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 */
377 static const struct processor_costs *sparc_costs = &cypress_costs;
379 #ifdef HAVE_AS_RELAX_OPTION
380 /* If 'as' and 'ld' are relaxing tail call insns into branch always, use
381 "or %o7,%g0,X; call Y; or X,%g0,%o7" always, so that it can be optimized.
382 With sethi/jmp, neither 'as' nor 'ld' has an easy way how to find out if
383 somebody does not branch between the sethi and jmp. */
384 #define LEAF_SIBCALL_SLOT_RESERVED_P 1
386 #define LEAF_SIBCALL_SLOT_RESERVED_P \
387 ((TARGET_ARCH64 && !TARGET_CM_MEDLOW) || flag_pic)
390 /* Vector to say how input registers are mapped to output registers.
391 HARD_FRAME_POINTER_REGNUM cannot be remapped by this function to
392 eliminate it. You must use -fomit-frame-pointer to get that. */
393 char leaf_reg_remap[] =
394 { 0, 1, 2, 3, 4, 5, 6, 7,
395 -1, -1, -1, -1, -1, -1, 14, -1,
396 -1, -1, -1, -1, -1, -1, -1, -1,
397 8, 9, 10, 11, 12, 13, -1, 15,
399 32, 33, 34, 35, 36, 37, 38, 39,
400 40, 41, 42, 43, 44, 45, 46, 47,
401 48, 49, 50, 51, 52, 53, 54, 55,
402 56, 57, 58, 59, 60, 61, 62, 63,
403 64, 65, 66, 67, 68, 69, 70, 71,
404 72, 73, 74, 75, 76, 77, 78, 79,
405 80, 81, 82, 83, 84, 85, 86, 87,
406 88, 89, 90, 91, 92, 93, 94, 95,
407 96, 97, 98, 99, 100, 101, 102};
409 /* Vector, indexed by hard register number, which contains 1
410 for a register that is allowable in a candidate for leaf
411 function treatment. */
412 char sparc_leaf_regs[] =
413 { 1, 1, 1, 1, 1, 1, 1, 1,
414 0, 0, 0, 0, 0, 0, 1, 0,
415 0, 0, 0, 0, 0, 0, 0, 0,
416 1, 1, 1, 1, 1, 1, 0, 1,
417 1, 1, 1, 1, 1, 1, 1, 1,
418 1, 1, 1, 1, 1, 1, 1, 1,
419 1, 1, 1, 1, 1, 1, 1, 1,
420 1, 1, 1, 1, 1, 1, 1, 1,
421 1, 1, 1, 1, 1, 1, 1, 1,
422 1, 1, 1, 1, 1, 1, 1, 1,
423 1, 1, 1, 1, 1, 1, 1, 1,
424 1, 1, 1, 1, 1, 1, 1, 1,
425 1, 1, 1, 1, 1, 1, 1};
427 struct GTY(()) machine_function
429 /* Size of the frame of the function. */
430 HOST_WIDE_INT frame_size;
432 /* Size of the frame of the function minus the register window save area
433 and the outgoing argument area. */
434 HOST_WIDE_INT apparent_frame_size;
436 /* Register we pretend the frame pointer is allocated to. Normally, this
437 is %fp, but if we are in a leaf procedure, this is (%sp + offset). We
438 record "offset" separately as it may be too big for (reg + disp). */
440 HOST_WIDE_INT frame_base_offset;
442 /* Some local-dynamic TLS symbol name. */
443 const char *some_ld_name;
445 /* Number of global or FP registers to be saved (as 4-byte quantities). */
446 int n_global_fp_regs;
448 /* True if the current function is leaf and uses only leaf regs,
449 so that the SPARC leaf function optimization can be applied.
450 Private version of current_function_uses_only_leaf_regs, see
451 sparc_expand_prologue for the rationale. */
454 /* True if the prologue saves local or in registers. */
455 bool save_local_in_regs_p;
457 /* True if the data calculated by sparc_expand_prologue are valid. */
458 bool prologue_data_valid_p;
461 #define sparc_frame_size cfun->machine->frame_size
462 #define sparc_apparent_frame_size cfun->machine->apparent_frame_size
463 #define sparc_frame_base_reg cfun->machine->frame_base_reg
464 #define sparc_frame_base_offset cfun->machine->frame_base_offset
465 #define sparc_n_global_fp_regs cfun->machine->n_global_fp_regs
466 #define sparc_leaf_function_p cfun->machine->leaf_function_p
467 #define sparc_save_local_in_regs_p cfun->machine->save_local_in_regs_p
468 #define sparc_prologue_data_valid_p cfun->machine->prologue_data_valid_p
470 /* 1 if the next opcode is to be specially indented. */
471 int sparc_indent_opcode = 0;
473 static void sparc_option_override (void);
474 static void sparc_init_modes (void);
475 static void scan_record_type (const_tree, int *, int *, int *);
476 static int function_arg_slotno (const CUMULATIVE_ARGS *, enum machine_mode,
477 const_tree, bool, bool, int *, int *);
479 static int supersparc_adjust_cost (rtx, rtx, rtx, int);
480 static int hypersparc_adjust_cost (rtx, rtx, rtx, int);
482 static void sparc_emit_set_const32 (rtx, rtx);
483 static void sparc_emit_set_const64 (rtx, rtx);
484 static void sparc_output_addr_vec (rtx);
485 static void sparc_output_addr_diff_vec (rtx);
486 static void sparc_output_deferred_case_vectors (void);
487 static bool sparc_legitimate_address_p (enum machine_mode, rtx, bool);
488 static bool sparc_legitimate_constant_p (enum machine_mode, rtx);
489 static rtx sparc_builtin_saveregs (void);
490 static int epilogue_renumber (rtx *, int);
491 static bool sparc_assemble_integer (rtx, unsigned int, int);
492 static int set_extends (rtx);
493 static void sparc_asm_function_prologue (FILE *, HOST_WIDE_INT);
494 static void sparc_asm_function_epilogue (FILE *, HOST_WIDE_INT);
495 #ifdef TARGET_SOLARIS
496 static void sparc_solaris_elf_asm_named_section (const char *, unsigned int,
497 tree) ATTRIBUTE_UNUSED;
499 static int sparc_adjust_cost (rtx, rtx, rtx, int);
500 static int sparc_issue_rate (void);
501 static void sparc_sched_init (FILE *, int, int);
502 static int sparc_use_sched_lookahead (void);
504 static void emit_soft_tfmode_libcall (const char *, int, rtx *);
505 static void emit_soft_tfmode_binop (enum rtx_code, rtx *);
506 static void emit_soft_tfmode_unop (enum rtx_code, rtx *);
507 static void emit_soft_tfmode_cvt (enum rtx_code, rtx *);
508 static void emit_hard_tfmode_operation (enum rtx_code, rtx *);
510 static bool sparc_function_ok_for_sibcall (tree, tree);
511 static void sparc_init_libfuncs (void);
512 static void sparc_init_builtins (void);
513 static void sparc_vis_init_builtins (void);
514 static rtx sparc_expand_builtin (tree, rtx, rtx, enum machine_mode, int);
515 static tree sparc_fold_builtin (tree, int, tree *, bool);
516 static int sparc_vis_mul8x16 (int, int);
517 static tree sparc_handle_vis_mul8x16 (int, tree, tree, tree);
518 static void sparc_output_mi_thunk (FILE *, tree, HOST_WIDE_INT,
519 HOST_WIDE_INT, tree);
520 static bool sparc_can_output_mi_thunk (const_tree, HOST_WIDE_INT,
521 HOST_WIDE_INT, const_tree);
522 static void sparc_reorg (void);
523 static struct machine_function * sparc_init_machine_status (void);
524 static bool sparc_cannot_force_const_mem (enum machine_mode, rtx);
525 static rtx sparc_tls_get_addr (void);
526 static rtx sparc_tls_got (void);
527 static const char *get_some_local_dynamic_name (void);
528 static int get_some_local_dynamic_name_1 (rtx *, void *);
529 static int sparc_register_move_cost (enum machine_mode,
530 reg_class_t, reg_class_t);
531 static bool sparc_rtx_costs (rtx, int, int, int, int *, bool);
532 static rtx sparc_function_value (const_tree, const_tree, bool);
533 static rtx sparc_libcall_value (enum machine_mode, const_rtx);
534 static bool sparc_function_value_regno_p (const unsigned int);
535 static rtx sparc_struct_value_rtx (tree, int);
536 static enum machine_mode sparc_promote_function_mode (const_tree, enum machine_mode,
537 int *, const_tree, int);
538 static bool sparc_return_in_memory (const_tree, const_tree);
539 static bool sparc_strict_argument_naming (cumulative_args_t);
540 static void sparc_va_start (tree, rtx);
541 static tree sparc_gimplify_va_arg (tree, tree, gimple_seq *, gimple_seq *);
542 static bool sparc_vector_mode_supported_p (enum machine_mode);
543 static bool sparc_tls_referenced_p (rtx);
544 static rtx sparc_legitimize_tls_address (rtx);
545 static rtx sparc_legitimize_pic_address (rtx, rtx);
546 static rtx sparc_legitimize_address (rtx, rtx, enum machine_mode);
547 static rtx sparc_delegitimize_address (rtx);
548 static bool sparc_mode_dependent_address_p (const_rtx);
549 static bool sparc_pass_by_reference (cumulative_args_t,
550 enum machine_mode, const_tree, bool);
551 static void sparc_function_arg_advance (cumulative_args_t,
552 enum machine_mode, const_tree, bool);
553 static rtx sparc_function_arg_1 (cumulative_args_t,
554 enum machine_mode, const_tree, bool, bool);
555 static rtx sparc_function_arg (cumulative_args_t,
556 enum machine_mode, const_tree, bool);
557 static rtx sparc_function_incoming_arg (cumulative_args_t,
558 enum machine_mode, const_tree, bool);
559 static unsigned int sparc_function_arg_boundary (enum machine_mode,
561 static int sparc_arg_partial_bytes (cumulative_args_t,
562 enum machine_mode, tree, bool);
563 static void sparc_output_dwarf_dtprel (FILE *, int, rtx) ATTRIBUTE_UNUSED;
564 static void sparc_file_end (void);
565 static bool sparc_frame_pointer_required (void);
566 static bool sparc_can_eliminate (const int, const int);
567 static rtx sparc_builtin_setjmp_frame_value (void);
568 static void sparc_conditional_register_usage (void);
569 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
570 static const char *sparc_mangle_type (const_tree);
572 static void sparc_trampoline_init (rtx, tree, rtx);
573 static enum machine_mode sparc_preferred_simd_mode (enum machine_mode);
574 static reg_class_t sparc_preferred_reload_class (rtx x, reg_class_t rclass);
575 static bool sparc_print_operand_punct_valid_p (unsigned char);
576 static void sparc_print_operand (FILE *, rtx, int);
577 static void sparc_print_operand_address (FILE *, rtx);
578 static reg_class_t sparc_secondary_reload (bool, rtx, reg_class_t,
580 secondary_reload_info *);
582 #ifdef SUBTARGET_ATTRIBUTE_TABLE
583 /* Table of valid machine attributes. */
584 static const struct attribute_spec sparc_attribute_table[] =
586 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler,
588 SUBTARGET_ATTRIBUTE_TABLE,
589 { NULL, 0, 0, false, false, false, NULL, false }
593 /* Option handling. */
596 enum cmodel sparc_cmodel;
598 char sparc_hard_reg_printed[8];
600 /* Initialize the GCC target structure. */
602 /* The default is to use .half rather than .short for aligned HI objects. */
603 #undef TARGET_ASM_ALIGNED_HI_OP
604 #define TARGET_ASM_ALIGNED_HI_OP "\t.half\t"
606 #undef TARGET_ASM_UNALIGNED_HI_OP
607 #define TARGET_ASM_UNALIGNED_HI_OP "\t.uahalf\t"
608 #undef TARGET_ASM_UNALIGNED_SI_OP
609 #define TARGET_ASM_UNALIGNED_SI_OP "\t.uaword\t"
610 #undef TARGET_ASM_UNALIGNED_DI_OP
611 #define TARGET_ASM_UNALIGNED_DI_OP "\t.uaxword\t"
613 /* The target hook has to handle DI-mode values. */
614 #undef TARGET_ASM_INTEGER
615 #define TARGET_ASM_INTEGER sparc_assemble_integer
617 #undef TARGET_ASM_FUNCTION_PROLOGUE
618 #define TARGET_ASM_FUNCTION_PROLOGUE sparc_asm_function_prologue
619 #undef TARGET_ASM_FUNCTION_EPILOGUE
620 #define TARGET_ASM_FUNCTION_EPILOGUE sparc_asm_function_epilogue
622 #undef TARGET_SCHED_ADJUST_COST
623 #define TARGET_SCHED_ADJUST_COST sparc_adjust_cost
624 #undef TARGET_SCHED_ISSUE_RATE
625 #define TARGET_SCHED_ISSUE_RATE sparc_issue_rate
626 #undef TARGET_SCHED_INIT
627 #define TARGET_SCHED_INIT sparc_sched_init
628 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
629 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD sparc_use_sched_lookahead
631 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
632 #define TARGET_FUNCTION_OK_FOR_SIBCALL sparc_function_ok_for_sibcall
634 #undef TARGET_INIT_LIBFUNCS
635 #define TARGET_INIT_LIBFUNCS sparc_init_libfuncs
636 #undef TARGET_INIT_BUILTINS
637 #define TARGET_INIT_BUILTINS sparc_init_builtins
639 #undef TARGET_LEGITIMIZE_ADDRESS
640 #define TARGET_LEGITIMIZE_ADDRESS sparc_legitimize_address
641 #undef TARGET_DELEGITIMIZE_ADDRESS
642 #define TARGET_DELEGITIMIZE_ADDRESS sparc_delegitimize_address
643 #undef TARGET_MODE_DEPENDENT_ADDRESS_P
644 #define TARGET_MODE_DEPENDENT_ADDRESS_P sparc_mode_dependent_address_p
646 #undef TARGET_EXPAND_BUILTIN
647 #define TARGET_EXPAND_BUILTIN sparc_expand_builtin
648 #undef TARGET_FOLD_BUILTIN
649 #define TARGET_FOLD_BUILTIN sparc_fold_builtin
652 #undef TARGET_HAVE_TLS
653 #define TARGET_HAVE_TLS true
656 #undef TARGET_CANNOT_FORCE_CONST_MEM
657 #define TARGET_CANNOT_FORCE_CONST_MEM sparc_cannot_force_const_mem
659 #undef TARGET_ASM_OUTPUT_MI_THUNK
660 #define TARGET_ASM_OUTPUT_MI_THUNK sparc_output_mi_thunk
661 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
662 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK sparc_can_output_mi_thunk
664 #undef TARGET_MACHINE_DEPENDENT_REORG
665 #define TARGET_MACHINE_DEPENDENT_REORG sparc_reorg
667 #undef TARGET_RTX_COSTS
668 #define TARGET_RTX_COSTS sparc_rtx_costs
669 #undef TARGET_ADDRESS_COST
670 #define TARGET_ADDRESS_COST hook_int_rtx_bool_0
671 #undef TARGET_REGISTER_MOVE_COST
672 #define TARGET_REGISTER_MOVE_COST sparc_register_move_cost
674 #undef TARGET_PROMOTE_FUNCTION_MODE
675 #define TARGET_PROMOTE_FUNCTION_MODE sparc_promote_function_mode
677 #undef TARGET_FUNCTION_VALUE
678 #define TARGET_FUNCTION_VALUE sparc_function_value
679 #undef TARGET_LIBCALL_VALUE
680 #define TARGET_LIBCALL_VALUE sparc_libcall_value
681 #undef TARGET_FUNCTION_VALUE_REGNO_P
682 #define TARGET_FUNCTION_VALUE_REGNO_P sparc_function_value_regno_p
684 #undef TARGET_STRUCT_VALUE_RTX
685 #define TARGET_STRUCT_VALUE_RTX sparc_struct_value_rtx
686 #undef TARGET_RETURN_IN_MEMORY
687 #define TARGET_RETURN_IN_MEMORY sparc_return_in_memory
688 #undef TARGET_MUST_PASS_IN_STACK
689 #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size
690 #undef TARGET_PASS_BY_REFERENCE
691 #define TARGET_PASS_BY_REFERENCE sparc_pass_by_reference
692 #undef TARGET_ARG_PARTIAL_BYTES
693 #define TARGET_ARG_PARTIAL_BYTES sparc_arg_partial_bytes
694 #undef TARGET_FUNCTION_ARG_ADVANCE
695 #define TARGET_FUNCTION_ARG_ADVANCE sparc_function_arg_advance
696 #undef TARGET_FUNCTION_ARG
697 #define TARGET_FUNCTION_ARG sparc_function_arg
698 #undef TARGET_FUNCTION_INCOMING_ARG
699 #define TARGET_FUNCTION_INCOMING_ARG sparc_function_incoming_arg
700 #undef TARGET_FUNCTION_ARG_BOUNDARY
701 #define TARGET_FUNCTION_ARG_BOUNDARY sparc_function_arg_boundary
703 #undef TARGET_EXPAND_BUILTIN_SAVEREGS
704 #define TARGET_EXPAND_BUILTIN_SAVEREGS sparc_builtin_saveregs
705 #undef TARGET_STRICT_ARGUMENT_NAMING
706 #define TARGET_STRICT_ARGUMENT_NAMING sparc_strict_argument_naming
708 #undef TARGET_EXPAND_BUILTIN_VA_START
709 #define TARGET_EXPAND_BUILTIN_VA_START sparc_va_start
710 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
711 #define TARGET_GIMPLIFY_VA_ARG_EXPR sparc_gimplify_va_arg
713 #undef TARGET_VECTOR_MODE_SUPPORTED_P
714 #define TARGET_VECTOR_MODE_SUPPORTED_P sparc_vector_mode_supported_p
716 #undef TARGET_VECTORIZE_PREFERRED_SIMD_MODE
717 #define TARGET_VECTORIZE_PREFERRED_SIMD_MODE sparc_preferred_simd_mode
719 #ifdef SUBTARGET_INSERT_ATTRIBUTES
720 #undef TARGET_INSERT_ATTRIBUTES
721 #define TARGET_INSERT_ATTRIBUTES SUBTARGET_INSERT_ATTRIBUTES
724 #ifdef SUBTARGET_ATTRIBUTE_TABLE
725 #undef TARGET_ATTRIBUTE_TABLE
726 #define TARGET_ATTRIBUTE_TABLE sparc_attribute_table
729 #undef TARGET_RELAXED_ORDERING
730 #define TARGET_RELAXED_ORDERING SPARC_RELAXED_ORDERING
732 #undef TARGET_OPTION_OVERRIDE
733 #define TARGET_OPTION_OVERRIDE sparc_option_override
735 #if TARGET_GNU_TLS && defined(HAVE_AS_SPARC_UA_PCREL)
736 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
737 #define TARGET_ASM_OUTPUT_DWARF_DTPREL sparc_output_dwarf_dtprel
740 #undef TARGET_ASM_FILE_END
741 #define TARGET_ASM_FILE_END sparc_file_end
743 #undef TARGET_FRAME_POINTER_REQUIRED
744 #define TARGET_FRAME_POINTER_REQUIRED sparc_frame_pointer_required
746 #undef TARGET_BUILTIN_SETJMP_FRAME_VALUE
747 #define TARGET_BUILTIN_SETJMP_FRAME_VALUE sparc_builtin_setjmp_frame_value
749 #undef TARGET_CAN_ELIMINATE
750 #define TARGET_CAN_ELIMINATE sparc_can_eliminate
752 #undef TARGET_PREFERRED_RELOAD_CLASS
753 #define TARGET_PREFERRED_RELOAD_CLASS sparc_preferred_reload_class
755 #undef TARGET_SECONDARY_RELOAD
756 #define TARGET_SECONDARY_RELOAD sparc_secondary_reload
758 #undef TARGET_CONDITIONAL_REGISTER_USAGE
759 #define TARGET_CONDITIONAL_REGISTER_USAGE sparc_conditional_register_usage
761 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
762 #undef TARGET_MANGLE_TYPE
763 #define TARGET_MANGLE_TYPE sparc_mangle_type
766 #undef TARGET_LEGITIMATE_ADDRESS_P
767 #define TARGET_LEGITIMATE_ADDRESS_P sparc_legitimate_address_p
769 #undef TARGET_LEGITIMATE_CONSTANT_P
770 #define TARGET_LEGITIMATE_CONSTANT_P sparc_legitimate_constant_p
772 #undef TARGET_TRAMPOLINE_INIT
773 #define TARGET_TRAMPOLINE_INIT sparc_trampoline_init
775 #undef TARGET_PRINT_OPERAND_PUNCT_VALID_P
776 #define TARGET_PRINT_OPERAND_PUNCT_VALID_P sparc_print_operand_punct_valid_p
777 #undef TARGET_PRINT_OPERAND
778 #define TARGET_PRINT_OPERAND sparc_print_operand
779 #undef TARGET_PRINT_OPERAND_ADDRESS
780 #define TARGET_PRINT_OPERAND_ADDRESS sparc_print_operand_address
782 /* The value stored by LDSTUB. */
783 #undef TARGET_ATOMIC_TEST_AND_SET_TRUEVAL
784 #define TARGET_ATOMIC_TEST_AND_SET_TRUEVAL 0xff
786 struct gcc_target targetm = TARGET_INITIALIZER;
789 dump_target_flag_bits (const int flags)
791 if (flags & MASK_64BIT)
792 fprintf (stderr, "64BIT ");
793 if (flags & MASK_APP_REGS)
794 fprintf (stderr, "APP_REGS ");
795 if (flags & MASK_FASTER_STRUCTS)
796 fprintf (stderr, "FASTER_STRUCTS ");
797 if (flags & MASK_FLAT)
798 fprintf (stderr, "FLAT ");
799 if (flags & MASK_FMAF)
800 fprintf (stderr, "FMAF ");
801 if (flags & MASK_FPU)
802 fprintf (stderr, "FPU ");
803 if (flags & MASK_HARD_QUAD)
804 fprintf (stderr, "HARD_QUAD ");
805 if (flags & MASK_POPC)
806 fprintf (stderr, "POPC ");
807 if (flags & MASK_PTR64)
808 fprintf (stderr, "PTR64 ");
809 if (flags & MASK_STACK_BIAS)
810 fprintf (stderr, "STACK_BIAS ");
811 if (flags & MASK_UNALIGNED_DOUBLES)
812 fprintf (stderr, "UNALIGNED_DOUBLES ");
813 if (flags & MASK_V8PLUS)
814 fprintf (stderr, "V8PLUS ");
815 if (flags & MASK_VIS)
816 fprintf (stderr, "VIS ");
817 if (flags & MASK_VIS2)
818 fprintf (stderr, "VIS2 ");
819 if (flags & MASK_VIS3)
820 fprintf (stderr, "VIS3 ");
821 if (flags & MASK_DEPRECATED_V8_INSNS)
822 fprintf (stderr, "DEPRECATED_V8_INSNS ");
823 if (flags & MASK_SPARCLET)
824 fprintf (stderr, "SPARCLET ");
825 if (flags & MASK_SPARCLITE)
826 fprintf (stderr, "SPARCLITE ");
828 fprintf (stderr, "V8 ");
830 fprintf (stderr, "V9 ");
834 dump_target_flags (const char *prefix, const int flags)
836 fprintf (stderr, "%s: (%08x) [ ", prefix, flags);
837 dump_target_flag_bits (flags);
838 fprintf(stderr, "]\n");
841 /* Validate and override various options, and do some machine dependent
845 sparc_option_override (void)
847 static struct code_model {
848 const char *const name;
849 const enum cmodel value;
850 } const cmodels[] = {
852 { "medlow", CM_MEDLOW },
853 { "medmid", CM_MEDMID },
854 { "medany", CM_MEDANY },
855 { "embmedany", CM_EMBMEDANY },
856 { NULL, (enum cmodel) 0 }
858 const struct code_model *cmodel;
859 /* Map TARGET_CPU_DEFAULT to value for -m{cpu,tune}=. */
860 static struct cpu_default {
862 const enum processor_type processor;
863 } const cpu_default[] = {
864 /* There must be one entry here for each TARGET_CPU value. */
865 { TARGET_CPU_sparc, PROCESSOR_CYPRESS },
866 { TARGET_CPU_v8, PROCESSOR_V8 },
867 { TARGET_CPU_supersparc, PROCESSOR_SUPERSPARC },
868 { TARGET_CPU_hypersparc, PROCESSOR_HYPERSPARC },
869 { TARGET_CPU_leon, PROCESSOR_LEON },
870 { TARGET_CPU_sparclite, PROCESSOR_F930 },
871 { TARGET_CPU_sparclite86x, PROCESSOR_SPARCLITE86X },
872 { TARGET_CPU_sparclet, PROCESSOR_TSC701 },
873 { TARGET_CPU_v9, PROCESSOR_V9 },
874 { TARGET_CPU_ultrasparc, PROCESSOR_ULTRASPARC },
875 { TARGET_CPU_ultrasparc3, PROCESSOR_ULTRASPARC3 },
876 { TARGET_CPU_niagara, PROCESSOR_NIAGARA },
877 { TARGET_CPU_niagara2, PROCESSOR_NIAGARA2 },
878 { TARGET_CPU_niagara3, PROCESSOR_NIAGARA3 },
879 { TARGET_CPU_niagara4, PROCESSOR_NIAGARA4 },
882 const struct cpu_default *def;
883 /* Table of values for -m{cpu,tune}=. This must match the order of
884 the PROCESSOR_* enumeration. */
885 static struct cpu_table {
886 const char *const name;
889 } const cpu_table[] = {
890 { "v7", MASK_ISA, 0 },
891 { "cypress", MASK_ISA, 0 },
892 { "v8", MASK_ISA, MASK_V8 },
893 /* TI TMS390Z55 supersparc */
894 { "supersparc", MASK_ISA, MASK_V8 },
895 { "hypersparc", MASK_ISA, MASK_V8|MASK_FPU },
897 { "leon", MASK_ISA, MASK_V8|MASK_FPU },
898 { "sparclite", MASK_ISA, MASK_SPARCLITE },
899 /* The Fujitsu MB86930 is the original sparclite chip, with no FPU. */
900 { "f930", MASK_ISA|MASK_FPU, MASK_SPARCLITE },
901 /* The Fujitsu MB86934 is the recent sparclite chip, with an FPU. */
902 { "f934", MASK_ISA, MASK_SPARCLITE|MASK_FPU },
903 { "sparclite86x", MASK_ISA|MASK_FPU, MASK_SPARCLITE },
904 { "sparclet", MASK_ISA, MASK_SPARCLET },
906 { "tsc701", MASK_ISA, MASK_SPARCLET },
907 { "v9", MASK_ISA, MASK_V9 },
908 /* UltraSPARC I, II, IIi */
909 { "ultrasparc", MASK_ISA,
910 /* Although insns using %y are deprecated, it is a clear win. */
911 MASK_V9|MASK_DEPRECATED_V8_INSNS },
913 /* ??? Check if %y issue still holds true. */
914 { "ultrasparc3", MASK_ISA,
915 MASK_V9|MASK_DEPRECATED_V8_INSNS|MASK_VIS2 },
917 { "niagara", MASK_ISA,
918 MASK_V9|MASK_DEPRECATED_V8_INSNS },
920 { "niagara2", MASK_ISA,
921 MASK_V9|MASK_POPC|MASK_VIS2 },
923 { "niagara3", MASK_ISA,
924 MASK_V9|MASK_POPC|MASK_VIS2|MASK_VIS3|MASK_FMAF },
926 { "niagara4", MASK_ISA,
927 MASK_V9|MASK_POPC|MASK_VIS2|MASK_VIS3|MASK_FMAF },
929 const struct cpu_table *cpu;
933 if (sparc_debug_string != NULL)
938 p = ASTRDUP (sparc_debug_string);
939 while ((q = strtok (p, ",")) != NULL)
953 if (! strcmp (q, "all"))
954 mask = MASK_DEBUG_ALL;
955 else if (! strcmp (q, "options"))
956 mask = MASK_DEBUG_OPTIONS;
958 error ("unknown -mdebug-%s switch", q);
961 sparc_debug &= ~mask;
967 if (TARGET_DEBUG_OPTIONS)
969 dump_target_flags("Initial target_flags", target_flags);
970 dump_target_flags("target_flags_explicit", target_flags_explicit);
973 #ifdef SUBTARGET_OVERRIDE_OPTIONS
974 SUBTARGET_OVERRIDE_OPTIONS;
977 #ifndef SPARC_BI_ARCH
978 /* Check for unsupported architecture size. */
979 if (! TARGET_64BIT != DEFAULT_ARCH32_P)
980 error ("%s is not supported by this configuration",
981 DEFAULT_ARCH32_P ? "-m64" : "-m32");
984 /* We force all 64bit archs to use 128 bit long double */
985 if (TARGET_64BIT && ! TARGET_LONG_DOUBLE_128)
987 error ("-mlong-double-64 not allowed with -m64");
988 target_flags |= MASK_LONG_DOUBLE_128;
991 /* Code model selection. */
992 sparc_cmodel = SPARC_DEFAULT_CMODEL;
996 sparc_cmodel = CM_32;
999 if (sparc_cmodel_string != NULL)
1003 for (cmodel = &cmodels[0]; cmodel->name; cmodel++)
1004 if (strcmp (sparc_cmodel_string, cmodel->name) == 0)
1006 if (cmodel->name == NULL)
1007 error ("bad value (%s) for -mcmodel= switch", sparc_cmodel_string);
1009 sparc_cmodel = cmodel->value;
1012 error ("-mcmodel= is not supported on 32 bit systems");
1015 /* Check that -fcall-saved-REG wasn't specified for out registers. */
1016 for (i = 8; i < 16; i++)
1017 if (!call_used_regs [i])
1019 error ("-fcall-saved-REG is not supported for out registers");
1020 call_used_regs [i] = 1;
1023 fpu = target_flags & MASK_FPU; /* save current -mfpu status */
1025 /* Set the default CPU. */
1026 if (!global_options_set.x_sparc_cpu_and_features)
1028 for (def = &cpu_default[0]; def->cpu != -1; ++def)
1029 if (def->cpu == TARGET_CPU_DEFAULT)
1031 gcc_assert (def->cpu != -1);
1032 sparc_cpu_and_features = def->processor;
1035 if (!global_options_set.x_sparc_cpu)
1036 sparc_cpu = sparc_cpu_and_features;
1038 cpu = &cpu_table[(int) sparc_cpu_and_features];
1040 if (TARGET_DEBUG_OPTIONS)
1042 fprintf (stderr, "sparc_cpu_and_features: %s\n", cpu->name);
1043 fprintf (stderr, "sparc_cpu: %s\n",
1044 cpu_table[(int) sparc_cpu].name);
1045 dump_target_flags ("cpu->disable", cpu->disable);
1046 dump_target_flags ("cpu->enable", cpu->enable);
1049 target_flags &= ~cpu->disable;
1050 target_flags |= (cpu->enable
1051 #ifndef HAVE_AS_FMAF_HPC_VIS3
1052 & ~(MASK_FMAF | MASK_VIS3)
1056 /* If -mfpu or -mno-fpu was explicitly used, don't override with
1057 the processor default. */
1058 if (target_flags_explicit & MASK_FPU)
1059 target_flags = (target_flags & ~MASK_FPU) | fpu;
1061 /* -mvis2 implies -mvis */
1063 target_flags |= MASK_VIS;
1065 /* -mvis3 implies -mvis2 and -mvis */
1067 target_flags |= MASK_VIS2 | MASK_VIS;
1069 /* Don't allow -mvis, -mvis2, -mvis3, or -mfmaf if FPU is disabled. */
1071 target_flags &= ~(MASK_VIS | MASK_VIS2 | MASK_VIS3 | MASK_FMAF);
1073 /* -mvis assumes UltraSPARC+, so we are sure v9 instructions
1075 -m64 also implies v9. */
1076 if (TARGET_VIS || TARGET_ARCH64)
1078 target_flags |= MASK_V9;
1079 target_flags &= ~(MASK_V8 | MASK_SPARCLET | MASK_SPARCLITE);
1082 /* -mvis also implies -mv8plus on 32-bit */
1083 if (TARGET_VIS && ! TARGET_ARCH64)
1084 target_flags |= MASK_V8PLUS;
1086 /* Use the deprecated v8 insns for sparc64 in 32 bit mode. */
1087 if (TARGET_V9 && TARGET_ARCH32)
1088 target_flags |= MASK_DEPRECATED_V8_INSNS;
1090 /* V8PLUS requires V9, makes no sense in 64 bit mode. */
1091 if (! TARGET_V9 || TARGET_ARCH64)
1092 target_flags &= ~MASK_V8PLUS;
1094 /* Don't use stack biasing in 32 bit mode. */
1096 target_flags &= ~MASK_STACK_BIAS;
1098 /* Supply a default value for align_functions. */
1099 if (align_functions == 0
1100 && (sparc_cpu == PROCESSOR_ULTRASPARC
1101 || sparc_cpu == PROCESSOR_ULTRASPARC3
1102 || sparc_cpu == PROCESSOR_NIAGARA
1103 || sparc_cpu == PROCESSOR_NIAGARA2
1104 || sparc_cpu == PROCESSOR_NIAGARA3
1105 || sparc_cpu == PROCESSOR_NIAGARA4))
1106 align_functions = 32;
1108 /* Validate PCC_STRUCT_RETURN. */
1109 if (flag_pcc_struct_return == DEFAULT_PCC_STRUCT_RETURN)
1110 flag_pcc_struct_return = (TARGET_ARCH64 ? 0 : 1);
1112 /* Only use .uaxword when compiling for a 64-bit target. */
1114 targetm.asm_out.unaligned_op.di = NULL;
1116 /* Do various machine dependent initializations. */
1117 sparc_init_modes ();
1119 /* Set up function hooks. */
1120 init_machine_status = sparc_init_machine_status;
1125 case PROCESSOR_CYPRESS:
1126 sparc_costs = &cypress_costs;
1129 case PROCESSOR_SPARCLITE:
1130 case PROCESSOR_SUPERSPARC:
1131 sparc_costs = &supersparc_costs;
1133 case PROCESSOR_F930:
1134 case PROCESSOR_F934:
1135 case PROCESSOR_HYPERSPARC:
1136 case PROCESSOR_SPARCLITE86X:
1137 sparc_costs = &hypersparc_costs;
1139 case PROCESSOR_LEON:
1140 sparc_costs = &leon_costs;
1142 case PROCESSOR_SPARCLET:
1143 case PROCESSOR_TSC701:
1144 sparc_costs = &sparclet_costs;
1147 case PROCESSOR_ULTRASPARC:
1148 sparc_costs = &ultrasparc_costs;
1150 case PROCESSOR_ULTRASPARC3:
1151 sparc_costs = &ultrasparc3_costs;
1153 case PROCESSOR_NIAGARA:
1154 sparc_costs = &niagara_costs;
1156 case PROCESSOR_NIAGARA2:
1157 sparc_costs = &niagara2_costs;
1159 case PROCESSOR_NIAGARA3:
1160 case PROCESSOR_NIAGARA4:
1161 sparc_costs = &niagara3_costs;
1163 case PROCESSOR_NATIVE:
1167 if (sparc_memory_model == SMM_DEFAULT)
1169 /* Choose the memory model for the operating system. */
1170 enum sparc_memory_model_type os_default = SUBTARGET_DEFAULT_MEMORY_MODEL;
1171 if (os_default != SMM_DEFAULT)
1172 sparc_memory_model = os_default;
1173 /* Choose the most relaxed model for the processor. */
1175 sparc_memory_model = SMM_RMO;
1177 sparc_memory_model = SMM_PSO;
1179 sparc_memory_model = SMM_SC;
1182 #ifdef TARGET_DEFAULT_LONG_DOUBLE_128
1183 if (!(target_flags_explicit & MASK_LONG_DOUBLE_128))
1184 target_flags |= MASK_LONG_DOUBLE_128;
1187 if (TARGET_DEBUG_OPTIONS)
1188 dump_target_flags ("Final target_flags", target_flags);
1190 maybe_set_param_value (PARAM_SIMULTANEOUS_PREFETCHES,
1191 ((sparc_cpu == PROCESSOR_ULTRASPARC
1192 || sparc_cpu == PROCESSOR_NIAGARA
1193 || sparc_cpu == PROCESSOR_NIAGARA2
1194 || sparc_cpu == PROCESSOR_NIAGARA3
1195 || sparc_cpu == PROCESSOR_NIAGARA4)
1197 : (sparc_cpu == PROCESSOR_ULTRASPARC3
1199 global_options.x_param_values,
1200 global_options_set.x_param_values);
1201 maybe_set_param_value (PARAM_L1_CACHE_LINE_SIZE,
1202 ((sparc_cpu == PROCESSOR_ULTRASPARC
1203 || sparc_cpu == PROCESSOR_ULTRASPARC3
1204 || sparc_cpu == PROCESSOR_NIAGARA
1205 || sparc_cpu == PROCESSOR_NIAGARA2
1206 || sparc_cpu == PROCESSOR_NIAGARA3
1207 || sparc_cpu == PROCESSOR_NIAGARA4)
1209 global_options.x_param_values,
1210 global_options_set.x_param_values);
1212 /* Disable save slot sharing for call-clobbered registers by default.
1213 The IRA sharing algorithm works on single registers only and this
1214 pessimizes for double floating-point registers. */
1215 if (!global_options_set.x_flag_ira_share_save_slots)
1216 flag_ira_share_save_slots = 0;
1219 /* Miscellaneous utilities. */
1221 /* Nonzero if CODE, a comparison, is suitable for use in v9 conditional move
1222 or branch on register contents instructions. */
1225 v9_regcmp_p (enum rtx_code code)
1227 return (code == EQ || code == NE || code == GE || code == LT
1228 || code == LE || code == GT);
1231 /* Nonzero if OP is a floating point constant which can
1232 be loaded into an integer register using a single
1233 sethi instruction. */
1238 if (GET_CODE (op) == CONST_DOUBLE)
1243 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
1244 REAL_VALUE_TO_TARGET_SINGLE (r, i);
1245 return !SPARC_SIMM13_P (i) && SPARC_SETHI_P (i);
1251 /* Nonzero if OP is a floating point constant which can
1252 be loaded into an integer register using a single
1258 if (GET_CODE (op) == CONST_DOUBLE)
1263 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
1264 REAL_VALUE_TO_TARGET_SINGLE (r, i);
1265 return SPARC_SIMM13_P (i);
1271 /* Nonzero if OP is a floating point constant which can
1272 be loaded into an integer register using a high/losum
1273 instruction sequence. */
1276 fp_high_losum_p (rtx op)
1278 /* The constraints calling this should only be in
1279 SFmode move insns, so any constant which cannot
1280 be moved using a single insn will do. */
1281 if (GET_CODE (op) == CONST_DOUBLE)
1286 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
1287 REAL_VALUE_TO_TARGET_SINGLE (r, i);
1288 return !SPARC_SIMM13_P (i) && !SPARC_SETHI_P (i);
1294 /* Return true if the address of LABEL can be loaded by means of the
1295 mov{si,di}_pic_label_ref patterns in PIC mode. */
1298 can_use_mov_pic_label_ref (rtx label)
1300 /* VxWorks does not impose a fixed gap between segments; the run-time
1301 gap can be different from the object-file gap. We therefore can't
1302 assume X - _GLOBAL_OFFSET_TABLE_ is a link-time constant unless we
1303 are absolutely sure that X is in the same segment as the GOT.
1304 Unfortunately, the flexibility of linker scripts means that we
1305 can't be sure of that in general, so assume that GOT-relative
1306 accesses are never valid on VxWorks. */
1307 if (TARGET_VXWORKS_RTP)
1310 /* Similarly, if the label is non-local, it might end up being placed
1311 in a different section than the current one; now mov_pic_label_ref
1312 requires the label and the code to be in the same section. */
1313 if (LABEL_REF_NONLOCAL_P (label))
1316 /* Finally, if we are reordering basic blocks and partition into hot
1317 and cold sections, this might happen for any label. */
1318 if (flag_reorder_blocks_and_partition)
1324 /* Expand a move instruction. Return true if all work is done. */
1327 sparc_expand_move (enum machine_mode mode, rtx *operands)
1329 /* Handle sets of MEM first. */
1330 if (GET_CODE (operands[0]) == MEM)
1332 /* 0 is a register (or a pair of registers) on SPARC. */
1333 if (register_or_zero_operand (operands[1], mode))
1336 if (!reload_in_progress)
1338 operands[0] = validize_mem (operands[0]);
1339 operands[1] = force_reg (mode, operands[1]);
1343 /* Fixup TLS cases. */
1345 && CONSTANT_P (operands[1])
1346 && sparc_tls_referenced_p (operands [1]))
1348 operands[1] = sparc_legitimize_tls_address (operands[1]);
1352 /* Fixup PIC cases. */
1353 if (flag_pic && CONSTANT_P (operands[1]))
1355 if (pic_address_needs_scratch (operands[1]))
1356 operands[1] = sparc_legitimize_pic_address (operands[1], NULL_RTX);
1358 /* We cannot use the mov{si,di}_pic_label_ref patterns in all cases. */
1359 if (GET_CODE (operands[1]) == LABEL_REF
1360 && can_use_mov_pic_label_ref (operands[1]))
1364 emit_insn (gen_movsi_pic_label_ref (operands[0], operands[1]));
1370 gcc_assert (TARGET_ARCH64);
1371 emit_insn (gen_movdi_pic_label_ref (operands[0], operands[1]));
1376 if (symbolic_operand (operands[1], mode))
1379 = sparc_legitimize_pic_address (operands[1],
1381 ? operands[0] : NULL_RTX);
1386 /* If we are trying to toss an integer constant into FP registers,
1387 or loading a FP or vector constant, force it into memory. */
1388 if (CONSTANT_P (operands[1])
1389 && REG_P (operands[0])
1390 && (SPARC_FP_REG_P (REGNO (operands[0]))
1391 || SCALAR_FLOAT_MODE_P (mode)
1392 || VECTOR_MODE_P (mode)))
1394 /* emit_group_store will send such bogosity to us when it is
1395 not storing directly into memory. So fix this up to avoid
1396 crashes in output_constant_pool. */
1397 if (operands [1] == const0_rtx)
1398 operands[1] = CONST0_RTX (mode);
1400 /* We can clear or set to all-ones FP registers if TARGET_VIS, and
1401 always other regs. */
1402 if ((TARGET_VIS || REGNO (operands[0]) < SPARC_FIRST_FP_REG)
1403 && (const_zero_operand (operands[1], mode)
1404 || const_all_ones_operand (operands[1], mode)))
1407 if (REGNO (operands[0]) < SPARC_FIRST_FP_REG
1408 /* We are able to build any SF constant in integer registers
1409 with at most 2 instructions. */
1411 /* And any DF constant in integer registers. */
1413 && ! can_create_pseudo_p ())))
1416 operands[1] = force_const_mem (mode, operands[1]);
1417 if (!reload_in_progress)
1418 operands[1] = validize_mem (operands[1]);
1422 /* Accept non-constants and valid constants unmodified. */
1423 if (!CONSTANT_P (operands[1])
1424 || GET_CODE (operands[1]) == HIGH
1425 || input_operand (operands[1], mode))
1431 /* All QImode constants require only one insn, so proceed. */
1436 sparc_emit_set_const32 (operands[0], operands[1]);
1440 /* input_operand should have filtered out 32-bit mode. */
1441 sparc_emit_set_const64 (operands[0], operands[1]);
1451 /* Load OP1, a 32-bit constant, into OP0, a register.
1452 We know it can't be done in one insn when we get
1453 here, the move expander guarantees this. */
1456 sparc_emit_set_const32 (rtx op0, rtx op1)
1458 enum machine_mode mode = GET_MODE (op0);
1461 if (can_create_pseudo_p ())
1462 temp = gen_reg_rtx (mode);
1464 if (GET_CODE (op1) == CONST_INT)
1466 gcc_assert (!small_int_operand (op1, mode)
1467 && !const_high_operand (op1, mode));
1469 /* Emit them as real moves instead of a HIGH/LO_SUM,
1470 this way CSE can see everything and reuse intermediate
1471 values if it wants. */
1472 emit_insn (gen_rtx_SET (VOIDmode, temp,
1473 GEN_INT (INTVAL (op1)
1474 & ~(HOST_WIDE_INT)0x3ff)));
1476 emit_insn (gen_rtx_SET (VOIDmode,
1478 gen_rtx_IOR (mode, temp,
1479 GEN_INT (INTVAL (op1) & 0x3ff))));
1483 /* A symbol, emit in the traditional way. */
1484 emit_insn (gen_rtx_SET (VOIDmode, temp,
1485 gen_rtx_HIGH (mode, op1)));
1486 emit_insn (gen_rtx_SET (VOIDmode,
1487 op0, gen_rtx_LO_SUM (mode, temp, op1)));
1491 /* Load OP1, a symbolic 64-bit constant, into OP0, a DImode register.
