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 struct gcc_target targetm = TARGET_INITIALIZER;
785 dump_target_flag_bits (const int flags)
787 if (flags & MASK_64BIT)
788 fprintf (stderr, "64BIT ");
789 if (flags & MASK_APP_REGS)
790 fprintf (stderr, "APP_REGS ");
791 if (flags & MASK_FASTER_STRUCTS)
792 fprintf (stderr, "FASTER_STRUCTS ");
793 if (flags & MASK_FLAT)
794 fprintf (stderr, "FLAT ");
795 if (flags & MASK_FMAF)
796 fprintf (stderr, "FMAF ");
797 if (flags & MASK_FPU)
798 fprintf (stderr, "FPU ");
799 if (flags & MASK_HARD_QUAD)
800 fprintf (stderr, "HARD_QUAD ");
801 if (flags & MASK_POPC)
802 fprintf (stderr, "POPC ");
803 if (flags & MASK_PTR64)
804 fprintf (stderr, "PTR64 ");
805 if (flags & MASK_STACK_BIAS)
806 fprintf (stderr, "STACK_BIAS ");
807 if (flags & MASK_UNALIGNED_DOUBLES)
808 fprintf (stderr, "UNALIGNED_DOUBLES ");
809 if (flags & MASK_V8PLUS)
810 fprintf (stderr, "V8PLUS ");
811 if (flags & MASK_VIS)
812 fprintf (stderr, "VIS ");
813 if (flags & MASK_VIS2)
814 fprintf (stderr, "VIS2 ");
815 if (flags & MASK_VIS3)
816 fprintf (stderr, "VIS3 ");
817 if (flags & MASK_DEPRECATED_V8_INSNS)
818 fprintf (stderr, "DEPRECATED_V8_INSNS ");
819 if (flags & MASK_SPARCLET)
820 fprintf (stderr, "SPARCLET ");
821 if (flags & MASK_SPARCLITE)
822 fprintf (stderr, "SPARCLITE ");
824 fprintf (stderr, "V8 ");
826 fprintf (stderr, "V9 ");
830 dump_target_flags (const char *prefix, const int flags)
832 fprintf (stderr, "%s: (%08x) [ ", prefix, flags);
833 dump_target_flag_bits (flags);
834 fprintf(stderr, "]\n");
837 /* Validate and override various options, and do some machine dependent
841 sparc_option_override (void)
843 static struct code_model {
844 const char *const name;
845 const enum cmodel value;
846 } const cmodels[] = {
848 { "medlow", CM_MEDLOW },
849 { "medmid", CM_MEDMID },
850 { "medany", CM_MEDANY },
851 { "embmedany", CM_EMBMEDANY },
852 { NULL, (enum cmodel) 0 }
854 const struct code_model *cmodel;
855 /* Map TARGET_CPU_DEFAULT to value for -m{cpu,tune}=. */
856 static struct cpu_default {
858 const enum processor_type processor;
859 } const cpu_default[] = {
860 /* There must be one entry here for each TARGET_CPU value. */
861 { TARGET_CPU_sparc, PROCESSOR_CYPRESS },
862 { TARGET_CPU_v8, PROCESSOR_V8 },
863 { TARGET_CPU_supersparc, PROCESSOR_SUPERSPARC },
864 { TARGET_CPU_hypersparc, PROCESSOR_HYPERSPARC },
865 { TARGET_CPU_leon, PROCESSOR_LEON },
866 { TARGET_CPU_sparclite, PROCESSOR_F930 },
867 { TARGET_CPU_sparclite86x, PROCESSOR_SPARCLITE86X },
868 { TARGET_CPU_sparclet, PROCESSOR_TSC701 },
869 { TARGET_CPU_v9, PROCESSOR_V9 },
870 { TARGET_CPU_ultrasparc, PROCESSOR_ULTRASPARC },
871 { TARGET_CPU_ultrasparc3, PROCESSOR_ULTRASPARC3 },
872 { TARGET_CPU_niagara, PROCESSOR_NIAGARA },
873 { TARGET_CPU_niagara2, PROCESSOR_NIAGARA2 },
874 { TARGET_CPU_niagara3, PROCESSOR_NIAGARA3 },
875 { TARGET_CPU_niagara4, PROCESSOR_NIAGARA4 },
878 const struct cpu_default *def;
879 /* Table of values for -m{cpu,tune}=. This must match the order of
880 the PROCESSOR_* enumeration. */
881 static struct cpu_table {
882 const char *const name;
885 } const cpu_table[] = {
886 { "v7", MASK_ISA, 0 },
887 { "cypress", MASK_ISA, 0 },
888 { "v8", MASK_ISA, MASK_V8 },
889 /* TI TMS390Z55 supersparc */
890 { "supersparc", MASK_ISA, MASK_V8 },
891 { "hypersparc", MASK_ISA, MASK_V8|MASK_FPU },
893 { "leon", MASK_ISA, MASK_V8|MASK_FPU },
894 { "sparclite", MASK_ISA, MASK_SPARCLITE },
895 /* The Fujitsu MB86930 is the original sparclite chip, with no FPU. */
896 { "f930", MASK_ISA|MASK_FPU, MASK_SPARCLITE },
897 /* The Fujitsu MB86934 is the recent sparclite chip, with an FPU. */
898 { "f934", MASK_ISA, MASK_SPARCLITE|MASK_FPU },
899 { "sparclite86x", MASK_ISA|MASK_FPU, MASK_SPARCLITE },
900 { "sparclet", MASK_ISA, MASK_SPARCLET },
902 { "tsc701", MASK_ISA, MASK_SPARCLET },
903 { "v9", MASK_ISA, MASK_V9 },
904 /* UltraSPARC I, II, IIi */
905 { "ultrasparc", MASK_ISA,
906 /* Although insns using %y are deprecated, it is a clear win. */
907 MASK_V9|MASK_DEPRECATED_V8_INSNS },
909 /* ??? Check if %y issue still holds true. */
910 { "ultrasparc3", MASK_ISA,
911 MASK_V9|MASK_DEPRECATED_V8_INSNS|MASK_VIS2 },
913 { "niagara", MASK_ISA,
914 MASK_V9|MASK_DEPRECATED_V8_INSNS },
916 { "niagara2", MASK_ISA,
917 MASK_V9|MASK_POPC|MASK_VIS2 },
919 { "niagara3", MASK_ISA,
920 MASK_V9|MASK_POPC|MASK_VIS2|MASK_VIS3|MASK_FMAF },
922 { "niagara4", MASK_ISA,
923 MASK_V9|MASK_POPC|MASK_VIS2|MASK_VIS3|MASK_FMAF },
925 const struct cpu_table *cpu;
929 if (sparc_debug_string != NULL)
934 p = ASTRDUP (sparc_debug_string);
935 while ((q = strtok (p, ",")) != NULL)
949 if (! strcmp (q, "all"))
950 mask = MASK_DEBUG_ALL;
951 else if (! strcmp (q, "options"))
952 mask = MASK_DEBUG_OPTIONS;
954 error ("unknown -mdebug-%s switch", q);
957 sparc_debug &= ~mask;
963 if (TARGET_DEBUG_OPTIONS)
965 dump_target_flags("Initial target_flags", target_flags);
966 dump_target_flags("target_flags_explicit", target_flags_explicit);
969 #ifdef SUBTARGET_OVERRIDE_OPTIONS
970 SUBTARGET_OVERRIDE_OPTIONS;
973 #ifndef SPARC_BI_ARCH
974 /* Check for unsupported architecture size. */
975 if (! TARGET_64BIT != DEFAULT_ARCH32_P)
976 error ("%s is not supported by this configuration",
977 DEFAULT_ARCH32_P ? "-m64" : "-m32");
980 /* We force all 64bit archs to use 128 bit long double */
981 if (TARGET_64BIT && ! TARGET_LONG_DOUBLE_128)
983 error ("-mlong-double-64 not allowed with -m64");
984 target_flags |= MASK_LONG_DOUBLE_128;
987 /* Code model selection. */
988 sparc_cmodel = SPARC_DEFAULT_CMODEL;
992 sparc_cmodel = CM_32;
995 if (sparc_cmodel_string != NULL)
999 for (cmodel = &cmodels[0]; cmodel->name; cmodel++)
1000 if (strcmp (sparc_cmodel_string, cmodel->name) == 0)
1002 if (cmodel->name == NULL)
1003 error ("bad value (%s) for -mcmodel= switch", sparc_cmodel_string);
1005 sparc_cmodel = cmodel->value;
1008 error ("-mcmodel= is not supported on 32 bit systems");
1011 /* Check that -fcall-saved-REG wasn't specified for out registers. */
1012 for (i = 8; i < 16; i++)
1013 if (!call_used_regs [i])
1015 error ("-fcall-saved-REG is not supported for out registers");
1016 call_used_regs [i] = 1;
1019 fpu = target_flags & MASK_FPU; /* save current -mfpu status */
1021 /* Set the default CPU. */
1022 if (!global_options_set.x_sparc_cpu_and_features)
1024 for (def = &cpu_default[0]; def->cpu != -1; ++def)
1025 if (def->cpu == TARGET_CPU_DEFAULT)
1027 gcc_assert (def->cpu != -1);
1028 sparc_cpu_and_features = def->processor;
1031 if (!global_options_set.x_sparc_cpu)
1032 sparc_cpu = sparc_cpu_and_features;
1034 cpu = &cpu_table[(int) sparc_cpu_and_features];
1036 if (TARGET_DEBUG_OPTIONS)
1038 fprintf (stderr, "sparc_cpu_and_features: %s\n", cpu->name);
1039 fprintf (stderr, "sparc_cpu: %s\n",
1040 cpu_table[(int) sparc_cpu].name);
1041 dump_target_flags ("cpu->disable", cpu->disable);
1042 dump_target_flags ("cpu->enable", cpu->enable);
1045 target_flags &= ~cpu->disable;
1046 target_flags |= (cpu->enable
1047 #ifndef HAVE_AS_FMAF_HPC_VIS3
1048 & ~(MASK_FMAF | MASK_VIS3)
1052 /* If -mfpu or -mno-fpu was explicitly used, don't override with
1053 the processor default. */
1054 if (target_flags_explicit & MASK_FPU)
1055 target_flags = (target_flags & ~MASK_FPU) | fpu;
1057 /* -mvis2 implies -mvis */
1059 target_flags |= MASK_VIS;
1061 /* -mvis3 implies -mvis2 and -mvis */
1063 target_flags |= MASK_VIS2 | MASK_VIS;
1065 /* Don't allow -mvis, -mvis2, -mvis3, or -mfmaf if FPU is disabled. */
1067 target_flags &= ~(MASK_VIS | MASK_VIS2 | MASK_VIS3 | MASK_FMAF);
1069 /* -mvis assumes UltraSPARC+, so we are sure v9 instructions
1071 -m64 also implies v9. */
1072 if (TARGET_VIS || TARGET_ARCH64)
1074 target_flags |= MASK_V9;
1075 target_flags &= ~(MASK_V8 | MASK_SPARCLET | MASK_SPARCLITE);
1078 /* -mvis also implies -mv8plus on 32-bit */
1079 if (TARGET_VIS && ! TARGET_ARCH64)
1080 target_flags |= MASK_V8PLUS;
1082 /* Use the deprecated v8 insns for sparc64 in 32 bit mode. */
1083 if (TARGET_V9 && TARGET_ARCH32)
1084 target_flags |= MASK_DEPRECATED_V8_INSNS;
1086 /* V8PLUS requires V9, makes no sense in 64 bit mode. */
1087 if (! TARGET_V9 || TARGET_ARCH64)
1088 target_flags &= ~MASK_V8PLUS;
1090 /* Don't use stack biasing in 32 bit mode. */
1092 target_flags &= ~MASK_STACK_BIAS;
1094 /* Supply a default value for align_functions. */
1095 if (align_functions == 0
1096 && (sparc_cpu == PROCESSOR_ULTRASPARC
1097 || sparc_cpu == PROCESSOR_ULTRASPARC3
1098 || sparc_cpu == PROCESSOR_NIAGARA
1099 || sparc_cpu == PROCESSOR_NIAGARA2
1100 || sparc_cpu == PROCESSOR_NIAGARA3
1101 || sparc_cpu == PROCESSOR_NIAGARA4))
1102 align_functions = 32;
1104 /* Validate PCC_STRUCT_RETURN. */
1105 if (flag_pcc_struct_return == DEFAULT_PCC_STRUCT_RETURN)
1106 flag_pcc_struct_return = (TARGET_ARCH64 ? 0 : 1);
1108 /* Only use .uaxword when compiling for a 64-bit target. */
1110 targetm.asm_out.unaligned_op.di = NULL;
1112 /* Do various machine dependent initializations. */
1113 sparc_init_modes ();
1115 /* Set up function hooks. */
1116 init_machine_status = sparc_init_machine_status;
1121 case PROCESSOR_CYPRESS:
1122 sparc_costs = &cypress_costs;
1125 case PROCESSOR_SPARCLITE:
1126 case PROCESSOR_SUPERSPARC:
1127 sparc_costs = &supersparc_costs;
1129 case PROCESSOR_F930:
1130 case PROCESSOR_F934:
1131 case PROCESSOR_HYPERSPARC:
1132 case PROCESSOR_SPARCLITE86X:
1133 sparc_costs = &hypersparc_costs;
1135 case PROCESSOR_LEON:
1136 sparc_costs = &leon_costs;
1138 case PROCESSOR_SPARCLET:
1139 case PROCESSOR_TSC701:
1140 sparc_costs = &sparclet_costs;
1143 case PROCESSOR_ULTRASPARC:
1144 sparc_costs = &ultrasparc_costs;
1146 case PROCESSOR_ULTRASPARC3:
1147 sparc_costs = &ultrasparc3_costs;
1149 case PROCESSOR_NIAGARA:
1150 sparc_costs = &niagara_costs;
1152 case PROCESSOR_NIAGARA2:
1153 sparc_costs = &niagara2_costs;
1155 case PROCESSOR_NIAGARA3:
1156 case PROCESSOR_NIAGARA4:
1157 sparc_costs = &niagara3_costs;
1159 case PROCESSOR_NATIVE:
1163 #ifdef TARGET_DEFAULT_LONG_DOUBLE_128
1164 if (!(target_flags_explicit & MASK_LONG_DOUBLE_128))
1165 target_flags |= MASK_LONG_DOUBLE_128;
1168 if (TARGET_DEBUG_OPTIONS)
1169 dump_target_flags ("Final target_flags", target_flags);
1171 maybe_set_param_value (PARAM_SIMULTANEOUS_PREFETCHES,
1172 ((sparc_cpu == PROCESSOR_ULTRASPARC
1173 || sparc_cpu == PROCESSOR_NIAGARA
1174 || sparc_cpu == PROCESSOR_NIAGARA2
1175 || sparc_cpu == PROCESSOR_NIAGARA3
1176 || sparc_cpu == PROCESSOR_NIAGARA4)
1178 : (sparc_cpu == PROCESSOR_ULTRASPARC3
1180 global_options.x_param_values,
1181 global_options_set.x_param_values);
1182 maybe_set_param_value (PARAM_L1_CACHE_LINE_SIZE,
1183 ((sparc_cpu == PROCESSOR_ULTRASPARC
1184 || sparc_cpu == PROCESSOR_ULTRASPARC3
1185 || sparc_cpu == PROCESSOR_NIAGARA
1186 || sparc_cpu == PROCESSOR_NIAGARA2
1187 || sparc_cpu == PROCESSOR_NIAGARA3
1188 || sparc_cpu == PROCESSOR_NIAGARA4)
1190 global_options.x_param_values,
1191 global_options_set.x_param_values);
1193 /* Disable save slot sharing for call-clobbered registers by default.
1194 The IRA sharing algorithm works on single registers only and this
1195 pessimizes for double floating-point registers. */
1196 if (!global_options_set.x_flag_ira_share_save_slots)
1197 flag_ira_share_save_slots = 0;
1200 /* Miscellaneous utilities. */
1202 /* Nonzero if CODE, a comparison, is suitable for use in v9 conditional move
1203 or branch on register contents instructions. */
1206 v9_regcmp_p (enum rtx_code code)
1208 return (code == EQ || code == NE || code == GE || code == LT
1209 || code == LE || code == GT);
1212 /* Nonzero if OP is a floating point constant which can
1213 be loaded into an integer register using a single
1214 sethi instruction. */
1219 if (GET_CODE (op) == CONST_DOUBLE)
1224 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
1225 REAL_VALUE_TO_TARGET_SINGLE (r, i);
1226 return !SPARC_SIMM13_P (i) && SPARC_SETHI_P (i);
1232 /* Nonzero if OP is a floating point constant which can
1233 be loaded into an integer register using a single
1239 if (GET_CODE (op) == CONST_DOUBLE)
1244 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
1245 REAL_VALUE_TO_TARGET_SINGLE (r, i);
1246 return SPARC_SIMM13_P (i);
1252 /* Nonzero if OP is a floating point constant which can
1253 be loaded into an integer register using a high/losum
1254 instruction sequence. */
1257 fp_high_losum_p (rtx op)
1259 /* The constraints calling this should only be in
1260 SFmode move insns, so any constant which cannot
1261 be moved using a single insn will do. */
1262 if (GET_CODE (op) == CONST_DOUBLE)
1267 REAL_VALUE_FROM_CONST_DOUBLE (r, op);
1268 REAL_VALUE_TO_TARGET_SINGLE (r, i);
1269 return !SPARC_SIMM13_P (i) && !SPARC_SETHI_P (i);
1275 /* Return true if the address of LABEL can be loaded by means of the
1276 mov{si,di}_pic_label_ref patterns in PIC mode. */
1279 can_use_mov_pic_label_ref (rtx label)
1281 /* VxWorks does not impose a fixed gap between segments; the run-time
1282 gap can be different from the object-file gap. We therefore can't
1283 assume X - _GLOBAL_OFFSET_TABLE_ is a link-time constant unless we
1284 are absolutely sure that X is in the same segment as the GOT.
1285 Unfortunately, the flexibility of linker scripts means that we
1286 can't be sure of that in general, so assume that GOT-relative
1287 accesses are never valid on VxWorks. */
1288 if (TARGET_VXWORKS_RTP)
1291 /* Similarly, if the label is non-local, it might end up being placed
1292 in a different section than the current one; now mov_pic_label_ref
1293 requires the label and the code to be in the same section. */
1294 if (LABEL_REF_NONLOCAL_P (label))
1297 /* Finally, if we are reordering basic blocks and partition into hot
1298 and cold sections, this might happen for any label. */
1299 if (flag_reorder_blocks_and_partition)
1305 /* Expand a move instruction. Return true if all work is done. */
1308 sparc_expand_move (enum machine_mode mode, rtx *operands)
1310 /* Handle sets of MEM first. */
1311 if (GET_CODE (operands[0]) == MEM)
1313 /* 0 is a register (or a pair of registers) on SPARC. */
1314 if (register_or_zero_operand (operands[1], mode))
1317 if (!reload_in_progress)
1319 operands[0] = validize_mem (operands[0]);
1320 operands[1] = force_reg (mode, operands[1]);
1324 /* Fixup TLS cases. */
1326 && CONSTANT_P (operands[1])
1327 && sparc_tls_referenced_p (operands [1]))
1329 operands[1] = sparc_legitimize_tls_address (operands[1]);
1333 /* Fixup PIC cases. */
1334 if (flag_pic && CONSTANT_P (operands[1]))
1336 if (pic_address_needs_scratch (operands[1]))
1337 operands[1] = sparc_legitimize_pic_address (operands[1], NULL_RTX);
1339 /* We cannot use the mov{si,di}_pic_label_ref patterns in all cases. */
1340 if (GET_CODE (operands[1]) == LABEL_REF
1341 && can_use_mov_pic_label_ref (operands[1]))
1345 emit_insn (gen_movsi_pic_label_ref (operands[0], operands[1]));
1351 gcc_assert (TARGET_ARCH64);
1352 emit_insn (gen_movdi_pic_label_ref (operands[0], operands[1]));
1357 if (symbolic_operand (operands[1], mode))
1360 = sparc_legitimize_pic_address (operands[1],
1362 ? operands[0] : NULL_RTX);
1367 /* If we are trying to toss an integer constant into FP registers,
1368 or loading a FP or vector constant, force it into memory. */
1369 if (CONSTANT_P (operands[1])
1370 && REG_P (operands[0])
1371 && (SPARC_FP_REG_P (REGNO (operands[0]))
1372 || SCALAR_FLOAT_MODE_P (mode)
1373 || VECTOR_MODE_P (mode)))
1375 /* emit_group_store will send such bogosity to us when it is
1376 not storing directly into memory. So fix this up to avoid
1377 crashes in output_constant_pool. */
1378 if (operands [1] == const0_rtx)
1379 operands[1] = CONST0_RTX (mode);
1381 /* We can clear or set to all-ones FP registers if TARGET_VIS, and
1382 always other regs. */
1383 if ((TARGET_VIS || REGNO (operands[0]) < SPARC_FIRST_FP_REG)
1384 && (const_zero_operand (operands[1], mode)
1385 || const_all_ones_operand (operands[1], mode)))
1388 if (REGNO (operands[0]) < SPARC_FIRST_FP_REG
1389 /* We are able to build any SF constant in integer registers
1390 with at most 2 instructions. */
1392 /* And any DF constant in integer registers. */
1394 && ! can_create_pseudo_p ())))
1397 operands[1] = force_const_mem (mode, operands[1]);
1398 if (!reload_in_progress)
1399 operands[1] = validize_mem (operands[1]);
1403 /* Accept non-constants and valid constants unmodified. */
1404 if (!CONSTANT_P (operands[1])
1405 || GET_CODE (operands[1]) == HIGH
1406 || input_operand (operands[1], mode))
1412 /* All QImode constants require only one insn, so proceed. */
1417 sparc_emit_set_const32 (operands[0], operands[1]);
1421 /* input_operand should have filtered out 32-bit mode. */
1422 sparc_emit_set_const64 (operands[0], operands[1]);
1432 /* Load OP1, a 32-bit constant, into OP0, a register.
1433 We know it can't be done in one insn when we get
1434 here, the move expander guarantees this. */
1437 sparc_emit_set_const32 (rtx op0, rtx op1)
1439 enum machine_mode mode = GET_MODE (op0);
1442 if (can_create_pseudo_p ())
1443 temp = gen_reg_rtx (mode);
1445 if (GET_CODE (op1) == CONST_INT)
1447 gcc_assert (!small_int_operand (op1, mode)
1448 && !const_high_operand (op1, mode));
1450 /* Emit them as real moves instead of a HIGH/LO_SUM,
1451 this way CSE can see everything and reuse intermediate
1452 values if it wants. */
1453 emit_insn (gen_rtx_SET (VOIDmode, temp,
1454 GEN_INT (INTVAL (op1)
1455 & ~(HOST_WIDE_INT)0x3ff)));
1457 emit_insn (gen_rtx_SET (VOIDmode,
1459 gen_rtx_IOR (mode, temp,
1460 GEN_INT (INTVAL (op1) & 0x3ff))));
1464 /* A symbol, emit in the traditional way. */
1465 emit_insn (gen_rtx_SET (VOIDmode, temp,
1466 gen_rtx_HIGH (mode, op1)));
1467 emit_insn (gen_rtx_SET (VOIDmode,
1468 op0, gen_rtx_LO_SUM (mode, temp, op1)));
1472 /* Load OP1, a symbolic 64-bit constant, into OP0, a DImode register.
1473 If TEMP is nonzero, we are forbidden to use any other scratch
1474 registers. Otherwise, we are allowed to generate them as needed.
1476 Note that TEMP may have TImode if the code model is TARGET_CM_MEDANY
1477 or TARGET_CM_EMBMEDANY (see the reload_indi and reload_outdi patterns). */
1480 sparc_emit_set_symbolic_const64 (rtx op0, rtx op1, rtx temp)
1482 rtx temp1, temp2, temp3, temp4, temp5;
1485 if (temp && GET_MODE (temp) == TImode)
1488 temp = gen_rtx_REG (DImode, REGNO (temp));
1491 /* SPARC-V9 code-model support. */
1492 switch (sparc_cmodel)
1495 /* The range spanned by all instructions in the object is less
1496 than 2^31 bytes (2GB) and the distance from any instruction
1497 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1498 than 2^31 bytes (2GB).
1500 The executable must be in the low 4TB of the virtual address
1503 sethi %hi(symbol), %temp1
1504 or %temp1, %lo(symbol), %reg */
1506 temp1 = temp; /* op0 is allowed. */
1508 temp1 = gen_reg_rtx (DImode);
1510 emit_insn (gen_rtx_SET (VOIDmode, temp1, gen_rtx_HIGH (DImode, op1)));
1511 emit_insn (gen_rtx_SET (VOIDmode, op0, gen_rtx_LO_SUM (DImode, temp1, op1)));
1515 /* The range spanned by all instructions in the object is less
1516 than 2^31 bytes (2GB) and the distance from any instruction
1517 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1518 than 2^31 bytes (2GB).
1520 The executable must be in the low 16TB of the virtual address
1523 sethi %h44(symbol), %temp1
1524 or %temp1, %m44(symbol), %temp2
1525 sllx %temp2, 12, %temp3
1526 or %temp3, %l44(symbol), %reg */
1531 temp3 = temp; /* op0 is allowed. */
1535 temp1 = gen_reg_rtx (DImode);
1536 temp2 = gen_reg_rtx (DImode);
1537 temp3 = gen_reg_rtx (DImode);
1540 emit_insn (gen_seth44 (temp1, op1));
1541 emit_insn (gen_setm44 (temp2, temp1, op1));
1542 emit_insn (gen_rtx_SET (VOIDmode, temp3,
1543 gen_rtx_ASHIFT (DImode, temp2, GEN_INT (12))));
1544 emit_insn (gen_setl44 (op0, temp3, op1));
1548 /* The range spanned by all instructions in the object is less
1549 than 2^31 bytes (2GB) and the distance from any instruction
1550 to the location of the label _GLOBAL_OFFSET_TABLE_ is less
1551 than 2^31 bytes (2GB).
1553 The executable can be placed anywhere in the virtual address
1556 sethi %hh(symbol), %temp1
1557 sethi %lm(symbol), %temp2
1558 or %temp1, %hm(symbol), %temp3
1559 sllx %temp3, 32, %temp4
1560 or %temp4, %temp2, %temp5
1561 or %temp5, %lo(symbol), %reg */
1564 /* It is possible that one of the registers we got for operands[2]
1565 might coincide with that of operands[0] (which is why we made
1566 it TImode). Pick the other one to use as our scratch. */
1567 if (rtx_equal_p (temp, op0))
1569 gcc_assert (ti_temp);
1570 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1573 temp2 = temp; /* op0 is _not_ allowed, see above. */
1580 temp1 = gen_reg_rtx (DImode);
1581 temp2 = gen_reg_rtx (DImode);
1582 temp3 = gen_reg_rtx (DImode);
1583 temp4 = gen_reg_rtx (DImode);
1584 temp5 = gen_reg_rtx (DImode);
1587 emit_insn (gen_sethh (temp1, op1));
1588 emit_insn (gen_setlm (temp2, op1));
1589 emit_insn (gen_sethm (temp3, temp1, op1));
1590 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1591 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1592 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1593 gen_rtx_PLUS (DImode, temp4, temp2)));
1594 emit_insn (gen_setlo (op0, temp5, op1));
1598 /* Old old old backwards compatibility kruft here.
1599 Essentially it is MEDLOW with a fixed 64-bit
1600 virtual base added to all data segment addresses.
1601 Text-segment stuff is computed like MEDANY, we can't
1602 reuse the code above because the relocation knobs
1605 Data segment: sethi %hi(symbol), %temp1
1606 add %temp1, EMBMEDANY_BASE_REG, %temp2
1607 or %temp2, %lo(symbol), %reg */
1608 if (data_segment_operand (op1, GET_MODE (op1)))
1612 temp1 = temp; /* op0 is allowed. */
1617 temp1 = gen_reg_rtx (DImode);
1618 temp2 = gen_reg_rtx (DImode);
1621 emit_insn (gen_embmedany_sethi (temp1, op1));
1622 emit_insn (gen_embmedany_brsum (temp2, temp1));
1623 emit_insn (gen_embmedany_losum (op0, temp2, op1));
1626 /* Text segment: sethi %uhi(symbol), %temp1
1627 sethi %hi(symbol), %temp2
1628 or %temp1, %ulo(symbol), %temp3
1629 sllx %temp3, 32, %temp4
1630 or %temp4, %temp2, %temp5
1631 or %temp5, %lo(symbol), %reg */
1636 /* It is possible that one of the registers we got for operands[2]
1637 might coincide with that of operands[0] (which is why we made
1638 it TImode). Pick the other one to use as our scratch. */
1639 if (rtx_equal_p (temp, op0))
1641 gcc_assert (ti_temp);
1642 temp = gen_rtx_REG (DImode, REGNO (temp) + 1);
1645 temp2 = temp; /* op0 is _not_ allowed, see above. */
1652 temp1 = gen_reg_rtx (DImode);
1653 temp2 = gen_reg_rtx (DImode);
1654 temp3 = gen_reg_rtx (DImode);
1655 temp4 = gen_reg_rtx (DImode);
1656 temp5 = gen_reg_rtx (DImode);
1659 emit_insn (gen_embmedany_textuhi (temp1, op1));
1660 emit_insn (gen_embmedany_texthi (temp2, op1));
1661 emit_insn (gen_embmedany_textulo (temp3, temp1, op1));
1662 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1663 gen_rtx_ASHIFT (DImode, temp3, GEN_INT (32))));
1664 emit_insn (gen_rtx_SET (VOIDmode, temp5,
1665 gen_rtx_PLUS (DImode, temp4, temp2)));
1666 emit_insn (gen_embmedany_textlo (op0, temp5, op1));
1675 #if HOST_BITS_PER_WIDE_INT == 32
1677 sparc_emit_set_const64 (rtx op0 ATTRIBUTE_UNUSED, rtx op1 ATTRIBUTE_UNUSED)
1682 /* These avoid problems when cross compiling. If we do not
1683 go through all this hair then the optimizer will see
1684 invalid REG_EQUAL notes or in some cases none at all. */
1685 static rtx gen_safe_HIGH64 (rtx, HOST_WIDE_INT);
1686 static rtx gen_safe_SET64 (rtx, HOST_WIDE_INT);
1687 static rtx gen_safe_OR64 (rtx, HOST_WIDE_INT);
1688 static rtx gen_safe_XOR64 (rtx, HOST_WIDE_INT);
1690 /* The optimizer is not to assume anything about exactly
1691 which bits are set for a HIGH, they are unspecified.
1692 Unfortunately this leads to many missed optimizations
1693 during CSE. We mask out the non-HIGH bits, and matches
1694 a plain movdi, to alleviate this problem. */
1696 gen_safe_HIGH64 (rtx dest, HOST_WIDE_INT val)
1698 return gen_rtx_SET (VOIDmode, dest, GEN_INT (val & ~(HOST_WIDE_INT)0x3ff));
1702 gen_safe_SET64 (rtx dest, HOST_WIDE_INT val)
1704 return gen_rtx_SET (VOIDmode, dest, GEN_INT (val));
1708 gen_safe_OR64 (rtx src, HOST_WIDE_INT val)
1710 return gen_rtx_IOR (DImode, src, GEN_INT (val));
1714 gen_safe_XOR64 (rtx src, HOST_WIDE_INT val)
1716 return gen_rtx_XOR (DImode, src, GEN_INT (val));
1719 /* Worker routines for 64-bit constant formation on arch64.
1720 One of the key things to be doing in these emissions is
1721 to create as many temp REGs as possible. This makes it
1722 possible for half-built constants to be used later when
1723 such values are similar to something required later on.
