); break;
8265 case LT: fputs ("l", file); break;
8266 case GEU: fputs ("geu", file); break;
8267 case GTU: fputs ("gu", file); break;
8268 case LEU: fputs ("leu", file); break;
8269 case LTU: fputs ("lu", file); break;
8270 case LTGT: fputs ("lg", file); break;
8271 case UNORDERED: fputs ("u", file); break;
8272 case ORDERED: fputs ("o", file); break;
8273 case UNLT: fputs ("ul", file); break;
8274 case UNLE: fputs ("ule", file); break;
8275 case UNGT: fputs ("ug", file); break;
8276 case UNGE: fputs ("uge", file); break;
8277 case UNEQ: fputs ("ue", file); break;
8278 default: output_operand_lossage ("invalid %%C operand");
8283 /* This are used by the movr instruction pattern. */
8286 enum rtx_code rc = GET_CODE (x);
8289 case NE: fputs ("ne", file); break;
8290 case EQ: fputs ("e", file); break;
8291 case GE: fputs ("gez", file); break;
8292 case LT: fputs ("lz", file); break;
8293 case LE: fputs ("lez", file); break;
8294 case GT: fputs ("gz", file); break;
8295 default: output_operand_lossage ("invalid %%D operand");
8302 /* Print a sign-extended character. */
8303 int i = trunc_int_for_mode (INTVAL (x), QImode);
8304 fprintf (file, "%d", i);
8309 /* Operand must be a MEM; write its address. */
8310 if (GET_CODE (x) != MEM)
8311 output_operand_lossage ("invalid %%f operand");
8312 output_address (XEXP (x, 0));
8317 /* Print a sign-extended 32-bit value. */
8319 if (GET_CODE(x) == CONST_INT)
8321 else if (GET_CODE(x) == CONST_DOUBLE)
8322 i = CONST_DOUBLE_LOW (x);
8325 output_operand_lossage ("invalid %%s operand");
8328 i = trunc_int_for_mode (i, SImode);
8329 fprintf (file, HOST_WIDE_INT_PRINT_DEC, i);
8334 /* Do nothing special. */
8338 /* Undocumented flag. */
8339 output_operand_lossage ("invalid operand output code");
8342 if (GET_CODE (x) == REG)
8343 fputs (reg_names[REGNO (x)], file);
8344 else if (GET_CODE (x) == MEM)
8347 /* Poor Sun assembler doesn't understand absolute addressing. */
8348 if (CONSTANT_P (XEXP (x, 0)))
8349 fputs ("%g0+", file);
8350 output_address (XEXP (x, 0));
8353 else if (GET_CODE (x) == HIGH)
8355 fputs ("%hi(", file);
8356 output_addr_const (file, XEXP (x, 0));
8359 else if (GET_CODE (x) == LO_SUM)
8361 sparc_print_operand (file, XEXP (x, 0), 0);
8362 if (TARGET_CM_MEDMID)
8363 fputs ("+%l44(", file);
8365 fputs ("+%lo(", file);
8366 output_addr_const (file, XEXP (x, 1));
8369 else if (GET_CODE (x) == CONST_DOUBLE
8370 && (GET_MODE (x) == VOIDmode
8371 || GET_MODE_CLASS (GET_MODE (x)) == MODE_INT))
8373 if (CONST_DOUBLE_HIGH (x) == 0)
8374 fprintf (file, "%u", (unsigned int) CONST_DOUBLE_LOW (x));
8375 else if (CONST_DOUBLE_HIGH (x) == -1
8376 && CONST_DOUBLE_LOW (x) < 0)
8377 fprintf (file, "%d", (int) CONST_DOUBLE_LOW (x));
8379 output_operand_lossage ("long long constant not a valid immediate operand");
8381 else if (GET_CODE (x) == CONST_DOUBLE)
8382 output_operand_lossage ("floating point constant not a valid immediate operand");
8383 else { output_addr_const (file, x); }
8386 /* Implement TARGET_PRINT_OPERAND_ADDRESS. */
8389 sparc_print_operand_address (FILE *file, rtx x)
8391 register rtx base, index = 0;
8393 register rtx addr = x;
8396 fputs (reg_names[REGNO (addr)], file);
8397 else if (GET_CODE (addr) == PLUS)
8399 if (CONST_INT_P (XEXP (addr, 0)))
8400 offset = INTVAL (XEXP (addr, 0)), base = XEXP (addr, 1);
8401 else if (CONST_INT_P (XEXP (addr, 1)))
8402 offset = INTVAL (XEXP (addr, 1)), base = XEXP (addr, 0);
8404 base = XEXP (addr, 0), index = XEXP (addr, 1);
8405 if (GET_CODE (base) == LO_SUM)
8407 gcc_assert (USE_AS_OFFSETABLE_LO10
8409 && ! TARGET_CM_MEDMID);
8410 output_operand (XEXP (base, 0), 0);
8411 fputs ("+%lo(", file);
8412 output_address (XEXP (base, 1));
8413 fprintf (file, ")+%d", offset);
8417 fputs (reg_names[REGNO (base)], file);
8419 fprintf (file, "%+d", offset);
8420 else if (REG_P (index))
8421 fprintf (file, "+%s", reg_names[REGNO (index)]);
8422 else if (GET_CODE (index) == SYMBOL_REF
8423 || GET_CODE (index) == LABEL_REF
8424 || GET_CODE (index) == CONST)
8425 fputc ('+', file), output_addr_const (file, index);
8426 else gcc_unreachable ();
8429 else if (GET_CODE (addr) == MINUS
8430 && GET_CODE (XEXP (addr, 1)) == LABEL_REF)
8432 output_addr_const (file, XEXP (addr, 0));
8434 output_addr_const (file, XEXP (addr, 1));
8435 fputs ("-.)", file);
8437 else if (GET_CODE (addr) == LO_SUM)
8439 output_operand (XEXP (addr, 0), 0);
8440 if (TARGET_CM_MEDMID)
8441 fputs ("+%l44(", file);
8443 fputs ("+%lo(", file);
8444 output_address (XEXP (addr, 1));
8448 && GET_CODE (addr) == CONST
8449 && GET_CODE (XEXP (addr, 0)) == MINUS
8450 && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST
8451 && GET_CODE (XEXP (XEXP (XEXP (addr, 0), 1), 0)) == MINUS
8452 && XEXP (XEXP (XEXP (XEXP (addr, 0), 1), 0), 1) == pc_rtx)
8454 addr = XEXP (addr, 0);
8455 output_addr_const (file, XEXP (addr, 0));
8456 /* Group the args of the second CONST in parenthesis. */
8458 /* Skip past the second CONST--it does nothing for us. */
8459 output_addr_const (file, XEXP (XEXP (addr, 1), 0));
8460 /* Close the parenthesis. */
8465 output_addr_const (file, addr);
8469 /* Target hook for assembling integer objects. The sparc version has
8470 special handling for aligned DI-mode objects. */
8473 sparc_assemble_integer (rtx x, unsigned int size, int aligned_p)
8475 /* ??? We only output .xword's for symbols and only then in environments
8476 where the assembler can handle them. */
8477 if (aligned_p && size == 8
8478 && (GET_CODE (x) != CONST_INT && GET_CODE (x) != CONST_DOUBLE))
8482 assemble_integer_with_op ("\t.xword\t", x);
8487 assemble_aligned_integer (4, const0_rtx);
8488 assemble_aligned_integer (4, x);
8492 return default_assemble_integer (x, size, aligned_p);
8495 /* Return the value of a code used in the .proc pseudo-op that says
8496 what kind of result this function returns. For non-C types, we pick
8497 the closest C type. */
8499 #ifndef SHORT_TYPE_SIZE
8500 #define SHORT_TYPE_SIZE (BITS_PER_UNIT * 2)
8503 #ifndef INT_TYPE_SIZE
8504 #define INT_TYPE_SIZE BITS_PER_WORD
8507 #ifndef LONG_TYPE_SIZE
8508 #define LONG_TYPE_SIZE BITS_PER_WORD
8511 #ifndef LONG_LONG_TYPE_SIZE
8512 #define LONG_LONG_TYPE_SIZE (BITS_PER_WORD * 2)
8515 #ifndef FLOAT_TYPE_SIZE
8516 #define FLOAT_TYPE_SIZE BITS_PER_WORD
8519 #ifndef DOUBLE_TYPE_SIZE
8520 #define DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
8523 #ifndef LONG_DOUBLE_TYPE_SIZE
8524 #define LONG_DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
8528 sparc_type_code (register tree type)
8530 register unsigned long qualifiers = 0;
8531 register unsigned shift;
8533 /* Only the first 30 bits of the qualifier are valid. We must refrain from
8534 setting more, since some assemblers will give an error for this. Also,
8535 we must be careful to avoid shifts of 32 bits or more to avoid getting
8536 unpredictable results. */
8538 for (shift = 6; shift < 30; shift += 2, type = TREE_TYPE (type))
8540 switch (TREE_CODE (type))
8546 qualifiers |= (3 << shift);
8551 qualifiers |= (2 << shift);
8555 case REFERENCE_TYPE:
8557 qualifiers |= (1 << shift);
8561 return (qualifiers | 8);
8564 case QUAL_UNION_TYPE:
8565 return (qualifiers | 9);
8568 return (qualifiers | 10);
8571 return (qualifiers | 16);
8574 /* If this is a range type, consider it to be the underlying
8576 if (TREE_TYPE (type) != 0)
8579 /* Carefully distinguish all the standard types of C,
8580 without messing up if the language is not C. We do this by
8581 testing TYPE_PRECISION and TYPE_UNSIGNED. The old code used to
8582 look at both the names and the above fields, but that's redundant.
8583 Any type whose size is between two C types will be considered
8584 to be the wider of the two types. Also, we do not have a
8585 special code to use for "long long", so anything wider than
8586 long is treated the same. Note that we can't distinguish
8587 between "int" and "long" in this code if they are the same
8588 size, but that's fine, since neither can the assembler. */
8590 if (TYPE_PRECISION (type) <= CHAR_TYPE_SIZE)
8591 return (qualifiers | (TYPE_UNSIGNED (type) ? 12 : 2));
8593 else if (TYPE_PRECISION (type) <= SHORT_TYPE_SIZE)
8594 return (qualifiers | (TYPE_UNSIGNED (type) ? 13 : 3));
8596 else if (TYPE_PRECISION (type) <= INT_TYPE_SIZE)
8597 return (qualifiers | (TYPE_UNSIGNED (type) ? 14 : 4));
8600 return (qualifiers | (TYPE_UNSIGNED (type) ? 15 : 5));
8603 /* If this is a range type, consider it to be the underlying
8605 if (TREE_TYPE (type) != 0)
8608 /* Carefully distinguish all the standard types of C,
8609 without messing up if the language is not C. */
8611 if (TYPE_PRECISION (type) == FLOAT_TYPE_SIZE)
8612 return (qualifiers | 6);
8615 return (qualifiers | 7);
8617 case COMPLEX_TYPE: /* GNU Fortran COMPLEX type. */
8618 /* ??? We need to distinguish between double and float complex types,
8619 but I don't know how yet because I can't reach this code from
8620 existing front-ends. */
8621 return (qualifiers | 7); /* Who knows? */
8624 case BOOLEAN_TYPE: /* Boolean truth value type. */
8630 gcc_unreachable (); /* Not a type! */
8637 /* Nested function support. */
8639 /* Emit RTL insns to initialize the variable parts of a trampoline.
8640 FNADDR is an RTX for the address of the function's pure code.
8641 CXT is an RTX for the static chain value for the function.
8643 This takes 16 insns: 2 shifts & 2 ands (to split up addresses), 4 sethi
8644 (to load in opcodes), 4 iors (to merge address and opcodes), and 4 writes
8645 (to store insns). This is a bit excessive. Perhaps a different
8646 mechanism would be better here.