1492 If TEMP is nonzero, we are forbidden to use any other scratch
1493 registers. Otherwise, we are allowed to generate them as needed.
1495 Note that TEMP may have TImode if the code model is TARGET_CM_MEDANY
1496 or TARGET_CM_EMBMEDANY (see the reload_indi and reload_outdi patterns). */
1499 sparc_emit_set_symbolic_const64 (rtx op0, rtx op1, rtx temp)
1501 rtx temp1, temp2, temp3, temp4, temp5;
1504 if (temp && GET_MODE (temp) == TImode)
1507 temp = gen_rtx_REG (DImode, REGNO (temp));
1510 /* SPARC-V9 code-model support. */
1511 switch (sparc_cmodel)
1514 /* The range spanned by all instructions in the object is less
1515 than 2^31 bytes (2GB) and the distance from any instruction
1516 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1517 than 2^31 bytes (2GB).
1519 The executable must be in the low 4TB of the virtual address
1522 sethi %hi(symbol), %temp1
1523 or %temp1, %lo(symbol), %reg */
1525 temp1 = temp; /* op0 is allowed. */
1527 temp1 = gen_reg_rtx (DImode);
1529 emit_insn (gen_rtx_SET (VOIDmode, temp1, gen_rtx_HIGH (DImode, op1)));
1530 emit_insn (gen_rtx_SET (VOIDmode, op0, gen_rtx_LO_SUM (DImode, temp1, op1)));
1534 /* The range spanned by all instructions in the object is less
1535 than 2^31 bytes (2GB) and the distance from any instruction
1536 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1537 than 2^31 bytes (2GB).
1539 The executable must be in the low 16TB of the virtual address
1542 sethi %h44(symbol), %temp1
1543 or %temp1, %m44(symbol), %temp2
1544 sllx %temp2, 12, %temp3
1545 or %temp3, %l44(symbol), %reg */
1550 temp3 = temp; /* op0 is allowed. */
1554 temp1 = gen_reg_rtx (DImode);
1555 temp2 = gen_reg_rtx (DImode);
1556 temp3 = gen_reg_rtx (DImode);
1559 emit_insn (gen_seth44 (temp1, op1));
1560 emit_insn (gen_setm44 (temp2, temp1, op1));
1561 emit_insn (gen_rtx_SET (VOIDmode, temp3,
1562 gen_rtx_ASHIFT (DImode, temp2, GEN_INT (12))));
1563 emit_insn (gen_setl44 (op0, temp3, op1));
1567 /* The range spanned by all instructions in the object is less
1568 than 2^31 bytes (2GB) and the distance from any instruction
1569 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1570 than 2^31 bytes (2GB).
1572 The executable can be placed anywhere in the virtual address
1575 sethi %hh(symbol), %temp1
1576 sethi %lm(symbol), %temp2
1577 or %temp1, %hm(symbol), %temp3
1578 sllx %temp3, 32, %temp4
1579 or %temp4, %temp2, %temp5
1580 or %temp5, %lo(symbol), %reg */
1583 /* It is possible that one of the registers we got for operands[2]
1584 might coincide with that of operands[0] (which is why we made
1585 it TImode). Pick the other one to use as our scratch. */
1586 if (rtx_equal_p (temp, op0))
1588 gcc_assert (ti_temp);
1589 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1592 temp2 = temp; /* op0 is _not_ allowed, see above. */
1599 temp1 = gen_reg_rtx (DImode);
1600 temp2 = gen_reg_rtx (DImode);
1601 temp3 = gen_reg_rtx (DImode);
1602 temp4 = gen_reg_rtx (DImode);
1603 temp5 = gen_reg_rtx (DImode);
1606 emit_insn (gen_sethh (temp1, op1));
1607 emit_insn (gen_setlm (temp2, op1));
1608 emit_insn (gen_sethm (temp3, temp1, op1));
1609 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1610 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1611 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1612 gen_rtx_PLUS (DImode, temp4, temp2)));
1613 emit_insn (gen_setlo (op0, temp5, op1));
1617 /* Old old old backwards compatibility kruft here.
1618 Essentially it is MEDLOW with a fixed 64-bit
1619 virtual base added to all data segment addresses.
1620 Text-segment stuff is computed like MEDANY, we can't
1621 reuse the code above because the relocation knobs
1624 Data segment: sethi %hi(symbol), %temp1
1625 add %temp1, EMBMEDANY_BASE_REG, %temp2
1626 or %temp2, %lo(symbol), %reg */
1627 if (data_segment_operand (op1, GET_MODE (op1)))
1631 temp1 = temp; /* op0 is allowed. */
1636 temp1 = gen_reg_rtx (DImode);
1637 temp2 = gen_reg_rtx (DImode);
1640 emit_insn (gen_embmedany_sethi (temp1, op1));
1641 emit_insn (gen_embmedany_brsum (temp2, temp1));
1642 emit_insn (gen_embmedany_losum (op0, temp2, op1));
1645 /* Text segment: sethi %uhi(symbol), %temp1
1646 sethi %hi(symbol), %temp2
1647 or %temp1, %ulo(symbol), %temp3
1648 sllx %temp3, 32, %temp4
1649 or %temp4, %temp2, %temp5
1650 or %temp5, %lo(symbol), %reg */
1655 /* It is possible that one of the registers we got for operands[2]
1656 might coincide with that of operands[0] (which is why we made
1657 it TImode). Pick the other one to use as our scratch. */
1658 if (rtx_equal_p (temp, op0))
1660 gcc_assert (ti_temp);
1661 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1664 temp2 = temp; /* op0 is _not_ allowed, see above. */
1671 temp1 = gen_reg_rtx (DImode);
1672 temp2 = gen_reg_rtx (DImode);
1673 temp3 = gen_reg_rtx (DImode);
1674 temp4 = gen_reg_rtx (DImode);
1675 temp5 = gen_reg_rtx (DImode);
1678 emit_insn (gen_embmedany_textuhi (temp1, op1));
1679 emit_insn (gen_embmedany_texthi (temp2, op1));
1680 emit_insn (gen_embmedany_textulo (temp3, temp1, op1));
1681 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1682 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1683 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1684 gen_rtx_PLUS (DImode, temp4, temp2)));
1685 emit_insn (gen_embmedany_textlo (op0, temp5, op1));
1694 #if HOST_BITS_PER_WIDE_INT == 32
1696 sparc_emit_set_const64 (rtx op0 ATTRIBUTE_UNUSED, rtx op1 ATTRIBUTE_UNUSED)
1701 /* These avoid problems when cross compiling. If we do not
1702 go through all this hair then the optimizer will see
1703 invalid REG_EQUAL notes or in some cases none at all. */
1704 static rtx gen_safe_HIGH64 (rtx, HOST_WIDE_INT);
1705 static rtx gen_safe_SET64 (rtx, HOST_WIDE_INT);
1706 static rtx gen_safe_OR64 (rtx, HOST_WIDE_INT);
1707 static rtx gen_safe_XOR64 (rtx, HOST_WIDE_INT);
1709 /* The optimizer is not to assume anything about exactly
1710 which bits are set for a HIGH, they are unspecified.
1711 Unfortunately this leads to many missed optimizations
1712 during CSE. We mask out the non-HIGH bits, and matches
1713 a plain movdi, to alleviate this problem. */
1715 gen_safe_HIGH64 (rtx dest, HOST_WIDE_INT val)
1717 return gen_rtx_SET (VOIDmode, dest, GEN_INT (val & ~(HOST_WIDE_INT)0x3ff));
1721 gen_safe_SET64 (rtx dest, HOST_WIDE_INT val)
1723 return gen_rtx_SET (VOIDmode, dest, GEN_INT (val));
1727 gen_safe_OR64 (rtx src, HOST_WIDE_INT val)
1729 return gen_rtx_IOR (DImode, src, GEN_INT (val));
1733 gen_safe_XOR64 (rtx src, HOST_WIDE_INT val)
1735 return gen_rtx_XOR (DImode, src, GEN_INT (val));
1738 /* Worker routines for 64-bit constant formation on arch64.
1739 One of the key things to be doing in these emissions is
1740 to create as many temp REGs as possible. This makes it
1741 possible for half-built constants to be used later when
1742 such values are similar to something required later on.
1743 Without doing this, the optimizer cannot see such
1746 static void sparc_emit_set_const64_quick1 (rtx, rtx,
1747 unsigned HOST_WIDE_INT, int);
1750 sparc_emit_set_const64_quick1 (rtx op0, rtx temp,
1751 unsigned HOST_WIDE_INT low_bits, int is_neg)
1753 unsigned HOST_WIDE_INT high_bits;
1756 high_bits = (~low_bits) & 0xffffffff;
1758 high_bits = low_bits;
1760 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1763 emit_insn (gen_rtx_SET (VOIDmode, op0,
1764 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1768 /* If we are XOR'ing with -1, then we should emit a one's complement
1769 instead. This way the combiner will notice logical operations
1770 such as ANDN later on and substitute. */
1771 if ((low_bits & 0x3ff) == 0x3ff)
1773 emit_insn (gen_rtx_SET (VOIDmode, op0,
1774 gen_rtx_NOT (DImode, temp)));
1778 emit_insn (gen_rtx_SET (VOIDmode, op0,
1779 gen_safe_XOR64 (temp,
1780 (-(HOST_WIDE_INT)0x400
1781 | (low_bits & 0x3ff)))));
1786 static void sparc_emit_set_const64_quick2 (rtx, rtx, unsigned HOST_WIDE_INT,
1787 unsigned HOST_WIDE_INT, int);
1790 sparc_emit_set_const64_quick2 (rtx op0, rtx temp,
1791 unsigned HOST_WIDE_INT high_bits,
1792 unsigned HOST_WIDE_INT low_immediate,
1797 if ((high_bits & 0xfffffc00) != 0)
1799 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1800 if ((high_bits & ~0xfffffc00) != 0)
1801 emit_insn (gen_rtx_SET (VOIDmode, op0,
1802 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1808 emit_insn (gen_safe_SET64 (temp, high_bits));
1812 /* Now shift it up into place. */
1813 emit_insn (gen_rtx_SET (VOIDmode, op0,
1814 gen_rtx_ASHIFT (DImode, temp2,
1815 GEN_INT (shift_count))));
1817 /* If there is a low immediate part piece, finish up by
1818 putting that in as well. */
1819 if (low_immediate != 0)
1820 emit_insn (gen_rtx_SET (VOIDmode, op0,
1821 gen_safe_OR64 (op0, low_immediate)));
1824 static void sparc_emit_set_const64_longway (rtx, rtx, unsigned HOST_WIDE_INT,
1825 unsigned HOST_WIDE_INT);
1827 /* Full 64-bit constant decomposition. Even though this is the
1828 'worst' case, we still optimize a few things away. */
1830 sparc_emit_set_const64_longway (rtx op0, rtx temp,
1831 unsigned HOST_WIDE_INT high_bits,
1832 unsigned HOST_WIDE_INT low_bits)
1836 if (can_create_pseudo_p ())
1837 sub_temp = gen_reg_rtx (DImode);
1839 if ((high_bits & 0xfffffc00) != 0)
1841 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1842 if ((high_bits & ~0xfffffc00) != 0)
1843 emit_insn (gen_rtx_SET (VOIDmode,
1845 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1851 emit_insn (gen_safe_SET64 (temp, high_bits));
1855 if (can_create_pseudo_p ())
1857 rtx temp2 = gen_reg_rtx (DImode);
1858 rtx temp3 = gen_reg_rtx (DImode);
1859 rtx temp4 = gen_reg_rtx (DImode);
1861 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1862 gen_rtx_ASHIFT (DImode, sub_temp,
1865 emit_insn (gen_safe_HIGH64 (temp2, low_bits));
1866 if ((low_bits & ~0xfffffc00) != 0)
1868 emit_insn (gen_rtx_SET (VOIDmode, temp3,
1869 gen_safe_OR64 (temp2, (low_bits & 0x3ff))));
1870 emit_insn (gen_rtx_SET (VOIDmode, op0,
1871 gen_rtx_PLUS (DImode, temp4, temp3)));
1875 emit_insn (gen_rtx_SET (VOIDmode, op0,
1876 gen_rtx_PLUS (DImode, temp4, temp2)));
1881 rtx low1 = GEN_INT ((low_bits >> (32 - 12)) & 0xfff);
1882 rtx low2 = GEN_INT ((low_bits >> (32 - 12 - 12)) & 0xfff);
1883 rtx low3 = GEN_INT ((low_bits >> (32 - 12 - 12 - 8)) & 0x0ff);
1886 /* We are in the middle of reload, so this is really
1887 painful. However we do still make an attempt to
1888 avoid emitting truly stupid code. */
1889 if (low1 != const0_rtx)
1891 emit_insn (gen_rtx_SET (VOIDmode, op0,
1892 gen_rtx_ASHIFT (DImode, sub_temp,
1893 GEN_INT (to_shift))));
1894 emit_insn (gen_rtx_SET (VOIDmode, op0,
1895 gen_rtx_IOR (DImode, op0, low1)));
1903 if (low2 != const0_rtx)
1905 emit_insn (gen_rtx_SET (VOIDmode, op0,
1906 gen_rtx_ASHIFT (DImode, sub_temp,
1907 GEN_INT (to_shift))));
1908 emit_insn (gen_rtx_SET (VOIDmode, op0,
1909 gen_rtx_IOR (DImode, op0, low2)));
1917 emit_insn (gen_rtx_SET (VOIDmode, op0,
1918 gen_rtx_ASHIFT (DImode, sub_temp,
1919 GEN_INT (to_shift))));
1920 if (low3 != const0_rtx)
1921 emit_insn (gen_rtx_SET (VOIDmode, op0,
1922 gen_rtx_IOR (DImode, op0, low3)));
1927 /* Analyze a 64-bit constant for certain properties. */
1928 static void analyze_64bit_constant (unsigned HOST_WIDE_INT,
1929 unsigned HOST_WIDE_INT,
1930 int *, int *, int *);
1933 analyze_64bit_constant (unsigned HOST_WIDE_INT high_bits,
1934 unsigned HOST_WIDE_INT low_bits,
1935 int *hbsp, int *lbsp, int *abbasp)
1937 int lowest_bit_set, highest_bit_set, all_bits_between_are_set;
1940 lowest_bit_set = highest_bit_set = -1;
1944 if ((lowest_bit_set == -1)
1945 && ((low_bits >> i) & 1))
1947 if ((highest_bit_set == -1)
1948 && ((high_bits >> (32 - i - 1)) & 1))
1949 highest_bit_set = (64 - i - 1);
1952 && ((highest_bit_set == -1)
1953 || (lowest_bit_set == -1)));
1959 if ((lowest_bit_set == -1)
1960 && ((high_bits >> i) & 1))
1961 lowest_bit_set = i + 32;
1962 if ((highest_bit_set == -1)
1963 && ((low_bits >> (32 - i - 1)) & 1))
1964 highest_bit_set = 32 - i - 1;
1967 && ((highest_bit_set == -1)
1968 || (lowest_bit_set == -1)));
1970 /* If there are no bits set this should have gone out
1971 as one instruction! */
1972 gcc_assert (lowest_bit_set != -1 && highest_bit_set != -1);
1973 all_bits_between_are_set = 1;
1974 for (i = lowest_bit_set; i <= highest_bit_set; i++)
1978 if ((low_bits & (1 << i)) != 0)
1983 if ((high_bits & (1 << (i - 32))) != 0)
1986 all_bits_between_are_set = 0;
1989 *hbsp = highest_bit_set;
1990 *lbsp = lowest_bit_set;
1991 *abbasp = all_bits_between_are_set;
1994 static int const64_is_2insns (unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT);
1997 const64_is_2insns (unsigned HOST_WIDE_INT high_bits,
1998 unsigned HOST_WIDE_INT low_bits)
2000 int highest_bit_set, lowest_bit_set, all_bits_between_are_set;
2003 || high_bits == 0xffffffff)
2006 analyze_64bit_constant (high_bits, low_bits,
2007 &highest_bit_set, &lowest_bit_set,
2008 &all_bits_between_are_set);
2010 if ((highest_bit_set == 63
2011 || lowest_bit_set == 0)
2012 && all_bits_between_are_set != 0)
2015 if ((highest_bit_set - lowest_bit_set) < 21)
2021 static unsigned HOST_WIDE_INT create_simple_focus_bits (unsigned HOST_WIDE_INT,
2022 unsigned HOST_WIDE_INT,
2025 static unsigned HOST_WIDE_INT
2026 create_simple_focus_bits (unsigned HOST_WIDE_INT high_bits,
2027 unsigned HOST_WIDE_INT low_bits,
2028 int lowest_bit_set, int shift)
2030 HOST_WIDE_INT hi, lo;
2032 if (lowest_bit_set < 32)
2034 lo = (low_bits >> lowest_bit_set) << shift;
2035 hi = ((high_bits << (32 - lowest_bit_set)) << shift);
2040 hi = ((high_bits >> (lowest_bit_set - 32)) << shift);
2042 gcc_assert (! (hi & lo));
2046 /* Here we are sure to be arch64 and this is an integer constant
2047 being loaded into a register. Emit the most efficient
2048 insn sequence possible. Detection of all the 1-insn cases
2049 has been done already. */
2051 sparc_emit_set_const64 (rtx op0, rtx op1)
2053 unsigned HOST_WIDE_INT high_bits, low_bits;
2054 int lowest_bit_set, highest_bit_set;
2055 int all_bits_between_are_set;
2058 /* Sanity check that we know what we are working with. */
2059 gcc_assert (TARGET_ARCH64
2060 && (GET_CODE (op0) == SUBREG
2061 || (REG_P (op0) && ! SPARC_FP_REG_P (REGNO (op0)))));
2063 if (! can_create_pseudo_p ())
2066 if (GET_CODE (op1) != CONST_INT)
2068 sparc_emit_set_symbolic_const64 (op0, op1, temp);
2073 temp = gen_reg_rtx (DImode);
2075 high_bits = ((INTVAL (op1) >> 32) & 0xffffffff);
2076 low_bits = (INTVAL (op1) & 0xffffffff);
2078 /* low_bits bits 0 --> 31
2079 high_bits bits 32 --> 63 */
2081 analyze_64bit_constant (high_bits, low_bits,
2082 &highest_bit_set, &lowest_bit_set,
2083 &all_bits_between_are_set);
2085 /* First try for a 2-insn sequence. */
2087 /* These situations are preferred because the optimizer can
2088 * do more things with them:
2090 * sllx %reg, shift, %reg
2092 * srlx %reg, shift, %reg
2093 * 3) mov some_small_const, %reg
2094 * sllx %reg, shift, %reg
2096 if (((highest_bit_set == 63
2097 || lowest_bit_set == 0)
2098 && all_bits_between_are_set != 0)
2099 || ((highest_bit_set - lowest_bit_set) < 12))
2101 HOST_WIDE_INT the_const = -1;
2102 int shift = lowest_bit_set;
2104 if ((highest_bit_set != 63
2105 && lowest_bit_set != 0)
2106 || all_bits_between_are_set == 0)
2109 create_simple_focus_bits (high_bits, low_bits,
2112 else if (lowest_bit_set == 0)
2113 shift = -(63 - highest_bit_set);
2115 gcc_assert (SPARC_SIMM13_P (the_const));
2116 gcc_assert (shift != 0);
2118 emit_insn (gen_safe_SET64 (temp, the_const));
2120 emit_insn (gen_rtx_SET (VOIDmode,
2122 gen_rtx_ASHIFT (DImode,
2126 emit_insn (gen_rtx_SET (VOIDmode,
2128 gen_rtx_LSHIFTRT (DImode,
2130 GEN_INT (-shift))));
2134 /* Now a range of 22 or less bits set somewhere.
2135 * 1) sethi %hi(focus_bits), %reg
2136 * sllx %reg, shift, %reg
2137 * 2) sethi %hi(focus_bits), %reg
2138 * srlx %reg, shift, %reg
2140 if ((highest_bit_set - lowest_bit_set) < 21)
2142 unsigned HOST_WIDE_INT focus_bits =
2143 create_simple_focus_bits (high_bits, low_bits,
2144 lowest_bit_set, 10);
2146 gcc_assert (SPARC_SETHI_P (focus_bits));
2147 gcc_assert (lowest_bit_set != 10);
2149 emit_insn (gen_safe_HIGH64 (temp, focus_bits));
2151 /* If lowest_bit_set == 10 then a sethi alone could have done it. */
2152 if (lowest_bit_set < 10)
2153 emit_insn (gen_rtx_SET (VOIDmode,
2155 gen_rtx_LSHIFTRT (DImode, temp,
2156 GEN_INT (10 - lowest_bit_set))));
2157 else if (lowest_bit_set > 10)
2158 emit_insn (gen_rtx_SET (VOIDmode,
2160 gen_rtx_ASHIFT (DImode, temp,
2161 GEN_INT (lowest_bit_set - 10))));
2165 /* 1) sethi %hi(low_bits), %reg
2166 * or %reg, %lo(low_bits), %reg
2167 * 2) sethi %hi(~low_bits), %reg
2168 * xor %reg, %lo(-0x400 | (low_bits & 0x3ff)), %reg
2171 || high_bits == 0xffffffff)
2173 sparc_emit_set_const64_quick1 (op0, temp, low_bits,
2174 (high_bits == 0xffffffff));
2178 /* Now, try 3-insn sequences. */
2180 /* 1) sethi %hi(high_bits), %reg
2181 * or %reg, %lo(high_bits), %reg
2182 * sllx %reg, 32, %reg
2186 sparc_emit_set_const64_quick2 (op0, temp, high_bits, 0, 32);
2190 /* We may be able to do something quick
2191 when the constant is negated, so try that. */
2192 if (const64_is_2insns ((~high_bits) & 0xffffffff,
2193 (~low_bits) & 0xfffffc00))
2195 /* NOTE: The trailing bits get XOR'd so we need the
2196 non-negated bits, not the negated ones. */
2197 unsigned HOST_WIDE_INT trailing_bits = low_bits & 0x3ff;
2199 if ((((~high_bits) & 0xffffffff) == 0
2200 && ((~low_bits) & 0x80000000) == 0)
2201 || (((~high_bits) & 0xffffffff) == 0xffffffff
2202 && ((~low_bits) & 0x80000000) != 0))
2204 unsigned HOST_WIDE_INT fast_int = (~low_bits & 0xffffffff);
2206 if ((SPARC_SETHI_P (fast_int)
2207 && (~high_bits & 0xffffffff) == 0)
2208 || SPARC_SIMM13_P (fast_int))
2209 emit_insn (gen_safe_SET64 (temp, fast_int));
2211 sparc_emit_set_const64 (temp, GEN_INT (fast_int));
2216 negated_const = GEN_INT (((~low_bits) & 0xfffffc00) |
2217 (((HOST_WIDE_INT)((~high_bits) & 0xffffffff))<<32));
2218 sparc_emit_set_const64 (temp, negated_const);
2221 /* If we are XOR'ing with -1, then we should emit a one's complement
2222 instead. This way the combiner will notice logical operations
2223 such as ANDN later on and substitute. */
2224 if (trailing_bits == 0x3ff)
2226 emit_insn (gen_rtx_SET (VOIDmode, op0,
2227 gen_rtx_NOT (DImode, temp)));
2231 emit_insn (gen_rtx_SET (VOIDmode,
2233 gen_safe_XOR64 (temp,
2234 (-0x400 | trailing_bits))));
2239 /* 1) sethi %hi(xxx), %reg
2240 * or %reg, %lo(xxx), %reg
2241 * sllx %reg, yyy, %reg
2243 * ??? This is just a generalized version of the low_bits==0
2244 * thing above, FIXME...
2246 if ((highest_bit_set - lowest_bit_set) < 32)
2248 unsigned HOST_WIDE_INT focus_bits =
2249 create_simple_focus_bits (high_bits, low_bits,
2252 /* We can't get here in this state. */
2253 gcc_assert (highest_bit_set >= 32 && lowest_bit_set < 32);
2255 /* So what we know is that the set bits straddle the
2256 middle of the 64-bit word. */
2257 sparc_emit_set_const64_quick2 (op0, temp,
2263 /* 1) sethi %hi(high_bits), %reg
2264 * or %reg, %lo(high_bits), %reg
2265 * sllx %reg, 32, %reg
2266 * or %reg, low_bits, %reg
2268 if (SPARC_SIMM13_P(low_bits)
2269 && ((int)low_bits > 0))
2271 sparc_emit_set_const64_quick2 (op0, temp, high_bits, low_bits, 32);
2275 /* The easiest way when all else fails, is full decomposition. */
2276 sparc_emit_set_const64_longway (op0, temp, high_bits, low_bits);
2278 #endif /* HOST_BITS_PER_WIDE_INT == 32 */
2280 /* Given a comparison code (EQ, NE, etc.) and the first operand of a COMPARE,
2281 return the mode to be used for the comparison. For floating-point,
2282 CCFP[E]mode is used. CC_NOOVmode should be used when the first operand
2283 is a PLUS, MINUS, NEG, or ASHIFT. CCmode should be used when no special
2284 processing is needed. */
2287 select_cc_mode (enum rtx_code op, rtx x, rtx y ATTRIBUTE_UNUSED)
2289 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2315 else if (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS
2316 || GET_CODE (x) == NEG || GET_CODE (x) == ASHIFT)
2318 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
2319 return CCX_NOOVmode;
2325 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
2332 /* Emit the compare insn and return the CC reg for a CODE comparison
2333 with operands X and Y. */
2336 gen_compare_reg_1 (enum rtx_code code, rtx x, rtx y)
2338 enum machine_mode mode;
2341 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2344 mode = SELECT_CC_MODE (code, x, y);
2346 /* ??? We don't have movcc patterns so we cannot generate pseudo regs for the
2347 fcc regs (cse can't tell they're really call clobbered regs and will
2348 remove a duplicate comparison even if there is an intervening function
2349 call - it will then try to reload the cc reg via an int reg which is why
2350 we need the movcc patterns). It is possible to provide the movcc
2351 patterns by using the ldxfsr/stxfsr v9 insns. I tried it: you need two
2352 registers (say %g1,%g5) and it takes about 6 insns. A better fix would be
2353 to tell cse that CCFPE mode registers (even pseudos) are call
2356 /* ??? This is an experiment. Rather than making changes to cse which may
2357 or may not be easy/clean, we do our own cse. This is possible because
2358 we will generate hard registers. Cse knows they're call clobbered (it
2359 doesn't know the same thing about pseudos). If we guess wrong, no big
2360 deal, but if we win, great! */
2362 if (TARGET_V9 && GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2363 #if 1 /* experiment */
2366 /* We cycle through the registers to ensure they're all exercised. */
2367 static int next_fcc_reg = 0;
2368 /* Previous x,y for each fcc reg. */
2369 static rtx prev_args[4][2];
2371 /* Scan prev_args for x,y. */
2372 for (reg = 0; reg < 4; reg++)
2373 if (prev_args[reg][0] == x && prev_args[reg][1] == y)
2378 prev_args[reg][0] = x;
2379 prev_args[reg][1] = y;
2380 next_fcc_reg = (next_fcc_reg + 1) & 3;
2382 cc_reg = gen_rtx_REG (mode, reg + SPARC_FIRST_V9_FCC_REG);
2385 cc_reg = gen_reg_rtx (mode);
2386 #endif /* ! experiment */
2387 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2388 cc_reg = gen_rtx_REG (mode, SPARC_FCC_REG);
2390 cc_reg = gen_rtx_REG (mode, SPARC_ICC_REG);
2392 /* We shouldn't get there for TFmode if !TARGET_HARD_QUAD. If we do, this
2393 will only result in an unrecognizable insn so no point in asserting. */
2394 emit_insn (gen_rtx_SET (VOIDmode, cc_reg, gen_rtx_COMPARE (mode, x, y)));
2400 /* Emit the compare insn and return the CC reg for the comparison in CMP. */
2403 gen_compare_reg (rtx cmp)
2405 return gen_compare_reg_1 (GET_CODE (cmp), XEXP (cmp, 0), XEXP (cmp, 1));
2408 /* This function is used for v9 only.
2409 DEST is the target of the Scc insn.
2410 CODE is the code for an Scc's comparison.
2411 X and Y are the values we compare.
2413 This function is needed to turn
2416 (gt (reg:CCX 100 %icc)
2420 (gt:DI (reg:CCX 100 %icc)
2423 IE: The instruction recognizer needs to see the mode of the comparison to
2424 find the right instruction. We could use "gt:DI" right in the
2425 define_expand, but leaving it out allows us to handle DI, SI, etc. */
2428 gen_v9_scc (rtx dest, enum rtx_code compare_code, rtx x, rtx y)
2431 && (GET_MODE (x) == DImode
2432 || GET_MODE (dest) == DImode))
2435 /* Try to use the movrCC insns. */
2437 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
2439 && v9_regcmp_p (compare_code))
2444 /* Special case for op0 != 0. This can be done with one instruction if
2447 if (compare_code == NE
2448 && GET_MODE (dest) == DImode
2449 && rtx_equal_p (op0, dest))
2451 emit_insn (gen_rtx_SET (VOIDmode, dest,
2452 gen_rtx_IF_THEN_ELSE (DImode,
2453 gen_rtx_fmt_ee (compare_code, DImode,
2460 if (reg_overlap_mentioned_p (dest, op0))
2462 /* Handle the case where dest == x.
2463 We "early clobber" the result. */
2464 op0 = gen_reg_rtx (GET_MODE (x));
2465 emit_move_insn (op0, x);
2468 emit_insn (gen_rtx_SET (VOIDmode, dest, const0_rtx));
2469 if (GET_MODE (op0) != DImode)
2471 temp = gen_reg_rtx (DImode);
2472 convert_move (temp, op0, 0);
2476 emit_insn (gen_rtx_SET (VOIDmode, dest,
2477 gen_rtx_IF_THEN_ELSE (GET_MODE (dest),
2478 gen_rtx_fmt_ee (compare_code, DImode,
2486 x = gen_compare_reg_1 (compare_code, x, y);
2489 gcc_assert (GET_MODE (x) != CC_NOOVmode
2490 && GET_MODE (x) != CCX_NOOVmode);
2492 emit_insn (gen_rtx_SET (VOIDmode, dest, const0_rtx));
2493 emit_insn (gen_rtx_SET (VOIDmode, dest,
2494 gen_rtx_IF_THEN_ELSE (GET_MODE (dest),
2495 gen_rtx_fmt_ee (compare_code,
2496 GET_MODE (x), x, y),
2497 const1_rtx, dest)));
2503 /* Emit an scc insn. For seq, sne, sgeu, and sltu, we can do this
2504 without jumps using the addx/subx instructions. */
2507 emit_scc_insn (rtx operands[])
2514 /* The quad-word fp compare library routines all return nonzero to indicate
2515 true, which is different from the equivalent libgcc routines, so we must
2516 handle them specially here. */
2517 if (GET_MODE (operands[2]) == TFmode && ! TARGET_HARD_QUAD)
2519 operands[1] = sparc_emit_float_lib_cmp (operands[2], operands[3],
2520 GET_CODE (operands[1]));
2521 operands[2] = XEXP (operands[1], 0);
2522 operands[3] = XEXP (operands[1], 1);
2525 code = GET_CODE (operands[1]);
2529 /* For seq/sne on v9 we use the same code as v8 (the addx/subx method has
2530 more applications). The exception to this is "reg != 0" which can
2531 be done in one instruction on v9 (so we do it). */
2534 if (GET_MODE (x) == SImode)
2536 rtx pat = gen_seqsi_special (operands[0], x, y);
2540 else if (GET_MODE (x) == DImode)
2542 rtx pat = gen_seqdi_special (operands[0], x, y);
2550 if (GET_MODE (x) == SImode)
2552 rtx pat = gen_snesi_special (operands[0], x, y);
2556 else if (GET_MODE (x) == DImode)
2560 pat = gen_snedi_special_vis3 (operands[0], x, y);
2562 pat = gen_snedi_special (operands[0], x, y);
2570 && GET_MODE (x) == DImode
2572 && (code == GTU || code == LTU))
2573 && gen_v9_scc (operands[0], code, x, y))
2576 /* We can do LTU and GEU using the addx/subx instructions too. And
2577 for GTU/LEU, if both operands are registers swap them and fall
2578 back to the easy case. */
2579 if (code == GTU || code == LEU)
2581 if ((GET_CODE (x) == REG || GET_CODE (x) == SUBREG)
2582 && (GET_CODE (y) == REG || GET_CODE (y) == SUBREG))
2587 code = swap_condition (code);
2592 || (!TARGET_VIS3 && code == GEU))
2594 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
2595 gen_rtx_fmt_ee (code, SImode,
2596 gen_compare_reg_1 (code, x, y),
2601 /* All the posibilities to use addx/subx based sequences has been
2602 exhausted, try for a 3 instruction sequence using v9 conditional
2604 if (TARGET_V9 && gen_v9_scc (operands[0], code, x, y))
2607 /* Nope, do branches. */
2611 /* Emit a conditional jump insn for the v9 architecture using comparison code
2612 CODE and jump target LABEL.
2613 This function exists to take advantage of the v9 brxx insns. */
2616 emit_v9_brxx_insn (enum rtx_code code, rtx op0, rtx label)
2618 emit_jump_insn (gen_rtx_SET (VOIDmode,
2620 gen_rtx_IF_THEN_ELSE (VOIDmode,
2621 gen_rtx_fmt_ee (code, GET_MODE (op0),
2623 gen_rtx_LABEL_REF (VOIDmode, label),
2628 emit_conditional_branch_insn (rtx operands[])
2630 /* The quad-word fp compare library routines all return nonzero to indicate
2631 true, which is different from the equivalent libgcc routines, so we must
2632 handle them specially here. */
2633 if (GET_MODE (operands[1]) == TFmode && ! TARGET_HARD_QUAD)
2635 operands[0] = sparc_emit_float_lib_cmp (operands[1], operands[2],
2636 GET_CODE (operands[0]));
2637 operands[1] = XEXP (operands[0], 0);
2638 operands[2] = XEXP (operands[0], 1);
2641 if (TARGET_ARCH64 && operands[2] == const0_rtx
2642 && GET_CODE (operands[1]) == REG
2643 && GET_MODE (operands[1]) == DImode)
2645 emit_v9_brxx_insn (GET_CODE (operands[0]), operands[1], operands[3]);
2649 operands[1] = gen_compare_reg (operands[0]);
2650 operands[2] = const0_rtx;
2651 operands[0] = gen_rtx_fmt_ee (GET_CODE (operands[0]), VOIDmode,
2652 operands[1], operands[2]);
2653 emit_jump_insn (gen_cbranchcc4 (operands[0], operands[1], operands[2],
2658 /* Generate a DFmode part of a hard TFmode register.
2659 REG is the TFmode hard register, LOW is 1 for the
2660 low 64bit of the register and 0 otherwise.
2663 gen_df_reg (rtx reg, int low)
2665 int regno = REGNO (reg);
2667 if ((WORDS_BIG_ENDIAN == 0) ^ (low != 0))
2668 regno += (TARGET_ARCH64 && SPARC_INT_REG_P (regno)) ? 1 : 2;
2669 return gen_rtx_REG (DFmode, regno);
2672 /* Generate a call to FUNC with OPERANDS. Operand 0 is the return value.
2673 Unlike normal calls, TFmode operands are passed by reference. It is
2674 assumed that no more than 3 operands are required. */
2677 emit_soft_tfmode_libcall (const char *func_name, int nargs, rtx *operands)
2679 rtx ret_slot = NULL, arg[3], func_sym;
2682 /* We only expect to be called for conversions, unary, and binary ops. */
2683 gcc_assert (nargs == 2 || nargs == 3);
2685 for (i = 0; i < nargs; ++i)
2687 rtx this_arg = operands[i];
2690 /* TFmode arguments and return values are passed by reference. */
2691 if (GET_MODE (this_arg) == TFmode)
2693 int force_stack_temp;
2695 force_stack_temp = 0;
2696 if (TARGET_BUGGY_QP_LIB && i == 0)
2697 force_stack_temp = 1;
2699 if (GET_CODE (this_arg) == MEM
2700 && ! force_stack_temp)
2702 tree expr = MEM_EXPR (this_arg);
2704 mark_addressable (expr);
2705 this_arg = XEXP (this_arg, 0);
2707 else if (CONSTANT_P (this_arg)
2708 && ! force_stack_temp)
2710 this_slot = force_const_mem (TFmode, this_arg);
2711 this_arg = XEXP (this_slot, 0);
2715 this_slot = assign_stack_temp (TFmode, GET_MODE_SIZE (TFmode), 0);
2717 /* Operand 0 is the return value. We'll copy it out later. */
2719 emit_move_insn (this_slot, this_arg);
2721 ret_slot = this_slot;
2723 this_arg = XEXP (this_slot, 0);
2730 func_sym = gen_rtx_SYMBOL_REF (Pmode, func_name);
2732 if (GET_MODE (operands[0]) == TFmode)
2735 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 2,
2736 arg[0], GET_MODE (arg[0]),
2737 arg[1], GET_MODE (arg[1]));
2739 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 3,
2740 arg[0], GET_MODE (arg[0]),
2741 arg[1], GET_MODE (arg[1]),
2742 arg[2], GET_MODE (arg[2]));
2745 emit_move_insn (operands[0], ret_slot);
2751 gcc_assert (nargs == 2);
2753 ret = emit_library_call_value (func_sym, operands[0], LCT_NORMAL,
2754 GET_MODE (operands[0]), 1,
2755 arg[1], GET_MODE (arg[1]));
2757 if (ret != operands[0])
2758 emit_move_insn (operands[0], ret);
2762 /* Expand soft-float TFmode calls to sparc abi routines. */
2765 emit_soft_tfmode_binop (enum rtx_code code, rtx *operands)
2787 emit_soft_tfmode_libcall (func, 3, operands);
2791 emit_soft_tfmode_unop (enum rtx_code code, rtx *operands)
2795 gcc_assert (code == SQRT);
2798 emit_soft_tfmode_libcall (func, 2, operands);
2802 emit_soft_tfmode_cvt (enum rtx_code code, rtx *operands)
2809 switch (GET_MODE (operands[1]))
2822 case FLOAT_TRUNCATE:
2823 switch (GET_MODE (operands[0]))
2837 switch (GET_MODE (operands[1]))
2842 operands[1] = gen_rtx_SIGN_EXTEND (DImode, operands[1]);
2852 case UNSIGNED_FLOAT:
2853 switch (GET_MODE (operands[1]))
2858 operands[1] = gen_rtx_ZERO_EXTEND (DImode, operands[1]);
2869 switch (GET_MODE (operands[0]))
2883 switch (GET_MODE (operands[0]))
2900 emit_soft_tfmode_libcall (func, 2, operands);
2903 /* Expand a hard-float tfmode operation. All arguments must be in
2907 emit_hard_tfmode_operation (enum rtx_code code, rtx *operands)
2911 if (GET_RTX_CLASS (code) == RTX_UNARY)
2913 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2914 op = gen_rtx_fmt_e (code, GET_MODE (operands[0]), operands[1]);
2918 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2919 operands[2] = force_reg (GET_MODE (operands[2]), operands[2]);
2920 op = gen_rtx_fmt_ee (code, GET_MODE (operands[0]),
2921 operands[1], operands[2]);
2924 if (register_operand (operands[0], VOIDmode))
2927 dest = gen_reg_rtx (GET_MODE (operands[0]));
2929 emit_insn (gen_rtx_SET (VOIDmode, dest, op));
2931 if (dest != operands[0])
2932 emit_move_insn (operands[0], dest);
2936 emit_tfmode_binop (enum rtx_code code, rtx *operands)
2938 if (TARGET_HARD_QUAD)
2939 emit_hard_tfmode_operation (code, operands);
2941 emit_soft_tfmode_binop (code, operands);
2945 emit_tfmode_unop (enum rtx_code code, rtx *operands)
2947 if (TARGET_HARD_QUAD)
2948 emit_hard_tfmode_operation (code, operands);
2950 emit_soft_tfmode_unop (code, operands);
2954 emit_tfmode_cvt (enum rtx_code code, rtx *operands)
2956 if (TARGET_HARD_QUAD)
2957 emit_hard_tfmode_operation (code, operands);
2959 emit_soft_tfmode_cvt (code, operands);
2962 /* Return nonzero if a branch/jump/call instruction will be emitting
2963 nop into its delay slot. */
2966 empty_delay_slot (rtx insn)
2970 /* If no previous instruction (should not happen), return true. */
2971 if (PREV_INSN (insn) == NULL)
2974 seq = NEXT_INSN (PREV_INSN (insn));
2975 if (GET_CODE (PATTERN (seq)) == SEQUENCE)
2981 /* Return nonzero if TRIAL can go into the call delay slot. */
2984 tls_call_delay (rtx trial)
2989 call __tls_get_addr, %tgd_call (foo)
2990 add %l7, %o0, %o0, %tgd_add (foo)
2991 while Sun as/ld does not. */
2992 if (TARGET_GNU_TLS || !TARGET_TLS)
2995 pat = PATTERN (trial);
2997 /* We must reject tgd_add{32|64}, i.e.
2998 (set (reg) (plus (reg) (unspec [(reg) (symbol_ref)] UNSPEC_TLSGD)))
2999 and tldm_add{32|64}, i.e.
3000 (set (reg) (plus (reg) (unspec [(reg) (symbol_ref)] UNSPEC_TLSLDM)))
3002 if (GET_CODE (pat) == SET
3003 && GET_CODE (SET_SRC (pat)) == PLUS)
3005 rtx unspec = XEXP (SET_SRC (pat), 1);
3007 if (GET_CODE (unspec) == UNSPEC
3008 && (XINT (unspec, 1) == UNSPEC_TLSGD
3009 || XINT (unspec, 1) == UNSPEC_TLSLDM))
3016 /* Return nonzero if TRIAL, an insn, can be combined with a 'restore'
3017 instruction. RETURN_P is true if the v9 variant 'return' is to be
3018 considered in the test too.