1724 Without doing this, the optimizer cannot see such
1727 static void sparc_emit_set_const64_quick1 (rtx, rtx,
1728 unsigned HOST_WIDE_INT, int);
1731 sparc_emit_set_const64_quick1 (rtx op0, rtx temp,
1732 unsigned HOST_WIDE_INT low_bits, int is_neg)
1734 unsigned HOST_WIDE_INT high_bits;
1737 high_bits = (~low_bits) & 0xffffffff;
1739 high_bits = low_bits;
1741 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1744 emit_insn (gen_rtx_SET (VOIDmode, op0,
1745 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1749 /* If we are XOR'ing with -1, then we should emit a one's complement
1750 instead. This way the combiner will notice logical operations
1751 such as ANDN later on and substitute. */
1752 if ((low_bits & 0x3ff) == 0x3ff)
1754 emit_insn (gen_rtx_SET (VOIDmode, op0,
1755 gen_rtx_NOT (DImode, temp)));
1759 emit_insn (gen_rtx_SET (VOIDmode, op0,
1760 gen_safe_XOR64 (temp,
1761 (-(HOST_WIDE_INT)0x400
1762 | (low_bits & 0x3ff)))));
1767 static void sparc_emit_set_const64_quick2 (rtx, rtx, unsigned HOST_WIDE_INT,
1768 unsigned HOST_WIDE_INT, int);
1771 sparc_emit_set_const64_quick2 (rtx op0, rtx temp,
1772 unsigned HOST_WIDE_INT high_bits,
1773 unsigned HOST_WIDE_INT low_immediate,
1778 if ((high_bits & 0xfffffc00) != 0)
1780 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1781 if ((high_bits & ~0xfffffc00) != 0)
1782 emit_insn (gen_rtx_SET (VOIDmode, op0,
1783 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1789 emit_insn (gen_safe_SET64 (temp, high_bits));
1793 /* Now shift it up into place. */
1794 emit_insn (gen_rtx_SET (VOIDmode, op0,
1795 gen_rtx_ASHIFT (DImode, temp2,
1796 GEN_INT (shift_count))));
1798 /* If there is a low immediate part piece, finish up by
1799 putting that in as well. */
1800 if (low_immediate != 0)
1801 emit_insn (gen_rtx_SET (VOIDmode, op0,
1802 gen_safe_OR64 (op0, low_immediate)));
1805 static void sparc_emit_set_const64_longway (rtx, rtx, unsigned HOST_WIDE_INT,
1806 unsigned HOST_WIDE_INT);
1808 /* Full 64-bit constant decomposition. Even though this is the
1809 'worst' case, we still optimize a few things away. */
1811 sparc_emit_set_const64_longway (rtx op0, rtx temp,
1812 unsigned HOST_WIDE_INT high_bits,
1813 unsigned HOST_WIDE_INT low_bits)
1817 if (can_create_pseudo_p ())
1818 sub_temp = gen_reg_rtx (DImode);
1820 if ((high_bits & 0xfffffc00) != 0)
1822 emit_insn (gen_safe_HIGH64 (temp, high_bits));
1823 if ((high_bits & ~0xfffffc00) != 0)
1824 emit_insn (gen_rtx_SET (VOIDmode,
1826 gen_safe_OR64 (temp, (high_bits & 0x3ff))));
1832 emit_insn (gen_safe_SET64 (temp, high_bits));
1836 if (can_create_pseudo_p ())
1838 rtx temp2 = gen_reg_rtx (DImode);
1839 rtx temp3 = gen_reg_rtx (DImode);
1840 rtx temp4 = gen_reg_rtx (DImode);
1842 emit_insn (gen_rtx_SET (VOIDmode, temp4,
1843 gen_rtx_ASHIFT (DImode, sub_temp,
1846 emit_insn (gen_safe_HIGH64 (temp2, low_bits));
1847 if ((low_bits & ~0xfffffc00) != 0)
1849 emit_insn (gen_rtx_SET (VOIDmode, temp3,
1850 gen_safe_OR64 (temp2, (low_bits & 0x3ff))));
1851 emit_insn (gen_rtx_SET (VOIDmode, op0,
1852 gen_rtx_PLUS (DImode, temp4, temp3)));
1856 emit_insn (gen_rtx_SET (VOIDmode, op0,
1857 gen_rtx_PLUS (DImode, temp4, temp2)));
1862 rtx low1 = GEN_INT ((low_bits >> (32 - 12)) & 0xfff);
1863 rtx low2 = GEN_INT ((low_bits >> (32 - 12 - 12)) & 0xfff);
1864 rtx low3 = GEN_INT ((low_bits >> (32 - 12 - 12 - 8)) & 0x0ff);
1867 /* We are in the middle of reload, so this is really
1868 painful. However we do still make an attempt to
1869 avoid emitting truly stupid code. */
1870 if (low1 != const0_rtx)
1872 emit_insn (gen_rtx_SET (VOIDmode, op0,
1873 gen_rtx_ASHIFT (DImode, sub_temp,
1874 GEN_INT (to_shift))));
1875 emit_insn (gen_rtx_SET (VOIDmode, op0,
1876 gen_rtx_IOR (DImode, op0, low1)));
1884 if (low2 != const0_rtx)
1886 emit_insn (gen_rtx_SET (VOIDmode, op0,
1887 gen_rtx_ASHIFT (DImode, sub_temp,
1888 GEN_INT (to_shift))));
1889 emit_insn (gen_rtx_SET (VOIDmode, op0,
1890 gen_rtx_IOR (DImode, op0, low2)));
1898 emit_insn (gen_rtx_SET (VOIDmode, op0,
1899 gen_rtx_ASHIFT (DImode, sub_temp,
1900 GEN_INT (to_shift))));
1901 if (low3 != const0_rtx)
1902 emit_insn (gen_rtx_SET (VOIDmode, op0,
1903 gen_rtx_IOR (DImode, op0, low3)));
1908 /* Analyze a 64-bit constant for certain properties. */
1909 static void analyze_64bit_constant (unsigned HOST_WIDE_INT,
1910 unsigned HOST_WIDE_INT,
1911 int *, int *, int *);
1914 analyze_64bit_constant (unsigned HOST_WIDE_INT high_bits,
1915 unsigned HOST_WIDE_INT low_bits,
1916 int *hbsp, int *lbsp, int *abbasp)
1918 int lowest_bit_set, highest_bit_set, all_bits_between_are_set;
1921 lowest_bit_set = highest_bit_set = -1;
1925 if ((lowest_bit_set == -1)
1926 && ((low_bits >> i) & 1))
1928 if ((highest_bit_set == -1)
1929 && ((high_bits >> (32 - i - 1)) & 1))
1930 highest_bit_set = (64 - i - 1);
1933 && ((highest_bit_set == -1)
1934 || (lowest_bit_set == -1)));
1940 if ((lowest_bit_set == -1)
1941 && ((high_bits >> i) & 1))
1942 lowest_bit_set = i + 32;
1943 if ((highest_bit_set == -1)
1944 && ((low_bits >> (32 - i - 1)) & 1))
1945 highest_bit_set = 32 - i - 1;
1948 && ((highest_bit_set == -1)
1949 || (lowest_bit_set == -1)));
1951 /* If there are no bits set this should have gone out
1952 as one instruction! */
1953 gcc_assert (lowest_bit_set != -1 && highest_bit_set != -1);
1954 all_bits_between_are_set = 1;
1955 for (i = lowest_bit_set; i <= highest_bit_set; i++)
1959 if ((low_bits & (1 << i)) != 0)
1964 if ((high_bits & (1 << (i - 32))) != 0)
1967 all_bits_between_are_set = 0;
1970 *hbsp = highest_bit_set;
1971 *lbsp = lowest_bit_set;
1972 *abbasp = all_bits_between_are_set;
1975 static int const64_is_2insns (unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT);
1978 const64_is_2insns (unsigned HOST_WIDE_INT high_bits,
1979 unsigned HOST_WIDE_INT low_bits)
1981 int highest_bit_set, lowest_bit_set, all_bits_between_are_set;
1984 || high_bits == 0xffffffff)
1987 analyze_64bit_constant (high_bits, low_bits,
1988 &highest_bit_set, &lowest_bit_set,
1989 &all_bits_between_are_set);
1991 if ((highest_bit_set == 63
1992 || lowest_bit_set == 0)
1993 && all_bits_between_are_set != 0)
1996 if ((highest_bit_set - lowest_bit_set) < 21)
2002 static unsigned HOST_WIDE_INT create_simple_focus_bits (unsigned HOST_WIDE_INT,
2003 unsigned HOST_WIDE_INT,
2006 static unsigned HOST_WIDE_INT
2007 create_simple_focus_bits (unsigned HOST_WIDE_INT high_bits,
2008 unsigned HOST_WIDE_INT low_bits,
2009 int lowest_bit_set, int shift)
2011 HOST_WIDE_INT hi, lo;
2013 if (lowest_bit_set < 32)
2015 lo = (low_bits >> lowest_bit_set) << shift;
2016 hi = ((high_bits << (32 - lowest_bit_set)) << shift);
2021 hi = ((high_bits >> (lowest_bit_set - 32)) << shift);
2023 gcc_assert (! (hi & lo));
2027 /* Here we are sure to be arch64 and this is an integer constant
2028 being loaded into a register. Emit the most efficient
2029 insn sequence possible. Detection of all the 1-insn cases
2030 has been done already. */
2032 sparc_emit_set_const64 (rtx op0, rtx op1)
2034 unsigned HOST_WIDE_INT high_bits, low_bits;
2035 int lowest_bit_set, highest_bit_set;
2036 int all_bits_between_are_set;
2039 /* Sanity check that we know what we are working with. */
2040 gcc_assert (TARGET_ARCH64
2041 && (GET_CODE (op0) == SUBREG
2042 || (REG_P (op0) && ! SPARC_FP_REG_P (REGNO (op0)))));
2044 if (! can_create_pseudo_p ())
2047 if (GET_CODE (op1) != CONST_INT)
2049 sparc_emit_set_symbolic_const64 (op0, op1, temp);
2054 temp = gen_reg_rtx (DImode);
2056 high_bits = ((INTVAL (op1) >> 32) & 0xffffffff);
2057 low_bits = (INTVAL (op1) & 0xffffffff);
2059 /* low_bits bits 0 --> 31
2060 high_bits bits 32 --> 63 */
2062 analyze_64bit_constant (high_bits, low_bits,
2063 &highest_bit_set, &lowest_bit_set,
2064 &all_bits_between_are_set);
2066 /* First try for a 2-insn sequence. */
2068 /* These situations are preferred because the optimizer can
2069 * do more things with them:
2071 * sllx %reg, shift, %reg
2073 * srlx %reg, shift, %reg
2074 * 3) mov some_small_const, %reg
2075 * sllx %reg, shift, %reg
2077 if (((highest_bit_set == 63
2078 || lowest_bit_set == 0)
2079 && all_bits_between_are_set != 0)
2080 || ((highest_bit_set - lowest_bit_set) < 12))
2082 HOST_WIDE_INT the_const = -1;
2083 int shift = lowest_bit_set;
2085 if ((highest_bit_set != 63
2086 && lowest_bit_set != 0)
2087 || all_bits_between_are_set == 0)
2090 create_simple_focus_bits (high_bits, low_bits,
2093 else if (lowest_bit_set == 0)
2094 shift = -(63 - highest_bit_set);
2096 gcc_assert (SPARC_SIMM13_P (the_const));
2097 gcc_assert (shift != 0);
2099 emit_insn (gen_safe_SET64 (temp, the_const));
2101 emit_insn (gen_rtx_SET (VOIDmode,
2103 gen_rtx_ASHIFT (DImode,
2107 emit_insn (gen_rtx_SET (VOIDmode,
2109 gen_rtx_LSHIFTRT (DImode,
2111 GEN_INT (-shift))));
2115 /* Now a range of 22 or less bits set somewhere.
2116 * 1) sethi %hi(focus_bits), %reg
2117 * sllx %reg, shift, %reg
2118 * 2) sethi %hi(focus_bits), %reg
2119 * srlx %reg, shift, %reg
2121 if ((highest_bit_set - lowest_bit_set) < 21)
2123 unsigned HOST_WIDE_INT focus_bits =
2124 create_simple_focus_bits (high_bits, low_bits,
2125 lowest_bit_set, 10);
2127 gcc_assert (SPARC_SETHI_P (focus_bits));
2128 gcc_assert (lowest_bit_set != 10);
2130 emit_insn (gen_safe_HIGH64 (temp, focus_bits));
2132 /* If lowest_bit_set == 10 then a sethi alone could have done it. */
2133 if (lowest_bit_set < 10)
2134 emit_insn (gen_rtx_SET (VOIDmode,
2136 gen_rtx_LSHIFTRT (DImode, temp,
2137 GEN_INT (10 - lowest_bit_set))));
2138 else if (lowest_bit_set > 10)
2139 emit_insn (gen_rtx_SET (VOIDmode,
2141 gen_rtx_ASHIFT (DImode, temp,
2142 GEN_INT (lowest_bit_set - 10))));
2146 /* 1) sethi %hi(low_bits), %reg
2147 * or %reg, %lo(low_bits), %reg
2148 * 2) sethi %hi(~low_bits), %reg
2149 * xor %reg, %lo(-0x400 | (low_bits & 0x3ff)), %reg
2152 || high_bits == 0xffffffff)
2154 sparc_emit_set_const64_quick1 (op0, temp, low_bits,
2155 (high_bits == 0xffffffff));
2159 /* Now, try 3-insn sequences. */
2161 /* 1) sethi %hi(high_bits), %reg
2162 * or %reg, %lo(high_bits), %reg
2163 * sllx %reg, 32, %reg
2167 sparc_emit_set_const64_quick2 (op0, temp, high_bits, 0, 32);
2171 /* We may be able to do something quick
2172 when the constant is negated, so try that. */
2173 if (const64_is_2insns ((~high_bits) & 0xffffffff,
2174 (~low_bits) & 0xfffffc00))
2176 /* NOTE: The trailing bits get XOR'd so we need the
2177 non-negated bits, not the negated ones. */
2178 unsigned HOST_WIDE_INT trailing_bits = low_bits & 0x3ff;
2180 if ((((~high_bits) & 0xffffffff) == 0
2181 && ((~low_bits) & 0x80000000) == 0)
2182 || (((~high_bits) & 0xffffffff) == 0xffffffff
2183 && ((~low_bits) & 0x80000000) != 0))
2185 unsigned HOST_WIDE_INT fast_int = (~low_bits & 0xffffffff);
2187 if ((SPARC_SETHI_P (fast_int)
2188 && (~high_bits & 0xffffffff) == 0)
2189 || SPARC_SIMM13_P (fast_int))
2190 emit_insn (gen_safe_SET64 (temp, fast_int));
2192 sparc_emit_set_const64 (temp, GEN_INT (fast_int));
2197 negated_const = GEN_INT (((~low_bits) & 0xfffffc00) |
2198 (((HOST_WIDE_INT)((~high_bits) & 0xffffffff))<<32));
2199 sparc_emit_set_const64 (temp, negated_const);
2202 /* If we are XOR'ing with -1, then we should emit a one's complement
2203 instead. This way the combiner will notice logical operations
2204 such as ANDN later on and substitute. */
2205 if (trailing_bits == 0x3ff)
2207 emit_insn (gen_rtx_SET (VOIDmode, op0,
2208 gen_rtx_NOT (DImode, temp)));
2212 emit_insn (gen_rtx_SET (VOIDmode,
2214 gen_safe_XOR64 (temp,
2215 (-0x400 | trailing_bits))));
2220 /* 1) sethi %hi(xxx), %reg
2221 * or %reg, %lo(xxx), %reg
2222 * sllx %reg, yyy, %reg
2224 * ??? This is just a generalized version of the low_bits==0
2225 * thing above, FIXME...
2227 if ((highest_bit_set - lowest_bit_set) < 32)
2229 unsigned HOST_WIDE_INT focus_bits =
2230 create_simple_focus_bits (high_bits, low_bits,
2233 /* We can't get here in this state. */
2234 gcc_assert (highest_bit_set >= 32 && lowest_bit_set < 32);
2236 /* So what we know is that the set bits straddle the
2237 middle of the 64-bit word. */
2238 sparc_emit_set_const64_quick2 (op0, temp,
2244 /* 1) sethi %hi(high_bits), %reg
2245 * or %reg, %lo(high_bits), %reg
2246 * sllx %reg, 32, %reg
2247 * or %reg, low_bits, %reg
2249 if (SPARC_SIMM13_P(low_bits)
2250 && ((int)low_bits > 0))
2252 sparc_emit_set_const64_quick2 (op0, temp, high_bits, low_bits, 32);
2256 /* The easiest way when all else fails, is full decomposition. */
2257 sparc_emit_set_const64_longway (op0, temp, high_bits, low_bits);
2259 #endif /* HOST_BITS_PER_WIDE_INT == 32 */
2261 /* Given a comparison code (EQ, NE, etc.) and the first operand of a COMPARE,
2262 return the mode to be used for the comparison. For floating-point,
2263 CCFP[E]mode is used. CC_NOOVmode should be used when the first operand
2264 is a PLUS, MINUS, NEG, or ASHIFT. CCmode should be used when no special
2265 processing is needed. */
2268 select_cc_mode (enum rtx_code op, rtx x, rtx y ATTRIBUTE_UNUSED)
2270 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2296 else if (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS
2297 || GET_CODE (x) == NEG || GET_CODE (x) == ASHIFT)
2299 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
2300 return CCX_NOOVmode;
2306 if (TARGET_ARCH64 && GET_MODE (x) == DImode)
2313 /* Emit the compare insn and return the CC reg for a CODE comparison
2314 with operands X and Y. */
2317 gen_compare_reg_1 (enum rtx_code code, rtx x, rtx y)
2319 enum machine_mode mode;
2322 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2325 mode = SELECT_CC_MODE (code, x, y);
2327 /* ??? We don't have movcc patterns so we cannot generate pseudo regs for the
2328 fcc regs (cse can't tell they're really call clobbered regs and will
2329 remove a duplicate comparison even if there is an intervening function
2330 call - it will then try to reload the cc reg via an int reg which is why
2331 we need the movcc patterns). It is possible to provide the movcc
2332 patterns by using the ldxfsr/stxfsr v9 insns. I tried it: you need two
2333 registers (say %g1,%g5) and it takes about 6 insns. A better fix would be
2334 to tell cse that CCFPE mode registers (even pseudos) are call
2337 /* ??? This is an experiment. Rather than making changes to cse which may
2338 or may not be easy/clean, we do our own cse. This is possible because
2339 we will generate hard registers. Cse knows they're call clobbered (it
2340 doesn't know the same thing about pseudos). If we guess wrong, no big
2341 deal, but if we win, great! */
2343 if (TARGET_V9 && GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2344 #if 1 /* experiment */
2347 /* We cycle through the registers to ensure they're all exercised. */
2348 static int next_fcc_reg = 0;
2349 /* Previous x,y for each fcc reg. */
2350 static rtx prev_args[4][2];
2352 /* Scan prev_args for x,y. */
2353 for (reg = 0; reg < 4; reg++)
2354 if (prev_args[reg][0] == x && prev_args[reg][1] == y)
2359 prev_args[reg][0] = x;
2360 prev_args[reg][1] = y;
2361 next_fcc_reg = (next_fcc_reg + 1) & 3;
2363 cc_reg = gen_rtx_REG (mode, reg + SPARC_FIRST_V9_FCC_REG);
2366 cc_reg = gen_reg_rtx (mode);
2367 #endif /* ! experiment */
2368 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
2369 cc_reg = gen_rtx_REG (mode, SPARC_FCC_REG);
2371 cc_reg = gen_rtx_REG (mode, SPARC_ICC_REG);
2373 /* We shouldn't get there for TFmode if !TARGET_HARD_QUAD. If we do, this
2374 will only result in an unrecognizable insn so no point in asserting. */
2375 emit_insn (gen_rtx_SET (VOIDmode, cc_reg, gen_rtx_COMPARE (mode, x, y)));
2381 /* Emit the compare insn and return the CC reg for the comparison in CMP. */
2384 gen_compare_reg (rtx cmp)
2386 return gen_compare_reg_1 (GET_CODE (cmp), XEXP (cmp, 0), XEXP (cmp, 1));
2389 /* This function is used for v9 only.
2390 DEST is the target of the Scc insn.
2391 CODE is the code for an Scc's comparison.
2392 X and Y are the values we compare.
2394 This function is needed to turn
2397 (gt (reg:CCX 100 %icc)
2401 (gt:DI (reg:CCX 100 %icc)
2404 IE: The instruction recognizer needs to see the mode of the comparison to
2405 find the right instruction. We could use "gt:DI" right in the
2406 define_expand, but leaving it out allows us to handle DI, SI, etc. */
2409 gen_v9_scc (rtx dest, enum rtx_code compare_code, rtx x, rtx y)
2412 && (GET_MODE (x) == DImode
2413 || GET_MODE (dest) == DImode))
2416 /* Try to use the movrCC insns. */
2418 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
2420 && v9_regcmp_p (compare_code))
2425 /* Special case for op0 != 0. This can be done with one instruction if
2428 if (compare_code == NE
2429 && GET_MODE (dest) == DImode
2430 && rtx_equal_p (op0, dest))
2432 emit_insn (gen_rtx_SET (VOIDmode, dest,
2433 gen_rtx_IF_THEN_ELSE (DImode,
2434 gen_rtx_fmt_ee (compare_code, DImode,
2441 if (reg_overlap_mentioned_p (dest, op0))
2443 /* Handle the case where dest == x.
2444 We "early clobber" the result. */
2445 op0 = gen_reg_rtx (GET_MODE (x));
2446 emit_move_insn (op0, x);
2449 emit_insn (gen_rtx_SET (VOIDmode, dest, const0_rtx));
2450 if (GET_MODE (op0) != DImode)
2452 temp = gen_reg_rtx (DImode);
2453 convert_move (temp, op0, 0);
2457 emit_insn (gen_rtx_SET (VOIDmode, dest,
2458 gen_rtx_IF_THEN_ELSE (GET_MODE (dest),
2459 gen_rtx_fmt_ee (compare_code, DImode,
2467 x = gen_compare_reg_1 (compare_code, x, y);
2470 gcc_assert (GET_MODE (x) != CC_NOOVmode
2471 && GET_MODE (x) != CCX_NOOVmode);
2473 emit_insn (gen_rtx_SET (VOIDmode, dest, const0_rtx));
2474 emit_insn (gen_rtx_SET (VOIDmode, dest,
2475 gen_rtx_IF_THEN_ELSE (GET_MODE (dest),
2476 gen_rtx_fmt_ee (compare_code,
2477 GET_MODE (x), x, y),
2478 const1_rtx, dest)));
2484 /* Emit an scc insn. For seq, sne, sgeu, and sltu, we can do this
2485 without jumps using the addx/subx instructions. */
2488 emit_scc_insn (rtx operands[])
2495 /* The quad-word fp compare library routines all return nonzero to indicate
2496 true, which is different from the equivalent libgcc routines, so we must
2497 handle them specially here. */
2498 if (GET_MODE (operands[2]) == TFmode && ! TARGET_HARD_QUAD)
2500 operands[1] = sparc_emit_float_lib_cmp (operands[2], operands[3],
2501 GET_CODE (operands[1]));
2502 operands[2] = XEXP (operands[1], 0);
2503 operands[3] = XEXP (operands[1], 1);
2506 code = GET_CODE (operands[1]);
2510 /* For seq/sne on v9 we use the same code as v8 (the addx/subx method has
2511 more applications). The exception to this is "reg != 0" which can
2512 be done in one instruction on v9 (so we do it). */
2515 if (GET_MODE (x) == SImode)
2517 rtx pat = gen_seqsi_special (operands[0], x, y);
2521 else if (GET_MODE (x) == DImode)
2523 rtx pat = gen_seqdi_special (operands[0], x, y);
2531 if (GET_MODE (x) == SImode)
2533 rtx pat = gen_snesi_special (operands[0], x, y);
2537 else if (GET_MODE (x) == DImode)
2541 pat = gen_snedi_special_vis3 (operands[0], x, y);
2543 pat = gen_snedi_special (operands[0], x, y);
2551 && GET_MODE (x) == DImode
2553 && (code == GTU || code == LTU))
2554 && gen_v9_scc (operands[0], code, x, y))
2557 /* We can do LTU and GEU using the addx/subx instructions too. And
2558 for GTU/LEU, if both operands are registers swap them and fall
2559 back to the easy case. */
2560 if (code == GTU || code == LEU)
2562 if ((GET_CODE (x) == REG || GET_CODE (x) == SUBREG)
2563 && (GET_CODE (y) == REG || GET_CODE (y) == SUBREG))
2568 code = swap_condition (code);
2573 || (!TARGET_VIS3 && code == GEU))
2575 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
2576 gen_rtx_fmt_ee (code, SImode,
2577 gen_compare_reg_1 (code, x, y),
2582 /* All the posibilities to use addx/subx based sequences has been
2583 exhausted, try for a 3 instruction sequence using v9 conditional
2585 if (TARGET_V9 && gen_v9_scc (operands[0], code, x, y))
2588 /* Nope, do branches. */
2592 /* Emit a conditional jump insn for the v9 architecture using comparison code
2593 CODE and jump target LABEL.
2594 This function exists to take advantage of the v9 brxx insns. */
2597 emit_v9_brxx_insn (enum rtx_code code, rtx op0, rtx label)
2599 emit_jump_insn (gen_rtx_SET (VOIDmode,
2601 gen_rtx_IF_THEN_ELSE (VOIDmode,
2602 gen_rtx_fmt_ee (code, GET_MODE (op0),
2604 gen_rtx_LABEL_REF (VOIDmode, label),
2609 emit_conditional_branch_insn (rtx operands[])
2611 /* The quad-word fp compare library routines all return nonzero to indicate
2612 true, which is different from the equivalent libgcc routines, so we must
2613 handle them specially here. */
2614 if (GET_MODE (operands[1]) == TFmode && ! TARGET_HARD_QUAD)
2616 operands[0] = sparc_emit_float_lib_cmp (operands[1], operands[2],
2617 GET_CODE (operands[0]));
2618 operands[1] = XEXP (operands[0], 0);
2619 operands[2] = XEXP (operands[0], 1);
2622 if (TARGET_ARCH64 && operands[2] == const0_rtx
2623 && GET_CODE (operands[1]) == REG
2624 && GET_MODE (operands[1]) == DImode)
2626 emit_v9_brxx_insn (GET_CODE (operands[0]), operands[1], operands[3]);
2630 operands[1] = gen_compare_reg (operands[0]);
2631 operands[2] = const0_rtx;
2632 operands[0] = gen_rtx_fmt_ee (GET_CODE (operands[0]), VOIDmode,
2633 operands[1], operands[2]);
2634 emit_jump_insn (gen_cbranchcc4 (operands[0], operands[1], operands[2],
2639 /* Generate a DFmode part of a hard TFmode register.
2640 REG is the TFmode hard register, LOW is 1 for the
2641 low 64bit of the register and 0 otherwise.
2644 gen_df_reg (rtx reg, int low)
2646 int regno = REGNO (reg);
2648 if ((WORDS_BIG_ENDIAN == 0) ^ (low != 0))
2649 regno += (TARGET_ARCH64 && SPARC_INT_REG_P (regno)) ? 1 : 2;
2650 return gen_rtx_REG (DFmode, regno);
2653 /* Generate a call to FUNC with OPERANDS. Operand 0 is the return value.
2654 Unlike normal calls, TFmode operands are passed by reference. It is
2655 assumed that no more than 3 operands are required. */
2658 emit_soft_tfmode_libcall (const char *func_name, int nargs, rtx *operands)
2660 rtx ret_slot = NULL, arg[3], func_sym;
2663 /* We only expect to be called for conversions, unary, and binary ops. */
2664 gcc_assert (nargs == 2 || nargs == 3);
2666 for (i = 0; i < nargs; ++i)
2668 rtx this_arg = operands[i];
2671 /* TFmode arguments and return values are passed by reference. */
2672 if (GET_MODE (this_arg) == TFmode)
2674 int force_stack_temp;
2676 force_stack_temp = 0;
2677 if (TARGET_BUGGY_QP_LIB && i == 0)
2678 force_stack_temp = 1;
2680 if (GET_CODE (this_arg) == MEM
2681 && ! force_stack_temp)
2682 this_arg = XEXP (this_arg, 0);
2683 else if (CONSTANT_P (this_arg)
2684 && ! force_stack_temp)
2686 this_slot = force_const_mem (TFmode, this_arg);
2687 this_arg = XEXP (this_slot, 0);
2691 this_slot = assign_stack_temp (TFmode, GET_MODE_SIZE (TFmode), 0);
2693 /* Operand 0 is the return value. We'll copy it out later. */
2695 emit_move_insn (this_slot, this_arg);
2697 ret_slot = this_slot;
2699 this_arg = XEXP (this_slot, 0);
2706 func_sym = gen_rtx_SYMBOL_REF (Pmode, func_name);
2708 if (GET_MODE (operands[0]) == TFmode)
2711 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 2,
2712 arg[0], GET_MODE (arg[0]),
2713 arg[1], GET_MODE (arg[1]));
2715 emit_library_call (func_sym, LCT_NORMAL, VOIDmode, 3,
2716 arg[0], GET_MODE (arg[0]),
2717 arg[1], GET_MODE (arg[1]),
2718 arg[2], GET_MODE (arg[2]));
2721 emit_move_insn (operands[0], ret_slot);
2727 gcc_assert (nargs == 2);
2729 ret = emit_library_call_value (func_sym, operands[0], LCT_NORMAL,
2730 GET_MODE (operands[0]), 1,
2731 arg[1], GET_MODE (arg[1]));
2733 if (ret != operands[0])
2734 emit_move_insn (operands[0], ret);
2738 /* Expand soft-float TFmode calls to sparc abi routines. */
2741 emit_soft_tfmode_binop (enum rtx_code code, rtx *operands)
2763 emit_soft_tfmode_libcall (func, 3, operands);
2767 emit_soft_tfmode_unop (enum rtx_code code, rtx *operands)
2771 gcc_assert (code == SQRT);
2774 emit_soft_tfmode_libcall (func, 2, operands);
2778 emit_soft_tfmode_cvt (enum rtx_code code, rtx *operands)
2785 switch (GET_MODE (operands[1]))
2798 case FLOAT_TRUNCATE:
2799 switch (GET_MODE (operands[0]))
2813 switch (GET_MODE (operands[1]))
2818 operands[1] = gen_rtx_SIGN_EXTEND (DImode, operands[1]);
2828 case UNSIGNED_FLOAT:
2829 switch (GET_MODE (operands[1]))
2834 operands[1] = gen_rtx_ZERO_EXTEND (DImode, operands[1]);
2845 switch (GET_MODE (operands[0]))
2859 switch (GET_MODE (operands[0]))
2876 emit_soft_tfmode_libcall (func, 2, operands);
2879 /* Expand a hard-float tfmode operation. All arguments must be in
2883 emit_hard_tfmode_operation (enum rtx_code code, rtx *operands)
2887 if (GET_RTX_CLASS (code) == RTX_UNARY)
2889 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2890 op = gen_rtx_fmt_e (code, GET_MODE (operands[0]), operands[1]);
2894 operands[1] = force_reg (GET_MODE (operands[1]), operands[1]);
2895 operands[2] = force_reg (GET_MODE (operands[2]), operands[2]);
2896 op = gen_rtx_fmt_ee (code, GET_MODE (operands[0]),
2897 operands[1], operands[2]);
2900 if (register_operand (operands[0], VOIDmode))
2903 dest = gen_reg_rtx (GET_MODE (operands[0]));
2905 emit_insn (gen_rtx_SET (VOIDmode, dest, op));
2907 if (dest != operands[0])
2908 emit_move_insn (operands[0], dest);
2912 emit_tfmode_binop (enum rtx_code code, rtx *operands)
2914 if (TARGET_HARD_QUAD)
2915 emit_hard_tfmode_operation (code, operands);
2917 emit_soft_tfmode_binop (code, operands);
2921 emit_tfmode_unop (enum rtx_code code, rtx *operands)
2923 if (TARGET_HARD_QUAD)
2924 emit_hard_tfmode_operation (code, operands);
2926 emit_soft_tfmode_unop (code, operands);
2930 emit_tfmode_cvt (enum rtx_code code, rtx *operands)
2932 if (TARGET_HARD_QUAD)
2933 emit_hard_tfmode_operation (code, operands);
2935 emit_soft_tfmode_cvt (code, operands);
2938 /* Return nonzero if a branch/jump/call instruction will be emitting
2939 nop into its delay slot. */
2942 empty_delay_slot (rtx insn)
2946 /* If no previous instruction (should not happen), return true. */
2947 if (PREV_INSN (insn) == NULL)
2950 seq = NEXT_INSN (PREV_INSN (insn));
2951 if (GET_CODE (PATTERN (seq)) == SEQUENCE)
2957 /* Return nonzero if TRIAL can go into the call delay slot. */
2960 tls_call_delay (rtx trial)
2965 call __tls_get_addr, %tgd_call (foo)
2966 add %l7, %o0, %o0, %tgd_add (foo)
2967 while Sun as/ld does not. */
2968 if (TARGET_GNU_TLS || !TARGET_TLS)
2971 pat = PATTERN (trial);
2973 /* We must reject tgd_add{32|64}, i.e.
2974 (set (reg) (plus (reg) (unspec [(reg) (symbol_ref)] UNSPEC_TLSGD)))
2975 and tldm_add{32|64}, i.e.
2976 (set (reg) (plus (reg) (unspec [(reg) (symbol_ref)] UNSPEC_TLSLDM)))
2978 if (GET_CODE (pat) == SET
2979 && GET_CODE (SET_SRC (pat)) == PLUS)
2981 rtx unspec = XEXP (SET_SRC (pat), 1);
2983 if (GET_CODE (unspec) == UNSPEC
2984 && (XINT (unspec, 1) == UNSPEC_TLSGD
2985 || XINT (unspec, 1) == UNSPEC_TLSLDM))
2992 /* Return nonzero if TRIAL, an insn, can be combined with a 'restore'
2993 instruction. RETURN_P is true if the v9 variant 'return' is to be
2994 considered in the test too.