8648 Emit enough FLUSH insns to synchronize the data and instruction caches. */
8651 sparc32_initialize_trampoline (rtx m_tramp, rtx fnaddr, rtx cxt)
8653 /* SPARC 32-bit trampoline:
8656 sethi %hi(static), %g2
8658 or %g2, %lo(static), %g2
8660 SETHI i,r = 00rr rrr1 00ii iiii iiii iiii iiii iiii
8661 JMPL r+i,d = 10dd ddd1 1100 0rrr rr1i iiii iiii iiii
8665 (adjust_address (m_tramp, SImode, 0),
8666 expand_binop (SImode, ior_optab,
8667 expand_shift (RSHIFT_EXPR, SImode, fnaddr, 10, 0, 1),
8668 GEN_INT (trunc_int_for_mode (0x03000000, SImode)),
8669 NULL_RTX, 1, OPTAB_DIRECT));
8672 (adjust_address (m_tramp, SImode, 4),
8673 expand_binop (SImode, ior_optab,
8674 expand_shift (RSHIFT_EXPR, SImode, cxt, 10, 0, 1),
8675 GEN_INT (trunc_int_for_mode (0x05000000, SImode)),
8676 NULL_RTX, 1, OPTAB_DIRECT));
8679 (adjust_address (m_tramp, SImode, 8),
8680 expand_binop (SImode, ior_optab,
8681 expand_and (SImode, fnaddr, GEN_INT (0x3ff), NULL_RTX),
8682 GEN_INT (trunc_int_for_mode (0x81c06000, SImode)),
8683 NULL_RTX, 1, OPTAB_DIRECT));
8686 (adjust_address (m_tramp, SImode, 12),
8687 expand_binop (SImode, ior_optab,
8688 expand_and (SImode, cxt, GEN_INT (0x3ff), NULL_RTX),
8689 GEN_INT (trunc_int_for_mode (0x8410a000, SImode)),
8690 NULL_RTX, 1, OPTAB_DIRECT));
8692 /* On UltraSPARC a flush flushes an entire cache line. The trampoline is
8693 aligned on a 16 byte boundary so one flush clears it all. */
8694 emit_insn (gen_flush (validize_mem (adjust_address (m_tramp, SImode, 0))));
8695 if (sparc_cpu != PROCESSOR_ULTRASPARC
8696 && sparc_cpu != PROCESSOR_ULTRASPARC3
8697 && sparc_cpu != PROCESSOR_NIAGARA
8698 && sparc_cpu != PROCESSOR_NIAGARA2
8699 && sparc_cpu != PROCESSOR_NIAGARA3
8700 && sparc_cpu != PROCESSOR_NIAGARA4)
8701 emit_insn (gen_flush (validize_mem (adjust_address (m_tramp, SImode, 8))));
8703 /* Call __enable_execute_stack after writing onto the stack to make sure
8704 the stack address is accessible. */
8705 #ifdef HAVE_ENABLE_EXECUTE_STACK
8706 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
8707 LCT_NORMAL, VOIDmode, 1, XEXP (m_tramp, 0), Pmode);
8712 /* The 64-bit version is simpler because it makes more sense to load the
8713 values as "immediate" data out of the trampoline. It's also easier since
8714 we can read the PC without clobbering a register. */
8717 sparc64_initialize_trampoline (rtx m_tramp, rtx fnaddr, rtx cxt)
8719 /* SPARC 64-bit trampoline:
8728 emit_move_insn (adjust_address (m_tramp, SImode, 0),
8729 GEN_INT (trunc_int_for_mode (0x83414000, SImode)));
8730 emit_move_insn (adjust_address (m_tramp, SImode, 4),
8731 GEN_INT (trunc_int_for_mode (0xca586018, SImode)));
8732 emit_move_insn (adjust_address (m_tramp, SImode, 8),
8733 GEN_INT (trunc_int_for_mode (0x81c14000, SImode)));
8734 emit_move_insn (adjust_address (m_tramp, SImode, 12),
8735 GEN_INT (trunc_int_for_mode (0xca586010, SImode)));
8736 emit_move_insn (adjust_address (m_tramp, DImode, 16), cxt);
8737 emit_move_insn (adjust_address (m_tramp, DImode, 24), fnaddr);
8738 emit_insn (gen_flushdi (validize_mem (adjust_address (m_tramp, DImode, 0))));
8740 if (sparc_cpu != PROCESSOR_ULTRASPARC
8741 && sparc_cpu != PROCESSOR_ULTRASPARC3
8742 && sparc_cpu != PROCESSOR_NIAGARA
8743 && sparc_cpu != PROCESSOR_NIAGARA2
8744 && sparc_cpu != PROCESSOR_NIAGARA3
8745 && sparc_cpu != PROCESSOR_NIAGARA4)
8746 emit_insn (gen_flushdi (validize_mem (adjust_address (m_tramp, DImode, 8))));
8748 /* Call __enable_execute_stack after writing onto the stack to make sure
8749 the stack address is accessible. */
8750 #ifdef HAVE_ENABLE_EXECUTE_STACK
8751 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
8752 LCT_NORMAL, VOIDmode, 1, XEXP (m_tramp, 0), Pmode);
8756 /* Worker for TARGET_TRAMPOLINE_INIT. */
8759 sparc_trampoline_init (rtx m_tramp, tree fndecl, rtx cxt)
8761 rtx fnaddr = force_reg (Pmode, XEXP (DECL_RTL (fndecl), 0));
8762 cxt = force_reg (Pmode, cxt);
8764 sparc64_initialize_trampoline (m_tramp, fnaddr, cxt);
8766 sparc32_initialize_trampoline (m_tramp, fnaddr, cxt);
8769 /* Adjust the cost of a scheduling dependency. Return the new cost of
8770 a dependency LINK or INSN on DEP_INSN. COST is the current cost. */
8773 supersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
8775 enum attr_type insn_type;
8777 if (! recog_memoized (insn))
8780 insn_type = get_attr_type (insn);
8782 if (REG_NOTE_KIND (link) == 0)
8784 /* Data dependency; DEP_INSN writes a register that INSN reads some
8787 /* if a load, then the dependence must be on the memory address;
8788 add an extra "cycle". Note that the cost could be two cycles
8789 if the reg was written late in an instruction group; we ca not tell
8791 if (insn_type == TYPE_LOAD || insn_type == TYPE_FPLOAD)
8794 /* Get the delay only if the address of the store is the dependence. */
8795 if (insn_type == TYPE_STORE || insn_type == TYPE_FPSTORE)
8797 rtx pat = PATTERN(insn);
8798 rtx dep_pat = PATTERN (dep_insn);
8800 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
8801 return cost; /* This should not happen! */
8803 /* The dependency between the two instructions was on the data that
8804 is being stored. Assume that this implies that the address of the
8805 store is not dependent. */
8806 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
8809 return cost + 3; /* An approximation. */
8812 /* A shift instruction cannot receive its data from an instruction
8813 in the same cycle; add a one cycle penalty. */
8814 if (insn_type == TYPE_SHIFT)
8815 return cost + 3; /* Split before cascade into shift. */
8819 /* Anti- or output- dependency; DEP_INSN reads/writes a register that
8820 INSN writes some cycles later. */
8822 /* These are only significant for the fpu unit; writing a fp reg before
8823 the fpu has finished with it stalls the processor. */
8825 /* Reusing an integer register causes no problems. */
8826 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
8834 hypersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
8836 enum attr_type insn_type, dep_type;
8837 rtx pat = PATTERN(insn);
8838 rtx dep_pat = PATTERN (dep_insn);
8840 if (recog_memoized (insn) < 0 || recog_memoized (dep_insn) < 0)
8843 insn_type = get_attr_type (insn);
8844 dep_type = get_attr_type (dep_insn);
8846 switch (REG_NOTE_KIND (link))
8849 /* Data dependency; DEP_INSN writes a register that INSN reads some
8856 /* Get the delay iff the address of the store is the dependence. */
8857 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
8860 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
8867 /* If a load, then the dependence must be on the memory address. If
8868 the addresses aren't equal, then it might be a false dependency */
8869 if (dep_type == TYPE_STORE || dep_type == TYPE_FPSTORE)
8871 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET
8872 || GET_CODE (SET_DEST (dep_pat)) != MEM
8873 || GET_CODE (SET_SRC (pat)) != MEM
8874 || ! rtx_equal_p (XEXP (SET_DEST (dep_pat), 0),
8875 XEXP (SET_SRC (pat), 0)))
8883 /* Compare to branch latency is 0. There is no benefit from
8884 separating compare and branch. */
8885 if (dep_type == TYPE_COMPARE)
8887 /* Floating point compare to branch latency is less than
8888 compare to conditional move. */
8889 if (dep_type == TYPE_FPCMP)
8898 /* Anti-dependencies only penalize the fpu unit. */
8899 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
8911 sparc_adjust_cost(rtx insn, rtx link, rtx dep, int cost)
8915 case PROCESSOR_SUPERSPARC:
8916 cost = supersparc_adjust_cost (insn, link, dep, cost);
8918 case PROCESSOR_HYPERSPARC:
8919 case PROCESSOR_SPARCLITE86X:
8920 cost = hypersparc_adjust_cost (insn, link, dep, cost);
8929 sparc_sched_init (FILE *dump ATTRIBUTE_UNUSED,
8930 int sched_verbose ATTRIBUTE_UNUSED,
8931 int max_ready ATTRIBUTE_UNUSED)
8935 sparc_use_sched_lookahead (void)
8937 if (sparc_cpu == PROCESSOR_NIAGARA
8938 || sparc_cpu == PROCESSOR_NIAGARA2
8939 || sparc_cpu == PROCESSOR_NIAGARA3)
8941 if (sparc_cpu == PROCESSOR_NIAGARA4)
8943 if (sparc_cpu == PROCESSOR_ULTRASPARC
8944 || sparc_cpu == PROCESSOR_ULTRASPARC3)
8946 if ((1 << sparc_cpu) &
8947 ((1 << PROCESSOR_SUPERSPARC) | (1 << PROCESSOR_HYPERSPARC) |
8948 (1 << PROCESSOR_SPARCLITE86X)))
8954 sparc_issue_rate (void)
8958 case PROCESSOR_NIAGARA:
8959 case PROCESSOR_NIAGARA2:
8960 case PROCESSOR_NIAGARA3:
8963 case PROCESSOR_NIAGARA4:
8965 /* Assume V9 processors are capable of at least dual-issue. */
8967 case PROCESSOR_SUPERSPARC:
8969 case PROCESSOR_HYPERSPARC:
8970 case PROCESSOR_SPARCLITE86X:
8972 case PROCESSOR_ULTRASPARC:
8973 case PROCESSOR_ULTRASPARC3:
8979 set_extends (rtx insn)
8981 register rtx pat = PATTERN (insn);
8983 switch (GET_CODE (SET_SRC (pat)))
8985 /* Load and some shift instructions zero extend. */
8988 /* sethi clears the high bits */
8990 /* LO_SUM is used with sethi. sethi cleared the high
8991 bits and the values used with lo_sum are positive */
8993 /* Store flag stores 0 or 1 */
9003 rtx op0 = XEXP (SET_SRC (pat), 0);
9004 rtx op1 = XEXP (SET_SRC (pat), 1);
9005 if (GET_CODE (op1) == CONST_INT)
9006 return INTVAL (op1) >= 0;
9007 if (GET_CODE (op0) != REG)
9009 if (sparc_check_64 (op0, insn) == 1)
9011 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
9016 rtx op0 = XEXP (SET_SRC (pat), 0);
9017 rtx op1 = XEXP (SET_SRC (pat), 1);
9018 if (GET_CODE (op0) != REG || sparc_check_64 (op0, insn) <= 0)
9020 if (GET_CODE (op1) == CONST_INT)
9021 return INTVAL (op1) >= 0;
9022 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
9025 return GET_MODE (SET_SRC (pat)) == SImode;
9026 /* Positive integers leave the high bits zero. */
9028 return ! (CONST_DOUBLE_LOW (SET_SRC (pat)) & 0x80000000);
9030 return ! (INTVAL (SET_SRC (pat)) & 0x80000000);
9033 return - (GET_MODE (SET_SRC (pat)) == SImode);
9035 return sparc_check_64 (SET_SRC (pat), insn);
9041 /* We _ought_ to have only one kind per function, but... */
9042 static GTY(()) rtx sparc_addr_diff_list;
9043 static GTY(()) rtx sparc_addr_list;
9046 sparc_defer_case_vector (rtx lab, rtx vec, int diff)
9048 vec = gen_rtx_EXPR_LIST (VOIDmode, lab, vec);
9050 sparc_addr_diff_list
9051 = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_diff_list);
9053 sparc_addr_list = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_list);
9057 sparc_output_addr_vec (rtx vec)
9059 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
9060 int idx, vlen = XVECLEN (body, 0);
9062 #ifdef ASM_OUTPUT_ADDR_VEC_START
9063 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
9066 #ifdef ASM_OUTPUT_CASE_LABEL
9067 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
9070 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
9073 for (idx = 0; idx < vlen; idx++)
9075 ASM_OUTPUT_ADDR_VEC_ELT
9076 (asm_out_file, CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 0, idx), 0)));
9079 #ifdef ASM_OUTPUT_ADDR_VEC_END
9080 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
9085 sparc_output_addr_diff_vec (rtx vec)
9087 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
9088 rtx base = XEXP (XEXP (body, 0), 0);
9089 int idx, vlen = XVECLEN (body, 1);
9091 #ifdef ASM_OUTPUT_ADDR_VEC_START
9092 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
9095 #ifdef ASM_OUTPUT_CASE_LABEL
9096 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
9099 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
9102 for (idx = 0; idx < vlen; idx++)
9104 ASM_OUTPUT_ADDR_DIFF_ELT
9107 CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 1, idx), 0)),
9108 CODE_LABEL_NUMBER (base));
9111 #ifdef ASM_OUTPUT_ADDR_VEC_END
9112 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
9117 sparc_output_deferred_case_vectors (void)
9122 if (sparc_addr_list == NULL_RTX
9123 && sparc_addr_diff_list == NULL_RTX)
9126 /* Align to cache line in the function's code section. */
9127 switch_to_section (current_function_section ());
9129 align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
9131 ASM_OUTPUT_ALIGN (asm_out_file, align);
9133 for (t = sparc_addr_list; t ; t = XEXP (t, 1))
9134 sparc_output_addr_vec (XEXP (t, 0));
9135 for (t = sparc_addr_diff_list; t ; t = XEXP (t, 1))
9136 sparc_output_addr_diff_vec (XEXP (t, 0));
9138 sparc_addr_list = sparc_addr_diff_list = NULL_RTX;
9141 /* Return 0 if the high 32 bits of X (the low word of X, if DImode) are
9142 unknown. Return 1 if the high bits are zero, -1 if the register is
9145 sparc_check_64 (rtx x, rtx insn)
9147 /* If a register is set only once it is safe to ignore insns this
9148 code does not know how to handle. The loop will either recognize
9149 the single set and return the correct value or fail to recognize
9154 gcc_assert (GET_CODE (x) == REG);
9156 if (GET_MODE (x) == DImode)
9157 y = gen_rtx_REG (SImode, REGNO (x) + WORDS_BIG_ENDIAN);
9159 if (flag_expensive_optimizations
9160 && df && DF_REG_DEF_COUNT (REGNO (y)) == 1)
9166 insn = get_last_insn_anywhere ();
9171 while ((insn = PREV_INSN (insn)))
9173 switch (GET_CODE (insn))
9186 rtx pat = PATTERN (insn);
9187 if (GET_CODE (pat) != SET)
9189 if (rtx_equal_p (x, SET_DEST (pat)))
9190 return set_extends (insn);
9191 if (y && rtx_equal_p (y, SET_DEST (pat)))
9192 return set_extends (insn);
9193 if (reg_overlap_mentioned_p (SET_DEST (pat), y))
9201 /* Output a wide shift instruction in V8+ mode. INSN is the instruction,
9202 OPERANDS are its operands and OPCODE is the mnemonic to be used. */
9205 output_v8plus_shift (rtx insn, rtx *operands, const char *opcode)
9207 static char asm_code[60];
9209 /* The scratch register is only required when the destination
9210 register is not a 64-bit global or out register. */
9211 if (which_alternative != 2)
9212 operands[3] = operands[0];
9214 /* We can only shift by constants <= 63. */
9215 if (GET_CODE (operands[2]) == CONST_INT)
9216 operands[2] = GEN_INT (INTVAL (operands[2]) & 0x3f);
9218 if (GET_CODE (operands[1]) == CONST_INT)
9220 output_asm_insn ("mov\t%1, %3", operands);
9224 output_asm_insn ("sllx\t%H1, 32, %3", operands);
9225 if (sparc_check_64 (operands[1], insn) <= 0)
9226 output_asm_insn ("srl\t%L1, 0, %L1", operands);
9227 output_asm_insn ("or\t%L1, %3, %3", operands);
9230 strcpy (asm_code, opcode);
9232 if (which_alternative != 2)
9233 return strcat (asm_code, "\t%0, %2, %L0\n\tsrlx\t%L0, 32, %H0");
9236 strcat (asm_code, "\t%3, %2, %3\n\tsrlx\t%3, 32, %H0\n\tmov\t%3, %L0");
9239 /* Output rtl to increment the profiler label LABELNO
9240 for profiling a function entry. */
9243 sparc_profile_hook (int labelno)
9248 fun = gen_rtx_SYMBOL_REF (Pmode, MCOUNT_FUNCTION);
9249 if (NO_PROFILE_COUNTERS)
9251 emit_library_call (fun, LCT_NORMAL, VOIDmode, 0);
9255 ASM_GENERATE_INTERNAL_LABEL (buf, "LP", labelno);
9256 lab = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (buf));
9257 emit_library_call (fun, LCT_NORMAL, VOIDmode, 1, lab, Pmode);
9261 #ifdef TARGET_SOLARIS
9262 /* Solaris implementation of TARGET_ASM_NAMED_SECTION. */
9265 sparc_solaris_elf_asm_named_section (const char *name, unsigned int flags,
9266 tree decl ATTRIBUTE_UNUSED)
9268 if (HAVE_COMDAT_GROUP && flags & SECTION_LINKONCE)
9270 solaris_elf_asm_comdat_section (name, flags, decl);
9274 fprintf (asm_out_file, "\t.section\t\"%s\"", name);
9276 if (!(flags & SECTION_DEBUG))
9277 fputs (",#alloc", asm_out_file);
9278 if (flags & SECTION_WRITE)
9279 fputs (",#write", asm_out_file);
9280 if (flags & SECTION_TLS)
9281 fputs (",#tls", asm_out_file);
9282 if (flags & SECTION_CODE)
9283 fputs (",#execinstr", asm_out_file);
9285 /* ??? Handle SECTION_BSS. */
9287 fputc ('\n', asm_out_file);
9289 #endif /* TARGET_SOLARIS */
9291 /* We do not allow indirect calls to be optimized into sibling calls.
9293 We cannot use sibling calls when delayed branches are disabled
9294 because they will likely require the call delay slot to be filled.