3020 TRIAL must be a SET whose destination is a REG appropriate for the
3021 'restore' instruction or, if RETURN_P is true, for the 'return'
3025 eligible_for_restore_insn (rtx trial, bool return_p)
3027 rtx pat = PATTERN (trial);
3028 rtx src = SET_SRC (pat);
3029 bool src_is_freg = false;
3032 /* Since we now can do moves between float and integer registers when
3033 VIS3 is enabled, we have to catch this case. We can allow such
3034 moves when doing a 'return' however. */
3036 if (GET_CODE (src_reg) == SUBREG)
3037 src_reg = SUBREG_REG (src_reg);
3038 if (GET_CODE (src_reg) == REG
3039 && SPARC_FP_REG_P (REGNO (src_reg)))
3042 /* The 'restore src,%g0,dest' pattern for word mode and below. */
3043 if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
3044 && arith_operand (src, GET_MODE (src))
3048 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
3050 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (SImode);
3053 /* The 'restore src,%g0,dest' pattern for double-word mode. */
3054 else if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
3055 && arith_double_operand (src, GET_MODE (src))
3057 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
3059 /* The 'restore src,%g0,dest' pattern for float if no FPU. */
3060 else if (! TARGET_FPU && register_operand (src, SFmode))
3063 /* The 'restore src,%g0,dest' pattern for double if no FPU. */
3064 else if (! TARGET_FPU && TARGET_ARCH64 && register_operand (src, DFmode))
3067 /* If we have the 'return' instruction, anything that does not use
3068 local or output registers and can go into a delay slot wins. */
3071 && !epilogue_renumber (&pat, 1)
3072 && get_attr_in_uncond_branch_delay (trial)
3073 == IN_UNCOND_BRANCH_DELAY_TRUE)
3076 /* The 'restore src1,src2,dest' pattern for SImode. */
3077 else if (GET_CODE (src) == PLUS
3078 && register_operand (XEXP (src, 0), SImode)
3079 && arith_operand (XEXP (src, 1), SImode))
3082 /* The 'restore src1,src2,dest' pattern for DImode. */
3083 else if (GET_CODE (src) == PLUS
3084 && register_operand (XEXP (src, 0), DImode)
3085 && arith_double_operand (XEXP (src, 1), DImode))
3088 /* The 'restore src1,%lo(src2),dest' pattern. */
3089 else if (GET_CODE (src) == LO_SUM
3090 && ! TARGET_CM_MEDMID
3091 && ((register_operand (XEXP (src, 0), SImode)
3092 && immediate_operand (XEXP (src, 1), SImode))
3094 && register_operand (XEXP (src, 0), DImode)
3095 && immediate_operand (XEXP (src, 1), DImode))))
3098 /* The 'restore src,src,dest' pattern. */
3099 else if (GET_CODE (src) == ASHIFT
3100 && (register_operand (XEXP (src, 0), SImode)
3101 || register_operand (XEXP (src, 0), DImode))
3102 && XEXP (src, 1) == const1_rtx)
3108 /* Return nonzero if TRIAL can go into the function return's delay slot. */
3111 eligible_for_return_delay (rtx trial)
3116 if (GET_CODE (trial) != INSN)
3119 if (get_attr_length (trial) != 1)
3122 /* If the function uses __builtin_eh_return, the eh_return machinery
3123 occupies the delay slot. */
3124 if (crtl->calls_eh_return)
3127 /* In the case of a leaf or flat function, anything can go into the slot. */
3128 if (sparc_leaf_function_p || TARGET_FLAT)
3130 get_attr_in_uncond_branch_delay (trial) == IN_UNCOND_BRANCH_DELAY_TRUE;
3132 pat = PATTERN (trial);
3133 if (GET_CODE (pat) == PARALLEL)
3139 for (i = XVECLEN (pat, 0) - 1; i >= 0; i--)
3141 rtx expr = XVECEXP (pat, 0, i);
3142 if (GET_CODE (expr) != SET)
3144 if (GET_CODE (SET_DEST (expr)) != REG)
3146 regno = REGNO (SET_DEST (expr));
3147 if (regno >= 8 && regno < 24)
3150 return !epilogue_renumber (&pat, 1)
3151 && (get_attr_in_uncond_branch_delay (trial)
3152 == IN_UNCOND_BRANCH_DELAY_TRUE);
3155 if (GET_CODE (pat) != SET)
3158 if (GET_CODE (SET_DEST (pat)) != REG)
3161 regno = REGNO (SET_DEST (pat));
3163 /* Otherwise, only operations which can be done in tandem with
3164 a `restore' or `return' insn can go into the delay slot. */
3165 if (regno >= 8 && regno < 24)
3168 /* If this instruction sets up floating point register and we have a return
3169 instruction, it can probably go in. But restore will not work
3171 if (! SPARC_INT_REG_P (regno))
3173 && !epilogue_renumber (&pat, 1)
3174 && get_attr_in_uncond_branch_delay (trial)
3175 == IN_UNCOND_BRANCH_DELAY_TRUE);
3177 return eligible_for_restore_insn (trial, true);
3180 /* Return nonzero if TRIAL can go into the sibling call's delay slot. */
3183 eligible_for_sibcall_delay (rtx trial)
3187 if (GET_CODE (trial) != INSN || GET_CODE (PATTERN (trial)) != SET)
3190 if (get_attr_length (trial) != 1)
3193 pat = PATTERN (trial);
3195 if (sparc_leaf_function_p || TARGET_FLAT)
3197 /* If the tail call is done using the call instruction,
3198 we have to restore %o7 in the delay slot. */
3199 if (LEAF_SIBCALL_SLOT_RESERVED_P)
3202 /* %g1 is used to build the function address */
3203 if (reg_mentioned_p (gen_rtx_REG (Pmode, 1), pat))
3209 /* Otherwise, only operations which can be done in tandem with
3210 a `restore' insn can go into the delay slot. */
3211 if (GET_CODE (SET_DEST (pat)) != REG
3212 || (REGNO (SET_DEST (pat)) >= 8 && REGNO (SET_DEST (pat)) < 24)
3213 || ! SPARC_INT_REG_P (REGNO (SET_DEST (pat))))
3216 /* If it mentions %o7, it can't go in, because sibcall will clobber it
3218 if (reg_mentioned_p (gen_rtx_REG (Pmode, 15), pat))
3221 return eligible_for_restore_insn (trial, false);
3224 /* Determine if it's legal to put X into the constant pool. This
3225 is not possible if X contains the address of a symbol that is
3226 not constant (TLS) or not known at final link time (PIC). */
3229 sparc_cannot_force_const_mem (enum machine_mode mode, rtx x)
3231 switch (GET_CODE (x))
3236 /* Accept all non-symbolic constants. */
3240 /* Labels are OK iff we are non-PIC. */
3241 return flag_pic != 0;
3244 /* 'Naked' TLS symbol references are never OK,
3245 non-TLS symbols are OK iff we are non-PIC. */
3246 if (SYMBOL_REF_TLS_MODEL (x))
3249 return flag_pic != 0;
3252 return sparc_cannot_force_const_mem (mode, XEXP (x, 0));
3255 return sparc_cannot_force_const_mem (mode, XEXP (x, 0))
3256 || sparc_cannot_force_const_mem (mode, XEXP (x, 1));
3264 /* Global Offset Table support. */
3265 static GTY(()) rtx got_helper_rtx = NULL_RTX;
3266 static GTY(()) rtx global_offset_table_rtx = NULL_RTX;
3268 /* Return the SYMBOL_REF for the Global Offset Table. */
3270 static GTY(()) rtx sparc_got_symbol = NULL_RTX;
3275 if (!sparc_got_symbol)
3276 sparc_got_symbol = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_");
3278 return sparc_got_symbol;
3281 /* Ensure that we are not using patterns that are not OK with PIC. */
3291 op = recog_data.operand[i];
3292 gcc_assert (GET_CODE (op) != SYMBOL_REF
3293 && (GET_CODE (op) != CONST
3294 || (GET_CODE (XEXP (op, 0)) == MINUS
3295 && XEXP (XEXP (op, 0), 0) == sparc_got ()
3296 && GET_CODE (XEXP (XEXP (op, 0), 1)) == CONST)));
3303 /* Return true if X is an address which needs a temporary register when
3304 reloaded while generating PIC code. */
3307 pic_address_needs_scratch (rtx x)
3309 /* An address which is a symbolic plus a non SMALL_INT needs a temp reg. */
3310 if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS
3311 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
3312 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
3313 && ! SMALL_INT (XEXP (XEXP (x, 0), 1)))
3319 /* Determine if a given RTX is a valid constant. We already know this
3320 satisfies CONSTANT_P. */
3323 sparc_legitimate_constant_p (enum machine_mode mode, rtx x)
3325 switch (GET_CODE (x))
3329 if (sparc_tls_referenced_p (x))
3334 if (GET_MODE (x) == VOIDmode)
3337 /* Floating point constants are generally not ok.
3338 The only exception is 0.0 and all-ones in VIS. */
3340 && SCALAR_FLOAT_MODE_P (mode)
3341 && (const_zero_operand (x, mode)
3342 || const_all_ones_operand (x, mode)))
3348 /* Vector constants are generally not ok.
3349 The only exception is 0 or -1 in VIS. */
3351 && (const_zero_operand (x, mode)
3352 || const_all_ones_operand (x, mode)))
3364 /* Determine if a given RTX is a valid constant address. */
3367 constant_address_p (rtx x)
3369 switch (GET_CODE (x))
3377 if (flag_pic && pic_address_needs_scratch (x))
3379 return sparc_legitimate_constant_p (Pmode, x);
3382 return !flag_pic && sparc_legitimate_constant_p (Pmode, x);
3389 /* Nonzero if the constant value X is a legitimate general operand
3390 when generating PIC code. It is given that flag_pic is on and
3391 that X satisfies CONSTANT_P or is a CONST_DOUBLE. */
3394 legitimate_pic_operand_p (rtx x)
3396 if (pic_address_needs_scratch (x))
3398 if (sparc_tls_referenced_p (x))
3403 #define RTX_OK_FOR_OFFSET_P(X, MODE) \
3405 && INTVAL (X) >= -0x1000 \
3406 && INTVAL (X) < (0x1000 - GET_MODE_SIZE (MODE)))
3408 #define RTX_OK_FOR_OLO10_P(X, MODE) \
3410 && INTVAL (X) >= -0x1000 \
3411 && INTVAL (X) < (0xc00 - GET_MODE_SIZE (MODE)))
3413 /* Handle the TARGET_LEGITIMATE_ADDRESS_P target hook.
3415 On SPARC, the actual legitimate addresses must be REG+REG or REG+SMALLINT
3416 ordinarily. This changes a bit when generating PIC. */
3419 sparc_legitimate_address_p (enum machine_mode mode, rtx addr, bool strict)
3421 rtx rs1 = NULL, rs2 = NULL, imm1 = NULL;
3423 if (REG_P (addr) || GET_CODE (addr) == SUBREG)
3425 else if (GET_CODE (addr) == PLUS)
3427 rs1 = XEXP (addr, 0);
3428 rs2 = XEXP (addr, 1);
3430 /* Canonicalize. REG comes first, if there are no regs,
3431 LO_SUM comes first. */
3433 && GET_CODE (rs1) != SUBREG
3435 || GET_CODE (rs2) == SUBREG
3436 || (GET_CODE (rs2) == LO_SUM && GET_CODE (rs1) != LO_SUM)))
3438 rs1 = XEXP (addr, 1);
3439 rs2 = XEXP (addr, 0);
3443 && rs1 == pic_offset_table_rtx
3445 && GET_CODE (rs2) != SUBREG
3446 && GET_CODE (rs2) != LO_SUM
3447 && GET_CODE (rs2) != MEM
3448 && !(GET_CODE (rs2) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs2))
3449 && (! symbolic_operand (rs2, VOIDmode) || mode == Pmode)
3450 && (GET_CODE (rs2) != CONST_INT || SMALL_INT (rs2)))
3452 || GET_CODE (rs1) == SUBREG)
3453 && RTX_OK_FOR_OFFSET_P (rs2, mode)))
3458 else if ((REG_P (rs1) || GET_CODE (rs1) == SUBREG)
3459 && (REG_P (rs2) || GET_CODE (rs2) == SUBREG))
3461 /* We prohibit REG + REG for TFmode when there are no quad move insns
3462 and we consequently need to split. We do this because REG+REG
3463 is not an offsettable address. If we get the situation in reload
3464 where source and destination of a movtf pattern are both MEMs with
3465 REG+REG address, then only one of them gets converted to an
3466 offsettable address. */
3468 && ! (TARGET_ARCH64 && TARGET_HARD_QUAD))
3471 /* We prohibit REG + REG on ARCH32 if not optimizing for
3472 DFmode/DImode because then mem_min_alignment is likely to be zero
3473 after reload and the forced split would lack a matching splitter
3475 if (TARGET_ARCH32 && !optimize
3476 && (mode == DFmode || mode == DImode))
3479 else if (USE_AS_OFFSETABLE_LO10
3480 && GET_CODE (rs1) == LO_SUM
3482 && ! TARGET_CM_MEDMID
3483 && RTX_OK_FOR_OLO10_P (rs2, mode))
3486 imm1 = XEXP (rs1, 1);
3487 rs1 = XEXP (rs1, 0);
3488 if (!CONSTANT_P (imm1)
3489 || (GET_CODE (rs1) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs1)))
3493 else if (GET_CODE (addr) == LO_SUM)
3495 rs1 = XEXP (addr, 0);
3496 imm1 = XEXP (addr, 1);
3498 if (!CONSTANT_P (imm1)
3499 || (GET_CODE (rs1) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs1)))
3502 /* We can't allow TFmode in 32-bit mode, because an offset greater
3503 than the alignment (8) may cause the LO_SUM to overflow. */
3504 if (mode == TFmode && TARGET_ARCH32)
3507 else if (GET_CODE (addr) == CONST_INT && SMALL_INT (addr))
3512 if (GET_CODE (rs1) == SUBREG)
3513 rs1 = SUBREG_REG (rs1);
3519 if (GET_CODE (rs2) == SUBREG)
3520 rs2 = SUBREG_REG (rs2);
3527 if (!REGNO_OK_FOR_BASE_P (REGNO (rs1))
3528 || (rs2 && !REGNO_OK_FOR_BASE_P (REGNO (rs2))))
3533 if ((! SPARC_INT_REG_P (REGNO (rs1))
3534 && REGNO (rs1) != FRAME_POINTER_REGNUM
3535 && REGNO (rs1) < FIRST_PSEUDO_REGISTER)
3537 && (! SPARC_INT_REG_P (REGNO (rs2))
3538 && REGNO (rs2) != FRAME_POINTER_REGNUM
3539 && REGNO (rs2) < FIRST_PSEUDO_REGISTER)))
3545 /* Return the SYMBOL_REF for the tls_get_addr function. */
3547 static GTY(()) rtx sparc_tls_symbol = NULL_RTX;
3550 sparc_tls_get_addr (void)
3552 if (!sparc_tls_symbol)
3553 sparc_tls_symbol = gen_rtx_SYMBOL_REF (Pmode, "__tls_get_addr");
3555 return sparc_tls_symbol;
3558 /* Return the Global Offset Table to be used in TLS mode. */
3561 sparc_tls_got (void)
3563 /* In PIC mode, this is just the PIC offset table. */
3566 crtl->uses_pic_offset_table = 1;
3567 return pic_offset_table_rtx;
3570 /* In non-PIC mode, Sun as (unlike GNU as) emits PC-relative relocations for
3571 the GOT symbol with the 32-bit ABI, so we reload the GOT register. */
3572 if (TARGET_SUN_TLS && TARGET_ARCH32)
3574 load_got_register ();
3575 return global_offset_table_rtx;
3578 /* In all other cases, we load a new pseudo with the GOT symbol. */
3579 return copy_to_reg (sparc_got ());
3582 /* Return true if X contains a thread-local symbol. */
3585 sparc_tls_referenced_p (rtx x)
3587 if (!TARGET_HAVE_TLS)
3590 if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS)
3591 x = XEXP (XEXP (x, 0), 0);
3593 if (GET_CODE (x) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (x))
3596 /* That's all we handle in sparc_legitimize_tls_address for now. */
3600 /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute
3601 this (thread-local) address. */
3604 sparc_legitimize_tls_address (rtx addr)
3606 rtx temp1, temp2, temp3, ret, o0, got, insn;
3608 gcc_assert (can_create_pseudo_p ());
3610 if (GET_CODE (addr) == SYMBOL_REF)
3611 switch (SYMBOL_REF_TLS_MODEL (addr))
3613 case TLS_MODEL_GLOBAL_DYNAMIC:
3615 temp1 = gen_reg_rtx (SImode);
3616 temp2 = gen_reg_rtx (SImode);
3617 ret = gen_reg_rtx (Pmode);
3618 o0 = gen_rtx_REG (Pmode, 8);
3619 got = sparc_tls_got ();
3620 emit_insn (gen_tgd_hi22 (temp1, addr));
3621 emit_insn (gen_tgd_lo10 (temp2, temp1, addr));
3624 emit_insn (gen_tgd_add32 (o0, got, temp2, addr));
3625 insn = emit_call_insn (gen_tgd_call32 (o0, sparc_tls_get_addr (),
3630 emit_insn (gen_tgd_add64 (o0, got, temp2, addr));
3631 insn = emit_call_insn (gen_tgd_call64 (o0, sparc_tls_get_addr (),
3634 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), o0);
3635 insn = get_insns ();
3637 emit_libcall_block (insn, ret, o0, addr);
3640 case TLS_MODEL_LOCAL_DYNAMIC:
3642 temp1 = gen_reg_rtx (SImode);
3643 temp2 = gen_reg_rtx (SImode);
3644 temp3 = gen_reg_rtx (Pmode);
3645 ret = gen_reg_rtx (Pmode);
3646 o0 = gen_rtx_REG (Pmode, 8);
3647 got = sparc_tls_got ();
3648 emit_insn (gen_tldm_hi22 (temp1));
3649 emit_insn (gen_tldm_lo10 (temp2, temp1));
3652 emit_insn (gen_tldm_add32 (o0, got, temp2));
3653 insn = emit_call_insn (gen_tldm_call32 (o0, sparc_tls_get_addr (),
3658 emit_insn (gen_tldm_add64 (o0, got, temp2));
3659 insn = emit_call_insn (gen_tldm_call64 (o0, sparc_tls_get_addr (),
3662 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), o0);
3663 insn = get_insns ();
3665 emit_libcall_block (insn, temp3, o0,
3666 gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
3667 UNSPEC_TLSLD_BASE));
3668 temp1 = gen_reg_rtx (SImode);
3669 temp2 = gen_reg_rtx (SImode);
3670 emit_insn (gen_tldo_hix22 (temp1, addr));
3671 emit_insn (gen_tldo_lox10 (temp2, temp1, addr));
3673 emit_insn (gen_tldo_add32 (ret, temp3, temp2, addr));
3675 emit_insn (gen_tldo_add64 (ret, temp3, temp2, addr));
3678 case TLS_MODEL_INITIAL_EXEC:
3679 temp1 = gen_reg_rtx (SImode);
3680 temp2 = gen_reg_rtx (SImode);
3681 temp3 = gen_reg_rtx (Pmode);
3682 got = sparc_tls_got ();
3683 emit_insn (gen_tie_hi22 (temp1, addr));
3684 emit_insn (gen_tie_lo10 (temp2, temp1, addr));
3686 emit_insn (gen_tie_ld32 (temp3, got, temp2, addr));
3688 emit_insn (gen_tie_ld64 (temp3, got, temp2, addr));
3691 ret = gen_reg_rtx (Pmode);
3693 emit_insn (gen_tie_add32 (ret, gen_rtx_REG (Pmode, 7),
3696 emit_insn (gen_tie_add64 (ret, gen_rtx_REG (Pmode, 7),
3700 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp3);
3703 case TLS_MODEL_LOCAL_EXEC:
3704 temp1 = gen_reg_rtx (Pmode);
3705 temp2 = gen_reg_rtx (Pmode);
3708 emit_insn (gen_tle_hix22_sp32 (temp1, addr));
3709 emit_insn (gen_tle_lox10_sp32 (temp2, temp1, addr));
3713 emit_insn (gen_tle_hix22_sp64 (temp1, addr));
3714 emit_insn (gen_tle_lox10_sp64 (temp2, temp1, addr));
3716 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp2);
3723 else if (GET_CODE (addr) == CONST)
3727 gcc_assert (GET_CODE (XEXP (addr, 0)) == PLUS);
3729 base = sparc_legitimize_tls_address (XEXP (XEXP (addr, 0), 0));
3730 offset = XEXP (XEXP (addr, 0), 1);
3732 base = force_operand (base, NULL_RTX);
3733 if (!(GET_CODE (offset) == CONST_INT && SMALL_INT (offset)))
3734 offset = force_reg (Pmode, offset);
3735 ret = gen_rtx_PLUS (Pmode, base, offset);
3739 gcc_unreachable (); /* for now ... */
3744 /* Legitimize PIC addresses. If the address is already position-independent,
3745 we return ORIG. Newly generated position-independent addresses go into a
3746 reg. This is REG if nonzero, otherwise we allocate register(s) as
3750 sparc_legitimize_pic_address (rtx orig, rtx reg)
3752 bool gotdata_op = false;
3754 if (GET_CODE (orig) == SYMBOL_REF
3755 /* See the comment in sparc_expand_move. */
3756 || (GET_CODE (orig) == LABEL_REF && !can_use_mov_pic_label_ref (orig)))
3758 rtx pic_ref, address;
3763 gcc_assert (can_create_pseudo_p ());
3764 reg = gen_reg_rtx (Pmode);
3769 /* If not during reload, allocate another temp reg here for loading
3770 in the address, so that these instructions can be optimized
3772 rtx temp_reg = (! can_create_pseudo_p ()
3773 ? reg : gen_reg_rtx (Pmode));
3775 /* Must put the SYMBOL_REF inside an UNSPEC here so that cse
3776 won't get confused into thinking that these two instructions
3777 are loading in the true address of the symbol. If in the
3778 future a PIC rtx exists, that should be used instead. */
3781 emit_insn (gen_movdi_high_pic (temp_reg, orig));
3782 emit_insn (gen_movdi_lo_sum_pic (temp_reg, temp_reg, orig));
3786 emit_insn (gen_movsi_high_pic (temp_reg, orig));
3787 emit_insn (gen_movsi_lo_sum_pic (temp_reg, temp_reg, orig));
3795 crtl->uses_pic_offset_table = 1;
3799 insn = emit_insn (gen_movdi_pic_gotdata_op (reg,
3800 pic_offset_table_rtx,
3803 insn = emit_insn (gen_movsi_pic_gotdata_op (reg,
3804 pic_offset_table_rtx,
3810 = gen_const_mem (Pmode,
3811 gen_rtx_PLUS (Pmode,
3812 pic_offset_table_rtx, address));
3813 insn = emit_move_insn (reg, pic_ref);
3816 /* Put a REG_EQUAL note on this insn, so that it can be optimized
3818 set_unique_reg_note (insn, REG_EQUAL, orig);
3821 else if (GET_CODE (orig) == CONST)
3825 if (GET_CODE (XEXP (orig, 0)) == PLUS
3826 && XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx)
3831 gcc_assert (can_create_pseudo_p ());
3832 reg = gen_reg_rtx (Pmode);
3835 gcc_assert (GET_CODE (XEXP (orig, 0)) == PLUS);
3836 base = sparc_legitimize_pic_address (XEXP (XEXP (orig, 0), 0), reg);
3837 offset = sparc_legitimize_pic_address (XEXP (XEXP (orig, 0), 1),
3838 base == reg ? NULL_RTX : reg);
3840 if (GET_CODE (offset) == CONST_INT)
3842 if (SMALL_INT (offset))
3843 return plus_constant (base, INTVAL (offset));
3844 else if (can_create_pseudo_p ())
3845 offset = force_reg (Pmode, offset);
3847 /* If we reach here, then something is seriously wrong. */
3850 return gen_rtx_PLUS (Pmode, base, offset);
3852 else if (GET_CODE (orig) == LABEL_REF)
3853 /* ??? We ought to be checking that the register is live instead, in case
3854 it is eliminated. */
3855 crtl->uses_pic_offset_table = 1;
3860 /* Try machine-dependent ways of modifying an illegitimate address X
3861 to be legitimate. If we find one, return the new, valid address.
3863 OLDX is the address as it was before break_out_memory_refs was called.
3864 In some cases it is useful to look at this to decide what needs to be done.
3866 MODE is the mode of the operand pointed to by X.
3868 On SPARC, change REG+N into REG+REG, and REG+(X*Y) into REG+REG. */
3871 sparc_legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED,
3872 enum machine_mode mode)
3876 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == MULT)
3877 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3878 force_operand (XEXP (x, 0), NULL_RTX));
3879 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == MULT)
3880 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3881 force_operand (XEXP (x, 1), NULL_RTX));
3882 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == PLUS)
3883 x = gen_rtx_PLUS (Pmode, force_operand (XEXP (x, 0), NULL_RTX),
3885 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == PLUS)
3886 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3887 force_operand (XEXP (x, 1), NULL_RTX));
3889 if (x != orig_x && sparc_legitimate_address_p (mode, x, FALSE))
3892 if (sparc_tls_referenced_p (x))
3893 x = sparc_legitimize_tls_address (x);
3895 x = sparc_legitimize_pic_address (x, NULL_RTX);
3896 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 1)))
3897 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3898 copy_to_mode_reg (Pmode, XEXP (x, 1)));
3899 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 0)))
3900 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3901 copy_to_mode_reg (Pmode, XEXP (x, 0)));
3902 else if (GET_CODE (x) == SYMBOL_REF
3903 || GET_CODE (x) == CONST
3904 || GET_CODE (x) == LABEL_REF)
3905 x = copy_to_suggested_reg (x, NULL_RTX, Pmode);
3910 /* Delegitimize an address that was legitimized by the above function. */
3913 sparc_delegitimize_address (rtx x)
3915 x = delegitimize_mem_from_attrs (x);
3917 if (GET_CODE (x) == LO_SUM && GET_CODE (XEXP (x, 1)) == UNSPEC)
3918 switch (XINT (XEXP (x, 1), 1))
3920 case UNSPEC_MOVE_PIC:
3922 x = XVECEXP (XEXP (x, 1), 0, 0);
3923 gcc_assert (GET_CODE (x) == SYMBOL_REF);
3929 /* This is generated by mov{si,di}_pic_label_ref in PIC mode. */
3930 if (GET_CODE (x) == MINUS
3931 && REG_P (XEXP (x, 0))
3932 && REGNO (XEXP (x, 0)) == PIC_OFFSET_TABLE_REGNUM
3933 && GET_CODE (XEXP (x, 1)) == LO_SUM
3934 && GET_CODE (XEXP (XEXP (x, 1), 1)) == UNSPEC
3935 && XINT (XEXP (XEXP (x, 1), 1), 1) == UNSPEC_MOVE_PIC_LABEL)
3937 x = XVECEXP (XEXP (XEXP (x, 1), 1), 0, 0);
3938 gcc_assert (GET_CODE (x) == LABEL_REF);
3944 /* SPARC implementation of LEGITIMIZE_RELOAD_ADDRESS. Returns a value to
3945 replace the input X, or the original X if no replacement is called for.
3946 The output parameter *WIN is 1 if the calling macro should goto WIN,
3949 For SPARC, we wish to handle addresses by splitting them into
3950 HIGH+LO_SUM pairs, retaining the LO_SUM in the memory reference.
3951 This cuts the number of extra insns by one.
3953 Do nothing when generating PIC code and the address is a symbolic
3954 operand or requires a scratch register. */
3957 sparc_legitimize_reload_address (rtx x, enum machine_mode mode,
3958 int opnum, int type,
3959 int ind_levels ATTRIBUTE_UNUSED, int *win)
3961 /* Decompose SImode constants into HIGH+LO_SUM. */
3963 && (mode != TFmode || TARGET_ARCH64)
3964 && GET_MODE (x) == SImode
3965 && GET_CODE (x) != LO_SUM
3966 && GET_CODE (x) != HIGH
3967 && sparc_cmodel <= CM_MEDLOW
3969 && (symbolic_operand (x, Pmode) || pic_address_needs_scratch (x))))
3971 x = gen_rtx_LO_SUM (GET_MODE (x), gen_rtx_HIGH (GET_MODE (x), x), x);
3972 push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL,
3973 BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0,
3974 opnum, (enum reload_type)type);
3979 /* We have to recognize what we have already generated above. */
3980 if (GET_CODE (x) == LO_SUM && GET_CODE (XEXP (x, 0)) == HIGH)
3982 push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL,
3983 BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0,
3984 opnum, (enum reload_type)type);
3993 /* Return true if ADDR (a legitimate address expression)
3994 has an effect that depends on the machine mode it is used for.
4000 is not equivalent to
4002 (mem:QI [%l7+a]) (mem:QI [%l7+a+1])
4004 because [%l7+a+1] is interpreted as the address of (a+1). */
4008 sparc_mode_dependent_address_p (const_rtx addr)
4010 if (flag_pic && GET_CODE (addr) == PLUS)
4012 rtx op0 = XEXP (addr, 0);
4013 rtx op1 = XEXP (addr, 1);
4014 if (op0 == pic_offset_table_rtx
4015 && symbolic_operand (op1, VOIDmode))
4022 #ifdef HAVE_GAS_HIDDEN
4023 # define USE_HIDDEN_LINKONCE 1
4025 # define USE_HIDDEN_LINKONCE 0
4029 get_pc_thunk_name (char name[32], unsigned int regno)
4031 const char *reg_name = reg_names[regno];
4033 /* Skip the leading '%' as that cannot be used in a
4037 if (USE_HIDDEN_LINKONCE)
4038 sprintf (name, "__sparc_get_pc_thunk.%s", reg_name);
4040 ASM_GENERATE_INTERNAL_LABEL (name, "LADDPC", regno);
4043 /* Wrapper around the load_pcrel_sym{si,di} patterns. */
4046 gen_load_pcrel_sym (rtx op0, rtx op1, rtx op2, rtx op3)
4048 int orig_flag_pic = flag_pic;
4051 /* The load_pcrel_sym{si,di} patterns require absolute addressing. */
4054 insn = gen_load_pcrel_symdi (op0, op1, op2, op3);
4056 insn = gen_load_pcrel_symsi (op0, op1, op2, op3);
4057 flag_pic = orig_flag_pic;
4062 /* Emit code to load the GOT register. */
4065 load_got_register (void)
4067 /* In PIC mode, this will retrieve pic_offset_table_rtx. */
4068 if (!global_offset_table_rtx)
4069 global_offset_table_rtx = gen_rtx_REG (Pmode, GLOBAL_OFFSET_TABLE_REGNUM);
4071 if (TARGET_VXWORKS_RTP)
4072 emit_insn (gen_vxworks_load_got ());
4075 /* The GOT symbol is subject to a PC-relative relocation so we need a
4076 helper function to add the PC value and thus get the final value. */
4077 if (!got_helper_rtx)
4080 get_pc_thunk_name (name, GLOBAL_OFFSET_TABLE_REGNUM);
4081 got_helper_rtx = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (name));
4084 emit_insn (gen_load_pcrel_sym (global_offset_table_rtx, sparc_got (),
4086 GEN_INT (GLOBAL_OFFSET_TABLE_REGNUM)));
4089 /* Need to emit this whether or not we obey regdecls,
4090 since setjmp/longjmp can cause life info to screw up.
4091 ??? In the case where we don't obey regdecls, this is not sufficient
4092 since we may not fall out the bottom. */
4093 emit_use (global_offset_table_rtx);
4096 /* Emit a call instruction with the pattern given by PAT. ADDR is the
4097 address of the call target. */
4100 sparc_emit_call_insn (rtx pat, rtx addr)
4104 insn = emit_call_insn (pat);
4106 /* The PIC register is live on entry to VxWorks PIC PLT entries. */
4107 if (TARGET_VXWORKS_RTP
4109 && GET_CODE (addr) == SYMBOL_REF
4110 && (SYMBOL_REF_DECL (addr)
4111 ? !targetm.binds_local_p (SYMBOL_REF_DECL (addr))
4112 : !SYMBOL_REF_LOCAL_P (addr)))
4114 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), pic_offset_table_rtx);
4115 crtl->uses_pic_offset_table = 1;
4119 /* Return 1 if RTX is a MEM which is known to be aligned to at
4120 least a DESIRED byte boundary. */
4123 mem_min_alignment (rtx mem, int desired)
4125 rtx addr, base, offset;
4127 /* If it's not a MEM we can't accept it. */
4128 if (GET_CODE (mem) != MEM)
4132 if (!TARGET_UNALIGNED_DOUBLES
4133 && MEM_ALIGN (mem) / BITS_PER_UNIT >= (unsigned)desired)
4136 /* ??? The rest of the function predates MEM_ALIGN so
4137 there is probably a bit of redundancy. */
4138 addr = XEXP (mem, 0);
4139 base = offset = NULL_RTX;
4140 if (GET_CODE (addr) == PLUS)
4142 if (GET_CODE (XEXP (addr, 0)) == REG)
4144 base = XEXP (addr, 0);
4146 /* What we are saying here is that if the base
4147 REG is aligned properly, the compiler will make
4148 sure any REG based index upon it will be so
4150 if (GET_CODE (XEXP (addr, 1)) == CONST_INT)
4151 offset = XEXP (addr, 1);
4153 offset = const0_rtx;
4156 else if (GET_CODE (addr) == REG)
4159 offset = const0_rtx;
4162 if (base != NULL_RTX)
4164 int regno = REGNO (base);
4166 if (regno != HARD_FRAME_POINTER_REGNUM && regno != STACK_POINTER_REGNUM)
4168 /* Check if the compiler has recorded some information
4169 about the alignment of the base REG. If reload has
4170 completed, we already matched with proper alignments.
4171 If not running global_alloc, reload might give us
4172 unaligned pointer to local stack though. */
4174 && REGNO_POINTER_ALIGN (regno) >= desired * BITS_PER_UNIT)
4175 || (optimize && reload_completed))
4176 && (INTVAL (offset) & (desired - 1)) == 0)
4181 if (((INTVAL (offset) - SPARC_STACK_BIAS) & (desired - 1)) == 0)
4185 else if (! TARGET_UNALIGNED_DOUBLES
4186 || CONSTANT_P (addr)
4187 || GET_CODE (addr) == LO_SUM)
4189 /* Anything else we know is properly aligned unless TARGET_UNALIGNED_DOUBLES
4190 is true, in which case we can only assume that an access is aligned if
4191 it is to a constant address, or the address involves a LO_SUM. */
4195 /* An obviously unaligned address. */
4200 /* Vectors to keep interesting information about registers where it can easily
4201 be got. We used to use the actual mode value as the bit number, but there
4202 are more than 32 modes now. Instead we use two tables: one indexed by
4203 hard register number, and one indexed by mode. */
4205 /* The purpose of sparc_mode_class is to shrink the range of modes so that
4206 they all fit (as bit numbers) in a 32-bit word (again). Each real mode is
4207 mapped into one sparc_mode_class mode. */
4209 enum sparc_mode_class {
4210 S_MODE, D_MODE, T_MODE, O_MODE,
4211 SF_MODE, DF_MODE, TF_MODE, OF_MODE,
4215 /* Modes for single-word and smaller quantities. */
4216 #define S_MODES ((1 << (int) S_MODE) | (1 << (int) SF_MODE))
4218 /* Modes for double-word and smaller quantities. */
4219 #define D_MODES (S_MODES | (1 << (int) D_MODE) | (1 << DF_MODE))
4221 /* Modes for quad-word and smaller quantities. */
4222 #define T_MODES (D_MODES | (1 << (int) T_MODE) | (1 << (int) TF_MODE))
4224 /* Modes for 8-word and smaller quantities. */
4225 #define O_MODES (T_MODES | (1 << (int) O_MODE) | (1 << (int) OF_MODE))
4227 /* Modes for single-float quantities. We must allow any single word or
4228 smaller quantity. This is because the fix/float conversion instructions
4229 take integer inputs/outputs from the float registers. */
4230 #define SF_MODES (S_MODES)
4232 /* Modes for double-float and smaller quantities. */
4233 #define DF_MODES (D_MODES)
4235 /* Modes for quad-float and smaller quantities. */
4236 #define TF_MODES (DF_MODES | (1 << (int) TF_MODE))
4238 /* Modes for quad-float pairs and smaller quantities. */
4239 #define OF_MODES (TF_MODES | (1 << (int) OF_MODE))
4241 /* Modes for double-float only quantities. */
4242 #define DF_MODES_NO_S ((1 << (int) D_MODE) | (1 << (int) DF_MODE))
4244 /* Modes for quad-float and double-float only quantities. */
4245 #define TF_MODES_NO_S (DF_MODES_NO_S | (1 << (int) TF_MODE))
4247 /* Modes for quad-float pairs and double-float only quantities. */
4248 #define OF_MODES_NO_S (TF_MODES_NO_S | (1 << (int) OF_MODE))
4250 /* Modes for condition codes. */
4251 #define CC_MODES (1 << (int) CC_MODE)
4252 #define CCFP_MODES (1 << (int) CCFP_MODE)
4254 /* Value is 1 if register/mode pair is acceptable on sparc.
4255 The funny mixture of D and T modes is because integer operations
4256 do not specially operate on tetra quantities, so non-quad-aligned
4257 registers can hold quadword quantities (except %o4 and %i4 because
4258 they cross fixed registers). */
4260 /* This points to either the 32 bit or the 64 bit version. */
4261 const int *hard_regno_mode_classes;
4263 static const int hard_32bit_mode_classes[] = {
4264 S_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
4265 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
4266 T_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
4267 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
4269 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4270 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4271 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4272 OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
4274 /* FP regs f32 to f63. Only the even numbered registers actually exist,
4275 and none can hold SFmode/SImode values. */
4276 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4277 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4278 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4279 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4282 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
4284 /* %icc, %sfp, %gsr */
4285 CC_MODES, 0, D_MODES
4288 static const int hard_64bit_mode_classes[] = {
4289 D_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4290 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4291 T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4292 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4294 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_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, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
4299 /* FP regs f32 to f63. Only the even numbered registers actually exist,
4300 and none can hold SFmode/SImode values. */
4301 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
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, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4307 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
4309 /* %icc, %sfp, %gsr */
4310 CC_MODES, 0, D_MODES
4313 int sparc_mode_class [NUM_MACHINE_MODES];
4315 enum reg_class sparc_regno_reg_class[FIRST_PSEUDO_REGISTER];
4318 sparc_init_modes (void)
4322 for (i = 0; i < NUM_MACHINE_MODES; i++)
4324 switch (GET_MODE_CLASS (i))
4327 case MODE_PARTIAL_INT:
4328 case MODE_COMPLEX_INT:
4329 if (GET_MODE_SIZE (i) <= 4)
4330 sparc_mode_class[i] = 1 << (int) S_MODE;
4331 else if (GET_MODE_SIZE (i) == 8)
4332 sparc_mode_class[i] = 1 << (int) D_MODE;
4333 else if (GET_MODE_SIZE (i) == 16)
4334 sparc_mode_class[i] = 1 << (int) T_MODE;
4335 else if (GET_MODE_SIZE (i) == 32)
4336 sparc_mode_class[i] = 1 << (int) O_MODE;
4338 sparc_mode_class[i] = 0;
4340 case MODE_VECTOR_INT:
4341 if (GET_MODE_SIZE (i) <= 4)
4342 sparc_mode_class[i] = 1 << (int)SF_MODE;
4343 else if (GET_MODE_SIZE (i) == 8)
4344 sparc_mode_class[i] = 1 << (int)DF_MODE;
4347 case MODE_COMPLEX_FLOAT:
4348 if (GET_MODE_SIZE (i) <= 4)
4349 sparc_mode_class[i] = 1 << (int) SF_MODE;
4350 else if (GET_MODE_SIZE (i) == 8)
4351 sparc_mode_class[i] = 1 << (int) DF_MODE;
4352 else if (GET_MODE_SIZE (i) == 16)
4353 sparc_mode_class[i] = 1 << (int) TF_MODE;
4354 else if (GET_MODE_SIZE (i) == 32)
4355 sparc_mode_class[i] = 1 << (int) OF_MODE;
4357 sparc_mode_class[i] = 0;
4360 if (i == (int) CCFPmode || i == (int) CCFPEmode)
4361 sparc_mode_class[i] = 1 << (int) CCFP_MODE;
4363 sparc_mode_class[i] = 1 << (int) CC_MODE;
4366 sparc_mode_class[i] = 0;
4372 hard_regno_mode_classes = hard_64bit_mode_classes;
4374 hard_regno_mode_classes = hard_32bit_mode_classes;
4376 /* Initialize the array used by REGNO_REG_CLASS. */
4377 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
4379 if (i < 16 && TARGET_V8PLUS)
4380 sparc_regno_reg_class[i] = I64_REGS;
4381 else if (i < 32 || i == FRAME_POINTER_REGNUM)
4382 sparc_regno_reg_class[i] = GENERAL_REGS;
4384 sparc_regno_reg_class[i] = FP_REGS;
4386 sparc_regno_reg_class[i] = EXTRA_FP_REGS;
4388 sparc_regno_reg_class[i] = FPCC_REGS;
4390 sparc_regno_reg_class[i] = NO_REGS;
4394 /* Return whether REGNO, a global or FP register, must be saved/restored. */
4397 save_global_or_fp_reg_p (unsigned int regno,
4398 int leaf_function ATTRIBUTE_UNUSED)
4400 return !call_used_regs[regno] && df_regs_ever_live_p (regno);
4403 /* Return whether the return address register (%i7) is needed. */
4406 return_addr_reg_needed_p (int leaf_function)
4408 /* If it is live, for example because of __builtin_return_address (0). */
4409 if (df_regs_ever_live_p (RETURN_ADDR_REGNUM))
4412 /* Otherwise, it is needed as save register if %o7 is clobbered. */
4414 /* Loading the GOT register clobbers %o7. */
4415 || crtl->uses_pic_offset_table
4416 || df_regs_ever_live_p (INCOMING_RETURN_ADDR_REGNUM))
4422 /* Return whether REGNO, a local or in register, must be saved/restored. */
4425 save_local_or_in_reg_p (unsigned int regno, int leaf_function)
4427 /* General case: call-saved registers live at some point. */
4428 if (!call_used_regs[regno] && df_regs_ever_live_p (regno))
4431 /* Frame pointer register (%fp) if needed. */
4432 if (regno == HARD_FRAME_POINTER_REGNUM && frame_pointer_needed)
4435 /* Return address register (%i7) if needed. */
4436 if (regno == RETURN_ADDR_REGNUM && return_addr_reg_needed_p (leaf_function))
4439 /* GOT register (%l7) if needed. */
4440 if (regno == PIC_OFFSET_TABLE_REGNUM && crtl->uses_pic_offset_table)
4443 /* If the function accesses prior frames, the frame pointer and the return
4444 address of the previous frame must be saved on the stack. */
4445 if (crtl->accesses_prior_frames
4446 && (regno == HARD_FRAME_POINTER_REGNUM || regno == RETURN_ADDR_REGNUM))
4452 /* Compute the frame size required by the function. This function is called
4453 during the reload pass and also by sparc_expand_prologue. */
4456 sparc_compute_frame_size (HOST_WIDE_INT size, int leaf_function)
4458 HOST_WIDE_INT frame_size, apparent_frame_size;
4459 int args_size, n_global_fp_regs = 0;
4460 bool save_local_in_regs_p = false;
4463 /* If the function allocates dynamic stack space, the dynamic offset is
4464 computed early and contains REG_PARM_STACK_SPACE, so we need to cope. */
4465 if (leaf_function && !cfun->calls_alloca)
4468 args_size = crtl->outgoing_args_size + REG_PARM_STACK_SPACE (cfun->decl);
4470 /* Calculate space needed for global registers. */
4472 for (i = 0; i < 8; i++)
4473 if (save_global_or_fp_reg_p (i, 0))
4474 n_global_fp_regs += 2;
4476 for (i = 0; i < 8; i += 2)
4477 if (save_global_or_fp_reg_p (i, 0) || save_global_or_fp_reg_p (i + 1, 0))
4478 n_global_fp_regs += 2;
4480 /* In the flat window model, find out which local and in registers need to
4481 be saved. We don't reserve space in the current frame for them as they
4482 will be spilled into the register window save area of the caller's frame.
4483 However, as soon as we use this register window save area, we must create
4484 that of the current frame to make it the live one. */
4486 for (i = 16; i < 32; i++)
4487 if (save_local_or_in_reg_p (i, leaf_function))
4489 save_local_in_regs_p = true;
4493 /* Calculate space needed for FP registers. */
4494 for (i = 32; i < (TARGET_V9 ? 96 : 64); i += 2)
4495 if (save_global_or_fp_reg_p (i, 0) || save_global_or_fp_reg_p (i + 1, 0))
4496 n_global_fp_regs += 2;
4499 && n_global_fp_regs == 0
4501 && !save_local_in_regs_p)
4502 frame_size = apparent_frame_size = 0;
4505 /* We subtract STARTING_FRAME_OFFSET, remember it's negative. */
4506 apparent_frame_size = (size - STARTING_FRAME_OFFSET + 7) & -8;
4507 apparent_frame_size += n_global_fp_regs * 4;
4509 /* We need to add the size of the outgoing argument area. */
4510 frame_size = apparent_frame_size + ((args_size + 7) & -8);
4512 /* And that of the register window save area. */
4513 frame_size += FIRST_PARM_OFFSET (cfun->decl);
4515 /* Finally, bump to the appropriate alignment. */
4516 frame_size = SPARC_STACK_ALIGN (frame_size);
4519 /* Set up values for use in prologue and epilogue. */
4520 sparc_frame_size = frame_size;
4521 sparc_apparent_frame_size = apparent_frame_size;
4522 sparc_n_global_fp_regs = n_global_fp_regs;
4523 sparc_save_local_in_regs_p = save_local_in_regs_p;
4528 /* Output any necessary .register pseudo-ops. */
4531 sparc_output_scratch_registers (FILE *file ATTRIBUTE_UNUSED)
4533 #ifdef HAVE_AS_REGISTER_PSEUDO_OP
4539 /* Check if %g[2367] were used without
4540 .register being printed for them already. */
4541 for (i = 2; i < 8; i++)
4543 if (df_regs_ever_live_p (i)
4544 && ! sparc_hard_reg_printed [i])
4546 sparc_hard_reg_printed [i] = 1;
4547 /* %g7 is used as TLS base register, use #ignore
4548 for it instead of #scratch. */
4549 fprintf (file, "\t.register\t%%g%d, #%s\n", i,
4550 i == 7 ? "ignore" : "scratch");
4557 #define PROBE_INTERVAL (1 << STACK_CHECK_PROBE_INTERVAL_EXP)
4559 #if PROBE_INTERVAL > 4096
4560 #error Cannot use indexed addressing mode for stack probing
4563 /* Emit code to probe a range of stack addresses from FIRST to FIRST+SIZE,
4564 inclusive. These are offsets from the current stack pointer.