2996 TRIAL must be a SET whose destination is a REG appropriate for the
2997 'restore' instruction or, if RETURN_P is true, for the 'return'
3001 eligible_for_restore_insn (rtx trial, bool return_p)
3003 rtx pat = PATTERN (trial);
3004 rtx src = SET_SRC (pat);
3005 bool src_is_freg = false;
3008 /* Since we now can do moves between float and integer registers when
3009 VIS3 is enabled, we have to catch this case. We can allow such
3010 moves when doing a 'return' however. */
3012 if (GET_CODE (src_reg) == SUBREG)
3013 src_reg = SUBREG_REG (src_reg);
3014 if (GET_CODE (src_reg) == REG
3015 && SPARC_FP_REG_P (REGNO (src_reg)))
3018 /* The 'restore src,%g0,dest' pattern for word mode and below. */
3019 if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
3020 && arith_operand (src, GET_MODE (src))
3024 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
3026 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (SImode);
3029 /* The 'restore src,%g0,dest' pattern for double-word mode. */
3030 else if (GET_MODE_CLASS (GET_MODE (src)) != MODE_FLOAT
3031 && arith_double_operand (src, GET_MODE (src))
3033 return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode);
3035 /* The 'restore src,%g0,dest' pattern for float if no FPU. */
3036 else if (! TARGET_FPU && register_operand (src, SFmode))
3039 /* The 'restore src,%g0,dest' pattern for double if no FPU. */
3040 else if (! TARGET_FPU && TARGET_ARCH64 && register_operand (src, DFmode))
3043 /* If we have the 'return' instruction, anything that does not use
3044 local or output registers and can go into a delay slot wins. */
3047 && !epilogue_renumber (&pat, 1)
3048 && get_attr_in_uncond_branch_delay (trial)
3049 == IN_UNCOND_BRANCH_DELAY_TRUE)
3052 /* The 'restore src1,src2,dest' pattern for SImode. */
3053 else if (GET_CODE (src) == PLUS
3054 && register_operand (XEXP (src, 0), SImode)
3055 && arith_operand (XEXP (src, 1), SImode))
3058 /* The 'restore src1,src2,dest' pattern for DImode. */
3059 else if (GET_CODE (src) == PLUS
3060 && register_operand (XEXP (src, 0), DImode)
3061 && arith_double_operand (XEXP (src, 1), DImode))
3064 /* The 'restore src1,%lo(src2),dest' pattern. */
3065 else if (GET_CODE (src) == LO_SUM
3066 && ! TARGET_CM_MEDMID
3067 && ((register_operand (XEXP (src, 0), SImode)
3068 && immediate_operand (XEXP (src, 1), SImode))
3070 && register_operand (XEXP (src, 0), DImode)
3071 && immediate_operand (XEXP (src, 1), DImode))))
3074 /* The 'restore src,src,dest' pattern. */
3075 else if (GET_CODE (src) == ASHIFT
3076 && (register_operand (XEXP (src, 0), SImode)
3077 || register_operand (XEXP (src, 0), DImode))
3078 && XEXP (src, 1) == const1_rtx)
3084 /* Return nonzero if TRIAL can go into the function return's delay slot. */
3087 eligible_for_return_delay (rtx trial)
3092 if (GET_CODE (trial) != INSN)
3095 if (get_attr_length (trial) != 1)
3098 /* If the function uses __builtin_eh_return, the eh_return machinery
3099 occupies the delay slot. */
3100 if (crtl->calls_eh_return)
3103 /* In the case of a leaf or flat function, anything can go into the slot. */
3104 if (sparc_leaf_function_p || TARGET_FLAT)
3106 get_attr_in_uncond_branch_delay (trial) == IN_UNCOND_BRANCH_DELAY_TRUE;
3108 pat = PATTERN (trial);
3109 if (GET_CODE (pat) == PARALLEL)
3115 for (i = XVECLEN (pat, 0) - 1; i >= 0; i--)
3117 rtx expr = XVECEXP (pat, 0, i);
3118 if (GET_CODE (expr) != SET)
3120 if (GET_CODE (SET_DEST (expr)) != REG)
3122 regno = REGNO (SET_DEST (expr));
3123 if (regno >= 8 && regno < 24)
3126 return !epilogue_renumber (&pat, 1)
3127 && (get_attr_in_uncond_branch_delay (trial)
3128 == IN_UNCOND_BRANCH_DELAY_TRUE);
3131 if (GET_CODE (pat) != SET)
3134 if (GET_CODE (SET_DEST (pat)) != REG)
3137 regno = REGNO (SET_DEST (pat));
3139 /* Otherwise, only operations which can be done in tandem with
3140 a `restore' or `return' insn can go into the delay slot. */
3141 if (regno >= 8 && regno < 24)
3144 /* If this instruction sets up floating point register and we have a return
3145 instruction, it can probably go in. But restore will not work
3147 if (! SPARC_INT_REG_P (regno))
3149 && !epilogue_renumber (&pat, 1)
3150 && get_attr_in_uncond_branch_delay (trial)
3151 == IN_UNCOND_BRANCH_DELAY_TRUE);
3153 return eligible_for_restore_insn (trial, true);
3156 /* Return nonzero if TRIAL can go into the sibling call's delay slot. */
3159 eligible_for_sibcall_delay (rtx trial)
3163 if (GET_CODE (trial) != INSN || GET_CODE (PATTERN (trial)) != SET)
3166 if (get_attr_length (trial) != 1)
3169 pat = PATTERN (trial);
3171 if (sparc_leaf_function_p || TARGET_FLAT)
3173 /* If the tail call is done using the call instruction,
3174 we have to restore %o7 in the delay slot. */
3175 if (LEAF_SIBCALL_SLOT_RESERVED_P)
3178 /* %g1 is used to build the function address */
3179 if (reg_mentioned_p (gen_rtx_REG (Pmode, 1), pat))
3185 /* Otherwise, only operations which can be done in tandem with
3186 a `restore' insn can go into the delay slot. */
3187 if (GET_CODE (SET_DEST (pat)) != REG
3188 || (REGNO (SET_DEST (pat)) >= 8 && REGNO (SET_DEST (pat)) < 24)
3189 || ! SPARC_INT_REG_P (REGNO (SET_DEST (pat))))
3192 /* If it mentions %o7, it can't go in, because sibcall will clobber it
3194 if (reg_mentioned_p (gen_rtx_REG (Pmode, 15), pat))
3197 return eligible_for_restore_insn (trial, false);
3200 /* Determine if it's legal to put X into the constant pool. This
3201 is not possible if X contains the address of a symbol that is
3202 not constant (TLS) or not known at final link time (PIC). */
3205 sparc_cannot_force_const_mem (enum machine_mode mode, rtx x)
3207 switch (GET_CODE (x))
3212 /* Accept all non-symbolic constants. */
3216 /* Labels are OK iff we are non-PIC. */
3217 return flag_pic != 0;
3220 /* 'Naked' TLS symbol references are never OK,
3221 non-TLS symbols are OK iff we are non-PIC. */
3222 if (SYMBOL_REF_TLS_MODEL (x))
3225 return flag_pic != 0;
3228 return sparc_cannot_force_const_mem (mode, XEXP (x, 0));
3231 return sparc_cannot_force_const_mem (mode, XEXP (x, 0))
3232 || sparc_cannot_force_const_mem (mode, XEXP (x, 1));
3240 /* Global Offset Table support. */
3241 static GTY(()) rtx got_helper_rtx = NULL_RTX;
3242 static GTY(()) rtx global_offset_table_rtx = NULL_RTX;
3244 /* Return the SYMBOL_REF for the Global Offset Table. */
3246 static GTY(()) rtx sparc_got_symbol = NULL_RTX;
3251 if (!sparc_got_symbol)
3252 sparc_got_symbol = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_");
3254 return sparc_got_symbol;
3257 /* Ensure that we are not using patterns that are not OK with PIC. */
3267 op = recog_data.operand[i];
3268 gcc_assert (GET_CODE (op) != SYMBOL_REF
3269 && (GET_CODE (op) != CONST
3270 || (GET_CODE (XEXP (op, 0)) == MINUS
3271 && XEXP (XEXP (op, 0), 0) == sparc_got ()
3272 && GET_CODE (XEXP (XEXP (op, 0), 1)) == CONST)));
3279 /* Return true if X is an address which needs a temporary register when
3280 reloaded while generating PIC code. */
3283 pic_address_needs_scratch (rtx x)
3285 /* An address which is a symbolic plus a non SMALL_INT needs a temp reg. */
3286 if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS
3287 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
3288 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
3289 && ! SMALL_INT (XEXP (XEXP (x, 0), 1)))
3295 /* Determine if a given RTX is a valid constant. We already know this
3296 satisfies CONSTANT_P. */
3299 sparc_legitimate_constant_p (enum machine_mode mode, rtx x)
3301 switch (GET_CODE (x))
3305 if (sparc_tls_referenced_p (x))
3310 if (GET_MODE (x) == VOIDmode)
3313 /* Floating point constants are generally not ok.
3314 The only exception is 0.0 and all-ones in VIS. */
3316 && SCALAR_FLOAT_MODE_P (mode)
3317 && (const_zero_operand (x, mode)
3318 || const_all_ones_operand (x, mode)))
3324 /* Vector constants are generally not ok.
3325 The only exception is 0 or -1 in VIS. */
3327 && (const_zero_operand (x, mode)
3328 || const_all_ones_operand (x, mode)))
3340 /* Determine if a given RTX is a valid constant address. */
3343 constant_address_p (rtx x)
3345 switch (GET_CODE (x))
3353 if (flag_pic && pic_address_needs_scratch (x))
3355 return sparc_legitimate_constant_p (Pmode, x);
3358 return !flag_pic && sparc_legitimate_constant_p (Pmode, x);
3365 /* Nonzero if the constant value X is a legitimate general operand
3366 when generating PIC code. It is given that flag_pic is on and
3367 that X satisfies CONSTANT_P or is a CONST_DOUBLE. */
3370 legitimate_pic_operand_p (rtx x)
3372 if (pic_address_needs_scratch (x))
3374 if (sparc_tls_referenced_p (x))
3379 #define RTX_OK_FOR_OFFSET_P(X, MODE) \
3381 && INTVAL (X) >= -0x1000 \
3382 && INTVAL (X) < (0x1000 - GET_MODE_SIZE (MODE)))
3384 #define RTX_OK_FOR_OLO10_P(X, MODE) \
3386 && INTVAL (X) >= -0x1000 \
3387 && INTVAL (X) < (0xc00 - GET_MODE_SIZE (MODE)))
3389 /* Handle the TARGET_LEGITIMATE_ADDRESS_P target hook.
3391 On SPARC, the actual legitimate addresses must be REG+REG or REG+SMALLINT
3392 ordinarily. This changes a bit when generating PIC. */
3395 sparc_legitimate_address_p (enum machine_mode mode, rtx addr, bool strict)
3397 rtx rs1 = NULL, rs2 = NULL, imm1 = NULL;
3399 if (REG_P (addr) || GET_CODE (addr) == SUBREG)
3401 else if (GET_CODE (addr) == PLUS)
3403 rs1 = XEXP (addr, 0);
3404 rs2 = XEXP (addr, 1);
3406 /* Canonicalize. REG comes first, if there are no regs,
3407 LO_SUM comes first. */
3409 && GET_CODE (rs1) != SUBREG
3411 || GET_CODE (rs2) == SUBREG
3412 || (GET_CODE (rs2) == LO_SUM && GET_CODE (rs1) != LO_SUM)))
3414 rs1 = XEXP (addr, 1);
3415 rs2 = XEXP (addr, 0);
3419 && rs1 == pic_offset_table_rtx
3421 && GET_CODE (rs2) != SUBREG
3422 && GET_CODE (rs2) != LO_SUM
3423 && GET_CODE (rs2) != MEM
3424 && !(GET_CODE (rs2) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs2))
3425 && (! symbolic_operand (rs2, VOIDmode) || mode == Pmode)
3426 && (GET_CODE (rs2) != CONST_INT || SMALL_INT (rs2)))
3428 || GET_CODE (rs1) == SUBREG)
3429 && RTX_OK_FOR_OFFSET_P (rs2, mode)))
3434 else if ((REG_P (rs1) || GET_CODE (rs1) == SUBREG)
3435 && (REG_P (rs2) || GET_CODE (rs2) == SUBREG))
3437 /* We prohibit REG + REG for TFmode when there are no quad move insns
3438 and we consequently need to split. We do this because REG+REG
3439 is not an offsettable address. If we get the situation in reload
3440 where source and destination of a movtf pattern are both MEMs with
3441 REG+REG address, then only one of them gets converted to an
3442 offsettable address. */
3444 && ! (TARGET_ARCH64 && TARGET_HARD_QUAD))
3447 /* We prohibit REG + REG on ARCH32 if not optimizing for
3448 DFmode/DImode because then mem_min_alignment is likely to be zero
3449 after reload and the forced split would lack a matching splitter
3451 if (TARGET_ARCH32 && !optimize
3452 && (mode == DFmode || mode == DImode))
3455 else if (USE_AS_OFFSETABLE_LO10
3456 && GET_CODE (rs1) == LO_SUM
3458 && ! TARGET_CM_MEDMID
3459 && RTX_OK_FOR_OLO10_P (rs2, mode))
3462 imm1 = XEXP (rs1, 1);
3463 rs1 = XEXP (rs1, 0);
3464 if (!CONSTANT_P (imm1)
3465 || (GET_CODE (rs1) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs1)))
3469 else if (GET_CODE (addr) == LO_SUM)
3471 rs1 = XEXP (addr, 0);
3472 imm1 = XEXP (addr, 1);
3474 if (!CONSTANT_P (imm1)
3475 || (GET_CODE (rs1) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (rs1)))
3478 /* We can't allow TFmode in 32-bit mode, because an offset greater
3479 than the alignment (8) may cause the LO_SUM to overflow. */
3480 if (mode == TFmode && TARGET_ARCH32)
3483 else if (GET_CODE (addr) == CONST_INT && SMALL_INT (addr))
3488 if (GET_CODE (rs1) == SUBREG)
3489 rs1 = SUBREG_REG (rs1);
3495 if (GET_CODE (rs2) == SUBREG)
3496 rs2 = SUBREG_REG (rs2);
3503 if (!REGNO_OK_FOR_BASE_P (REGNO (rs1))
3504 || (rs2 && !REGNO_OK_FOR_BASE_P (REGNO (rs2))))
3509 if ((! SPARC_INT_REG_P (REGNO (rs1))
3510 && REGNO (rs1) != FRAME_POINTER_REGNUM
3511 && REGNO (rs1) < FIRST_PSEUDO_REGISTER)
3513 && (! SPARC_INT_REG_P (REGNO (rs2))
3514 && REGNO (rs2) != FRAME_POINTER_REGNUM
3515 && REGNO (rs2) < FIRST_PSEUDO_REGISTER)))
3521 /* Return the SYMBOL_REF for the tls_get_addr function. */
3523 static GTY(()) rtx sparc_tls_symbol = NULL_RTX;
3526 sparc_tls_get_addr (void)
3528 if (!sparc_tls_symbol)
3529 sparc_tls_symbol = gen_rtx_SYMBOL_REF (Pmode, "__tls_get_addr");
3531 return sparc_tls_symbol;
3534 /* Return the Global Offset Table to be used in TLS mode. */
3537 sparc_tls_got (void)
3539 /* In PIC mode, this is just the PIC offset table. */
3542 crtl->uses_pic_offset_table = 1;
3543 return pic_offset_table_rtx;
3546 /* In non-PIC mode, Sun as (unlike GNU as) emits PC-relative relocations for
3547 the GOT symbol with the 32-bit ABI, so we reload the GOT register. */
3548 if (TARGET_SUN_TLS && TARGET_ARCH32)
3550 load_got_register ();
3551 return global_offset_table_rtx;
3554 /* In all other cases, we load a new pseudo with the GOT symbol. */
3555 return copy_to_reg (sparc_got ());
3558 /* Return true if X contains a thread-local symbol. */
3561 sparc_tls_referenced_p (rtx x)
3563 if (!TARGET_HAVE_TLS)
3566 if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS)
3567 x = XEXP (XEXP (x, 0), 0);
3569 if (GET_CODE (x) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (x))
3572 /* That's all we handle in sparc_legitimize_tls_address for now. */
3576 /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute
3577 this (thread-local) address. */
3580 sparc_legitimize_tls_address (rtx addr)
3582 rtx temp1, temp2, temp3, ret, o0, got, insn;
3584 gcc_assert (can_create_pseudo_p ());
3586 if (GET_CODE (addr) == SYMBOL_REF)
3587 switch (SYMBOL_REF_TLS_MODEL (addr))
3589 case TLS_MODEL_GLOBAL_DYNAMIC:
3591 temp1 = gen_reg_rtx (SImode);
3592 temp2 = gen_reg_rtx (SImode);
3593 ret = gen_reg_rtx (Pmode);
3594 o0 = gen_rtx_REG (Pmode, 8);
3595 got = sparc_tls_got ();
3596 emit_insn (gen_tgd_hi22 (temp1, addr));
3597 emit_insn (gen_tgd_lo10 (temp2, temp1, addr));
3600 emit_insn (gen_tgd_add32 (o0, got, temp2, addr));
3601 insn = emit_call_insn (gen_tgd_call32 (o0, sparc_tls_get_addr (),
3606 emit_insn (gen_tgd_add64 (o0, got, temp2, addr));
3607 insn = emit_call_insn (gen_tgd_call64 (o0, sparc_tls_get_addr (),
3610 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), o0);
3611 insn = get_insns ();
3613 emit_libcall_block (insn, ret, o0, addr);
3616 case TLS_MODEL_LOCAL_DYNAMIC:
3618 temp1 = gen_reg_rtx (SImode);
3619 temp2 = gen_reg_rtx (SImode);
3620 temp3 = gen_reg_rtx (Pmode);
3621 ret = gen_reg_rtx (Pmode);
3622 o0 = gen_rtx_REG (Pmode, 8);
3623 got = sparc_tls_got ();
3624 emit_insn (gen_tldm_hi22 (temp1));
3625 emit_insn (gen_tldm_lo10 (temp2, temp1));
3628 emit_insn (gen_tldm_add32 (o0, got, temp2));
3629 insn = emit_call_insn (gen_tldm_call32 (o0, sparc_tls_get_addr (),
3634 emit_insn (gen_tldm_add64 (o0, got, temp2));
3635 insn = emit_call_insn (gen_tldm_call64 (o0, sparc_tls_get_addr (),
3638 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), o0);
3639 insn = get_insns ();
3641 emit_libcall_block (insn, temp3, o0,
3642 gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
3643 UNSPEC_TLSLD_BASE));
3644 temp1 = gen_reg_rtx (SImode);
3645 temp2 = gen_reg_rtx (SImode);
3646 emit_insn (gen_tldo_hix22 (temp1, addr));
3647 emit_insn (gen_tldo_lox10 (temp2, temp1, addr));
3649 emit_insn (gen_tldo_add32 (ret, temp3, temp2, addr));
3651 emit_insn (gen_tldo_add64 (ret, temp3, temp2, addr));
3654 case TLS_MODEL_INITIAL_EXEC:
3655 temp1 = gen_reg_rtx (SImode);
3656 temp2 = gen_reg_rtx (SImode);
3657 temp3 = gen_reg_rtx (Pmode);
3658 got = sparc_tls_got ();
3659 emit_insn (gen_tie_hi22 (temp1, addr));
3660 emit_insn (gen_tie_lo10 (temp2, temp1, addr));
3662 emit_insn (gen_tie_ld32 (temp3, got, temp2, addr));
3664 emit_insn (gen_tie_ld64 (temp3, got, temp2, addr));
3667 ret = gen_reg_rtx (Pmode);
3669 emit_insn (gen_tie_add32 (ret, gen_rtx_REG (Pmode, 7),
3672 emit_insn (gen_tie_add64 (ret, gen_rtx_REG (Pmode, 7),
3676 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp3);
3679 case TLS_MODEL_LOCAL_EXEC:
3680 temp1 = gen_reg_rtx (Pmode);
3681 temp2 = gen_reg_rtx (Pmode);
3684 emit_insn (gen_tle_hix22_sp32 (temp1, addr));
3685 emit_insn (gen_tle_lox10_sp32 (temp2, temp1, addr));
3689 emit_insn (gen_tle_hix22_sp64 (temp1, addr));
3690 emit_insn (gen_tle_lox10_sp64 (temp2, temp1, addr));
3692 ret = gen_rtx_PLUS (Pmode, gen_rtx_REG (Pmode, 7), temp2);
3699 else if (GET_CODE (addr) == CONST)
3703 gcc_assert (GET_CODE (XEXP (addr, 0)) == PLUS);
3705 base = sparc_legitimize_tls_address (XEXP (XEXP (addr, 0), 0));
3706 offset = XEXP (XEXP (addr, 0), 1);
3708 base = force_operand (base, NULL_RTX);
3709 if (!(GET_CODE (offset) == CONST_INT && SMALL_INT (offset)))
3710 offset = force_reg (Pmode, offset);
3711 ret = gen_rtx_PLUS (Pmode, base, offset);
3715 gcc_unreachable (); /* for now ... */
3720 /* Legitimize PIC addresses. If the address is already position-independent,
3721 we return ORIG. Newly generated position-independent addresses go into a
3722 reg. This is REG if nonzero, otherwise we allocate register(s) as
3726 sparc_legitimize_pic_address (rtx orig, rtx reg)
3728 bool gotdata_op = false;
3730 if (GET_CODE (orig) == SYMBOL_REF
3731 /* See the comment in sparc_expand_move. */
3732 || (GET_CODE (orig) == LABEL_REF && !can_use_mov_pic_label_ref (orig)))
3734 rtx pic_ref, address;
3739 gcc_assert (can_create_pseudo_p ());
3740 reg = gen_reg_rtx (Pmode);
3745 /* If not during reload, allocate another temp reg here for loading
3746 in the address, so that these instructions can be optimized
3748 rtx temp_reg = (! can_create_pseudo_p ()
3749 ? reg : gen_reg_rtx (Pmode));
3751 /* Must put the SYMBOL_REF inside an UNSPEC here so that cse
3752 won't get confused into thinking that these two instructions
3753 are loading in the true address of the symbol. If in the
3754 future a PIC rtx exists, that should be used instead. */
3757 emit_insn (gen_movdi_high_pic (temp_reg, orig));
3758 emit_insn (gen_movdi_lo_sum_pic (temp_reg, temp_reg, orig));
3762 emit_insn (gen_movsi_high_pic (temp_reg, orig));
3763 emit_insn (gen_movsi_lo_sum_pic (temp_reg, temp_reg, orig));
3771 crtl->uses_pic_offset_table = 1;
3775 insn = emit_insn (gen_movdi_pic_gotdata_op (reg,
3776 pic_offset_table_rtx,
3779 insn = emit_insn (gen_movsi_pic_gotdata_op (reg,
3780 pic_offset_table_rtx,
3786 = gen_const_mem (Pmode,
3787 gen_rtx_PLUS (Pmode,
3788 pic_offset_table_rtx, address));
3789 insn = emit_move_insn (reg, pic_ref);
3792 /* Put a REG_EQUAL note on this insn, so that it can be optimized
3794 set_unique_reg_note (insn, REG_EQUAL, orig);
3797 else if (GET_CODE (orig) == CONST)
3801 if (GET_CODE (XEXP (orig, 0)) == PLUS
3802 && XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx)
3807 gcc_assert (can_create_pseudo_p ());
3808 reg = gen_reg_rtx (Pmode);
3811 gcc_assert (GET_CODE (XEXP (orig, 0)) == PLUS);
3812 base = sparc_legitimize_pic_address (XEXP (XEXP (orig, 0), 0), reg);
3813 offset = sparc_legitimize_pic_address (XEXP (XEXP (orig, 0), 1),
3814 base == reg ? NULL_RTX : reg);
3816 if (GET_CODE (offset) == CONST_INT)
3818 if (SMALL_INT (offset))
3819 return plus_constant (base, INTVAL (offset));
3820 else if (can_create_pseudo_p ())
3821 offset = force_reg (Pmode, offset);
3823 /* If we reach here, then something is seriously wrong. */
3826 return gen_rtx_PLUS (Pmode, base, offset);
3828 else if (GET_CODE (orig) == LABEL_REF)
3829 /* ??? We ought to be checking that the register is live instead, in case
3830 it is eliminated. */
3831 crtl->uses_pic_offset_table = 1;
3836 /* Try machine-dependent ways of modifying an illegitimate address X
3837 to be legitimate. If we find one, return the new, valid address.
3839 OLDX is the address as it was before break_out_memory_refs was called.
3840 In some cases it is useful to look at this to decide what needs to be done.
3842 MODE is the mode of the operand pointed to by X.
3844 On SPARC, change REG+N into REG+REG, and REG+(X*Y) into REG+REG. */
3847 sparc_legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED,
3848 enum machine_mode mode)
3852 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == MULT)
3853 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3854 force_operand (XEXP (x, 0), NULL_RTX));
3855 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == MULT)
3856 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3857 force_operand (XEXP (x, 1), NULL_RTX));
3858 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == PLUS)
3859 x = gen_rtx_PLUS (Pmode, force_operand (XEXP (x, 0), NULL_RTX),
3861 if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == PLUS)
3862 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3863 force_operand (XEXP (x, 1), NULL_RTX));
3865 if (x != orig_x && sparc_legitimate_address_p (mode, x, FALSE))
3868 if (sparc_tls_referenced_p (x))
3869 x = sparc_legitimize_tls_address (x);
3871 x = sparc_legitimize_pic_address (x, NULL_RTX);
3872 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 1)))
3873 x = gen_rtx_PLUS (Pmode, XEXP (x, 0),
3874 copy_to_mode_reg (Pmode, XEXP (x, 1)));
3875 else if (GET_CODE (x) == PLUS && CONSTANT_ADDRESS_P (XEXP (x, 0)))
3876 x = gen_rtx_PLUS (Pmode, XEXP (x, 1),
3877 copy_to_mode_reg (Pmode, XEXP (x, 0)));
3878 else if (GET_CODE (x) == SYMBOL_REF
3879 || GET_CODE (x) == CONST
3880 || GET_CODE (x) == LABEL_REF)
3881 x = copy_to_suggested_reg (x, NULL_RTX, Pmode);
3886 /* Delegitimize an address that was legitimized by the above function. */
3889 sparc_delegitimize_address (rtx x)
3891 x = delegitimize_mem_from_attrs (x);
3893 if (GET_CODE (x) == LO_SUM && GET_CODE (XEXP (x, 1)) == UNSPEC)
3894 switch (XINT (XEXP (x, 1), 1))
3896 case UNSPEC_MOVE_PIC:
3898 x = XVECEXP (XEXP (x, 1), 0, 0);
3899 gcc_assert (GET_CODE (x) == SYMBOL_REF);
3905 /* This is generated by mov{si,di}_pic_label_ref in PIC mode. */
3906 if (GET_CODE (x) == MINUS
3907 && REG_P (XEXP (x, 0))
3908 && REGNO (XEXP (x, 0)) == PIC_OFFSET_TABLE_REGNUM
3909 && GET_CODE (XEXP (x, 1)) == LO_SUM
3910 && GET_CODE (XEXP (XEXP (x, 1), 1)) == UNSPEC
3911 && XINT (XEXP (XEXP (x, 1), 1), 1) == UNSPEC_MOVE_PIC_LABEL)
3913 x = XVECEXP (XEXP (XEXP (x, 1), 1), 0, 0);
3914 gcc_assert (GET_CODE (x) == LABEL_REF);
3920 /* SPARC implementation of LEGITIMIZE_RELOAD_ADDRESS. Returns a value to
3921 replace the input X, or the original X if no replacement is called for.
3922 The output parameter *WIN is 1 if the calling macro should goto WIN,
3925 For SPARC, we wish to handle addresses by splitting them into
3926 HIGH+LO_SUM pairs, retaining the LO_SUM in the memory reference.
3927 This cuts the number of extra insns by one.
3929 Do nothing when generating PIC code and the address is a symbolic
3930 operand or requires a scratch register. */
3933 sparc_legitimize_reload_address (rtx x, enum machine_mode mode,
3934 int opnum, int type,
3935 int ind_levels ATTRIBUTE_UNUSED, int *win)
3937 /* Decompose SImode constants into HIGH+LO_SUM. */
3939 && (mode != TFmode || TARGET_ARCH64)
3940 && GET_MODE (x) == SImode
3941 && GET_CODE (x) != LO_SUM
3942 && GET_CODE (x) != HIGH
3943 && sparc_cmodel <= CM_MEDLOW
3945 && (symbolic_operand (x, Pmode) || pic_address_needs_scratch (x))))
3947 x = gen_rtx_LO_SUM (GET_MODE (x), gen_rtx_HIGH (GET_MODE (x), x), x);
3948 push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL,
3949 BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0,
3950 opnum, (enum reload_type)type);
3955 /* We have to recognize what we have already generated above. */
3956 if (GET_CODE (x) == LO_SUM && GET_CODE (XEXP (x, 0)) == HIGH)
3958 push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL,
3959 BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0,
3960 opnum, (enum reload_type)type);
3969 /* Return true if ADDR (a legitimate address expression)
3970 has an effect that depends on the machine mode it is used for.
3976 is not equivalent to
3978 (mem:QI [%l7+a]) (mem:QI [%l7+a+1])
3980 because [%l7+a+1] is interpreted as the address of (a+1). */
3984 sparc_mode_dependent_address_p (const_rtx addr)
3986 if (flag_pic && GET_CODE (addr) == PLUS)
3988 rtx op0 = XEXP (addr, 0);
3989 rtx op1 = XEXP (addr, 1);
3990 if (op0 == pic_offset_table_rtx
3991 && symbolic_operand (op1, VOIDmode))
3998 #ifdef HAVE_GAS_HIDDEN
3999 # define USE_HIDDEN_LINKONCE 1
4001 # define USE_HIDDEN_LINKONCE 0
4005 get_pc_thunk_name (char name[32], unsigned int regno)
4007 const char *reg_name = reg_names[regno];
4009 /* Skip the leading '%' as that cannot be used in a
4013 if (USE_HIDDEN_LINKONCE)
4014 sprintf (name, "__sparc_get_pc_thunk.%s", reg_name);
4016 ASM_GENERATE_INTERNAL_LABEL (name, "LADDPC", regno);
4019 /* Wrapper around the load_pcrel_sym{si,di} patterns. */
4022 gen_load_pcrel_sym (rtx op0, rtx op1, rtx op2, rtx op3)
4024 int orig_flag_pic = flag_pic;
4027 /* The load_pcrel_sym{si,di} patterns require absolute addressing. */
4030 insn = gen_load_pcrel_symdi (op0, op1, op2, op3);
4032 insn = gen_load_pcrel_symsi (op0, op1, op2, op3);
4033 flag_pic = orig_flag_pic;
4038 /* Emit code to load the GOT register. */
4041 load_got_register (void)
4043 /* In PIC mode, this will retrieve pic_offset_table_rtx. */
4044 if (!global_offset_table_rtx)
4045 global_offset_table_rtx = gen_rtx_REG (Pmode, GLOBAL_OFFSET_TABLE_REGNUM);
4047 if (TARGET_VXWORKS_RTP)
4048 emit_insn (gen_vxworks_load_got ());
4051 /* The GOT symbol is subject to a PC-relative relocation so we need a
4052 helper function to add the PC value and thus get the final value. */
4053 if (!got_helper_rtx)
4056 get_pc_thunk_name (name, GLOBAL_OFFSET_TABLE_REGNUM);
4057 got_helper_rtx = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (name));
4060 emit_insn (gen_load_pcrel_sym (global_offset_table_rtx, sparc_got (),
4062 GEN_INT (GLOBAL_OFFSET_TABLE_REGNUM)));
4065 /* Need to emit this whether or not we obey regdecls,
4066 since setjmp/longjmp can cause life info to screw up.
4067 ??? In the case where we don't obey regdecls, this is not sufficient
4068 since we may not fall out the bottom. */
4069 emit_use (global_offset_table_rtx);
4072 /* Emit a call instruction with the pattern given by PAT. ADDR is the
4073 address of the call target. */
4076 sparc_emit_call_insn (rtx pat, rtx addr)
4080 insn = emit_call_insn (pat);
4082 /* The PIC register is live on entry to VxWorks PIC PLT entries. */
4083 if (TARGET_VXWORKS_RTP
4085 && GET_CODE (addr) == SYMBOL_REF
4086 && (SYMBOL_REF_DECL (addr)
4087 ? !targetm.binds_local_p (SYMBOL_REF_DECL (addr))
4088 : !SYMBOL_REF_LOCAL_P (addr)))
4090 use_reg (&CALL_INSN_FUNCTION_USAGE (insn), pic_offset_table_rtx);
4091 crtl->uses_pic_offset_table = 1;
4095 /* Return 1 if RTX is a MEM which is known to be aligned to at
4096 least a DESIRED byte boundary. */
4099 mem_min_alignment (rtx mem, int desired)
4101 rtx addr, base, offset;
4103 /* If it's not a MEM we can't accept it. */
4104 if (GET_CODE (mem) != MEM)
4108 if (!TARGET_UNALIGNED_DOUBLES
4109 && MEM_ALIGN (mem) / BITS_PER_UNIT >= (unsigned)desired)
4112 /* ??? The rest of the function predates MEM_ALIGN so
4113 there is probably a bit of redundancy. */
4114 addr = XEXP (mem, 0);
4115 base = offset = NULL_RTX;
4116 if (GET_CODE (addr) == PLUS)
4118 if (GET_CODE (XEXP (addr, 0)) == REG)
4120 base = XEXP (addr, 0);
4122 /* What we are saying here is that if the base
4123 REG is aligned properly, the compiler will make
4124 sure any REG based index upon it will be so
4126 if (GET_CODE (XEXP (addr, 1)) == CONST_INT)
4127 offset = XEXP (addr, 1);
4129 offset = const0_rtx;
4132 else if (GET_CODE (addr) == REG)
4135 offset = const0_rtx;
4138 if (base != NULL_RTX)
4140 int regno = REGNO (base);
4142 if (regno != HARD_FRAME_POINTER_REGNUM && regno != STACK_POINTER_REGNUM)
4144 /* Check if the compiler has recorded some information
4145 about the alignment of the base REG. If reload has
4146 completed, we already matched with proper alignments.
4147 If not running global_alloc, reload might give us
4148 unaligned pointer to local stack though. */
4150 && REGNO_POINTER_ALIGN (regno) >= desired * BITS_PER_UNIT)
4151 || (optimize && reload_completed))
4152 && (INTVAL (offset) & (desired - 1)) == 0)
4157 if (((INTVAL (offset) - SPARC_STACK_BIAS) & (desired - 1)) == 0)
4161 else if (! TARGET_UNALIGNED_DOUBLES
4162 || CONSTANT_P (addr)
4163 || GET_CODE (addr) == LO_SUM)
4165 /* Anything else we know is properly aligned unless TARGET_UNALIGNED_DOUBLES
4166 is true, in which case we can only assume that an access is aligned if
4167 it is to a constant address, or the address involves a LO_SUM. */
4171 /* An obviously unaligned address. */
4176 /* Vectors to keep interesting information about registers where it can easily
4177 be got. We used to use the actual mode value as the bit number, but there
4178 are more than 32 modes now. Instead we use two tables: one indexed by
4179 hard register number, and one indexed by mode. */
4181 /* The purpose of sparc_mode_class is to shrink the range of modes so that
4182 they all fit (as bit numbers) in a 32-bit word (again). Each real mode is
4183 mapped into one sparc_mode_class mode. */
4185 enum sparc_mode_class {
4186 S_MODE, D_MODE, T_MODE, O_MODE,
4187 SF_MODE, DF_MODE, TF_MODE, OF_MODE,
4191 /* Modes for single-word and smaller quantities. */
4192 #define S_MODES ((1 << (int) S_MODE) | (1 << (int) SF_MODE))
4194 /* Modes for double-word and smaller quantities. */
4195 #define D_MODES (S_MODES | (1 << (int) D_MODE) | (1 << DF_MODE))
4197 /* Modes for quad-word and smaller quantities. */
4198 #define T_MODES (D_MODES | (1 << (int) T_MODE) | (1 << (int) TF_MODE))
4200 /* Modes for 8-word and smaller quantities. */
4201 #define O_MODES (T_MODES | (1 << (int) O_MODE) | (1 << (int) OF_MODE))
4203 /* Modes for single-float quantities. We must allow any single word or
4204 smaller quantity. This is because the fix/float conversion instructions
4205 take integer inputs/outputs from the float registers. */
4206 #define SF_MODES (S_MODES)
4208 /* Modes for double-float and smaller quantities. */
4209 #define DF_MODES (D_MODES)
4211 /* Modes for quad-float and smaller quantities. */
4212 #define TF_MODES (DF_MODES | (1 << (int) TF_MODE))
4214 /* Modes for quad-float pairs and smaller quantities. */
4215 #define OF_MODES (TF_MODES | (1 << (int) OF_MODE))
4217 /* Modes for double-float only quantities. */
4218 #define DF_MODES_NO_S ((1 << (int) D_MODE) | (1 << (int) DF_MODE))
4220 /* Modes for quad-float and double-float only quantities. */
4221 #define TF_MODES_NO_S (DF_MODES_NO_S | (1 << (int) TF_MODE))
4223 /* Modes for quad-float pairs and double-float only quantities. */
4224 #define OF_MODES_NO_S (TF_MODES_NO_S | (1 << (int) OF_MODE))
4226 /* Modes for condition codes. */
4227 #define CC_MODES (1 << (int) CC_MODE)
4228 #define CCFP_MODES (1 << (int) CCFP_MODE)
4230 /* Value is 1 if register/mode pair is acceptable on sparc.