9296 Also, on SPARC 32-bit we cannot emit a sibling call when the
9297 current function returns a structure. This is because the "unimp
9298 after call" convention would cause the callee to return to the
9299 wrong place. The generic code already disallows cases where the
9300 function being called returns a structure.
9302 It may seem strange how this last case could occur. Usually there
9303 is code after the call which jumps to epilogue code which dumps the
9304 return value into the struct return area. That ought to invalidate
9305 the sibling call right? Well, in the C++ case we can end up passing
9306 the pointer to the struct return area to a constructor (which returns
9307 void) and then nothing else happens. Such a sibling call would look
9308 valid without the added check here.
9310 VxWorks PIC PLT entries require the global pointer to be initialized
9311 on entry. We therefore can't emit sibling calls to them. */
9313 sparc_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
9316 && flag_delayed_branch
9317 && (TARGET_ARCH64 || ! cfun->returns_struct)
9318 && !(TARGET_VXWORKS_RTP
9320 && !targetm.binds_local_p (decl)));
9323 /* libfunc renaming. */
9326 sparc_init_libfuncs (void)
9330 /* Use the subroutines that Sun's library provides for integer
9331 multiply and divide. The `*' prevents an underscore from
9332 being prepended by the compiler. .umul is a little faster
9334 set_optab_libfunc (smul_optab, SImode, "*.umul");
9335 set_optab_libfunc (sdiv_optab, SImode, "*.div");
9336 set_optab_libfunc (udiv_optab, SImode, "*.udiv");
9337 set_optab_libfunc (smod_optab, SImode, "*.rem");
9338 set_optab_libfunc (umod_optab, SImode, "*.urem");
9340 /* TFmode arithmetic. These names are part of the SPARC 32bit ABI. */
9341 set_optab_libfunc (add_optab, TFmode, "_Q_add");
9342 set_optab_libfunc (sub_optab, TFmode, "_Q_sub");
9343 set_optab_libfunc (neg_optab, TFmode, "_Q_neg");
9344 set_optab_libfunc (smul_optab, TFmode, "_Q_mul");
9345 set_optab_libfunc (sdiv_optab, TFmode, "_Q_div");
9347 /* We can define the TFmode sqrt optab only if TARGET_FPU. This
9348 is because with soft-float, the SFmode and DFmode sqrt
9349 instructions will be absent, and the compiler will notice and
9350 try to use the TFmode sqrt instruction for calls to the
9351 builtin function sqrt, but this fails. */
9353 set_optab_libfunc (sqrt_optab, TFmode, "_Q_sqrt");
9355 set_optab_libfunc (eq_optab, TFmode, "_Q_feq");
9356 set_optab_libfunc (ne_optab, TFmode, "_Q_fne");
9357 set_optab_libfunc (gt_optab, TFmode, "_Q_fgt");
9358 set_optab_libfunc (ge_optab, TFmode, "_Q_fge");
9359 set_optab_libfunc (lt_optab, TFmode, "_Q_flt");
9360 set_optab_libfunc (le_optab, TFmode, "_Q_fle");
9362 set_conv_libfunc (sext_optab, TFmode, SFmode, "_Q_stoq");
9363 set_conv_libfunc (sext_optab, TFmode, DFmode, "_Q_dtoq");
9364 set_conv_libfunc (trunc_optab, SFmode, TFmode, "_Q_qtos");
9365 set_conv_libfunc (trunc_optab, DFmode, TFmode, "_Q_qtod");
9367 set_conv_libfunc (sfix_optab, SImode, TFmode, "_Q_qtoi");
9368 set_conv_libfunc (ufix_optab, SImode, TFmode, "_Q_qtou");
9369 set_conv_libfunc (sfloat_optab, TFmode, SImode, "_Q_itoq");
9370 set_conv_libfunc (ufloat_optab, TFmode, SImode, "_Q_utoq");
9372 if (DITF_CONVERSION_LIBFUNCS)
9374 set_conv_libfunc (sfix_optab, DImode, TFmode, "_Q_qtoll");
9375 set_conv_libfunc (ufix_optab, DImode, TFmode, "_Q_qtoull");
9376 set_conv_libfunc (sfloat_optab, TFmode, DImode, "_Q_lltoq");
9377 set_conv_libfunc (ufloat_optab, TFmode, DImode, "_Q_ulltoq");
9380 if (SUN_CONVERSION_LIBFUNCS)
9382 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftoll");
9383 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoull");
9384 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtoll");
9385 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoull");
9390 /* In the SPARC 64bit ABI, SImode multiply and divide functions
9391 do not exist in the library. Make sure the compiler does not
9392 emit calls to them by accident. (It should always use the
9393 hardware instructions.) */
9394 set_optab_libfunc (smul_optab, SImode, 0);
9395 set_optab_libfunc (sdiv_optab, SImode, 0);
9396 set_optab_libfunc (udiv_optab, SImode, 0);
9397 set_optab_libfunc (smod_optab, SImode, 0);
9398 set_optab_libfunc (umod_optab, SImode, 0);
9400 if (SUN_INTEGER_MULTIPLY_64)
9402 set_optab_libfunc (smul_optab, DImode, "__mul64");
9403 set_optab_libfunc (sdiv_optab, DImode, "__div64");
9404 set_optab_libfunc (udiv_optab, DImode, "__udiv64");
9405 set_optab_libfunc (smod_optab, DImode, "__rem64");
9406 set_optab_libfunc (umod_optab, DImode, "__urem64");
9409 if (SUN_CONVERSION_LIBFUNCS)
9411 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftol");
9412 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoul");
9413 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtol");
9414 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoul");
9419 static tree def_builtin(const char *name, int code, tree type)
9421 return add_builtin_function(name, type, code, BUILT_IN_MD, NULL,
9425 static tree def_builtin_const(const char *name, int code, tree type)
9427 tree t = def_builtin(name, code, type);
9430 TREE_READONLY (t) = 1;
9435 /* Implement the TARGET_INIT_BUILTINS target hook.
9436 Create builtin functions for special SPARC instructions. */
9439 sparc_init_builtins (void)
9442 sparc_vis_init_builtins ();
9445 /* Create builtin functions for VIS 1.0 instructions. */
9448 sparc_vis_init_builtins (void)
9450 tree v4qi = build_vector_type (unsigned_intQI_type_node, 4);
9451 tree v8qi = build_vector_type (unsigned_intQI_type_node, 8);
9452 tree v4hi = build_vector_type (intHI_type_node, 4);
9453 tree v2hi = build_vector_type (intHI_type_node, 2);
9454 tree v2si = build_vector_type (intSI_type_node, 2);
9455 tree v1si = build_vector_type (intSI_type_node, 1);
9457 tree v4qi_ftype_v4hi = build_function_type_list (v4qi, v4hi, 0);
9458 tree v8qi_ftype_v2si_v8qi = build_function_type_list (v8qi, v2si, v8qi, 0);
9459 tree v2hi_ftype_v2si = build_function_type_list (v2hi, v2si, 0);
9460 tree v4hi_ftype_v4qi = build_function_type_list (v4hi, v4qi, 0);
9461 tree v8qi_ftype_v4qi_v4qi = build_function_type_list (v8qi, v4qi, v4qi, 0);
9462 tree v4hi_ftype_v4qi_v4hi = build_function_type_list (v4hi, v4qi, v4hi, 0);
9463 tree v4hi_ftype_v4qi_v2hi = build_function_type_list (v4hi, v4qi, v2hi, 0);
9464 tree v2si_ftype_v4qi_v2hi = build_function_type_list (v2si, v4qi, v2hi, 0);
9465 tree v4hi_ftype_v8qi_v4hi = build_function_type_list (v4hi, v8qi, v4hi, 0);
9466 tree v4hi_ftype_v4hi_v4hi = build_function_type_list (v4hi, v4hi, v4hi, 0);
9467 tree v2si_ftype_v2si_v2si = build_function_type_list (v2si, v2si, v2si, 0);
9468 tree v8qi_ftype_v8qi_v8qi = build_function_type_list (v8qi, v8qi, v8qi, 0);
9469 tree v2hi_ftype_v2hi_v2hi = build_function_type_list (v2hi, v2hi, v2hi, 0);
9470 tree v1si_ftype_v1si_v1si = build_function_type_list (v1si, v1si, v1si, 0);
9471 tree di_ftype_v8qi_v8qi_di = build_function_type_list (intDI_type_node,
9473 intDI_type_node, 0);
9474 tree di_ftype_v8qi_v8qi = build_function_type_list (intDI_type_node,
9476 tree si_ftype_v8qi_v8qi = build_function_type_list (intSI_type_node,
9478 tree di_ftype_di_di = build_function_type_list (intDI_type_node,
9480 intDI_type_node, 0);
9481 tree si_ftype_si_si = build_function_type_list (intSI_type_node,
9483 intSI_type_node, 0);
9484 tree ptr_ftype_ptr_si = build_function_type_list (ptr_type_node,
9486 intSI_type_node, 0);
9487 tree ptr_ftype_ptr_di = build_function_type_list (ptr_type_node,
9489 intDI_type_node, 0);
9490 tree si_ftype_ptr_ptr = build_function_type_list (intSI_type_node,
9493 tree di_ftype_ptr_ptr = build_function_type_list (intDI_type_node,
9496 tree si_ftype_v4hi_v4hi = build_function_type_list (intSI_type_node,
9498 tree si_ftype_v2si_v2si = build_function_type_list (intSI_type_node,
9500 tree di_ftype_v4hi_v4hi = build_function_type_list (intDI_type_node,
9502 tree di_ftype_v2si_v2si = build_function_type_list (intDI_type_node,
9504 tree void_ftype_di = build_function_type_list (void_type_node,
9505 intDI_type_node, 0);
9506 tree di_ftype_void = build_function_type_list (intDI_type_node,
9508 tree void_ftype_si = build_function_type_list (void_type_node,
9509 intSI_type_node, 0);
9510 tree sf_ftype_sf_sf = build_function_type_list (float_type_node,
9512 float_type_node, 0);
9513 tree df_ftype_df_df = build_function_type_list (double_type_node,
9515 double_type_node, 0);
9517 /* Packing and expanding vectors. */
9518 def_builtin ("__builtin_vis_fpack16", CODE_FOR_fpack16_vis,
9520 def_builtin ("__builtin_vis_fpack32", CODE_FOR_fpack32_vis,
9521 v8qi_ftype_v2si_v8qi);
9522 def_builtin ("__builtin_vis_fpackfix", CODE_FOR_fpackfix_vis,
9524 def_builtin_const ("__builtin_vis_fexpand", CODE_FOR_fexpand_vis,
9526 def_builtin_const ("__builtin_vis_fpmerge", CODE_FOR_fpmerge_vis,
9527 v8qi_ftype_v4qi_v4qi);
9529 /* Multiplications. */
9530 def_builtin_const ("__builtin_vis_fmul8x16", CODE_FOR_fmul8x16_vis,
9531 v4hi_ftype_v4qi_v4hi);
9532 def_builtin_const ("__builtin_vis_fmul8x16au", CODE_FOR_fmul8x16au_vis,
9533 v4hi_ftype_v4qi_v2hi);
9534 def_builtin_const ("__builtin_vis_fmul8x16al", CODE_FOR_fmul8x16al_vis,
9535 v4hi_ftype_v4qi_v2hi);
9536 def_builtin_const ("__builtin_vis_fmul8sux16", CODE_FOR_fmul8sux16_vis,
9537 v4hi_ftype_v8qi_v4hi);
9538 def_builtin_const ("__builtin_vis_fmul8ulx16", CODE_FOR_fmul8ulx16_vis,
9539 v4hi_ftype_v8qi_v4hi);
9540 def_builtin_const ("__builtin_vis_fmuld8sux16", CODE_FOR_fmuld8sux16_vis,
9541 v2si_ftype_v4qi_v2hi);
9542 def_builtin_const ("__builtin_vis_fmuld8ulx16", CODE_FOR_fmuld8ulx16_vis,
9543 v2si_ftype_v4qi_v2hi);
9545 /* Data aligning. */
9546 def_builtin ("__builtin_vis_faligndatav4hi", CODE_FOR_faligndatav4hi_vis,
9547 v4hi_ftype_v4hi_v4hi);
9548 def_builtin ("__builtin_vis_faligndatav8qi", CODE_FOR_faligndatav8qi_vis,
9549 v8qi_ftype_v8qi_v8qi);
9550 def_builtin ("__builtin_vis_faligndatav2si", CODE_FOR_faligndatav2si_vis,
9551 v2si_ftype_v2si_v2si);
9552 def_builtin ("__builtin_vis_faligndatadi", CODE_FOR_faligndatav1di_vis,
9555 def_builtin ("__builtin_vis_write_gsr", CODE_FOR_wrgsr_vis,
9557 def_builtin ("__builtin_vis_read_gsr", CODE_FOR_rdgsr_vis,
9562 def_builtin ("__builtin_vis_alignaddr", CODE_FOR_alignaddrdi_vis,
9564 def_builtin ("__builtin_vis_alignaddrl", CODE_FOR_alignaddrldi_vis,
9569 def_builtin ("__builtin_vis_alignaddr", CODE_FOR_alignaddrsi_vis,
9571 def_builtin ("__builtin_vis_alignaddrl", CODE_FOR_alignaddrlsi_vis,
9575 /* Pixel distance. */
9576 def_builtin_const ("__builtin_vis_pdist", CODE_FOR_pdist_vis,
9577 di_ftype_v8qi_v8qi_di);
9579 /* Edge handling. */
9582 def_builtin_const ("__builtin_vis_edge8", CODE_FOR_edge8di_vis,
9584 def_builtin_const ("__builtin_vis_edge8l", CODE_FOR_edge8ldi_vis,
9586 def_builtin_const ("__builtin_vis_edge16", CODE_FOR_edge16di_vis,
9588 def_builtin_const ("__builtin_vis_edge16l", CODE_FOR_edge16ldi_vis,
9590 def_builtin_const ("__builtin_vis_edge32", CODE_FOR_edge32di_vis,
9592 def_builtin_const ("__builtin_vis_edge32l", CODE_FOR_edge32ldi_vis,
9596 def_builtin_const ("__builtin_vis_edge8n", CODE_FOR_edge8ndi_vis,
9598 def_builtin_const ("__builtin_vis_edge8ln", CODE_FOR_edge8lndi_vis,
9600 def_builtin_const ("__builtin_vis_edge16n", CODE_FOR_edge16ndi_vis,