4566 Note that we don't use the REG+REG addressing mode for the probes because
4567 of the stack bias in 64-bit mode. And it doesn't really buy us anything
4568 so the advantages of having a single code win here. */
4571 sparc_emit_probe_stack_range (HOST_WIDE_INT first, HOST_WIDE_INT size)
4573 rtx g1 = gen_rtx_REG (Pmode, 1);
4575 /* See if we have a constant small number of probes to generate. If so,
4576 that's the easy case. */
4577 if (size <= PROBE_INTERVAL)
4579 emit_move_insn (g1, GEN_INT (first));
4580 emit_insn (gen_rtx_SET (VOIDmode, g1,
4581 gen_rtx_MINUS (Pmode, stack_pointer_rtx, g1)));
4582 emit_stack_probe (plus_constant (g1, -size));
4585 /* The run-time loop is made up of 10 insns in the generic case while the
4586 compile-time loop is made up of 4+2*(n-2) insns for n # of intervals. */
4587 else if (size <= 5 * PROBE_INTERVAL)
4591 emit_move_insn (g1, GEN_INT (first + PROBE_INTERVAL));
4592 emit_insn (gen_rtx_SET (VOIDmode, g1,
4593 gen_rtx_MINUS (Pmode, stack_pointer_rtx, g1)));
4594 emit_stack_probe (g1);
4596 /* Probe at FIRST + N * PROBE_INTERVAL for values of N from 2 until
4597 it exceeds SIZE. If only two probes are needed, this will not
4598 generate any code. Then probe at FIRST + SIZE. */
4599 for (i = 2 * PROBE_INTERVAL; i < size; i += PROBE_INTERVAL)
4601 emit_insn (gen_rtx_SET (VOIDmode, g1,
4602 plus_constant (g1, -PROBE_INTERVAL)));
4603 emit_stack_probe (g1);
4606 emit_stack_probe (plus_constant (g1, (i - PROBE_INTERVAL) - size));
4609 /* Otherwise, do the same as above, but in a loop. Note that we must be
4610 extra careful with variables wrapping around because we might be at
4611 the very top (or the very bottom) of the address space and we have
4612 to be able to handle this case properly; in particular, we use an
4613 equality test for the loop condition. */
4616 HOST_WIDE_INT rounded_size;
4617 rtx g4 = gen_rtx_REG (Pmode, 4);
4619 emit_move_insn (g1, GEN_INT (first));
4622 /* Step 1: round SIZE to the previous multiple of the interval. */
4624 rounded_size = size & -PROBE_INTERVAL;
4625 emit_move_insn (g4, GEN_INT (rounded_size));
4628 /* Step 2: compute initial and final value of the loop counter. */
4630 /* TEST_ADDR = SP + FIRST. */
4631 emit_insn (gen_rtx_SET (VOIDmode, g1,
4632 gen_rtx_MINUS (Pmode, stack_pointer_rtx, g1)));
4634 /* LAST_ADDR = SP + FIRST + ROUNDED_SIZE. */
4635 emit_insn (gen_rtx_SET (VOIDmode, g4, gen_rtx_MINUS (Pmode, g1, g4)));
4640 while (TEST_ADDR != LAST_ADDR)
4642 TEST_ADDR = TEST_ADDR + PROBE_INTERVAL
4646 probes at FIRST + N * PROBE_INTERVAL for values of N from 1
4647 until it is equal to ROUNDED_SIZE. */
4650 emit_insn (gen_probe_stack_rangedi (g1, g1, g4));
4652 emit_insn (gen_probe_stack_rangesi (g1, g1, g4));
4655 /* Step 4: probe at FIRST + SIZE if we cannot assert at compile-time
4656 that SIZE is equal to ROUNDED_SIZE. */
4658 if (size != rounded_size)
4659 emit_stack_probe (plus_constant (g4, rounded_size - size));
4662 /* Make sure nothing is scheduled before we are done. */
4663 emit_insn (gen_blockage ());
4666 /* Probe a range of stack addresses from REG1 to REG2 inclusive. These are
4667 absolute addresses. */
4670 output_probe_stack_range (rtx reg1, rtx reg2)
4672 static int labelno = 0;
4673 char loop_lab[32], end_lab[32];
4676 ASM_GENERATE_INTERNAL_LABEL (loop_lab, "LPSRL", labelno);
4677 ASM_GENERATE_INTERNAL_LABEL (end_lab, "LPSRE", labelno++);
4679 ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, loop_lab);
4681 /* Jump to END_LAB if TEST_ADDR == LAST_ADDR. */
4684 output_asm_insn ("cmp\t%0, %1", xops);
4686 fputs ("\tbe,pn\t%xcc,", asm_out_file);
4688 fputs ("\tbe\t", asm_out_file);
4689 assemble_name_raw (asm_out_file, end_lab);
4690 fputc ('\n', asm_out_file);
4692 /* TEST_ADDR = TEST_ADDR + PROBE_INTERVAL. */
4693 xops[1] = GEN_INT (-PROBE_INTERVAL);
4694 output_asm_insn (" add\t%0, %1, %0", xops);
4696 /* Probe at TEST_ADDR and branch. */
4698 fputs ("\tba,pt\t%xcc,", asm_out_file);
4700 fputs ("\tba\t", asm_out_file);
4701 assemble_name_raw (asm_out_file, loop_lab);
4702 fputc ('\n', asm_out_file);
4703 xops[1] = GEN_INT (SPARC_STACK_BIAS);
4704 output_asm_insn (" st\t%%g0, [%0+%1]", xops);
4706 ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, end_lab);
4711 /* Emit code to save/restore registers from LOW to HIGH at BASE+OFFSET as
4712 needed. LOW is supposed to be double-word aligned for 32-bit registers.
4713 SAVE_P decides whether a register must be saved/restored. ACTION_TRUE
4714 is the action to be performed if SAVE_P returns true and ACTION_FALSE
4715 the action to be performed if it returns false. Return the new offset. */
4717 typedef bool (*sorr_pred_t) (unsigned int, int);
4718 typedef enum { SORR_NONE, SORR_ADVANCE, SORR_SAVE, SORR_RESTORE } sorr_act_t;
4721 emit_save_or_restore_regs (unsigned int low, unsigned int high, rtx base,
4722 int offset, int leaf_function, sorr_pred_t save_p,
4723 sorr_act_t action_true, sorr_act_t action_false)
4728 if (TARGET_ARCH64 && high <= 32)
4732 for (i = low; i < high; i++)
4734 if (save_p (i, leaf_function))
4736 mem = gen_frame_mem (DImode, plus_constant (base, offset));
4737 if (action_true == SORR_SAVE)
4739 insn = emit_move_insn (mem, gen_rtx_REG (DImode, i));
4740 RTX_FRAME_RELATED_P (insn) = 1;
4742 else /* action_true == SORR_RESTORE */
4744 /* The frame pointer must be restored last since its old
4745 value may be used as base address for the frame. This
4746 is problematic in 64-bit mode only because of the lack
4747 of double-word load instruction. */
4748 if (i == HARD_FRAME_POINTER_REGNUM)
4751 emit_move_insn (gen_rtx_REG (DImode, i), mem);
4755 else if (action_false == SORR_ADVANCE)
4761 mem = gen_frame_mem (DImode, plus_constant (base, fp_offset));
4762 emit_move_insn (hard_frame_pointer_rtx, mem);
4767 for (i = low; i < high; i += 2)
4769 bool reg0 = save_p (i, leaf_function);
4770 bool reg1 = save_p (i + 1, leaf_function);
4771 enum machine_mode mode;
4776 mode = SPARC_INT_REG_P (i) ? DImode : DFmode;
4781 mode = SPARC_INT_REG_P (i) ? SImode : SFmode;
4786 mode = SPARC_INT_REG_P (i) ? SImode : SFmode;
4792 if (action_false == SORR_ADVANCE)
4797 mem = gen_frame_mem (mode, plus_constant (base, offset));
4798 if (action_true == SORR_SAVE)
4800 insn = emit_move_insn (mem, gen_rtx_REG (mode, regno));
4801 RTX_FRAME_RELATED_P (insn) = 1;
4805 mem = gen_frame_mem (SImode, plus_constant (base, offset));
4806 set1 = gen_rtx_SET (VOIDmode, mem,
4807 gen_rtx_REG (SImode, regno));
4808 RTX_FRAME_RELATED_P (set1) = 1;
4810 = gen_frame_mem (SImode, plus_constant (base, offset + 4));
4811 set2 = gen_rtx_SET (VOIDmode, mem,
4812 gen_rtx_REG (SImode, regno + 1));
4813 RTX_FRAME_RELATED_P (set2) = 1;
4814 add_reg_note (insn, REG_FRAME_RELATED_EXPR,
4815 gen_rtx_PARALLEL (VOIDmode,
4816 gen_rtvec (2, set1, set2)));
4819 else /* action_true == SORR_RESTORE */
4820 emit_move_insn (gen_rtx_REG (mode, regno), mem);
4822 /* Always preserve double-word alignment. */
4823 offset = (offset + 8) & -8;
4830 /* Emit code to adjust BASE to OFFSET. Return the new base. */
4833 emit_adjust_base_to_offset (rtx base, int offset)
4835 /* ??? This might be optimized a little as %g1 might already have a
4836 value close enough that a single add insn will do. */
4837 /* ??? Although, all of this is probably only a temporary fix because
4838 if %g1 can hold a function result, then sparc_expand_epilogue will
4839 lose (the result will be clobbered). */
4840 rtx new_base = gen_rtx_REG (Pmode, 1);
4841 emit_move_insn (new_base, GEN_INT (offset));
4842 emit_insn (gen_rtx_SET (VOIDmode,
4843 new_base, gen_rtx_PLUS (Pmode, base, new_base)));
4847 /* Emit code to save/restore call-saved global and FP registers. */
4850 emit_save_or_restore_global_fp_regs (rtx base, int offset, sorr_act_t action)
4852 if (offset < -4096 || offset + sparc_n_global_fp_regs * 4 > 4095)
4854 base = emit_adjust_base_to_offset (base, offset);
4859 = emit_save_or_restore_regs (0, 8, base, offset, 0,
4860 save_global_or_fp_reg_p, action, SORR_NONE);
4861 emit_save_or_restore_regs (32, TARGET_V9 ? 96 : 64, base, offset, 0,
4862 save_global_or_fp_reg_p, action, SORR_NONE);
4865 /* Emit code to save/restore call-saved local and in registers. */
4868 emit_save_or_restore_local_in_regs (rtx base, int offset, sorr_act_t action)
4870 if (offset < -4096 || offset + 16 * UNITS_PER_WORD > 4095)
4872 base = emit_adjust_base_to_offset (base, offset);
4876 emit_save_or_restore_regs (16, 32, base, offset, sparc_leaf_function_p,
4877 save_local_or_in_reg_p, action, SORR_ADVANCE);
4880 /* Emit a window_save insn. */
4883 emit_window_save (rtx increment)
4885 rtx insn = emit_insn (gen_window_save (increment));
4886 RTX_FRAME_RELATED_P (insn) = 1;
4888 /* The incoming return address (%o7) is saved in %i7. */
4889 add_reg_note (insn, REG_CFA_REGISTER,
4890 gen_rtx_SET (VOIDmode,
4891 gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM),
4893 INCOMING_RETURN_ADDR_REGNUM)));
4895 /* The window save event. */
4896 add_reg_note (insn, REG_CFA_WINDOW_SAVE, const0_rtx);
4898 /* The CFA is %fp, the hard frame pointer. */
4899 add_reg_note (insn, REG_CFA_DEF_CFA,
4900 plus_constant (hard_frame_pointer_rtx,
4901 INCOMING_FRAME_SP_OFFSET));
4906 /* Generate an increment for the stack pointer. */
4909 gen_stack_pointer_inc (rtx increment)
4911 return gen_rtx_SET (VOIDmode,
4913 gen_rtx_PLUS (Pmode,
4918 /* Generate a decrement for the stack pointer. */
4921 gen_stack_pointer_dec (rtx decrement)
4923 return gen_rtx_SET (VOIDmode,
4925 gen_rtx_MINUS (Pmode,
4930 /* Expand the function prologue. The prologue is responsible for reserving
4931 storage for the frame, saving the call-saved registers and loading the
4932 GOT register if needed. */
4935 sparc_expand_prologue (void)
4940 /* Compute a snapshot of current_function_uses_only_leaf_regs. Relying
4941 on the final value of the flag means deferring the prologue/epilogue
4942 expansion until just before the second scheduling pass, which is too
4943 late to emit multiple epilogues or return insns.
4945 Of course we are making the assumption that the value of the flag
4946 will not change between now and its final value. Of the three parts
4947 of the formula, only the last one can reasonably vary. Let's take a
4948 closer look, after assuming that the first two ones are set to true
4949 (otherwise the last value is effectively silenced).
4951 If only_leaf_regs_used returns false, the global predicate will also
4952 be false so the actual frame size calculated below will be positive.
4953 As a consequence, the save_register_window insn will be emitted in
4954 the instruction stream; now this insn explicitly references %fp
4955 which is not a leaf register so only_leaf_regs_used will always
4956 return false subsequently.
4958 If only_leaf_regs_used returns true, we hope that the subsequent
4959 optimization passes won't cause non-leaf registers to pop up. For
4960 example, the regrename pass has special provisions to not rename to
4961 non-leaf registers in a leaf function. */
4962 sparc_leaf_function_p
4963 = optimize > 0 && current_function_is_leaf && only_leaf_regs_used ();
4965 size = sparc_compute_frame_size (get_frame_size(), sparc_leaf_function_p);
4967 if (flag_stack_usage_info)
4968 current_function_static_stack_size = size;
4970 if (flag_stack_check == STATIC_BUILTIN_STACK_CHECK && size)
4971 sparc_emit_probe_stack_range (STACK_CHECK_PROTECT, size);
4975 else if (sparc_leaf_function_p)
4977 rtx size_int_rtx = GEN_INT (-size);
4980 insn = emit_insn (gen_stack_pointer_inc (size_int_rtx));
4981 else if (size <= 8192)
4983 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (-4096)));
4984 RTX_FRAME_RELATED_P (insn) = 1;
4986 /* %sp is still the CFA register. */
4988 = emit_insn (gen_stack_pointer_inc (GEN_INT (4096 - size)));
4992 rtx size_rtx = gen_rtx_REG (Pmode, 1);
4993 emit_move_insn (size_rtx, size_int_rtx);
4994 insn = emit_insn (gen_stack_pointer_inc (size_rtx));
4995 add_reg_note (insn, REG_FRAME_RELATED_EXPR,
4996 gen_stack_pointer_inc (size_int_rtx));
4999 RTX_FRAME_RELATED_P (insn) = 1;
5003 rtx size_int_rtx = GEN_INT (-size);
5006 emit_window_save (size_int_rtx);
5007 else if (size <= 8192)
5009 emit_window_save (GEN_INT (-4096));
5012 /* %sp is not the CFA register anymore. */
5013 emit_insn (gen_stack_pointer_inc (GEN_INT (4096 - size)));
5015 /* Make sure no %fp-based store is issued until after the frame is
5016 established. The offset between the frame pointer and the stack
5017 pointer is calculated relative to the value of the stack pointer
5018 at the end of the function prologue, and moving instructions that
5019 access the stack via the frame pointer between the instructions
5020 that decrement the stack pointer could result in accessing the
5021 register window save area, which is volatile. */
5022 emit_insn (gen_frame_blockage ());
5026 rtx size_rtx = gen_rtx_REG (Pmode, 1);
5027 emit_move_insn (size_rtx, size_int_rtx);
5028 emit_window_save (size_rtx);
5032 if (sparc_leaf_function_p)
5034 sparc_frame_base_reg = stack_pointer_rtx;
5035 sparc_frame_base_offset = size + SPARC_STACK_BIAS;
5039 sparc_frame_base_reg = hard_frame_pointer_rtx;
5040 sparc_frame_base_offset = SPARC_STACK_BIAS;
5043 if (sparc_n_global_fp_regs > 0)
5044 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg,
5045 sparc_frame_base_offset
5046 - sparc_apparent_frame_size,
5049 /* Load the GOT register if needed. */
5050 if (crtl->uses_pic_offset_table)
5051 load_got_register ();
5053 /* Advertise that the data calculated just above are now valid. */
5054 sparc_prologue_data_valid_p = true;
5057 /* Expand the function prologue. The prologue is responsible for reserving
5058 storage for the frame, saving the call-saved registers and loading the
5059 GOT register if needed. */
5062 sparc_flat_expand_prologue (void)
5067 sparc_leaf_function_p = optimize > 0 && current_function_is_leaf;
5069 size = sparc_compute_frame_size (get_frame_size(), sparc_leaf_function_p);
5071 if (flag_stack_usage_info)
5072 current_function_static_stack_size = size;
5074 if (flag_stack_check == STATIC_BUILTIN_STACK_CHECK && size)
5075 sparc_emit_probe_stack_range (STACK_CHECK_PROTECT, size);
5077 if (sparc_save_local_in_regs_p)
5078 emit_save_or_restore_local_in_regs (stack_pointer_rtx, SPARC_STACK_BIAS,
5085 rtx size_int_rtx, size_rtx;
5087 size_rtx = size_int_rtx = GEN_INT (-size);
5089 /* We establish the frame (i.e. decrement the stack pointer) first, even
5090 if we use a frame pointer, because we cannot clobber any call-saved
5091 registers, including the frame pointer, if we haven't created a new
5092 register save area, for the sake of compatibility with the ABI. */
5094 insn = emit_insn (gen_stack_pointer_inc (size_int_rtx));
5095 else if (size <= 8192 && !frame_pointer_needed)
5097 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (-4096)));
5098 RTX_FRAME_RELATED_P (insn) = 1;
5099 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (4096 - size)));
5103 size_rtx = gen_rtx_REG (Pmode, 1);
5104 emit_move_insn (size_rtx, size_int_rtx);
5105 insn = emit_insn (gen_stack_pointer_inc (size_rtx));
5106 add_reg_note (insn, REG_CFA_ADJUST_CFA,
5107 gen_stack_pointer_inc (size_int_rtx));
5109 RTX_FRAME_RELATED_P (insn) = 1;
5111 /* Ensure nothing is scheduled until after the frame is established. */
5112 emit_insn (gen_blockage ());
5114 if (frame_pointer_needed)
5116 insn = emit_insn (gen_rtx_SET (VOIDmode, hard_frame_pointer_rtx,
5117 gen_rtx_MINUS (Pmode,
5120 RTX_FRAME_RELATED_P (insn) = 1;
5122 add_reg_note (insn, REG_CFA_ADJUST_CFA,
5123 gen_rtx_SET (VOIDmode, hard_frame_pointer_rtx,
5124 plus_constant (stack_pointer_rtx,
5128 if (return_addr_reg_needed_p (sparc_leaf_function_p))
5130 rtx o7 = gen_rtx_REG (Pmode, INCOMING_RETURN_ADDR_REGNUM);
5131 rtx i7 = gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM);
5133 insn = emit_move_insn (i7, o7);
5134 RTX_FRAME_RELATED_P (insn) = 1;
5136 add_reg_note (insn, REG_CFA_REGISTER,
5137 gen_rtx_SET (VOIDmode, i7, o7));
5139 /* Prevent this instruction from ever being considered dead,
5140 even if this function has no epilogue. */
5145 if (frame_pointer_needed)
5147 sparc_frame_base_reg = hard_frame_pointer_rtx;
5148 sparc_frame_base_offset = SPARC_STACK_BIAS;
5152 sparc_frame_base_reg = stack_pointer_rtx;
5153 sparc_frame_base_offset = size + SPARC_STACK_BIAS;
5156 if (sparc_n_global_fp_regs > 0)
5157 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg,
5158 sparc_frame_base_offset
5159 - sparc_apparent_frame_size,
5162 /* Load the GOT register if needed. */
5163 if (crtl->uses_pic_offset_table)
5164 load_got_register ();
5166 /* Advertise that the data calculated just above are now valid. */
5167 sparc_prologue_data_valid_p = true;
5170 /* This function generates the assembly code for function entry, which boils
5171 down to emitting the necessary .register directives. */
5174 sparc_asm_function_prologue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
5176 /* Check that the assumption we made in sparc_expand_prologue is valid. */
5178 gcc_assert (sparc_leaf_function_p == current_function_uses_only_leaf_regs);
5180 sparc_output_scratch_registers (file);
5183 /* Expand the function epilogue, either normal or part of a sibcall.
5184 We emit all the instructions except the return or the call. */
5187 sparc_expand_epilogue (bool for_eh)
5189 HOST_WIDE_INT size = sparc_frame_size;
5191 if (sparc_n_global_fp_regs > 0)
5192 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg,
5193 sparc_frame_base_offset
5194 - sparc_apparent_frame_size,
5197 if (size == 0 || for_eh)
5199 else if (sparc_leaf_function_p)
5202 emit_insn (gen_stack_pointer_dec (GEN_INT (-size)));
5203 else if (size <= 8192)
5205 emit_insn (gen_stack_pointer_dec (GEN_INT (-4096)));
5206 emit_insn (gen_stack_pointer_dec (GEN_INT (4096 - size)));
5210 rtx reg = gen_rtx_REG (Pmode, 1);
5211 emit_move_insn (reg, GEN_INT (-size));
5212 emit_insn (gen_stack_pointer_dec (reg));
5217 /* Expand the function epilogue, either normal or part of a sibcall.
5218 We emit all the instructions except the return or the call. */
5221 sparc_flat_expand_epilogue (bool for_eh)
5223 HOST_WIDE_INT size = sparc_frame_size;
5225 if (sparc_n_global_fp_regs > 0)
5226 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg,
5227 sparc_frame_base_offset
5228 - sparc_apparent_frame_size,
5231 /* If we have a frame pointer, we'll need both to restore it before the
5232 frame is destroyed and use its current value in destroying the frame.
5233 Since we don't have an atomic way to do that in the flat window model,
5234 we save the current value into a temporary register (%g1). */
5235 if (frame_pointer_needed && !for_eh)
5236 emit_move_insn (gen_rtx_REG (Pmode, 1), hard_frame_pointer_rtx);
5238 if (return_addr_reg_needed_p (sparc_leaf_function_p))
5239 emit_move_insn (gen_rtx_REG (Pmode, INCOMING_RETURN_ADDR_REGNUM),
5240 gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM));
5242 if (sparc_save_local_in_regs_p)
5243 emit_save_or_restore_local_in_regs (sparc_frame_base_reg,
5244 sparc_frame_base_offset,
5247 if (size == 0 || for_eh)
5249 else if (frame_pointer_needed)
5251 /* Make sure the frame is destroyed after everything else is done. */
5252 emit_insn (gen_blockage ());
5254 emit_move_insn (stack_pointer_rtx, gen_rtx_REG (Pmode, 1));
5259 emit_insn (gen_blockage ());
5262 emit_insn (gen_stack_pointer_dec (GEN_INT (-size)));
5263 else if (size <= 8192)
5265 emit_insn (gen_stack_pointer_dec (GEN_INT (-4096)));
5266 emit_insn (gen_stack_pointer_dec (GEN_INT (4096 - size)));
5270 rtx reg = gen_rtx_REG (Pmode, 1);
5271 emit_move_insn (reg, GEN_INT (-size));
5272 emit_insn (gen_stack_pointer_dec (reg));
5277 /* Return true if it is appropriate to emit `return' instructions in the
5278 body of a function. */
5281 sparc_can_use_return_insn_p (void)
5283 return sparc_prologue_data_valid_p
5284 && sparc_n_global_fp_regs == 0
5286 ? (sparc_frame_size == 0 && !sparc_save_local_in_regs_p)
5287 : (sparc_frame_size == 0 || !sparc_leaf_function_p);
5290 /* This function generates the assembly code for function exit. */
5293 sparc_asm_function_epilogue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
5295 /* If the last two instructions of a function are "call foo; dslot;"
5296 the return address might point to the first instruction in the next
5297 function and we have to output a dummy nop for the sake of sane
5298 backtraces in such cases. This is pointless for sibling calls since
5299 the return address is explicitly adjusted. */
5301 rtx insn, last_real_insn;
5303 insn = get_last_insn ();
5305 last_real_insn = prev_real_insn (insn);
5307 && GET_CODE (last_real_insn) == INSN
5308 && GET_CODE (PATTERN (last_real_insn)) == SEQUENCE)
5309 last_real_insn = XVECEXP (PATTERN (last_real_insn), 0, 0);
5312 && CALL_P (last_real_insn)
5313 && !SIBLING_CALL_P (last_real_insn))
5314 fputs("\tnop\n", file);
5316 sparc_output_deferred_case_vectors ();
5319 /* Output a 'restore' instruction. */
5322 output_restore (rtx pat)
5328 fputs ("\t restore\n", asm_out_file);
5332 gcc_assert (GET_CODE (pat) == SET);
5334 operands[0] = SET_DEST (pat);
5335 pat = SET_SRC (pat);
5337 switch (GET_CODE (pat))
5340 operands[1] = XEXP (pat, 0);
5341 operands[2] = XEXP (pat, 1);
5342 output_asm_insn (" restore %r1, %2, %Y0", operands);
5345 operands[1] = XEXP (pat, 0);
5346 operands[2] = XEXP (pat, 1);
5347 output_asm_insn (" restore %r1, %%lo(%a2), %Y0", operands);
5350 operands[1] = XEXP (pat, 0);
5351 gcc_assert (XEXP (pat, 1) == const1_rtx);
5352 output_asm_insn (" restore %r1, %r1, %Y0", operands);
5356 output_asm_insn (" restore %%g0, %1, %Y0", operands);
5361 /* Output a return. */
5364 output_return (rtx insn)
5366 if (crtl->calls_eh_return)
5368 /* If the function uses __builtin_eh_return, the eh_return
5369 machinery occupies the delay slot. */
5370 gcc_assert (!final_sequence);
5372 if (flag_delayed_branch)
5374 if (!TARGET_FLAT && TARGET_V9)
5375 fputs ("\treturn\t%i7+8\n", asm_out_file);
5379 fputs ("\trestore\n", asm_out_file);
5381 fputs ("\tjmp\t%o7+8\n", asm_out_file);
5384 fputs ("\t add\t%sp, %g1, %sp\n", asm_out_file);
5389 fputs ("\trestore\n", asm_out_file);
5391 fputs ("\tadd\t%sp, %g1, %sp\n", asm_out_file);
5392 fputs ("\tjmp\t%o7+8\n\t nop\n", asm_out_file);
5395 else if (sparc_leaf_function_p || TARGET_FLAT)
5397 /* This is a leaf or flat function so we don't have to bother restoring
5398 the register window, which frees us from dealing with the convoluted
5399 semantics of restore/return. We simply output the jump to the
5400 return address and the insn in the delay slot (if any). */
5402 return "jmp\t%%o7+%)%#";
5406 /* This is a regular function so we have to restore the register window.
5407 We may have a pending insn for the delay slot, which will be either
5408 combined with the 'restore' instruction or put in the delay slot of
5409 the 'return' instruction. */
5415 delay = NEXT_INSN (insn);
5418 pat = PATTERN (delay);
5420 if (TARGET_V9 && ! epilogue_renumber (&pat, 1))
5422 epilogue_renumber (&pat, 0);
5423 return "return\t%%i7+%)%#";
5427 output_asm_insn ("jmp\t%%i7+%)", NULL);
5428 output_restore (pat);
5429 PATTERN (delay) = gen_blockage ();
5430 INSN_CODE (delay) = -1;
5435 /* The delay slot is empty. */
5437 return "return\t%%i7+%)\n\t nop";
5438 else if (flag_delayed_branch)
5439 return "jmp\t%%i7+%)\n\t restore";
5441 return "restore\n\tjmp\t%%o7+%)\n\t nop";
5448 /* Output a sibling call. */
5451 output_sibcall (rtx insn, rtx call_operand)
5455 gcc_assert (flag_delayed_branch);
5457 operands[0] = call_operand;
5459 if (sparc_leaf_function_p || TARGET_FLAT)
5461 /* This is a leaf or flat function so we don't have to bother restoring
5462 the register window. We simply output the jump to the function and
5463 the insn in the delay slot (if any). */
5465 gcc_assert (!(LEAF_SIBCALL_SLOT_RESERVED_P && final_sequence));
5468 output_asm_insn ("sethi\t%%hi(%a0), %%g1\n\tjmp\t%%g1 + %%lo(%a0)%#",
5471 /* Use or with rs2 %%g0 instead of mov, so that as/ld can optimize
5472 it into branch if possible. */
5473 output_asm_insn ("or\t%%o7, %%g0, %%g1\n\tcall\t%a0, 0\n\t or\t%%g1, %%g0, %%o7",
5478 /* This is a regular function so we have to restore the register window.
5479 We may have a pending insn for the delay slot, which will be combined
5480 with the 'restore' instruction. */
5482 output_asm_insn ("call\t%a0, 0", operands);
5486 rtx delay = NEXT_INSN (insn);
5489 output_restore (PATTERN (delay));
5491 PATTERN (delay) = gen_blockage ();
5492 INSN_CODE (delay) = -1;
5495 output_restore (NULL_RTX);
5501 /* Functions for handling argument passing.
5503 For 32-bit, the first 6 args are normally in registers and the rest are
5504 pushed. Any arg that starts within the first 6 words is at least
5505 partially passed in a register unless its data type forbids.
5507 For 64-bit, the argument registers are laid out as an array of 16 elements
5508 and arguments are added sequentially. The first 6 int args and up to the
5509 first 16 fp args (depending on size) are passed in regs.
5511 Slot Stack Integral Float Float in structure Double Long Double
5512 ---- ----- -------- ----- ------------------ ------ -----------
5513 15 [SP+248] %f31 %f30,%f31 %d30
5514 14 [SP+240] %f29 %f28,%f29 %d28 %q28
5515 13 [SP+232] %f27 %f26,%f27 %d26
5516 12 [SP+224] %f25 %f24,%f25 %d24 %q24
5517 11 [SP+216] %f23 %f22,%f23 %d22
5518 10 [SP+208] %f21 %f20,%f21 %d20 %q20
5519 9 [SP+200] %f19 %f18,%f19 %d18
5520 8 [SP+192] %f17 %f16,%f17 %d16 %q16
5521 7 [SP+184] %f15 %f14,%f15 %d14
5522 6 [SP+176] %f13 %f12,%f13 %d12 %q12
5523 5 [SP+168] %o5 %f11 %f10,%f11 %d10
5524 4 [SP+160] %o4 %f9 %f8,%f9 %d8 %q8
5525 3 [SP+152] %o3 %f7 %f6,%f7 %d6
5526 2 [SP+144] %o2 %f5 %f4,%f5 %d4 %q4
5527 1 [SP+136] %o1 %f3 %f2,%f3 %d2
5528 0 [SP+128] %o0 %f1 %f0,%f1 %d0 %q0
5530 Here SP = %sp if -mno-stack-bias or %sp+stack_bias otherwise.
5532 Integral arguments are always passed as 64-bit quantities appropriately
5535 Passing of floating point values is handled as follows.
5536 If a prototype is in scope:
5537 If the value is in a named argument (i.e. not a stdarg function or a
5538 value not part of the `...') then the value is passed in the appropriate
5540 If the value is part of the `...' and is passed in one of the first 6
5541 slots then the value is passed in the appropriate int reg.
5542 If the value is part of the `...' and is not passed in one of the first 6
5543 slots then the value is passed in memory.
5544 If a prototype is not in scope:
5545 If the value is one of the first 6 arguments the value is passed in the
5546 appropriate integer reg and the appropriate fp reg.
5547 If the value is not one of the first 6 arguments the value is passed in
5548 the appropriate fp reg and in memory.
5551 Summary of the calling conventions implemented by GCC on the SPARC:
5554 size argument return value
5556 small integer <4 int. reg. int. reg.
5557 word 4 int. reg. int. reg.
5558 double word 8 int. reg. int. reg.
5560 _Complex small integer <8 int. reg. int. reg.
5561 _Complex word 8 int. reg. int. reg.
5562 _Complex double word 16 memory int. reg.
5564 vector integer <=8 int. reg. FP reg.
5565 vector integer >8 memory memory
5567 float 4 int. reg. FP reg.
5568 double 8 int. reg. FP reg.
5569 long double 16 memory memory
5571 _Complex float 8 memory FP reg.
5572 _Complex double 16 memory FP reg.
5573 _Complex long double 32 memory FP reg.
5575 vector float any memory memory
5577 aggregate any memory memory
5582 size argument return value
5584 small integer <8 int. reg. int. reg.
5585 word 8 int. reg. int. reg.
5586 double word 16 int. reg. int. reg.
5588 _Complex small integer <16 int. reg. int. reg.
5589 _Complex word 16 int. reg. int. reg.
5590 _Complex double word 32 memory int. reg.
5592 vector integer <=16 FP reg. FP reg.
5593 vector integer 16<s<=32 memory FP reg.
5594 vector integer >32 memory memory
5596 float 4 FP reg. FP reg.
5597 double 8 FP reg. FP reg.
5598 long double 16 FP reg. FP reg.
5600 _Complex float 8 FP reg. FP reg.
5601 _Complex double 16 FP reg. FP reg.
5602 _Complex long double 32 memory FP reg.
5604 vector float <=16 FP reg. FP reg.
5605 vector float 16<s<=32 memory FP reg.
5606 vector float >32 memory memory
5608 aggregate <=16 reg. reg.
5609 aggregate 16<s<=32 memory reg.
5610 aggregate >32 memory memory
5614 Note #1: complex floating-point types follow the extended SPARC ABIs as
5615 implemented by the Sun compiler.
5617 Note #2: integral vector types follow the scalar floating-point types
5618 conventions to match what is implemented by the Sun VIS SDK.
5620 Note #3: floating-point vector types follow the aggregate types
5624 /* Maximum number of int regs for args. */
5625 #define SPARC_INT_ARG_MAX 6
5626 /* Maximum number of fp regs for args. */
5627 #define SPARC_FP_ARG_MAX 16
5629 #define ROUND_ADVANCE(SIZE) (((SIZE) + UNITS_PER_WORD - 1) / UNITS_PER_WORD)
5631 /* Handle the INIT_CUMULATIVE_ARGS macro.
5632 Initialize a variable CUM of type CUMULATIVE_ARGS
5633 for a call to a function whose data type is FNTYPE.
5634 For a library call, FNTYPE is 0. */
5637 init_cumulative_args (struct sparc_args *cum, tree fntype,
5638 rtx libname ATTRIBUTE_UNUSED,
5639 tree fndecl ATTRIBUTE_UNUSED)
5642 cum->prototype_p = fntype && prototype_p (fntype);
5643 cum->libcall_p = fntype == 0;
5646 /* Handle promotion of pointer and integer arguments. */
5648 static enum machine_mode
5649 sparc_promote_function_mode (const_tree type,
5650 enum machine_mode mode,
5652 const_tree fntype ATTRIBUTE_UNUSED,
5653 int for_return ATTRIBUTE_UNUSED)
5655 if (type != NULL_TREE && POINTER_TYPE_P (type))
5657 *punsignedp = POINTERS_EXTEND_UNSIGNED;
5661 /* Integral arguments are passed as full words, as per the ABI. */
5662 if (GET_MODE_CLASS (mode) == MODE_INT
5663 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5669 /* Handle the TARGET_STRICT_ARGUMENT_NAMING target hook. */
5672 sparc_strict_argument_naming (cumulative_args_t ca ATTRIBUTE_UNUSED)
5674 return TARGET_ARCH64 ? true : false;
5677 /* Scan the record type TYPE and return the following predicates:
5678 - INTREGS_P: the record contains at least one field or sub-field
5679 that is eligible for promotion in integer registers.
5680 - FP_REGS_P: the record contains at least one field or sub-field
5681 that is eligible for promotion in floating-point registers.
5682 - PACKED_P: the record contains at least one field that is packed.
5684 Sub-fields are not taken into account for the PACKED_P predicate. */
5687 scan_record_type (const_tree type, int *intregs_p, int *fpregs_p,
5692 for (field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field))
5694 if (TREE_CODE (field) == FIELD_DECL)
5696 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
5697 scan_record_type (TREE_TYPE (field), intregs_p, fpregs_p, 0);
5698 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
5699 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
5705 if (packed_p && DECL_PACKED (field))
5711 /* Compute the slot number to pass an argument in.
5712 Return the slot number or -1 if passing on the stack.
5714 CUM is a variable of type CUMULATIVE_ARGS which gives info about
5715 the preceding args and about the function being called.
5716 MODE is the argument's machine mode.
5717 TYPE is the data type of the argument (as a tree).
5718 This is null for libcalls where that information may
5720 NAMED is nonzero if this argument is a named parameter
5721 (otherwise it is an extra parameter matching an ellipsis).
5722 INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG.
5723 *PREGNO records the register number to use if scalar type.
5724 *PPADDING records the amount of padding needed in words. */
5727 function_arg_slotno (const struct sparc_args *cum, enum machine_mode mode,
5728 const_tree type, bool named, bool incoming_p,
5729 int *pregno, int *ppadding)
5731 int regbase = (incoming_p
5732 ? SPARC_INCOMING_INT_ARG_FIRST
5733 : SPARC_OUTGOING_INT_ARG_FIRST);
5734 int slotno = cum->words;
5735 enum mode_class mclass;
5740 if (type && TREE_ADDRESSABLE (type))
5746 && TYPE_ALIGN (type) % PARM_BOUNDARY != 0)
5749 /* For SPARC64, objects requiring 16-byte alignment get it. */
5751 && (type ? TYPE_ALIGN (type) : GET_MODE_ALIGNMENT (mode)) >= 128
5752 && (slotno & 1) != 0)
5753 slotno++, *ppadding = 1;
5755 mclass = GET_MODE_CLASS (mode);
5756 if (type && TREE_CODE (type) == VECTOR_TYPE)
5758 /* Vector types deserve special treatment because they are
5759 polymorphic wrt their mode, depending upon whether VIS
5760 instructions are enabled. */
5761 if (TREE_CODE (TREE_TYPE (type)) == REAL_TYPE)
5763 /* The SPARC port defines no floating-point vector modes. */
5764 gcc_assert (mode == BLKmode);
5768 /* Integral vector types should either have a vector
5769 mode or an integral mode, because we are guaranteed
5770 by pass_by_reference that their size is not greater
5771 than 16 bytes and TImode is 16-byte wide. */
5772 gcc_assert (mode != BLKmode);
5774 /* Vector integers are handled like floats according to
5776 mclass = MODE_FLOAT;
5783 case MODE_COMPLEX_FLOAT:
5784 case MODE_VECTOR_INT:
5785 if (TARGET_ARCH64 && TARGET_FPU && named)
5787 if (slotno >= SPARC_FP_ARG_MAX)
5789 regno = SPARC_FP_ARG_FIRST + slotno * 2;
5790 /* Arguments filling only one single FP register are
5791 right-justified in the outer double FP register. */
5792 if (GET_MODE_SIZE (mode) <= 4)
5799 case MODE_COMPLEX_INT:
5800 if (slotno >= SPARC_INT_ARG_MAX)
5802 regno = regbase + slotno;
5806 if (mode == VOIDmode)
5807 /* MODE is VOIDmode when generating the actual call. */
5810 gcc_assert (mode == BLKmode);
5814 || (TREE_CODE (type) != VECTOR_TYPE
5815 && TREE_CODE (type) != RECORD_TYPE))
5817 if (slotno >= SPARC_INT_ARG_MAX)
5819 regno = regbase + slotno;
5821 else /* TARGET_ARCH64 && type */
5823 int intregs_p = 0, fpregs_p = 0, packed_p = 0;
5825 /* First see what kinds of registers we would need. */
5826 if (TREE_CODE (type) == VECTOR_TYPE)
5829 scan_record_type (type, &intregs_p, &fpregs_p, &packed_p);
5831 /* The ABI obviously doesn't specify how packed structures
5832 are passed. These are defined to be passed in int regs
5833 if possible, otherwise memory. */
5834 if (packed_p || !named)
5835 fpregs_p = 0, intregs_p = 1;
5837 /* If all arg slots are filled, then must pass on stack. */
5838 if (fpregs_p && slotno >= SPARC_FP_ARG_MAX)
5841 /* If there are only int args and all int arg slots are filled,
5842 then must pass on stack. */
5843 if (!fpregs_p && intregs_p && slotno >= SPARC_INT_ARG_MAX)
5846 /* Note that even if all int arg slots are filled, fp members may
5847 still be passed in regs if such regs are available.
5848 *PREGNO isn't set because there may be more than one, it's up
5849 to the caller to compute them. */
5862 /* Handle recursive register counting for structure field layout. */
5864 struct function_arg_record_value_parms
5866 rtx ret; /* return expression being built. */
5867 int slotno; /* slot number of the argument. */
5868 int named; /* whether the argument is named. */
5869 int regbase; /* regno of the base register. */
5870 int stack; /* 1 if part of the argument is on the stack. */
5871 int intoffset; /* offset of the first pending integer field. */
5872 unsigned int nregs; /* number of words passed in registers. */
5875 static void function_arg_record_value_3
5876 (HOST_WIDE_INT, struct function_arg_record_value_parms *);
5877 static void function_arg_record_value_2
5878 (const_tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
5879 static void function_arg_record_value_1
5880 (const_tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
5881 static rtx function_arg_record_value (const_tree, enum machine_mode, int, int, int);
5882 static rtx function_arg_union_value (int, enum machine_mode, int, int);
5884 /* A subroutine of function_arg_record_value. Traverse the structure
5885 recursively and determine how many registers will be required. */
5888 function_arg_record_value_1 (const_tree type, HOST_WIDE_INT startbitpos,
5889 struct function_arg_record_value_parms *parms,
5894 /* We need to compute how many registers are needed so we can
5895 allocate the PARALLEL but before we can do that we need to know
5896 whether there are any packed fields. The ABI obviously doesn't
5897 specify how structures are passed in this case, so they are
5898 defined to be passed in int regs if possible, otherwise memory,
5899 regardless of whether there are fp values present. */
5902 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5904 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
5911 /* Compute how many registers we need. */
5912 for (field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field))
5914 if (TREE_CODE (field) == FIELD_DECL)
5916 HOST_WIDE_INT bitpos = startbitpos;
5918 if (DECL_SIZE (field) != 0)
5920 if (integer_zerop (DECL_SIZE (field)))
5923 if (host_integerp (bit_position (field), 1))
5924 bitpos += int_bit_position (field);
5927 /* ??? FIXME: else assume zero offset. */
5929 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
5930 function_arg_record_value_1 (TREE_TYPE (field),
5934 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
5935 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
5940 if (parms->intoffset != -1)
5942 unsigned int startbit, endbit;
5943 int intslots, this_slotno;
5945 startbit = parms->intoffset & -BITS_PER_WORD;
5946 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
5948 intslots = (endbit - startbit) / BITS_PER_WORD;
5949 this_slotno = parms->slotno + parms->intoffset
5952 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
5954 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
5955 /* We need to pass this field on the stack. */
5959 parms->nregs += intslots;
5960 parms->intoffset = -1;
5963 /* There's no need to check this_slotno < SPARC_FP_ARG MAX.
5964 If it wasn't true we wouldn't be here. */
5965 if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE
5966 && DECL_MODE (field) == BLKmode)
5967 parms->nregs += TYPE_VECTOR_SUBPARTS (TREE_TYPE (field));
5968 else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
5975 if (parms->intoffset == -1)
5976 parms->intoffset = bitpos;
5982 /* A subroutine of function_arg_record_value. Assign the bits of the
5983 structure between parms->intoffset and bitpos to integer registers. */
5986 function_arg_record_value_3 (HOST_WIDE_INT bitpos,
5987 struct function_arg_record_value_parms *parms)
5989 enum machine_mode mode;
5991 unsigned int startbit, endbit;
5992 int this_slotno, intslots, intoffset;
5995 if (parms->intoffset == -1)
5998 intoffset = parms->intoffset;
5999 parms->intoffset = -1;
6001 startbit = intoffset & -BITS_PER_WORD;
6002 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
6003 intslots = (endbit - startbit) / BITS_PER_WORD;
6004 this_slotno = parms->slotno + intoffset / BITS_PER_WORD;
6006 intslots = MIN (intslots, SPARC_INT_ARG_MAX - this_slotno);
6010 /* If this is the trailing part of a word, only load that much into
6011 the register. Otherwise load the whole register. Note that in
6012 the latter case we may pick up unwanted bits. It's not a problem
6013 at the moment but may wish to revisit. */
6015 if (intoffset % BITS_PER_WORD != 0)
6016 mode = smallest_mode_for_size (BITS_PER_WORD - intoffset % BITS_PER_WORD,
6021 intoffset /= BITS_PER_UNIT;
6024 regno = parms->regbase + this_slotno;
6025 reg = gen_rtx_REG (mode, regno);
6026 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
6027 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (intoffset));
6030 intoffset = (intoffset | (UNITS_PER_WORD-1)) + 1;
6035 while (intslots > 0);
6038 /* A subroutine of function_arg_record_value. Traverse the structure
6039 recursively and assign bits to floating point registers. Track which
6040 bits in between need integer registers; invoke function_arg_record_value_3
6041 to make that happen. */
6044 function_arg_record_value_2 (const_tree type, HOST_WIDE_INT startbitpos,
6045 struct function_arg_record_value_parms *parms,
6051 for (field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field))
6053 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
6060 for (field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field))
6062 if (TREE_CODE (field) == FIELD_DECL)
6064 HOST_WIDE_INT bitpos = startbitpos;
6066 if (DECL_SIZE (field) != 0)
6068 if (integer_zerop (DECL_SIZE (field)))
6071 if (host_integerp (bit_position (field), 1))
6072 bitpos += int_bit_position (field);
6075 /* ??? FIXME: else assume zero offset. */
6077 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
6078 function_arg_record_value_2 (TREE_TYPE (field),
6082 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
6083 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
6088 int this_slotno = parms->slotno + bitpos / BITS_PER_WORD;
6089 int regno, nregs, pos;
6090 enum machine_mode mode = DECL_MODE (field);
6093 function_arg_record_value_3 (bitpos, parms);
6095 if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE
6098 mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field)));
6099 nregs = TYPE_VECTOR_SUBPARTS (TREE_TYPE (field));
6101 else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
6103 mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field)));
6109 regno = SPARC_FP_ARG_FIRST + this_slotno * 2;
6110 if (GET_MODE_SIZE (mode) <= 4 && (bitpos & 32) != 0)
6112 reg = gen_rtx_REG (mode, regno);
6113 pos = bitpos / BITS_PER_UNIT;
6114 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
6115 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos));
6119 regno += GET_MODE_SIZE (mode) / 4;
6120 reg = gen_rtx_REG (mode, regno);
6121 pos += GET_MODE_SIZE (mode);
6122 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
6123 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos));
6129 if (parms->intoffset == -1)
6130 parms->intoffset = bitpos;
6136 /* Used by function_arg and sparc_function_value_1 to implement the complex
6137 conventions of the 64-bit ABI for passing and returning structures.