4231 The funny mixture of D and T modes is because integer operations
4232 do not specially operate on tetra quantities, so non-quad-aligned
4233 registers can hold quadword quantities (except %o4 and %i4 because
4234 they cross fixed registers). */
4236 /* This points to either the 32 bit or the 64 bit version. */
4237 const int *hard_regno_mode_classes;
4239 static const int hard_32bit_mode_classes[] = {
4240 S_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
4241 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
4242 T_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES,
4243 T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES,
4245 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4246 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4247 OF_MODES, SF_MODES, DF_MODES, SF_MODES, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4248 OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
4250 /* FP regs f32 to f63. Only the even numbered registers actually exist,
4251 and none can hold SFmode/SImode values. */
4252 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4253 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4254 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4255 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4258 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
4260 /* %icc, %sfp, %gsr */
4261 CC_MODES, 0, D_MODES
4264 static const int hard_64bit_mode_classes[] = {
4265 D_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4266 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4267 T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES,
4268 O_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_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, OF_MODES, SF_MODES, DF_MODES, SF_MODES,
4273 OF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES,
4275 /* FP regs f32 to f63. Only the even numbered registers actually exist,
4276 and none can hold SFmode/SImode values. */
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, OF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4280 OF_MODES_NO_S, 0, DF_MODES_NO_S, 0, TF_MODES_NO_S, 0, DF_MODES_NO_S, 0,
4283 CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES,
4285 /* %icc, %sfp, %gsr */
4286 CC_MODES, 0, D_MODES
4289 int sparc_mode_class [NUM_MACHINE_MODES];
4291 enum reg_class sparc_regno_reg_class[FIRST_PSEUDO_REGISTER];
4294 sparc_init_modes (void)
4298 for (i = 0; i < NUM_MACHINE_MODES; i++)
4300 switch (GET_MODE_CLASS (i))
4303 case MODE_PARTIAL_INT:
4304 case MODE_COMPLEX_INT:
4305 if (GET_MODE_SIZE (i) <= 4)
4306 sparc_mode_class[i] = 1 << (int) S_MODE;
4307 else if (GET_MODE_SIZE (i) == 8)
4308 sparc_mode_class[i] = 1 << (int) D_MODE;
4309 else if (GET_MODE_SIZE (i) == 16)
4310 sparc_mode_class[i] = 1 << (int) T_MODE;
4311 else if (GET_MODE_SIZE (i) == 32)
4312 sparc_mode_class[i] = 1 << (int) O_MODE;
4314 sparc_mode_class[i] = 0;
4316 case MODE_VECTOR_INT:
4317 if (GET_MODE_SIZE (i) <= 4)
4318 sparc_mode_class[i] = 1 << (int)SF_MODE;
4319 else if (GET_MODE_SIZE (i) == 8)
4320 sparc_mode_class[i] = 1 << (int)DF_MODE;
4323 case MODE_COMPLEX_FLOAT:
4324 if (GET_MODE_SIZE (i) <= 4)
4325 sparc_mode_class[i] = 1 << (int) SF_MODE;
4326 else if (GET_MODE_SIZE (i) == 8)
4327 sparc_mode_class[i] = 1 << (int) DF_MODE;
4328 else if (GET_MODE_SIZE (i) == 16)
4329 sparc_mode_class[i] = 1 << (int) TF_MODE;
4330 else if (GET_MODE_SIZE (i) == 32)
4331 sparc_mode_class[i] = 1 << (int) OF_MODE;
4333 sparc_mode_class[i] = 0;
4336 if (i == (int) CCFPmode || i == (int) CCFPEmode)
4337 sparc_mode_class[i] = 1 << (int) CCFP_MODE;
4339 sparc_mode_class[i] = 1 << (int) CC_MODE;
4342 sparc_mode_class[i] = 0;
4348 hard_regno_mode_classes = hard_64bit_mode_classes;
4350 hard_regno_mode_classes = hard_32bit_mode_classes;
4352 /* Initialize the array used by REGNO_REG_CLASS. */
4353 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
4355 if (i < 16 && TARGET_V8PLUS)
4356 sparc_regno_reg_class[i] = I64_REGS;
4357 else if (i < 32 || i == FRAME_POINTER_REGNUM)
4358 sparc_regno_reg_class[i] = GENERAL_REGS;
4360 sparc_regno_reg_class[i] = FP_REGS;
4362 sparc_regno_reg_class[i] = EXTRA_FP_REGS;
4364 sparc_regno_reg_class[i] = FPCC_REGS;
4366 sparc_regno_reg_class[i] = NO_REGS;
4370 /* Return whether REGNO, a global or FP register, must be saved/restored. */
4373 save_global_or_fp_reg_p (unsigned int regno,
4374 int leaf_function ATTRIBUTE_UNUSED)
4376 return !call_used_regs[regno] && df_regs_ever_live_p (regno);
4379 /* Return whether the return address register (%i7) is needed. */
4382 return_addr_reg_needed_p (int leaf_function)
4384 /* If it is live, for example because of __builtin_return_address (0). */
4385 if (df_regs_ever_live_p (RETURN_ADDR_REGNUM))
4388 /* Otherwise, it is needed as save register if %o7 is clobbered. */
4390 /* Loading the GOT register clobbers %o7. */
4391 || crtl->uses_pic_offset_table
4392 || df_regs_ever_live_p (INCOMING_RETURN_ADDR_REGNUM))
4398 /* Return whether REGNO, a local or in register, must be saved/restored. */
4401 save_local_or_in_reg_p (unsigned int regno, int leaf_function)
4403 /* General case: call-saved registers live at some point. */
4404 if (!call_used_regs[regno] && df_regs_ever_live_p (regno))
4407 /* Frame pointer register (%fp) if needed. */
4408 if (regno == HARD_FRAME_POINTER_REGNUM && frame_pointer_needed)
4411 /* Return address register (%i7) if needed. */
4412 if (regno == RETURN_ADDR_REGNUM && return_addr_reg_needed_p (leaf_function))
4415 /* GOT register (%l7) if needed. */
4416 if (regno == PIC_OFFSET_TABLE_REGNUM && crtl->uses_pic_offset_table)
4419 /* If the function accesses prior frames, the frame pointer and the return
4420 address of the previous frame must be saved on the stack. */
4421 if (crtl->accesses_prior_frames
4422 && (regno == HARD_FRAME_POINTER_REGNUM || regno == RETURN_ADDR_REGNUM))
4428 /* Compute the frame size required by the function. This function is called
4429 during the reload pass and also by sparc_expand_prologue. */
4432 sparc_compute_frame_size (HOST_WIDE_INT size, int leaf_function)
4434 HOST_WIDE_INT frame_size, apparent_frame_size;
4435 int args_size, n_global_fp_regs = 0;
4436 bool save_local_in_regs_p = false;
4439 /* If the function allocates dynamic stack space, the dynamic offset is
4440 computed early and contains REG_PARM_STACK_SPACE, so we need to cope. */
4441 if (leaf_function && !cfun->calls_alloca)
4444 args_size = crtl->outgoing_args_size + REG_PARM_STACK_SPACE (cfun->decl);
4446 /* Calculate space needed for global registers. */
4448 for (i = 0; i < 8; i++)
4449 if (save_global_or_fp_reg_p (i, 0))
4450 n_global_fp_regs += 2;
4452 for (i = 0; i < 8; i += 2)
4453 if (save_global_or_fp_reg_p (i, 0) || save_global_or_fp_reg_p (i + 1, 0))
4454 n_global_fp_regs += 2;
4456 /* In the flat window model, find out which local and in registers need to
4457 be saved. We don't reserve space in the current frame for them as they
4458 will be spilled into the register window save area of the caller's frame.
4459 However, as soon as we use this register window save area, we must create
4460 that of the current frame to make it the live one. */
4462 for (i = 16; i < 32; i++)
4463 if (save_local_or_in_reg_p (i, leaf_function))
4465 save_local_in_regs_p = true;
4469 /* Calculate space needed for FP registers. */
4470 for (i = 32; i < (TARGET_V9 ? 96 : 64); i += 2)
4471 if (save_global_or_fp_reg_p (i, 0) || save_global_or_fp_reg_p (i + 1, 0))
4472 n_global_fp_regs += 2;
4475 && n_global_fp_regs == 0
4477 && !save_local_in_regs_p)
4478 frame_size = apparent_frame_size = 0;
4481 /* We subtract STARTING_FRAME_OFFSET, remember it's negative. */
4482 apparent_frame_size = (size - STARTING_FRAME_OFFSET + 7) & -8;
4483 apparent_frame_size += n_global_fp_regs * 4;
4485 /* We need to add the size of the outgoing argument area. */
4486 frame_size = apparent_frame_size + ((args_size + 7) & -8);
4488 /* And that of the register window save area. */
4489 frame_size += FIRST_PARM_OFFSET (cfun->decl);
4491 /* Finally, bump to the appropriate alignment. */
4492 frame_size = SPARC_STACK_ALIGN (frame_size);
4495 /* Set up values for use in prologue and epilogue. */
4496 sparc_frame_size = frame_size;
4497 sparc_apparent_frame_size = apparent_frame_size;
4498 sparc_n_global_fp_regs = n_global_fp_regs;
4499 sparc_save_local_in_regs_p = save_local_in_regs_p;
4504 /* Output any necessary .register pseudo-ops. */
4507 sparc_output_scratch_registers (FILE *file ATTRIBUTE_UNUSED)
4509 #ifdef HAVE_AS_REGISTER_PSEUDO_OP
4515 /* Check if %g[2367] were used without
4516 .register being printed for them already. */
4517 for (i = 2; i < 8; i++)
4519 if (df_regs_ever_live_p (i)
4520 && ! sparc_hard_reg_printed [i])
4522 sparc_hard_reg_printed [i] = 1;
4523 /* %g7 is used as TLS base register, use #ignore
4524 for it instead of #scratch. */
4525 fprintf (file, "\t.register\t%%g%d, #%s\n", i,
4526 i == 7 ? "ignore" : "scratch");
4533 #define PROBE_INTERVAL (1 << STACK_CHECK_PROBE_INTERVAL_EXP)
4535 #if PROBE_INTERVAL > 4096
4536 #error Cannot use indexed addressing mode for stack probing
4539 /* Emit code to probe a range of stack addresses from FIRST to FIRST+SIZE,
4540 inclusive. These are offsets from the current stack pointer.
4542 Note that we don't use the REG+REG addressing mode for the probes because
4543 of the stack bias in 64-bit mode. And it doesn't really buy us anything
4544 so the advantages of having a single code win here. */
4547 sparc_emit_probe_stack_range (HOST_WIDE_INT first, HOST_WIDE_INT size)
4549 rtx g1 = gen_rtx_REG (Pmode, 1);
4551 /* See if we have a constant small number of probes to generate. If so,
4552 that's the easy case. */
4553 if (size <= PROBE_INTERVAL)
4555 emit_move_insn (g1, GEN_INT (first));
4556 emit_insn (gen_rtx_SET (VOIDmode, g1,
4557 gen_rtx_MINUS (Pmode, stack_pointer_rtx, g1)));
4558 emit_stack_probe (plus_constant (g1, -size));
4561 /* The run-time loop is made up of 10 insns in the generic case while the
4562 compile-time loop is made up of 4+2*(n-2) insns for n # of intervals. */
4563 else if (size <= 5 * PROBE_INTERVAL)
4567 emit_move_insn (g1, GEN_INT (first + PROBE_INTERVAL));
4568 emit_insn (gen_rtx_SET (VOIDmode, g1,
4569 gen_rtx_MINUS (Pmode, stack_pointer_rtx, g1)));
4570 emit_stack_probe (g1);
4572 /* Probe at FIRST + N * PROBE_INTERVAL for values of N from 2 until
4573 it exceeds SIZE. If only two probes are needed, this will not
4574 generate any code. Then probe at FIRST + SIZE. */
4575 for (i = 2 * PROBE_INTERVAL; i < size; i += PROBE_INTERVAL)
4577 emit_insn (gen_rtx_SET (VOIDmode, g1,
4578 plus_constant (g1, -PROBE_INTERVAL)));
4579 emit_stack_probe (g1);
4582 emit_stack_probe (plus_constant (g1, (i - PROBE_INTERVAL) - size));
4585 /* Otherwise, do the same as above, but in a loop. Note that we must be
4586 extra careful with variables wrapping around because we might be at
4587 the very top (or the very bottom) of the address space and we have
4588 to be able to handle this case properly; in particular, we use an
4589 equality test for the loop condition. */
4592 HOST_WIDE_INT rounded_size;
4593 rtx g4 = gen_rtx_REG (Pmode, 4);
4595 emit_move_insn (g1, GEN_INT (first));
4598 /* Step 1: round SIZE to the previous multiple of the interval. */
4600 rounded_size = size & -PROBE_INTERVAL;
4601 emit_move_insn (g4, GEN_INT (rounded_size));
4604 /* Step 2: compute initial and final value of the loop counter. */
4606 /* TEST_ADDR = SP + FIRST. */
4607 emit_insn (gen_rtx_SET (VOIDmode, g1,
4608 gen_rtx_MINUS (Pmode, stack_pointer_rtx, g1)));
4610 /* LAST_ADDR = SP + FIRST + ROUNDED_SIZE. */
4611 emit_insn (gen_rtx_SET (VOIDmode, g4, gen_rtx_MINUS (Pmode, g1, g4)));
4616 while (TEST_ADDR != LAST_ADDR)
4618 TEST_ADDR = TEST_ADDR + PROBE_INTERVAL
4622 probes at FIRST + N * PROBE_INTERVAL for values of N from 1
4623 until it is equal to ROUNDED_SIZE. */
4626 emit_insn (gen_probe_stack_rangedi (g1, g1, g4));
4628 emit_insn (gen_probe_stack_rangesi (g1, g1, g4));
4631 /* Step 4: probe at FIRST + SIZE if we cannot assert at compile-time
4632 that SIZE is equal to ROUNDED_SIZE. */
4634 if (size != rounded_size)
4635 emit_stack_probe (plus_constant (g4, rounded_size - size));
4638 /* Make sure nothing is scheduled before we are done. */
4639 emit_insn (gen_blockage ());
4642 /* Probe a range of stack addresses from REG1 to REG2 inclusive. These are
4643 absolute addresses. */
4646 output_probe_stack_range (rtx reg1, rtx reg2)
4648 static int labelno = 0;
4649 char loop_lab[32], end_lab[32];
4652 ASM_GENERATE_INTERNAL_LABEL (loop_lab, "LPSRL", labelno);
4653 ASM_GENERATE_INTERNAL_LABEL (end_lab, "LPSRE", labelno++);
4655 ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, loop_lab);
4657 /* Jump to END_LAB if TEST_ADDR == LAST_ADDR. */
4660 output_asm_insn ("cmp\t%0, %1", xops);
4662 fputs ("\tbe,pn\t%xcc,", asm_out_file);
4664 fputs ("\tbe\t", asm_out_file);
4665 assemble_name_raw (asm_out_file, end_lab);
4666 fputc ('\n', asm_out_file);
4668 /* TEST_ADDR = TEST_ADDR + PROBE_INTERVAL. */
4669 xops[1] = GEN_INT (-PROBE_INTERVAL);
4670 output_asm_insn (" add\t%0, %1, %0", xops);
4672 /* Probe at TEST_ADDR and branch. */
4674 fputs ("\tba,pt\t%xcc,", asm_out_file);
4676 fputs ("\tba\t", asm_out_file);
4677 assemble_name_raw (asm_out_file, loop_lab);
4678 fputc ('\n', asm_out_file);
4679 xops[1] = GEN_INT (SPARC_STACK_BIAS);
4680 output_asm_insn (" st\t%%g0, [%0+%1]", xops);
4682 ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, end_lab);
4687 /* Emit code to save/restore registers from LOW to HIGH at BASE+OFFSET as
4688 needed. LOW is supposed to be double-word aligned for 32-bit registers.
4689 SAVE_P decides whether a register must be saved/restored. ACTION_TRUE
4690 is the action to be performed if SAVE_P returns true and ACTION_FALSE
4691 the action to be performed if it returns false. Return the new offset. */
4693 typedef bool (*sorr_pred_t) (unsigned int, int);
4694 typedef enum { SORR_NONE, SORR_ADVANCE, SORR_SAVE, SORR_RESTORE } sorr_act_t;
4697 emit_save_or_restore_regs (unsigned int low, unsigned int high, rtx base,
4698 int offset, int leaf_function, sorr_pred_t save_p,
4699 sorr_act_t action_true, sorr_act_t action_false)
4704 if (TARGET_ARCH64 && high <= 32)
4708 for (i = low; i < high; i++)
4710 if (save_p (i, leaf_function))
4712 mem = gen_frame_mem (DImode, plus_constant (base, offset));
4713 if (action_true == SORR_SAVE)
4715 insn = emit_move_insn (mem, gen_rtx_REG (DImode, i));
4716 RTX_FRAME_RELATED_P (insn) = 1;
4718 else /* action_true == SORR_RESTORE */
4720 /* The frame pointer must be restored last since its old
4721 value may be used as base address for the frame. This
4722 is problematic in 64-bit mode only because of the lack
4723 of double-word load instruction. */
4724 if (i == HARD_FRAME_POINTER_REGNUM)
4727 emit_move_insn (gen_rtx_REG (DImode, i), mem);
4731 else if (action_false == SORR_ADVANCE)
4737 mem = gen_frame_mem (DImode, plus_constant (base, fp_offset));
4738 emit_move_insn (hard_frame_pointer_rtx, mem);
4743 for (i = low; i < high; i += 2)
4745 bool reg0 = save_p (i, leaf_function);
4746 bool reg1 = save_p (i + 1, leaf_function);
4747 enum machine_mode mode;
4752 mode = SPARC_INT_REG_P (i) ? DImode : DFmode;
4757 mode = SPARC_INT_REG_P (i) ? SImode : SFmode;
4762 mode = SPARC_INT_REG_P (i) ? SImode : SFmode;
4768 if (action_false == SORR_ADVANCE)
4773 mem = gen_frame_mem (mode, plus_constant (base, offset));
4774 if (action_true == SORR_SAVE)
4776 insn = emit_move_insn (mem, gen_rtx_REG (mode, regno));
4777 RTX_FRAME_RELATED_P (insn) = 1;
4781 mem = gen_frame_mem (SImode, plus_constant (base, offset));
4782 set1 = gen_rtx_SET (VOIDmode, mem,
4783 gen_rtx_REG (SImode, regno));
4784 RTX_FRAME_RELATED_P (set1) = 1;
4786 = gen_frame_mem (SImode, plus_constant (base, offset + 4));
4787 set2 = gen_rtx_SET (VOIDmode, mem,
4788 gen_rtx_REG (SImode, regno + 1));
4789 RTX_FRAME_RELATED_P (set2) = 1;
4790 add_reg_note (insn, REG_FRAME_RELATED_EXPR,
4791 gen_rtx_PARALLEL (VOIDmode,
4792 gen_rtvec (2, set1, set2)));
4795 else /* action_true == SORR_RESTORE */
4796 emit_move_insn (gen_rtx_REG (mode, regno), mem);
4798 /* Always preserve double-word alignment. */
4799 offset = (offset + 8) & -8;
4806 /* Emit code to adjust BASE to OFFSET. Return the new base. */
4809 emit_adjust_base_to_offset (rtx base, int offset)
4811 /* ??? This might be optimized a little as %g1 might already have a
4812 value close enough that a single add insn will do. */
4813 /* ??? Although, all of this is probably only a temporary fix because
4814 if %g1 can hold a function result, then sparc_expand_epilogue will
4815 lose (the result will be clobbered). */
4816 rtx new_base = gen_rtx_REG (Pmode, 1);
4817 emit_move_insn (new_base, GEN_INT (offset));
4818 emit_insn (gen_rtx_SET (VOIDmode,
4819 new_base, gen_rtx_PLUS (Pmode, base, new_base)));
4823 /* Emit code to save/restore call-saved global and FP registers. */
4826 emit_save_or_restore_global_fp_regs (rtx base, int offset, sorr_act_t action)
4828 if (offset < -4096 || offset + sparc_n_global_fp_regs * 4 > 4095)
4830 base = emit_adjust_base_to_offset (base, offset);
4835 = emit_save_or_restore_regs (0, 8, base, offset, 0,
4836 save_global_or_fp_reg_p, action, SORR_NONE);
4837 emit_save_or_restore_regs (32, TARGET_V9 ? 96 : 64, base, offset, 0,
4838 save_global_or_fp_reg_p, action, SORR_NONE);
4841 /* Emit code to save/restore call-saved local and in registers. */
4844 emit_save_or_restore_local_in_regs (rtx base, int offset, sorr_act_t action)
4846 if (offset < -4096 || offset + 16 * UNITS_PER_WORD > 4095)
4848 base = emit_adjust_base_to_offset (base, offset);
4852 emit_save_or_restore_regs (16, 32, base, offset, sparc_leaf_function_p,
4853 save_local_or_in_reg_p, action, SORR_ADVANCE);
4856 /* Emit a window_save insn. */
4859 emit_window_save (rtx increment)
4861 rtx insn = emit_insn (gen_window_save (increment));
4862 RTX_FRAME_RELATED_P (insn) = 1;
4864 /* The incoming return address (%o7) is saved in %i7. */
4865 add_reg_note (insn, REG_CFA_REGISTER,
4866 gen_rtx_SET (VOIDmode,
4867 gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM),
4869 INCOMING_RETURN_ADDR_REGNUM)));
4871 /* The window save event. */
4872 add_reg_note (insn, REG_CFA_WINDOW_SAVE, const0_rtx);
4874 /* The CFA is %fp, the hard frame pointer. */
4875 add_reg_note (insn, REG_CFA_DEF_CFA,
4876 plus_constant (hard_frame_pointer_rtx,
4877 INCOMING_FRAME_SP_OFFSET));
4882 /* Generate an increment for the stack pointer. */
4885 gen_stack_pointer_inc (rtx increment)
4887 return gen_rtx_SET (VOIDmode,
4889 gen_rtx_PLUS (Pmode,
4894 /* Generate a decrement for the stack pointer. */
4897 gen_stack_pointer_dec (rtx decrement)
4899 return gen_rtx_SET (VOIDmode,
4901 gen_rtx_MINUS (Pmode,
4906 /* Expand the function prologue. The prologue is responsible for reserving
4907 storage for the frame, saving the call-saved registers and loading the
4908 GOT register if needed. */
4911 sparc_expand_prologue (void)
4916 /* Compute a snapshot of current_function_uses_only_leaf_regs. Relying
4917 on the final value of the flag means deferring the prologue/epilogue
4918 expansion until just before the second scheduling pass, which is too
4919 late to emit multiple epilogues or return insns.
4921 Of course we are making the assumption that the value of the flag
4922 will not change between now and its final value. Of the three parts
4923 of the formula, only the last one can reasonably vary. Let's take a
4924 closer look, after assuming that the first two ones are set to true
4925 (otherwise the last value is effectively silenced).
4927 If only_leaf_regs_used returns false, the global predicate will also
4928 be false so the actual frame size calculated below will be positive.
4929 As a consequence, the save_register_window insn will be emitted in
4930 the instruction stream; now this insn explicitly references %fp
4931 which is not a leaf register so only_leaf_regs_used will always
4932 return false subsequently.
4934 If only_leaf_regs_used returns true, we hope that the subsequent
4935 optimization passes won't cause non-leaf registers to pop up. For
4936 example, the regrename pass has special provisions to not rename to
4937 non-leaf registers in a leaf function. */
4938 sparc_leaf_function_p
4939 = optimize > 0 && current_function_is_leaf && only_leaf_regs_used ();
4941 size = sparc_compute_frame_size (get_frame_size(), sparc_leaf_function_p);
4943 if (flag_stack_usage_info)
4944 current_function_static_stack_size = size;
4946 if (flag_stack_check == STATIC_BUILTIN_STACK_CHECK && size)
4947 sparc_emit_probe_stack_range (STACK_CHECK_PROTECT, size);
4951 else if (sparc_leaf_function_p)
4953 rtx size_int_rtx = GEN_INT (-size);
4956 insn = emit_insn (gen_stack_pointer_inc (size_int_rtx));
4957 else if (size <= 8192)
4959 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (-4096)));
4960 /* %sp is still the CFA register. */
4961 RTX_FRAME_RELATED_P (insn) = 1;
4962 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (4096 - size)));
4966 rtx size_rtx = gen_rtx_REG (Pmode, 1);
4967 emit_move_insn (size_rtx, size_int_rtx);
4968 insn = emit_insn (gen_stack_pointer_inc (size_rtx));
4969 add_reg_note (insn, REG_FRAME_RELATED_EXPR,
4970 gen_stack_pointer_inc (size_int_rtx));
4973 RTX_FRAME_RELATED_P (insn) = 1;
4977 rtx size_int_rtx = GEN_INT (-size);
4980 emit_window_save (size_int_rtx);
4981 else if (size <= 8192)
4983 emit_window_save (GEN_INT (-4096));
4984 /* %sp is not the CFA register anymore. */
4985 emit_insn (gen_stack_pointer_inc (GEN_INT (4096 - size)));
4989 rtx size_rtx = gen_rtx_REG (Pmode, 1);
4990 emit_move_insn (size_rtx, size_int_rtx);
4991 emit_window_save (size_rtx);
4995 if (sparc_leaf_function_p)
4997 sparc_frame_base_reg = stack_pointer_rtx;
4998 sparc_frame_base_offset = size + SPARC_STACK_BIAS;
5002 sparc_frame_base_reg = hard_frame_pointer_rtx;
5003 sparc_frame_base_offset = SPARC_STACK_BIAS;
5006 if (sparc_n_global_fp_regs > 0)
5007 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg,
5008 sparc_frame_base_offset
5009 - sparc_apparent_frame_size,
5012 /* Load the GOT register if needed. */
5013 if (crtl->uses_pic_offset_table)
5014 load_got_register ();
5016 /* Advertise that the data calculated just above are now valid. */
5017 sparc_prologue_data_valid_p = true;
5020 /* Expand the function prologue. The prologue is responsible for reserving
5021 storage for the frame, saving the call-saved registers and loading the
5022 GOT register if needed. */
5025 sparc_flat_expand_prologue (void)
5030 sparc_leaf_function_p = optimize > 0 && current_function_is_leaf;
5032 size = sparc_compute_frame_size (get_frame_size(), sparc_leaf_function_p);
5034 if (flag_stack_usage_info)
5035 current_function_static_stack_size = size;
5037 if (flag_stack_check == STATIC_BUILTIN_STACK_CHECK && size)
5038 sparc_emit_probe_stack_range (STACK_CHECK_PROTECT, size);
5040 if (sparc_save_local_in_regs_p)
5041 emit_save_or_restore_local_in_regs (stack_pointer_rtx, SPARC_STACK_BIAS,
5048 rtx size_int_rtx, size_rtx;
5050 size_rtx = size_int_rtx = GEN_INT (-size);
5052 /* We establish the frame (i.e. decrement the stack pointer) first, even
5053 if we use a frame pointer, because we cannot clobber any call-saved
5054 registers, including the frame pointer, if we haven't created a new
5055 register save area, for the sake of compatibility with the ABI. */
5057 insn = emit_insn (gen_stack_pointer_inc (size_int_rtx));
5058 else if (size <= 8192 && !frame_pointer_needed)
5060 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (-4096)));
5061 RTX_FRAME_RELATED_P (insn) = 1;
5062 insn = emit_insn (gen_stack_pointer_inc (GEN_INT (4096 - size)));
5066 size_rtx = gen_rtx_REG (Pmode, 1);
5067 emit_move_insn (size_rtx, size_int_rtx);
5068 insn = emit_insn (gen_stack_pointer_inc (size_rtx));
5069 add_reg_note (insn, REG_CFA_ADJUST_CFA,
5070 gen_stack_pointer_inc (size_int_rtx));
5072 RTX_FRAME_RELATED_P (insn) = 1;
5074 /* Ensure nothing is scheduled until after the frame is established. */
5075 emit_insn (gen_blockage ());
5077 if (frame_pointer_needed)
5079 insn = emit_insn (gen_rtx_SET (VOIDmode, hard_frame_pointer_rtx,
5080 gen_rtx_MINUS (Pmode,
5083 RTX_FRAME_RELATED_P (insn) = 1;
5085 add_reg_note (insn, REG_CFA_ADJUST_CFA,
5086 gen_rtx_SET (VOIDmode, hard_frame_pointer_rtx,
5087 plus_constant (stack_pointer_rtx,
5091 if (return_addr_reg_needed_p (sparc_leaf_function_p))
5093 rtx o7 = gen_rtx_REG (Pmode, INCOMING_RETURN_ADDR_REGNUM);
5094 rtx i7 = gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM);
5096 insn = emit_move_insn (i7, o7);
5097 RTX_FRAME_RELATED_P (insn) = 1;
5099 add_reg_note (insn, REG_CFA_REGISTER,
5100 gen_rtx_SET (VOIDmode, i7, o7));
5102 /* Prevent this instruction from ever being considered dead,
5103 even if this function has no epilogue. */
5104 emit_insn (gen_rtx_USE (VOIDmode, i7));
5108 if (frame_pointer_needed)
5110 sparc_frame_base_reg = hard_frame_pointer_rtx;
5111 sparc_frame_base_offset = SPARC_STACK_BIAS;
5115 sparc_frame_base_reg = stack_pointer_rtx;
5116 sparc_frame_base_offset = size + SPARC_STACK_BIAS;
5119 if (sparc_n_global_fp_regs > 0)
5120 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg,
5121 sparc_frame_base_offset
5122 - sparc_apparent_frame_size,
5125 /* Load the GOT register if needed. */
5126 if (crtl->uses_pic_offset_table)
5127 load_got_register ();
5129 /* Advertise that the data calculated just above are now valid. */
5130 sparc_prologue_data_valid_p = true;
5133 /* This function generates the assembly code for function entry, which boils
5134 down to emitting the necessary .register directives. */
5137 sparc_asm_function_prologue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
5139 /* Check that the assumption we made in sparc_expand_prologue is valid. */
5141 gcc_assert (sparc_leaf_function_p == current_function_uses_only_leaf_regs);
5143 sparc_output_scratch_registers (file);
5146 /* Expand the function epilogue, either normal or part of a sibcall.
5147 We emit all the instructions except the return or the call. */
5150 sparc_expand_epilogue (bool for_eh)
5152 HOST_WIDE_INT size = sparc_frame_size;
5154 if (sparc_n_global_fp_regs > 0)
5155 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg,
5156 sparc_frame_base_offset
5157 - sparc_apparent_frame_size,
5160 if (size == 0 || for_eh)
5162 else if (sparc_leaf_function_p)
5165 emit_insn (gen_stack_pointer_dec (GEN_INT (-size)));
5166 else if (size <= 8192)
5168 emit_insn (gen_stack_pointer_dec (GEN_INT (-4096)));
5169 emit_insn (gen_stack_pointer_dec (GEN_INT (4096 - size)));
5173 rtx reg = gen_rtx_REG (Pmode, 1);
5174 emit_move_insn (reg, GEN_INT (-size));
5175 emit_insn (gen_stack_pointer_dec (reg));
5180 /* Expand the function epilogue, either normal or part of a sibcall.
5181 We emit all the instructions except the return or the call. */
5184 sparc_flat_expand_epilogue (bool for_eh)
5186 HOST_WIDE_INT size = sparc_frame_size;
5188 if (sparc_n_global_fp_regs > 0)
5189 emit_save_or_restore_global_fp_regs (sparc_frame_base_reg,
5190 sparc_frame_base_offset
5191 - sparc_apparent_frame_size,
5194 /* If we have a frame pointer, we'll need both to restore it before the
5195 frame is destroyed and use its current value in destroying the frame.
5196 Since we don't have an atomic way to do that in the flat window model,
5197 we save the current value into a temporary register (%g1). */
5198 if (frame_pointer_needed && !for_eh)
5199 emit_move_insn (gen_rtx_REG (Pmode, 1), hard_frame_pointer_rtx);
5201 if (return_addr_reg_needed_p (sparc_leaf_function_p))
5202 emit_move_insn (gen_rtx_REG (Pmode, INCOMING_RETURN_ADDR_REGNUM),
5203 gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM));
5205 if (sparc_save_local_in_regs_p)
5206 emit_save_or_restore_local_in_regs (sparc_frame_base_reg,
5207 sparc_frame_base_offset,
5210 if (size == 0 || for_eh)
5212 else if (frame_pointer_needed)
5214 /* Make sure the frame is destroyed after everything else is done. */
5215 emit_insn (gen_blockage ());
5217 emit_move_insn (stack_pointer_rtx, gen_rtx_REG (Pmode, 1));
5222 emit_insn (gen_blockage ());
5225 emit_insn (gen_stack_pointer_dec (GEN_INT (-size)));
5226 else if (size <= 8192)
5228 emit_insn (gen_stack_pointer_dec (GEN_INT (-4096)));
5229 emit_insn (gen_stack_pointer_dec (GEN_INT (4096 - size)));
5233 rtx reg = gen_rtx_REG (Pmode, 1);
5234 emit_move_insn (reg, GEN_INT (-size));
5235 emit_insn (gen_stack_pointer_dec (reg));
5240 /* Return true if it is appropriate to emit `return' instructions in the
5241 body of a function. */
5244 sparc_can_use_return_insn_p (void)
5246 return sparc_prologue_data_valid_p
5247 && sparc_n_global_fp_regs == 0
5249 ? (sparc_frame_size == 0 && !sparc_save_local_in_regs_p)
5250 : (sparc_frame_size == 0 || !sparc_leaf_function_p);
5253 /* This function generates the assembly code for function exit. */
5256 sparc_asm_function_epilogue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
5258 /* If the last two instructions of a function are "call foo; dslot;"
5259 the return address might point to the first instruction in the next
5260 function and we have to output a dummy nop for the sake of sane
5261 backtraces in such cases. This is pointless for sibling calls since
5262 the return address is explicitly adjusted. */
5264 rtx insn, last_real_insn;
5266 insn = get_last_insn ();
5268 last_real_insn = prev_real_insn (insn);
5270 && GET_CODE (last_real_insn) == INSN
5271 && GET_CODE (PATTERN (last_real_insn)) == SEQUENCE)
5272 last_real_insn = XVECEXP (PATTERN (last_real_insn), 0, 0);
5275 && CALL_P (last_real_insn)
5276 && !SIBLING_CALL_P (last_real_insn))
5277 fputs("\tnop\n", file);
5279 sparc_output_deferred_case_vectors ();
5282 /* Output a 'restore' instruction. */
5285 output_restore (rtx pat)
5291 fputs ("\t restore\n", asm_out_file);
5295 gcc_assert (GET_CODE (pat) == SET);
5297 operands[0] = SET_DEST (pat);
5298 pat = SET_SRC (pat);
5300 switch (GET_CODE (pat))
5303 operands[1] = XEXP (pat, 0);
5304 operands[2] = XEXP (pat, 1);
5305 output_asm_insn (" restore %r1, %2, %Y0", operands);
5308 operands[1] = XEXP (pat, 0);
5309 operands[2] = XEXP (pat, 1);
5310 output_asm_insn (" restore %r1, %%lo(%a2), %Y0", operands);
5313 operands[1] = XEXP (pat, 0);
5314 gcc_assert (XEXP (pat, 1) == const1_rtx);
5315 output_asm_insn (" restore %r1, %r1, %Y0", operands);
5319 output_asm_insn (" restore %%g0, %1, %Y0", operands);
5324 /* Output a return. */
5327 output_return (rtx insn)
5329 if (crtl->calls_eh_return)
5331 /* If the function uses __builtin_eh_return, the eh_return
5332 machinery occupies the delay slot. */
5333 gcc_assert (!final_sequence);
5335 if (flag_delayed_branch)
5337 if (!TARGET_FLAT && TARGET_V9)
5338 fputs ("\treturn\t%i7+8\n", asm_out_file);
5342 fputs ("\trestore\n", asm_out_file);
5344 fputs ("\tjmp\t%o7+8\n", asm_out_file);
5347 fputs ("\t add\t%sp, %g1, %sp\n", asm_out_file);
5352 fputs ("\trestore\n", asm_out_file);
5354 fputs ("\tadd\t%sp, %g1, %sp\n", asm_out_file);
5355 fputs ("\tjmp\t%o7+8\n\t nop\n", asm_out_file);
5358 else if (sparc_leaf_function_p || TARGET_FLAT)
5360 /* This is a leaf or flat function so we don't have to bother restoring
5361 the register window, which frees us from dealing with the convoluted
5362 semantics of restore/return. We simply output the jump to the
5363 return address and the insn in the delay slot (if any). */
5365 return "jmp\t%%o7+%)%#";
5369 /* This is a regular function so we have to restore the register window.