9602 def_builtin_const ("__builtin_vis_edge16ln", CODE_FOR_edge16lndi_vis,
9604 def_builtin_const ("__builtin_vis_edge32n", CODE_FOR_edge32ndi_vis,
9606 def_builtin_const ("__builtin_vis_edge32ln", CODE_FOR_edge32lndi_vis,
9612 def_builtin_const ("__builtin_vis_edge8", CODE_FOR_edge8si_vis,
9614 def_builtin_const ("__builtin_vis_edge8l", CODE_FOR_edge8lsi_vis,
9616 def_builtin_const ("__builtin_vis_edge16", CODE_FOR_edge16si_vis,
9618 def_builtin_const ("__builtin_vis_edge16l", CODE_FOR_edge16lsi_vis,
9620 def_builtin_const ("__builtin_vis_edge32", CODE_FOR_edge32si_vis,
9622 def_builtin_const ("__builtin_vis_edge32l", CODE_FOR_edge32lsi_vis,
9626 def_builtin_const ("__builtin_vis_edge8n", CODE_FOR_edge8nsi_vis,
9628 def_builtin_const ("__builtin_vis_edge8ln", CODE_FOR_edge8lnsi_vis,
9630 def_builtin_const ("__builtin_vis_edge16n", CODE_FOR_edge16nsi_vis,
9632 def_builtin_const ("__builtin_vis_edge16ln", CODE_FOR_edge16lnsi_vis,
9634 def_builtin_const ("__builtin_vis_edge32n", CODE_FOR_edge32nsi_vis,
9636 def_builtin_const ("__builtin_vis_edge32ln", CODE_FOR_edge32lnsi_vis,
9641 /* Pixel compare. */
9644 def_builtin_const ("__builtin_vis_fcmple16", CODE_FOR_fcmple16di_vis,
9645 di_ftype_v4hi_v4hi);
9646 def_builtin_const ("__builtin_vis_fcmple32", CODE_FOR_fcmple32di_vis,
9647 di_ftype_v2si_v2si);
9648 def_builtin_const ("__builtin_vis_fcmpne16", CODE_FOR_fcmpne16di_vis,
9649 di_ftype_v4hi_v4hi);
9650 def_builtin_const ("__builtin_vis_fcmpne32", CODE_FOR_fcmpne32di_vis,
9651 di_ftype_v2si_v2si);
9652 def_builtin_const ("__builtin_vis_fcmpgt16", CODE_FOR_fcmpgt16di_vis,
9653 di_ftype_v4hi_v4hi);
9654 def_builtin_const ("__builtin_vis_fcmpgt32", CODE_FOR_fcmpgt32di_vis,
9655 di_ftype_v2si_v2si);
9656 def_builtin_const ("__builtin_vis_fcmpeq16", CODE_FOR_fcmpeq16di_vis,
9657 di_ftype_v4hi_v4hi);
9658 def_builtin_const ("__builtin_vis_fcmpeq32", CODE_FOR_fcmpeq32di_vis,
9659 di_ftype_v2si_v2si);
9663 def_builtin_const ("__builtin_vis_fcmple16", CODE_FOR_fcmple16si_vis,
9664 si_ftype_v4hi_v4hi);
9665 def_builtin_const ("__builtin_vis_fcmple32", CODE_FOR_fcmple32si_vis,
9666 si_ftype_v2si_v2si);
9667 def_builtin_const ("__builtin_vis_fcmpne16", CODE_FOR_fcmpne16si_vis,
9668 si_ftype_v4hi_v4hi);
9669 def_builtin_const ("__builtin_vis_fcmpne32", CODE_FOR_fcmpne32si_vis,
9670 si_ftype_v2si_v2si);
9671 def_builtin_const ("__builtin_vis_fcmpgt16", CODE_FOR_fcmpgt16si_vis,
9672 si_ftype_v4hi_v4hi);
9673 def_builtin_const ("__builtin_vis_fcmpgt32", CODE_FOR_fcmpgt32si_vis,
9674 si_ftype_v2si_v2si);
9675 def_builtin_const ("__builtin_vis_fcmpeq16", CODE_FOR_fcmpeq16si_vis,
9676 si_ftype_v4hi_v4hi);
9677 def_builtin_const ("__builtin_vis_fcmpeq32", CODE_FOR_fcmpeq32si_vis,
9678 si_ftype_v2si_v2si);
9681 /* Addition and subtraction. */
9682 def_builtin_const ("__builtin_vis_fpadd16", CODE_FOR_addv4hi3,
9683 v4hi_ftype_v4hi_v4hi);
9684 def_builtin_const ("__builtin_vis_fpadd16s", CODE_FOR_addv2hi3,
9685 v2hi_ftype_v2hi_v2hi);
9686 def_builtin_const ("__builtin_vis_fpadd32", CODE_FOR_addv2si3,
9687 v2si_ftype_v2si_v2si);
9688 def_builtin_const ("__builtin_vis_fpadd32s", CODE_FOR_addv1si3,
9689 v1si_ftype_v1si_v1si);
9690 def_builtin_const ("__builtin_vis_fpsub16", CODE_FOR_subv4hi3,
9691 v4hi_ftype_v4hi_v4hi);
9692 def_builtin_const ("__builtin_vis_fpsub16s", CODE_FOR_subv2hi3,
9693 v2hi_ftype_v2hi_v2hi);
9694 def_builtin_const ("__builtin_vis_fpsub32", CODE_FOR_subv2si3,
9695 v2si_ftype_v2si_v2si);
9696 def_builtin_const ("__builtin_vis_fpsub32s", CODE_FOR_subv1si3,
9697 v1si_ftype_v1si_v1si);
9699 /* Three-dimensional array addressing. */
9702 def_builtin_const ("__builtin_vis_array8", CODE_FOR_array8di_vis,
9704 def_builtin_const ("__builtin_vis_array16", CODE_FOR_array16di_vis,
9706 def_builtin_const ("__builtin_vis_array32", CODE_FOR_array32di_vis,
9711 def_builtin_const ("__builtin_vis_array8", CODE_FOR_array8si_vis,
9713 def_builtin_const ("__builtin_vis_array16", CODE_FOR_array16si_vis,
9715 def_builtin_const ("__builtin_vis_array32", CODE_FOR_array32si_vis,
9721 /* Byte mask and shuffle */
9723 def_builtin ("__builtin_vis_bmask", CODE_FOR_bmaskdi_vis,
9726 def_builtin ("__builtin_vis_bmask", CODE_FOR_bmasksi_vis,
9728 def_builtin ("__builtin_vis_bshufflev4hi", CODE_FOR_bshufflev4hi_vis,
9729 v4hi_ftype_v4hi_v4hi);
9730 def_builtin ("__builtin_vis_bshufflev8qi", CODE_FOR_bshufflev8qi_vis,
9731 v8qi_ftype_v8qi_v8qi);
9732 def_builtin ("__builtin_vis_bshufflev2si", CODE_FOR_bshufflev2si_vis,
9733 v2si_ftype_v2si_v2si);
9734 def_builtin ("__builtin_vis_bshuffledi", CODE_FOR_bshufflev1di_vis,
9742 def_builtin ("__builtin_vis_cmask8", CODE_FOR_cmask8di_vis,
9744 def_builtin ("__builtin_vis_cmask16", CODE_FOR_cmask16di_vis,
9746 def_builtin ("__builtin_vis_cmask32", CODE_FOR_cmask32di_vis,
9751 def_builtin ("__builtin_vis_cmask8", CODE_FOR_cmask8si_vis,
9753 def_builtin ("__builtin_vis_cmask16", CODE_FOR_cmask16si_vis,
9755 def_builtin ("__builtin_vis_cmask32", CODE_FOR_cmask32si_vis,
9759 def_builtin_const ("__builtin_vis_fchksm16", CODE_FOR_fchksm16_vis,
9760 v4hi_ftype_v4hi_v4hi);
9762 def_builtin_const ("__builtin_vis_fsll16", CODE_FOR_vashlv4hi3,
9763 v4hi_ftype_v4hi_v4hi);
9764 def_builtin_const ("__builtin_vis_fslas16", CODE_FOR_vssashlv4hi3,
9765 v4hi_ftype_v4hi_v4hi);
9766 def_builtin_const ("__builtin_vis_fsrl16", CODE_FOR_vlshrv4hi3,
9767 v4hi_ftype_v4hi_v4hi);
9768 def_builtin_const ("__builtin_vis_fsra16", CODE_FOR_vashrv4hi3,
9769 v4hi_ftype_v4hi_v4hi);
9770 def_builtin_const ("__builtin_vis_fsll32", CODE_FOR_vashlv2si3,
9771 v2si_ftype_v2si_v2si);
9772 def_builtin_const ("__builtin_vis_fslas32", CODE_FOR_vssashlv2si3,
9773 v2si_ftype_v2si_v2si);
9774 def_builtin_const ("__builtin_vis_fsrl32", CODE_FOR_vlshrv2si3,
9775 v2si_ftype_v2si_v2si);
9776 def_builtin_const ("__builtin_vis_fsra32", CODE_FOR_vashrv2si3,
9777 v2si_ftype_v2si_v2si);
9780 def_builtin_const ("__builtin_vis_pdistn", CODE_FOR_pdistndi_vis,
9781 di_ftype_v8qi_v8qi);
9783 def_builtin_const ("__builtin_vis_pdistn", CODE_FOR_pdistnsi_vis,
9784 si_ftype_v8qi_v8qi);
9786 def_builtin_const ("__builtin_vis_fmean16", CODE_FOR_fmean16_vis,
9787 v4hi_ftype_v4hi_v4hi);
9788 def_builtin_const ("__builtin_vis_fpadd64", CODE_FOR_fpadd64_vis,
9790 def_builtin_const ("__builtin_vis_fpsub64", CODE_FOR_fpsub64_vis,
9793 def_builtin_const ("__builtin_vis_fpadds16", CODE_FOR_ssaddv4hi3,
9794 v4hi_ftype_v4hi_v4hi);
9795 def_builtin_const ("__builtin_vis_fpadds16s", CODE_FOR_ssaddv2hi3,
9796 v2hi_ftype_v2hi_v2hi);
9797 def_builtin_const ("__builtin_vis_fpsubs16", CODE_FOR_sssubv4hi3,
9798 v4hi_ftype_v4hi_v4hi);
9799 def_builtin_const ("__builtin_vis_fpsubs16s", CODE_FOR_sssubv2hi3,
9800 v2hi_ftype_v2hi_v2hi);
9801 def_builtin_const ("__builtin_vis_fpadds32", CODE_FOR_ssaddv2si3,
9802 v2si_ftype_v2si_v2si);
9803 def_builtin_const ("__builtin_vis_fpadds32s", CODE_FOR_ssaddv1si3,
9804 v1si_ftype_v1si_v1si);
9805 def_builtin_const ("__builtin_vis_fpsubs32", CODE_FOR_sssubv2si3,
9806 v2si_ftype_v2si_v2si);
9807 def_builtin_const ("__builtin_vis_fpsubs32s", CODE_FOR_sssubv1si3,
9808 v1si_ftype_v1si_v1si);
9812 def_builtin_const ("__builtin_vis_fucmple8", CODE_FOR_fucmple8di_vis,
9813 di_ftype_v8qi_v8qi);
9814 def_builtin_const ("__builtin_vis_fucmpne8", CODE_FOR_fucmpne8di_vis,
9815 di_ftype_v8qi_v8qi);
9816 def_builtin_const ("__builtin_vis_fucmpgt8", CODE_FOR_fucmpgt8di_vis,
9817 di_ftype_v8qi_v8qi);
9818 def_builtin_const ("__builtin_vis_fucmpeq8", CODE_FOR_fucmpeq8di_vis,
9819 di_ftype_v8qi_v8qi);
9823 def_builtin_const ("__builtin_vis_fucmple8", CODE_FOR_fucmple8si_vis,
9824 si_ftype_v8qi_v8qi);
9825 def_builtin_const ("__builtin_vis_fucmpne8", CODE_FOR_fucmpne8si_vis,
9826 si_ftype_v8qi_v8qi);
9827 def_builtin_const ("__builtin_vis_fucmpgt8", CODE_FOR_fucmpgt8si_vis,
9828 si_ftype_v8qi_v8qi);
9829 def_builtin_const ("__builtin_vis_fucmpeq8", CODE_FOR_fucmpeq8si_vis,
9830 si_ftype_v8qi_v8qi);
9833 def_builtin_const ("__builtin_vis_fhadds", CODE_FOR_fhaddsf_vis,
9835 def_builtin_const ("__builtin_vis_fhaddd", CODE_FOR_fhadddf_vis,
9837 def_builtin_const ("__builtin_vis_fhsubs", CODE_FOR_fhsubsf_vis,
9839 def_builtin_const ("__builtin_vis_fhsubd", CODE_FOR_fhsubdf_vis,
9841 def_builtin_const ("__builtin_vis_fnhadds", CODE_FOR_fnhaddsf_vis,
9843 def_builtin_const ("__builtin_vis_fnhaddd", CODE_FOR_fnhadddf_vis,
9846 def_builtin_const ("__builtin_vis_umulxhi", CODE_FOR_umulxhi_vis,
9848 def_builtin_const ("__builtin_vis_xmulx", CODE_FOR_xmulx_vis,
9850 def_builtin_const ("__builtin_vis_xmulxhi", CODE_FOR_xmulxhi_vis,
9855 /* Handle TARGET_EXPAND_BUILTIN target hook.
9856 Expand builtin functions for sparc intrinsics. */
9859 sparc_expand_builtin (tree exp, rtx target,
9860 rtx subtarget ATTRIBUTE_UNUSED,
9861 enum machine_mode tmode ATTRIBUTE_UNUSED,
9862 int ignore ATTRIBUTE_UNUSED)
9865 call_expr_arg_iterator iter;
9866 tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0);
9867 unsigned int icode = DECL_FUNCTION_CODE (fndecl);
9872 nonvoid = TREE_TYPE (TREE_TYPE (fndecl)) != void_type_node;
9876 enum machine_mode tmode = insn_data[icode].operand[0].mode;
9878 || GET_MODE (target) != tmode
9879 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
9880 op[0] = gen_reg_rtx (tmode);
9884 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
9886 const struct insn_operand_data *insn_op;
9889 if (arg == error_mark_node)
9893 idx = arg_count - !nonvoid;
9894 insn_op = &insn_data[icode].operand[idx];
9895 op[arg_count] = expand_normal (arg);
9897 if (insn_op->mode == V1DImode
9898 && GET_MODE (op[arg_count]) == DImode)
9899 op[arg_count] = gen_lowpart (V1DImode, op[arg_count]);
9900 else if (insn_op->mode == V1SImode
9901 && GET_MODE (op[arg_count]) == SImode)
9902 op[arg_count] = gen_lowpart (V1SImode, op[arg_count]);
9904 if (! (*insn_data[icode].operand[idx].predicate) (op[arg_count],
9906 op[arg_count] = copy_to_mode_reg (insn_op->mode, op[arg_count]);
9912 pat = GEN_FCN (icode) (op[0]);
9916 pat = GEN_FCN (icode) (op[0], op[1]);
9918 pat = GEN_FCN (icode) (op[1]);
9921 pat = GEN_FCN (icode) (op[0], op[1], op[2]);
9924 pat = GEN_FCN (icode) (op[0], op[1], op[2], op[3]);
9942 sparc_vis_mul8x16 (int e8, int e16)
9944 return (e8 * e16 + 128) / 256;
9947 /* Multiply the VECTOR_CSTs CST0 and CST1 as specified by FNCODE and put
9948 the result into the array N_ELTS, whose elements are of INNER_TYPE. */
9951 sparc_handle_vis_mul8x16 (tree *n_elts, int fncode, tree inner_type,
9952 tree cst0, tree cst1)
9954 unsigned i, num = VECTOR_CST_NELTS (cst0);
9959 case CODE_FOR_fmul8x16_vis:
9960 for (i = 0; i < num; ++i)
9963 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (VECTOR_CST_ELT (cst0, i)),
9964 TREE_INT_CST_LOW (VECTOR_CST_ELT (cst1, i)));
9965 n_elts[i] = build_int_cst (inner_type, val);
9969 case CODE_FOR_fmul8x16au_vis:
9970 scale = TREE_INT_CST_LOW (VECTOR_CST_ELT (cst1, 0));
9972 for (i = 0; i < num; ++i)
9975 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (VECTOR_CST_ELT (cst0, i)),
9977 n_elts[i] = build_int_cst (inner_type, val);
9981 case CODE_FOR_fmul8x16al_vis:
9982 scale = TREE_INT_CST_LOW (VECTOR_CST_ELT (cst1, 1));
9984 for (i = 0; i < num; ++i)
9987 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (VECTOR_CST_ELT (cst0, i)),
9989 n_elts[i] = build_int_cst (inner_type, val);
9998 /* Handle TARGET_FOLD_BUILTIN target hook.