6138 Return an expression valid as a return value for the FUNCTION_ARG
6139 and TARGET_FUNCTION_VALUE.
6141 TYPE is the data type of the argument (as a tree).
6142 This is null for libcalls where that information may
6144 MODE is the argument's machine mode.
6145 SLOTNO is the index number of the argument's slot in the parameter array.
6146 NAMED is nonzero if this argument is a named parameter
6147 (otherwise it is an extra parameter matching an ellipsis).
6148 REGBASE is the regno of the base register for the parameter array. */
6151 function_arg_record_value (const_tree type, enum machine_mode mode,
6152 int slotno, int named, int regbase)
6154 HOST_WIDE_INT typesize = int_size_in_bytes (type);
6155 struct function_arg_record_value_parms parms;
6158 parms.ret = NULL_RTX;
6159 parms.slotno = slotno;
6160 parms.named = named;
6161 parms.regbase = regbase;
6164 /* Compute how many registers we need. */
6166 parms.intoffset = 0;
6167 function_arg_record_value_1 (type, 0, &parms, false);
6169 /* Take into account pending integer fields. */
6170 if (parms.intoffset != -1)
6172 unsigned int startbit, endbit;
6173 int intslots, this_slotno;
6175 startbit = parms.intoffset & -BITS_PER_WORD;
6176 endbit = (typesize*BITS_PER_UNIT + BITS_PER_WORD - 1) & -BITS_PER_WORD;
6177 intslots = (endbit - startbit) / BITS_PER_WORD;
6178 this_slotno = slotno + parms.intoffset / BITS_PER_WORD;
6180 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
6182 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
6183 /* We need to pass this field on the stack. */
6187 parms.nregs += intslots;
6189 nregs = parms.nregs;
6191 /* Allocate the vector and handle some annoying special cases. */
6194 /* ??? Empty structure has no value? Duh? */
6197 /* Though there's nothing really to store, return a word register
6198 anyway so the rest of gcc doesn't go nuts. Returning a PARALLEL
6199 leads to breakage due to the fact that there are zero bytes to
6201 return gen_rtx_REG (mode, regbase);
6205 /* ??? C++ has structures with no fields, and yet a size. Give up
6206 for now and pass everything back in integer registers. */
6207 nregs = (typesize + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
6209 if (nregs + slotno > SPARC_INT_ARG_MAX)
6210 nregs = SPARC_INT_ARG_MAX - slotno;
6212 gcc_assert (nregs != 0);
6214 parms.ret = gen_rtx_PARALLEL (mode, rtvec_alloc (parms.stack + nregs));
6216 /* If at least one field must be passed on the stack, generate
6217 (parallel [(expr_list (nil) ...) ...]) so that all fields will
6218 also be passed on the stack. We can't do much better because the
6219 semantics of TARGET_ARG_PARTIAL_BYTES doesn't handle the case
6220 of structures for which the fields passed exclusively in registers
6221 are not at the beginning of the structure. */
6223 XVECEXP (parms.ret, 0, 0)
6224 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
6226 /* Fill in the entries. */
6228 parms.intoffset = 0;
6229 function_arg_record_value_2 (type, 0, &parms, false);
6230 function_arg_record_value_3 (typesize * BITS_PER_UNIT, &parms);
6232 gcc_assert (parms.nregs == nregs);
6237 /* Used by function_arg and sparc_function_value_1 to implement the conventions
6238 of the 64-bit ABI for passing and returning unions.
6239 Return an expression valid as a return value for the FUNCTION_ARG
6240 and TARGET_FUNCTION_VALUE.
6242 SIZE is the size in bytes of the union.
6243 MODE is the argument's machine mode.
6244 REGNO is the hard register the union will be passed in. */
6247 function_arg_union_value (int size, enum machine_mode mode, int slotno,
6250 int nwords = ROUND_ADVANCE (size), i;
6253 /* See comment in previous function for empty structures. */
6255 return gen_rtx_REG (mode, regno);
6257 if (slotno == SPARC_INT_ARG_MAX - 1)
6260 regs = gen_rtx_PARALLEL (mode, rtvec_alloc (nwords));
6262 for (i = 0; i < nwords; i++)
6264 /* Unions are passed left-justified. */
6265 XVECEXP (regs, 0, i)
6266 = gen_rtx_EXPR_LIST (VOIDmode,
6267 gen_rtx_REG (word_mode, regno),
6268 GEN_INT (UNITS_PER_WORD * i));
6275 /* Used by function_arg and sparc_function_value_1 to implement the conventions
6276 for passing and returning large (BLKmode) vectors.
6277 Return an expression valid as a return value for the FUNCTION_ARG
6278 and TARGET_FUNCTION_VALUE.
6280 SIZE is the size in bytes of the vector (at least 8 bytes).
6281 REGNO is the FP hard register the vector will be passed in. */
6284 function_arg_vector_value (int size, int regno)
6286 int i, nregs = size / 8;
6289 regs = gen_rtx_PARALLEL (BLKmode, rtvec_alloc (nregs));
6291 for (i = 0; i < nregs; i++)
6293 XVECEXP (regs, 0, i)
6294 = gen_rtx_EXPR_LIST (VOIDmode,
6295 gen_rtx_REG (DImode, regno + 2*i),
6302 /* Determine where to put an argument to a function.
6303 Value is zero to push the argument on the stack,
6304 or a hard register in which to store the argument.
6306 CUM is a variable of type CUMULATIVE_ARGS which gives info about
6307 the preceding args and about the function being called.
6308 MODE is the argument's machine mode.
6309 TYPE is the data type of the argument (as a tree).
6310 This is null for libcalls where that information may
6312 NAMED is true if this argument is a named parameter
6313 (otherwise it is an extra parameter matching an ellipsis).
6314 INCOMING_P is false for TARGET_FUNCTION_ARG, true for
6315 TARGET_FUNCTION_INCOMING_ARG. */
6318 sparc_function_arg_1 (cumulative_args_t cum_v, enum machine_mode mode,
6319 const_tree type, bool named, bool incoming_p)
6321 const CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v);
6323 int regbase = (incoming_p
6324 ? SPARC_INCOMING_INT_ARG_FIRST
6325 : SPARC_OUTGOING_INT_ARG_FIRST);
6326 int slotno, regno, padding;
6327 enum mode_class mclass = GET_MODE_CLASS (mode);
6329 slotno = function_arg_slotno (cum, mode, type, named, incoming_p,
6334 /* Vector types deserve special treatment because they are polymorphic wrt
6335 their mode, depending upon whether VIS instructions are enabled. */
6336 if (type && TREE_CODE (type) == VECTOR_TYPE)
6338 HOST_WIDE_INT size = int_size_in_bytes (type);
6339 gcc_assert ((TARGET_ARCH32 && size <= 8)
6340 || (TARGET_ARCH64 && size <= 16));
6342 if (mode == BLKmode)
6343 return function_arg_vector_value (size,
6344 SPARC_FP_ARG_FIRST + 2*slotno);
6346 mclass = MODE_FLOAT;
6350 return gen_rtx_REG (mode, regno);
6352 /* Structures up to 16 bytes in size are passed in arg slots on the stack
6353 and are promoted to registers if possible. */
6354 if (type && TREE_CODE (type) == RECORD_TYPE)
6356 HOST_WIDE_INT size = int_size_in_bytes (type);
6357 gcc_assert (size <= 16);
6359 return function_arg_record_value (type, mode, slotno, named, regbase);
6362 /* Unions up to 16 bytes in size are passed in integer registers. */
6363 else if (type && TREE_CODE (type) == UNION_TYPE)
6365 HOST_WIDE_INT size = int_size_in_bytes (type);
6366 gcc_assert (size <= 16);
6368 return function_arg_union_value (size, mode, slotno, regno);
6371 /* v9 fp args in reg slots beyond the int reg slots get passed in regs
6372 but also have the slot allocated for them.
6373 If no prototype is in scope fp values in register slots get passed
6374 in two places, either fp regs and int regs or fp regs and memory. */
6375 else if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT)
6376 && SPARC_FP_REG_P (regno))
6378 rtx reg = gen_rtx_REG (mode, regno);
6379 if (cum->prototype_p || cum->libcall_p)
6381 /* "* 2" because fp reg numbers are recorded in 4 byte
6384 /* ??? This will cause the value to be passed in the fp reg and
6385 in the stack. When a prototype exists we want to pass the
6386 value in the reg but reserve space on the stack. That's an
6387 optimization, and is deferred [for a bit]. */
6388 if ((regno - SPARC_FP_ARG_FIRST) >= SPARC_INT_ARG_MAX * 2)
6389 return gen_rtx_PARALLEL (mode,
6391 gen_rtx_EXPR_LIST (VOIDmode,
6392 NULL_RTX, const0_rtx),
6393 gen_rtx_EXPR_LIST (VOIDmode,
6397 /* ??? It seems that passing back a register even when past
6398 the area declared by REG_PARM_STACK_SPACE will allocate
6399 space appropriately, and will not copy the data onto the
6400 stack, exactly as we desire.
6402 This is due to locate_and_pad_parm being called in
6403 expand_call whenever reg_parm_stack_space > 0, which
6404 while beneficial to our example here, would seem to be
6405 in error from what had been intended. Ho hum... -- r~ */
6413 if ((regno - SPARC_FP_ARG_FIRST) < SPARC_INT_ARG_MAX * 2)
6417 /* On incoming, we don't need to know that the value
6418 is passed in %f0 and %i0, and it confuses other parts
6419 causing needless spillage even on the simplest cases. */
6423 intreg = (SPARC_OUTGOING_INT_ARG_FIRST
6424 + (regno - SPARC_FP_ARG_FIRST) / 2);
6426 v0 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
6427 v1 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_REG (mode, intreg),
6429 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
6433 v0 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
6434 v1 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
6435 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
6440 /* All other aggregate types are passed in an integer register in a mode
6441 corresponding to the size of the type. */
6442 else if (type && AGGREGATE_TYPE_P (type))
6444 HOST_WIDE_INT size = int_size_in_bytes (type);
6445 gcc_assert (size <= 16);
6447 mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0);
6450 return gen_rtx_REG (mode, regno);
6453 /* Handle the TARGET_FUNCTION_ARG target hook. */
6456 sparc_function_arg (cumulative_args_t cum, enum machine_mode mode,
6457 const_tree type, bool named)
6459 return sparc_function_arg_1 (cum, mode, type, named, false);
6462 /* Handle the TARGET_FUNCTION_INCOMING_ARG target hook. */
6465 sparc_function_incoming_arg (cumulative_args_t cum, enum machine_mode mode,
6466 const_tree type, bool named)
6468 return sparc_function_arg_1 (cum, mode, type, named, true);
6471 /* For sparc64, objects requiring 16 byte alignment are passed that way. */
6474 sparc_function_arg_boundary (enum machine_mode mode, const_tree type)
6476 return ((TARGET_ARCH64
6477 && (GET_MODE_ALIGNMENT (mode) == 128
6478 || (type && TYPE_ALIGN (type) == 128)))
6483 /* For an arg passed partly in registers and partly in memory,
6484 this is the number of bytes of registers used.
6485 For args passed entirely in registers or entirely in memory, zero.
6487 Any arg that starts in the first 6 regs but won't entirely fit in them
6488 needs partial registers on v8. On v9, structures with integer
6489 values in arg slots 5,6 will be passed in %o5 and SP+176, and complex fp
6490 values that begin in the last fp reg [where "last fp reg" varies with the
6491 mode] will be split between that reg and memory. */
6494 sparc_arg_partial_bytes (cumulative_args_t cum, enum machine_mode mode,
6495 tree type, bool named)
6497 int slotno, regno, padding;
6499 /* We pass false for incoming_p here, it doesn't matter. */
6500 slotno = function_arg_slotno (get_cumulative_args (cum), mode, type, named,
6501 false, ®no, &padding);
6508 if ((slotno + (mode == BLKmode
6509 ? ROUND_ADVANCE (int_size_in_bytes (type))
6510 : ROUND_ADVANCE (GET_MODE_SIZE (mode))))
6511 > SPARC_INT_ARG_MAX)
6512 return (SPARC_INT_ARG_MAX - slotno) * UNITS_PER_WORD;
6516 /* We are guaranteed by pass_by_reference that the size of the
6517 argument is not greater than 16 bytes, so we only need to return
6518 one word if the argument is partially passed in registers. */
6520 if (type && AGGREGATE_TYPE_P (type))
6522 int size = int_size_in_bytes (type);
6524 if (size > UNITS_PER_WORD
6525 && slotno == SPARC_INT_ARG_MAX - 1)
6526 return UNITS_PER_WORD;
6528 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_INT
6529 || (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT
6530 && ! (TARGET_FPU && named)))
6532 /* The complex types are passed as packed types. */
6533 if (GET_MODE_SIZE (mode) > UNITS_PER_WORD
6534 && slotno == SPARC_INT_ARG_MAX - 1)
6535 return UNITS_PER_WORD;
6537 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT)
6539 if ((slotno + GET_MODE_SIZE (mode) / UNITS_PER_WORD)
6541 return UNITS_PER_WORD;
6548 /* Handle the TARGET_PASS_BY_REFERENCE target hook.
6549 Specify whether to pass the argument by reference. */
6552 sparc_pass_by_reference (cumulative_args_t cum ATTRIBUTE_UNUSED,
6553 enum machine_mode mode, const_tree type,
6554 bool named ATTRIBUTE_UNUSED)
6557 /* Original SPARC 32-bit ABI says that structures and unions,
6558 and quad-precision floats are passed by reference. For Pascal,
6559 also pass arrays by reference. All other base types are passed
6562 Extended ABI (as implemented by the Sun compiler) says that all
6563 complex floats are passed by reference. Pass complex integers
6564 in registers up to 8 bytes. More generally, enforce the 2-word
6565 cap for passing arguments in registers.
6567 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6568 integers are passed like floats of the same size, that is in
6569 registers up to 8 bytes. Pass all vector floats by reference
6570 like structure and unions. */
6571 return ((type && (AGGREGATE_TYPE_P (type) || VECTOR_FLOAT_TYPE_P (type)))
6573 /* Catch CDImode, TFmode, DCmode and TCmode. */
6574 || GET_MODE_SIZE (mode) > 8
6576 && TREE_CODE (type) == VECTOR_TYPE
6577 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8));
6579 /* Original SPARC 64-bit ABI says that structures and unions
6580 smaller than 16 bytes are passed in registers, as well as
6581 all other base types.
6583 Extended ABI (as implemented by the Sun compiler) says that
6584 complex floats are passed in registers up to 16 bytes. Pass
6585 all complex integers in registers up to 16 bytes. More generally,
6586 enforce the 2-word cap for passing arguments in registers.
6588 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6589 integers are passed like floats of the same size, that is in
6590 registers (up to 16 bytes). Pass all vector floats like structure
6593 && (AGGREGATE_TYPE_P (type) || TREE_CODE (type) == VECTOR_TYPE)
6594 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 16)
6595 /* Catch CTImode and TCmode. */
6596 || GET_MODE_SIZE (mode) > 16);
6599 /* Handle the TARGET_FUNCTION_ARG_ADVANCE hook.
6600 Update the data in CUM to advance over an argument
6601 of mode MODE and data type TYPE.
6602 TYPE is null for libcalls where that information may not be available. */
6605 sparc_function_arg_advance (cumulative_args_t cum_v, enum machine_mode mode,
6606 const_tree type, bool named)
6608 CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v);
6611 /* We pass false for incoming_p here, it doesn't matter. */
6612 function_arg_slotno (cum, mode, type, named, false, ®no, &padding);
6614 /* If argument requires leading padding, add it. */
6615 cum->words += padding;
6619 cum->words += (mode != BLKmode
6620 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
6621 : ROUND_ADVANCE (int_size_in_bytes (type)));
6625 if (type && AGGREGATE_TYPE_P (type))
6627 int size = int_size_in_bytes (type);
6631 else if (size <= 16)
6633 else /* passed by reference */
6638 cum->words += (mode != BLKmode
6639 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
6640 : ROUND_ADVANCE (int_size_in_bytes (type)));
6645 /* Handle the FUNCTION_ARG_PADDING macro.
6646 For the 64 bit ABI structs are always stored left shifted in their
6650 function_arg_padding (enum machine_mode mode, const_tree type)
6652 if (TARGET_ARCH64 && type != 0 && AGGREGATE_TYPE_P (type))
6655 /* Fall back to the default. */
6656 return DEFAULT_FUNCTION_ARG_PADDING (mode, type);
6659 /* Handle the TARGET_RETURN_IN_MEMORY target hook.
6660 Specify whether to return the return value in memory. */
6663 sparc_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
6666 /* Original SPARC 32-bit ABI says that structures and unions,
6667 and quad-precision floats are returned in memory. All other
6668 base types are returned in registers.
6670 Extended ABI (as implemented by the Sun compiler) says that
6671 all complex floats are returned in registers (8 FP registers
6672 at most for '_Complex long double'). Return all complex integers
6673 in registers (4 at most for '_Complex long long').
6675 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6676 integers are returned like floats of the same size, that is in
6677 registers up to 8 bytes and in memory otherwise. Return all
6678 vector floats in memory like structure and unions; note that
6679 they always have BLKmode like the latter. */
6680 return (TYPE_MODE (type) == BLKmode
6681 || TYPE_MODE (type) == TFmode
6682 || (TREE_CODE (type) == VECTOR_TYPE
6683 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8));
6685 /* Original SPARC 64-bit ABI says that structures and unions
6686 smaller than 32 bytes are returned in registers, as well as
6687 all other base types.
6689 Extended ABI (as implemented by the Sun compiler) says that all
6690 complex floats are returned in registers (8 FP registers at most
6691 for '_Complex long double'). Return all complex integers in
6692 registers (4 at most for '_Complex TItype').
6694 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6695 integers are returned like floats of the same size, that is in
6696 registers. Return all vector floats like structure and unions;
6697 note that they always have BLKmode like the latter. */
6698 return (TYPE_MODE (type) == BLKmode
6699 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 32);
6702 /* Handle the TARGET_STRUCT_VALUE target hook.
6703 Return where to find the structure return value address. */
6706 sparc_struct_value_rtx (tree fndecl, int incoming)
6715 mem = gen_frame_mem (Pmode, plus_constant (frame_pointer_rtx,
6716 STRUCT_VALUE_OFFSET));
6718 mem = gen_frame_mem (Pmode, plus_constant (stack_pointer_rtx,
6719 STRUCT_VALUE_OFFSET));
6721 /* Only follow the SPARC ABI for fixed-size structure returns.
6722 Variable size structure returns are handled per the normal
6723 procedures in GCC. This is enabled by -mstd-struct-return */
6725 && sparc_std_struct_return
6726 && TYPE_SIZE_UNIT (TREE_TYPE (fndecl))
6727 && TREE_CODE (TYPE_SIZE_UNIT (TREE_TYPE (fndecl))) == INTEGER_CST)
6729 /* We must check and adjust the return address, as it is
6730 optional as to whether the return object is really
6732 rtx ret_reg = gen_rtx_REG (Pmode, 31);
6733 rtx scratch = gen_reg_rtx (SImode);
6734 rtx endlab = gen_label_rtx ();
6736 /* Calculate the return object size */
6737 tree size = TYPE_SIZE_UNIT (TREE_TYPE (fndecl));
6738 rtx size_rtx = GEN_INT (TREE_INT_CST_LOW (size) & 0xfff);
6739 /* Construct a temporary return value */
6741 = assign_stack_local (Pmode, TREE_INT_CST_LOW (size), 0);
6743 /* Implement SPARC 32-bit psABI callee return struct checking:
6745 Fetch the instruction where we will return to and see if
6746 it's an unimp instruction (the most significant 10 bits
6748 emit_move_insn (scratch, gen_rtx_MEM (SImode,
6749 plus_constant (ret_reg, 8)));
6750 /* Assume the size is valid and pre-adjust */
6751 emit_insn (gen_add3_insn (ret_reg, ret_reg, GEN_INT (4)));
6752 emit_cmp_and_jump_insns (scratch, size_rtx, EQ, const0_rtx, SImode,
6754 emit_insn (gen_sub3_insn (ret_reg, ret_reg, GEN_INT (4)));
6755 /* Write the address of the memory pointed to by temp_val into
6756 the memory pointed to by mem */
6757 emit_move_insn (mem, XEXP (temp_val, 0));
6758 emit_label (endlab);
6765 /* Handle TARGET_FUNCTION_VALUE, and TARGET_LIBCALL_VALUE target hook.
6766 For v9, function return values are subject to the same rules as arguments,
6767 except that up to 32 bytes may be returned in registers. */
6770 sparc_function_value_1 (const_tree type, enum machine_mode mode,
6773 /* Beware that the two values are swapped here wrt function_arg. */
6774 int regbase = (outgoing
6775 ? SPARC_INCOMING_INT_ARG_FIRST
6776 : SPARC_OUTGOING_INT_ARG_FIRST);
6777 enum mode_class mclass = GET_MODE_CLASS (mode);
6780 /* Vector types deserve special treatment because they are polymorphic wrt
6781 their mode, depending upon whether VIS instructions are enabled. */
6782 if (type && TREE_CODE (type) == VECTOR_TYPE)
6784 HOST_WIDE_INT size = int_size_in_bytes (type);
6785 gcc_assert ((TARGET_ARCH32 && size <= 8)
6786 || (TARGET_ARCH64 && size <= 32));
6788 if (mode == BLKmode)
6789 return function_arg_vector_value (size,
6790 SPARC_FP_ARG_FIRST);
6792 mclass = MODE_FLOAT;
6795 if (TARGET_ARCH64 && type)
6797 /* Structures up to 32 bytes in size are returned in registers. */
6798 if (TREE_CODE (type) == RECORD_TYPE)
6800 HOST_WIDE_INT size = int_size_in_bytes (type);
6801 gcc_assert (size <= 32);
6803 return function_arg_record_value (type, mode, 0, 1, regbase);
6806 /* Unions up to 32 bytes in size are returned in integer registers. */
6807 else if (TREE_CODE (type) == UNION_TYPE)
6809 HOST_WIDE_INT size = int_size_in_bytes (type);
6810 gcc_assert (size <= 32);
6812 return function_arg_union_value (size, mode, 0, regbase);
6815 /* Objects that require it are returned in FP registers. */
6816 else if (mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT)
6819 /* All other aggregate types are returned in an integer register in a
6820 mode corresponding to the size of the type. */
6821 else if (AGGREGATE_TYPE_P (type))
6823 /* All other aggregate types are passed in an integer register
6824 in a mode corresponding to the size of the type. */
6825 HOST_WIDE_INT size = int_size_in_bytes (type);
6826 gcc_assert (size <= 32);
6828 mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0);
6830 /* ??? We probably should have made the same ABI change in
6831 3.4.0 as the one we made for unions. The latter was
6832 required by the SCD though, while the former is not
6833 specified, so we favored compatibility and efficiency.
6835 Now we're stuck for aggregates larger than 16 bytes,
6836 because OImode vanished in the meantime. Let's not
6837 try to be unduly clever, and simply follow the ABI
6838 for unions in that case. */
6839 if (mode == BLKmode)
6840 return function_arg_union_value (size, mode, 0, regbase);
6845 /* We should only have pointer and integer types at this point. This
6846 must match sparc_promote_function_mode. */
6847 else if (mclass == MODE_INT && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
6851 /* We should only have pointer and integer types at this point. This must
6852 match sparc_promote_function_mode. */
6853 else if (TARGET_ARCH32
6854 && mclass == MODE_INT
6855 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
6858 if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT) && TARGET_FPU)
6859 regno = SPARC_FP_ARG_FIRST;
6863 return gen_rtx_REG (mode, regno);
6866 /* Handle TARGET_FUNCTION_VALUE.
6867 On the SPARC, the value is found in the first "output" register, but the
6868 called function leaves it in the first "input" register. */
6871 sparc_function_value (const_tree valtype,
6872 const_tree fn_decl_or_type ATTRIBUTE_UNUSED,
6875 return sparc_function_value_1 (valtype, TYPE_MODE (valtype), outgoing);
6878 /* Handle TARGET_LIBCALL_VALUE. */
6881 sparc_libcall_value (enum machine_mode mode,
6882 const_rtx fun ATTRIBUTE_UNUSED)
6884 return sparc_function_value_1 (NULL_TREE, mode, false);
6887 /* Handle FUNCTION_VALUE_REGNO_P.
6888 On the SPARC, the first "output" reg is used for integer values, and the
6889 first floating point register is used for floating point values. */
6892 sparc_function_value_regno_p (const unsigned int regno)
6894 return (regno == 8 || regno == 32);
6897 /* Do what is necessary for `va_start'. We look at the current function
6898 to determine if stdarg or varargs is used and return the address of
6899 the first unnamed parameter. */
6902 sparc_builtin_saveregs (void)
6904 int first_reg = crtl->args.info.words;
6908 for (regno = first_reg; regno < SPARC_INT_ARG_MAX; regno++)
6909 emit_move_insn (gen_rtx_MEM (word_mode,
6910 gen_rtx_PLUS (Pmode,
6912 GEN_INT (FIRST_PARM_OFFSET (0)
6915 gen_rtx_REG (word_mode,
6916 SPARC_INCOMING_INT_ARG_FIRST + regno));
6918 address = gen_rtx_PLUS (Pmode,
6920 GEN_INT (FIRST_PARM_OFFSET (0)
6921 + UNITS_PER_WORD * first_reg));
6926 /* Implement `va_start' for stdarg. */
6929 sparc_va_start (tree valist, rtx nextarg)
6931 nextarg = expand_builtin_saveregs ();
6932 std_expand_builtin_va_start (valist, nextarg);
6935 /* Implement `va_arg' for stdarg. */
6938 sparc_gimplify_va_arg (tree valist, tree type, gimple_seq *pre_p,
6941 HOST_WIDE_INT size, rsize, align;
6944 tree ptrtype = build_pointer_type (type);
6946 if (pass_by_reference (NULL, TYPE_MODE (type), type, false))
6949 size = rsize = UNITS_PER_WORD;
6955 size = int_size_in_bytes (type);
6956 rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD;
6961 /* For SPARC64, objects requiring 16-byte alignment get it. */
6962 if (TYPE_ALIGN (type) >= 2 * (unsigned) BITS_PER_WORD)
6963 align = 2 * UNITS_PER_WORD;
6965 /* SPARC-V9 ABI states that structures up to 16 bytes in size
6966 are left-justified in their slots. */
6967 if (AGGREGATE_TYPE_P (type))
6970 size = rsize = UNITS_PER_WORD;
6980 incr = fold_build_pointer_plus_hwi (incr, align - 1);
6981 incr = fold_convert (sizetype, incr);
6982 incr = fold_build2 (BIT_AND_EXPR, sizetype, incr,
6984 incr = fold_convert (ptr_type_node, incr);
6987 gimplify_expr (&incr, pre_p, post_p, is_gimple_val, fb_rvalue);
6990 if (BYTES_BIG_ENDIAN && size < rsize)
6991 addr = fold_build_pointer_plus_hwi (incr, rsize - size);
6995 addr = fold_convert (build_pointer_type (ptrtype), addr);
6996 addr = build_va_arg_indirect_ref (addr);
6999 /* If the address isn't aligned properly for the type, we need a temporary.
7000 FIXME: This is inefficient, usually we can do this in registers. */
7001 else if (align == 0 && TYPE_ALIGN (type) > BITS_PER_WORD)
7003 tree tmp = create_tmp_var (type, "va_arg_tmp");
7004 tree dest_addr = build_fold_addr_expr (tmp);
7005 tree copy = build_call_expr (builtin_decl_implicit (BUILT_IN_MEMCPY),
7006 3, dest_addr, addr, size_int (rsize));
7007 TREE_ADDRESSABLE (tmp) = 1;
7008 gimplify_and_add (copy, pre_p);
7013 addr = fold_convert (ptrtype, addr);
7015 incr = fold_build_pointer_plus_hwi (incr, rsize);
7016 gimplify_assign (valist, incr, post_p);
7018 return build_va_arg_indirect_ref (addr);
7021 /* Implement the TARGET_VECTOR_MODE_SUPPORTED_P target hook.
7022 Specify whether the vector mode is supported by the hardware. */
7025 sparc_vector_mode_supported_p (enum machine_mode mode)
7027 return TARGET_VIS && VECTOR_MODE_P (mode) ? true : false;
7030 /* Implement the TARGET_VECTORIZE_PREFERRED_SIMD_MODE target hook. */
7032 static enum machine_mode
7033 sparc_preferred_simd_mode (enum machine_mode mode)
7051 /* Return the string to output an unconditional branch to LABEL, which is
7052 the operand number of the label.
7054 DEST is the destination insn (i.e. the label), INSN is the source. */
7057 output_ubranch (rtx dest, int label, rtx insn)
7059 static char string[64];
7060 bool v9_form = false;
7063 if (TARGET_V9 && INSN_ADDRESSES_SET_P ())
7065 int delta = (INSN_ADDRESSES (INSN_UID (dest))
7066 - INSN_ADDRESSES (INSN_UID (insn)));
7067 /* Leave some instructions for "slop". */
7068 if (delta >= -260000 && delta < 260000)
7073 strcpy (string, "ba%*,pt\t%%xcc, ");
7075 strcpy (string, "b%*\t");
7077 p = strchr (string, '\0');
7088 /* Return the string to output a conditional branch to LABEL, which is
7089 the operand number of the label. OP is the conditional expression.
7090 XEXP (OP, 0) is assumed to be a condition code register (integer or
7091 floating point) and its mode specifies what kind of comparison we made.
7093 DEST is the destination insn (i.e. the label), INSN is the source.
7095 REVERSED is nonzero if we should reverse the sense of the comparison.
7097 ANNUL is nonzero if we should generate an annulling branch. */
7100 output_cbranch (rtx op, rtx dest, int label, int reversed, int annul,
7103 static char string[64];
7104 enum rtx_code code = GET_CODE (op);
7105 rtx cc_reg = XEXP (op, 0);
7106 enum machine_mode mode = GET_MODE (cc_reg);
7107 const char *labelno, *branch;
7108 int spaces = 8, far;
7111 /* v9 branches are limited to +-1MB. If it is too far away,
7124 fbne,a,pn %fcc2, .LC29
7132 far = TARGET_V9 && (get_attr_length (insn) >= 3);
7135 /* Reversal of FP compares takes care -- an ordered compare
7136 becomes an unordered compare and vice versa. */
7137 if (mode == CCFPmode || mode == CCFPEmode)
7138 code = reverse_condition_maybe_unordered (code);
7140 code = reverse_condition (code);
7143 /* Start by writing the branch condition. */
7144 if (mode == CCFPmode || mode == CCFPEmode)
7195 /* ??? !v9: FP branches cannot be preceded by another floating point
7196 insn. Because there is currently no concept of pre-delay slots,
7197 we can fix this only by always emitting a nop before a floating
7202 strcpy (string, "nop\n\t");
7203 strcat (string, branch);
7216 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
7228 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
7249 strcpy (string, branch);
7251 spaces -= strlen (branch);
7252 p = strchr (string, '\0');
7254 /* Now add the annulling, the label, and a possible noop. */
7267 if (! far && insn && INSN_ADDRESSES_SET_P ())
7269 int delta = (INSN_ADDRESSES (INSN_UID (dest))
7270 - INSN_ADDRESSES (INSN_UID (insn)));
7271 /* Leave some instructions for "slop". */
7272 if (delta < -260000 || delta >= 260000)
7276 if (mode == CCFPmode || mode == CCFPEmode)
7278 static char v9_fcc_labelno[] = "%%fccX, ";
7279 /* Set the char indicating the number of the fcc reg to use. */
7280 v9_fcc_labelno[5] = REGNO (cc_reg) - SPARC_FIRST_V9_FCC_REG + '0';
7281 labelno = v9_fcc_labelno;
7284 gcc_assert (REGNO (cc_reg) == SPARC_FCC_REG);
7288 else if (mode == CCXmode || mode == CCX_NOOVmode)
7290 labelno = "%%xcc, ";
7295 labelno = "%%icc, ";
7300 if (*labelno && insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
7303 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
7316 strcpy (p, labelno);
7317 p = strchr (p, '\0');
7320 strcpy (p, ".+12\n\t nop\n\tb\t");
7321 /* Skip the next insn if requested or
7322 if we know that it will be a nop. */
7323 if (annul || ! final_sequence)
7337 /* Emit a library call comparison between floating point X and Y.
7338 COMPARISON is the operator to compare with (EQ, NE, GT, etc).
7339 Return the new operator to be used in the comparison sequence.
7341 TARGET_ARCH64 uses _Qp_* functions, which use pointers to TFmode
7342 values as arguments instead of the TFmode registers themselves,
7343 that's why we cannot call emit_float_lib_cmp. */
7346 sparc_emit_float_lib_cmp (rtx x, rtx y, enum rtx_code comparison)
7349 rtx slot0, slot1, result, tem, tem2, libfunc;
7350 enum machine_mode mode;
7351 enum rtx_code new_comparison;
7356 qpfunc = (TARGET_ARCH64 ? "_Qp_feq" : "_Q_feq");
7360 qpfunc = (TARGET_ARCH64 ? "_Qp_fne" : "_Q_fne");
7364 qpfunc = (TARGET_ARCH64 ? "_Qp_fgt" : "_Q_fgt");
7368 qpfunc = (TARGET_ARCH64 ? "_Qp_fge" : "_Q_fge");
7372 qpfunc = (TARGET_ARCH64 ? "_Qp_flt" : "_Q_flt");
7376 qpfunc = (TARGET_ARCH64 ? "_Qp_fle" : "_Q_fle");
7387 qpfunc = (TARGET_ARCH64 ? "_Qp_cmp" : "_Q_cmp");
7398 tree expr = MEM_EXPR (x);
7400 mark_addressable (expr);
7405 slot0 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0);
7406 emit_move_insn (slot0, x);
7411 tree expr = MEM_EXPR (y);
7413 mark_addressable (expr);
7418 slot1 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0);
7419 emit_move_insn (slot1, y);
7422 libfunc = gen_rtx_SYMBOL_REF (Pmode, qpfunc);
7423 emit_library_call (libfunc, LCT_NORMAL,
7425 XEXP (slot0, 0), Pmode,
7426 XEXP (slot1, 0), Pmode);
7431 libfunc = gen_rtx_SYMBOL_REF (Pmode, qpfunc);
7432 emit_library_call (libfunc, LCT_NORMAL,
7434 x, TFmode, y, TFmode);
7439 /* Immediately move the result of the libcall into a pseudo
7440 register so reload doesn't clobber the value if it needs
7441 the return register for a spill reg. */
7442 result = gen_reg_rtx (mode);
7443 emit_move_insn (result, hard_libcall_value (mode, libfunc));
7448 return gen_rtx_NE (VOIDmode, result, const0_rtx);
7451 new_comparison = (comparison == UNORDERED ? EQ : NE);
7452 return gen_rtx_fmt_ee (new_comparison, VOIDmode, result, GEN_INT(3));
7455 new_comparison = (comparison == UNGT ? GT : NE);
7456 return gen_rtx_fmt_ee (new_comparison, VOIDmode, result, const1_rtx);
7458 return gen_rtx_NE (VOIDmode, result, const2_rtx);
7460 tem = gen_reg_rtx (mode);
7462 emit_insn (gen_andsi3 (tem, result, const1_rtx));
7464 emit_insn (gen_anddi3 (tem, result, const1_rtx));
7465 return gen_rtx_NE (VOIDmode, tem, const0_rtx);
7468 tem = gen_reg_rtx (mode);
7470 emit_insn (gen_addsi3 (tem, result, const1_rtx));
7472 emit_insn (gen_adddi3 (tem, result, const1_rtx));
7473 tem2 = gen_reg_rtx (mode);
7475 emit_insn (gen_andsi3 (tem2, tem, const2_rtx));
7477 emit_insn (gen_anddi3 (tem2, tem, const2_rtx));
7478 new_comparison = (comparison == UNEQ ? EQ : NE);
7479 return gen_rtx_fmt_ee (new_comparison, VOIDmode, tem2, const0_rtx);
7485 /* Generate an unsigned DImode to FP conversion. This is the same code
7486 optabs would emit if we didn't have TFmode patterns. */
7489 sparc_emit_floatunsdi (rtx *operands, enum machine_mode mode)
7491 rtx neglab, donelab, i0, i1, f0, in, out;
7494 in = force_reg (DImode, operands[1]);
7495 neglab = gen_label_rtx ();
7496 donelab = gen_label_rtx ();
7497 i0 = gen_reg_rtx (DImode);
7498 i1 = gen_reg_rtx (DImode);
7499 f0 = gen_reg_rtx (mode);
7501 emit_cmp_and_jump_insns (in, const0_rtx, LT, const0_rtx, DImode, 0, neglab);
7503 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_FLOAT (mode, in)));
7504 emit_jump_insn (gen_jump (donelab));
7507 emit_label (neglab);
7509 emit_insn (gen_lshrdi3 (i0, in, const1_rtx));
7510 emit_insn (gen_anddi3 (i1, in, const1_rtx));
7511 emit_insn (gen_iordi3 (i0, i0, i1));
7512 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_FLOAT (mode, i0)));
7513 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_PLUS (mode, f0, f0)));
7515 emit_label (donelab);
7518 /* Generate an FP to unsigned DImode conversion. This is the same code
7519 optabs would emit if we didn't have TFmode patterns. */
7522 sparc_emit_fixunsdi (rtx *operands, enum machine_mode mode)
7524 rtx neglab, donelab, i0, i1, f0, in, out, limit;
7527 in = force_reg (mode, operands[1]);
7528 neglab = gen_label_rtx ();
7529 donelab = gen_label_rtx ();
7530 i0 = gen_reg_rtx (DImode);
7531 i1 = gen_reg_rtx (DImode);
7532 limit = gen_reg_rtx (mode);
7533 f0 = gen_reg_rtx (mode);
7535 emit_move_insn (limit,
7536 CONST_DOUBLE_FROM_REAL_VALUE (
7537 REAL_VALUE_ATOF ("9223372036854775808.0", mode), mode));
7538 emit_cmp_and_jump_insns (in, limit, GE, NULL_RTX, mode, 0, neglab);
7540 emit_insn (gen_rtx_SET (VOIDmode,
7542 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, in))));
7543 emit_jump_insn (gen_jump (donelab));
7546 emit_label (neglab);
7548 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_MINUS (mode, in, limit)));
7549 emit_insn (gen_rtx_SET (VOIDmode,
7551 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, f0))));
7552 emit_insn (gen_movdi (i1, const1_rtx));
7553 emit_insn (gen_ashldi3 (i1, i1, GEN_INT (63)));
7554 emit_insn (gen_xordi3 (out, i0, i1));
7556 emit_label (donelab);
7559 /* Return the string to output a conditional branch to LABEL, testing
7560 register REG. LABEL is the operand number of the label; REG is the
7561 operand number of the reg. OP is the conditional expression. The mode
7562 of REG says what kind of comparison we made.
7564 DEST is the destination insn (i.e. the label), INSN is the source.
7566 REVERSED is nonzero if we should reverse the sense of the comparison.
7568 ANNUL is nonzero if we should generate an annulling branch. */
7571 output_v9branch (rtx op, rtx dest, int reg, int label, int reversed,
7572 int annul, rtx insn)
7574 static char string[64];
7575 enum rtx_code code = GET_CODE (op);
7576 enum machine_mode mode = GET_MODE (XEXP (op, 0));
7581 /* branch on register are limited to +-128KB. If it is too far away,
7594 brgez,a,pn %o1, .LC29
7600 ba,pt %xcc, .LC29 */
7602 far = get_attr_length (insn) >= 3;
7604 /* If not floating-point or if EQ or NE, we can just reverse the code. */
7606 code = reverse_condition (code);
7608 /* Only 64 bit versions of these instructions exist. */
7609 gcc_assert (mode == DImode);
7611 /* Start by writing the branch condition. */
7616 strcpy (string, "brnz");
7620 strcpy (string, "brz");
7624 strcpy (string, "brgez");
7628 strcpy (string, "brlz");
7632 strcpy (string, "brlez");
7636 strcpy (string, "brgz");
7643 p = strchr (string, '\0');
7645 /* Now add the annulling, reg, label, and nop. */
7652 if (insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
7655 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
7660 *p = p < string + 8 ? '\t' : ' ';
7668 int veryfar = 1, delta;
7670 if (INSN_ADDRESSES_SET_P ())
7672 delta = (INSN_ADDRESSES (INSN_UID (dest))
7673 - INSN_ADDRESSES (INSN_UID (insn)));
7674 /* Leave some instructions for "slop". */
7675 if (delta >= -260000 && delta < 260000)
7679 strcpy (p, ".+12\n\t nop\n\t");
7680 /* Skip the next insn if requested or
7681 if we know that it will be a nop. */
7682 if (annul || ! final_sequence)
7692 strcpy (p, "ba,pt\t%%xcc, ");
7706 /* Return 1, if any of the registers of the instruction are %l[0-7] or %o[0-7].
7707 Such instructions cannot be used in the delay slot of return insn on v9.
7708 If TEST is 0, also rename all %i[0-7] registers to their %o[0-7] counterparts.
7712 epilogue_renumber (register rtx *where, int test)
7714 register const char *fmt;
7716 register enum rtx_code code;
7721 code = GET_CODE (*where);
7726 if (REGNO (*where) >= 8 && REGNO (*where) < 24) /* oX or lX */
7728 if (! test && REGNO (*where) >= 24 && REGNO (*where) < 32)
7729 *where = gen_rtx_REG (GET_MODE (*where), OUTGOING_REGNO (REGNO(*where)));
7737 /* Do not replace the frame pointer with the stack pointer because
7738 it can cause the delayed instruction to load below the stack.
7739 This occurs when instructions like:
7741 (set (reg/i:SI 24 %i0)
7742 (mem/f:SI (plus:SI (reg/f:SI 30 %fp)
7743 (const_int -20 [0xffffffec])) 0))
7745 are in the return delayed slot. */
7747 if (GET_CODE (XEXP (*where, 0)) == REG
7748 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM
7749 && (GET_CODE (XEXP (*where, 1)) != CONST_INT
7750 || INTVAL (XEXP (*where, 1)) < SPARC_STACK_BIAS))
7755 if (SPARC_STACK_BIAS
7756 && GET_CODE (XEXP (*where, 0)) == REG
7757 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM)
7765 fmt = GET_RTX_FORMAT (code);
7767 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7772 for (j = XVECLEN (*where, i) - 1; j >= 0; j--)
7773 if (epilogue_renumber (&(XVECEXP (*where, i, j)), test))
7776 else if (fmt[i] == 'e'
7777 && epilogue_renumber (&(XEXP (*where, i)), test))
7783 /* Leaf functions and non-leaf functions have different needs. */
7786 reg_leaf_alloc_order[] = REG_LEAF_ALLOC_ORDER;
7789 reg_nonleaf_alloc_order[] = REG_ALLOC_ORDER;
7791 static const int *const reg_alloc_orders[] = {
7792 reg_leaf_alloc_order,
7793 reg_nonleaf_alloc_order};
7796 order_regs_for_local_alloc (void)
7798 static int last_order_nonleaf = 1;
7800 if (df_regs_ever_live_p (15) != last_order_nonleaf)
7802 last_order_nonleaf = !last_order_nonleaf;
7803 memcpy ((char *) reg_alloc_order,
7804 (const char *) reg_alloc_orders[last_order_nonleaf],
7805 FIRST_PSEUDO_REGISTER * sizeof (int));
7809 /* Return 1 if REG and MEM are legitimate enough to allow the various
7810 mem<-->reg splits to be run. */
7813 sparc_splitdi_legitimate (rtx reg, rtx mem)
7815 /* Punt if we are here by mistake. */
7816 gcc_assert (reload_completed);
7818 /* We must have an offsettable memory reference. */
7819 if (! offsettable_memref_p (mem))
7822 /* If we have legitimate args for ldd/std, we do not want
7823 the split to happen. */
7824 if ((REGNO (reg) % 2) == 0
7825 && mem_min_alignment (mem, 8))
7832 /* Like sparc_splitdi_legitimate but for REG <--> REG moves. */
7835 sparc_split_regreg_legitimate (rtx reg1, rtx reg2)
7839 if (GET_CODE (reg1) == SUBREG)
7840 reg1 = SUBREG_REG (reg1);
7841 if (GET_CODE (reg1) != REG)
7843 regno1 = REGNO (reg1);
7845 if (GET_CODE (reg2) == SUBREG)
7846 reg2 = SUBREG_REG (reg2);
7847 if (GET_CODE (reg2) != REG)
7849 regno2 = REGNO (reg2);
7851 if (SPARC_INT_REG_P (regno1) && SPARC_INT_REG_P (regno2))
7856 if ((SPARC_INT_REG_P (regno1) && SPARC_FP_REG_P (regno2))
7857 || (SPARC_FP_REG_P (regno1) && SPARC_INT_REG_P (regno2)))
7864 /* Return 1 if x and y are some kind of REG and they refer to
7865 different hard registers. This test is guaranteed to be
7866 run after reload. */
7869 sparc_absnegfloat_split_legitimate (rtx x, rtx y)
7871 if (GET_CODE (x) != REG)
7873 if (GET_CODE (y) != REG)
7875 if (REGNO (x) == REGNO (y))
7880 /* Return 1 if REGNO (reg1) is even and REGNO (reg1) == REGNO (reg2) - 1.