5370 We may have a pending insn for the delay slot, which will be either
5371 combined with the 'restore' instruction or put in the delay slot of
5372 the 'return' instruction. */
5378 delay = NEXT_INSN (insn);
5381 pat = PATTERN (delay);
5383 if (TARGET_V9 && ! epilogue_renumber (&pat, 1))
5385 epilogue_renumber (&pat, 0);
5386 return "return\t%%i7+%)%#";
5390 output_asm_insn ("jmp\t%%i7+%)", NULL);
5391 output_restore (pat);
5392 PATTERN (delay) = gen_blockage ();
5393 INSN_CODE (delay) = -1;
5398 /* The delay slot is empty. */
5400 return "return\t%%i7+%)\n\t nop";
5401 else if (flag_delayed_branch)
5402 return "jmp\t%%i7+%)\n\t restore";
5404 return "restore\n\tjmp\t%%o7+%)\n\t nop";
5411 /* Output a sibling call. */
5414 output_sibcall (rtx insn, rtx call_operand)
5418 gcc_assert (flag_delayed_branch);
5420 operands[0] = call_operand;
5422 if (sparc_leaf_function_p || TARGET_FLAT)
5424 /* This is a leaf or flat function so we don't have to bother restoring
5425 the register window. We simply output the jump to the function and
5426 the insn in the delay slot (if any). */
5428 gcc_assert (!(LEAF_SIBCALL_SLOT_RESERVED_P && final_sequence));
5431 output_asm_insn ("sethi\t%%hi(%a0), %%g1\n\tjmp\t%%g1 + %%lo(%a0)%#",
5434 /* Use or with rs2 %%g0 instead of mov, so that as/ld can optimize
5435 it into branch if possible. */
5436 output_asm_insn ("or\t%%o7, %%g0, %%g1\n\tcall\t%a0, 0\n\t or\t%%g1, %%g0, %%o7",
5441 /* This is a regular function so we have to restore the register window.
5442 We may have a pending insn for the delay slot, which will be combined
5443 with the 'restore' instruction. */
5445 output_asm_insn ("call\t%a0, 0", operands);
5449 rtx delay = NEXT_INSN (insn);
5452 output_restore (PATTERN (delay));
5454 PATTERN (delay) = gen_blockage ();
5455 INSN_CODE (delay) = -1;
5458 output_restore (NULL_RTX);
5464 /* Functions for handling argument passing.
5466 For 32-bit, the first 6 args are normally in registers and the rest are
5467 pushed. Any arg that starts within the first 6 words is at least
5468 partially passed in a register unless its data type forbids.
5470 For 64-bit, the argument registers are laid out as an array of 16 elements
5471 and arguments are added sequentially. The first 6 int args and up to the
5472 first 16 fp args (depending on size) are passed in regs.
5474 Slot Stack Integral Float Float in structure Double Long Double
5475 ---- ----- -------- ----- ------------------ ------ -----------
5476 15 [SP+248] %f31 %f30,%f31 %d30
5477 14 [SP+240] %f29 %f28,%f29 %d28 %q28
5478 13 [SP+232] %f27 %f26,%f27 %d26
5479 12 [SP+224] %f25 %f24,%f25 %d24 %q24
5480 11 [SP+216] %f23 %f22,%f23 %d22
5481 10 [SP+208] %f21 %f20,%f21 %d20 %q20
5482 9 [SP+200] %f19 %f18,%f19 %d18
5483 8 [SP+192] %f17 %f16,%f17 %d16 %q16
5484 7 [SP+184] %f15 %f14,%f15 %d14
5485 6 [SP+176] %f13 %f12,%f13 %d12 %q12
5486 5 [SP+168] %o5 %f11 %f10,%f11 %d10
5487 4 [SP+160] %o4 %f9 %f8,%f9 %d8 %q8
5488 3 [SP+152] %o3 %f7 %f6,%f7 %d6
5489 2 [SP+144] %o2 %f5 %f4,%f5 %d4 %q4
5490 1 [SP+136] %o1 %f3 %f2,%f3 %d2
5491 0 [SP+128] %o0 %f1 %f0,%f1 %d0 %q0
5493 Here SP = %sp if -mno-stack-bias or %sp+stack_bias otherwise.
5495 Integral arguments are always passed as 64-bit quantities appropriately
5498 Passing of floating point values is handled as follows.
5499 If a prototype is in scope:
5500 If the value is in a named argument (i.e. not a stdarg function or a
5501 value not part of the `...') then the value is passed in the appropriate
5503 If the value is part of the `...' and is passed in one of the first 6
5504 slots then the value is passed in the appropriate int reg.
5505 If the value is part of the `...' and is not passed in one of the first 6
5506 slots then the value is passed in memory.
5507 If a prototype is not in scope:
5508 If the value is one of the first 6 arguments the value is passed in the
5509 appropriate integer reg and the appropriate fp reg.
5510 If the value is not one of the first 6 arguments the value is passed in
5511 the appropriate fp reg and in memory.
5514 Summary of the calling conventions implemented by GCC on the SPARC:
5517 size argument return value
5519 small integer <4 int. reg. int. reg.
5520 word 4 int. reg. int. reg.
5521 double word 8 int. reg. int. reg.
5523 _Complex small integer <8 int. reg. int. reg.
5524 _Complex word 8 int. reg. int. reg.
5525 _Complex double word 16 memory int. reg.
5527 vector integer <=8 int. reg. FP reg.
5528 vector integer >8 memory memory
5530 float 4 int. reg. FP reg.
5531 double 8 int. reg. FP reg.
5532 long double 16 memory memory
5534 _Complex float 8 memory FP reg.
5535 _Complex double 16 memory FP reg.
5536 _Complex long double 32 memory FP reg.
5538 vector float any memory memory
5540 aggregate any memory memory
5545 size argument return value
5547 small integer <8 int. reg. int. reg.
5548 word 8 int. reg. int. reg.
5549 double word 16 int. reg. int. reg.
5551 _Complex small integer <16 int. reg. int. reg.
5552 _Complex word 16 int. reg. int. reg.
5553 _Complex double word 32 memory int. reg.
5555 vector integer <=16 FP reg. FP reg.
5556 vector integer 16<s<=32 memory FP reg.
5557 vector integer >32 memory memory
5559 float 4 FP reg. FP reg.
5560 double 8 FP reg. FP reg.
5561 long double 16 FP reg. FP reg.
5563 _Complex float 8 FP reg. FP reg.
5564 _Complex double 16 FP reg. FP reg.
5565 _Complex long double 32 memory FP reg.
5567 vector float <=16 FP reg. FP reg.
5568 vector float 16<s<=32 memory FP reg.
5569 vector float >32 memory memory
5571 aggregate <=16 reg. reg.
5572 aggregate 16<s<=32 memory reg.
5573 aggregate >32 memory memory
5577 Note #1: complex floating-point types follow the extended SPARC ABIs as
5578 implemented by the Sun compiler.
5580 Note #2: integral vector types follow the scalar floating-point types
5581 conventions to match what is implemented by the Sun VIS SDK.
5583 Note #3: floating-point vector types follow the aggregate types
5587 /* Maximum number of int regs for args. */
5588 #define SPARC_INT_ARG_MAX 6
5589 /* Maximum number of fp regs for args. */
5590 #define SPARC_FP_ARG_MAX 16
5592 #define ROUND_ADVANCE(SIZE) (((SIZE) + UNITS_PER_WORD - 1) / UNITS_PER_WORD)
5594 /* Handle the INIT_CUMULATIVE_ARGS macro.
5595 Initialize a variable CUM of type CUMULATIVE_ARGS
5596 for a call to a function whose data type is FNTYPE.
5597 For a library call, FNTYPE is 0. */
5600 init_cumulative_args (struct sparc_args *cum, tree fntype,
5601 rtx libname ATTRIBUTE_UNUSED,
5602 tree fndecl ATTRIBUTE_UNUSED)
5605 cum->prototype_p = fntype && prototype_p (fntype);
5606 cum->libcall_p = fntype == 0;
5609 /* Handle promotion of pointer and integer arguments. */
5611 static enum machine_mode
5612 sparc_promote_function_mode (const_tree type,
5613 enum machine_mode mode,
5615 const_tree fntype ATTRIBUTE_UNUSED,
5616 int for_return ATTRIBUTE_UNUSED)
5618 if (type != NULL_TREE && POINTER_TYPE_P (type))
5620 *punsignedp = POINTERS_EXTEND_UNSIGNED;
5624 /* Integral arguments are passed as full words, as per the ABI. */
5625 if (GET_MODE_CLASS (mode) == MODE_INT
5626 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5632 /* Handle the TARGET_STRICT_ARGUMENT_NAMING target hook. */
5635 sparc_strict_argument_naming (cumulative_args_t ca ATTRIBUTE_UNUSED)
5637 return TARGET_ARCH64 ? true : false;
5640 /* Scan the record type TYPE and return the following predicates:
5641 - INTREGS_P: the record contains at least one field or sub-field
5642 that is eligible for promotion in integer registers.
5643 - FP_REGS_P: the record contains at least one field or sub-field
5644 that is eligible for promotion in floating-point registers.
5645 - PACKED_P: the record contains at least one field that is packed.
5647 Sub-fields are not taken into account for the PACKED_P predicate. */
5650 scan_record_type (const_tree type, int *intregs_p, int *fpregs_p,
5655 for (field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field))
5657 if (TREE_CODE (field) == FIELD_DECL)
5659 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
5660 scan_record_type (TREE_TYPE (field), intregs_p, fpregs_p, 0);
5661 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
5662 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
5668 if (packed_p && DECL_PACKED (field))
5674 /* Compute the slot number to pass an argument in.
5675 Return the slot number or -1 if passing on the stack.
5677 CUM is a variable of type CUMULATIVE_ARGS which gives info about
5678 the preceding args and about the function being called.
5679 MODE is the argument's machine mode.
5680 TYPE is the data type of the argument (as a tree).
5681 This is null for libcalls where that information may
5683 NAMED is nonzero if this argument is a named parameter
5684 (otherwise it is an extra parameter matching an ellipsis).
5685 INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG.
5686 *PREGNO records the register number to use if scalar type.
5687 *PPADDING records the amount of padding needed in words. */
5690 function_arg_slotno (const struct sparc_args *cum, enum machine_mode mode,
5691 const_tree type, bool named, bool incoming_p,
5692 int *pregno, int *ppadding)
5694 int regbase = (incoming_p
5695 ? SPARC_INCOMING_INT_ARG_FIRST
5696 : SPARC_OUTGOING_INT_ARG_FIRST);
5697 int slotno = cum->words;
5698 enum mode_class mclass;
5703 if (type && TREE_ADDRESSABLE (type))
5709 && TYPE_ALIGN (type) % PARM_BOUNDARY != 0)
5712 /* For SPARC64, objects requiring 16-byte alignment get it. */
5714 && (type ? TYPE_ALIGN (type) : GET_MODE_ALIGNMENT (mode)) >= 128
5715 && (slotno & 1) != 0)
5716 slotno++, *ppadding = 1;
5718 mclass = GET_MODE_CLASS (mode);
5719 if (type && TREE_CODE (type) == VECTOR_TYPE)
5721 /* Vector types deserve special treatment because they are
5722 polymorphic wrt their mode, depending upon whether VIS
5723 instructions are enabled. */
5724 if (TREE_CODE (TREE_TYPE (type)) == REAL_TYPE)
5726 /* The SPARC port defines no floating-point vector modes. */
5727 gcc_assert (mode == BLKmode);
5731 /* Integral vector types should either have a vector
5732 mode or an integral mode, because we are guaranteed
5733 by pass_by_reference that their size is not greater
5734 than 16 bytes and TImode is 16-byte wide. */
5735 gcc_assert (mode != BLKmode);
5737 /* Vector integers are handled like floats according to
5739 mclass = MODE_FLOAT;
5746 case MODE_COMPLEX_FLOAT:
5747 case MODE_VECTOR_INT:
5748 if (TARGET_ARCH64 && TARGET_FPU && named)
5750 if (slotno >= SPARC_FP_ARG_MAX)
5752 regno = SPARC_FP_ARG_FIRST + slotno * 2;
5753 /* Arguments filling only one single FP register are
5754 right-justified in the outer double FP register. */
5755 if (GET_MODE_SIZE (mode) <= 4)
5762 case MODE_COMPLEX_INT:
5763 if (slotno >= SPARC_INT_ARG_MAX)
5765 regno = regbase + slotno;
5769 if (mode == VOIDmode)
5770 /* MODE is VOIDmode when generating the actual call. */
5773 gcc_assert (mode == BLKmode);
5777 || (TREE_CODE (type) != VECTOR_TYPE
5778 && TREE_CODE (type) != RECORD_TYPE))
5780 if (slotno >= SPARC_INT_ARG_MAX)
5782 regno = regbase + slotno;
5784 else /* TARGET_ARCH64 && type */
5786 int intregs_p = 0, fpregs_p = 0, packed_p = 0;
5788 /* First see what kinds of registers we would need. */
5789 if (TREE_CODE (type) == VECTOR_TYPE)
5792 scan_record_type (type, &intregs_p, &fpregs_p, &packed_p);
5794 /* The ABI obviously doesn't specify how packed structures
5795 are passed. These are defined to be passed in int regs
5796 if possible, otherwise memory. */
5797 if (packed_p || !named)
5798 fpregs_p = 0, intregs_p = 1;
5800 /* If all arg slots are filled, then must pass on stack. */
5801 if (fpregs_p && slotno >= SPARC_FP_ARG_MAX)
5804 /* If there are only int args and all int arg slots are filled,
5805 then must pass on stack. */
5806 if (!fpregs_p && intregs_p && slotno >= SPARC_INT_ARG_MAX)
5809 /* Note that even if all int arg slots are filled, fp members may
5810 still be passed in regs if such regs are available.
5811 *PREGNO isn't set because there may be more than one, it's up
5812 to the caller to compute them. */
5825 /* Handle recursive register counting for structure field layout. */
5827 struct function_arg_record_value_parms
5829 rtx ret; /* return expression being built. */
5830 int slotno; /* slot number of the argument. */
5831 int named; /* whether the argument is named. */
5832 int regbase; /* regno of the base register. */
5833 int stack; /* 1 if part of the argument is on the stack. */
5834 int intoffset; /* offset of the first pending integer field. */
5835 unsigned int nregs; /* number of words passed in registers. */
5838 static void function_arg_record_value_3
5839 (HOST_WIDE_INT, struct function_arg_record_value_parms *);
5840 static void function_arg_record_value_2
5841 (const_tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
5842 static void function_arg_record_value_1
5843 (const_tree, HOST_WIDE_INT, struct function_arg_record_value_parms *, bool);
5844 static rtx function_arg_record_value (const_tree, enum machine_mode, int, int, int);
5845 static rtx function_arg_union_value (int, enum machine_mode, int, int);
5847 /* A subroutine of function_arg_record_value. Traverse the structure
5848 recursively and determine how many registers will be required. */
5851 function_arg_record_value_1 (const_tree type, HOST_WIDE_INT startbitpos,
5852 struct function_arg_record_value_parms *parms,
5857 /* We need to compute how many registers are needed so we can
5858 allocate the PARALLEL but before we can do that we need to know
5859 whether there are any packed fields. The ABI obviously doesn't
5860 specify how structures are passed in this case, so they are
5861 defined to be passed in int regs if possible, otherwise memory,
5862 regardless of whether there are fp values present. */
5865 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5867 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
5874 /* Compute how many registers we need. */
5875 for (field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field))
5877 if (TREE_CODE (field) == FIELD_DECL)
5879 HOST_WIDE_INT bitpos = startbitpos;
5881 if (DECL_SIZE (field) != 0)
5883 if (integer_zerop (DECL_SIZE (field)))
5886 if (host_integerp (bit_position (field), 1))
5887 bitpos += int_bit_position (field);
5890 /* ??? FIXME: else assume zero offset. */
5892 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
5893 function_arg_record_value_1 (TREE_TYPE (field),
5897 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
5898 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
5903 if (parms->intoffset != -1)
5905 unsigned int startbit, endbit;
5906 int intslots, this_slotno;
5908 startbit = parms->intoffset & -BITS_PER_WORD;
5909 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
5911 intslots = (endbit - startbit) / BITS_PER_WORD;
5912 this_slotno = parms->slotno + parms->intoffset
5915 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
5917 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
5918 /* We need to pass this field on the stack. */
5922 parms->nregs += intslots;
5923 parms->intoffset = -1;
5926 /* There's no need to check this_slotno < SPARC_FP_ARG MAX.
5927 If it wasn't true we wouldn't be here. */
5928 if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE
5929 && DECL_MODE (field) == BLKmode)
5930 parms->nregs += TYPE_VECTOR_SUBPARTS (TREE_TYPE (field));
5931 else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
5938 if (parms->intoffset == -1)
5939 parms->intoffset = bitpos;
5945 /* A subroutine of function_arg_record_value. Assign the bits of the
5946 structure between parms->intoffset and bitpos to integer registers. */
5949 function_arg_record_value_3 (HOST_WIDE_INT bitpos,
5950 struct function_arg_record_value_parms *parms)
5952 enum machine_mode mode;
5954 unsigned int startbit, endbit;
5955 int this_slotno, intslots, intoffset;
5958 if (parms->intoffset == -1)
5961 intoffset = parms->intoffset;
5962 parms->intoffset = -1;
5964 startbit = intoffset & -BITS_PER_WORD;
5965 endbit = (bitpos + BITS_PER_WORD - 1) & -BITS_PER_WORD;
5966 intslots = (endbit - startbit) / BITS_PER_WORD;
5967 this_slotno = parms->slotno + intoffset / BITS_PER_WORD;
5969 intslots = MIN (intslots, SPARC_INT_ARG_MAX - this_slotno);
5973 /* If this is the trailing part of a word, only load that much into
5974 the register. Otherwise load the whole register. Note that in
5975 the latter case we may pick up unwanted bits. It's not a problem
5976 at the moment but may wish to revisit. */
5978 if (intoffset % BITS_PER_WORD != 0)
5979 mode = smallest_mode_for_size (BITS_PER_WORD - intoffset % BITS_PER_WORD,
5984 intoffset /= BITS_PER_UNIT;
5987 regno = parms->regbase + this_slotno;
5988 reg = gen_rtx_REG (mode, regno);
5989 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
5990 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (intoffset));
5993 intoffset = (intoffset | (UNITS_PER_WORD-1)) + 1;
5998 while (intslots > 0);
6001 /* A subroutine of function_arg_record_value. Traverse the structure
6002 recursively and assign bits to floating point registers. Track which
6003 bits in between need integer registers; invoke function_arg_record_value_3
6004 to make that happen. */
6007 function_arg_record_value_2 (const_tree type, HOST_WIDE_INT startbitpos,
6008 struct function_arg_record_value_parms *parms,
6014 for (field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field))
6016 if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field))
6023 for (field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field))
6025 if (TREE_CODE (field) == FIELD_DECL)
6027 HOST_WIDE_INT bitpos = startbitpos;
6029 if (DECL_SIZE (field) != 0)
6031 if (integer_zerop (DECL_SIZE (field)))
6034 if (host_integerp (bit_position (field), 1))
6035 bitpos += int_bit_position (field);
6038 /* ??? FIXME: else assume zero offset. */
6040 if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE)
6041 function_arg_record_value_2 (TREE_TYPE (field),
6045 else if ((FLOAT_TYPE_P (TREE_TYPE (field))
6046 || TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
6051 int this_slotno = parms->slotno + bitpos / BITS_PER_WORD;
6052 int regno, nregs, pos;
6053 enum machine_mode mode = DECL_MODE (field);
6056 function_arg_record_value_3 (bitpos, parms);
6058 if (TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE
6061 mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field)));
6062 nregs = TYPE_VECTOR_SUBPARTS (TREE_TYPE (field));
6064 else if (TREE_CODE (TREE_TYPE (field)) == COMPLEX_TYPE)
6066 mode = TYPE_MODE (TREE_TYPE (TREE_TYPE (field)));
6072 regno = SPARC_FP_ARG_FIRST + this_slotno * 2;
6073 if (GET_MODE_SIZE (mode) <= 4 && (bitpos & 32) != 0)
6075 reg = gen_rtx_REG (mode, regno);
6076 pos = bitpos / BITS_PER_UNIT;
6077 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
6078 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos));
6082 regno += GET_MODE_SIZE (mode) / 4;
6083 reg = gen_rtx_REG (mode, regno);
6084 pos += GET_MODE_SIZE (mode);
6085 XVECEXP (parms->ret, 0, parms->stack + parms->nregs)
6086 = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (pos));
6092 if (parms->intoffset == -1)
6093 parms->intoffset = bitpos;
6099 /* Used by function_arg and sparc_function_value_1 to implement the complex
6100 conventions of the 64-bit ABI for passing and returning structures.
6101 Return an expression valid as a return value for the FUNCTION_ARG
6102 and TARGET_FUNCTION_VALUE.
6104 TYPE is the data type of the argument (as a tree).
6105 This is null for libcalls where that information may
6107 MODE is the argument's machine mode.
6108 SLOTNO is the index number of the argument's slot in the parameter array.
6109 NAMED is nonzero if this argument is a named parameter
6110 (otherwise it is an extra parameter matching an ellipsis).
6111 REGBASE is the regno of the base register for the parameter array. */
6114 function_arg_record_value (const_tree type, enum machine_mode mode,
6115 int slotno, int named, int regbase)
6117 HOST_WIDE_INT typesize = int_size_in_bytes (type);
6118 struct function_arg_record_value_parms parms;
6121 parms.ret = NULL_RTX;
6122 parms.slotno = slotno;
6123 parms.named = named;
6124 parms.regbase = regbase;
6127 /* Compute how many registers we need. */
6129 parms.intoffset = 0;
6130 function_arg_record_value_1 (type, 0, &parms, false);
6132 /* Take into account pending integer fields. */
6133 if (parms.intoffset != -1)
6135 unsigned int startbit, endbit;
6136 int intslots, this_slotno;
6138 startbit = parms.intoffset & -BITS_PER_WORD;
6139 endbit = (typesize*BITS_PER_UNIT + BITS_PER_WORD - 1) & -BITS_PER_WORD;
6140 intslots = (endbit - startbit) / BITS_PER_WORD;
6141 this_slotno = slotno + parms.intoffset / BITS_PER_WORD;
6143 if (intslots > 0 && intslots > SPARC_INT_ARG_MAX - this_slotno)
6145 intslots = MAX (0, SPARC_INT_ARG_MAX - this_slotno);
6146 /* We need to pass this field on the stack. */
6150 parms.nregs += intslots;
6152 nregs = parms.nregs;
6154 /* Allocate the vector and handle some annoying special cases. */
6157 /* ??? Empty structure has no value? Duh? */
6160 /* Though there's nothing really to store, return a word register
6161 anyway so the rest of gcc doesn't go nuts. Returning a PARALLEL
6162 leads to breakage due to the fact that there are zero bytes to
6164 return gen_rtx_REG (mode, regbase);
6168 /* ??? C++ has structures with no fields, and yet a size. Give up
6169 for now and pass everything back in integer registers. */
6170 nregs = (typesize + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
6172 if (nregs + slotno > SPARC_INT_ARG_MAX)
6173 nregs = SPARC_INT_ARG_MAX - slotno;
6175 gcc_assert (nregs != 0);
6177 parms.ret = gen_rtx_PARALLEL (mode, rtvec_alloc (parms.stack + nregs));
6179 /* If at least one field must be passed on the stack, generate
6180 (parallel [(expr_list (nil) ...) ...]) so that all fields will
6181 also be passed on the stack. We can't do much better because the
6182 semantics of TARGET_ARG_PARTIAL_BYTES doesn't handle the case
6183 of structures for which the fields passed exclusively in registers
6184 are not at the beginning of the structure. */
6186 XVECEXP (parms.ret, 0, 0)
6187 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
6189 /* Fill in the entries. */
6191 parms.intoffset = 0;
6192 function_arg_record_value_2 (type, 0, &parms, false);
6193 function_arg_record_value_3 (typesize * BITS_PER_UNIT, &parms);
6195 gcc_assert (parms.nregs == nregs);
6200 /* Used by function_arg and sparc_function_value_1 to implement the conventions
6201 of the 64-bit ABI for passing and returning unions.
6202 Return an expression valid as a return value for the FUNCTION_ARG
6203 and TARGET_FUNCTION_VALUE.
6205 SIZE is the size in bytes of the union.
6206 MODE is the argument's machine mode.
6207 REGNO is the hard register the union will be passed in. */
6210 function_arg_union_value (int size, enum machine_mode mode, int slotno,
6213 int nwords = ROUND_ADVANCE (size), i;
6216 /* See comment in previous function for empty structures. */
6218 return gen_rtx_REG (mode, regno);
6220 if (slotno == SPARC_INT_ARG_MAX - 1)
6223 regs = gen_rtx_PARALLEL (mode, rtvec_alloc (nwords));
6225 for (i = 0; i < nwords; i++)
6227 /* Unions are passed left-justified. */
6228 XVECEXP (regs, 0, i)
6229 = gen_rtx_EXPR_LIST (VOIDmode,
6230 gen_rtx_REG (word_mode, regno),
6231 GEN_INT (UNITS_PER_WORD * i));
6238 /* Used by function_arg and sparc_function_value_1 to implement the conventions
6239 for passing and returning large (BLKmode) vectors.
6240 Return an expression valid as a return value for the FUNCTION_ARG
6241 and TARGET_FUNCTION_VALUE.
6243 SIZE is the size in bytes of the vector (at least 8 bytes).
6244 REGNO is the FP hard register the vector will be passed in. */
6247 function_arg_vector_value (int size, int regno)
6249 int i, nregs = size / 8;
6252 regs = gen_rtx_PARALLEL (BLKmode, rtvec_alloc (nregs));
6254 for (i = 0; i < nregs; i++)
6256 XVECEXP (regs, 0, i)
6257 = gen_rtx_EXPR_LIST (VOIDmode,
6258 gen_rtx_REG (DImode, regno + 2*i),
6265 /* Determine where to put an argument to a function.
6266 Value is zero to push the argument on the stack,
6267 or a hard register in which to store the argument.
6269 CUM is a variable of type CUMULATIVE_ARGS which gives info about
6270 the preceding args and about the function being called.
6271 MODE is the argument's machine mode.
6272 TYPE is the data type of the argument (as a tree).
6273 This is null for libcalls where that information may
6275 NAMED is true if this argument is a named parameter
6276 (otherwise it is an extra parameter matching an ellipsis).
6277 INCOMING_P is false for TARGET_FUNCTION_ARG, true for
6278 TARGET_FUNCTION_INCOMING_ARG. */
6281 sparc_function_arg_1 (cumulative_args_t cum_v, enum machine_mode mode,
6282 const_tree type, bool named, bool incoming_p)
6284 const CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v);
6286 int regbase = (incoming_p
6287 ? SPARC_INCOMING_INT_ARG_FIRST
6288 : SPARC_OUTGOING_INT_ARG_FIRST);
6289 int slotno, regno, padding;
6290 enum mode_class mclass = GET_MODE_CLASS (mode);
6292 slotno = function_arg_slotno (cum, mode, type, named, incoming_p,
6297 /* Vector types deserve special treatment because they are polymorphic wrt
6298 their mode, depending upon whether VIS instructions are enabled. */
6299 if (type && TREE_CODE (type) == VECTOR_TYPE)
6301 HOST_WIDE_INT size = int_size_in_bytes (type);
6302 gcc_assert ((TARGET_ARCH32 && size <= 8)
6303 || (TARGET_ARCH64 && size <= 16));
6305 if (mode == BLKmode)
6306 return function_arg_vector_value (size,
6307 SPARC_FP_ARG_FIRST + 2*slotno);
6309 mclass = MODE_FLOAT;
6313 return gen_rtx_REG (mode, regno);
6315 /* Structures up to 16 bytes in size are passed in arg slots on the stack
6316 and are promoted to registers if possible. */
6317 if (type && TREE_CODE (type) == RECORD_TYPE)
6319 HOST_WIDE_INT size = int_size_in_bytes (type);
6320 gcc_assert (size <= 16);
6322 return function_arg_record_value (type, mode, slotno, named, regbase);
6325 /* Unions up to 16 bytes in size are passed in integer registers. */
6326 else if (type && TREE_CODE (type) == UNION_TYPE)
6328 HOST_WIDE_INT size = int_size_in_bytes (type);
6329 gcc_assert (size <= 16);
6331 return function_arg_union_value (size, mode, slotno, regno);
6334 /* v9 fp args in reg slots beyond the int reg slots get passed in regs
6335 but also have the slot allocated for them.
6336 If no prototype is in scope fp values in register slots get passed
6337 in two places, either fp regs and int regs or fp regs and memory. */
6338 else if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT)
6339 && SPARC_FP_REG_P (regno))
6341 rtx reg = gen_rtx_REG (mode, regno);
6342 if (cum->prototype_p || cum->libcall_p)
6344 /* "* 2" because fp reg numbers are recorded in 4 byte
6347 /* ??? This will cause the value to be passed in the fp reg and
6348 in the stack. When a prototype exists we want to pass the
6349 value in the reg but reserve space on the stack. That's an
6350 optimization, and is deferred [for a bit]. */
6351 if ((regno - SPARC_FP_ARG_FIRST) >= SPARC_INT_ARG_MAX * 2)
6352 return gen_rtx_PARALLEL (mode,
6354 gen_rtx_EXPR_LIST (VOIDmode,
6355 NULL_RTX, const0_rtx),
6356 gen_rtx_EXPR_LIST (VOIDmode,
6360 /* ??? It seems that passing back a register even when past
6361 the area declared by REG_PARM_STACK_SPACE will allocate
6362 space appropriately, and will not copy the data onto the
6363 stack, exactly as we desire.
6365 This is due to locate_and_pad_parm being called in
6366 expand_call whenever reg_parm_stack_space > 0, which
6367 while beneficial to our example here, would seem to be
6368 in error from what had been intended. Ho hum... -- r~ */
6376 if ((regno - SPARC_FP_ARG_FIRST) < SPARC_INT_ARG_MAX * 2)
6380 /* On incoming, we don't need to know that the value
6381 is passed in %f0 and %i0, and it confuses other parts
6382 causing needless spillage even on the simplest cases. */
6386 intreg = (SPARC_OUTGOING_INT_ARG_FIRST
6387 + (regno - SPARC_FP_ARG_FIRST) / 2);
6389 v0 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
6390 v1 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_REG (mode, intreg),
6392 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
6396 v0 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
6397 v1 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx);
6398 return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1));
6403 /* All other aggregate types are passed in an integer register in a mode
6404 corresponding to the size of the type. */
6405 else if (type && AGGREGATE_TYPE_P (type))
6407 HOST_WIDE_INT size = int_size_in_bytes (type);
6408 gcc_assert (size <= 16);
6410 mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0);
6413 return gen_rtx_REG (mode, regno);
6416 /* Handle the TARGET_FUNCTION_ARG target hook. */
6419 sparc_function_arg (cumulative_args_t cum, enum machine_mode mode,
6420 const_tree type, bool named)
6422 return sparc_function_arg_1 (cum, mode, type, named, false);
6425 /* Handle the TARGET_FUNCTION_INCOMING_ARG target hook. */
6428 sparc_function_incoming_arg (cumulative_args_t cum, enum machine_mode mode,
6429 const_tree type, bool named)
6431 return sparc_function_arg_1 (cum, mode, type, named, true);
6434 /* For sparc64, objects requiring 16 byte alignment are passed that way. */
6437 sparc_function_arg_boundary (enum machine_mode mode, const_tree type)
6439 return ((TARGET_ARCH64
6440 && (GET_MODE_ALIGNMENT (mode) == 128
6441 || (type && TYPE_ALIGN (type) == 128)))
6446 /* For an arg passed partly in registers and partly in memory,
6447 this is the number of bytes of registers used.