9999 Fold builtin functions for SPARC intrinsics. If IGNORE is true the
10000 result of the function call is ignored. NULL_TREE is returned if the
10001 function could not be folded. */
10004 sparc_fold_builtin (tree fndecl, int n_args ATTRIBUTE_UNUSED,
10005 tree *args, bool ignore)
10007 tree arg0, arg1, arg2;
10008 tree rtype = TREE_TYPE (TREE_TYPE (fndecl));
10009 enum insn_code icode = (enum insn_code) DECL_FUNCTION_CODE (fndecl);
10013 /* Note that a switch statement instead of the sequence of tests would
10014 be incorrect as many of the CODE_FOR values could be CODE_FOR_nothing
10015 and that would yield multiple alternatives with identical values. */
10016 if (icode == CODE_FOR_alignaddrsi_vis
10017 || icode == CODE_FOR_alignaddrdi_vis
10018 || icode == CODE_FOR_wrgsr_vis
10019 || icode == CODE_FOR_bmasksi_vis
10020 || icode == CODE_FOR_bmaskdi_vis
10021 || icode == CODE_FOR_cmask8si_vis
10022 || icode == CODE_FOR_cmask8di_vis
10023 || icode == CODE_FOR_cmask16si_vis
10024 || icode == CODE_FOR_cmask16di_vis
10025 || icode == CODE_FOR_cmask32si_vis
10026 || icode == CODE_FOR_cmask32di_vis)
10029 return build_zero_cst (rtype);
10034 case CODE_FOR_fexpand_vis:
10038 if (TREE_CODE (arg0) == VECTOR_CST)
10040 tree inner_type = TREE_TYPE (rtype);
10044 n_elts = XALLOCAVEC (tree, VECTOR_CST_NELTS (arg0));
10045 for (i = 0; i < VECTOR_CST_NELTS (arg0); ++i)
10046 n_elts[i] = build_int_cst (inner_type,
10048 (VECTOR_CST_ELT (arg0, i)) << 4);
10049 return build_vector (rtype, n_elts);
10053 case CODE_FOR_fmul8x16_vis:
10054 case CODE_FOR_fmul8x16au_vis:
10055 case CODE_FOR_fmul8x16al_vis:
10061 if (TREE_CODE (arg0) == VECTOR_CST && TREE_CODE (arg1) == VECTOR_CST)
10063 tree inner_type = TREE_TYPE (rtype);
10064 tree *n_elts = XALLOCAVEC (tree, VECTOR_CST_NELTS (arg0));
10065 sparc_handle_vis_mul8x16 (n_elts, icode, inner_type, arg0, arg1);
10066 return build_vector (rtype, n_elts);
10070 case CODE_FOR_fpmerge_vis:
10076 if (TREE_CODE (arg0) == VECTOR_CST && TREE_CODE (arg1) == VECTOR_CST)
10078 tree *n_elts = XALLOCAVEC (tree, 2 * VECTOR_CST_NELTS (arg0));
10080 for (i = 0; i < VECTOR_CST_NELTS (arg0); ++i)
10082 n_elts[2*i] = VECTOR_CST_ELT (arg0, i);
10083 n_elts[2*i+1] = VECTOR_CST_ELT (arg1, i);
10086 return build_vector (rtype, n_elts);
10090 case CODE_FOR_pdist_vis:
10098 if (TREE_CODE (arg0) == VECTOR_CST
10099 && TREE_CODE (arg1) == VECTOR_CST
10100 && TREE_CODE (arg2) == INTEGER_CST)
10103 unsigned HOST_WIDE_INT low = TREE_INT_CST_LOW (arg2);
10104 HOST_WIDE_INT high = TREE_INT_CST_HIGH (arg2);
10107 for (i = 0; i < VECTOR_CST_NELTS (arg0); ++i)
10109 unsigned HOST_WIDE_INT
10110 low0 = TREE_INT_CST_LOW (VECTOR_CST_ELT (arg0, i)),
10111 low1 = TREE_INT_CST_LOW (VECTOR_CST_ELT (arg1, i));
10113 high0 = TREE_INT_CST_HIGH (VECTOR_CST_ELT (arg0, i));
10115 high1 = TREE_INT_CST_HIGH (VECTOR_CST_ELT (arg1, i));
10117 unsigned HOST_WIDE_INT l;
10120 overflow |= neg_double (low1, high1, &l, &h);
10121 overflow |= add_double (low0, high0, l, h, &l, &h);
10123 overflow |= neg_double (l, h, &l, &h);
10125 overflow |= add_double (low, high, l, h, &low, &high);
10128 gcc_assert (overflow == 0);
10130 return build_int_cst_wide (rtype, low, high);
10140 /* ??? This duplicates information provided to the compiler by the
10141 ??? scheduler description. Some day, teach genautomata to output
10142 ??? the latencies and then CSE will just use that. */
10145 sparc_rtx_costs (rtx x, int code, int outer_code, int opno ATTRIBUTE_UNUSED,
10146 int *total, bool speed ATTRIBUTE_UNUSED)
10148 enum machine_mode mode = GET_MODE (x);
10149 bool float_mode_p = FLOAT_MODE_P (mode);
10154 if (INTVAL (x) < 0x1000 && INTVAL (x) >= -0x1000)
10172 if (GET_MODE (x) == VOIDmode
10173 && ((CONST_DOUBLE_HIGH (x) == 0
10174 && CONST_DOUBLE_LOW (x) < 0x1000)
10175 || (CONST_DOUBLE_HIGH (x) == -1
10176 && CONST_DOUBLE_LOW (x) < 0
10177 && CONST_DOUBLE_LOW (x) >= -0x1000)))
10184 /* If outer-code was a sign or zero extension, a cost
10185 of COSTS_N_INSNS (1) was already added in. This is
10186 why we are subtracting it back out. */
10187 if (outer_code == ZERO_EXTEND)
10189 *total = sparc_costs->int_zload - COSTS_N_INSNS (1);
10191 else if (outer_code == SIGN_EXTEND)
10193 *total = sparc_costs->int_sload - COSTS_N_INSNS (1);
10195 else if (float_mode_p)
10197 *total = sparc_costs->float_load;
10201 *total = sparc_costs->int_load;
10209 *total = sparc_costs->float_plusminus;
10211 *total = COSTS_N_INSNS (1);
10218 gcc_assert (float_mode_p);
10219 *total = sparc_costs->float_mul;
10222 if (GET_CODE (sub) == NEG)
10223 sub = XEXP (sub, 0);
10224 *total += rtx_cost (sub, FMA, 0, speed);
10227 if (GET_CODE (sub) == NEG)
10228 sub = XEXP (sub, 0);
10229 *total += rtx_cost (sub, FMA, 2, speed);
10235 *total = sparc_costs->float_mul;
10236 else if (! TARGET_HARD_MUL)
10237 *total = COSTS_N_INSNS (25);
10243 if (sparc_costs->int_mul_bit_factor)
10247 if (GET_CODE (XEXP (x, 1)) == CONST_INT)
10249 unsigned HOST_WIDE_INT value = INTVAL (XEXP (x, 1));
10250 for (nbits = 0; value != 0; value &= value - 1)
10253 else if (GET_CODE (XEXP (x, 1)) == CONST_DOUBLE
10254 && GET_MODE (XEXP (x, 1)) == VOIDmode)
10256 rtx x1 = XEXP (x, 1);
10257 unsigned HOST_WIDE_INT value1 = CONST_DOUBLE_LOW (x1);
10258 unsigned HOST_WIDE_INT value2 = CONST_DOUBLE_HIGH (x1);
10260 for (nbits = 0; value1 != 0; value1 &= value1 - 1)
10262 for (; value2 != 0; value2 &= value2 - 1)
10270 bit_cost = (nbits - 3) / sparc_costs->int_mul_bit_factor;
10271 bit_cost = COSTS_N_INSNS (bit_cost);
10274 if (mode == DImode)
10275 *total = sparc_costs->int_mulX + bit_cost;
10277 *total = sparc_costs->int_mul + bit_cost;
10284 *total = COSTS_N_INSNS (1) + sparc_costs->shift_penalty;
10293 if (mode == DFmode)
10294 *total = sparc_costs->float_div_df;
10296 *total = sparc_costs->float_div_sf;
10300 if (mode == DImode)
10301 *total = sparc_costs->int_divX;
10303 *total = sparc_costs->int_div;
10308 if (! float_mode_p)
10310 *total = COSTS_N_INSNS (1);
10317 case UNSIGNED_FLOAT:
10321 case FLOAT_TRUNCATE:
10322 *total = sparc_costs->float_move;
10326 if (mode == DFmode)
10327 *total = sparc_costs->float_sqrt_df;
10329 *total = sparc_costs->float_sqrt_sf;
10334 *total = sparc_costs->float_cmp;
10336 *total = COSTS_N_INSNS (1);
10341 *total = sparc_costs->float_cmove;
10343 *total = sparc_costs->int_cmove;
10347 /* Handle the NAND vector patterns. */
10348 if (sparc_vector_mode_supported_p (GET_MODE (x))
10349 && GET_CODE (XEXP (x, 0)) == NOT
10350 && GET_CODE (XEXP (x, 1)) == NOT)
10352 *total = COSTS_N_INSNS (1);
10363 /* Return true if CLASS is either GENERAL_REGS or I64_REGS. */
10366 general_or_i64_p (reg_class_t rclass)
10368 return (rclass == GENERAL_REGS || rclass == I64_REGS);
10371 /* Implement TARGET_REGISTER_MOVE_COST. */
10374 sparc_register_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED,
10375 reg_class_t from, reg_class_t to)
10377 bool need_memory = false;
10379 if (from == FPCC_REGS || to == FPCC_REGS)
10380 need_memory = true;
10381 else if ((FP_REG_CLASS_P (from) && general_or_i64_p (to))
10382 || (general_or_i64_p (from) && FP_REG_CLASS_P (to)))
10386 int size = GET_MODE_SIZE (mode);
10387 if (size == 8 || size == 4)
10389 if (! TARGET_ARCH32 || size == 4)
10395 need_memory = true;
10400 if (sparc_cpu == PROCESSOR_ULTRASPARC
10401 || sparc_cpu == PROCESSOR_ULTRASPARC3
10402 || sparc_cpu == PROCESSOR_NIAGARA
10403 || sparc_cpu == PROCESSOR_NIAGARA2
10404 || sparc_cpu == PROCESSOR_NIAGARA3
10405 || sparc_cpu == PROCESSOR_NIAGARA4)
10414 /* Emit the sequence of insns SEQ while preserving the registers REG and REG2.
10415 This is achieved by means of a manual dynamic stack space allocation in
10416 the current frame. We make the assumption that SEQ doesn't contain any
10417 function calls, with the possible exception of calls to the GOT helper. */
10420 emit_and_preserve (rtx seq, rtx reg, rtx reg2)
10422 /* We must preserve the lowest 16 words for the register save area. */
10423 HOST_WIDE_INT offset = 16*UNITS_PER_WORD;
10424 /* We really need only 2 words of fresh stack space. */
10425 HOST_WIDE_INT size = SPARC_STACK_ALIGN (offset + 2*UNITS_PER_WORD);
10428 = gen_rtx_MEM (word_mode, plus_constant (Pmode, stack_pointer_rtx,
10429 SPARC_STACK_BIAS + offset));
10431 emit_insn (gen_stack_pointer_dec (GEN_INT (size)));
10432 emit_insn (gen_rtx_SET (VOIDmode, slot, reg));
10434 emit_insn (gen_rtx_SET (VOIDmode,
10435 adjust_address (slot, word_mode, UNITS_PER_WORD),
10439 emit_insn (gen_rtx_SET (VOIDmode,
10441 adjust_address (slot, word_mode, UNITS_PER_WORD)));
10442 emit_insn (gen_rtx_SET (VOIDmode, reg, slot));
10443 emit_insn (gen_stack_pointer_inc (GEN_INT (size)));
10446 /* Output the assembler code for a thunk function. THUNK_DECL is the
10447 declaration for the thunk function itself, FUNCTION is the decl for
10448 the target function. DELTA is an immediate constant offset to be
10449 added to THIS. If VCALL_OFFSET is nonzero, the word at address
10450 (*THIS + VCALL_OFFSET) should be additionally added to THIS. */
10453 sparc_output_mi_thunk (FILE *file, tree thunk_fndecl ATTRIBUTE_UNUSED,
10454 HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset,
10457 rtx this_rtx, insn, funexp;
10458 unsigned int int_arg_first;
10460 reload_completed = 1;
10461 epilogue_completed = 1;
10463 emit_note (NOTE_INSN_PROLOGUE_END);
10467 sparc_leaf_function_p = 1;
10469 int_arg_first = SPARC_OUTGOING_INT_ARG_FIRST;
10471 else if (flag_delayed_branch)
10473 /* We will emit a regular sibcall below, so we need to instruct
10474 output_sibcall that we are in a leaf function. */
10475 sparc_leaf_function_p = current_function_uses_only_leaf_regs = 1;
10477 /* This will cause final.c to invoke leaf_renumber_regs so we
10478 must behave as if we were in a not-yet-leafified function. */
10479 int_arg_first = SPARC_INCOMING_INT_ARG_FIRST;
10483 /* We will emit the sibcall manually below, so we will need to
10484 manually spill non-leaf registers. */
10485 sparc_leaf_function_p = current_function_uses_only_leaf_regs = 0;
10487 /* We really are in a leaf function. */
10488 int_arg_first = SPARC_OUTGOING_INT_ARG_FIRST;
10491 /* Find the "this" pointer. Normally in %o0, but in ARCH64 if the function
10492 returns a structure, the structure return pointer is there instead. */
10494 && aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function))
10495 this_rtx = gen_rtx_REG (Pmode, int_arg_first + 1);
10497 this_rtx = gen_rtx_REG (Pmode, int_arg_first);
10499 /* Add DELTA. When possible use a plain add, otherwise load it into
10500 a register first. */
10503 rtx delta_rtx = GEN_INT (delta);
10505 if (! SPARC_SIMM13_P (delta))
10507 rtx scratch = gen_rtx_REG (Pmode, 1);
10508 emit_move_insn (scratch, delta_rtx);
10509 delta_rtx = scratch;
10512 /* THIS_RTX += DELTA. */
10513 emit_insn (gen_add2_insn (this_rtx, delta_rtx));
10516 /* Add the word at address (*THIS_RTX + VCALL_OFFSET). */
10519 rtx vcall_offset_rtx = GEN_INT (vcall_offset);
10520 rtx scratch = gen_rtx_REG (Pmode, 1);
10522 gcc_assert (vcall_offset < 0);
10524 /* SCRATCH = *THIS_RTX. */
10525 emit_move_insn (scratch, gen_rtx_MEM (Pmode, this_rtx));
10527 /* Prepare for adding VCALL_OFFSET. The difficulty is that we
10528 may not have any available scratch register at this point. */
10529 if (SPARC_SIMM13_P (vcall_offset))
10531 /* This is the case if ARCH64 (unless -ffixed-g5 is passed). */
10532 else if (! fixed_regs[5]
10533 /* The below sequence is made up of at least 2 insns,
10534 while the default method may need only one. */
10535 && vcall_offset < -8192)
10537 rtx scratch2 = gen_rtx_REG (Pmode, 5);
10538 emit_move_insn (scratch2, vcall_offset_rtx);
10539 vcall_offset_rtx = scratch2;
10543 rtx increment = GEN_INT (-4096);
10545 /* VCALL_OFFSET is a negative number whose typical range can be
10546 estimated as -32768..0 in 32-bit mode. In almost all cases
10547 it is therefore cheaper to emit multiple add insns than
10548 spilling and loading the constant into a register (at least
10550 while (! SPARC_SIMM13_P (vcall_offset))
10552 emit_insn (gen_add2_insn (scratch, increment));
10553 vcall_offset += 4096;
10555 vcall_offset_rtx = GEN_INT (vcall_offset); /* cannot be 0 */
10558 /* SCRATCH = *(*THIS_RTX + VCALL_OFFSET). */
10559 emit_move_insn (scratch, gen_rtx_MEM (Pmode,
10560 gen_rtx_PLUS (Pmode,
10562 vcall_offset_rtx)));
10564 /* THIS_RTX += *(*THIS_RTX + VCALL_OFFSET). */
10565 emit_insn (gen_add2_insn (this_rtx, scratch));
10568 /* Generate a tail call to the target function. */
10569 if (! TREE_USED (function))
10571 assemble_external (function);
10572 TREE_USED (function) = 1;
10574 funexp = XEXP (DECL_RTL (function), 0);
10576 if (flag_delayed_branch)
10578 funexp = gen_rtx_MEM (FUNCTION_MODE, funexp);
10579 insn = emit_call_insn (gen_sibcall (funexp));
10580 SIBLING_CALL_P (insn) = 1;
10584 /* The hoops we have to jump through in order to generate a sibcall
10585 without using delay slots... */
10586 rtx spill_reg, seq, scratch = gen_rtx_REG (Pmode, 1);
10590 spill_reg = gen_rtx_REG (word_mode, 15); /* %o7 */
10592 load_got_register (); /* clobbers %o7 */
10593 scratch = sparc_legitimize_pic_address (funexp, scratch);
10594 seq = get_insns ();
10596 emit_and_preserve (seq, spill_reg, pic_offset_table_rtx);
10598 else if (TARGET_ARCH32)
10600 emit_insn (gen_rtx_SET (VOIDmode,
10602 gen_rtx_HIGH (SImode, funexp)));
10603 emit_insn (gen_rtx_SET (VOIDmode,
10605 gen_rtx_LO_SUM (SImode, scratch, funexp)));
10607 else /* TARGET_ARCH64 */
10609 switch (sparc_cmodel)
10613 /* The destination can serve as a temporary. */
10614 sparc_emit_set_symbolic_const64 (scratch, funexp, scratch);
10619 /* The destination cannot serve as a temporary. */
10620 spill_reg = gen_rtx_REG (DImode, 15); /* %o7 */
10622 sparc_emit_set_symbolic_const64 (scratch, funexp, spill_reg);
10623 seq = get_insns ();
10625 emit_and_preserve (seq, spill_reg, 0);
10629 gcc_unreachable ();
10633 emit_jump_insn (gen_indirect_jump (scratch));
10638 /* Run just enough of rest_of_compilation to get the insns emitted.