7881 This makes them candidates for using ldd and std insns.
7883 Note reg1 and reg2 *must* be hard registers. */
7886 registers_ok_for_ldd_peep (rtx reg1, rtx reg2)
7888 /* We might have been passed a SUBREG. */
7889 if (GET_CODE (reg1) != REG || GET_CODE (reg2) != REG)
7892 if (REGNO (reg1) % 2 != 0)
7895 /* Integer ldd is deprecated in SPARC V9 */
7896 if (TARGET_V9 && SPARC_INT_REG_P (REGNO (reg1)))
7899 return (REGNO (reg1) == REGNO (reg2) - 1);
7902 /* Return 1 if the addresses in mem1 and mem2 are suitable for use in
7905 This can only happen when addr1 and addr2, the addresses in mem1
7906 and mem2, are consecutive memory locations (addr1 + 4 == addr2).
7907 addr1 must also be aligned on a 64-bit boundary.
7909 Also iff dependent_reg_rtx is not null it should not be used to
7910 compute the address for mem1, i.e. we cannot optimize a sequence
7922 But, note that the transformation from:
7927 is perfectly fine. Thus, the peephole2 patterns always pass us
7928 the destination register of the first load, never the second one.
7930 For stores we don't have a similar problem, so dependent_reg_rtx is
7934 mems_ok_for_ldd_peep (rtx mem1, rtx mem2, rtx dependent_reg_rtx)
7938 HOST_WIDE_INT offset1;
7940 /* The mems cannot be volatile. */
7941 if (MEM_VOLATILE_P (mem1) || MEM_VOLATILE_P (mem2))
7944 /* MEM1 should be aligned on a 64-bit boundary. */
7945 if (MEM_ALIGN (mem1) < 64)
7948 addr1 = XEXP (mem1, 0);
7949 addr2 = XEXP (mem2, 0);
7951 /* Extract a register number and offset (if used) from the first addr. */
7952 if (GET_CODE (addr1) == PLUS)
7954 /* If not a REG, return zero. */
7955 if (GET_CODE (XEXP (addr1, 0)) != REG)
7959 reg1 = REGNO (XEXP (addr1, 0));
7960 /* The offset must be constant! */
7961 if (GET_CODE (XEXP (addr1, 1)) != CONST_INT)
7963 offset1 = INTVAL (XEXP (addr1, 1));
7966 else if (GET_CODE (addr1) != REG)
7970 reg1 = REGNO (addr1);
7971 /* This was a simple (mem (reg)) expression. Offset is 0. */
7975 /* Make sure the second address is a (mem (plus (reg) (const_int). */
7976 if (GET_CODE (addr2) != PLUS)
7979 if (GET_CODE (XEXP (addr2, 0)) != REG
7980 || GET_CODE (XEXP (addr2, 1)) != CONST_INT)
7983 if (reg1 != REGNO (XEXP (addr2, 0)))
7986 if (dependent_reg_rtx != NULL_RTX && reg1 == REGNO (dependent_reg_rtx))
7989 /* The first offset must be evenly divisible by 8 to ensure the
7990 address is 64 bit aligned. */
7991 if (offset1 % 8 != 0)
7994 /* The offset for the second addr must be 4 more than the first addr. */
7995 if (INTVAL (XEXP (addr2, 1)) != offset1 + 4)
7998 /* All the tests passed. addr1 and addr2 are valid for ldd and std
8003 /* Return 1 if reg is a pseudo, or is the first register in
8004 a hard register pair. This makes it suitable for use in
8005 ldd and std insns. */
8008 register_ok_for_ldd (rtx reg)
8010 /* We might have been passed a SUBREG. */
8014 if (REGNO (reg) < FIRST_PSEUDO_REGISTER)
8015 return (REGNO (reg) % 2 == 0);
8020 /* Return 1 if OP is a memory whose address is known to be
8021 aligned to 8-byte boundary, or a pseudo during reload.
8022 This makes it suitable for use in ldd and std insns. */
8025 memory_ok_for_ldd (rtx op)
8029 /* In 64-bit mode, we assume that the address is word-aligned. */
8030 if (TARGET_ARCH32 && !mem_min_alignment (op, 8))
8033 if (! can_create_pseudo_p ()
8034 && !strict_memory_address_p (Pmode, XEXP (op, 0)))
8037 else if (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER)
8039 if (!(reload_in_progress && reg_renumber [REGNO (op)] < 0))
8048 /* Implement TARGET_PRINT_OPERAND_PUNCT_VALID_P. */
8051 sparc_print_operand_punct_valid_p (unsigned char code)
8064 /* Implement TARGET_PRINT_OPERAND.
8065 Print operand X (an rtx) in assembler syntax to file FILE.
8066 CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified.
8067 For `%' followed by punctuation, CODE is the punctuation and X is null. */
8070 sparc_print_operand (FILE *file, rtx x, int code)
8075 /* Output an insn in a delay slot. */
8077 sparc_indent_opcode = 1;
8079 fputs ("\n\t nop", file);
8082 /* Output an annul flag if there's nothing for the delay slot and we
8083 are optimizing. This is always used with '(' below.
8084 Sun OS 4.1.1 dbx can't handle an annulled unconditional branch;
8085 this is a dbx bug. So, we only do this when optimizing.
8086 On UltraSPARC, a branch in a delay slot causes a pipeline flush.
8087 Always emit a nop in case the next instruction is a branch. */
8088 if (! final_sequence && (optimize && (int)sparc_cpu < PROCESSOR_V9))
8092 /* Output a 'nop' if there's nothing for the delay slot and we are
8093 not optimizing. This is always used with '*' above. */
8094 if (! final_sequence && ! (optimize && (int)sparc_cpu < PROCESSOR_V9))
8095 fputs ("\n\t nop", file);
8096 else if (final_sequence)
8097 sparc_indent_opcode = 1;
8100 /* Output the right displacement from the saved PC on function return.
8101 The caller may have placed an "unimp" insn immediately after the call
8102 so we have to account for it. This insn is used in the 32-bit ABI
8103 when calling a function that returns a non zero-sized structure. The
8104 64-bit ABI doesn't have it. Be careful to have this test be the same
8105 as that for the call. The exception is when sparc_std_struct_return
8106 is enabled, the psABI is followed exactly and the adjustment is made
8107 by the code in sparc_struct_value_rtx. The call emitted is the same
8108 when sparc_std_struct_return is enabled. */
8110 && cfun->returns_struct
8111 && !sparc_std_struct_return
8112 && DECL_SIZE (DECL_RESULT (current_function_decl))
8113 && TREE_CODE (DECL_SIZE (DECL_RESULT (current_function_decl)))
8115 && !integer_zerop (DECL_SIZE (DECL_RESULT (current_function_decl))))
8121 /* Output the Embedded Medium/Anywhere code model base register. */
8122 fputs (EMBMEDANY_BASE_REG, file);
8125 /* Print some local dynamic TLS name. */
8126 assemble_name (file, get_some_local_dynamic_name ());
8130 /* Adjust the operand to take into account a RESTORE operation. */
8131 if (GET_CODE (x) == CONST_INT)
8133 else if (GET_CODE (x) != REG)
8134 output_operand_lossage ("invalid %%Y operand");
8135 else if (REGNO (x) < 8)
8136 fputs (reg_names[REGNO (x)], file);
8137 else if (REGNO (x) >= 24 && REGNO (x) < 32)
8138 fputs (reg_names[REGNO (x)-16], file);
8140 output_operand_lossage ("invalid %%Y operand");
8143 /* Print out the low order register name of a register pair. */
8144 if (WORDS_BIG_ENDIAN)
8145 fputs (reg_names[REGNO (x)+1], file);
8147 fputs (reg_names[REGNO (x)], file);
8150 /* Print out the high order register name of a register pair. */
8151 if (WORDS_BIG_ENDIAN)
8152 fputs (reg_names[REGNO (x)], file);
8154 fputs (reg_names[REGNO (x)+1], file);
8157 /* Print out the second register name of a register pair or quad.
8158 I.e., R (%o0) => %o1. */
8159 fputs (reg_names[REGNO (x)+1], file);
8162 /* Print out the third register name of a register quad.
8163 I.e., S (%o0) => %o2. */
8164 fputs (reg_names[REGNO (x)+2], file);
8167 /* Print out the fourth register name of a register quad.
8168 I.e., T (%o0) => %o3. */
8169 fputs (reg_names[REGNO (x)+3], file);
8172 /* Print a condition code register. */
8173 if (REGNO (x) == SPARC_ICC_REG)
8175 /* We don't handle CC[X]_NOOVmode because they're not supposed
8177 if (GET_MODE (x) == CCmode)
8178 fputs ("%icc", file);
8179 else if (GET_MODE (x) == CCXmode)
8180 fputs ("%xcc", file);
8185 /* %fccN register */
8186 fputs (reg_names[REGNO (x)], file);
8189 /* Print the operand's address only. */
8190 output_address (XEXP (x, 0));
8193 /* In this case we need a register. Use %g0 if the
8194 operand is const0_rtx. */
8196 || (GET_MODE (x) != VOIDmode && x == CONST0_RTX (GET_MODE (x))))
8198 fputs ("%g0", file);
8205 switch (GET_CODE (x))
8207 case IOR: fputs ("or", file); break;
8208 case AND: fputs ("and", file); break;
8209 case XOR: fputs ("xor", file); break;
8210 default: output_operand_lossage ("invalid %%A operand");
8215 switch (GET_CODE (x))
8217 case IOR: fputs ("orn", file); break;
8218 case AND: fputs ("andn", file); break;
8219 case XOR: fputs ("xnor", file); break;
8220 default: output_operand_lossage ("invalid %%B operand");
8224 /* This is used by the conditional move instructions. */
8227 enum rtx_code rc = GET_CODE (x);
8231 case NE: fputs ("ne", file); break;
8232 case EQ: fputs ("e", file); break;
8233 case GE: fputs ("ge", file); break;
8234 case GT: fputs ("g", file); break;
8235 case LE: fputs ("le", file); break;
8236 case LT: fputs ("l", file); break;
8237 case GEU: fputs ("geu", file); break;
8238 case GTU: fputs ("gu", file); break;
8239 case LEU: fputs ("leu", file); break;
8240 case LTU: fputs ("lu", file); break;
8241 case LTGT: fputs ("lg", file); break;
8242 case UNORDERED: fputs ("u", file); break;
8243 case ORDERED: fputs ("o", file); break;
8244 case UNLT: fputs ("ul", file); break;
8245 case UNLE: fputs ("ule", file); break;
8246 case UNGT: fputs ("ug", file); break;
8247 case UNGE: fputs ("uge", file); break;
8248 case UNEQ: fputs ("ue", file); break;
8249 default: output_operand_lossage ("invalid %%C operand");
8254 /* This are used by the movr instruction pattern. */
8257 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 ("gez", file); break;
8263 case LT: fputs ("lz", file); break;
8264 case LE: fputs ("lez", file); break;
8265 case GT: fputs ("gz", file); break;
8266 default: output_operand_lossage ("invalid %%D operand");
8273 /* Print a sign-extended character. */
8274 int i = trunc_int_for_mode (INTVAL (x), QImode);
8275 fprintf (file, "%d", i);
8280 /* Operand must be a MEM; write its address. */
8281 if (GET_CODE (x) != MEM)
8282 output_operand_lossage ("invalid %%f operand");
8283 output_address (XEXP (x, 0));
8288 /* Print a sign-extended 32-bit value. */
8290 if (GET_CODE(x) == CONST_INT)
8292 else if (GET_CODE(x) == CONST_DOUBLE)
8293 i = CONST_DOUBLE_LOW (x);
8296 output_operand_lossage ("invalid %%s operand");
8299 i = trunc_int_for_mode (i, SImode);
8300 fprintf (file, HOST_WIDE_INT_PRINT_DEC, i);
8305 /* Do nothing special. */
8309 /* Undocumented flag. */
8310 output_operand_lossage ("invalid operand output code");
8313 if (GET_CODE (x) == REG)
8314 fputs (reg_names[REGNO (x)], file);
8315 else if (GET_CODE (x) == MEM)
8318 /* Poor Sun assembler doesn't understand absolute addressing. */
8319 if (CONSTANT_P (XEXP (x, 0)))
8320 fputs ("%g0+", file);
8321 output_address (XEXP (x, 0));
8324 else if (GET_CODE (x) == HIGH)
8326 fputs ("%hi(", file);
8327 output_addr_const (file, XEXP (x, 0));
8330 else if (GET_CODE (x) == LO_SUM)
8332 sparc_print_operand (file, XEXP (x, 0), 0);
8333 if (TARGET_CM_MEDMID)
8334 fputs ("+%l44(", file);
8336 fputs ("+%lo(", file);
8337 output_addr_const (file, XEXP (x, 1));
8340 else if (GET_CODE (x) == CONST_DOUBLE
8341 && (GET_MODE (x) == VOIDmode
8342 || GET_MODE_CLASS (GET_MODE (x)) == MODE_INT))
8344 if (CONST_DOUBLE_HIGH (x) == 0)
8345 fprintf (file, "%u", (unsigned int) CONST_DOUBLE_LOW (x));
8346 else if (CONST_DOUBLE_HIGH (x) == -1
8347 && CONST_DOUBLE_LOW (x) < 0)
8348 fprintf (file, "%d", (int) CONST_DOUBLE_LOW (x));
8350 output_operand_lossage ("long long constant not a valid immediate operand");
8352 else if (GET_CODE (x) == CONST_DOUBLE)
8353 output_operand_lossage ("floating point constant not a valid immediate operand");
8354 else { output_addr_const (file, x); }
8357 /* Implement TARGET_PRINT_OPERAND_ADDRESS. */
8360 sparc_print_operand_address (FILE *file, rtx x)
8362 register rtx base, index = 0;
8364 register rtx addr = x;
8367 fputs (reg_names[REGNO (addr)], file);
8368 else if (GET_CODE (addr) == PLUS)
8370 if (CONST_INT_P (XEXP (addr, 0)))
8371 offset = INTVAL (XEXP (addr, 0)), base = XEXP (addr, 1);
8372 else if (CONST_INT_P (XEXP (addr, 1)))
8373 offset = INTVAL (XEXP (addr, 1)), base = XEXP (addr, 0);
8375 base = XEXP (addr, 0), index = XEXP (addr, 1);
8376 if (GET_CODE (base) == LO_SUM)
8378 gcc_assert (USE_AS_OFFSETABLE_LO10
8380 && ! TARGET_CM_MEDMID);
8381 output_operand (XEXP (base, 0), 0);
8382 fputs ("+%lo(", file);
8383 output_address (XEXP (base, 1));
8384 fprintf (file, ")+%d", offset);
8388 fputs (reg_names[REGNO (base)], file);
8390 fprintf (file, "%+d", offset);
8391 else if (REG_P (index))
8392 fprintf (file, "+%s", reg_names[REGNO (index)]);
8393 else if (GET_CODE (index) == SYMBOL_REF
8394 || GET_CODE (index) == LABEL_REF
8395 || GET_CODE (index) == CONST)
8396 fputc ('+', file), output_addr_const (file, index);
8397 else gcc_unreachable ();
8400 else if (GET_CODE (addr) == MINUS
8401 && GET_CODE (XEXP (addr, 1)) == LABEL_REF)
8403 output_addr_const (file, XEXP (addr, 0));
8405 output_addr_const (file, XEXP (addr, 1));
8406 fputs ("-.)", file);
8408 else if (GET_CODE (addr) == LO_SUM)
8410 output_operand (XEXP (addr, 0), 0);
8411 if (TARGET_CM_MEDMID)
8412 fputs ("+%l44(", file);
8414 fputs ("+%lo(", file);
8415 output_address (XEXP (addr, 1));
8419 && GET_CODE (addr) == CONST
8420 && GET_CODE (XEXP (addr, 0)) == MINUS
8421 && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST
8422 && GET_CODE (XEXP (XEXP (XEXP (addr, 0), 1), 0)) == MINUS
8423 && XEXP (XEXP (XEXP (XEXP (addr, 0), 1), 0), 1) == pc_rtx)
8425 addr = XEXP (addr, 0);
8426 output_addr_const (file, XEXP (addr, 0));
8427 /* Group the args of the second CONST in parenthesis. */
8429 /* Skip past the second CONST--it does nothing for us. */
8430 output_addr_const (file, XEXP (XEXP (addr, 1), 0));
8431 /* Close the parenthesis. */
8436 output_addr_const (file, addr);
8440 /* Target hook for assembling integer objects. The sparc version has
8441 special handling for aligned DI-mode objects. */
8444 sparc_assemble_integer (rtx x, unsigned int size, int aligned_p)
8446 /* ??? We only output .xword's for symbols and only then in environments
8447 where the assembler can handle them. */
8448 if (aligned_p && size == 8
8449 && (GET_CODE (x) != CONST_INT && GET_CODE (x) != CONST_DOUBLE))
8453 assemble_integer_with_op ("\t.xword\t", x);
8458 assemble_aligned_integer (4, const0_rtx);
8459 assemble_aligned_integer (4, x);
8463 return default_assemble_integer (x, size, aligned_p);
8466 /* Return the value of a code used in the .proc pseudo-op that says
8467 what kind of result this function returns. For non-C types, we pick
8468 the closest C type. */
8470 #ifndef SHORT_TYPE_SIZE
8471 #define SHORT_TYPE_SIZE (BITS_PER_UNIT * 2)
8474 #ifndef INT_TYPE_SIZE
8475 #define INT_TYPE_SIZE BITS_PER_WORD
8478 #ifndef LONG_TYPE_SIZE
8479 #define LONG_TYPE_SIZE BITS_PER_WORD
8482 #ifndef LONG_LONG_TYPE_SIZE
8483 #define LONG_LONG_TYPE_SIZE (BITS_PER_WORD * 2)
8486 #ifndef FLOAT_TYPE_SIZE
8487 #define FLOAT_TYPE_SIZE BITS_PER_WORD
8490 #ifndef DOUBLE_TYPE_SIZE
8491 #define DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
8494 #ifndef LONG_DOUBLE_TYPE_SIZE
8495 #define LONG_DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
8499 sparc_type_code (register tree type)
8501 register unsigned long qualifiers = 0;
8502 register unsigned shift;
8504 /* Only the first 30 bits of the qualifier are valid. We must refrain from
8505 setting more, since some assemblers will give an error for this. Also,
8506 we must be careful to avoid shifts of 32 bits or more to avoid getting
8507 unpredictable results. */
8509 for (shift = 6; shift < 30; shift += 2, type = TREE_TYPE (type))
8511 switch (TREE_CODE (type))
8517 qualifiers |= (3 << shift);
8522 qualifiers |= (2 << shift);
8526 case REFERENCE_TYPE:
8528 qualifiers |= (1 << shift);
8532 return (qualifiers | 8);
8535 case QUAL_UNION_TYPE:
8536 return (qualifiers | 9);
8539 return (qualifiers | 10);
8542 return (qualifiers | 16);
8545 /* If this is a range type, consider it to be the underlying
8547 if (TREE_TYPE (type) != 0)
8550 /* Carefully distinguish all the standard types of C,
8551 without messing up if the language is not C. We do this by
8552 testing TYPE_PRECISION and TYPE_UNSIGNED. The old code used to
8553 look at both the names and the above fields, but that's redundant.
8554 Any type whose size is between two C types will be considered
8555 to be the wider of the two types. Also, we do not have a
8556 special code to use for "long long", so anything wider than
8557 long is treated the same. Note that we can't distinguish
8558 between "int" and "long" in this code if they are the same
8559 size, but that's fine, since neither can the assembler. */
8561 if (TYPE_PRECISION (type) <= CHAR_TYPE_SIZE)
8562 return (qualifiers | (TYPE_UNSIGNED (type) ? 12 : 2));
8564 else if (TYPE_PRECISION (type) <= SHORT_TYPE_SIZE)
8565 return (qualifiers | (TYPE_UNSIGNED (type) ? 13 : 3));
8567 else if (TYPE_PRECISION (type) <= INT_TYPE_SIZE)
8568 return (qualifiers | (TYPE_UNSIGNED (type) ? 14 : 4));
8571 return (qualifiers | (TYPE_UNSIGNED (type) ? 15 : 5));
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. */
8582 if (TYPE_PRECISION (type) == FLOAT_TYPE_SIZE)
8583 return (qualifiers | 6);
8586 return (qualifiers | 7);
8588 case COMPLEX_TYPE: /* GNU Fortran COMPLEX type. */
8589 /* ??? We need to distinguish between double and float complex types,
8590 but I don't know how yet because I can't reach this code from
8591 existing front-ends. */
8592 return (qualifiers | 7); /* Who knows? */
8595 case BOOLEAN_TYPE: /* Boolean truth value type. */
8601 gcc_unreachable (); /* Not a type! */
8608 /* Nested function support. */
8610 /* Emit RTL insns to initialize the variable parts of a trampoline.
8611 FNADDR is an RTX for the address of the function's pure code.
8612 CXT is an RTX for the static chain value for the function.
8614 This takes 16 insns: 2 shifts & 2 ands (to split up addresses), 4 sethi
8615 (to load in opcodes), 4 iors (to merge address and opcodes), and 4 writes
8616 (to store insns). This is a bit excessive. Perhaps a different
8617 mechanism would be better here.
8619 Emit enough FLUSH insns to synchronize the data and instruction caches. */
8622 sparc32_initialize_trampoline (rtx m_tramp, rtx fnaddr, rtx cxt)
8624 /* SPARC 32-bit trampoline:
8627 sethi %hi(static), %g2
8629 or %g2, %lo(static), %g2
8631 SETHI i,r = 00rr rrr1 00ii iiii iiii iiii iiii iiii
8632 JMPL r+i,d = 10dd ddd1 1100 0rrr rr1i iiii iiii iiii
8636 (adjust_address (m_tramp, SImode, 0),
8637 expand_binop (SImode, ior_optab,
8638 expand_shift (RSHIFT_EXPR, SImode, fnaddr, 10, 0, 1),
8639 GEN_INT (trunc_int_for_mode (0x03000000, SImode)),
8640 NULL_RTX, 1, OPTAB_DIRECT));
8643 (adjust_address (m_tramp, SImode, 4),
8644 expand_binop (SImode, ior_optab,
8645 expand_shift (RSHIFT_EXPR, SImode, cxt, 10, 0, 1),
8646 GEN_INT (trunc_int_for_mode (0x05000000, SImode)),
8647 NULL_RTX, 1, OPTAB_DIRECT));
8650 (adjust_address (m_tramp, SImode, 8),
8651 expand_binop (SImode, ior_optab,
8652 expand_and (SImode, fnaddr, GEN_INT (0x3ff), NULL_RTX),
8653 GEN_INT (trunc_int_for_mode (0x81c06000, SImode)),
8654 NULL_RTX, 1, OPTAB_DIRECT));
8657 (adjust_address (m_tramp, SImode, 12),
8658 expand_binop (SImode, ior_optab,
8659 expand_and (SImode, cxt, GEN_INT (0x3ff), NULL_RTX),
8660 GEN_INT (trunc_int_for_mode (0x8410a000, SImode)),
8661 NULL_RTX, 1, OPTAB_DIRECT));
8663 /* On UltraSPARC a flush flushes an entire cache line. The trampoline is
8664 aligned on a 16 byte boundary so one flush clears it all. */
8665 emit_insn (gen_flush (validize_mem (adjust_address (m_tramp, SImode, 0))));
8666 if (sparc_cpu != PROCESSOR_ULTRASPARC
8667 && sparc_cpu != PROCESSOR_ULTRASPARC3
8668 && sparc_cpu != PROCESSOR_NIAGARA
8669 && sparc_cpu != PROCESSOR_NIAGARA2
8670 && sparc_cpu != PROCESSOR_NIAGARA3
8671 && sparc_cpu != PROCESSOR_NIAGARA4)
8672 emit_insn (gen_flush (validize_mem (adjust_address (m_tramp, SImode, 8))));
8674 /* Call __enable_execute_stack after writing onto the stack to make sure
8675 the stack address is accessible. */
8676 #ifdef HAVE_ENABLE_EXECUTE_STACK
8677 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
8678 LCT_NORMAL, VOIDmode, 1, XEXP (m_tramp, 0), Pmode);
8683 /* The 64-bit version is simpler because it makes more sense to load the
8684 values as "immediate" data out of the trampoline. It's also easier since
8685 we can read the PC without clobbering a register. */
8688 sparc64_initialize_trampoline (rtx m_tramp, rtx fnaddr, rtx cxt)
8690 /* SPARC 64-bit trampoline:
8699 emit_move_insn (adjust_address (m_tramp, SImode, 0),
8700 GEN_INT (trunc_int_for_mode (0x83414000, SImode)));
8701 emit_move_insn (adjust_address (m_tramp, SImode, 4),
8702 GEN_INT (trunc_int_for_mode (0xca586018, SImode)));
8703 emit_move_insn (adjust_address (m_tramp, SImode, 8),
8704 GEN_INT (trunc_int_for_mode (0x81c14000, SImode)));
8705 emit_move_insn (adjust_address (m_tramp, SImode, 12),
8706 GEN_INT (trunc_int_for_mode (0xca586010, SImode)));
8707 emit_move_insn (adjust_address (m_tramp, DImode, 16), cxt);
8708 emit_move_insn (adjust_address (m_tramp, DImode, 24), fnaddr);
8709 emit_insn (gen_flushdi (validize_mem (adjust_address (m_tramp, DImode, 0))));
8711 if (sparc_cpu != PROCESSOR_ULTRASPARC
8712 && sparc_cpu != PROCESSOR_ULTRASPARC3
8713 && sparc_cpu != PROCESSOR_NIAGARA
8714 && sparc_cpu != PROCESSOR_NIAGARA2
8715 && sparc_cpu != PROCESSOR_NIAGARA3
8716 && sparc_cpu != PROCESSOR_NIAGARA4)
8717 emit_insn (gen_flushdi (validize_mem (adjust_address (m_tramp, DImode, 8))));
8719 /* Call __enable_execute_stack after writing onto the stack to make sure
8720 the stack address is accessible. */
8721 #ifdef HAVE_ENABLE_EXECUTE_STACK
8722 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
8723 LCT_NORMAL, VOIDmode, 1, XEXP (m_tramp, 0), Pmode);
8727 /* Worker for TARGET_TRAMPOLINE_INIT. */
8730 sparc_trampoline_init (rtx m_tramp, tree fndecl, rtx cxt)
8732 rtx fnaddr = force_reg (Pmode, XEXP (DECL_RTL (fndecl), 0));
8733 cxt = force_reg (Pmode, cxt);
8735 sparc64_initialize_trampoline (m_tramp, fnaddr, cxt);
8737 sparc32_initialize_trampoline (m_tramp, fnaddr, cxt);
8740 /* Adjust the cost of a scheduling dependency. Return the new cost of
8741 a dependency LINK or INSN on DEP_INSN. COST is the current cost. */
8744 supersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
8746 enum attr_type insn_type;
8748 if (! recog_memoized (insn))
8751 insn_type = get_attr_type (insn);
8753 if (REG_NOTE_KIND (link) == 0)
8755 /* Data dependency; DEP_INSN writes a register that INSN reads some
8758 /* if a load, then the dependence must be on the memory address;
8759 add an extra "cycle". Note that the cost could be two cycles
8760 if the reg was written late in an instruction group; we ca not tell
8762 if (insn_type == TYPE_LOAD || insn_type == TYPE_FPLOAD)
8765 /* Get the delay only if the address of the store is the dependence. */
8766 if (insn_type == TYPE_STORE || insn_type == TYPE_FPSTORE)
8768 rtx pat = PATTERN(insn);
8769 rtx dep_pat = PATTERN (dep_insn);
8771 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
8772 return cost; /* This should not happen! */
8774 /* The dependency between the two instructions was on the data that
8775 is being stored. Assume that this implies that the address of the
8776 store is not dependent. */
8777 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
8780 return cost + 3; /* An approximation. */
8783 /* A shift instruction cannot receive its data from an instruction
8784 in the same cycle; add a one cycle penalty. */
8785 if (insn_type == TYPE_SHIFT)
8786 return cost + 3; /* Split before cascade into shift. */
8790 /* Anti- or output- dependency; DEP_INSN reads/writes a register that
8791 INSN writes some cycles later. */
8793 /* These are only significant for the fpu unit; writing a fp reg before
8794 the fpu has finished with it stalls the processor. */
8796 /* Reusing an integer register causes no problems. */
8797 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
8805 hypersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
8807 enum attr_type insn_type, dep_type;
8808 rtx pat = PATTERN(insn);
8809 rtx dep_pat = PATTERN (dep_insn);
8811 if (recog_memoized (insn) < 0 || recog_memoized (dep_insn) < 0)
8814 insn_type = get_attr_type (insn);
8815 dep_type = get_attr_type (dep_insn);
8817 switch (REG_NOTE_KIND (link))
8820 /* Data dependency; DEP_INSN writes a register that INSN reads some
8827 /* Get the delay iff the address of the store is the dependence. */
8828 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
8831 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
8838 /* If a load, then the dependence must be on the memory address. If
8839 the addresses aren't equal, then it might be a false dependency */
8840 if (dep_type == TYPE_STORE || dep_type == TYPE_FPSTORE)
8842 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET
8843 || GET_CODE (SET_DEST (dep_pat)) != MEM
8844 || GET_CODE (SET_SRC (pat)) != MEM
8845 || ! rtx_equal_p (XEXP (SET_DEST (dep_pat), 0),
8846 XEXP (SET_SRC (pat), 0)))
8854 /* Compare to branch latency is 0. There is no benefit from
8855 separating compare and branch. */
8856 if (dep_type == TYPE_COMPARE)
8858 /* Floating point compare to branch latency is less than
8859 compare to conditional move. */
8860 if (dep_type == TYPE_FPCMP)
8869 /* Anti-dependencies only penalize the fpu unit. */
8870 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
8882 sparc_adjust_cost(rtx insn, rtx link, rtx dep, int cost)
8886 case PROCESSOR_SUPERSPARC:
8887 cost = supersparc_adjust_cost (insn, link, dep, cost);
8889 case PROCESSOR_HYPERSPARC:
8890 case PROCESSOR_SPARCLITE86X:
8891 cost = hypersparc_adjust_cost (insn, link, dep, cost);
8900 sparc_sched_init (FILE *dump ATTRIBUTE_UNUSED,
8901 int sched_verbose ATTRIBUTE_UNUSED,
8902 int max_ready ATTRIBUTE_UNUSED)
8906 sparc_use_sched_lookahead (void)
8908 if (sparc_cpu == PROCESSOR_NIAGARA
8909 || sparc_cpu == PROCESSOR_NIAGARA2
8910 || sparc_cpu == PROCESSOR_NIAGARA3
8911 || sparc_cpu == PROCESSOR_NIAGARA4)
8913 if (sparc_cpu == PROCESSOR_ULTRASPARC
8914 || sparc_cpu == PROCESSOR_ULTRASPARC3)
8916 if ((1 << sparc_cpu) &
8917 ((1 << PROCESSOR_SUPERSPARC) | (1 << PROCESSOR_HYPERSPARC) |
8918 (1 << PROCESSOR_SPARCLITE86X)))
8924 sparc_issue_rate (void)
8928 case PROCESSOR_NIAGARA:
8929 case PROCESSOR_NIAGARA2:
8930 case PROCESSOR_NIAGARA3:
8931 case PROCESSOR_NIAGARA4:
8935 /* Assume V9 processors are capable of at least dual-issue. */
8937 case PROCESSOR_SUPERSPARC:
8939 case PROCESSOR_HYPERSPARC:
8940 case PROCESSOR_SPARCLITE86X:
8942 case PROCESSOR_ULTRASPARC:
8943 case PROCESSOR_ULTRASPARC3:
8949 set_extends (rtx insn)
8951 register rtx pat = PATTERN (insn);
8953 switch (GET_CODE (SET_SRC (pat)))
8955 /* Load and some shift instructions zero extend. */
8958 /* sethi clears the high bits */
8960 /* LO_SUM is used with sethi. sethi cleared the high
8961 bits and the values used with lo_sum are positive */
8963 /* Store flag stores 0 or 1 */
8973 rtx op0 = XEXP (SET_SRC (pat), 0);
8974 rtx op1 = XEXP (SET_SRC (pat), 1);
8975 if (GET_CODE (op1) == CONST_INT)
8976 return INTVAL (op1) >= 0;
8977 if (GET_CODE (op0) != REG)
8979 if (sparc_check_64 (op0, insn) == 1)
8981 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
8986 rtx op0 = XEXP (SET_SRC (pat), 0);
8987 rtx op1 = XEXP (SET_SRC (pat), 1);
8988 if (GET_CODE (op0) != REG || sparc_check_64 (op0, insn) <= 0)
8990 if (GET_CODE (op1) == CONST_INT)
8991 return INTVAL (op1) >= 0;
8992 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
8995 return GET_MODE (SET_SRC (pat)) == SImode;
8996 /* Positive integers leave the high bits zero. */
8998 return ! (CONST_DOUBLE_LOW (SET_SRC (pat)) & 0x80000000);
9000 return ! (INTVAL (SET_SRC (pat)) & 0x80000000);
9003 return - (GET_MODE (SET_SRC (pat)) == SImode);
9005 return sparc_check_64 (SET_SRC (pat), insn);
9011 /* We _ought_ to have only one kind per function, but... */
9012 static GTY(()) rtx sparc_addr_diff_list;
9013 static GTY(()) rtx sparc_addr_list;
9016 sparc_defer_case_vector (rtx lab, rtx vec, int diff)
9018 vec = gen_rtx_EXPR_LIST (VOIDmode, lab, vec);
9020 sparc_addr_diff_list
9021 = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_diff_list);
9023 sparc_addr_list = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_list);
9027 sparc_output_addr_vec (rtx vec)
9029 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
9030 int idx, vlen = XVECLEN (body, 0);
9032 #ifdef ASM_OUTPUT_ADDR_VEC_START
9033 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
9036 #ifdef ASM_OUTPUT_CASE_LABEL
9037 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
9040 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
9043 for (idx = 0; idx < vlen; idx++)
9045 ASM_OUTPUT_ADDR_VEC_ELT
9046 (asm_out_file, CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 0, idx), 0)));
9049 #ifdef ASM_OUTPUT_ADDR_VEC_END
9050 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
9055 sparc_output_addr_diff_vec (rtx vec)
9057 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
9058 rtx base = XEXP (XEXP (body, 0), 0);
9059 int idx, vlen = XVECLEN (body, 1);
9061 #ifdef ASM_OUTPUT_ADDR_VEC_START
9062 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
9065 #ifdef ASM_OUTPUT_CASE_LABEL
9066 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
9069 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
9072 for (idx = 0; idx < vlen; idx++)
9074 ASM_OUTPUT_ADDR_DIFF_ELT
9077 CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 1, idx), 0)),
9078 CODE_LABEL_NUMBER (base));
9081 #ifdef ASM_OUTPUT_ADDR_VEC_END
9082 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
9087 sparc_output_deferred_case_vectors (void)
9092 if (sparc_addr_list == NULL_RTX
9093 && sparc_addr_diff_list == NULL_RTX)
9096 /* Align to cache line in the function's code section. */
9097 switch_to_section (current_function_section ());
9099 align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
9101 ASM_OUTPUT_ALIGN (asm_out_file, align);
9103 for (t = sparc_addr_list; t ; t = XEXP (t, 1))
9104 sparc_output_addr_vec (XEXP (t, 0));
9105 for (t = sparc_addr_diff_list; t ; t = XEXP (t, 1))
9106 sparc_output_addr_diff_vec (XEXP (t, 0));
9108 sparc_addr_list = sparc_addr_diff_list = NULL_RTX;
9111 /* Return 0 if the high 32 bits of X (the low word of X, if DImode) are
9112 unknown. Return 1 if the high bits are zero, -1 if the register is
9115 sparc_check_64 (rtx x, rtx insn)
9117 /* If a register is set only once it is safe to ignore insns this
9118 code does not know how to handle. The loop will either recognize
9119 the single set and return the correct value or fail to recognize
9124 gcc_assert (GET_CODE (x) == REG);
9126 if (GET_MODE (x) == DImode)
9127 y = gen_rtx_REG (SImode, REGNO (x) + WORDS_BIG_ENDIAN);
9129 if (flag_expensive_optimizations
9130 && df && DF_REG_DEF_COUNT (REGNO (y)) == 1)
9136 insn = get_last_insn_anywhere ();
9141 while ((insn = PREV_INSN (insn)))
9143 switch (GET_CODE (insn))
9156 rtx pat = PATTERN (insn);
9157 if (GET_CODE (pat) != SET)
9159 if (rtx_equal_p (x, SET_DEST (pat)))
9160 return set_extends (insn);
9161 if (y && rtx_equal_p (y, SET_DEST (pat)))
9162 return set_extends (insn);
9163 if (reg_overlap_mentioned_p (SET_DEST (pat), y))
9171 /* Output a wide shift instruction in V8+ mode. INSN is the instruction,
9172 OPERANDS are its operands and OPCODE is the mnemonic to be used. */
9175 output_v8plus_shift (rtx insn, rtx *operands, const char *opcode)
9177 static char asm_code[60];
9179 /* The scratch register is only required when the destination
9180 register is not a 64-bit global or out register. */
9181 if (which_alternative != 2)
9182 operands[3] = operands[0];
9184 /* We can only shift by constants <= 63. */
9185 if (GET_CODE (operands[2]) == CONST_INT)
9186 operands[2] = GEN_INT (INTVAL (operands[2]) & 0x3f);
9188 if (GET_CODE (operands[1]) == CONST_INT)
9190 output_asm_insn ("mov\t%1, %3", operands);
9194 output_asm_insn ("sllx\t%H1, 32, %3", operands);
9195 if (sparc_check_64 (operands[1], insn) <= 0)
9196 output_asm_insn ("srl\t%L1, 0, %L1", operands);
9197 output_asm_insn ("or\t%L1, %3, %3", operands);
9200 strcpy (asm_code, opcode);
9202 if (which_alternative != 2)
9203 return strcat (asm_code, "\t%0, %2, %L0\n\tsrlx\t%L0, 32, %H0");
9206 strcat (asm_code, "\t%3, %2, %3\n\tsrlx\t%3, 32, %H0\n\tmov\t%3, %L0");
9209 /* Output rtl to increment the profiler label LABELNO
9210 for profiling a function entry. */
9213 sparc_profile_hook (int labelno)
9218 fun = gen_rtx_SYMBOL_REF (Pmode, MCOUNT_FUNCTION);
9219 if (NO_PROFILE_COUNTERS)
9221 emit_library_call (fun, LCT_NORMAL, VOIDmode, 0);
9225 ASM_GENERATE_INTERNAL_LABEL (buf, "LP", labelno);
9226 lab = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (buf));
9227 emit_library_call (fun, LCT_NORMAL, VOIDmode, 1, lab, Pmode);
9231 #ifdef TARGET_SOLARIS
9232 /* Solaris implementation of TARGET_ASM_NAMED_SECTION. */
9235 sparc_solaris_elf_asm_named_section (const char *name, unsigned int flags,
9236 tree decl ATTRIBUTE_UNUSED)
9238 if (HAVE_COMDAT_GROUP && flags & SECTION_LINKONCE)
9240 solaris_elf_asm_comdat_section (name, flags, decl);
9244 fprintf (asm_out_file, "\t.section\t\"%s\"", name);
9246 if (!(flags & SECTION_DEBUG))
9247 fputs (",#alloc", asm_out_file);
9248 if (flags & SECTION_WRITE)
9249 fputs (",#write", asm_out_file);
9250 if (flags & SECTION_TLS)
9251 fputs (",#tls", asm_out_file);
9252 if (flags & SECTION_CODE)
9253 fputs (",#execinstr", asm_out_file);
9255 /* ??? Handle SECTION_BSS. */
9257 fputc ('\n', asm_out_file);
9259 #endif /* TARGET_SOLARIS */
9261 /* We do not allow indirect calls to be optimized into sibling calls.
9263 We cannot use sibling calls when delayed branches are disabled
9264 because they will likely require the call delay slot to be filled.
9266 Also, on SPARC 32-bit we cannot emit a sibling call when the
9267 current function returns a structure. This is because the "unimp
9268 after call" convention would cause the callee to return to the
9269 wrong place. The generic code already disallows cases where the
9270 function being called returns a structure.
9272 It may seem strange how this last case could occur. Usually there
9273 is code after the call which jumps to epilogue code which dumps the
9274 return value into the struct return area. That ought to invalidate
9275 the sibling call right? Well, in the C++ case we can end up passing
9276 the pointer to the struct return area to a constructor (which returns
9277 void) and then nothing else happens. Such a sibling call would look
9278 valid without the added check here.