6448 For args passed entirely in registers or entirely in memory, zero.
6450 Any arg that starts in the first 6 regs but won't entirely fit in them
6451 needs partial registers on v8. On v9, structures with integer
6452 values in arg slots 5,6 will be passed in %o5 and SP+176, and complex fp
6453 values that begin in the last fp reg [where "last fp reg" varies with the
6454 mode] will be split between that reg and memory. */
6457 sparc_arg_partial_bytes (cumulative_args_t cum, enum machine_mode mode,
6458 tree type, bool named)
6460 int slotno, regno, padding;
6462 /* We pass false for incoming_p here, it doesn't matter. */
6463 slotno = function_arg_slotno (get_cumulative_args (cum), mode, type, named,
6464 false, ®no, &padding);
6471 if ((slotno + (mode == BLKmode
6472 ? ROUND_ADVANCE (int_size_in_bytes (type))
6473 : ROUND_ADVANCE (GET_MODE_SIZE (mode))))
6474 > SPARC_INT_ARG_MAX)
6475 return (SPARC_INT_ARG_MAX - slotno) * UNITS_PER_WORD;
6479 /* We are guaranteed by pass_by_reference that the size of the
6480 argument is not greater than 16 bytes, so we only need to return
6481 one word if the argument is partially passed in registers. */
6483 if (type && AGGREGATE_TYPE_P (type))
6485 int size = int_size_in_bytes (type);
6487 if (size > UNITS_PER_WORD
6488 && slotno == SPARC_INT_ARG_MAX - 1)
6489 return UNITS_PER_WORD;
6491 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_INT
6492 || (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT
6493 && ! (TARGET_FPU && named)))
6495 /* The complex types are passed as packed types. */
6496 if (GET_MODE_SIZE (mode) > UNITS_PER_WORD
6497 && slotno == SPARC_INT_ARG_MAX - 1)
6498 return UNITS_PER_WORD;
6500 else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT)
6502 if ((slotno + GET_MODE_SIZE (mode) / UNITS_PER_WORD)
6504 return UNITS_PER_WORD;
6511 /* Handle the TARGET_PASS_BY_REFERENCE target hook.
6512 Specify whether to pass the argument by reference. */
6515 sparc_pass_by_reference (cumulative_args_t cum ATTRIBUTE_UNUSED,
6516 enum machine_mode mode, const_tree type,
6517 bool named ATTRIBUTE_UNUSED)
6520 /* Original SPARC 32-bit ABI says that structures and unions,
6521 and quad-precision floats are passed by reference. For Pascal,
6522 also pass arrays by reference. All other base types are passed
6525 Extended ABI (as implemented by the Sun compiler) says that all
6526 complex floats are passed by reference. Pass complex integers
6527 in registers up to 8 bytes. More generally, enforce the 2-word
6528 cap for passing arguments in registers.
6530 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6531 integers are passed like floats of the same size, that is in
6532 registers up to 8 bytes. Pass all vector floats by reference
6533 like structure and unions. */
6534 return ((type && (AGGREGATE_TYPE_P (type) || VECTOR_FLOAT_TYPE_P (type)))
6536 /* Catch CDImode, TFmode, DCmode and TCmode. */
6537 || GET_MODE_SIZE (mode) > 8
6539 && TREE_CODE (type) == VECTOR_TYPE
6540 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8));
6542 /* Original SPARC 64-bit ABI says that structures and unions
6543 smaller than 16 bytes are passed in registers, as well as
6544 all other base types.
6546 Extended ABI (as implemented by the Sun compiler) says that
6547 complex floats are passed in registers up to 16 bytes. Pass
6548 all complex integers in registers up to 16 bytes. More generally,
6549 enforce the 2-word cap for passing arguments in registers.
6551 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6552 integers are passed like floats of the same size, that is in
6553 registers (up to 16 bytes). Pass all vector floats like structure
6556 && (AGGREGATE_TYPE_P (type) || TREE_CODE (type) == VECTOR_TYPE)
6557 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 16)
6558 /* Catch CTImode and TCmode. */
6559 || GET_MODE_SIZE (mode) > 16);
6562 /* Handle the TARGET_FUNCTION_ARG_ADVANCE hook.
6563 Update the data in CUM to advance over an argument
6564 of mode MODE and data type TYPE.
6565 TYPE is null for libcalls where that information may not be available. */
6568 sparc_function_arg_advance (cumulative_args_t cum_v, enum machine_mode mode,
6569 const_tree type, bool named)
6571 CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v);
6574 /* We pass false for incoming_p here, it doesn't matter. */
6575 function_arg_slotno (cum, mode, type, named, false, ®no, &padding);
6577 /* If argument requires leading padding, add it. */
6578 cum->words += padding;
6582 cum->words += (mode != BLKmode
6583 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
6584 : ROUND_ADVANCE (int_size_in_bytes (type)));
6588 if (type && AGGREGATE_TYPE_P (type))
6590 int size = int_size_in_bytes (type);
6594 else if (size <= 16)
6596 else /* passed by reference */
6601 cum->words += (mode != BLKmode
6602 ? ROUND_ADVANCE (GET_MODE_SIZE (mode))
6603 : ROUND_ADVANCE (int_size_in_bytes (type)));
6608 /* Handle the FUNCTION_ARG_PADDING macro.
6609 For the 64 bit ABI structs are always stored left shifted in their
6613 function_arg_padding (enum machine_mode mode, const_tree type)
6615 if (TARGET_ARCH64 && type != 0 && AGGREGATE_TYPE_P (type))
6618 /* Fall back to the default. */
6619 return DEFAULT_FUNCTION_ARG_PADDING (mode, type);
6622 /* Handle the TARGET_RETURN_IN_MEMORY target hook.
6623 Specify whether to return the return value in memory. */
6626 sparc_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
6629 /* Original SPARC 32-bit ABI says that structures and unions,
6630 and quad-precision floats are returned in memory. All other
6631 base types are returned in registers.
6633 Extended ABI (as implemented by the Sun compiler) says that
6634 all complex floats are returned in registers (8 FP registers
6635 at most for '_Complex long double'). Return all complex integers
6636 in registers (4 at most for '_Complex long long').
6638 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6639 integers are returned like floats of the same size, that is in
6640 registers up to 8 bytes and in memory otherwise. Return all
6641 vector floats in memory like structure and unions; note that
6642 they always have BLKmode like the latter. */
6643 return (TYPE_MODE (type) == BLKmode
6644 || TYPE_MODE (type) == TFmode
6645 || (TREE_CODE (type) == VECTOR_TYPE
6646 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 8));
6648 /* Original SPARC 64-bit ABI says that structures and unions
6649 smaller than 32 bytes are returned in registers, as well as
6650 all other base types.
6652 Extended ABI (as implemented by the Sun compiler) says that all
6653 complex floats are returned in registers (8 FP registers at most
6654 for '_Complex long double'). Return all complex integers in
6655 registers (4 at most for '_Complex TItype').
6657 Vector ABI (as implemented by the Sun VIS SDK) says that vector
6658 integers are returned like floats of the same size, that is in
6659 registers. Return all vector floats like structure and unions;
6660 note that they always have BLKmode like the latter. */
6661 return (TYPE_MODE (type) == BLKmode
6662 && (unsigned HOST_WIDE_INT) int_size_in_bytes (type) > 32);
6665 /* Handle the TARGET_STRUCT_VALUE target hook.
6666 Return where to find the structure return value address. */
6669 sparc_struct_value_rtx (tree fndecl, int incoming)
6678 mem = gen_frame_mem (Pmode, plus_constant (frame_pointer_rtx,
6679 STRUCT_VALUE_OFFSET));
6681 mem = gen_frame_mem (Pmode, plus_constant (stack_pointer_rtx,
6682 STRUCT_VALUE_OFFSET));
6684 /* Only follow the SPARC ABI for fixed-size structure returns.
6685 Variable size structure returns are handled per the normal
6686 procedures in GCC. This is enabled by -mstd-struct-return */
6688 && sparc_std_struct_return
6689 && TYPE_SIZE_UNIT (TREE_TYPE (fndecl))
6690 && TREE_CODE (TYPE_SIZE_UNIT (TREE_TYPE (fndecl))) == INTEGER_CST)
6692 /* We must check and adjust the return address, as it is
6693 optional as to whether the return object is really
6695 rtx ret_reg = gen_rtx_REG (Pmode, 31);
6696 rtx scratch = gen_reg_rtx (SImode);
6697 rtx endlab = gen_label_rtx ();
6699 /* Calculate the return object size */
6700 tree size = TYPE_SIZE_UNIT (TREE_TYPE (fndecl));
6701 rtx size_rtx = GEN_INT (TREE_INT_CST_LOW (size) & 0xfff);
6702 /* Construct a temporary return value */
6704 = assign_stack_local (Pmode, TREE_INT_CST_LOW (size), 0);
6706 /* Implement SPARC 32-bit psABI callee return struct checking:
6708 Fetch the instruction where we will return to and see if
6709 it's an unimp instruction (the most significant 10 bits
6711 emit_move_insn (scratch, gen_rtx_MEM (SImode,
6712 plus_constant (ret_reg, 8)));
6713 /* Assume the size is valid and pre-adjust */
6714 emit_insn (gen_add3_insn (ret_reg, ret_reg, GEN_INT (4)));
6715 emit_cmp_and_jump_insns (scratch, size_rtx, EQ, const0_rtx, SImode,
6717 emit_insn (gen_sub3_insn (ret_reg, ret_reg, GEN_INT (4)));
6718 /* Write the address of the memory pointed to by temp_val into
6719 the memory pointed to by mem */
6720 emit_move_insn (mem, XEXP (temp_val, 0));
6721 emit_label (endlab);
6728 /* Handle TARGET_FUNCTION_VALUE, and TARGET_LIBCALL_VALUE target hook.
6729 For v9, function return values are subject to the same rules as arguments,
6730 except that up to 32 bytes may be returned in registers. */
6733 sparc_function_value_1 (const_tree type, enum machine_mode mode,
6736 /* Beware that the two values are swapped here wrt function_arg. */
6737 int regbase = (outgoing
6738 ? SPARC_INCOMING_INT_ARG_FIRST
6739 : SPARC_OUTGOING_INT_ARG_FIRST);
6740 enum mode_class mclass = GET_MODE_CLASS (mode);
6743 /* Vector types deserve special treatment because they are polymorphic wrt
6744 their mode, depending upon whether VIS instructions are enabled. */
6745 if (type && TREE_CODE (type) == VECTOR_TYPE)
6747 HOST_WIDE_INT size = int_size_in_bytes (type);
6748 gcc_assert ((TARGET_ARCH32 && size <= 8)
6749 || (TARGET_ARCH64 && size <= 32));
6751 if (mode == BLKmode)
6752 return function_arg_vector_value (size,
6753 SPARC_FP_ARG_FIRST);
6755 mclass = MODE_FLOAT;
6758 if (TARGET_ARCH64 && type)
6760 /* Structures up to 32 bytes in size are returned in registers. */
6761 if (TREE_CODE (type) == RECORD_TYPE)
6763 HOST_WIDE_INT size = int_size_in_bytes (type);
6764 gcc_assert (size <= 32);
6766 return function_arg_record_value (type, mode, 0, 1, regbase);
6769 /* Unions up to 32 bytes in size are returned in integer registers. */
6770 else if (TREE_CODE (type) == UNION_TYPE)
6772 HOST_WIDE_INT size = int_size_in_bytes (type);
6773 gcc_assert (size <= 32);
6775 return function_arg_union_value (size, mode, 0, regbase);
6778 /* Objects that require it are returned in FP registers. */
6779 else if (mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT)
6782 /* All other aggregate types are returned in an integer register in a
6783 mode corresponding to the size of the type. */
6784 else if (AGGREGATE_TYPE_P (type))
6786 /* All other aggregate types are passed in an integer register
6787 in a mode corresponding to the size of the type. */
6788 HOST_WIDE_INT size = int_size_in_bytes (type);
6789 gcc_assert (size <= 32);
6791 mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0);
6793 /* ??? We probably should have made the same ABI change in
6794 3.4.0 as the one we made for unions. The latter was
6795 required by the SCD though, while the former is not
6796 specified, so we favored compatibility and efficiency.
6798 Now we're stuck for aggregates larger than 16 bytes,
6799 because OImode vanished in the meantime. Let's not
6800 try to be unduly clever, and simply follow the ABI
6801 for unions in that case. */
6802 if (mode == BLKmode)
6803 return function_arg_union_value (size, mode, 0, regbase);
6808 /* We should only have pointer and integer types at this point. This
6809 must match sparc_promote_function_mode. */
6810 else if (mclass == MODE_INT && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
6814 /* We should only have pointer and integer types at this point. This must
6815 match sparc_promote_function_mode. */
6816 else if (TARGET_ARCH32
6817 && mclass == MODE_INT
6818 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
6821 if ((mclass == MODE_FLOAT || mclass == MODE_COMPLEX_FLOAT) && TARGET_FPU)
6822 regno = SPARC_FP_ARG_FIRST;
6826 return gen_rtx_REG (mode, regno);
6829 /* Handle TARGET_FUNCTION_VALUE.
6830 On the SPARC, the value is found in the first "output" register, but the
6831 called function leaves it in the first "input" register. */
6834 sparc_function_value (const_tree valtype,
6835 const_tree fn_decl_or_type ATTRIBUTE_UNUSED,
6838 return sparc_function_value_1 (valtype, TYPE_MODE (valtype), outgoing);
6841 /* Handle TARGET_LIBCALL_VALUE. */
6844 sparc_libcall_value (enum machine_mode mode,
6845 const_rtx fun ATTRIBUTE_UNUSED)
6847 return sparc_function_value_1 (NULL_TREE, mode, false);
6850 /* Handle FUNCTION_VALUE_REGNO_P.
6851 On the SPARC, the first "output" reg is used for integer values, and the
6852 first floating point register is used for floating point values. */
6855 sparc_function_value_regno_p (const unsigned int regno)
6857 return (regno == 8 || regno == 32);
6860 /* Do what is necessary for `va_start'. We look at the current function
6861 to determine if stdarg or varargs is used and return the address of
6862 the first unnamed parameter. */
6865 sparc_builtin_saveregs (void)
6867 int first_reg = crtl->args.info.words;
6871 for (regno = first_reg; regno < SPARC_INT_ARG_MAX; regno++)
6872 emit_move_insn (gen_rtx_MEM (word_mode,
6873 gen_rtx_PLUS (Pmode,
6875 GEN_INT (FIRST_PARM_OFFSET (0)
6878 gen_rtx_REG (word_mode,
6879 SPARC_INCOMING_INT_ARG_FIRST + regno));
6881 address = gen_rtx_PLUS (Pmode,
6883 GEN_INT (FIRST_PARM_OFFSET (0)
6884 + UNITS_PER_WORD * first_reg));
6889 /* Implement `va_start' for stdarg. */
6892 sparc_va_start (tree valist, rtx nextarg)
6894 nextarg = expand_builtin_saveregs ();
6895 std_expand_builtin_va_start (valist, nextarg);
6898 /* Implement `va_arg' for stdarg. */
6901 sparc_gimplify_va_arg (tree valist, tree type, gimple_seq *pre_p,
6904 HOST_WIDE_INT size, rsize, align;
6907 tree ptrtype = build_pointer_type (type);
6909 if (pass_by_reference (NULL, TYPE_MODE (type), type, false))
6912 size = rsize = UNITS_PER_WORD;
6918 size = int_size_in_bytes (type);
6919 rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD;
6924 /* For SPARC64, objects requiring 16-byte alignment get it. */
6925 if (TYPE_ALIGN (type) >= 2 * (unsigned) BITS_PER_WORD)
6926 align = 2 * UNITS_PER_WORD;
6928 /* SPARC-V9 ABI states that structures up to 16 bytes in size
6929 are left-justified in their slots. */
6930 if (AGGREGATE_TYPE_P (type))
6933 size = rsize = UNITS_PER_WORD;
6943 incr = fold_build_pointer_plus_hwi (incr, align - 1);
6944 incr = fold_convert (sizetype, incr);
6945 incr = fold_build2 (BIT_AND_EXPR, sizetype, incr,
6947 incr = fold_convert (ptr_type_node, incr);
6950 gimplify_expr (&incr, pre_p, post_p, is_gimple_val, fb_rvalue);
6953 if (BYTES_BIG_ENDIAN && size < rsize)
6954 addr = fold_build_pointer_plus_hwi (incr, rsize - size);
6958 addr = fold_convert (build_pointer_type (ptrtype), addr);
6959 addr = build_va_arg_indirect_ref (addr);
6962 /* If the address isn't aligned properly for the type, we need a temporary.
6963 FIXME: This is inefficient, usually we can do this in registers. */
6964 else if (align == 0 && TYPE_ALIGN (type) > BITS_PER_WORD)
6966 tree tmp = create_tmp_var (type, "va_arg_tmp");
6967 tree dest_addr = build_fold_addr_expr (tmp);
6968 tree copy = build_call_expr (builtin_decl_implicit (BUILT_IN_MEMCPY),
6969 3, dest_addr, addr, size_int (rsize));
6970 TREE_ADDRESSABLE (tmp) = 1;
6971 gimplify_and_add (copy, pre_p);
6976 addr = fold_convert (ptrtype, addr);
6978 incr = fold_build_pointer_plus_hwi (incr, rsize);
6979 gimplify_assign (valist, incr, post_p);
6981 return build_va_arg_indirect_ref (addr);
6984 /* Implement the TARGET_VECTOR_MODE_SUPPORTED_P target hook.
6985 Specify whether the vector mode is supported by the hardware. */
6988 sparc_vector_mode_supported_p (enum machine_mode mode)
6990 return TARGET_VIS && VECTOR_MODE_P (mode) ? true : false;
6993 /* Implement the TARGET_VECTORIZE_PREFERRED_SIMD_MODE target hook. */
6995 static enum machine_mode
6996 sparc_preferred_simd_mode (enum machine_mode mode)
7014 /* Return the string to output an unconditional branch to LABEL, which is
7015 the operand number of the label.
7017 DEST is the destination insn (i.e. the label), INSN is the source. */
7020 output_ubranch (rtx dest, int label, rtx insn)
7022 static char string[64];
7023 bool v9_form = false;
7026 if (TARGET_V9 && INSN_ADDRESSES_SET_P ())
7028 int delta = (INSN_ADDRESSES (INSN_UID (dest))
7029 - INSN_ADDRESSES (INSN_UID (insn)));
7030 /* Leave some instructions for "slop". */
7031 if (delta >= -260000 && delta < 260000)
7036 strcpy (string, "ba%*,pt\t%%xcc, ");
7038 strcpy (string, "b%*\t");
7040 p = strchr (string, '\0');
7051 /* Return the string to output a conditional branch to LABEL, which is
7052 the operand number of the label. OP is the conditional expression.
7053 XEXP (OP, 0) is assumed to be a condition code register (integer or
7054 floating point) and its mode specifies what kind of comparison we made.
7056 DEST is the destination insn (i.e. the label), INSN is the source.
7058 REVERSED is nonzero if we should reverse the sense of the comparison.
7060 ANNUL is nonzero if we should generate an annulling branch. */
7063 output_cbranch (rtx op, rtx dest, int label, int reversed, int annul,
7066 static char string[64];
7067 enum rtx_code code = GET_CODE (op);
7068 rtx cc_reg = XEXP (op, 0);
7069 enum machine_mode mode = GET_MODE (cc_reg);
7070 const char *labelno, *branch;
7071 int spaces = 8, far;
7074 /* v9 branches are limited to +-1MB. If it is too far away,
7087 fbne,a,pn %fcc2, .LC29
7095 far = TARGET_V9 && (get_attr_length (insn) >= 3);
7098 /* Reversal of FP compares takes care -- an ordered compare
7099 becomes an unordered compare and vice versa. */
7100 if (mode == CCFPmode || mode == CCFPEmode)
7101 code = reverse_condition_maybe_unordered (code);
7103 code = reverse_condition (code);
7106 /* Start by writing the branch condition. */
7107 if (mode == CCFPmode || mode == CCFPEmode)
7158 /* ??? !v9: FP branches cannot be preceded by another floating point
7159 insn. Because there is currently no concept of pre-delay slots,
7160 we can fix this only by always emitting a nop before a floating
7165 strcpy (string, "nop\n\t");
7166 strcat (string, branch);
7179 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
7191 if (mode == CC_NOOVmode || mode == CCX_NOOVmode)
7212 strcpy (string, branch);
7214 spaces -= strlen (branch);
7215 p = strchr (string, '\0');
7217 /* Now add the annulling, the label, and a possible noop. */
7230 if (! far && insn && INSN_ADDRESSES_SET_P ())
7232 int delta = (INSN_ADDRESSES (INSN_UID (dest))
7233 - INSN_ADDRESSES (INSN_UID (insn)));
7234 /* Leave some instructions for "slop". */
7235 if (delta < -260000 || delta >= 260000)
7239 if (mode == CCFPmode || mode == CCFPEmode)
7241 static char v9_fcc_labelno[] = "%%fccX, ";
7242 /* Set the char indicating the number of the fcc reg to use. */
7243 v9_fcc_labelno[5] = REGNO (cc_reg) - SPARC_FIRST_V9_FCC_REG + '0';
7244 labelno = v9_fcc_labelno;
7247 gcc_assert (REGNO (cc_reg) == SPARC_FCC_REG);
7251 else if (mode == CCXmode || mode == CCX_NOOVmode)
7253 labelno = "%%xcc, ";
7258 labelno = "%%icc, ";
7263 if (*labelno && insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
7266 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
7279 strcpy (p, labelno);
7280 p = strchr (p, '\0');
7283 strcpy (p, ".+12\n\t nop\n\tb\t");
7284 /* Skip the next insn if requested or
7285 if we know that it will be a nop. */
7286 if (annul || ! final_sequence)
7300 /* Emit a library call comparison between floating point X and Y.
7301 COMPARISON is the operator to compare with (EQ, NE, GT, etc).
7302 Return the new operator to be used in the comparison sequence.
7304 TARGET_ARCH64 uses _Qp_* functions, which use pointers to TFmode
7305 values as arguments instead of the TFmode registers themselves,
7306 that's why we cannot call emit_float_lib_cmp. */
7309 sparc_emit_float_lib_cmp (rtx x, rtx y, enum rtx_code comparison)
7312 rtx slot0, slot1, result, tem, tem2, libfunc;
7313 enum machine_mode mode;
7314 enum rtx_code new_comparison;
7319 qpfunc = (TARGET_ARCH64 ? "_Qp_feq" : "_Q_feq");
7323 qpfunc = (TARGET_ARCH64 ? "_Qp_fne" : "_Q_fne");
7327 qpfunc = (TARGET_ARCH64 ? "_Qp_fgt" : "_Q_fgt");
7331 qpfunc = (TARGET_ARCH64 ? "_Qp_fge" : "_Q_fge");
7335 qpfunc = (TARGET_ARCH64 ? "_Qp_flt" : "_Q_flt");
7339 qpfunc = (TARGET_ARCH64 ? "_Qp_fle" : "_Q_fle");
7350 qpfunc = (TARGET_ARCH64 ? "_Qp_cmp" : "_Q_cmp");
7363 slot0 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0);
7364 emit_move_insn (slot0, x);
7371 slot1 = assign_stack_temp (TFmode, GET_MODE_SIZE(TFmode), 0);
7372 emit_move_insn (slot1, y);
7375 libfunc = gen_rtx_SYMBOL_REF (Pmode, qpfunc);
7376 emit_library_call (libfunc, LCT_NORMAL,
7378 XEXP (slot0, 0), Pmode,
7379 XEXP (slot1, 0), Pmode);
7384 libfunc = gen_rtx_SYMBOL_REF (Pmode, qpfunc);
7385 emit_library_call (libfunc, LCT_NORMAL,
7387 x, TFmode, y, TFmode);
7392 /* Immediately move the result of the libcall into a pseudo
7393 register so reload doesn't clobber the value if it needs
7394 the return register for a spill reg. */
7395 result = gen_reg_rtx (mode);
7396 emit_move_insn (result, hard_libcall_value (mode, libfunc));
7401 return gen_rtx_NE (VOIDmode, result, const0_rtx);
7404 new_comparison = (comparison == UNORDERED ? EQ : NE);
7405 return gen_rtx_fmt_ee (new_comparison, VOIDmode, result, GEN_INT(3));
7408 new_comparison = (comparison == UNGT ? GT : NE);
7409 return gen_rtx_fmt_ee (new_comparison, VOIDmode, result, const1_rtx);
7411 return gen_rtx_NE (VOIDmode, result, const2_rtx);
7413 tem = gen_reg_rtx (mode);
7415 emit_insn (gen_andsi3 (tem, result, const1_rtx));
7417 emit_insn (gen_anddi3 (tem, result, const1_rtx));
7418 return gen_rtx_NE (VOIDmode, tem, const0_rtx);
7421 tem = gen_reg_rtx (mode);
7423 emit_insn (gen_addsi3 (tem, result, const1_rtx));
7425 emit_insn (gen_adddi3 (tem, result, const1_rtx));
7426 tem2 = gen_reg_rtx (mode);
7428 emit_insn (gen_andsi3 (tem2, tem, const2_rtx));
7430 emit_insn (gen_anddi3 (tem2, tem, const2_rtx));
7431 new_comparison = (comparison == UNEQ ? EQ : NE);
7432 return gen_rtx_fmt_ee (new_comparison, VOIDmode, tem2, const0_rtx);
7438 /* Generate an unsigned DImode to FP conversion. This is the same code
7439 optabs would emit if we didn't have TFmode patterns. */
7442 sparc_emit_floatunsdi (rtx *operands, enum machine_mode mode)
7444 rtx neglab, donelab, i0, i1, f0, in, out;
7447 in = force_reg (DImode, operands[1]);
7448 neglab = gen_label_rtx ();
7449 donelab = gen_label_rtx ();
7450 i0 = gen_reg_rtx (DImode);
7451 i1 = gen_reg_rtx (DImode);
7452 f0 = gen_reg_rtx (mode);
7454 emit_cmp_and_jump_insns (in, const0_rtx, LT, const0_rtx, DImode, 0, neglab);
7456 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_FLOAT (mode, in)));
7457 emit_jump_insn (gen_jump (donelab));
7460 emit_label (neglab);
7462 emit_insn (gen_lshrdi3 (i0, in, const1_rtx));
7463 emit_insn (gen_anddi3 (i1, in, const1_rtx));
7464 emit_insn (gen_iordi3 (i0, i0, i1));
7465 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_FLOAT (mode, i0)));
7466 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_PLUS (mode, f0, f0)));
7468 emit_label (donelab);
7471 /* Generate an FP to unsigned DImode conversion. This is the same code
7472 optabs would emit if we didn't have TFmode patterns. */
7475 sparc_emit_fixunsdi (rtx *operands, enum machine_mode mode)
7477 rtx neglab, donelab, i0, i1, f0, in, out, limit;
7480 in = force_reg (mode, operands[1]);
7481 neglab = gen_label_rtx ();
7482 donelab = gen_label_rtx ();
7483 i0 = gen_reg_rtx (DImode);
7484 i1 = gen_reg_rtx (DImode);
7485 limit = gen_reg_rtx (mode);
7486 f0 = gen_reg_rtx (mode);
7488 emit_move_insn (limit,
7489 CONST_DOUBLE_FROM_REAL_VALUE (
7490 REAL_VALUE_ATOF ("9223372036854775808.0", mode), mode));
7491 emit_cmp_and_jump_insns (in, limit, GE, NULL_RTX, mode, 0, neglab);
7493 emit_insn (gen_rtx_SET (VOIDmode,
7495 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, in))));
7496 emit_jump_insn (gen_jump (donelab));
7499 emit_label (neglab);
7501 emit_insn (gen_rtx_SET (VOIDmode, f0, gen_rtx_MINUS (mode, in, limit)));
7502 emit_insn (gen_rtx_SET (VOIDmode,
7504 gen_rtx_FIX (DImode, gen_rtx_FIX (mode, f0))));
7505 emit_insn (gen_movdi (i1, const1_rtx));
7506 emit_insn (gen_ashldi3 (i1, i1, GEN_INT (63)));
7507 emit_insn (gen_xordi3 (out, i0, i1));
7509 emit_label (donelab);
7512 /* Return the string to output a conditional branch to LABEL, testing
7513 register REG. LABEL is the operand number of the label; REG is the
7514 operand number of the reg. OP is the conditional expression. The mode
7515 of REG says what kind of comparison we made.
7517 DEST is the destination insn (i.e. the label), INSN is the source.
7519 REVERSED is nonzero if we should reverse the sense of the comparison.
7521 ANNUL is nonzero if we should generate an annulling branch. */
7524 output_v9branch (rtx op, rtx dest, int reg, int label, int reversed,
7525 int annul, rtx insn)
7527 static char string[64];
7528 enum rtx_code code = GET_CODE (op);
7529 enum machine_mode mode = GET_MODE (XEXP (op, 0));
7534 /* branch on register are limited to +-128KB. If it is too far away,
7547 brgez,a,pn %o1, .LC29
7553 ba,pt %xcc, .LC29 */
7555 far = get_attr_length (insn) >= 3;
7557 /* If not floating-point or if EQ or NE, we can just reverse the code. */
7559 code = reverse_condition (code);
7561 /* Only 64 bit versions of these instructions exist. */
7562 gcc_assert (mode == DImode);
7564 /* Start by writing the branch condition. */
7569 strcpy (string, "brnz");
7573 strcpy (string, "brz");
7577 strcpy (string, "brgez");
7581 strcpy (string, "brlz");
7585 strcpy (string, "brlez");
7589 strcpy (string, "brgz");
7596 p = strchr (string, '\0');
7598 /* Now add the annulling, reg, label, and nop. */
7605 if (insn && (note = find_reg_note (insn, REG_BR_PROB, NULL_RTX)))
7608 ((INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2) ^ far)
7613 *p = p < string + 8 ? '\t' : ' ';
7621 int veryfar = 1, delta;
7623 if (INSN_ADDRESSES_SET_P ())
7625 delta = (INSN_ADDRESSES (INSN_UID (dest))
7626 - INSN_ADDRESSES (INSN_UID (insn)));
7627 /* Leave some instructions for "slop". */
7628 if (delta >= -260000 && delta < 260000)
7632 strcpy (p, ".+12\n\t nop\n\t");
7633 /* Skip the next insn if requested or
7634 if we know that it will be a nop. */
7635 if (annul || ! final_sequence)
7645 strcpy (p, "ba,pt\t%%xcc, ");
7659 /* Return 1, if any of the registers of the instruction are %l[0-7] or %o[0-7].
7660 Such instructions cannot be used in the delay slot of return insn on v9.
7661 If TEST is 0, also rename all %i[0-7] registers to their %o[0-7] counterparts.
7665 epilogue_renumber (register rtx *where, int test)
7667 register const char *fmt;
7669 register enum rtx_code code;
7674 code = GET_CODE (*where);
7679 if (REGNO (*where) >= 8 && REGNO (*where) < 24) /* oX or lX */
7681 if (! test && REGNO (*where) >= 24 && REGNO (*where) < 32)
7682 *where = gen_rtx_REG (GET_MODE (*where), OUTGOING_REGNO (REGNO(*where)));
7690 /* Do not replace the frame pointer with the stack pointer because
7691 it can cause the delayed instruction to load below the stack.
7692 This occurs when instructions like:
7694 (set (reg/i:SI 24 %i0)
7695 (mem/f:SI (plus:SI (reg/f:SI 30 %fp)
7696 (const_int -20 [0xffffffec])) 0))
7698 are in the return delayed slot. */
7700 if (GET_CODE (XEXP (*where, 0)) == REG
7701 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM
7702 && (GET_CODE (XEXP (*where, 1)) != CONST_INT
7703 || INTVAL (XEXP (*where, 1)) < SPARC_STACK_BIAS))
7708 if (SPARC_STACK_BIAS
7709 && GET_CODE (XEXP (*where, 0)) == REG
7710 && REGNO (XEXP (*where, 0)) == HARD_FRAME_POINTER_REGNUM)
7718 fmt = GET_RTX_FORMAT (code);
7720 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7725 for (j = XVECLEN (*where, i) - 1; j >= 0; j--)
7726 if (epilogue_renumber (&(XVECEXP (*where, i, j)), test))
7729 else if (fmt[i] == 'e'
7730 && epilogue_renumber (&(XEXP (*where, i)), test))
7736 /* Leaf functions and non-leaf functions have different needs. */
7739 reg_leaf_alloc_order[] = REG_LEAF_ALLOC_ORDER;
7742 reg_nonleaf_alloc_order[] = REG_ALLOC_ORDER;
7744 static const int *const reg_alloc_orders[] = {
7745 reg_leaf_alloc_order,
7746 reg_nonleaf_alloc_order};
7749 order_regs_for_local_alloc (void)
7751 static int last_order_nonleaf = 1;
7753 if (df_regs_ever_live_p (15) != last_order_nonleaf)
7755 last_order_nonleaf = !last_order_nonleaf;
7756 memcpy ((char *) reg_alloc_order,
7757 (const char *) reg_alloc_orders[last_order_nonleaf],
7758 FIRST_PSEUDO_REGISTER * sizeof (int));
7762 /* Return 1 if REG and MEM are legitimate enough to allow the various
7763 mem<-->reg splits to be run. */
7766 sparc_splitdi_legitimate (rtx reg, rtx mem)
7768 /* Punt if we are here by mistake. */
7769 gcc_assert (reload_completed);
7771 /* We must have an offsettable memory reference. */
7772 if (! offsettable_memref_p (mem))
7775 /* If we have legitimate args for ldd/std, we do not want
7776 the split to happen. */
7777 if ((REGNO (reg) % 2) == 0
7778 && mem_min_alignment (mem, 8))
7785 /* Like sparc_splitdi_legitimate but for REG <--> REG moves. */
7788 sparc_split_regreg_legitimate (rtx reg1, rtx reg2)
7792 if (GET_CODE (reg1) == SUBREG)
7793 reg1 = SUBREG_REG (reg1);
7794 if (GET_CODE (reg1) != REG)
7796 regno1 = REGNO (reg1);
7798 if (GET_CODE (reg2) == SUBREG)
7799 reg2 = SUBREG_REG (reg2);
7800 if (GET_CODE (reg2) != REG)
7802 regno2 = REGNO (reg2);
7804 if (SPARC_INT_REG_P (regno1) && SPARC_INT_REG_P (regno2))
7809 if ((SPARC_INT_REG_P (regno1) && SPARC_FP_REG_P (regno2))
7810 || (SPARC_FP_REG_P (regno1) && SPARC_INT_REG_P (regno2)))
7817 /* Return 1 if x and y are some kind of REG and they refer to
7818 different hard registers. This test is guaranteed to be
7819 run after reload. */
7822 sparc_absnegfloat_split_legitimate (rtx x, rtx y)
7824 if (GET_CODE (x) != REG)
7826 if (GET_CODE (y) != REG)
7828 if (REGNO (x) == REGNO (y))
7833 /* Return 1 if REGNO (reg1) is even and REGNO (reg1) == REGNO (reg2) - 1.