10639 There's not really enough bulk here to make other passes such as
10640 instruction scheduling worth while. Note that use_thunk calls
10641 assemble_start_function and assemble_end_function. */
10642 insn = get_insns ();
10643 insn_locators_alloc ();
10644 shorten_branches (insn);
10645 final_start_function (insn, file, 1);
10646 final (insn, file, 1);
10647 final_end_function ();
10649 reload_completed = 0;
10650 epilogue_completed = 0;
10653 /* Return true if sparc_output_mi_thunk would be able to output the
10654 assembler code for the thunk function specified by the arguments
10655 it is passed, and false otherwise. */
10657 sparc_can_output_mi_thunk (const_tree thunk_fndecl ATTRIBUTE_UNUSED,
10658 HOST_WIDE_INT delta ATTRIBUTE_UNUSED,
10659 HOST_WIDE_INT vcall_offset,
10660 const_tree function ATTRIBUTE_UNUSED)
10662 /* Bound the loop used in the default method above. */
10663 return (vcall_offset >= -32768 || ! fixed_regs[5]);
10666 /* We use the machine specific reorg pass to enable workarounds for errata. */
10673 /* The only erratum we handle for now is that of the AT697F processor. */
10674 if (!sparc_fix_at697f)
10677 /* We need to have the (essentially) final form of the insn stream in order
10678 to properly detect the various hazards. Run delay slot scheduling. */
10679 if (optimize > 0 && flag_delayed_branch)
10680 dbr_schedule (get_insns ());
10682 /* Now look for specific patterns in the insn stream. */
10683 for (insn = get_insns (); insn; insn = next)
10685 bool insert_nop = false;
10688 /* Look for a single-word load into an odd-numbered FP register. */
10689 if (NONJUMP_INSN_P (insn)
10690 && (set = single_set (insn)) != NULL_RTX
10691 && GET_MODE_SIZE (GET_MODE (SET_SRC (set))) == 4
10692 && MEM_P (SET_SRC (set))
10693 && REG_P (SET_DEST (set))
10694 && REGNO (SET_DEST (set)) > 31
10695 && REGNO (SET_DEST (set)) % 2 != 0)
10697 /* The wrong dependency is on the enclosing double register. */
10698 unsigned int x = REGNO (SET_DEST (set)) - 1;
10699 unsigned int src1, src2, dest;
10702 /* If the insn has a delay slot, then it cannot be problematic. */
10703 next = next_active_insn (insn);
10704 if (NONJUMP_INSN_P (next) && GET_CODE (PATTERN (next)) == SEQUENCE)
10708 extract_insn (next);
10709 code = INSN_CODE (next);
10714 case CODE_FOR_adddf3:
10715 case CODE_FOR_subdf3:
10716 case CODE_FOR_muldf3:
10717 case CODE_FOR_divdf3:
10718 dest = REGNO (recog_data.operand[0]);
10719 src1 = REGNO (recog_data.operand[1]);
10720 src2 = REGNO (recog_data.operand[2]);
10724 ld [address], %fx+1
10725 FPOPd %f{x,y}, %f{y,x}, %f{x,y} */
10726 if ((src1 == x || src2 == x)
10727 && (dest == src1 || dest == src2))
10733 ld [address], %fx+1
10734 FPOPd %fx, %fx, %fx */
10737 && (code == CODE_FOR_adddf3 || code == CODE_FOR_muldf3))
10742 case CODE_FOR_sqrtdf2:
10743 dest = REGNO (recog_data.operand[0]);
10744 src1 = REGNO (recog_data.operand[1]);
10746 ld [address], %fx+1
10748 if (src1 == x && dest == src1)
10757 next = NEXT_INSN (insn);
10760 emit_insn_after (gen_nop (), insn);
10764 /* How to allocate a 'struct machine_function'. */
10766 static struct machine_function *
10767 sparc_init_machine_status (void)
10769 return ggc_alloc_cleared_machine_function ();
10772 /* Locate some local-dynamic symbol still in use by this function
10773 so that we can print its name in local-dynamic base patterns. */
10775 static const char *
10776 get_some_local_dynamic_name (void)
10780 if (cfun->machine->some_ld_name)
10781 return cfun->machine->some_ld_name;
10783 for (insn = get_insns (); insn ; insn = NEXT_INSN (insn))
10785 && for_each_rtx (&PATTERN (insn), get_some_local_dynamic_name_1, 0))
10786 return cfun->machine->some_ld_name;
10788 gcc_unreachable ();
10792 get_some_local_dynamic_name_1 (rtx *px, void *data ATTRIBUTE_UNUSED)
10797 && GET_CODE (x) == SYMBOL_REF
10798 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_DYNAMIC)
10800 cfun->machine->some_ld_name = XSTR (x, 0);
10807 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
10808 We need to emit DTP-relative relocations. */
10811 sparc_output_dwarf_dtprel (FILE *file, int size, rtx x)
10816 fputs ("\t.word\t%r_tls_dtpoff32(", file);
10819 fputs ("\t.xword\t%r_tls_dtpoff64(", file);
10822 gcc_unreachable ();
10824 output_addr_const (file, x);
10828 /* Do whatever processing is required at the end of a file. */
10831 sparc_file_end (void)
10833 /* If we need to emit the special GOT helper function, do so now. */
10834 if (got_helper_rtx)
10836 const char *name = XSTR (got_helper_rtx, 0);
10837 const char *reg_name = reg_names[GLOBAL_OFFSET_TABLE_REGNUM];
10838 #ifdef DWARF2_UNWIND_INFO
10842 if (USE_HIDDEN_LINKONCE)
10844 tree decl = build_decl (BUILTINS_LOCATION, FUNCTION_DECL,
10845 get_identifier (name),
10846 build_function_type_list (void_type_node,
10848 DECL_RESULT (decl) = build_decl (BUILTINS_LOCATION, RESULT_DECL,
10849 NULL_TREE, void_type_node);
10850 TREE_PUBLIC (decl) = 1;
10851 TREE_STATIC (decl) = 1;
10852 make_decl_one_only (decl, DECL_ASSEMBLER_NAME (decl));
10853 DECL_VISIBILITY (decl) = VISIBILITY_HIDDEN;
10854 DECL_VISIBILITY_SPECIFIED (decl) = 1;
10855 resolve_unique_section (decl, 0, flag_function_sections);
10856 allocate_struct_function (decl, true);
10857 cfun->is_thunk = 1;
10858 current_function_decl = decl;
10859 init_varasm_status ();
10860 assemble_start_function (decl, name);
10864 const int align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
10865 switch_to_section (text_section);
10867 ASM_OUTPUT_ALIGN (asm_out_file, align);
10868 ASM_OUTPUT_LABEL (asm_out_file, name);
10871 #ifdef DWARF2_UNWIND_INFO
10872 do_cfi = dwarf2out_do_cfi_asm ();
10874 fprintf (asm_out_file, "\t.cfi_startproc\n");
10876 if (flag_delayed_branch)
10877 fprintf (asm_out_file, "\tjmp\t%%o7+8\n\t add\t%%o7, %s, %s\n",
10878 reg_name, reg_name);
10880 fprintf (asm_out_file, "\tadd\t%%o7, %s, %s\n\tjmp\t%%o7+8\n\t nop\n",
10881 reg_name, reg_name);
10882 #ifdef DWARF2_UNWIND_INFO
10884 fprintf (asm_out_file, "\t.cfi_endproc\n");
10888 if (NEED_INDICATE_EXEC_STACK)
10889 file_end_indicate_exec_stack ();
10891 #ifdef TARGET_SOLARIS
10892 solaris_file_end ();
10896 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
10897 /* Implement TARGET_MANGLE_TYPE. */
10899 static const char *
10900 sparc_mangle_type (const_tree type)
10903 && TYPE_MAIN_VARIANT (type) == long_double_type_node
10904 && TARGET_LONG_DOUBLE_128)
10907 /* For all other types, use normal C++ mangling. */
10912 /* Expand a membar instruction for various use cases. Both the LOAD_STORE
10913 and BEFORE_AFTER arguments of the form X_Y. They are two-bit masks where
10914 bit 0 indicates that X is true, and bit 1 indicates Y is true. */
10917 sparc_emit_membar_for_model (enum memmodel model,
10918 int load_store, int before_after)
10920 /* Bits for the MEMBAR mmask field. */
10921 const int LoadLoad = 1;
10922 const int StoreLoad = 2;
10923 const int LoadStore = 4;
10924 const int StoreStore = 8;
10926 int mm = 0, implied = 0;
10928 switch (sparc_memory_model)
10931 /* Sequential Consistency. All memory transactions are immediately
10932 visible in sequential execution order. No barriers needed. */
10933 implied = LoadLoad | StoreLoad | LoadStore | StoreStore;
10937 /* Total Store Ordering: all memory transactions with store semantics
10938 are followed by an implied StoreStore. */
10939 implied |= StoreStore;
10943 /* Partial Store Ordering: all memory transactions with load semantics
10944 are followed by an implied LoadLoad | LoadStore. */
10945 implied |= LoadLoad | LoadStore;
10947 /* If we're not looking for a raw barrer (before+after), then atomic
10948 operations get the benefit of being both load and store. */
10949 if (load_store == 3 && before_after == 2)
10950 implied |= StoreLoad | StoreStore;
10954 /* Relaxed Memory Ordering: no implicit bits. */
10958 gcc_unreachable ();
10961 if (before_after & 1)
10963 if (model == MEMMODEL_ACQUIRE
10964 || model == MEMMODEL_ACQ_REL
10965 || model == MEMMODEL_SEQ_CST)
10967 if (load_store & 1)
10968 mm |= LoadLoad | LoadStore;
10969 if (load_store & 2)
10970 mm |= StoreLoad | StoreStore;
10973 if (before_after & 2)
10975 if (model == MEMMODEL_RELEASE
10976 || model == MEMMODEL_ACQ_REL
10977 || model == MEMMODEL_SEQ_CST)
10979 if (load_store & 1)
10980 mm |= LoadLoad | StoreLoad;
10981 if (load_store & 2)
10982 mm |= LoadStore | StoreStore;
10986 /* Remove the bits implied by the system memory model. */
10989 /* For raw barriers (before+after), always emit a barrier.