9280 VxWorks PIC PLT entries require the global pointer to be initialized
9281 on entry. We therefore can't emit sibling calls to them. */
9283 sparc_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
9286 && flag_delayed_branch
9287 && (TARGET_ARCH64 || ! cfun->returns_struct)
9288 && !(TARGET_VXWORKS_RTP
9290 && !targetm.binds_local_p (decl)));
9293 /* libfunc renaming. */
9296 sparc_init_libfuncs (void)
9300 /* Use the subroutines that Sun's library provides for integer
9301 multiply and divide. The `*' prevents an underscore from
9302 being prepended by the compiler. .umul is a little faster
9304 set_optab_libfunc (smul_optab, SImode, "*.umul");
9305 set_optab_libfunc (sdiv_optab, SImode, "*.div");
9306 set_optab_libfunc (udiv_optab, SImode, "*.udiv");
9307 set_optab_libfunc (smod_optab, SImode, "*.rem");
9308 set_optab_libfunc (umod_optab, SImode, "*.urem");
9310 /* TFmode arithmetic. These names are part of the SPARC 32bit ABI. */
9311 set_optab_libfunc (add_optab, TFmode, "_Q_add");
9312 set_optab_libfunc (sub_optab, TFmode, "_Q_sub");
9313 set_optab_libfunc (neg_optab, TFmode, "_Q_neg");
9314 set_optab_libfunc (smul_optab, TFmode, "_Q_mul");
9315 set_optab_libfunc (sdiv_optab, TFmode, "_Q_div");
9317 /* We can define the TFmode sqrt optab only if TARGET_FPU. This
9318 is because with soft-float, the SFmode and DFmode sqrt
9319 instructions will be absent, and the compiler will notice and
9320 try to use the TFmode sqrt instruction for calls to the
9321 builtin function sqrt, but this fails. */
9323 set_optab_libfunc (sqrt_optab, TFmode, "_Q_sqrt");
9325 set_optab_libfunc (eq_optab, TFmode, "_Q_feq");
9326 set_optab_libfunc (ne_optab, TFmode, "_Q_fne");
9327 set_optab_libfunc (gt_optab, TFmode, "_Q_fgt");
9328 set_optab_libfunc (ge_optab, TFmode, "_Q_fge");
9329 set_optab_libfunc (lt_optab, TFmode, "_Q_flt");
9330 set_optab_libfunc (le_optab, TFmode, "_Q_fle");
9332 set_conv_libfunc (sext_optab, TFmode, SFmode, "_Q_stoq");
9333 set_conv_libfunc (sext_optab, TFmode, DFmode, "_Q_dtoq");
9334 set_conv_libfunc (trunc_optab, SFmode, TFmode, "_Q_qtos");
9335 set_conv_libfunc (trunc_optab, DFmode, TFmode, "_Q_qtod");
9337 set_conv_libfunc (sfix_optab, SImode, TFmode, "_Q_qtoi");
9338 set_conv_libfunc (ufix_optab, SImode, TFmode, "_Q_qtou");
9339 set_conv_libfunc (sfloat_optab, TFmode, SImode, "_Q_itoq");
9340 set_conv_libfunc (ufloat_optab, TFmode, SImode, "_Q_utoq");
9342 if (DITF_CONVERSION_LIBFUNCS)
9344 set_conv_libfunc (sfix_optab, DImode, TFmode, "_Q_qtoll");
9345 set_conv_libfunc (ufix_optab, DImode, TFmode, "_Q_qtoull");
9346 set_conv_libfunc (sfloat_optab, TFmode, DImode, "_Q_lltoq");
9347 set_conv_libfunc (ufloat_optab, TFmode, DImode, "_Q_ulltoq");
9350 if (SUN_CONVERSION_LIBFUNCS)
9352 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftoll");
9353 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoull");
9354 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtoll");
9355 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoull");
9360 /* In the SPARC 64bit ABI, SImode multiply and divide functions
9361 do not exist in the library. Make sure the compiler does not
9362 emit calls to them by accident. (It should always use the
9363 hardware instructions.) */
9364 set_optab_libfunc (smul_optab, SImode, 0);
9365 set_optab_libfunc (sdiv_optab, SImode, 0);
9366 set_optab_libfunc (udiv_optab, SImode, 0);
9367 set_optab_libfunc (smod_optab, SImode, 0);
9368 set_optab_libfunc (umod_optab, SImode, 0);
9370 if (SUN_INTEGER_MULTIPLY_64)
9372 set_optab_libfunc (smul_optab, DImode, "__mul64");
9373 set_optab_libfunc (sdiv_optab, DImode, "__div64");
9374 set_optab_libfunc (udiv_optab, DImode, "__udiv64");
9375 set_optab_libfunc (smod_optab, DImode, "__rem64");
9376 set_optab_libfunc (umod_optab, DImode, "__urem64");
9379 if (SUN_CONVERSION_LIBFUNCS)
9381 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftol");
9382 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoul");
9383 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtol");
9384 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoul");
9389 static tree def_builtin(const char *name, int code, tree type)
9391 return add_builtin_function(name, type, code, BUILT_IN_MD, NULL,
9395 static tree def_builtin_const(const char *name, int code, tree type)
9397 tree t = def_builtin(name, code, type);
9400 TREE_READONLY (t) = 1;
9405 /* Implement the TARGET_INIT_BUILTINS target hook.
9406 Create builtin functions for special SPARC instructions. */
9409 sparc_init_builtins (void)
9412 sparc_vis_init_builtins ();
9415 /* Create builtin functions for VIS 1.0 instructions. */
9418 sparc_vis_init_builtins (void)
9420 tree v4qi = build_vector_type (unsigned_intQI_type_node, 4);
9421 tree v8qi = build_vector_type (unsigned_intQI_type_node, 8);
9422 tree v4hi = build_vector_type (intHI_type_node, 4);
9423 tree v2hi = build_vector_type (intHI_type_node, 2);
9424 tree v2si = build_vector_type (intSI_type_node, 2);
9425 tree v1si = build_vector_type (intSI_type_node, 1);
9427 tree v4qi_ftype_v4hi = build_function_type_list (v4qi, v4hi, 0);
9428 tree v8qi_ftype_v2si_v8qi = build_function_type_list (v8qi, v2si, v8qi, 0);
9429 tree v2hi_ftype_v2si = build_function_type_list (v2hi, v2si, 0);
9430 tree v4hi_ftype_v4qi = build_function_type_list (v4hi, v4qi, 0);
9431 tree v8qi_ftype_v4qi_v4qi = build_function_type_list (v8qi, v4qi, v4qi, 0);
9432 tree v4hi_ftype_v4qi_v4hi = build_function_type_list (v4hi, v4qi, v4hi, 0);
9433 tree v4hi_ftype_v4qi_v2hi = build_function_type_list (v4hi, v4qi, v2hi, 0);
9434 tree v2si_ftype_v4qi_v2hi = build_function_type_list (v2si, v4qi, v2hi, 0);
9435 tree v4hi_ftype_v8qi_v4hi = build_function_type_list (v4hi, v8qi, v4hi, 0);
9436 tree v4hi_ftype_v4hi_v4hi = build_function_type_list (v4hi, v4hi, v4hi, 0);
9437 tree v2si_ftype_v2si_v2si = build_function_type_list (v2si, v2si, v2si, 0);
9438 tree v8qi_ftype_v8qi_v8qi = build_function_type_list (v8qi, v8qi, v8qi, 0);
9439 tree v2hi_ftype_v2hi_v2hi = build_function_type_list (v2hi, v2hi, v2hi, 0);
9440 tree v1si_ftype_v1si_v1si = build_function_type_list (v1si, v1si, v1si, 0);
9441 tree di_ftype_v8qi_v8qi_di = build_function_type_list (intDI_type_node,
9443 intDI_type_node, 0);
9444 tree di_ftype_v8qi_v8qi = build_function_type_list (intDI_type_node,
9446 tree si_ftype_v8qi_v8qi = build_function_type_list (intSI_type_node,
9448 tree di_ftype_di_di = build_function_type_list (intDI_type_node,
9450 intDI_type_node, 0);
9451 tree si_ftype_si_si = build_function_type_list (intSI_type_node,
9453 intSI_type_node, 0);
9454 tree ptr_ftype_ptr_si = build_function_type_list (ptr_type_node,
9456 intSI_type_node, 0);
9457 tree ptr_ftype_ptr_di = build_function_type_list (ptr_type_node,
9459 intDI_type_node, 0);
9460 tree si_ftype_ptr_ptr = build_function_type_list (intSI_type_node,
9463 tree di_ftype_ptr_ptr = build_function_type_list (intDI_type_node,
9466 tree si_ftype_v4hi_v4hi = build_function_type_list (intSI_type_node,
9468 tree si_ftype_v2si_v2si = build_function_type_list (intSI_type_node,
9470 tree di_ftype_v4hi_v4hi = build_function_type_list (intDI_type_node,
9472 tree di_ftype_v2si_v2si = build_function_type_list (intDI_type_node,
9474 tree void_ftype_di = build_function_type_list (void_type_node,
9475 intDI_type_node, 0);
9476 tree di_ftype_void = build_function_type_list (intDI_type_node,
9478 tree void_ftype_si = build_function_type_list (void_type_node,
9479 intSI_type_node, 0);
9480 tree sf_ftype_sf_sf = build_function_type_list (float_type_node,
9482 float_type_node, 0);
9483 tree df_ftype_df_df = build_function_type_list (double_type_node,
9485 double_type_node, 0);
9487 /* Packing and expanding vectors. */
9488 def_builtin ("__builtin_vis_fpack16", CODE_FOR_fpack16_vis,
9490 def_builtin ("__builtin_vis_fpack32", CODE_FOR_fpack32_vis,
9491 v8qi_ftype_v2si_v8qi);
9492 def_builtin ("__builtin_vis_fpackfix", CODE_FOR_fpackfix_vis,
9494 def_builtin_const ("__builtin_vis_fexpand", CODE_FOR_fexpand_vis,
9496 def_builtin_const ("__builtin_vis_fpmerge", CODE_FOR_fpmerge_vis,
9497 v8qi_ftype_v4qi_v4qi);
9499 /* Multiplications. */
9500 def_builtin_const ("__builtin_vis_fmul8x16", CODE_FOR_fmul8x16_vis,
9501 v4hi_ftype_v4qi_v4hi);
9502 def_builtin_const ("__builtin_vis_fmul8x16au", CODE_FOR_fmul8x16au_vis,
9503 v4hi_ftype_v4qi_v2hi);
9504 def_builtin_const ("__builtin_vis_fmul8x16al", CODE_FOR_fmul8x16al_vis,
9505 v4hi_ftype_v4qi_v2hi);
9506 def_builtin_const ("__builtin_vis_fmul8sux16", CODE_FOR_fmul8sux16_vis,
9507 v4hi_ftype_v8qi_v4hi);
9508 def_builtin_const ("__builtin_vis_fmul8ulx16", CODE_FOR_fmul8ulx16_vis,
9509 v4hi_ftype_v8qi_v4hi);
9510 def_builtin_const ("__builtin_vis_fmuld8sux16", CODE_FOR_fmuld8sux16_vis,
9511 v2si_ftype_v4qi_v2hi);
9512 def_builtin_const ("__builtin_vis_fmuld8ulx16", CODE_FOR_fmuld8ulx16_vis,
9513 v2si_ftype_v4qi_v2hi);
9515 /* Data aligning. */
9516 def_builtin ("__builtin_vis_faligndatav4hi", CODE_FOR_faligndatav4hi_vis,
9517 v4hi_ftype_v4hi_v4hi);
9518 def_builtin ("__builtin_vis_faligndatav8qi", CODE_FOR_faligndatav8qi_vis,
9519 v8qi_ftype_v8qi_v8qi);
9520 def_builtin ("__builtin_vis_faligndatav2si", CODE_FOR_faligndatav2si_vis,
9521 v2si_ftype_v2si_v2si);
9522 def_builtin ("__builtin_vis_faligndatadi", CODE_FOR_faligndatav1di_vis,
9525 def_builtin ("__builtin_vis_write_gsr", CODE_FOR_wrgsr_vis,
9527 def_builtin ("__builtin_vis_read_gsr", CODE_FOR_rdgsr_vis,
9532 def_builtin ("__builtin_vis_alignaddr", CODE_FOR_alignaddrdi_vis,
9534 def_builtin ("__builtin_vis_alignaddrl", CODE_FOR_alignaddrldi_vis,
9539 def_builtin ("__builtin_vis_alignaddr", CODE_FOR_alignaddrsi_vis,
9541 def_builtin ("__builtin_vis_alignaddrl", CODE_FOR_alignaddrlsi_vis,
9545 /* Pixel distance. */
9546 def_builtin_const ("__builtin_vis_pdist", CODE_FOR_pdist_vis,
9547 di_ftype_v8qi_v8qi_di);
9549 /* Edge handling. */
9552 def_builtin_const ("__builtin_vis_edge8", CODE_FOR_edge8di_vis,
9554 def_builtin_const ("__builtin_vis_edge8l", CODE_FOR_edge8ldi_vis,
9556 def_builtin_const ("__builtin_vis_edge16", CODE_FOR_edge16di_vis,
9558 def_builtin_const ("__builtin_vis_edge16l", CODE_FOR_edge16ldi_vis,
9560 def_builtin_const ("__builtin_vis_edge32", CODE_FOR_edge32di_vis,
9562 def_builtin_const ("__builtin_vis_edge32l", CODE_FOR_edge32ldi_vis,
9566 def_builtin_const ("__builtin_vis_edge8n", CODE_FOR_edge8ndi_vis,
9568 def_builtin_const ("__builtin_vis_edge8ln", CODE_FOR_edge8lndi_vis,
9570 def_builtin_const ("__builtin_vis_edge16n", CODE_FOR_edge16ndi_vis,
9572 def_builtin_const ("__builtin_vis_edge16ln", CODE_FOR_edge16lndi_vis,
9574 def_builtin_const ("__builtin_vis_edge32n", CODE_FOR_edge32ndi_vis,
9576 def_builtin_const ("__builtin_vis_edge32ln", CODE_FOR_edge32lndi_vis,
9582 def_builtin_const ("__builtin_vis_edge8", CODE_FOR_edge8si_vis,
9584 def_builtin_const ("__builtin_vis_edge8l", CODE_FOR_edge8lsi_vis,
9586 def_builtin_const ("__builtin_vis_edge16", CODE_FOR_edge16si_vis,
9588 def_builtin_const ("__builtin_vis_edge16l", CODE_FOR_edge16lsi_vis,
9590 def_builtin_const ("__builtin_vis_edge32", CODE_FOR_edge32si_vis,
9592 def_builtin_const ("__builtin_vis_edge32l", CODE_FOR_edge32lsi_vis,
9596 def_builtin_const ("__builtin_vis_edge8n", CODE_FOR_edge8nsi_vis,
9598 def_builtin_const ("__builtin_vis_edge8ln", CODE_FOR_edge8lnsi_vis,
9600 def_builtin_const ("__builtin_vis_edge16n", CODE_FOR_edge16nsi_vis,
9602 def_builtin_const ("__builtin_vis_edge16ln", CODE_FOR_edge16lnsi_vis,
9604 def_builtin_const ("__builtin_vis_edge32n", CODE_FOR_edge32nsi_vis,
9606 def_builtin_const ("__builtin_vis_edge32ln", CODE_FOR_edge32lnsi_vis,
9611 /* Pixel compare. */
9614 def_builtin_const ("__builtin_vis_fcmple16", CODE_FOR_fcmple16di_vis,
9615 di_ftype_v4hi_v4hi);
9616 def_builtin_const ("__builtin_vis_fcmple32", CODE_FOR_fcmple32di_vis,
9617 di_ftype_v2si_v2si);
9618 def_builtin_const ("__builtin_vis_fcmpne16", CODE_FOR_fcmpne16di_vis,
9619 di_ftype_v4hi_v4hi);
9620 def_builtin_const ("__builtin_vis_fcmpne32", CODE_FOR_fcmpne32di_vis,
9621 di_ftype_v2si_v2si);
9622 def_builtin_const ("__builtin_vis_fcmpgt16", CODE_FOR_fcmpgt16di_vis,
9623 di_ftype_v4hi_v4hi);
9624 def_builtin_const ("__builtin_vis_fcmpgt32", CODE_FOR_fcmpgt32di_vis,
9625 di_ftype_v2si_v2si);
9626 def_builtin_const ("__builtin_vis_fcmpeq16", CODE_FOR_fcmpeq16di_vis,
9627 di_ftype_v4hi_v4hi);
9628 def_builtin_const ("__builtin_vis_fcmpeq32", CODE_FOR_fcmpeq32di_vis,
9629 di_ftype_v2si_v2si);
9633 def_builtin_const ("__builtin_vis_fcmple16", CODE_FOR_fcmple16si_vis,
9634 si_ftype_v4hi_v4hi);
9635 def_builtin_const ("__builtin_vis_fcmple32", CODE_FOR_fcmple32si_vis,
9636 si_ftype_v2si_v2si);
9637 def_builtin_const ("__builtin_vis_fcmpne16", CODE_FOR_fcmpne16si_vis,
9638 si_ftype_v4hi_v4hi);
9639 def_builtin_const ("__builtin_vis_fcmpne32", CODE_FOR_fcmpne32si_vis,
9640 si_ftype_v2si_v2si);
9641 def_builtin_const ("__builtin_vis_fcmpgt16", CODE_FOR_fcmpgt16si_vis,
9642 si_ftype_v4hi_v4hi);
9643 def_builtin_const ("__builtin_vis_fcmpgt32", CODE_FOR_fcmpgt32si_vis,
9644 si_ftype_v2si_v2si);
9645 def_builtin_const ("__builtin_vis_fcmpeq16", CODE_FOR_fcmpeq16si_vis,
9646 si_ftype_v4hi_v4hi);
9647 def_builtin_const ("__builtin_vis_fcmpeq32", CODE_FOR_fcmpeq32si_vis,
9648 si_ftype_v2si_v2si);
9651 /* Addition and subtraction. */
9652 def_builtin_const ("__builtin_vis_fpadd16", CODE_FOR_addv4hi3,
9653 v4hi_ftype_v4hi_v4hi);
9654 def_builtin_const ("__builtin_vis_fpadd16s", CODE_FOR_addv2hi3,
9655 v2hi_ftype_v2hi_v2hi);
9656 def_builtin_const ("__builtin_vis_fpadd32", CODE_FOR_addv2si3,
9657 v2si_ftype_v2si_v2si);
9658 def_builtin_const ("__builtin_vis_fpadd32s", CODE_FOR_addv1si3,
9659 v1si_ftype_v1si_v1si);
9660 def_builtin_const ("__builtin_vis_fpsub16", CODE_FOR_subv4hi3,
9661 v4hi_ftype_v4hi_v4hi);
9662 def_builtin_const ("__builtin_vis_fpsub16s", CODE_FOR_subv2hi3,
9663 v2hi_ftype_v2hi_v2hi);
9664 def_builtin_const ("__builtin_vis_fpsub32", CODE_FOR_subv2si3,
9665 v2si_ftype_v2si_v2si);
9666 def_builtin_const ("__builtin_vis_fpsub32s", CODE_FOR_subv1si3,
9667 v1si_ftype_v1si_v1si);
9669 /* Three-dimensional array addressing. */
9672 def_builtin_const ("__builtin_vis_array8", CODE_FOR_array8di_vis,
9674 def_builtin_const ("__builtin_vis_array16", CODE_FOR_array16di_vis,
9676 def_builtin_const ("__builtin_vis_array32", CODE_FOR_array32di_vis,
9681 def_builtin_const ("__builtin_vis_array8", CODE_FOR_array8si_vis,
9683 def_builtin_const ("__builtin_vis_array16", CODE_FOR_array16si_vis,
9685 def_builtin_const ("__builtin_vis_array32", CODE_FOR_array32si_vis,
9691 /* Byte mask and shuffle */
9693 def_builtin ("__builtin_vis_bmask", CODE_FOR_bmaskdi_vis,
9696 def_builtin ("__builtin_vis_bmask", CODE_FOR_bmasksi_vis,
9698 def_builtin ("__builtin_vis_bshufflev4hi", CODE_FOR_bshufflev4hi_vis,
9699 v4hi_ftype_v4hi_v4hi);
9700 def_builtin ("__builtin_vis_bshufflev8qi", CODE_FOR_bshufflev8qi_vis,
9701 v8qi_ftype_v8qi_v8qi);
9702 def_builtin ("__builtin_vis_bshufflev2si", CODE_FOR_bshufflev2si_vis,
9703 v2si_ftype_v2si_v2si);
9704 def_builtin ("__builtin_vis_bshuffledi", CODE_FOR_bshufflev1di_vis,
9712 def_builtin ("__builtin_vis_cmask8", CODE_FOR_cmask8di_vis,
9714 def_builtin ("__builtin_vis_cmask16", CODE_FOR_cmask16di_vis,
9716 def_builtin ("__builtin_vis_cmask32", CODE_FOR_cmask32di_vis,
9721 def_builtin ("__builtin_vis_cmask8", CODE_FOR_cmask8si_vis,
9723 def_builtin ("__builtin_vis_cmask16", CODE_FOR_cmask16si_vis,
9725 def_builtin ("__builtin_vis_cmask32", CODE_FOR_cmask32si_vis,
9729 def_builtin_const ("__builtin_vis_fchksm16", CODE_FOR_fchksm16_vis,
9730 v4hi_ftype_v4hi_v4hi);
9732 def_builtin_const ("__builtin_vis_fsll16", CODE_FOR_vashlv4hi3,
9733 v4hi_ftype_v4hi_v4hi);
9734 def_builtin_const ("__builtin_vis_fslas16", CODE_FOR_vssashlv4hi3,
9735 v4hi_ftype_v4hi_v4hi);
9736 def_builtin_const ("__builtin_vis_fsrl16", CODE_FOR_vlshrv4hi3,
9737 v4hi_ftype_v4hi_v4hi);
9738 def_builtin_const ("__builtin_vis_fsra16", CODE_FOR_vashrv4hi3,
9739 v4hi_ftype_v4hi_v4hi);
9740 def_builtin_const ("__builtin_vis_fsll32", CODE_FOR_vashlv2si3,
9741 v2si_ftype_v2si_v2si);
9742 def_builtin_const ("__builtin_vis_fslas32", CODE_FOR_vssashlv2si3,
9743 v2si_ftype_v2si_v2si);
9744 def_builtin_const ("__builtin_vis_fsrl32", CODE_FOR_vlshrv2si3,
9745 v2si_ftype_v2si_v2si);
9746 def_builtin_const ("__builtin_vis_fsra32", CODE_FOR_vashrv2si3,
9747 v2si_ftype_v2si_v2si);
9750 def_builtin_const ("__builtin_vis_pdistn", CODE_FOR_pdistndi_vis,
9751 di_ftype_v8qi_v8qi);
9753 def_builtin_const ("__builtin_vis_pdistn", CODE_FOR_pdistnsi_vis,
9754 si_ftype_v8qi_v8qi);
9756 def_builtin_const ("__builtin_vis_fmean16", CODE_FOR_fmean16_vis,
9757 v4hi_ftype_v4hi_v4hi);
9758 def_builtin_const ("__builtin_vis_fpadd64", CODE_FOR_fpadd64_vis,
9760 def_builtin_const ("__builtin_vis_fpsub64", CODE_FOR_fpsub64_vis,
9763 def_builtin_const ("__builtin_vis_fpadds16", CODE_FOR_ssaddv4hi3,
9764 v4hi_ftype_v4hi_v4hi);
9765 def_builtin_const ("__builtin_vis_fpadds16s", CODE_FOR_ssaddv2hi3,
9766 v2hi_ftype_v2hi_v2hi);
9767 def_builtin_const ("__builtin_vis_fpsubs16", CODE_FOR_sssubv4hi3,
9768 v4hi_ftype_v4hi_v4hi);
9769 def_builtin_const ("__builtin_vis_fpsubs16s", CODE_FOR_sssubv2hi3,
9770 v2hi_ftype_v2hi_v2hi);
9771 def_builtin_const ("__builtin_vis_fpadds32", CODE_FOR_ssaddv2si3,
9772 v2si_ftype_v2si_v2si);
9773 def_builtin_const ("__builtin_vis_fpadds32s", CODE_FOR_ssaddv1si3,
9774 v1si_ftype_v1si_v1si);
9775 def_builtin_const ("__builtin_vis_fpsubs32", CODE_FOR_sssubv2si3,
9776 v2si_ftype_v2si_v2si);
9777 def_builtin_const ("__builtin_vis_fpsubs32s", CODE_FOR_sssubv1si3,
9778 v1si_ftype_v1si_v1si);
9782 def_builtin_const ("__builtin_vis_fucmple8", CODE_FOR_fucmple8di_vis,
9783 di_ftype_v8qi_v8qi);
9784 def_builtin_const ("__builtin_vis_fucmpne8", CODE_FOR_fucmpne8di_vis,
9785 di_ftype_v8qi_v8qi);
9786 def_builtin_const ("__builtin_vis_fucmpgt8", CODE_FOR_fucmpgt8di_vis,
9787 di_ftype_v8qi_v8qi);
9788 def_builtin_const ("__builtin_vis_fucmpeq8", CODE_FOR_fucmpeq8di_vis,
9789 di_ftype_v8qi_v8qi);
9793 def_builtin_const ("__builtin_vis_fucmple8", CODE_FOR_fucmple8si_vis,
9794 si_ftype_v8qi_v8qi);
9795 def_builtin_const ("__builtin_vis_fucmpne8", CODE_FOR_fucmpne8si_vis,
9796 si_ftype_v8qi_v8qi);
9797 def_builtin_const ("__builtin_vis_fucmpgt8", CODE_FOR_fucmpgt8si_vis,
9798 si_ftype_v8qi_v8qi);
9799 def_builtin_const ("__builtin_vis_fucmpeq8", CODE_FOR_fucmpeq8si_vis,
9800 si_ftype_v8qi_v8qi);
9803 def_builtin_const ("__builtin_vis_fhadds", CODE_FOR_fhaddsf_vis,
9805 def_builtin_const ("__builtin_vis_fhaddd", CODE_FOR_fhadddf_vis,
9807 def_builtin_const ("__builtin_vis_fhsubs", CODE_FOR_fhsubsf_vis,
9809 def_builtin_const ("__builtin_vis_fhsubd", CODE_FOR_fhsubdf_vis,
9811 def_builtin_const ("__builtin_vis_fnhadds", CODE_FOR_fnhaddsf_vis,
9813 def_builtin_const ("__builtin_vis_fnhaddd", CODE_FOR_fnhadddf_vis,
9816 def_builtin_const ("__builtin_vis_umulxhi", CODE_FOR_umulxhi_vis,
9818 def_builtin_const ("__builtin_vis_xmulx", CODE_FOR_xmulx_vis,
9820 def_builtin_const ("__builtin_vis_xmulxhi", CODE_FOR_xmulxhi_vis,
9825 /* Handle TARGET_EXPAND_BUILTIN target hook.
9826 Expand builtin functions for sparc intrinsics. */
9829 sparc_expand_builtin (tree exp, rtx target,
9830 rtx subtarget ATTRIBUTE_UNUSED,
9831 enum machine_mode tmode ATTRIBUTE_UNUSED,
9832 int ignore ATTRIBUTE_UNUSED)
9835 call_expr_arg_iterator iter;
9836 tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0);
9837 unsigned int icode = DECL_FUNCTION_CODE (fndecl);
9842 nonvoid = TREE_TYPE (TREE_TYPE (fndecl)) != void_type_node;
9846 enum machine_mode tmode = insn_data[icode].operand[0].mode;
9848 || GET_MODE (target) != tmode
9849 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
9850 op[0] = gen_reg_rtx (tmode);
9854 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
9856 const struct insn_operand_data *insn_op;
9859 if (arg == error_mark_node)
9863 idx = arg_count - !nonvoid;
9864 insn_op = &insn_data[icode].operand[idx];
9865 op[arg_count] = expand_normal (arg);
9867 if (insn_op->mode == V1DImode
9868 && GET_MODE (op[arg_count]) == DImode)
9869 op[arg_count] = gen_lowpart (V1DImode, op[arg_count]);
9870 else if (insn_op->mode == V1SImode
9871 && GET_MODE (op[arg_count]) == SImode)
9872 op[arg_count] = gen_lowpart (V1SImode, op[arg_count]);
9874 if (! (*insn_data[icode].operand[idx].predicate) (op[arg_count],
9876 op[arg_count] = copy_to_mode_reg (insn_op->mode, op[arg_count]);
9882 pat = GEN_FCN (icode) (op[0]);
9886 pat = GEN_FCN (icode) (op[0], op[1]);
9888 pat = GEN_FCN (icode) (op[1]);
9891 pat = GEN_FCN (icode) (op[0], op[1], op[2]);
9894 pat = GEN_FCN (icode) (op[0], op[1], op[2], op[3]);
9912 sparc_vis_mul8x16 (int e8, int e16)
9914 return (e8 * e16 + 128) / 256;
9917 /* Multiply the vector elements in ELTS0 to the elements in ELTS1 as specified
9918 by FNCODE. All of the elements in ELTS0 and ELTS1 lists must be integer
9919 constants. A tree list with the results of the multiplications is returned,
9920 and each element in the list is of INNER_TYPE. */
9923 sparc_handle_vis_mul8x16 (int fncode, tree inner_type, tree elts0, tree elts1)
9925 tree n_elts = NULL_TREE;
9930 case CODE_FOR_fmul8x16_vis:
9931 for (; elts0 && elts1;
9932 elts0 = TREE_CHAIN (elts0), elts1 = TREE_CHAIN (elts1))
9935 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (TREE_VALUE (elts0)),
9936 TREE_INT_CST_LOW (TREE_VALUE (elts1)));
9937 n_elts = tree_cons (NULL_TREE,
9938 build_int_cst (inner_type, val),
9943 case CODE_FOR_fmul8x16au_vis:
9944 scale = TREE_INT_CST_LOW (TREE_VALUE (elts1));
9946 for (; elts0; elts0 = TREE_CHAIN (elts0))
9949 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (TREE_VALUE (elts0)),
9951 n_elts = tree_cons (NULL_TREE,
9952 build_int_cst (inner_type, val),
9957 case CODE_FOR_fmul8x16al_vis:
9958 scale = TREE_INT_CST_LOW (TREE_VALUE (TREE_CHAIN (elts1)));
9960 for (; elts0; elts0 = TREE_CHAIN (elts0))
9963 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (TREE_VALUE (elts0)),
9965 n_elts = tree_cons (NULL_TREE,
9966 build_int_cst (inner_type, val),
9975 return nreverse (n_elts);
9978 /* Handle TARGET_FOLD_BUILTIN target hook.
9979 Fold builtin functions for SPARC intrinsics. If IGNORE is true the
9980 result of the function call is ignored. NULL_TREE is returned if the
9981 function could not be folded. */
9984 sparc_fold_builtin (tree fndecl, int n_args ATTRIBUTE_UNUSED,
9985 tree *args, bool ignore)
9987 tree arg0, arg1, arg2;
9988 tree rtype = TREE_TYPE (TREE_TYPE (fndecl));
9989 enum insn_code icode = (enum insn_code) DECL_FUNCTION_CODE (fndecl);
9993 /* Note that a switch statement instead of the sequence of tests would
9994 be incorrect as many of the CODE_FOR values could be CODE_FOR_nothing
9995 and that would yield multiple alternatives with identical values. */
9996 if (icode == CODE_FOR_alignaddrsi_vis
9997 || icode == CODE_FOR_alignaddrdi_vis
9998 || icode == CODE_FOR_wrgsr_vis
9999 || icode == CODE_FOR_bmasksi_vis
10000 || icode == CODE_FOR_bmaskdi_vis
10001 || icode == CODE_FOR_cmask8si_vis
10002 || icode == CODE_FOR_cmask8di_vis
10003 || icode == CODE_FOR_cmask16si_vis
10004 || icode == CODE_FOR_cmask16di_vis
10005 || icode == CODE_FOR_cmask32si_vis
10006 || icode == CODE_FOR_cmask32di_vis)
10009 return build_zero_cst (rtype);
10014 case CODE_FOR_fexpand_vis:
10018 if (TREE_CODE (arg0) == VECTOR_CST)
10020 tree inner_type = TREE_TYPE (rtype);
10021 tree elts = TREE_VECTOR_CST_ELTS (arg0);
10022 tree n_elts = NULL_TREE;
10024 for (; elts; elts = TREE_CHAIN (elts))
10026 unsigned int val = TREE_INT_CST_LOW (TREE_VALUE (elts)) << 4;
10027 n_elts = tree_cons (NULL_TREE,
10028 build_int_cst (inner_type, val),
10031 return build_vector (rtype, nreverse (n_elts));
10035 case CODE_FOR_fmul8x16_vis:
10036 case CODE_FOR_fmul8x16au_vis:
10037 case CODE_FOR_fmul8x16al_vis:
10043 if (TREE_CODE (arg0) == VECTOR_CST && TREE_CODE (arg1) == VECTOR_CST)
10045 tree inner_type = TREE_TYPE (rtype);
10046 tree elts0 = TREE_VECTOR_CST_ELTS (arg0);
10047 tree elts1 = TREE_VECTOR_CST_ELTS (arg1);
10048 tree n_elts = sparc_handle_vis_mul8x16 (icode, inner_type, elts0,
10051 return build_vector (rtype, n_elts);
10055 case CODE_FOR_fpmerge_vis:
10061 if (TREE_CODE (arg0) == VECTOR_CST && TREE_CODE (arg1) == VECTOR_CST)
10063 tree elts0 = TREE_VECTOR_CST_ELTS (arg0);
10064 tree elts1 = TREE_VECTOR_CST_ELTS (arg1);
10065 tree n_elts = NULL_TREE;
10067 for (; elts0 && elts1;
10068 elts0 = TREE_CHAIN (elts0), elts1 = TREE_CHAIN (elts1))
10070 n_elts = tree_cons (NULL_TREE, TREE_VALUE (elts0), n_elts);
10071 n_elts = tree_cons (NULL_TREE, TREE_VALUE (elts1), n_elts);
10074 return build_vector (rtype, nreverse (n_elts));
10078 case CODE_FOR_pdist_vis:
10086 if (TREE_CODE (arg0) == VECTOR_CST
10087 && TREE_CODE (arg1) == VECTOR_CST
10088 && TREE_CODE (arg2) == INTEGER_CST)
10091 unsigned HOST_WIDE_INT low = TREE_INT_CST_LOW (arg2);
10092 HOST_WIDE_INT high = TREE_INT_CST_HIGH (arg2);
10093 tree elts0 = TREE_VECTOR_CST_ELTS (arg0);
10094 tree elts1 = TREE_VECTOR_CST_ELTS (arg1);
10096 for (; elts0 && elts1;
10097 elts0 = TREE_CHAIN (elts0), elts1 = TREE_CHAIN (elts1))
10099 unsigned HOST_WIDE_INT
10100 low0 = TREE_INT_CST_LOW (TREE_VALUE (elts0)),
10101 low1 = TREE_INT_CST_LOW (TREE_VALUE (elts1));
10102 HOST_WIDE_INT high0 = TREE_INT_CST_HIGH (TREE_VALUE (elts0));
10103 HOST_WIDE_INT high1 = TREE_INT_CST_HIGH (TREE_VALUE (elts1));
10105 unsigned HOST_WIDE_INT l;
10108 overflow |= neg_double (low1, high1, &l, &h);
10109 overflow |= add_double (low0, high0, l, h, &l, &h);
10111 overflow |= neg_double (l, h, &l, &h);
10113 overflow |= add_double (low, high, l, h, &low, &high);
10116 gcc_assert (overflow == 0);
10118 return build_int_cst_wide (rtype, low, high);
10128 /* ??? This duplicates information provided to the compiler by the
10129 ??? scheduler description. Some day, teach genautomata to output
10130 ??? the latencies and then CSE will just use that. */
10133 sparc_rtx_costs (rtx x, int code, int outer_code, int opno ATTRIBUTE_UNUSED,
10134 int *total, bool speed ATTRIBUTE_UNUSED)
10136 enum machine_mode mode = GET_MODE (x);
10137 bool float_mode_p = FLOAT_MODE_P (mode);
10142 if (INTVAL (x) < 0x1000 && INTVAL (x) >= -0x1000)
10160 if (GET_MODE (x) == VOIDmode
10161 && ((CONST_DOUBLE_HIGH (x) == 0
10162 && CONST_DOUBLE_LOW (x) < 0x1000)
10163 || (CONST_DOUBLE_HIGH (x) == -1
10164 && CONST_DOUBLE_LOW (x) < 0
10165 && CONST_DOUBLE_LOW (x) >= -0x1000)))
10172 /* If outer-code was a sign or zero extension, a cost
10173 of COSTS_N_INSNS (1) was already added in. This is
10174 why we are subtracting it back out. */
10175 if (outer_code == ZERO_EXTEND)
10177 *total = sparc_costs->int_zload - COSTS_N_INSNS (1);
10179 else if (outer_code == SIGN_EXTEND)
10181 *total = sparc_costs->int_sload - COSTS_N_INSNS (1);
10183 else if (float_mode_p)
10185 *total = sparc_costs->float_load;
10189 *total = sparc_costs->int_load;
10197 *total = sparc_costs->float_plusminus;
10199 *total = COSTS_N_INSNS (1);
10206 gcc_assert (float_mode_p);
10207 *total = sparc_costs->float_mul;
10210 if (GET_CODE (sub) == NEG)
10211 sub = XEXP (sub, 0);
10212 *total += rtx_cost (sub, FMA, 0, speed);
10215 if (GET_CODE (sub) == NEG)
10216 sub = XEXP (sub, 0);
10217 *total += rtx_cost (sub, FMA, 2, speed);
10223 *total = sparc_costs->float_mul;
10224 else if (! TARGET_HARD_MUL)
10225 *total = COSTS_N_INSNS (25);
10231 if (sparc_costs->int_mul_bit_factor)
10235 if (GET_CODE (XEXP (x, 1)) == CONST_INT)
10237 unsigned HOST_WIDE_INT value = INTVAL (XEXP (x, 1));
10238 for (nbits = 0; value != 0; value &= value - 1)
10241 else if (GET_CODE (XEXP (x, 1)) == CONST_DOUBLE
10242 && GET_MODE (XEXP (x, 1)) == VOIDmode)
10244 rtx x1 = XEXP (x, 1);
10245 unsigned HOST_WIDE_INT value1 = CONST_DOUBLE_LOW (x1);
10246 unsigned HOST_WIDE_INT value2 = CONST_DOUBLE_HIGH (x1);
10248 for (nbits = 0; value1 != 0; value1 &= value1 - 1)
10250 for (; value2 != 0; value2 &= value2 - 1)
10258 bit_cost = (nbits - 3) / sparc_costs->int_mul_bit_factor;
10259 bit_cost = COSTS_N_INSNS (bit_cost);
10262 if (mode == DImode)
10263 *total = sparc_costs->int_mulX + bit_cost;
10265 *total = sparc_costs->int_mul + bit_cost;
10272 *total = COSTS_N_INSNS (1) + sparc_costs->shift_penalty;
10281 if (mode == DFmode)
10282 *total = sparc_costs->float_div_df;
10284 *total = sparc_costs->float_div_sf;
10288 if (mode == DImode)
10289 *total = sparc_costs->int_divX;
10291 *total = sparc_costs->int_div;
10296 if (! float_mode_p)
10298 *total = COSTS_N_INSNS (1);
10305 case UNSIGNED_FLOAT:
10309 case FLOAT_TRUNCATE:
10310 *total = sparc_costs->float_move;
10314 if (mode == DFmode)
10315 *total = sparc_costs->float_sqrt_df;
10317 *total = sparc_costs->float_sqrt_sf;
10322 *total = sparc_costs->float_cmp;
10324 *total = COSTS_N_INSNS (1);
10329 *total = sparc_costs->float_cmove;
10331 *total = sparc_costs->int_cmove;
10335 /* Handle the NAND vector patterns. */
10336 if (sparc_vector_mode_supported_p (GET_MODE (x))
10337 && GET_CODE (XEXP (x, 0)) == NOT
10338 && GET_CODE (XEXP (x, 1)) == NOT)
10340 *total = COSTS_N_INSNS (1);
10351 /* Return true if CLASS is either GENERAL_REGS or I64_REGS. */
10354 general_or_i64_p (reg_class_t rclass)
10356 return (rclass == GENERAL_REGS || rclass == I64_REGS);
10359 /* Implement TARGET_REGISTER_MOVE_COST. */
10362 sparc_register_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED,
10363 reg_class_t from, reg_class_t to)
10365 bool need_memory = false;
10367 if (from == FPCC_REGS || to == FPCC_REGS)
10368 need_memory = true;
10369 else if ((FP_REG_CLASS_P (from) && general_or_i64_p (to))
10370 || (general_or_i64_p (from) && FP_REG_CLASS_P (to)))
10374 int size = GET_MODE_SIZE (mode);
10375 if (size == 8 || size == 4)
10377 if (! TARGET_ARCH32 || size == 4)
10383 need_memory = true;
10388 if (sparc_cpu == PROCESSOR_ULTRASPARC
10389 || sparc_cpu == PROCESSOR_ULTRASPARC3
10390 || sparc_cpu == PROCESSOR_NIAGARA
10391 || sparc_cpu == PROCESSOR_NIAGARA2
10392 || sparc_cpu == PROCESSOR_NIAGARA3
10393 || sparc_cpu == PROCESSOR_NIAGARA4)
10402 /* Emit the sequence of insns SEQ while preserving the registers REG and REG2.
10403 This is achieved by means of a manual dynamic stack space allocation in
10404 the current frame. We make the assumption that SEQ doesn't contain any
10405 function calls, with the possible exception of calls to the GOT helper. */
10408 emit_and_preserve (rtx seq, rtx reg, rtx reg2)
10410 /* We must preserve the lowest 16 words for the register save area. */
10411 HOST_WIDE_INT offset = 16*UNITS_PER_WORD;
10412 /* We really need only 2 words of fresh stack space. */
10413 HOST_WIDE_INT size = SPARC_STACK_ALIGN (offset + 2*UNITS_PER_WORD);
10416 = gen_rtx_MEM (word_mode, plus_constant (stack_pointer_rtx,
10417 SPARC_STACK_BIAS + offset));
10419 emit_insn (gen_stack_pointer_dec (GEN_INT (size)));
10420 emit_insn (gen_rtx_SET (VOIDmode, slot, reg));
10422 emit_insn (gen_rtx_SET (VOIDmode,
10423 adjust_address (slot, word_mode, UNITS_PER_WORD),
10427 emit_insn (gen_rtx_SET (VOIDmode,
10429 adjust_address (slot, word_mode, UNITS_PER_WORD)));
10430 emit_insn (gen_rtx_SET (VOIDmode, reg, slot));
10431 emit_insn (gen_stack_pointer_inc (GEN_INT (size)));
10434 /* Output the assembler code for a thunk function. THUNK_DECL is the
10435 declaration for the thunk function itself, FUNCTION is the decl for
10436 the target function. DELTA is an immediate constant offset to be
10437 added to THIS. If VCALL_OFFSET is nonzero, the word at address
10438 (*THIS + VCALL_OFFSET) should be additionally added to THIS. */
10441 sparc_output_mi_thunk (FILE *file, tree thunk_fndecl ATTRIBUTE_UNUSED,
10442 HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset,
10445 rtx this_rtx, insn, funexp;
10446 unsigned int int_arg_first;
10448 reload_completed = 1;
10449 epilogue_completed = 1;
10451 emit_note (NOTE_INSN_PROLOGUE_END);
10455 sparc_leaf_function_p = 1;
10457 int_arg_first = SPARC_OUTGOING_INT_ARG_FIRST;
10459 else if (flag_delayed_branch)
10461 /* We will emit a regular sibcall below, so we need to instruct
10462 output_sibcall that we are in a leaf function. */
10463 sparc_leaf_function_p = current_function_uses_only_leaf_regs = 1;
10465 /* This will cause final.c to invoke leaf_renumber_regs so we
10466 must behave as if we were in a not-yet-leafified function. */
10467 int_arg_first = SPARC_INCOMING_INT_ARG_FIRST;
10471 /* We will emit the sibcall manually below, so we will need to
10472 manually spill non-leaf registers. */
10473 sparc_leaf_function_p = current_function_uses_only_leaf_regs = 0;
10475 /* We really are in a leaf function. */
10476 int_arg_first = SPARC_OUTGOING_INT_ARG_FIRST;
10479 /* Find the "this" pointer. Normally in %o0, but in ARCH64 if the function
10480 returns a structure, the structure return pointer is there instead. */
10482 && aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function))
10483 this_rtx = gen_rtx_REG (Pmode, int_arg_first + 1);
10485 this_rtx = gen_rtx_REG (Pmode, int_arg_first);
10487 /* Add DELTA. When possible use a plain add, otherwise load it into
10488 a register first. */
10491 rtx delta_rtx = GEN_INT (delta);
10493 if (! SPARC_SIMM13_P (delta))
10495 rtx scratch = gen_rtx_REG (Pmode, 1);
10496 emit_move_insn (scratch, delta_rtx);
10497 delta_rtx = scratch;
10500 /* THIS_RTX += DELTA. */
10501 emit_insn (gen_add2_insn (this_rtx, delta_rtx));
10504 /* Add the word at address (*THIS_RTX + VCALL_OFFSET). */
10507 rtx vcall_offset_rtx = GEN_INT (vcall_offset);
10508 rtx scratch = gen_rtx_REG (Pmode, 1);
10510 gcc_assert (vcall_offset < 0);
10512 /* SCRATCH = *THIS_RTX. */
10513 emit_move_insn (scratch, gen_rtx_MEM (Pmode, this_rtx));
10515 /* Prepare for adding VCALL_OFFSET. The difficulty is that we
10516 may not have any available scratch register at this point. */
10517 if (SPARC_SIMM13_P (vcall_offset))
10519 /* This is the case if ARCH64 (unless -ffixed-g5 is passed). */
10520 else if (! fixed_regs[5]
10521 /* The below sequence is made up of at least 2 insns,
10522 while the default method may need only one. */
10523 && vcall_offset < -8192)
10525 rtx scratch2 = gen_rtx_REG (Pmode, 5);
10526 emit_move_insn (scratch2, vcall_offset_rtx);
10527 vcall_offset_rtx = scratch2;
10531 rtx increment = GEN_INT (-4096);
10533 /* VCALL_OFFSET is a negative number whose typical range can be
10534 estimated as -32768..0 in 32-bit mode. In almost all cases
10535 it is therefore cheaper to emit multiple add insns than
10536 spilling and loading the constant into a register (at least
10538 while (! SPARC_SIMM13_P (vcall_offset))
10540 emit_insn (gen_add2_insn (scratch, increment));
10541 vcall_offset += 4096;
10543 vcall_offset_rtx = GEN_INT (vcall_offset); /* cannot be 0 */
10546 /* SCRATCH = *(*THIS_RTX + VCALL_OFFSET). */
10547 emit_move_insn (scratch, gen_rtx_MEM (Pmode,
10548 gen_rtx_PLUS (Pmode,
10550 vcall_offset_rtx)));
10552 /* THIS_RTX += *(*THIS_RTX + VCALL_OFFSET). */
10553 emit_insn (gen_add2_insn (this_rtx, scratch));
10556 /* Generate a tail call to the target function. */
10557 if (! TREE_USED (function))
10559 assemble_external (function);
10560 TREE_USED (function) = 1;
10562 funexp = XEXP (DECL_RTL (function), 0);
10564 if (flag_delayed_branch)
10566 funexp = gen_rtx_MEM (FUNCTION_MODE, funexp);
10567 insn = emit_call_insn (gen_sibcall (funexp));
10568 SIBLING_CALL_P (insn) = 1;
10572 /* The hoops we have to jump through in order to generate a sibcall
10573 without using delay slots... */
10574 rtx spill_reg, seq, scratch = gen_rtx_REG (Pmode, 1);
10578 spill_reg = gen_rtx_REG (word_mode, 15); /* %o7 */
10580 load_got_register (); /* clobbers %o7 */
10581 scratch = sparc_legitimize_pic_address (funexp, scratch);
10582 seq = get_insns ();
10584 emit_and_preserve (seq, spill_reg, pic_offset_table_rtx);
10586 else if (TARGET_ARCH32)
10588 emit_insn (gen_rtx_SET (VOIDmode,
10590 gen_rtx_HIGH (SImode, funexp)));
10591 emit_insn (gen_rtx_SET (VOIDmode,
10593 gen_rtx_LO_SUM (SImode, scratch, funexp)));
10595 else /* TARGET_ARCH64 */
10597 switch (sparc_cmodel)
10601 /* The destination can serve as a temporary. */
10602 sparc_emit_set_symbolic_const64 (scratch, funexp, scratch);
10607 /* The destination cannot serve as a temporary. */
10608 spill_reg = gen_rtx_REG (DImode, 15); /* %o7 */
10610 sparc_emit_set_symbolic_const64 (scratch, funexp, spill_reg);
10611 seq = get_insns ();
10613 emit_and_preserve (seq, spill_reg, 0);
10617 gcc_unreachable ();
10621 emit_jump_insn (gen_indirect_jump (scratch));
10626 /* Run just enough of rest_of_compilation to get the insns emitted.