7834 This makes them candidates for using ldd and std insns.
7836 Note reg1 and reg2 *must* be hard registers. */
7839 registers_ok_for_ldd_peep (rtx reg1, rtx reg2)
7841 /* We might have been passed a SUBREG. */
7842 if (GET_CODE (reg1) != REG || GET_CODE (reg2) != REG)
7845 if (REGNO (reg1) % 2 != 0)
7848 /* Integer ldd is deprecated in SPARC V9 */
7849 if (TARGET_V9 && SPARC_INT_REG_P (REGNO (reg1)))
7852 return (REGNO (reg1) == REGNO (reg2) - 1);
7855 /* Return 1 if the addresses in mem1 and mem2 are suitable for use in
7858 This can only happen when addr1 and addr2, the addresses in mem1
7859 and mem2, are consecutive memory locations (addr1 + 4 == addr2).
7860 addr1 must also be aligned on a 64-bit boundary.
7862 Also iff dependent_reg_rtx is not null it should not be used to
7863 compute the address for mem1, i.e. we cannot optimize a sequence
7875 But, note that the transformation from:
7880 is perfectly fine. Thus, the peephole2 patterns always pass us
7881 the destination register of the first load, never the second one.
7883 For stores we don't have a similar problem, so dependent_reg_rtx is
7887 mems_ok_for_ldd_peep (rtx mem1, rtx mem2, rtx dependent_reg_rtx)
7891 HOST_WIDE_INT offset1;
7893 /* The mems cannot be volatile. */
7894 if (MEM_VOLATILE_P (mem1) || MEM_VOLATILE_P (mem2))
7897 /* MEM1 should be aligned on a 64-bit boundary. */
7898 if (MEM_ALIGN (mem1) < 64)
7901 addr1 = XEXP (mem1, 0);
7902 addr2 = XEXP (mem2, 0);
7904 /* Extract a register number and offset (if used) from the first addr. */
7905 if (GET_CODE (addr1) == PLUS)
7907 /* If not a REG, return zero. */
7908 if (GET_CODE (XEXP (addr1, 0)) != REG)
7912 reg1 = REGNO (XEXP (addr1, 0));
7913 /* The offset must be constant! */
7914 if (GET_CODE (XEXP (addr1, 1)) != CONST_INT)
7916 offset1 = INTVAL (XEXP (addr1, 1));
7919 else if (GET_CODE (addr1) != REG)
7923 reg1 = REGNO (addr1);
7924 /* This was a simple (mem (reg)) expression. Offset is 0. */
7928 /* Make sure the second address is a (mem (plus (reg) (const_int). */
7929 if (GET_CODE (addr2) != PLUS)
7932 if (GET_CODE (XEXP (addr2, 0)) != REG
7933 || GET_CODE (XEXP (addr2, 1)) != CONST_INT)
7936 if (reg1 != REGNO (XEXP (addr2, 0)))
7939 if (dependent_reg_rtx != NULL_RTX && reg1 == REGNO (dependent_reg_rtx))
7942 /* The first offset must be evenly divisible by 8 to ensure the
7943 address is 64 bit aligned. */
7944 if (offset1 % 8 != 0)
7947 /* The offset for the second addr must be 4 more than the first addr. */
7948 if (INTVAL (XEXP (addr2, 1)) != offset1 + 4)
7951 /* All the tests passed. addr1 and addr2 are valid for ldd and std
7956 /* Return 1 if reg is a pseudo, or is the first register in
7957 a hard register pair. This makes it suitable for use in
7958 ldd and std insns. */
7961 register_ok_for_ldd (rtx reg)
7963 /* We might have been passed a SUBREG. */
7967 if (REGNO (reg) < FIRST_PSEUDO_REGISTER)
7968 return (REGNO (reg) % 2 == 0);
7973 /* Return 1 if OP is a memory whose address is known to be
7974 aligned to 8-byte boundary, or a pseudo during reload.
7975 This makes it suitable for use in ldd and std insns. */
7978 memory_ok_for_ldd (rtx op)
7982 /* In 64-bit mode, we assume that the address is word-aligned. */
7983 if (TARGET_ARCH32 && !mem_min_alignment (op, 8))
7986 if (! can_create_pseudo_p ()
7987 && !strict_memory_address_p (Pmode, XEXP (op, 0)))
7990 else if (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER)
7992 if (!(reload_in_progress && reg_renumber [REGNO (op)] < 0))
8001 /* Implement TARGET_PRINT_OPERAND_PUNCT_VALID_P. */
8004 sparc_print_operand_punct_valid_p (unsigned char code)
8017 /* Implement TARGET_PRINT_OPERAND.
8018 Print operand X (an rtx) in assembler syntax to file FILE.
8019 CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified.
8020 For `%' followed by punctuation, CODE is the punctuation and X is null. */
8023 sparc_print_operand (FILE *file, rtx x, int code)
8028 /* Output an insn in a delay slot. */
8030 sparc_indent_opcode = 1;
8032 fputs ("\n\t nop", file);
8035 /* Output an annul flag if there's nothing for the delay slot and we
8036 are optimizing. This is always used with '(' below.
8037 Sun OS 4.1.1 dbx can't handle an annulled unconditional branch;
8038 this is a dbx bug. So, we only do this when optimizing.
8039 On UltraSPARC, a branch in a delay slot causes a pipeline flush.
8040 Always emit a nop in case the next instruction is a branch. */
8041 if (! final_sequence && (optimize && (int)sparc_cpu < PROCESSOR_V9))
8045 /* Output a 'nop' if there's nothing for the delay slot and we are
8046 not optimizing. This is always used with '*' above. */
8047 if (! final_sequence && ! (optimize && (int)sparc_cpu < PROCESSOR_V9))
8048 fputs ("\n\t nop", file);
8049 else if (final_sequence)
8050 sparc_indent_opcode = 1;
8053 /* Output the right displacement from the saved PC on function return.
8054 The caller may have placed an "unimp" insn immediately after the call
8055 so we have to account for it. This insn is used in the 32-bit ABI
8056 when calling a function that returns a non zero-sized structure. The
8057 64-bit ABI doesn't have it. Be careful to have this test be the same
8058 as that for the call. The exception is when sparc_std_struct_return
8059 is enabled, the psABI is followed exactly and the adjustment is made
8060 by the code in sparc_struct_value_rtx. The call emitted is the same
8061 when sparc_std_struct_return is enabled. */
8063 && cfun->returns_struct
8064 && !sparc_std_struct_return
8065 && DECL_SIZE (DECL_RESULT (current_function_decl))
8066 && TREE_CODE (DECL_SIZE (DECL_RESULT (current_function_decl)))
8068 && !integer_zerop (DECL_SIZE (DECL_RESULT (current_function_decl))))
8074 /* Output the Embedded Medium/Anywhere code model base register. */
8075 fputs (EMBMEDANY_BASE_REG, file);
8078 /* Print some local dynamic TLS name. */
8079 assemble_name (file, get_some_local_dynamic_name ());
8083 /* Adjust the operand to take into account a RESTORE operation. */
8084 if (GET_CODE (x) == CONST_INT)
8086 else if (GET_CODE (x) != REG)
8087 output_operand_lossage ("invalid %%Y operand");
8088 else if (REGNO (x) < 8)
8089 fputs (reg_names[REGNO (x)], file);
8090 else if (REGNO (x) >= 24 && REGNO (x) < 32)
8091 fputs (reg_names[REGNO (x)-16], file);
8093 output_operand_lossage ("invalid %%Y operand");
8096 /* Print out the low order register name of a register pair. */
8097 if (WORDS_BIG_ENDIAN)
8098 fputs (reg_names[REGNO (x)+1], file);
8100 fputs (reg_names[REGNO (x)], file);
8103 /* Print out the high order register name of a register pair. */
8104 if (WORDS_BIG_ENDIAN)
8105 fputs (reg_names[REGNO (x)], file);
8107 fputs (reg_names[REGNO (x)+1], file);
8110 /* Print out the second register name of a register pair or quad.
8111 I.e., R (%o0) => %o1. */
8112 fputs (reg_names[REGNO (x)+1], file);
8115 /* Print out the third register name of a register quad.
8116 I.e., S (%o0) => %o2. */
8117 fputs (reg_names[REGNO (x)+2], file);
8120 /* Print out the fourth register name of a register quad.
8121 I.e., T (%o0) => %o3. */
8122 fputs (reg_names[REGNO (x)+3], file);
8125 /* Print a condition code register. */
8126 if (REGNO (x) == SPARC_ICC_REG)
8128 /* We don't handle CC[X]_NOOVmode because they're not supposed
8130 if (GET_MODE (x) == CCmode)
8131 fputs ("%icc", file);
8132 else if (GET_MODE (x) == CCXmode)
8133 fputs ("%xcc", file);
8138 /* %fccN register */
8139 fputs (reg_names[REGNO (x)], file);
8142 /* Print the operand's address only. */
8143 output_address (XEXP (x, 0));
8146 /* In this case we need a register. Use %g0 if the
8147 operand is const0_rtx. */
8149 || (GET_MODE (x) != VOIDmode && x == CONST0_RTX (GET_MODE (x))))
8151 fputs ("%g0", file);
8158 switch (GET_CODE (x))
8160 case IOR: fputs ("or", file); break;
8161 case AND: fputs ("and", file); break;
8162 case XOR: fputs ("xor", file); break;
8163 default: output_operand_lossage ("invalid %%A operand");
8168 switch (GET_CODE (x))
8170 case IOR: fputs ("orn", file); break;
8171 case AND: fputs ("andn", file); break;
8172 case XOR: fputs ("xnor", file); break;
8173 default: output_operand_lossage ("invalid %%B operand");
8177 /* This is used by the conditional move instructions. */
8180 enum rtx_code rc = GET_CODE (x);
8184 case NE: fputs ("ne", file); break;
8185 case EQ: fputs ("e", file); break;
8186 case GE: fputs ("ge", file); break;
8187 case GT: fputs ("g", file); break;
8188 case LE: fputs ("le", file); break;
8189 case LT: fputs ("l", file); break;
8190 case GEU: fputs ("geu", file); break;
8191 case GTU: fputs ("gu", file); break;
8192 case LEU: fputs ("leu", file); break;
8193 case LTU: fputs ("lu", file); break;
8194 case LTGT: fputs ("lg", file); break;
8195 case UNORDERED: fputs ("u", file); break;
8196 case ORDERED: fputs ("o", file); break;
8197 case UNLT: fputs ("ul", file); break;
8198 case UNLE: fputs ("ule", file); break;
8199 case UNGT: fputs ("ug", file); break;
8200 case UNGE: fputs ("uge", file); break;
8201 case UNEQ: fputs ("ue", file); break;
8202 default: output_operand_lossage ("invalid %%C operand");
8207 /* This are used by the movr instruction pattern. */
8210 enum rtx_code rc = GET_CODE (x);
8213 case NE: fputs ("ne", file); break;
8214 case EQ: fputs ("e", file); break;
8215 case GE: fputs ("gez", file); break;
8216 case LT: fputs ("lz", file); break;
8217 case LE: fputs ("lez", file); break;
8218 case GT: fputs ("gz", file); break;
8219 default: output_operand_lossage ("invalid %%D operand");
8226 /* Print a sign-extended character. */
8227 int i = trunc_int_for_mode (INTVAL (x), QImode);
8228 fprintf (file, "%d", i);
8233 /* Operand must be a MEM; write its address. */
8234 if (GET_CODE (x) != MEM)
8235 output_operand_lossage ("invalid %%f operand");
8236 output_address (XEXP (x, 0));
8241 /* Print a sign-extended 32-bit value. */
8243 if (GET_CODE(x) == CONST_INT)
8245 else if (GET_CODE(x) == CONST_DOUBLE)
8246 i = CONST_DOUBLE_LOW (x);
8249 output_operand_lossage ("invalid %%s operand");
8252 i = trunc_int_for_mode (i, SImode);
8253 fprintf (file, HOST_WIDE_INT_PRINT_DEC, i);
8258 /* Do nothing special. */
8262 /* Undocumented flag. */
8263 output_operand_lossage ("invalid operand output code");
8266 if (GET_CODE (x) == REG)
8267 fputs (reg_names[REGNO (x)], file);
8268 else if (GET_CODE (x) == MEM)
8271 /* Poor Sun assembler doesn't understand absolute addressing. */
8272 if (CONSTANT_P (XEXP (x, 0)))
8273 fputs ("%g0+", file);
8274 output_address (XEXP (x, 0));
8277 else if (GET_CODE (x) == HIGH)
8279 fputs ("%hi(", file);
8280 output_addr_const (file, XEXP (x, 0));
8283 else if (GET_CODE (x) == LO_SUM)
8285 sparc_print_operand (file, XEXP (x, 0), 0);
8286 if (TARGET_CM_MEDMID)
8287 fputs ("+%l44(", file);
8289 fputs ("+%lo(", file);
8290 output_addr_const (file, XEXP (x, 1));
8293 else if (GET_CODE (x) == CONST_DOUBLE
8294 && (GET_MODE (x) == VOIDmode
8295 || GET_MODE_CLASS (GET_MODE (x)) == MODE_INT))
8297 if (CONST_DOUBLE_HIGH (x) == 0)
8298 fprintf (file, "%u", (unsigned int) CONST_DOUBLE_LOW (x));
8299 else if (CONST_DOUBLE_HIGH (x) == -1
8300 && CONST_DOUBLE_LOW (x) < 0)
8301 fprintf (file, "%d", (int) CONST_DOUBLE_LOW (x));
8303 output_operand_lossage ("long long constant not a valid immediate operand");
8305 else if (GET_CODE (x) == CONST_DOUBLE)
8306 output_operand_lossage ("floating point constant not a valid immediate operand");
8307 else { output_addr_const (file, x); }
8310 /* Implement TARGET_PRINT_OPERAND_ADDRESS. */
8313 sparc_print_operand_address (FILE *file, rtx x)
8315 register rtx base, index = 0;
8317 register rtx addr = x;
8320 fputs (reg_names[REGNO (addr)], file);
8321 else if (GET_CODE (addr) == PLUS)
8323 if (CONST_INT_P (XEXP (addr, 0)))
8324 offset = INTVAL (XEXP (addr, 0)), base = XEXP (addr, 1);
8325 else if (CONST_INT_P (XEXP (addr, 1)))
8326 offset = INTVAL (XEXP (addr, 1)), base = XEXP (addr, 0);
8328 base = XEXP (addr, 0), index = XEXP (addr, 1);
8329 if (GET_CODE (base) == LO_SUM)
8331 gcc_assert (USE_AS_OFFSETABLE_LO10
8333 && ! TARGET_CM_MEDMID);
8334 output_operand (XEXP (base, 0), 0);
8335 fputs ("+%lo(", file);
8336 output_address (XEXP (base, 1));
8337 fprintf (file, ")+%d", offset);
8341 fputs (reg_names[REGNO (base)], file);
8343 fprintf (file, "%+d", offset);
8344 else if (REG_P (index))
8345 fprintf (file, "+%s", reg_names[REGNO (index)]);
8346 else if (GET_CODE (index) == SYMBOL_REF
8347 || GET_CODE (index) == LABEL_REF
8348 || GET_CODE (index) == CONST)
8349 fputc ('+', file), output_addr_const (file, index);
8350 else gcc_unreachable ();
8353 else if (GET_CODE (addr) == MINUS
8354 && GET_CODE (XEXP (addr, 1)) == LABEL_REF)
8356 output_addr_const (file, XEXP (addr, 0));
8358 output_addr_const (file, XEXP (addr, 1));
8359 fputs ("-.)", file);
8361 else if (GET_CODE (addr) == LO_SUM)
8363 output_operand (XEXP (addr, 0), 0);
8364 if (TARGET_CM_MEDMID)
8365 fputs ("+%l44(", file);
8367 fputs ("+%lo(", file);
8368 output_address (XEXP (addr, 1));
8372 && GET_CODE (addr) == CONST
8373 && GET_CODE (XEXP (addr, 0)) == MINUS
8374 && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST
8375 && GET_CODE (XEXP (XEXP (XEXP (addr, 0), 1), 0)) == MINUS
8376 && XEXP (XEXP (XEXP (XEXP (addr, 0), 1), 0), 1) == pc_rtx)
8378 addr = XEXP (addr, 0);
8379 output_addr_const (file, XEXP (addr, 0));
8380 /* Group the args of the second CONST in parenthesis. */
8382 /* Skip past the second CONST--it does nothing for us. */
8383 output_addr_const (file, XEXP (XEXP (addr, 1), 0));
8384 /* Close the parenthesis. */
8389 output_addr_const (file, addr);
8393 /* Target hook for assembling integer objects. The sparc version has
8394 special handling for aligned DI-mode objects. */
8397 sparc_assemble_integer (rtx x, unsigned int size, int aligned_p)
8399 /* ??? We only output .xword's for symbols and only then in environments
8400 where the assembler can handle them. */
8401 if (aligned_p && size == 8
8402 && (GET_CODE (x) != CONST_INT && GET_CODE (x) != CONST_DOUBLE))
8406 assemble_integer_with_op ("\t.xword\t", x);
8411 assemble_aligned_integer (4, const0_rtx);
8412 assemble_aligned_integer (4, x);
8416 return default_assemble_integer (x, size, aligned_p);
8419 /* Return the value of a code used in the .proc pseudo-op that says
8420 what kind of result this function returns. For non-C types, we pick
8421 the closest C type. */
8423 #ifndef SHORT_TYPE_SIZE
8424 #define SHORT_TYPE_SIZE (BITS_PER_UNIT * 2)
8427 #ifndef INT_TYPE_SIZE
8428 #define INT_TYPE_SIZE BITS_PER_WORD
8431 #ifndef LONG_TYPE_SIZE
8432 #define LONG_TYPE_SIZE BITS_PER_WORD
8435 #ifndef LONG_LONG_TYPE_SIZE
8436 #define LONG_LONG_TYPE_SIZE (BITS_PER_WORD * 2)
8439 #ifndef FLOAT_TYPE_SIZE
8440 #define FLOAT_TYPE_SIZE BITS_PER_WORD
8443 #ifndef DOUBLE_TYPE_SIZE
8444 #define DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
8447 #ifndef LONG_DOUBLE_TYPE_SIZE
8448 #define LONG_DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2)
8452 sparc_type_code (register tree type)
8454 register unsigned long qualifiers = 0;
8455 register unsigned shift;
8457 /* Only the first 30 bits of the qualifier are valid. We must refrain from
8458 setting more, since some assemblers will give an error for this. Also,
8459 we must be careful to avoid shifts of 32 bits or more to avoid getting
8460 unpredictable results. */
8462 for (shift = 6; shift < 30; shift += 2, type = TREE_TYPE (type))
8464 switch (TREE_CODE (type))
8470 qualifiers |= (3 << shift);
8475 qualifiers |= (2 << shift);
8479 case REFERENCE_TYPE:
8481 qualifiers |= (1 << shift);
8485 return (qualifiers | 8);
8488 case QUAL_UNION_TYPE:
8489 return (qualifiers | 9);
8492 return (qualifiers | 10);
8495 return (qualifiers | 16);
8498 /* If this is a range type, consider it to be the underlying
8500 if (TREE_TYPE (type) != 0)
8503 /* Carefully distinguish all the standard types of C,
8504 without messing up if the language is not C. We do this by
8505 testing TYPE_PRECISION and TYPE_UNSIGNED. The old code used to
8506 look at both the names and the above fields, but that's redundant.
8507 Any type whose size is between two C types will be considered
8508 to be the wider of the two types. Also, we do not have a
8509 special code to use for "long long", so anything wider than
8510 long is treated the same. Note that we can't distinguish
8511 between "int" and "long" in this code if they are the same
8512 size, but that's fine, since neither can the assembler. */
8514 if (TYPE_PRECISION (type) <= CHAR_TYPE_SIZE)
8515 return (qualifiers | (TYPE_UNSIGNED (type) ? 12 : 2));
8517 else if (TYPE_PRECISION (type) <= SHORT_TYPE_SIZE)
8518 return (qualifiers | (TYPE_UNSIGNED (type) ? 13 : 3));
8520 else if (TYPE_PRECISION (type) <= INT_TYPE_SIZE)
8521 return (qualifiers | (TYPE_UNSIGNED (type) ? 14 : 4));
8524 return (qualifiers | (TYPE_UNSIGNED (type) ? 15 : 5));
8527 /* If this is a range type, consider it to be the underlying
8529 if (TREE_TYPE (type) != 0)
8532 /* Carefully distinguish all the standard types of C,
8533 without messing up if the language is not C. */
8535 if (TYPE_PRECISION (type) == FLOAT_TYPE_SIZE)
8536 return (qualifiers | 6);
8539 return (qualifiers | 7);
8541 case COMPLEX_TYPE: /* GNU Fortran COMPLEX type. */
8542 /* ??? We need to distinguish between double and float complex types,
8543 but I don't know how yet because I can't reach this code from
8544 existing front-ends. */
8545 return (qualifiers | 7); /* Who knows? */
8548 case BOOLEAN_TYPE: /* Boolean truth value type. */
8554 gcc_unreachable (); /* Not a type! */
8561 /* Nested function support. */
8563 /* Emit RTL insns to initialize the variable parts of a trampoline.
8564 FNADDR is an RTX for the address of the function's pure code.
8565 CXT is an RTX for the static chain value for the function.
8567 This takes 16 insns: 2 shifts & 2 ands (to split up addresses), 4 sethi
8568 (to load in opcodes), 4 iors (to merge address and opcodes), and 4 writes
8569 (to store insns). This is a bit excessive. Perhaps a different
8570 mechanism would be better here.
8572 Emit enough FLUSH insns to synchronize the data and instruction caches. */
8575 sparc32_initialize_trampoline (rtx m_tramp, rtx fnaddr, rtx cxt)
8577 /* SPARC 32-bit trampoline:
8580 sethi %hi(static), %g2
8582 or %g2, %lo(static), %g2
8584 SETHI i,r = 00rr rrr1 00ii iiii iiii iiii iiii iiii
8585 JMPL r+i,d = 10dd ddd1 1100 0rrr rr1i iiii iiii iiii
8589 (adjust_address (m_tramp, SImode, 0),
8590 expand_binop (SImode, ior_optab,
8591 expand_shift (RSHIFT_EXPR, SImode, fnaddr, 10, 0, 1),
8592 GEN_INT (trunc_int_for_mode (0x03000000, SImode)),
8593 NULL_RTX, 1, OPTAB_DIRECT));
8596 (adjust_address (m_tramp, SImode, 4),
8597 expand_binop (SImode, ior_optab,
8598 expand_shift (RSHIFT_EXPR, SImode, cxt, 10, 0, 1),
8599 GEN_INT (trunc_int_for_mode (0x05000000, SImode)),
8600 NULL_RTX, 1, OPTAB_DIRECT));
8603 (adjust_address (m_tramp, SImode, 8),
8604 expand_binop (SImode, ior_optab,
8605 expand_and (SImode, fnaddr, GEN_INT (0x3ff), NULL_RTX),
8606 GEN_INT (trunc_int_for_mode (0x81c06000, SImode)),
8607 NULL_RTX, 1, OPTAB_DIRECT));
8610 (adjust_address (m_tramp, SImode, 12),
8611 expand_binop (SImode, ior_optab,
8612 expand_and (SImode, cxt, GEN_INT (0x3ff), NULL_RTX),
8613 GEN_INT (trunc_int_for_mode (0x8410a000, SImode)),
8614 NULL_RTX, 1, OPTAB_DIRECT));
8616 /* On UltraSPARC a flush flushes an entire cache line. The trampoline is
8617 aligned on a 16 byte boundary so one flush clears it all. */
8618 emit_insn (gen_flush (validize_mem (adjust_address (m_tramp, SImode, 0))));
8619 if (sparc_cpu != PROCESSOR_ULTRASPARC
8620 && sparc_cpu != PROCESSOR_ULTRASPARC3
8621 && sparc_cpu != PROCESSOR_NIAGARA
8622 && sparc_cpu != PROCESSOR_NIAGARA2
8623 && sparc_cpu != PROCESSOR_NIAGARA3
8624 && sparc_cpu != PROCESSOR_NIAGARA4)
8625 emit_insn (gen_flush (validize_mem (adjust_address (m_tramp, SImode, 8))));
8627 /* Call __enable_execute_stack after writing onto the stack to make sure
8628 the stack address is accessible. */
8629 #ifdef HAVE_ENABLE_EXECUTE_STACK
8630 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
8631 LCT_NORMAL, VOIDmode, 1, XEXP (m_tramp, 0), Pmode);
8636 /* The 64-bit version is simpler because it makes more sense to load the
8637 values as "immediate" data out of the trampoline. It's also easier since
8638 we can read the PC without clobbering a register. */
8641 sparc64_initialize_trampoline (rtx m_tramp, rtx fnaddr, rtx cxt)
8643 /* SPARC 64-bit trampoline:
8652 emit_move_insn (adjust_address (m_tramp, SImode, 0),
8653 GEN_INT (trunc_int_for_mode (0x83414000, SImode)));
8654 emit_move_insn (adjust_address (m_tramp, SImode, 4),
8655 GEN_INT (trunc_int_for_mode (0xca586018, SImode)));
8656 emit_move_insn (adjust_address (m_tramp, SImode, 8),
8657 GEN_INT (trunc_int_for_mode (0x81c14000, SImode)));
8658 emit_move_insn (adjust_address (m_tramp, SImode, 12),
8659 GEN_INT (trunc_int_for_mode (0xca586010, SImode)));
8660 emit_move_insn (adjust_address (m_tramp, DImode, 16), cxt);
8661 emit_move_insn (adjust_address (m_tramp, DImode, 24), fnaddr);
8662 emit_insn (gen_flushdi (validize_mem (adjust_address (m_tramp, DImode, 0))));
8664 if (sparc_cpu != PROCESSOR_ULTRASPARC
8665 && sparc_cpu != PROCESSOR_ULTRASPARC3
8666 && sparc_cpu != PROCESSOR_NIAGARA
8667 && sparc_cpu != PROCESSOR_NIAGARA2
8668 && sparc_cpu != PROCESSOR_NIAGARA3
8669 && sparc_cpu != PROCESSOR_NIAGARA4)
8670 emit_insn (gen_flushdi (validize_mem (adjust_address (m_tramp, DImode, 8))));
8672 /* Call __enable_execute_stack after writing onto the stack to make sure
8673 the stack address is accessible. */
8674 #ifdef HAVE_ENABLE_EXECUTE_STACK
8675 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
8676 LCT_NORMAL, VOIDmode, 1, XEXP (m_tramp, 0), Pmode);
8680 /* Worker for TARGET_TRAMPOLINE_INIT. */
8683 sparc_trampoline_init (rtx m_tramp, tree fndecl, rtx cxt)
8685 rtx fnaddr = force_reg (Pmode, XEXP (DECL_RTL (fndecl), 0));
8686 cxt = force_reg (Pmode, cxt);
8688 sparc64_initialize_trampoline (m_tramp, fnaddr, cxt);
8690 sparc32_initialize_trampoline (m_tramp, fnaddr, cxt);
8693 /* Adjust the cost of a scheduling dependency. Return the new cost of
8694 a dependency LINK or INSN on DEP_INSN. COST is the current cost. */
8697 supersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
8699 enum attr_type insn_type;
8701 if (! recog_memoized (insn))
8704 insn_type = get_attr_type (insn);
8706 if (REG_NOTE_KIND (link) == 0)
8708 /* Data dependency; DEP_INSN writes a register that INSN reads some
8711 /* if a load, then the dependence must be on the memory address;
8712 add an extra "cycle". Note that the cost could be two cycles
8713 if the reg was written late in an instruction group; we ca not tell
8715 if (insn_type == TYPE_LOAD || insn_type == TYPE_FPLOAD)
8718 /* Get the delay only if the address of the store is the dependence. */
8719 if (insn_type == TYPE_STORE || insn_type == TYPE_FPSTORE)
8721 rtx pat = PATTERN(insn);
8722 rtx dep_pat = PATTERN (dep_insn);
8724 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
8725 return cost; /* This should not happen! */
8727 /* The dependency between the two instructions was on the data that
8728 is being stored. Assume that this implies that the address of the
8729 store is not dependent. */
8730 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
8733 return cost + 3; /* An approximation. */
8736 /* A shift instruction cannot receive its data from an instruction
8737 in the same cycle; add a one cycle penalty. */
8738 if (insn_type == TYPE_SHIFT)
8739 return cost + 3; /* Split before cascade into shift. */
8743 /* Anti- or output- dependency; DEP_INSN reads/writes a register that
8744 INSN writes some cycles later. */
8746 /* These are only significant for the fpu unit; writing a fp reg before
8747 the fpu has finished with it stalls the processor. */
8749 /* Reusing an integer register causes no problems. */
8750 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
8758 hypersparc_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
8760 enum attr_type insn_type, dep_type;
8761 rtx pat = PATTERN(insn);
8762 rtx dep_pat = PATTERN (dep_insn);
8764 if (recog_memoized (insn) < 0 || recog_memoized (dep_insn) < 0)
8767 insn_type = get_attr_type (insn);
8768 dep_type = get_attr_type (dep_insn);
8770 switch (REG_NOTE_KIND (link))
8773 /* Data dependency; DEP_INSN writes a register that INSN reads some
8780 /* Get the delay iff the address of the store is the dependence. */
8781 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET)
8784 if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat)))
8791 /* If a load, then the dependence must be on the memory address. If
8792 the addresses aren't equal, then it might be a false dependency */
8793 if (dep_type == TYPE_STORE || dep_type == TYPE_FPSTORE)
8795 if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET
8796 || GET_CODE (SET_DEST (dep_pat)) != MEM
8797 || GET_CODE (SET_SRC (pat)) != MEM
8798 || ! rtx_equal_p (XEXP (SET_DEST (dep_pat), 0),
8799 XEXP (SET_SRC (pat), 0)))
8807 /* Compare to branch latency is 0. There is no benefit from
8808 separating compare and branch. */
8809 if (dep_type == TYPE_COMPARE)
8811 /* Floating point compare to branch latency is less than
8812 compare to conditional move. */
8813 if (dep_type == TYPE_FPCMP)
8822 /* Anti-dependencies only penalize the fpu unit. */
8823 if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT)
8835 sparc_adjust_cost(rtx insn, rtx link, rtx dep, int cost)
8839 case PROCESSOR_SUPERSPARC:
8840 cost = supersparc_adjust_cost (insn, link, dep, cost);
8842 case PROCESSOR_HYPERSPARC:
8843 case PROCESSOR_SPARCLITE86X:
8844 cost = hypersparc_adjust_cost (insn, link, dep, cost);
8853 sparc_sched_init (FILE *dump ATTRIBUTE_UNUSED,
8854 int sched_verbose ATTRIBUTE_UNUSED,
8855 int max_ready ATTRIBUTE_UNUSED)
8859 sparc_use_sched_lookahead (void)
8861 if (sparc_cpu == PROCESSOR_NIAGARA
8862 || sparc_cpu == PROCESSOR_NIAGARA2
8863 || sparc_cpu == PROCESSOR_NIAGARA3
8864 || sparc_cpu == PROCESSOR_NIAGARA4)
8866 if (sparc_cpu == PROCESSOR_ULTRASPARC
8867 || sparc_cpu == PROCESSOR_ULTRASPARC3)
8869 if ((1 << sparc_cpu) &
8870 ((1 << PROCESSOR_SUPERSPARC) | (1 << PROCESSOR_HYPERSPARC) |
8871 (1 << PROCESSOR_SPARCLITE86X)))
8877 sparc_issue_rate (void)
8881 case PROCESSOR_NIAGARA:
8882 case PROCESSOR_NIAGARA2:
8883 case PROCESSOR_NIAGARA3:
8884 case PROCESSOR_NIAGARA4:
8888 /* Assume V9 processors are capable of at least dual-issue. */
8890 case PROCESSOR_SUPERSPARC:
8892 case PROCESSOR_HYPERSPARC:
8893 case PROCESSOR_SPARCLITE86X:
8895 case PROCESSOR_ULTRASPARC:
8896 case PROCESSOR_ULTRASPARC3:
8902 set_extends (rtx insn)
8904 register rtx pat = PATTERN (insn);
8906 switch (GET_CODE (SET_SRC (pat)))
8908 /* Load and some shift instructions zero extend. */
8911 /* sethi clears the high bits */
8913 /* LO_SUM is used with sethi. sethi cleared the high
8914 bits and the values used with lo_sum are positive */
8916 /* Store flag stores 0 or 1 */
8926 rtx op0 = XEXP (SET_SRC (pat), 0);
8927 rtx op1 = XEXP (SET_SRC (pat), 1);
8928 if (GET_CODE (op1) == CONST_INT)
8929 return INTVAL (op1) >= 0;
8930 if (GET_CODE (op0) != REG)
8932 if (sparc_check_64 (op0, insn) == 1)
8934 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
8939 rtx op0 = XEXP (SET_SRC (pat), 0);
8940 rtx op1 = XEXP (SET_SRC (pat), 1);
8941 if (GET_CODE (op0) != REG || sparc_check_64 (op0, insn) <= 0)
8943 if (GET_CODE (op1) == CONST_INT)
8944 return INTVAL (op1) >= 0;
8945 return (GET_CODE (op1) == REG && sparc_check_64 (op1, insn) == 1);
8948 return GET_MODE (SET_SRC (pat)) == SImode;
8949 /* Positive integers leave the high bits zero. */
8951 return ! (CONST_DOUBLE_LOW (SET_SRC (pat)) & 0x80000000);
8953 return ! (INTVAL (SET_SRC (pat)) & 0x80000000);
8956 return - (GET_MODE (SET_SRC (pat)) == SImode);
8958 return sparc_check_64 (SET_SRC (pat), insn);
8964 /* We _ought_ to have only one kind per function, but... */
8965 static GTY(()) rtx sparc_addr_diff_list;
8966 static GTY(()) rtx sparc_addr_list;
8969 sparc_defer_case_vector (rtx lab, rtx vec, int diff)
8971 vec = gen_rtx_EXPR_LIST (VOIDmode, lab, vec);
8973 sparc_addr_diff_list
8974 = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_diff_list);
8976 sparc_addr_list = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_list);
8980 sparc_output_addr_vec (rtx vec)
8982 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
8983 int idx, vlen = XVECLEN (body, 0);
8985 #ifdef ASM_OUTPUT_ADDR_VEC_START
8986 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
8989 #ifdef ASM_OUTPUT_CASE_LABEL
8990 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
8993 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
8996 for (idx = 0; idx < vlen; idx++)
8998 ASM_OUTPUT_ADDR_VEC_ELT
8999 (asm_out_file, CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 0, idx), 0)));
9002 #ifdef ASM_OUTPUT_ADDR_VEC_END
9003 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
9008 sparc_output_addr_diff_vec (rtx vec)
9010 rtx lab = XEXP (vec, 0), body = XEXP (vec, 1);
9011 rtx base = XEXP (XEXP (body, 0), 0);
9012 int idx, vlen = XVECLEN (body, 1);
9014 #ifdef ASM_OUTPUT_ADDR_VEC_START
9015 ASM_OUTPUT_ADDR_VEC_START (asm_out_file);
9018 #ifdef ASM_OUTPUT_CASE_LABEL
9019 ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab),
9022 (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (lab));
9025 for (idx = 0; idx < vlen; idx++)
9027 ASM_OUTPUT_ADDR_DIFF_ELT
9030 CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 1, idx), 0)),
9031 CODE_LABEL_NUMBER (base));
9034 #ifdef ASM_OUTPUT_ADDR_VEC_END
9035 ASM_OUTPUT_ADDR_VEC_END (asm_out_file);
9040 sparc_output_deferred_case_vectors (void)
9045 if (sparc_addr_list == NULL_RTX
9046 && sparc_addr_diff_list == NULL_RTX)
9049 /* Align to cache line in the function's code section. */
9050 switch_to_section (current_function_section ());
9052 align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
9054 ASM_OUTPUT_ALIGN (asm_out_file, align);
9056 for (t = sparc_addr_list; t ; t = XEXP (t, 1))
9057 sparc_output_addr_vec (XEXP (t, 0));
9058 for (t = sparc_addr_diff_list; t ; t = XEXP (t, 1))
9059 sparc_output_addr_diff_vec (XEXP (t, 0));
9061 sparc_addr_list = sparc_addr_diff_list = NULL_RTX;
9064 /* Return 0 if the high 32 bits of X (the low word of X, if DImode) are
9065 unknown. Return 1 if the high bits are zero, -1 if the register is
9068 sparc_check_64 (rtx x, rtx insn)
9070 /* If a register is set only once it is safe to ignore insns this
9071 code does not know how to handle. The loop will either recognize
9072 the single set and return the correct value or fail to recognize
9077 gcc_assert (GET_CODE (x) == REG);
9079 if (GET_MODE (x) == DImode)
9080 y = gen_rtx_REG (SImode, REGNO (x) + WORDS_BIG_ENDIAN);
9082 if (flag_expensive_optimizations
9083 && df && DF_REG_DEF_COUNT (REGNO (y)) == 1)
9089 insn = get_last_insn_anywhere ();
9094 while ((insn = PREV_INSN (insn)))
9096 switch (GET_CODE (insn))
9109 rtx pat = PATTERN (insn);
9110 if (GET_CODE (pat) != SET)
9112 if (rtx_equal_p (x, SET_DEST (pat)))
9113 return set_extends (insn);
9114 if (y && rtx_equal_p (y, SET_DEST (pat)))
9115 return set_extends (insn);
9116 if (reg_overlap_mentioned_p (SET_DEST (pat), y))
9124 /* Output a wide shift instruction in V8+ mode. INSN is the instruction,
9125 OPERANDS are its operands and OPCODE is the mnemonic to be used. */
9128 output_v8plus_shift (rtx insn, rtx *operands, const char *opcode)
9130 static char asm_code[60];
9132 /* The scratch register is only required when the destination
9133 register is not a 64-bit global or out register. */
9134 if (which_alternative != 2)
9135 operands[3] = operands[0];
9137 /* We can only shift by constants <= 63. */
9138 if (GET_CODE (operands[2]) == CONST_INT)
9139 operands[2] = GEN_INT (INTVAL (operands[2]) & 0x3f);
9141 if (GET_CODE (operands[1]) == CONST_INT)
9143 output_asm_insn ("mov\t%1, %3", operands);
9147 output_asm_insn ("sllx\t%H1, 32, %3", operands);
9148 if (sparc_check_64 (operands[1], insn) <= 0)
9149 output_asm_insn ("srl\t%L1, 0, %L1", operands);
9150 output_asm_insn ("or\t%L1, %3, %3", operands);
9153 strcpy (asm_code, opcode);
9155 if (which_alternative != 2)
9156 return strcat (asm_code, "\t%0, %2, %L0\n\tsrlx\t%L0, 32, %H0");
9159 strcat (asm_code, "\t%3, %2, %3\n\tsrlx\t%3, 32, %H0\n\tmov\t%3, %L0");
9162 /* Output rtl to increment the profiler label LABELNO
9163 for profiling a function entry. */
9166 sparc_profile_hook (int labelno)
9171 fun = gen_rtx_SYMBOL_REF (Pmode, MCOUNT_FUNCTION);
9172 if (NO_PROFILE_COUNTERS)
9174 emit_library_call (fun, LCT_NORMAL, VOIDmode, 0);
9178 ASM_GENERATE_INTERNAL_LABEL (buf, "LP", labelno);
9179 lab = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (buf));
9180 emit_library_call (fun, LCT_NORMAL, VOIDmode, 1, lab, Pmode);
9184 #ifdef TARGET_SOLARIS
9185 /* Solaris implementation of TARGET_ASM_NAMED_SECTION. */
9188 sparc_solaris_elf_asm_named_section (const char *name, unsigned int flags,
9189 tree decl ATTRIBUTE_UNUSED)
9191 if (HAVE_COMDAT_GROUP && flags & SECTION_LINKONCE)
9193 solaris_elf_asm_comdat_section (name, flags, decl);
9197 fprintf (asm_out_file, "\t.section\t\"%s\"", name);
9199 if (!(flags & SECTION_DEBUG))
9200 fputs (",#alloc", asm_out_file);
9201 if (flags & SECTION_WRITE)
9202 fputs (",#write", asm_out_file);
9203 if (flags & SECTION_TLS)
9204 fputs (",#tls", asm_out_file);
9205 if (flags & SECTION_CODE)
9206 fputs (",#execinstr", asm_out_file);
9208 /* ??? Handle SECTION_BSS. */
9210 fputc ('\n', asm_out_file);
9212 #endif /* TARGET_SOLARIS */
9214 /* We do not allow indirect calls to be optimized into sibling calls.