10990 This will become a compile-time barrier if needed. */
10991 if (mm || before_after == 3)
10992 emit_insn (gen_membar (GEN_INT (mm)));
10995 /* Expand code to perform a 8 or 16-bit compare and swap by doing 32-bit
10996 compare and swap on the word containing the byte or half-word. */
10999 sparc_expand_compare_and_swap_12 (rtx bool_result, rtx result, rtx mem,
11000 rtx oldval, rtx newval)
11002 rtx addr1 = force_reg (Pmode, XEXP (mem, 0));
11003 rtx addr = gen_reg_rtx (Pmode);
11004 rtx off = gen_reg_rtx (SImode);
11005 rtx oldv = gen_reg_rtx (SImode);
11006 rtx newv = gen_reg_rtx (SImode);
11007 rtx oldvalue = gen_reg_rtx (SImode);
11008 rtx newvalue = gen_reg_rtx (SImode);
11009 rtx res = gen_reg_rtx (SImode);
11010 rtx resv = gen_reg_rtx (SImode);
11011 rtx memsi, val, mask, end_label, loop_label, cc;
11013 emit_insn (gen_rtx_SET (VOIDmode, addr,
11014 gen_rtx_AND (Pmode, addr1, GEN_INT (-4))));
11016 if (Pmode != SImode)
11017 addr1 = gen_lowpart (SImode, addr1);
11018 emit_insn (gen_rtx_SET (VOIDmode, off,
11019 gen_rtx_AND (SImode, addr1, GEN_INT (3))));
11021 memsi = gen_rtx_MEM (SImode, addr);
11022 set_mem_alias_set (memsi, ALIAS_SET_MEMORY_BARRIER);
11023 MEM_VOLATILE_P (memsi) = MEM_VOLATILE_P (mem);
11025 val = copy_to_reg (memsi);
11027 emit_insn (gen_rtx_SET (VOIDmode, off,
11028 gen_rtx_XOR (SImode, off,
11029 GEN_INT (GET_MODE (mem) == QImode
11032 emit_insn (gen_rtx_SET (VOIDmode, off,
11033 gen_rtx_ASHIFT (SImode, off, GEN_INT (3))));
11035 if (GET_MODE (mem) == QImode)
11036 mask = force_reg (SImode, GEN_INT (0xff));
11038 mask = force_reg (SImode, GEN_INT (0xffff));
11040 emit_insn (gen_rtx_SET (VOIDmode, mask,
11041 gen_rtx_ASHIFT (SImode, mask, off)));
11043 emit_insn (gen_rtx_SET (VOIDmode, val,
11044 gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask),
11047 oldval = gen_lowpart (SImode, oldval);
11048 emit_insn (gen_rtx_SET (VOIDmode, oldv,
11049 gen_rtx_ASHIFT (SImode, oldval, off)));
11051 newval = gen_lowpart_common (SImode, newval);
11052 emit_insn (gen_rtx_SET (VOIDmode, newv,
11053 gen_rtx_ASHIFT (SImode, newval, off)));
11055 emit_insn (gen_rtx_SET (VOIDmode, oldv,
11056 gen_rtx_AND (SImode, oldv, mask)));
11058 emit_insn (gen_rtx_SET (VOIDmode, newv,
11059 gen_rtx_AND (SImode, newv, mask)));
11061 end_label = gen_label_rtx ();
11062 loop_label = gen_label_rtx ();
11063 emit_label (loop_label);
11065 emit_insn (gen_rtx_SET (VOIDmode, oldvalue,
11066 gen_rtx_IOR (SImode, oldv, val)));
11068 emit_insn (gen_rtx_SET (VOIDmode, newvalue,
11069 gen_rtx_IOR (SImode, newv, val)));
11071 emit_move_insn (bool_result, const1_rtx);
11073 emit_insn (gen_atomic_compare_and_swapsi_1 (res, memsi, oldvalue, newvalue));
11075 emit_cmp_and_jump_insns (res, oldvalue, EQ, NULL, SImode, 0, end_label);
11077 emit_insn (gen_rtx_SET (VOIDmode, resv,
11078 gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask),
11081 emit_move_insn (bool_result, const0_rtx);
11083 cc = gen_compare_reg_1 (NE, resv, val);
11084 emit_insn (gen_rtx_SET (VOIDmode, val, resv));
11086 /* Use cbranchcc4 to separate the compare and branch! */
11087 emit_jump_insn (gen_cbranchcc4 (gen_rtx_NE (VOIDmode, cc, const0_rtx),
11088 cc, const0_rtx, loop_label));
11090 emit_label (end_label);
11092 emit_insn (gen_rtx_SET (VOIDmode, res,
11093 gen_rtx_AND (SImode, res, mask)));
11095 emit_insn (gen_rtx_SET (VOIDmode, res,
11096 gen_rtx_LSHIFTRT (SImode, res, off)));
11098 emit_move_insn (result, gen_lowpart (GET_MODE (result), res));
11101 /* Expand code to perform a compare-and-swap. */
11104 sparc_expand_compare_and_swap (rtx operands[])
11106 rtx bval, retval, mem, oldval, newval;
11107 enum machine_mode mode;
11108 enum memmodel model;
11110 bval = operands[0];
11111 retval = operands[1];
11113 oldval = operands[3];
11114 newval = operands[4];
11115 model = (enum memmodel) INTVAL (operands[6]);
11116 mode = GET_MODE (mem);
11118 sparc_emit_membar_for_model (model, 3, 1);
11120 if (reg_overlap_mentioned_p (retval, oldval))
11121 oldval = copy_to_reg (oldval);
11123 if (mode == QImode || mode == HImode)
11124 sparc_expand_compare_and_swap_12 (bval, retval, mem, oldval, newval);
11127 rtx (*gen) (rtx, rtx, rtx, rtx);
11130 if (mode == SImode)
11131 gen = gen_atomic_compare_and_swapsi_1;
11133 gen = gen_atomic_compare_and_swapdi_1;
11134 emit_insn (gen (retval, mem, oldval, newval));
11136 x = emit_store_flag (bval, EQ, retval, oldval, mode, 1, 1);
11138 convert_move (bval, x, 1);
11141 sparc_emit_membar_for_model (model, 3, 2);
11145 sparc_expand_vec_perm_bmask (enum machine_mode vmode, rtx sel)
11149 sel = gen_lowpart (DImode, sel);
11153 /* inp = xxxxxxxAxxxxxxxB */
11154 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (16),
11155 NULL_RTX, 1, OPTAB_DIRECT);
11156 /* t_1 = ....xxxxxxxAxxx. */
11157 sel = expand_simple_binop (SImode, AND, gen_lowpart (SImode, sel),
11158 GEN_INT (3), NULL_RTX, 1, OPTAB_DIRECT);
11159 t_1 = expand_simple_binop (SImode, AND, gen_lowpart (SImode, t_1),
11160 GEN_INT (0x30000), NULL_RTX, 1, OPTAB_DIRECT);
11161 /* sel = .......B */
11162 /* t_1 = ...A.... */
11163 sel = expand_simple_binop (SImode, IOR, sel, t_1, sel, 1, OPTAB_DIRECT);
11164 /* sel = ...A...B */
11165 sel = expand_mult (SImode, sel, GEN_INT (0x4444), sel, 1);
11166 /* sel = AAAABBBB * 4 */
11167 t_1 = force_reg (SImode, GEN_INT (0x01230123));
11168 /* sel = { A*4, A*4+1, A*4+2, ... } */
11172 /* inp = xxxAxxxBxxxCxxxD */
11173 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (8),
11174 NULL_RTX, 1, OPTAB_DIRECT);
11175 t_2 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (16),
11176 NULL_RTX, 1, OPTAB_DIRECT);
11177 t_3 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (24),
11178 NULL_RTX, 1, OPTAB_DIRECT);
11179 /* t_1 = ..xxxAxxxBxxxCxx */
11180 /* t_2 = ....xxxAxxxBxxxC */
11181 /* t_3 = ......xxxAxxxBxx */
11182 sel = expand_simple_binop (SImode, AND, gen_lowpart (SImode, sel),
11184 NULL_RTX, 1, OPTAB_DIRECT);
11185 t_1 = expand_simple_binop (SImode, AND, gen_lowpart (SImode, t_1),
11187 NULL_RTX, 1, OPTAB_DIRECT);
11188 t_2 = expand_simple_binop (SImode, AND, gen_lowpart (SImode, t_2),
11189 GEN_INT (0x070000),
11190 NULL_RTX, 1, OPTAB_DIRECT);
11191 t_3 = expand_simple_binop (SImode, AND, gen_lowpart (SImode, t_3),
11192 GEN_INT (0x07000000),
11193 NULL_RTX, 1, OPTAB_DIRECT);
11194 /* sel = .......D */
11195 /* t_1 = .....C.. */
11196 /* t_2 = ...B.... */
11197 /* t_3 = .A...... */
11198 sel = expand_simple_binop (SImode, IOR, sel, t_1, sel, 1, OPTAB_DIRECT);
11199 t_2 = expand_simple_binop (SImode, IOR, t_2, t_3, t_2, 1, OPTAB_DIRECT);
11200 sel = expand_simple_binop (SImode, IOR, sel, t_2, sel, 1, OPTAB_DIRECT);
11201 /* sel = .A.B.C.D */
11202 sel = expand_mult (SImode, sel, GEN_INT (0x22), sel, 1);
11203 /* sel = AABBCCDD * 2 */
11204 t_1 = force_reg (SImode, GEN_INT (0x01010101));
11205 /* sel = { A*2, A*2+1, B*2, B*2+1, ... } */
11209 /* input = xAxBxCxDxExFxGxH */
11210 sel = expand_simple_binop (DImode, AND, sel,
11211 GEN_INT ((HOST_WIDE_INT)0x0f0f0f0f << 32
11213 NULL_RTX, 1, OPTAB_DIRECT);
11214 /* sel = .A.B.C.D.E.F.G.H */
11215 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (4),
11216 NULL_RTX, 1, OPTAB_DIRECT);
11217 /* t_1 = ..A.B.C.D.E.F.G. */
11218 sel = expand_simple_binop (DImode, IOR, sel, t_1,
11219 NULL_RTX, 1, OPTAB_DIRECT);
11220 /* sel = .AABBCCDDEEFFGGH */
11221 sel = expand_simple_binop (DImode, AND, sel,
11222 GEN_INT ((HOST_WIDE_INT)0xff00ff << 32
11224 NULL_RTX, 1, OPTAB_DIRECT);
11225 /* sel = ..AB..CD..EF..GH */
11226 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (8),
11227 NULL_RTX, 1, OPTAB_DIRECT);
11228 /* t_1 = ....AB..CD..EF.. */
11229 sel = expand_simple_binop (DImode, IOR, sel, t_1,
11230 NULL_RTX, 1, OPTAB_DIRECT);
11231 /* sel = ..ABABCDCDEFEFGH */
11232 sel = expand_simple_binop (DImode, AND, sel,
11233 GEN_INT ((HOST_WIDE_INT)0xffff << 32 | 0xffff),
11234 NULL_RTX, 1, OPTAB_DIRECT);
11235 /* sel = ....ABCD....EFGH */
11236 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (16),
11237 NULL_RTX, 1, OPTAB_DIRECT);
11238 /* t_1 = ........ABCD.... */
11239 sel = gen_lowpart (SImode, sel);
11240 t_1 = gen_lowpart (SImode, t_1);
11244 gcc_unreachable ();
11247 /* Always perform the final addition/merge within the bmask insn. */
11248 emit_insn (gen_bmasksi_vis (gen_reg_rtx (SImode), sel, t_1));
11251 /* Implement TARGET_FRAME_POINTER_REQUIRED. */
11254 sparc_frame_pointer_required (void)
11256 /* If the stack pointer is dynamically modified in the function, it cannot
11257 serve as the frame pointer. */
11258 if (cfun->calls_alloca)
11261 /* If the function receives nonlocal gotos, it needs to save the frame
11262 pointer in the nonlocal_goto_save_area object. */
11263 if (cfun->has_nonlocal_label)
11266 /* In flat mode, that's it. */
11270 /* Otherwise, the frame pointer is required if the function isn't leaf. */
11271 return !(current_function_is_leaf && only_leaf_regs_used ());
11274 /* The way this is structured, we can't eliminate SFP in favor of SP
11275 if the frame pointer is required: we want to use the SFP->HFP elimination
11276 in that case. But the test in update_eliminables doesn't know we are
11277 assuming below that we only do the former elimination. */
11280 sparc_can_eliminate (const int from ATTRIBUTE_UNUSED, const int to)
11282 return to == HARD_FRAME_POINTER_REGNUM || !sparc_frame_pointer_required ();
11285 /* Return the hard frame pointer directly to bypass the stack bias. */
11288 sparc_builtin_setjmp_frame_value (void)
11290 return hard_frame_pointer_rtx;
11293 /* If !TARGET_FPU, then make the fp registers and fp cc regs fixed so that
11294 they won't be allocated. */
11297 sparc_conditional_register_usage (void)
11299 if (PIC_OFFSET_TABLE_REGNUM != INVALID_REGNUM)
11301 fixed_regs[PIC_OFFSET_TABLE_REGNUM] = 1;
11302 call_used_regs[PIC_OFFSET_TABLE_REGNUM] = 1;
11304 /* If the user has passed -f{fixed,call-{used,saved}}-g5 */
11305 /* then honor it. */
11306 if (TARGET_ARCH32 && fixed_regs[5])
11308 else if (TARGET_ARCH64 && fixed_regs[5] == 2)
11313 for (regno = SPARC_FIRST_V9_FP_REG;
11314 regno <= SPARC_LAST_V9_FP_REG;
11316 fixed_regs[regno] = 1;
11317 /* %fcc0 is used by v8 and v9. */
11318 for (regno = SPARC_FIRST_V9_FCC_REG + 1;
11319 regno <= SPARC_LAST_V9_FCC_REG;
11321 fixed_regs[regno] = 1;
11326 for (regno = 32; regno < SPARC_LAST_V9_FCC_REG; regno++)
11327 fixed_regs[regno] = 1;
11329 /* If the user has passed -f{fixed,call-{used,saved}}-g2 */
11330 /* then honor it. Likewise with g3 and g4. */
11331 if (fixed_regs[2] == 2)
11332 fixed_regs[2] = ! TARGET_APP_REGS;
11333 if (fixed_regs[3] == 2)
11334 fixed_regs[3] = ! TARGET_APP_REGS;
11335 if (TARGET_ARCH32 && fixed_regs[4] == 2)
11336 fixed_regs[4] = ! TARGET_APP_REGS;
11337 else if (TARGET_CM_EMBMEDANY)
11339 else if (fixed_regs[4] == 2)
11344 /* Disable leaf functions. */
11345 memset (sparc_leaf_regs, 0, FIRST_PSEUDO_REGISTER);
11346 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
11347 leaf_reg_remap [regno] = regno;
11350 global_regs[SPARC_GSR_REG] = 1;
11353 /* Implement TARGET_PREFERRED_RELOAD_CLASS:
11355 - We can't load constants into FP registers.
11356 - We can't load FP constants into integer registers when soft-float,
11357 because there is no soft-float pattern with a r/F constraint.
11358 - We can't load FP constants into integer registers for TFmode unless
11359 it is 0.0L, because there is no movtf pattern with a r/F constraint.