10627 There's not really enough bulk here to make other passes such as
10628 instruction scheduling worth while. Note that use_thunk calls
10629 assemble_start_function and assemble_end_function. */
10630 insn = get_insns ();
10631 insn_locators_alloc ();
10632 shorten_branches (insn);
10633 final_start_function (insn, file, 1);
10634 final (insn, file, 1);
10635 final_end_function ();
10637 reload_completed = 0;
10638 epilogue_completed = 0;
10641 /* Return true if sparc_output_mi_thunk would be able to output the
10642 assembler code for the thunk function specified by the arguments
10643 it is passed, and false otherwise. */
10645 sparc_can_output_mi_thunk (const_tree thunk_fndecl ATTRIBUTE_UNUSED,
10646 HOST_WIDE_INT delta ATTRIBUTE_UNUSED,
10647 HOST_WIDE_INT vcall_offset,
10648 const_tree function ATTRIBUTE_UNUSED)
10650 /* Bound the loop used in the default method above. */
10651 return (vcall_offset >= -32768 || ! fixed_regs[5]);
10654 /* We use the machine specific reorg pass to enable workarounds for errata. */
10661 /* The only erratum we handle for now is that of the AT697F processor. */
10662 if (!sparc_fix_at697f)
10665 /* We need to have the (essentially) final form of the insn stream in order
10666 to properly detect the various hazards. Run delay slot scheduling. */
10667 if (optimize > 0 && flag_delayed_branch)
10668 dbr_schedule (get_insns ());
10670 /* Now look for specific patterns in the insn stream. */
10671 for (insn = get_insns (); insn; insn = next)
10673 bool insert_nop = false;
10676 /* Look for a single-word load into an odd-numbered FP register. */
10677 if (NONJUMP_INSN_P (insn)
10678 && (set = single_set (insn)) != NULL_RTX
10679 && GET_MODE_SIZE (GET_MODE (SET_SRC (set))) == 4
10680 && MEM_P (SET_SRC (set))
10681 && REG_P (SET_DEST (set))
10682 && REGNO (SET_DEST (set)) > 31
10683 && REGNO (SET_DEST (set)) % 2 != 0)
10685 /* The wrong dependency is on the enclosing double register. */
10686 unsigned int x = REGNO (SET_DEST (set)) - 1;
10687 unsigned int src1, src2, dest;
10690 /* If the insn has a delay slot, then it cannot be problematic. */
10691 next = next_active_insn (insn);
10692 if (NONJUMP_INSN_P (next) && GET_CODE (PATTERN (next)) == SEQUENCE)
10696 extract_insn (next);
10697 code = INSN_CODE (next);
10702 case CODE_FOR_adddf3:
10703 case CODE_FOR_subdf3:
10704 case CODE_FOR_muldf3:
10705 case CODE_FOR_divdf3:
10706 dest = REGNO (recog_data.operand[0]);
10707 src1 = REGNO (recog_data.operand[1]);
10708 src2 = REGNO (recog_data.operand[2]);
10712 ld [address], %fx+1
10713 FPOPd %f{x,y}, %f{y,x}, %f{x,y} */
10714 if ((src1 == x || src2 == x)
10715 && (dest == src1 || dest == src2))
10721 ld [address], %fx+1
10722 FPOPd %fx, %fx, %fx */
10725 && (code == CODE_FOR_adddf3 || code == CODE_FOR_muldf3))
10730 case CODE_FOR_sqrtdf2:
10731 dest = REGNO (recog_data.operand[0]);
10732 src1 = REGNO (recog_data.operand[1]);
10734 ld [address], %fx+1
10736 if (src1 == x && dest == src1)
10745 next = NEXT_INSN (insn);
10748 emit_insn_after (gen_nop (), insn);
10752 /* How to allocate a 'struct machine_function'. */
10754 static struct machine_function *
10755 sparc_init_machine_status (void)
10757 return ggc_alloc_cleared_machine_function ();
10760 /* Locate some local-dynamic symbol still in use by this function
10761 so that we can print its name in local-dynamic base patterns. */
10763 static const char *
10764 get_some_local_dynamic_name (void)
10768 if (cfun->machine->some_ld_name)
10769 return cfun->machine->some_ld_name;
10771 for (insn = get_insns (); insn ; insn = NEXT_INSN (insn))
10773 && for_each_rtx (&PATTERN (insn), get_some_local_dynamic_name_1, 0))
10774 return cfun->machine->some_ld_name;
10776 gcc_unreachable ();
10780 get_some_local_dynamic_name_1 (rtx *px, void *data ATTRIBUTE_UNUSED)
10785 && GET_CODE (x) == SYMBOL_REF
10786 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_DYNAMIC)
10788 cfun->machine->some_ld_name = XSTR (x, 0);
10795 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
10796 We need to emit DTP-relative relocations. */
10799 sparc_output_dwarf_dtprel (FILE *file, int size, rtx x)
10804 fputs ("\t.word\t%r_tls_dtpoff32(", file);
10807 fputs ("\t.xword\t%r_tls_dtpoff64(", file);
10810 gcc_unreachable ();
10812 output_addr_const (file, x);
10816 /* Do whatever processing is required at the end of a file. */
10819 sparc_file_end (void)
10821 /* If we need to emit the special GOT helper function, do so now. */
10822 if (got_helper_rtx)
10824 const char *name = XSTR (got_helper_rtx, 0);
10825 const char *reg_name = reg_names[GLOBAL_OFFSET_TABLE_REGNUM];
10826 #ifdef DWARF2_UNWIND_INFO
10830 if (USE_HIDDEN_LINKONCE)
10832 tree decl = build_decl (BUILTINS_LOCATION, FUNCTION_DECL,
10833 get_identifier (name),
10834 build_function_type_list (void_type_node,
10836 DECL_RESULT (decl) = build_decl (BUILTINS_LOCATION, RESULT_DECL,
10837 NULL_TREE, void_type_node);
10838 TREE_PUBLIC (decl) = 1;
10839 TREE_STATIC (decl) = 1;
10840 make_decl_one_only (decl, DECL_ASSEMBLER_NAME (decl));
10841 DECL_VISIBILITY (decl) = VISIBILITY_HIDDEN;
10842 DECL_VISIBILITY_SPECIFIED (decl) = 1;
10843 resolve_unique_section (decl, 0, flag_function_sections);
10844 allocate_struct_function (decl, true);
10845 cfun->is_thunk = 1;
10846 current_function_decl = decl;
10847 init_varasm_status ();
10848 assemble_start_function (decl, name);
10852 const int align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
10853 switch_to_section (text_section);
10855 ASM_OUTPUT_ALIGN (asm_out_file, align);
10856 ASM_OUTPUT_LABEL (asm_out_file, name);
10859 #ifdef DWARF2_UNWIND_INFO
10860 do_cfi = dwarf2out_do_cfi_asm ();
10862 fprintf (asm_out_file, "\t.cfi_startproc\n");
10864 if (flag_delayed_branch)
10865 fprintf (asm_out_file, "\tjmp\t%%o7+8\n\t add\t%%o7, %s, %s\n",
10866 reg_name, reg_name);
10868 fprintf (asm_out_file, "\tadd\t%%o7, %s, %s\n\tjmp\t%%o7+8\n\t nop\n",
10869 reg_name, reg_name);
10870 #ifdef DWARF2_UNWIND_INFO
10872 fprintf (asm_out_file, "\t.cfi_endproc\n");
10876 if (NEED_INDICATE_EXEC_STACK)
10877 file_end_indicate_exec_stack ();
10879 #ifdef TARGET_SOLARIS
10880 solaris_file_end ();
10884 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
10885 /* Implement TARGET_MANGLE_TYPE. */
10887 static const char *
10888 sparc_mangle_type (const_tree type)
10891 && TYPE_MAIN_VARIANT (type) == long_double_type_node
10892 && TARGET_LONG_DOUBLE_128)
10895 /* For all other types, use normal C++ mangling. */
10900 /* Expand a membar instruction for various use cases. Both the LOAD_STORE
10901 and BEFORE_AFTER arguments of the form X_Y. They are two-bit masks where
10902 bit 0 indicates that X is true, and bit 1 indicates Y is true. */
10905 sparc_emit_membar_for_model (enum memmodel model,
10906 int load_store, int before_after)
10908 /* Bits for the MEMBAR mmask field. */
10909 const int LoadLoad = 1;
10910 const int StoreLoad = 2;
10911 const int LoadStore = 4;
10912 const int StoreStore = 8;
10914 int mm = 0, implied = 0;
10916 switch (sparc_memory_model)
10919 /* Sequential Consistency. All memory transactions are immediately
10920 visible in sequential execution order. No barriers needed. */
10921 implied = LoadLoad | StoreLoad | LoadStore | StoreStore;
10925 /* Total Store Ordering: all memory transactions with store semantics
10926 are followed by an implied StoreStore. */
10927 implied |= StoreStore;
10931 /* Partial Store Ordering: all memory transactions with load semantics
10932 are followed by an implied LoadLoad | LoadStore. */
10933 implied |= LoadLoad | LoadStore;
10935 /* If we're not looking for a raw barrer (before+after), then atomic
10936 operations get the benefit of being both load and store. */
10937 if (load_store == 3 && before_after == 2)
10938 implied |= StoreLoad | StoreStore;
10942 /* Relaxed Memory Ordering: no implicit bits. */
10946 gcc_unreachable ();
10949 if (before_after & 1)
10951 if (model == MEMMODEL_ACQUIRE
10952 || model == MEMMODEL_ACQ_REL
10953 || model == MEMMODEL_SEQ_CST)
10955 if (load_store & 1)
10956 mm |= LoadLoad | LoadStore;
10957 if (load_store & 2)
10958 mm |= StoreLoad | StoreStore;
10961 if (before_after & 2)
10963 if (model == MEMMODEL_RELEASE
10964 || model == MEMMODEL_ACQ_REL
10965 || model == MEMMODEL_SEQ_CST)
10967 if (load_store & 1)
10968 mm |= LoadLoad | StoreLoad;
10969 if (load_store & 2)
10970 mm |= LoadStore | StoreStore;
10974 /* Remove the bits implied by the system memory model. */
10977 /* For raw barriers (before+after), always emit a barrier.
10978 This will become a compile-time barrier if needed. */
10979 if (mm || before_after == 3)
10980 emit_insn (gen_membar (GEN_INT (mm)));
10983 /* Expand code to perform a 8 or 16-bit compare and swap by doing 32-bit
10984 compare and swap on the word containing the byte or half-word. */
10987 sparc_expand_compare_and_swap_12 (rtx bool_result, rtx result, rtx mem,
10988 rtx oldval, rtx newval)
10990 rtx addr1 = force_reg (Pmode, XEXP (mem, 0));
10991 rtx addr = gen_reg_rtx (Pmode);
10992 rtx off = gen_reg_rtx (SImode);
10993 rtx oldv = gen_reg_rtx (SImode);
10994 rtx newv = gen_reg_rtx (SImode);
10995 rtx oldvalue = gen_reg_rtx (SImode);
10996 rtx newvalue = gen_reg_rtx (SImode);
10997 rtx res = gen_reg_rtx (SImode);
10998 rtx resv = gen_reg_rtx (SImode);
10999 rtx memsi, val, mask, end_label, loop_label, cc;
11001 emit_insn (gen_rtx_SET (VOIDmode, addr,
11002 gen_rtx_AND (Pmode, addr1, GEN_INT (-4))));
11004 if (Pmode != SImode)
11005 addr1 = gen_lowpart (SImode, addr1);
11006 emit_insn (gen_rtx_SET (VOIDmode, off,
11007 gen_rtx_AND (SImode, addr1, GEN_INT (3))));
11009 memsi = gen_rtx_MEM (SImode, addr);
11010 set_mem_alias_set (memsi, ALIAS_SET_MEMORY_BARRIER);
11011 MEM_VOLATILE_P (memsi) = MEM_VOLATILE_P (mem);
11013 val = copy_to_reg (memsi);
11015 emit_insn (gen_rtx_SET (VOIDmode, off,
11016 gen_rtx_XOR (SImode, off,
11017 GEN_INT (GET_MODE (mem) == QImode
11020 emit_insn (gen_rtx_SET (VOIDmode, off,
11021 gen_rtx_ASHIFT (SImode, off, GEN_INT (3))));
11023 if (GET_MODE (mem) == QImode)
11024 mask = force_reg (SImode, GEN_INT (0xff));
11026 mask = force_reg (SImode, GEN_INT (0xffff));
11028 emit_insn (gen_rtx_SET (VOIDmode, mask,
11029 gen_rtx_ASHIFT (SImode, mask, off)));
11031 emit_insn (gen_rtx_SET (VOIDmode, val,
11032 gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask),
11035 oldval = gen_lowpart (SImode, oldval);
11036 emit_insn (gen_rtx_SET (VOIDmode, oldv,
11037 gen_rtx_ASHIFT (SImode, oldval, off)));
11039 newval = gen_lowpart_common (SImode, newval);
11040 emit_insn (gen_rtx_SET (VOIDmode, newv,
11041 gen_rtx_ASHIFT (SImode, newval, off)));
11043 emit_insn (gen_rtx_SET (VOIDmode, oldv,
11044 gen_rtx_AND (SImode, oldv, mask)));
11046 emit_insn (gen_rtx_SET (VOIDmode, newv,
11047 gen_rtx_AND (SImode, newv, mask)));
11049 end_label = gen_label_rtx ();
11050 loop_label = gen_label_rtx ();
11051 emit_label (loop_label);
11053 emit_insn (gen_rtx_SET (VOIDmode, oldvalue,
11054 gen_rtx_IOR (SImode, oldv, val)));
11056 emit_insn (gen_rtx_SET (VOIDmode, newvalue,
11057 gen_rtx_IOR (SImode, newv, val)));
11059 emit_move_insn (bool_result, const1_rtx);
11061 emit_insn (gen_atomic_compare_and_swapsi_1 (res, memsi, oldvalue, newvalue));
11063 emit_cmp_and_jump_insns (res, oldvalue, EQ, NULL, SImode, 0, end_label);
11065 emit_insn (gen_rtx_SET (VOIDmode, resv,
11066 gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask),
11069 emit_move_insn (bool_result, const0_rtx);
11071 cc = gen_compare_reg_1 (NE, resv, val);
11072 emit_insn (gen_rtx_SET (VOIDmode, val, resv));
11074 /* Use cbranchcc4 to separate the compare and branch! */
11075 emit_jump_insn (gen_cbranchcc4 (gen_rtx_NE (VOIDmode, cc, const0_rtx),
11076 cc, const0_rtx, loop_label));
11078 emit_label (end_label);
11080 emit_insn (gen_rtx_SET (VOIDmode, res,
11081 gen_rtx_AND (SImode, res, mask)));
11083 emit_insn (gen_rtx_SET (VOIDmode, res,
11084 gen_rtx_LSHIFTRT (SImode, res, off)));
11086 emit_move_insn (result, gen_lowpart (GET_MODE (result), res));
11089 /* Expand code to perform a compare-and-swap. */
11092 sparc_expand_compare_and_swap (rtx operands[])
11094 rtx bval, retval, mem, oldval, newval;
11095 enum machine_mode mode;
11096 enum memmodel model;
11098 bval = operands[0];
11099 retval = operands[1];
11101 oldval = operands[3];
11102 newval = operands[4];
11103 model = (enum memmodel) INTVAL (operands[6]);
11104 mode = GET_MODE (mem);
11106 sparc_emit_membar_for_model (model, 3, 1);
11108 if (reg_overlap_mentioned_p (retval, oldval))
11109 oldval = copy_to_reg (oldval);
11111 if (mode == QImode || mode == HImode)
11112 sparc_expand_compare_and_swap_12 (bval, retval, mem, oldval, newval);
11115 rtx (*gen) (rtx, rtx, rtx, rtx);
11118 if (mode == SImode)
11119 gen = gen_atomic_compare_and_swapsi_1;
11121 gen = gen_atomic_compare_and_swapdi_1;
11122 emit_insn (gen (retval, mem, oldval, newval));
11124 x = emit_store_flag (bval, EQ, retval, oldval, mode, 1, 1);
11126 convert_move (bval, x, 1);
11129 sparc_emit_membar_for_model (model, 3, 2);
11133 sparc_expand_vec_perm_bmask (enum machine_mode vmode, rtx sel)
11137 sel = gen_lowpart (DImode, sel);
11141 /* inp = xxxxxxxAxxxxxxxB */
11142 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (16),
11143 NULL_RTX, 1, OPTAB_DIRECT);
11144 /* t_1 = ....xxxxxxxAxxx. */
11145 sel = expand_simple_binop (SImode, AND, gen_lowpart (SImode, sel),
11146 GEN_INT (3), NULL_RTX, 1, OPTAB_DIRECT);
11147 t_1 = expand_simple_binop (SImode, AND, gen_lowpart (SImode, t_1),
11148 GEN_INT (0x30000), NULL_RTX, 1, OPTAB_DIRECT);
11149 /* sel = .......B */
11150 /* t_1 = ...A.... */
11151 sel = expand_simple_binop (SImode, IOR, sel, t_1, sel, 1, OPTAB_DIRECT);
11152 /* sel = ...A...B */
11153 sel = expand_mult (SImode, sel, GEN_INT (0x4444), sel, 1);
11154 /* sel = AAAABBBB * 4 */
11155 t_1 = force_reg (SImode, GEN_INT (0x01230123));
11156 /* sel = { A*4, A*4+1, A*4+2, ... } */
11160 /* inp = xxxAxxxBxxxCxxxD */
11161 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (8),
11162 NULL_RTX, 1, OPTAB_DIRECT);
11163 t_2 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (16),
11164 NULL_RTX, 1, OPTAB_DIRECT);
11165 t_3 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (24),
11166 NULL_RTX, 1, OPTAB_DIRECT);
11167 /* t_1 = ..xxxAxxxBxxxCxx */
11168 /* t_2 = ....xxxAxxxBxxxC */
11169 /* t_3 = ......xxxAxxxBxx */
11170 sel = expand_simple_binop (SImode, AND, gen_lowpart (SImode, sel),
11172 NULL_RTX, 1, OPTAB_DIRECT);
11173 t_1 = expand_simple_binop (SImode, AND, gen_lowpart (SImode, t_1),
11175 NULL_RTX, 1, OPTAB_DIRECT);
11176 t_2 = expand_simple_binop (SImode, AND, gen_lowpart (SImode, t_2),
11177 GEN_INT (0x070000),
11178 NULL_RTX, 1, OPTAB_DIRECT);
11179 t_3 = expand_simple_binop (SImode, AND, gen_lowpart (SImode, t_3),
11180 GEN_INT (0x07000000),
11181 NULL_RTX, 1, OPTAB_DIRECT);
11182 /* sel = .......D */
11183 /* t_1 = .....C.. */
11184 /* t_2 = ...B.... */
11185 /* t_3 = .A...... */
11186 sel = expand_simple_binop (SImode, IOR, sel, t_1, sel, 1, OPTAB_DIRECT);
11187 t_2 = expand_simple_binop (SImode, IOR, t_2, t_3, t_2, 1, OPTAB_DIRECT);
11188 sel = expand_simple_binop (SImode, IOR, sel, t_2, sel, 1, OPTAB_DIRECT);
11189 /* sel = .A.B.C.D */
11190 sel = expand_mult (SImode, sel, GEN_INT (0x22), sel, 1);
11191 /* sel = AABBCCDD * 2 */
11192 t_1 = force_reg (SImode, GEN_INT (0x01010101));
11193 /* sel = { A*2, A*2+1, B*2, B*2+1, ... } */
11197 /* input = xAxBxCxDxExFxGxH */
11198 sel = expand_simple_binop (DImode, AND, sel,
11199 GEN_INT ((HOST_WIDE_INT)0x0f0f0f0f << 32
11201 NULL_RTX, 1, OPTAB_DIRECT);
11202 /* sel = .A.B.C.D.E.F.G.H */
11203 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (4),
11204 NULL_RTX, 1, OPTAB_DIRECT);
11205 /* t_1 = ..A.B.C.D.E.F.G. */
11206 sel = expand_simple_binop (DImode, IOR, sel, t_1,
11207 NULL_RTX, 1, OPTAB_DIRECT);
11208 /* sel = .AABBCCDDEEFFGGH */
11209 sel = expand_simple_binop (DImode, AND, sel,
11210 GEN_INT ((HOST_WIDE_INT)0xff00ff << 32
11212 NULL_RTX, 1, OPTAB_DIRECT);
11213 /* sel = ..AB..CD..EF..GH */
11214 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (8),
11215 NULL_RTX, 1, OPTAB_DIRECT);
11216 /* t_1 = ....AB..CD..EF.. */
11217 sel = expand_simple_binop (DImode, IOR, sel, t_1,
11218 NULL_RTX, 1, OPTAB_DIRECT);
11219 /* sel = ..ABABCDCDEFEFGH */
11220 sel = expand_simple_binop (DImode, AND, sel,
11221 GEN_INT ((HOST_WIDE_INT)0xffff << 32 | 0xffff),
11222 NULL_RTX, 1, OPTAB_DIRECT);
11223 /* sel = ....ABCD....EFGH */
11224 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (16),
11225 NULL_RTX, 1, OPTAB_DIRECT);
11226 /* t_1 = ........ABCD.... */
11227 sel = gen_lowpart (SImode, sel);
11228 t_1 = gen_lowpart (SImode, t_1);
11232 gcc_unreachable ();
11235 /* Always perform the final addition/merge within the bmask insn. */
11236 emit_insn (gen_bmasksi_vis (gen_reg_rtx (SImode), sel, t_1));
11239 /* Implement TARGET_FRAME_POINTER_REQUIRED. */
11242 sparc_frame_pointer_required (void)
11244 /* If the stack pointer is dynamically modified in the function, it cannot
11245 serve as the frame pointer. */
11246 if (cfun->calls_alloca)
11249 /* If the function receives nonlocal gotos, it needs to save the frame
11250 pointer in the nonlocal_goto_save_area object. */
11251 if (cfun->has_nonlocal_label)
11254 /* In flat mode, that's it. */
11258 /* Otherwise, the frame pointer is required if the function isn't leaf. */
11259 return !(current_function_is_leaf && only_leaf_regs_used ());
11262 /* The way this is structured, we can't eliminate SFP in favor of SP
11263 if the frame pointer is required: we want to use the SFP->HFP elimination
11264 in that case. But the test in update_eliminables doesn't know we are
11265 assuming below that we only do the former elimination. */
11268 sparc_can_eliminate (const int from ATTRIBUTE_UNUSED, const int to)
11270 return to == HARD_FRAME_POINTER_REGNUM || !sparc_frame_pointer_required ();
11273 /* Return the hard frame pointer directly to bypass the stack bias. */
11276 sparc_builtin_setjmp_frame_value (void)
11278 return hard_frame_pointer_rtx;
11281 /* If !TARGET_FPU, then make the fp registers and fp cc regs fixed so that
11282 they won't be allocated. */
11285 sparc_conditional_register_usage (void)
11287 if (PIC_OFFSET_TABLE_REGNUM != INVALID_REGNUM)
11289 fixed_regs[PIC_OFFSET_TABLE_REGNUM] = 1;
11290 call_used_regs[PIC_OFFSET_TABLE_REGNUM] = 1;
11292 /* If the user has passed -f{fixed,call-{used,saved}}-g5 */
11293 /* then honor it. */
11294 if (TARGET_ARCH32 && fixed_regs[5])
11296 else if (TARGET_ARCH64 && fixed_regs[5] == 2)
11301 for (regno = SPARC_FIRST_V9_FP_REG;
11302 regno <= SPARC_LAST_V9_FP_REG;
11304 fixed_regs[regno] = 1;
11305 /* %fcc0 is used by v8 and v9. */
11306 for (regno = SPARC_FIRST_V9_FCC_REG + 1;
11307 regno <= SPARC_LAST_V9_FCC_REG;
11309 fixed_regs[regno] = 1;
11314 for (regno = 32; regno < SPARC_LAST_V9_FCC_REG; regno++)
11315 fixed_regs[regno] = 1;
11317 /* If the user has passed -f{fixed,call-{used,saved}}-g2 */
11318 /* then honor it. Likewise with g3 and g4. */
11319 if (fixed_regs[2] == 2)
11320 fixed_regs[2] = ! TARGET_APP_REGS;
11321 if (fixed_regs[3] == 2)
11322 fixed_regs[3] = ! TARGET_APP_REGS;
11323 if (TARGET_ARCH32 && fixed_regs[4] == 2)
11324 fixed_regs[4] = ! TARGET_APP_REGS;
11325 else if (TARGET_CM_EMBMEDANY)
11327 else if (fixed_regs[4] == 2)
11332 /* Disable leaf functions. */
11333 memset (sparc_leaf_regs, 0, FIRST_PSEUDO_REGISTER);
11334 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
11335 leaf_reg_remap [regno] = regno;
11338 global_regs[SPARC_GSR_REG] = 1;
11341 /* Implement TARGET_PREFERRED_RELOAD_CLASS:
11343 - We can't load constants into FP registers.
11344 - We can't load FP constants into integer registers when soft-float,
11345 because there is no soft-float pattern with a r/F constraint.
11346 - We can't load FP constants into integer registers for TFmode unless
11347 it is 0.0L, because there is no movtf pattern with a r/F constraint.
11348 - Try and reload integer constants (symbolic or otherwise) back into
11349 registers directly, rather than having them dumped to memory. */
11352 sparc_preferred_reload_class (rtx x, reg_class_t rclass)
11354 enum machine_mode mode = GET_MODE (x);
11355 if (CONSTANT_P (x))
11357 if (FP_REG_CLASS_P (rclass)
11358 || rclass == GENERAL_OR_FP_REGS
11359 || rclass == GENERAL_OR_EXTRA_FP_REGS
11360 || (GET_MODE_CLASS (mode) == MODE_FLOAT && ! TARGET_FPU)
11361 || (mode == TFmode && ! const_zero_operand (x, mode)))
11364 if (GET_MODE_CLASS (mode) == MODE_INT)
11365 return GENERAL_REGS;
11367 if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT)
11369 if (! FP_REG_CLASS_P (rclass)
11370 || !(const_zero_operand (x, mode)
11371 || const_all_ones_operand (x, mode)))
11378 && (rclass == EXTRA_FP_REGS
11379 || rclass == GENERAL_OR_EXTRA_FP_REGS))
11381 int regno = true_regnum (x);
11383 if (SPARC_INT_REG_P (regno))
11384 return (rclass == EXTRA_FP_REGS
11385 ? FP_REGS : GENERAL_OR_FP_REGS);
11391 /* Output a wide multiply instruction in V8+ mode. INSN is the instruction,
11392 OPERANDS are its operands and OPCODE is the mnemonic to be used. */
11395 output_v8plus_mult (rtx insn, rtx *operands, const char *opcode)
11399 gcc_assert (! TARGET_ARCH64);
11401 if (sparc_check_64 (operands[1], insn) <= 0)
11402 output_asm_insn ("srl\t%L1, 0, %L1", operands);
11403 if (which_alternative == 1)
11404 output_asm_insn ("sllx\t%H1, 32, %H1", operands);
11405 if (GET_CODE (operands[2]) == CONST_INT)
11407 if (which_alternative == 1)
11409 output_asm_insn ("or\t%L1, %H1, %H1", operands);
11410 sprintf (mulstr, "%s\t%%H1, %%2, %%L0", opcode);
11411 output_asm_insn (mulstr, operands);
11412 return "srlx\t%L0, 32, %H0";
11416 output_asm_insn ("sllx\t%H1, 32, %3", operands);
11417 output_asm_insn ("or\t%L1, %3, %3", operands);
11418 sprintf (mulstr, "%s\t%%3, %%2, %%3", opcode);
11419 output_asm_insn (mulstr, operands);
11420 output_asm_insn ("srlx\t%3, 32, %H0", operands);
11421 return "mov\t%3, %L0";
11424 else if (rtx_equal_p (operands[1], operands[2]))
11426 if (which_alternative == 1)
11428 output_asm_insn ("or\t%L1, %H1, %H1", operands);
11429 sprintf (mulstr, "%s\t%%H1, %%H1, %%L0", opcode);
11430 output_asm_insn (mulstr, operands);
11431 return "srlx\t%L0, 32, %H0";
11435 output_asm_insn ("sllx\t%H1, 32, %3", operands);
11436 output_asm_insn ("or\t%L1, %3, %3", operands);
11437 sprintf (mulstr, "%s\t%%3, %%3, %%3", opcode);
11438 output_asm_insn (mulstr, operands);
11439 output_asm_insn ("srlx\t%3, 32, %H0", operands);
11440 return "mov\t%3, %L0";
11443 if (sparc_check_64 (operands[2], insn) <= 0)
11444 output_asm_insn ("srl\t%L2, 0, %L2", operands);
11445 if (which_alternative == 1)
11447 output_asm_insn ("or\t%L1, %H1, %H1", operands);
11448 output_asm_insn ("sllx\t%H2, 32, %L1", operands);
11449 output_asm_insn ("or\t%L2, %L1, %L1", operands);
11450 sprintf (mulstr, "%s\t%%H1, %%L1, %%L0", opcode);
11451 output_asm_insn (mulstr, operands);
11452 return "srlx\t%L0, 32, %H0";
11456 output_asm_insn ("sllx\t%H1, 32, %3", operands);
11457 output_asm_insn ("sllx\t%H2, 32, %4", operands);
11458 output_asm_insn ("or\t%L1, %3, %3", operands);
11459 output_asm_insn ("or\t%L2, %4, %4", operands);
11460 sprintf (mulstr, "%s\t%%3, %%4, %%3", opcode);
11461 output_asm_insn (mulstr, operands);
11462 output_asm_insn ("srlx\t%3, 32, %H0", operands);
11463 return "mov\t%3, %L0";
11467 /* Subroutine of sparc_expand_vector_init. Emit code to initialize
11468 all fields of TARGET to ELT by means of VIS2 BSHUFFLE insn. MODE
11469 and INNER_MODE are the modes describing TARGET. */
11472 vector_init_bshuffle (rtx target, rtx elt, enum machine_mode mode,
11473 enum machine_mode inner_mode)
11475 rtx t1, final_insn;
11478 t1 = gen_reg_rtx (mode);
11480 elt = convert_modes (SImode, inner_mode, elt, true);
11481 emit_move_insn (gen_lowpart(SImode, t1), elt);
11486 final_insn = gen_bshufflev2si_vis (target, t1, t1);
11487 bmask = 0x45674567;
11490 final_insn = gen_bshufflev4hi_vis (target, t1, t1);
11491 bmask = 0x67676767;
11494 final_insn = gen_bshufflev8qi_vis (target, t1, t1);
11495 bmask = 0x77777777;
11498 gcc_unreachable ();
11501 emit_insn (gen_bmasksi_vis (gen_reg_rtx (SImode), CONST0_RTX (SImode),
11502 force_reg (SImode, GEN_INT (bmask))));
11503 emit_insn (final_insn);
11506 /* Subroutine of sparc_expand_vector_init. Emit code to initialize
11507 all fields of TARGET to ELT in V8QI by means of VIS FPMERGE insn. */
11510 vector_init_fpmerge (rtx target, rtx elt)
11512 rtx t1, t2, t2_low, t3, t3_low;
11514 t1 = gen_reg_rtx (V4QImode);
11515 elt = convert_modes (SImode, QImode, elt, true);
11516 emit_move_insn (gen_lowpart (SImode, t1), elt);
11518 t2 = gen_reg_rtx (V8QImode);
11519 t2_low = gen_lowpart (V4QImode, t2);
11520 emit_insn (gen_fpmerge_vis (t2, t1, t1));
11522 t3 = gen_reg_rtx (V8QImode);
11523 t3_low = gen_lowpart (V4QImode, t3);
11524 emit_insn (gen_fpmerge_vis (t3, t2_low, t2_low));
11526 emit_insn (gen_fpmerge_vis (target, t3_low, t3_low));
11529 /* Subroutine of sparc_expand_vector_init. Emit code to initialize
11530 all fields of TARGET to ELT in V4HI by means of VIS FALIGNDATA insn. */
11533 vector_init_faligndata (rtx target, rtx elt)
11535 rtx t1 = gen_reg_rtx (V4HImode);
11538 elt = convert_modes (SImode, HImode, elt, true);
11539 emit_move_insn (gen_lowpart (SImode, t1), elt);
11541 emit_insn (gen_alignaddrsi_vis (gen_reg_rtx (SImode),
11542 force_reg (SImode, GEN_INT (6)),
11545 for (i = 0; i < 4; i++)
11546 emit_insn (gen_faligndatav4hi_vis (target, t1, target));
11549 /* Emit code to initialize TARGET to values for individual fields VALS. */
11552 sparc_expand_vector_init (rtx target, rtx vals)
11554 const enum machine_mode mode = GET_MODE (target);
11555 const enum machine_mode inner_mode = GET_MODE_INNER (mode);
11556 const int n_elts = GET_MODE_NUNITS (mode);
11562 for (i = 0; i < n_elts; i++)
11564 rtx x = XVECEXP (vals, 0, i);
11565 if (!CONSTANT_P (x))
11568 if (i > 0 && !rtx_equal_p (x, XVECEXP (vals, 0, 0)))
11574 emit_move_insn (target, gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0)));
11578 if (GET_MODE_SIZE (inner_mode) == GET_MODE_SIZE (mode))
11580 if (GET_MODE_SIZE (inner_mode) == 4)
11582 emit_move_insn (gen_lowpart (SImode, target),
11583 gen_lowpart (SImode, XVECEXP (vals, 0, 0)));
11586 else if (GET_MODE_SIZE (inner_mode) == 8)
11588 emit_move_insn (gen_lowpart (DImode, target),
11589 gen_lowpart (DImode, XVECEXP (vals, 0, 0)));
11593 else if (GET_MODE_SIZE (inner_mode) == GET_MODE_SIZE (word_mode)
11594 && GET_MODE_SIZE (mode) == 2 * GET_MODE_SIZE (word_mode))
11596 emit_move_insn (gen_highpart (word_mode, target),
11597 gen_lowpart (word_mode, XVECEXP (vals, 0, 0)));
11598 emit_move_insn (gen_lowpart (word_mode, target),
11599 gen_lowpart (word_mode, XVECEXP (vals, 0, 1)));
11603 if (all_same && GET_MODE_SIZE (mode) == 8)
11607 vector_init_bshuffle (target, XVECEXP (vals, 0, 0), mode, inner_mode);
11610 if (mode == V8QImode)
11612 vector_init_fpmerge (target, XVECEXP (vals, 0, 0));
11615 if (mode == V4HImode)
11617 vector_init_faligndata (target, XVECEXP (vals, 0, 0));
11622 mem = assign_stack_temp (mode, GET_MODE_SIZE (mode), 0);
11623 for (i = 0; i < n_elts; i++)
11624 emit_move_insn (adjust_address_nv (mem, inner_mode,
11625 i * GET_MODE_SIZE (inner_mode)),
11626 XVECEXP (vals, 0, i));
11627 emit_move_insn (target, mem);
11630 /* Implement TARGET_SECONDARY_RELOAD. */
11633 sparc_secondary_reload (bool in_p, rtx x, reg_class_t rclass_i,
11634 enum machine_mode mode, secondary_reload_info *sri)
11636 enum reg_class rclass = (enum reg_class) rclass_i;
11638 sri->icode = CODE_FOR_nothing;
11639 sri->extra_cost = 0;
11641 /* We need a temporary when loading/storing a HImode/QImode value
11642 between memory and the FPU registers. This can happen when combine puts
11643 a paradoxical subreg in a float/fix conversion insn. */
11644 if (FP_REG_CLASS_P (rclass)
11645 && (mode == HImode || mode == QImode)
11646 && (GET_CODE (x) == MEM
11647 || ((GET_CODE (x) == REG || GET_CODE (x) == SUBREG)
11648 && true_regnum (x) == -1)))
11649 return GENERAL_REGS;
11651 /* On 32-bit we need a temporary when loading/storing a DFmode value
11652 between unaligned memory and the upper FPU registers. */
11654 && rclass == EXTRA_FP_REGS
11656 && GET_CODE (x) == MEM
11657 && ! mem_min_alignment (x, 8))
11660 if (((TARGET_CM_MEDANY
11661 && symbolic_operand (x, mode))
11662 || (TARGET_CM_EMBMEDANY
11663 && text_segment_operand (x, mode)))
11667 sri->icode = direct_optab_handler (reload_in_optab, mode);
11669 sri->icode = direct_optab_handler (reload_out_optab, mode);
11673 if (TARGET_VIS3 && TARGET_ARCH32)
11675 int regno = true_regnum (x);
11677 /* When using VIS3 fp<-->int register moves, on 32-bit we have
11678 to move 8-byte values in 4-byte pieces. This only works via
11679 FP_REGS, and not via EXTRA_FP_REGS. Therefore if we try to
11680 move between EXTRA_FP_REGS and GENERAL_REGS, we will need
11681 an FP_REGS intermediate move. */
11682 if ((rclass == EXTRA_FP_REGS && SPARC_INT_REG_P (regno))
11683 || ((general_or_i64_p (rclass)
11684 || rclass == GENERAL_OR_FP_REGS)
11685 && SPARC_FP_REG_P (regno)))
11687 sri->extra_cost = 2;
11695 /* Emit code to conditionally move either OPERANDS[2] or OPERANDS[3] into
11696 OPERANDS[0] in MODE. OPERANDS[1] is the operator of the condition. */
11699 sparc_expand_conditional_move (enum machine_mode mode, rtx *operands)
11701 enum rtx_code rc = GET_CODE (operands[1]);
11702 enum machine_mode cmp_mode;
11703 rtx cc_reg, dst, cmp;
11706 if (GET_MODE (XEXP (cmp, 0)) == DImode && !TARGET_ARCH64)
11709 if (GET_MODE (XEXP (cmp, 0)) == TFmode && !TARGET_HARD_QUAD)
11710 cmp = sparc_emit_float_lib_cmp (XEXP (cmp, 0), XEXP (cmp, 1), rc);
11712 cmp_mode = GET_MODE (XEXP (cmp, 0));
11713 rc = GET_CODE (cmp);
11716 if (! rtx_equal_p (operands[2], dst)
11717 && ! rtx_equal_p (operands[3], dst))
11719 if (reg_overlap_mentioned_p (dst, cmp))
11720 dst = gen_reg_rtx (mode);
11722 emit_move_insn (dst, operands[3]);
11724 else if (operands[2] == dst)
11726 operands[2] = operands[3];
11728 if (GET_MODE_CLASS (cmp_mode) == MODE_FLOAT)
11729 rc = reverse_condition_maybe_unordered (rc);
11731 rc = reverse_condition (rc);
11734 if (XEXP (cmp, 1) == const0_rtx
11735 && GET_CODE (XEXP (cmp, 0)) == REG
11736 && cmp_mode == DImode
11737 && v9_regcmp_p (rc))
11738 cc_reg = XEXP (cmp, 0);
11740 cc_reg = gen_compare_reg_1 (rc, XEXP (cmp, 0), XEXP (cmp, 1));
11742 cmp = gen_rtx_fmt_ee (rc, GET_MODE (cc_reg), cc_reg, const0_rtx);
11744 emit_insn (gen_rtx_SET (VOIDmode, dst,
11745 gen_rtx_IF_THEN_ELSE (mode, cmp, operands[2], dst)));
11747 if (dst != operands[0])
11748 emit_move_insn (operands[0], dst);
11753 /* Emit code to conditionally move a combination of OPERANDS[1] and OPERANDS[2]
11754 into OPERANDS[0] in MODE, depending on the outcome of the comparison of
11755 OPERANDS[4] and OPERANDS[5]. OPERANDS[3] is the operator of the condition.
11756 FCODE is the machine code to be used for OPERANDS[3] and CCODE the machine
11757 code to be used for the condition mask. */
11760 sparc_expand_vcond (enum machine_mode mode, rtx *operands, int ccode, int fcode)
11762 rtx mask, cop0, cop1, fcmp, cmask, bshuf, gsr;
11763 enum rtx_code code = GET_CODE (operands[3]);
11765 mask = gen_reg_rtx (Pmode);
11766 cop0 = operands[4];
11767 cop1 = operands[5];
11768 if (code == LT || code == GE)
11772 code = swap_condition (code);
11773 t = cop0; cop0 = cop1; cop1 = t;
11776 gsr = gen_rtx_REG (DImode, SPARC_GSR_REG);
11778 fcmp = gen_rtx_UNSPEC (Pmode,
11779 gen_rtvec (1, gen_rtx_fmt_ee (code, mode, cop0, cop1)),
11782 cmask = gen_rtx_UNSPEC (DImode,
11783 gen_rtvec (2, mask, gsr),
11786 bshuf = gen_rtx_UNSPEC (mode,
11787 gen_rtvec (3, operands[1], operands[2], gsr),
11790 emit_insn (gen_rtx_SET (VOIDmode, mask, fcmp));
11791 emit_insn (gen_rtx_SET (VOIDmode, gsr, cmask));
11793 emit_insn (gen_rtx_SET (VOIDmode, operands[0], bshuf));
11796 /* On sparc, any mode which naturally allocates into the float
11797 registers should return 4 here. */
11800 sparc_regmode_natural_size (enum machine_mode mode)
11802 int size = UNITS_PER_WORD;
11806 enum mode_class mclass = GET_MODE_CLASS (mode);
11808 if (mclass == MODE_FLOAT || mclass == MODE_VECTOR_INT)
11815 /* Return TRUE if it is a good idea to tie two pseudo registers
11816 when one has mode MODE1 and one has mode MODE2.
11817 If HARD_REGNO_MODE_OK could produce different values for MODE1 and MODE2,
11818 for any hard reg, then this must be FALSE for correct output.
11820 For V9 we have to deal with the fact that only the lower 32 floating
11821 point registers are 32-bit addressable. */
11824 sparc_modes_tieable_p (enum machine_mode mode1, enum machine_mode mode2)
11826 enum mode_class mclass1, mclass2;
11827 unsigned short size1, size2;
11829 if (mode1 == mode2)
11832 mclass1 = GET_MODE_CLASS (mode1);
11833 mclass2 = GET_MODE_CLASS (mode2);
11834 if (mclass1 != mclass2)
11840 /* Classes are the same and we are V9 so we have to deal with upper
11841 vs. lower floating point registers. If one of the modes is a
11842 4-byte mode, and the other is not, we have to mark them as not
11843 tieable because only the lower 32 floating point register are
11844 addressable 32-bits at a time.
11846 We can't just test explicitly for SFmode, otherwise we won't
11847 cover the vector mode cases properly. */
11849 if (mclass1 != MODE_FLOAT && mclass1 != MODE_VECTOR_INT)
11852 size1 = GET_MODE_SIZE (mode1);
11853 size2 = GET_MODE_SIZE (mode2);
11854 if ((size1 > 4 && size2 == 4)
11855 || (size2 > 4 && size1 == 4))
11861 #include "gt-sparc.h"