11360 - Try and reload integer constants (symbolic or otherwise) back into
11361 registers directly, rather than having them dumped to memory. */
11364 sparc_preferred_reload_class (rtx x, reg_class_t rclass)
11366 enum machine_mode mode = GET_MODE (x);
11367 if (CONSTANT_P (x))
11369 if (FP_REG_CLASS_P (rclass)
11370 || rclass == GENERAL_OR_FP_REGS
11371 || rclass == GENERAL_OR_EXTRA_FP_REGS
11372 || (GET_MODE_CLASS (mode) == MODE_FLOAT && ! TARGET_FPU)
11373 || (mode == TFmode && ! const_zero_operand (x, mode)))
11376 if (GET_MODE_CLASS (mode) == MODE_INT)
11377 return GENERAL_REGS;
11379 if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT)
11381 if (! FP_REG_CLASS_P (rclass)
11382 || !(const_zero_operand (x, mode)
11383 || const_all_ones_operand (x, mode)))
11390 && (rclass == EXTRA_FP_REGS
11391 || rclass == GENERAL_OR_EXTRA_FP_REGS))
11393 int regno = true_regnum (x);
11395 if (SPARC_INT_REG_P (regno))
11396 return (rclass == EXTRA_FP_REGS
11397 ? FP_REGS : GENERAL_OR_FP_REGS);
11403 /* Output a wide multiply instruction in V8+ mode. INSN is the instruction,
11404 OPERANDS are its operands and OPCODE is the mnemonic to be used. */
11407 output_v8plus_mult (rtx insn, rtx *operands, const char *opcode)
11411 gcc_assert (! TARGET_ARCH64);
11413 if (sparc_check_64 (operands[1], insn) <= 0)
11414 output_asm_insn ("srl\t%L1, 0, %L1", operands);
11415 if (which_alternative == 1)
11416 output_asm_insn ("sllx\t%H1, 32, %H1", operands);
11417 if (GET_CODE (operands[2]) == CONST_INT)
11419 if (which_alternative == 1)
11421 output_asm_insn ("or\t%L1, %H1, %H1", operands);
11422 sprintf (mulstr, "%s\t%%H1, %%2, %%L0", opcode);
11423 output_asm_insn (mulstr, operands);
11424 return "srlx\t%L0, 32, %H0";
11428 output_asm_insn ("sllx\t%H1, 32, %3", operands);
11429 output_asm_insn ("or\t%L1, %3, %3", operands);
11430 sprintf (mulstr, "%s\t%%3, %%2, %%3", opcode);
11431 output_asm_insn (mulstr, operands);
11432 output_asm_insn ("srlx\t%3, 32, %H0", operands);
11433 return "mov\t%3, %L0";
11436 else if (rtx_equal_p (operands[1], operands[2]))
11438 if (which_alternative == 1)
11440 output_asm_insn ("or\t%L1, %H1, %H1", operands);
11441 sprintf (mulstr, "%s\t%%H1, %%H1, %%L0", opcode);
11442 output_asm_insn (mulstr, operands);
11443 return "srlx\t%L0, 32, %H0";
11447 output_asm_insn ("sllx\t%H1, 32, %3", operands);
11448 output_asm_insn ("or\t%L1, %3, %3", operands);
11449 sprintf (mulstr, "%s\t%%3, %%3, %%3", opcode);
11450 output_asm_insn (mulstr, operands);
11451 output_asm_insn ("srlx\t%3, 32, %H0", operands);
11452 return "mov\t%3, %L0";
11455 if (sparc_check_64 (operands[2], insn) <= 0)
11456 output_asm_insn ("srl\t%L2, 0, %L2", operands);
11457 if (which_alternative == 1)
11459 output_asm_insn ("or\t%L1, %H1, %H1", operands);
11460 output_asm_insn ("sllx\t%H2, 32, %L1", operands);
11461 output_asm_insn ("or\t%L2, %L1, %L1", operands);
11462 sprintf (mulstr, "%s\t%%H1, %%L1, %%L0", opcode);
11463 output_asm_insn (mulstr, operands);
11464 return "srlx\t%L0, 32, %H0";
11468 output_asm_insn ("sllx\t%H1, 32, %3", operands);
11469 output_asm_insn ("sllx\t%H2, 32, %4", operands);
11470 output_asm_insn ("or\t%L1, %3, %3", operands);
11471 output_asm_insn ("or\t%L2, %4, %4", operands);
11472 sprintf (mulstr, "%s\t%%3, %%4, %%3", opcode);
11473 output_asm_insn (mulstr, operands);
11474 output_asm_insn ("srlx\t%3, 32, %H0", operands);
11475 return "mov\t%3, %L0";
11479 /* Subroutine of sparc_expand_vector_init. Emit code to initialize
11480 all fields of TARGET to ELT by means of VIS2 BSHUFFLE insn. MODE
11481 and INNER_MODE are the modes describing TARGET. */
11484 vector_init_bshuffle (rtx target, rtx elt, enum machine_mode mode,
11485 enum machine_mode inner_mode)
11487 rtx t1, final_insn;
11490 t1 = gen_reg_rtx (mode);
11492 elt = convert_modes (SImode, inner_mode, elt, true);
11493 emit_move_insn (gen_lowpart(SImode, t1), elt);
11498 final_insn = gen_bshufflev2si_vis (target, t1, t1);
11499 bmask = 0x45674567;
11502 final_insn = gen_bshufflev4hi_vis (target, t1, t1);
11503 bmask = 0x67676767;
11506 final_insn = gen_bshufflev8qi_vis (target, t1, t1);
11507 bmask = 0x77777777;
11510 gcc_unreachable ();
11513 emit_insn (gen_bmasksi_vis (gen_reg_rtx (SImode), CONST0_RTX (SImode),
11514 force_reg (SImode, GEN_INT (bmask))));
11515 emit_insn (final_insn);
11518 /* Subroutine of sparc_expand_vector_init. Emit code to initialize
11519 all fields of TARGET to ELT in V8QI by means of VIS FPMERGE insn. */
11522 vector_init_fpmerge (rtx target, rtx elt)
11524 rtx t1, t2, t2_low, t3, t3_low;
11526 t1 = gen_reg_rtx (V4QImode);
11527 elt = convert_modes (SImode, QImode, elt, true);
11528 emit_move_insn (gen_lowpart (SImode, t1), elt);
11530 t2 = gen_reg_rtx (V8QImode);
11531 t2_low = gen_lowpart (V4QImode, t2);
11532 emit_insn (gen_fpmerge_vis (t2, t1, t1));
11534 t3 = gen_reg_rtx (V8QImode);
11535 t3_low = gen_lowpart (V4QImode, t3);
11536 emit_insn (gen_fpmerge_vis (t3, t2_low, t2_low));
11538 emit_insn (gen_fpmerge_vis (target, t3_low, t3_low));
11541 /* Subroutine of sparc_expand_vector_init. Emit code to initialize
11542 all fields of TARGET to ELT in V4HI by means of VIS FALIGNDATA insn. */
11545 vector_init_faligndata (rtx target, rtx elt)
11547 rtx t1 = gen_reg_rtx (V4HImode);
11550 elt = convert_modes (SImode, HImode, elt, true);
11551 emit_move_insn (gen_lowpart (SImode, t1), elt);
11553 emit_insn (gen_alignaddrsi_vis (gen_reg_rtx (SImode),
11554 force_reg (SImode, GEN_INT (6)),
11557 for (i = 0; i < 4; i++)
11558 emit_insn (gen_faligndatav4hi_vis (target, t1, target));
11561 /* Emit code to initialize TARGET to values for individual fields VALS. */
11564 sparc_expand_vector_init (rtx target, rtx vals)
11566 const enum machine_mode mode = GET_MODE (target);
11567 const enum machine_mode inner_mode = GET_MODE_INNER (mode);
11568 const int n_elts = GET_MODE_NUNITS (mode);
11574 for (i = 0; i < n_elts; i++)
11576 rtx x = XVECEXP (vals, 0, i);
11577 if (!CONSTANT_P (x))
11580 if (i > 0 && !rtx_equal_p (x, XVECEXP (vals, 0, 0)))
11586 emit_move_insn (target, gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0)));
11590 if (GET_MODE_SIZE (inner_mode) == GET_MODE_SIZE (mode))
11592 if (GET_MODE_SIZE (inner_mode) == 4)
11594 emit_move_insn (gen_lowpart (SImode, target),
11595 gen_lowpart (SImode, XVECEXP (vals, 0, 0)));
11598 else if (GET_MODE_SIZE (inner_mode) == 8)
11600 emit_move_insn (gen_lowpart (DImode, target),
11601 gen_lowpart (DImode, XVECEXP (vals, 0, 0)));
11605 else if (GET_MODE_SIZE (inner_mode) == GET_MODE_SIZE (word_mode)
11606 && GET_MODE_SIZE (mode) == 2 * GET_MODE_SIZE (word_mode))
11608 emit_move_insn (gen_highpart (word_mode, target),
11609 gen_lowpart (word_mode, XVECEXP (vals, 0, 0)));
11610 emit_move_insn (gen_lowpart (word_mode, target),
11611 gen_lowpart (word_mode, XVECEXP (vals, 0, 1)));
11615 if (all_same && GET_MODE_SIZE (mode) == 8)
11619 vector_init_bshuffle (target, XVECEXP (vals, 0, 0), mode, inner_mode);
11622 if (mode == V8QImode)
11624 vector_init_fpmerge (target, XVECEXP (vals, 0, 0));
11627 if (mode == V4HImode)
11629 vector_init_faligndata (target, XVECEXP (vals, 0, 0));
11634 mem = assign_stack_temp (mode, GET_MODE_SIZE (mode), 0);
11635 for (i = 0; i < n_elts; i++)
11636 emit_move_insn (adjust_address_nv (mem, inner_mode,
11637 i * GET_MODE_SIZE (inner_mode)),
11638 XVECEXP (vals, 0, i));
11639 emit_move_insn (target, mem);
11642 /* Implement TARGET_SECONDARY_RELOAD. */
11645 sparc_secondary_reload (bool in_p, rtx x, reg_class_t rclass_i,
11646 enum machine_mode mode, secondary_reload_info *sri)
11648 enum reg_class rclass = (enum reg_class) rclass_i;
11650 sri->icode = CODE_FOR_nothing;
11651 sri->extra_cost = 0;
11653 /* We need a temporary when loading/storing a HImode/QImode value
11654 between memory and the FPU registers. This can happen when combine puts
11655 a paradoxical subreg in a float/fix conversion insn. */
11656 if (FP_REG_CLASS_P (rclass)
11657 && (mode == HImode || mode == QImode)
11658 && (GET_CODE (x) == MEM
11659 || ((GET_CODE (x) == REG || GET_CODE (x) == SUBREG)
11660 && true_regnum (x) == -1)))
11661 return GENERAL_REGS;
11663 /* On 32-bit we need a temporary when loading/storing a DFmode value
11664 between unaligned memory and the upper FPU registers. */
11666 && rclass == EXTRA_FP_REGS
11668 && GET_CODE (x) == MEM
11669 && ! mem_min_alignment (x, 8))
11672 if (((TARGET_CM_MEDANY
11673 && symbolic_operand (x, mode))
11674 || (TARGET_CM_EMBMEDANY
11675 && text_segment_operand (x, mode)))
11679 sri->icode = direct_optab_handler (reload_in_optab, mode);
11681 sri->icode = direct_optab_handler (reload_out_optab, mode);
11685 if (TARGET_VIS3 && TARGET_ARCH32)
11687 int regno = true_regnum (x);
11689 /* When using VIS3 fp<-->int register moves, on 32-bit we have
11690 to move 8-byte values in 4-byte pieces. This only works via
11691 FP_REGS, and not via EXTRA_FP_REGS. Therefore if we try to
11692 move between EXTRA_FP_REGS and GENERAL_REGS, we will need
11693 an FP_REGS intermediate move. */
11694 if ((rclass == EXTRA_FP_REGS && SPARC_INT_REG_P (regno))
11695 || ((general_or_i64_p (rclass)
11696 || rclass == GENERAL_OR_FP_REGS)
11697 && SPARC_FP_REG_P (regno)))
11699 sri->extra_cost = 2;
11707 /* Emit code to conditionally move either OPERANDS[2] or OPERANDS[3] into
11708 OPERANDS[0] in MODE. OPERANDS[1] is the operator of the condition. */
11711 sparc_expand_conditional_move (enum machine_mode mode, rtx *operands)
11713 enum rtx_code rc = GET_CODE (operands[1]);
11714 enum machine_mode cmp_mode;
11715 rtx cc_reg, dst, cmp;
11718 if (GET_MODE (XEXP (cmp, 0)) == DImode && !TARGET_ARCH64)
11721 if (GET_MODE (XEXP (cmp, 0)) == TFmode && !TARGET_HARD_QUAD)
11722 cmp = sparc_emit_float_lib_cmp (XEXP (cmp, 0), XEXP (cmp, 1), rc);
11724 cmp_mode = GET_MODE (XEXP (cmp, 0));
11725 rc = GET_CODE (cmp);
11728 if (! rtx_equal_p (operands[2], dst)
11729 && ! rtx_equal_p (operands[3], dst))
11731 if (reg_overlap_mentioned_p (dst, cmp))
11732 dst = gen_reg_rtx (mode);
11734 emit_move_insn (dst, operands[3]);
11736 else if (operands[2] == dst)
11738 operands[2] = operands[3];
11740 if (GET_MODE_CLASS (cmp_mode) == MODE_FLOAT)
11741 rc = reverse_condition_maybe_unordered (rc);
11743 rc = reverse_condition (rc);
11746 if (XEXP (cmp, 1) == const0_rtx
11747 && GET_CODE (XEXP (cmp, 0)) == REG
11748 && cmp_mode == DImode
11749 && v9_regcmp_p (rc))
11750 cc_reg = XEXP (cmp, 0);
11752 cc_reg = gen_compare_reg_1 (rc, XEXP (cmp, 0), XEXP (cmp, 1));
11754 cmp = gen_rtx_fmt_ee (rc, GET_MODE (cc_reg), cc_reg, const0_rtx);
11756 emit_insn (gen_rtx_SET (VOIDmode, dst,
11757 gen_rtx_IF_THEN_ELSE (mode, cmp, operands[2], dst)));
11759 if (dst != operands[0])
11760 emit_move_insn (operands[0], dst);
11765 /* Emit code to conditionally move a combination of OPERANDS[1] and OPERANDS[2]
11766 into OPERANDS[0] in MODE, depending on the outcome of the comparison of
11767 OPERANDS[4] and OPERANDS[5]. OPERANDS[3] is the operator of the condition.
11768 FCODE is the machine code to be used for OPERANDS[3] and CCODE the machine
11769 code to be used for the condition mask. */
11772 sparc_expand_vcond (enum machine_mode mode, rtx *operands, int ccode, int fcode)
11774 rtx mask, cop0, cop1, fcmp, cmask, bshuf, gsr;
11775 enum rtx_code code = GET_CODE (operands[3]);
11777 mask = gen_reg_rtx (Pmode);
11778 cop0 = operands[4];
11779 cop1 = operands[5];
11780 if (code == LT || code == GE)
11784 code = swap_condition (code);
11785 t = cop0; cop0 = cop1; cop1 = t;
11788 gsr = gen_rtx_REG (DImode, SPARC_GSR_REG);
11790 fcmp = gen_rtx_UNSPEC (Pmode,
11791 gen_rtvec (1, gen_rtx_fmt_ee (code, mode, cop0, cop1)),
11794 cmask = gen_rtx_UNSPEC (DImode,
11795 gen_rtvec (2, mask, gsr),
11798 bshuf = gen_rtx_UNSPEC (mode,
11799 gen_rtvec (3, operands[1], operands[2], gsr),
11802 emit_insn (gen_rtx_SET (VOIDmode, mask, fcmp));
11803 emit_insn (gen_rtx_SET (VOIDmode, gsr, cmask));
11805 emit_insn (gen_rtx_SET (VOIDmode, operands[0], bshuf));
11808 /* On sparc, any mode which naturally allocates into the float
11809 registers should return 4 here. */
11812 sparc_regmode_natural_size (enum machine_mode mode)
11814 int size = UNITS_PER_WORD;
11818 enum mode_class mclass = GET_MODE_CLASS (mode);
11820 if (mclass == MODE_FLOAT || mclass == MODE_VECTOR_INT)
11827 /* Return TRUE if it is a good idea to tie two pseudo registers
11828 when one has mode MODE1 and one has mode MODE2.
11829 If HARD_REGNO_MODE_OK could produce different values for MODE1 and MODE2,
11830 for any hard reg, then this must be FALSE for correct output.
11832 For V9 we have to deal with the fact that only the lower 32 floating
11833 point registers are 32-bit addressable. */
11836 sparc_modes_tieable_p (enum machine_mode mode1, enum machine_mode mode2)
11838 enum mode_class mclass1, mclass2;
11839 unsigned short size1, size2;
11841 if (mode1 == mode2)
11844 mclass1 = GET_MODE_CLASS (mode1);
11845 mclass2 = GET_MODE_CLASS (mode2);
11846 if (mclass1 != mclass2)
11852 /* Classes are the same and we are V9 so we have to deal with upper
11853 vs. lower floating point registers. If one of the modes is a
11854 4-byte mode, and the other is not, we have to mark them as not
11855 tieable because only the lower 32 floating point register are
11856 addressable 32-bits at a time.
11858 We can't just test explicitly for SFmode, otherwise we won't
11859 cover the vector mode cases properly. */
11861 if (mclass1 != MODE_FLOAT && mclass1 != MODE_VECTOR_INT)
11864 size1 = GET_MODE_SIZE (mode1);
11865 size2 = GET_MODE_SIZE (mode2);
11866 if ((size1 > 4 && size2 == 4)
11867 || (size2 > 4 && size1 == 4))
11873 #include "gt-sparc.h"