9216 We cannot use sibling calls when delayed branches are disabled
9217 because they will likely require the call delay slot to be filled.
9219 Also, on SPARC 32-bit we cannot emit a sibling call when the
9220 current function returns a structure. This is because the "unimp
9221 after call" convention would cause the callee to return to the
9222 wrong place. The generic code already disallows cases where the
9223 function being called returns a structure.
9225 It may seem strange how this last case could occur. Usually there
9226 is code after the call which jumps to epilogue code which dumps the
9227 return value into the struct return area. That ought to invalidate
9228 the sibling call right? Well, in the C++ case we can end up passing
9229 the pointer to the struct return area to a constructor (which returns
9230 void) and then nothing else happens. Such a sibling call would look
9231 valid without the added check here.
9233 VxWorks PIC PLT entries require the global pointer to be initialized
9234 on entry. We therefore can't emit sibling calls to them. */
9236 sparc_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
9239 && flag_delayed_branch
9240 && (TARGET_ARCH64 || ! cfun->returns_struct)
9241 && !(TARGET_VXWORKS_RTP
9243 && !targetm.binds_local_p (decl)));
9246 /* libfunc renaming. */
9249 sparc_init_libfuncs (void)
9253 /* Use the subroutines that Sun's library provides for integer
9254 multiply and divide. The `*' prevents an underscore from
9255 being prepended by the compiler. .umul is a little faster
9257 set_optab_libfunc (smul_optab, SImode, "*.umul");
9258 set_optab_libfunc (sdiv_optab, SImode, "*.div");
9259 set_optab_libfunc (udiv_optab, SImode, "*.udiv");
9260 set_optab_libfunc (smod_optab, SImode, "*.rem");
9261 set_optab_libfunc (umod_optab, SImode, "*.urem");
9263 /* TFmode arithmetic. These names are part of the SPARC 32bit ABI. */
9264 set_optab_libfunc (add_optab, TFmode, "_Q_add");
9265 set_optab_libfunc (sub_optab, TFmode, "_Q_sub");
9266 set_optab_libfunc (neg_optab, TFmode, "_Q_neg");
9267 set_optab_libfunc (smul_optab, TFmode, "_Q_mul");
9268 set_optab_libfunc (sdiv_optab, TFmode, "_Q_div");
9270 /* We can define the TFmode sqrt optab only if TARGET_FPU. This
9271 is because with soft-float, the SFmode and DFmode sqrt
9272 instructions will be absent, and the compiler will notice and
9273 try to use the TFmode sqrt instruction for calls to the
9274 builtin function sqrt, but this fails. */
9276 set_optab_libfunc (sqrt_optab, TFmode, "_Q_sqrt");
9278 set_optab_libfunc (eq_optab, TFmode, "_Q_feq");
9279 set_optab_libfunc (ne_optab, TFmode, "_Q_fne");
9280 set_optab_libfunc (gt_optab, TFmode, "_Q_fgt");
9281 set_optab_libfunc (ge_optab, TFmode, "_Q_fge");
9282 set_optab_libfunc (lt_optab, TFmode, "_Q_flt");
9283 set_optab_libfunc (le_optab, TFmode, "_Q_fle");
9285 set_conv_libfunc (sext_optab, TFmode, SFmode, "_Q_stoq");
9286 set_conv_libfunc (sext_optab, TFmode, DFmode, "_Q_dtoq");
9287 set_conv_libfunc (trunc_optab, SFmode, TFmode, "_Q_qtos");
9288 set_conv_libfunc (trunc_optab, DFmode, TFmode, "_Q_qtod");
9290 set_conv_libfunc (sfix_optab, SImode, TFmode, "_Q_qtoi");
9291 set_conv_libfunc (ufix_optab, SImode, TFmode, "_Q_qtou");
9292 set_conv_libfunc (sfloat_optab, TFmode, SImode, "_Q_itoq");
9293 set_conv_libfunc (ufloat_optab, TFmode, SImode, "_Q_utoq");
9295 if (DITF_CONVERSION_LIBFUNCS)
9297 set_conv_libfunc (sfix_optab, DImode, TFmode, "_Q_qtoll");
9298 set_conv_libfunc (ufix_optab, DImode, TFmode, "_Q_qtoull");
9299 set_conv_libfunc (sfloat_optab, TFmode, DImode, "_Q_lltoq");
9300 set_conv_libfunc (ufloat_optab, TFmode, DImode, "_Q_ulltoq");
9303 if (SUN_CONVERSION_LIBFUNCS)
9305 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftoll");
9306 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoull");
9307 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtoll");
9308 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoull");
9313 /* In the SPARC 64bit ABI, SImode multiply and divide functions
9314 do not exist in the library. Make sure the compiler does not
9315 emit calls to them by accident. (It should always use the
9316 hardware instructions.) */
9317 set_optab_libfunc (smul_optab, SImode, 0);
9318 set_optab_libfunc (sdiv_optab, SImode, 0);
9319 set_optab_libfunc (udiv_optab, SImode, 0);
9320 set_optab_libfunc (smod_optab, SImode, 0);
9321 set_optab_libfunc (umod_optab, SImode, 0);
9323 if (SUN_INTEGER_MULTIPLY_64)
9325 set_optab_libfunc (smul_optab, DImode, "__mul64");
9326 set_optab_libfunc (sdiv_optab, DImode, "__div64");
9327 set_optab_libfunc (udiv_optab, DImode, "__udiv64");
9328 set_optab_libfunc (smod_optab, DImode, "__rem64");
9329 set_optab_libfunc (umod_optab, DImode, "__urem64");
9332 if (SUN_CONVERSION_LIBFUNCS)
9334 set_conv_libfunc (sfix_optab, DImode, SFmode, "__ftol");
9335 set_conv_libfunc (ufix_optab, DImode, SFmode, "__ftoul");
9336 set_conv_libfunc (sfix_optab, DImode, DFmode, "__dtol");
9337 set_conv_libfunc (ufix_optab, DImode, DFmode, "__dtoul");
9342 static tree def_builtin(const char *name, int code, tree type)
9344 return add_builtin_function(name, type, code, BUILT_IN_MD, NULL,
9348 static tree def_builtin_const(const char *name, int code, tree type)
9350 tree t = def_builtin(name, code, type);
9353 TREE_READONLY (t) = 1;
9358 /* Implement the TARGET_INIT_BUILTINS target hook.
9359 Create builtin functions for special SPARC instructions. */
9362 sparc_init_builtins (void)
9365 sparc_vis_init_builtins ();
9368 /* Create builtin functions for VIS 1.0 instructions. */
9371 sparc_vis_init_builtins (void)
9373 tree v4qi = build_vector_type (unsigned_intQI_type_node, 4);
9374 tree v8qi = build_vector_type (unsigned_intQI_type_node, 8);
9375 tree v4hi = build_vector_type (intHI_type_node, 4);
9376 tree v2hi = build_vector_type (intHI_type_node, 2);
9377 tree v2si = build_vector_type (intSI_type_node, 2);
9378 tree v1si = build_vector_type (intSI_type_node, 1);
9380 tree v4qi_ftype_v4hi = build_function_type_list (v4qi, v4hi, 0);
9381 tree v8qi_ftype_v2si_v8qi = build_function_type_list (v8qi, v2si, v8qi, 0);
9382 tree v2hi_ftype_v2si = build_function_type_list (v2hi, v2si, 0);
9383 tree v4hi_ftype_v4qi = build_function_type_list (v4hi, v4qi, 0);
9384 tree v8qi_ftype_v4qi_v4qi = build_function_type_list (v8qi, v4qi, v4qi, 0);
9385 tree v4hi_ftype_v4qi_v4hi = build_function_type_list (v4hi, v4qi, v4hi, 0);
9386 tree v4hi_ftype_v4qi_v2hi = build_function_type_list (v4hi, v4qi, v2hi, 0);
9387 tree v2si_ftype_v4qi_v2hi = build_function_type_list (v2si, v4qi, v2hi, 0);
9388 tree v4hi_ftype_v8qi_v4hi = build_function_type_list (v4hi, v8qi, v4hi, 0);
9389 tree v4hi_ftype_v4hi_v4hi = build_function_type_list (v4hi, v4hi, v4hi, 0);
9390 tree v2si_ftype_v2si_v2si = build_function_type_list (v2si, v2si, v2si, 0);
9391 tree v8qi_ftype_v8qi_v8qi = build_function_type_list (v8qi, v8qi, v8qi, 0);
9392 tree v2hi_ftype_v2hi_v2hi = build_function_type_list (v2hi, v2hi, v2hi, 0);
9393 tree v1si_ftype_v1si_v1si = build_function_type_list (v1si, v1si, v1si, 0);
9394 tree di_ftype_v8qi_v8qi_di = build_function_type_list (intDI_type_node,
9396 intDI_type_node, 0);
9397 tree di_ftype_v8qi_v8qi = build_function_type_list (intDI_type_node,
9399 tree si_ftype_v8qi_v8qi = build_function_type_list (intSI_type_node,
9401 tree di_ftype_di_di = build_function_type_list (intDI_type_node,
9403 intDI_type_node, 0);
9404 tree si_ftype_si_si = build_function_type_list (intSI_type_node,
9406 intSI_type_node, 0);
9407 tree ptr_ftype_ptr_si = build_function_type_list (ptr_type_node,
9409 intSI_type_node, 0);
9410 tree ptr_ftype_ptr_di = build_function_type_list (ptr_type_node,
9412 intDI_type_node, 0);
9413 tree si_ftype_ptr_ptr = build_function_type_list (intSI_type_node,
9416 tree di_ftype_ptr_ptr = build_function_type_list (intDI_type_node,
9419 tree si_ftype_v4hi_v4hi = build_function_type_list (intSI_type_node,
9421 tree si_ftype_v2si_v2si = build_function_type_list (intSI_type_node,
9423 tree di_ftype_v4hi_v4hi = build_function_type_list (intDI_type_node,
9425 tree di_ftype_v2si_v2si = build_function_type_list (intDI_type_node,
9427 tree void_ftype_di = build_function_type_list (void_type_node,
9428 intDI_type_node, 0);
9429 tree di_ftype_void = build_function_type_list (intDI_type_node,
9431 tree void_ftype_si = build_function_type_list (void_type_node,
9432 intSI_type_node, 0);
9433 tree sf_ftype_sf_sf = build_function_type_list (float_type_node,
9435 float_type_node, 0);
9436 tree df_ftype_df_df = build_function_type_list (double_type_node,
9438 double_type_node, 0);
9440 /* Packing and expanding vectors. */
9441 def_builtin ("__builtin_vis_fpack16", CODE_FOR_fpack16_vis,
9443 def_builtin ("__builtin_vis_fpack32", CODE_FOR_fpack32_vis,
9444 v8qi_ftype_v2si_v8qi);
9445 def_builtin ("__builtin_vis_fpackfix", CODE_FOR_fpackfix_vis,
9447 def_builtin_const ("__builtin_vis_fexpand", CODE_FOR_fexpand_vis,
9449 def_builtin_const ("__builtin_vis_fpmerge", CODE_FOR_fpmerge_vis,
9450 v8qi_ftype_v4qi_v4qi);
9452 /* Multiplications. */
9453 def_builtin_const ("__builtin_vis_fmul8x16", CODE_FOR_fmul8x16_vis,
9454 v4hi_ftype_v4qi_v4hi);
9455 def_builtin_const ("__builtin_vis_fmul8x16au", CODE_FOR_fmul8x16au_vis,
9456 v4hi_ftype_v4qi_v2hi);
9457 def_builtin_const ("__builtin_vis_fmul8x16al", CODE_FOR_fmul8x16al_vis,
9458 v4hi_ftype_v4qi_v2hi);
9459 def_builtin_const ("__builtin_vis_fmul8sux16", CODE_FOR_fmul8sux16_vis,
9460 v4hi_ftype_v8qi_v4hi);
9461 def_builtin_const ("__builtin_vis_fmul8ulx16", CODE_FOR_fmul8ulx16_vis,
9462 v4hi_ftype_v8qi_v4hi);
9463 def_builtin_const ("__builtin_vis_fmuld8sux16", CODE_FOR_fmuld8sux16_vis,
9464 v2si_ftype_v4qi_v2hi);
9465 def_builtin_const ("__builtin_vis_fmuld8ulx16", CODE_FOR_fmuld8ulx16_vis,
9466 v2si_ftype_v4qi_v2hi);
9468 /* Data aligning. */
9469 def_builtin ("__builtin_vis_faligndatav4hi", CODE_FOR_faligndatav4hi_vis,
9470 v4hi_ftype_v4hi_v4hi);
9471 def_builtin ("__builtin_vis_faligndatav8qi", CODE_FOR_faligndatav8qi_vis,
9472 v8qi_ftype_v8qi_v8qi);
9473 def_builtin ("__builtin_vis_faligndatav2si", CODE_FOR_faligndatav2si_vis,
9474 v2si_ftype_v2si_v2si);
9475 def_builtin ("__builtin_vis_faligndatadi", CODE_FOR_faligndatav1di_vis,
9478 def_builtin ("__builtin_vis_write_gsr", CODE_FOR_wrgsr_vis,
9480 def_builtin ("__builtin_vis_read_gsr", CODE_FOR_rdgsr_vis,
9485 def_builtin ("__builtin_vis_alignaddr", CODE_FOR_alignaddrdi_vis,
9487 def_builtin ("__builtin_vis_alignaddrl", CODE_FOR_alignaddrldi_vis,
9492 def_builtin ("__builtin_vis_alignaddr", CODE_FOR_alignaddrsi_vis,
9494 def_builtin ("__builtin_vis_alignaddrl", CODE_FOR_alignaddrlsi_vis,
9498 /* Pixel distance. */
9499 def_builtin_const ("__builtin_vis_pdist", CODE_FOR_pdist_vis,
9500 di_ftype_v8qi_v8qi_di);
9502 /* Edge handling. */
9505 def_builtin_const ("__builtin_vis_edge8", CODE_FOR_edge8di_vis,
9507 def_builtin_const ("__builtin_vis_edge8l", CODE_FOR_edge8ldi_vis,
9509 def_builtin_const ("__builtin_vis_edge16", CODE_FOR_edge16di_vis,
9511 def_builtin_const ("__builtin_vis_edge16l", CODE_FOR_edge16ldi_vis,
9513 def_builtin_const ("__builtin_vis_edge32", CODE_FOR_edge32di_vis,
9515 def_builtin_const ("__builtin_vis_edge32l", CODE_FOR_edge32ldi_vis,
9519 def_builtin_const ("__builtin_vis_edge8n", CODE_FOR_edge8ndi_vis,
9521 def_builtin_const ("__builtin_vis_edge8ln", CODE_FOR_edge8lndi_vis,
9523 def_builtin_const ("__builtin_vis_edge16n", CODE_FOR_edge16ndi_vis,
9525 def_builtin_const ("__builtin_vis_edge16ln", CODE_FOR_edge16lndi_vis,
9527 def_builtin_const ("__builtin_vis_edge32n", CODE_FOR_edge32ndi_vis,
9529 def_builtin_const ("__builtin_vis_edge32ln", CODE_FOR_edge32lndi_vis,
9535 def_builtin_const ("__builtin_vis_edge8", CODE_FOR_edge8si_vis,
9537 def_builtin_const ("__builtin_vis_edge8l", CODE_FOR_edge8lsi_vis,
9539 def_builtin_const ("__builtin_vis_edge16", CODE_FOR_edge16si_vis,
9541 def_builtin_const ("__builtin_vis_edge16l", CODE_FOR_edge16lsi_vis,
9543 def_builtin_const ("__builtin_vis_edge32", CODE_FOR_edge32si_vis,
9545 def_builtin_const ("__builtin_vis_edge32l", CODE_FOR_edge32lsi_vis,
9549 def_builtin_const ("__builtin_vis_edge8n", CODE_FOR_edge8nsi_vis,
9551 def_builtin_const ("__builtin_vis_edge8ln", CODE_FOR_edge8lnsi_vis,
9553 def_builtin_const ("__builtin_vis_edge16n", CODE_FOR_edge16nsi_vis,
9555 def_builtin_const ("__builtin_vis_edge16ln", CODE_FOR_edge16lnsi_vis,
9557 def_builtin_const ("__builtin_vis_edge32n", CODE_FOR_edge32nsi_vis,
9559 def_builtin_const ("__builtin_vis_edge32ln", CODE_FOR_edge32lnsi_vis,
9564 /* Pixel compare. */
9567 def_builtin_const ("__builtin_vis_fcmple16", CODE_FOR_fcmple16di_vis,
9568 di_ftype_v4hi_v4hi);
9569 def_builtin_const ("__builtin_vis_fcmple32", CODE_FOR_fcmple32di_vis,
9570 di_ftype_v2si_v2si);
9571 def_builtin_const ("__builtin_vis_fcmpne16", CODE_FOR_fcmpne16di_vis,
9572 di_ftype_v4hi_v4hi);
9573 def_builtin_const ("__builtin_vis_fcmpne32", CODE_FOR_fcmpne32di_vis,
9574 di_ftype_v2si_v2si);
9575 def_builtin_const ("__builtin_vis_fcmpgt16", CODE_FOR_fcmpgt16di_vis,
9576 di_ftype_v4hi_v4hi);
9577 def_builtin_const ("__builtin_vis_fcmpgt32", CODE_FOR_fcmpgt32di_vis,
9578 di_ftype_v2si_v2si);
9579 def_builtin_const ("__builtin_vis_fcmpeq16", CODE_FOR_fcmpeq16di_vis,
9580 di_ftype_v4hi_v4hi);
9581 def_builtin_const ("__builtin_vis_fcmpeq32", CODE_FOR_fcmpeq32di_vis,
9582 di_ftype_v2si_v2si);
9586 def_builtin_const ("__builtin_vis_fcmple16", CODE_FOR_fcmple16si_vis,
9587 si_ftype_v4hi_v4hi);
9588 def_builtin_const ("__builtin_vis_fcmple32", CODE_FOR_fcmple32si_vis,
9589 si_ftype_v2si_v2si);
9590 def_builtin_const ("__builtin_vis_fcmpne16", CODE_FOR_fcmpne16si_vis,
9591 si_ftype_v4hi_v4hi);
9592 def_builtin_const ("__builtin_vis_fcmpne32", CODE_FOR_fcmpne32si_vis,
9593 si_ftype_v2si_v2si);
9594 def_builtin_const ("__builtin_vis_fcmpgt16", CODE_FOR_fcmpgt16si_vis,
9595 si_ftype_v4hi_v4hi);
9596 def_builtin_const ("__builtin_vis_fcmpgt32", CODE_FOR_fcmpgt32si_vis,
9597 si_ftype_v2si_v2si);
9598 def_builtin_const ("__builtin_vis_fcmpeq16", CODE_FOR_fcmpeq16si_vis,
9599 si_ftype_v4hi_v4hi);
9600 def_builtin_const ("__builtin_vis_fcmpeq32", CODE_FOR_fcmpeq32si_vis,
9601 si_ftype_v2si_v2si);
9604 /* Addition and subtraction. */
9605 def_builtin_const ("__builtin_vis_fpadd16", CODE_FOR_addv4hi3,
9606 v4hi_ftype_v4hi_v4hi);
9607 def_builtin_const ("__builtin_vis_fpadd16s", CODE_FOR_addv2hi3,
9608 v2hi_ftype_v2hi_v2hi);
9609 def_builtin_const ("__builtin_vis_fpadd32", CODE_FOR_addv2si3,
9610 v2si_ftype_v2si_v2si);
9611 def_builtin_const ("__builtin_vis_fpadd32s", CODE_FOR_addv1si3,
9612 v1si_ftype_v1si_v1si);
9613 def_builtin_const ("__builtin_vis_fpsub16", CODE_FOR_subv4hi3,
9614 v4hi_ftype_v4hi_v4hi);
9615 def_builtin_const ("__builtin_vis_fpsub16s", CODE_FOR_subv2hi3,
9616 v2hi_ftype_v2hi_v2hi);
9617 def_builtin_const ("__builtin_vis_fpsub32", CODE_FOR_subv2si3,
9618 v2si_ftype_v2si_v2si);
9619 def_builtin_const ("__builtin_vis_fpsub32s", CODE_FOR_subv1si3,
9620 v1si_ftype_v1si_v1si);
9622 /* Three-dimensional array addressing. */
9625 def_builtin_const ("__builtin_vis_array8", CODE_FOR_array8di_vis,
9627 def_builtin_const ("__builtin_vis_array16", CODE_FOR_array16di_vis,
9629 def_builtin_const ("__builtin_vis_array32", CODE_FOR_array32di_vis,
9634 def_builtin_const ("__builtin_vis_array8", CODE_FOR_array8si_vis,
9636 def_builtin_const ("__builtin_vis_array16", CODE_FOR_array16si_vis,
9638 def_builtin_const ("__builtin_vis_array32", CODE_FOR_array32si_vis,
9644 /* Byte mask and shuffle */
9646 def_builtin ("__builtin_vis_bmask", CODE_FOR_bmaskdi_vis,
9649 def_builtin ("__builtin_vis_bmask", CODE_FOR_bmasksi_vis,
9651 def_builtin ("__builtin_vis_bshufflev4hi", CODE_FOR_bshufflev4hi_vis,
9652 v4hi_ftype_v4hi_v4hi);
9653 def_builtin ("__builtin_vis_bshufflev8qi", CODE_FOR_bshufflev8qi_vis,
9654 v8qi_ftype_v8qi_v8qi);
9655 def_builtin ("__builtin_vis_bshufflev2si", CODE_FOR_bshufflev2si_vis,
9656 v2si_ftype_v2si_v2si);
9657 def_builtin ("__builtin_vis_bshuffledi", CODE_FOR_bshufflev1di_vis,
9665 def_builtin ("__builtin_vis_cmask8", CODE_FOR_cmask8di_vis,
9667 def_builtin ("__builtin_vis_cmask16", CODE_FOR_cmask16di_vis,
9669 def_builtin ("__builtin_vis_cmask32", CODE_FOR_cmask32di_vis,
9674 def_builtin ("__builtin_vis_cmask8", CODE_FOR_cmask8si_vis,
9676 def_builtin ("__builtin_vis_cmask16", CODE_FOR_cmask16si_vis,
9678 def_builtin ("__builtin_vis_cmask32", CODE_FOR_cmask32si_vis,
9682 def_builtin_const ("__builtin_vis_fchksm16", CODE_FOR_fchksm16_vis,
9683 v4hi_ftype_v4hi_v4hi);
9685 def_builtin_const ("__builtin_vis_fsll16", CODE_FOR_vashlv4hi3,
9686 v4hi_ftype_v4hi_v4hi);
9687 def_builtin_const ("__builtin_vis_fslas16", CODE_FOR_vssashlv4hi3,
9688 v4hi_ftype_v4hi_v4hi);
9689 def_builtin_const ("__builtin_vis_fsrl16", CODE_FOR_vlshrv4hi3,
9690 v4hi_ftype_v4hi_v4hi);
9691 def_builtin_const ("__builtin_vis_fsra16", CODE_FOR_vashrv4hi3,
9692 v4hi_ftype_v4hi_v4hi);
9693 def_builtin_const ("__builtin_vis_fsll32", CODE_FOR_vashlv2si3,
9694 v2si_ftype_v2si_v2si);
9695 def_builtin_const ("__builtin_vis_fslas32", CODE_FOR_vssashlv2si3,
9696 v2si_ftype_v2si_v2si);
9697 def_builtin_const ("__builtin_vis_fsrl32", CODE_FOR_vlshrv2si3,
9698 v2si_ftype_v2si_v2si);
9699 def_builtin_const ("__builtin_vis_fsra32", CODE_FOR_vashrv2si3,
9700 v2si_ftype_v2si_v2si);
9703 def_builtin_const ("__builtin_vis_pdistn", CODE_FOR_pdistndi_vis,
9704 di_ftype_v8qi_v8qi);
9706 def_builtin_const ("__builtin_vis_pdistn", CODE_FOR_pdistnsi_vis,
9707 si_ftype_v8qi_v8qi);
9709 def_builtin_const ("__builtin_vis_fmean16", CODE_FOR_fmean16_vis,
9710 v4hi_ftype_v4hi_v4hi);
9711 def_builtin_const ("__builtin_vis_fpadd64", CODE_FOR_fpadd64_vis,
9713 def_builtin_const ("__builtin_vis_fpsub64", CODE_FOR_fpsub64_vis,
9716 def_builtin_const ("__builtin_vis_fpadds16", CODE_FOR_ssaddv4hi3,
9717 v4hi_ftype_v4hi_v4hi);
9718 def_builtin_const ("__builtin_vis_fpadds16s", CODE_FOR_ssaddv2hi3,
9719 v2hi_ftype_v2hi_v2hi);
9720 def_builtin_const ("__builtin_vis_fpsubs16", CODE_FOR_sssubv4hi3,
9721 v4hi_ftype_v4hi_v4hi);
9722 def_builtin_const ("__builtin_vis_fpsubs16s", CODE_FOR_sssubv2hi3,
9723 v2hi_ftype_v2hi_v2hi);
9724 def_builtin_const ("__builtin_vis_fpadds32", CODE_FOR_ssaddv2si3,
9725 v2si_ftype_v2si_v2si);
9726 def_builtin_const ("__builtin_vis_fpadds32s", CODE_FOR_ssaddv1si3,
9727 v1si_ftype_v1si_v1si);
9728 def_builtin_const ("__builtin_vis_fpsubs32", CODE_FOR_sssubv2si3,
9729 v2si_ftype_v2si_v2si);
9730 def_builtin_const ("__builtin_vis_fpsubs32s", CODE_FOR_sssubv1si3,
9731 v1si_ftype_v1si_v1si);
9735 def_builtin_const ("__builtin_vis_fucmple8", CODE_FOR_fucmple8di_vis,
9736 di_ftype_v8qi_v8qi);
9737 def_builtin_const ("__builtin_vis_fucmpne8", CODE_FOR_fucmpne8di_vis,
9738 di_ftype_v8qi_v8qi);
9739 def_builtin_const ("__builtin_vis_fucmpgt8", CODE_FOR_fucmpgt8di_vis,
9740 di_ftype_v8qi_v8qi);
9741 def_builtin_const ("__builtin_vis_fucmpeq8", CODE_FOR_fucmpeq8di_vis,
9742 di_ftype_v8qi_v8qi);
9746 def_builtin_const ("__builtin_vis_fucmple8", CODE_FOR_fucmple8si_vis,
9747 si_ftype_v8qi_v8qi);
9748 def_builtin_const ("__builtin_vis_fucmpne8", CODE_FOR_fucmpne8si_vis,
9749 si_ftype_v8qi_v8qi);
9750 def_builtin_const ("__builtin_vis_fucmpgt8", CODE_FOR_fucmpgt8si_vis,
9751 si_ftype_v8qi_v8qi);
9752 def_builtin_const ("__builtin_vis_fucmpeq8", CODE_FOR_fucmpeq8si_vis,
9753 si_ftype_v8qi_v8qi);
9756 def_builtin_const ("__builtin_vis_fhadds", CODE_FOR_fhaddsf_vis,
9758 def_builtin_const ("__builtin_vis_fhaddd", CODE_FOR_fhadddf_vis,
9760 def_builtin_const ("__builtin_vis_fhsubs", CODE_FOR_fhsubsf_vis,
9762 def_builtin_const ("__builtin_vis_fhsubd", CODE_FOR_fhsubdf_vis,
9764 def_builtin_const ("__builtin_vis_fnhadds", CODE_FOR_fnhaddsf_vis,
9766 def_builtin_const ("__builtin_vis_fnhaddd", CODE_FOR_fnhadddf_vis,
9769 def_builtin_const ("__builtin_vis_umulxhi", CODE_FOR_umulxhi_vis,
9771 def_builtin_const ("__builtin_vis_xmulx", CODE_FOR_xmulx_vis,
9773 def_builtin_const ("__builtin_vis_xmulxhi", CODE_FOR_xmulxhi_vis,
9778 /* Handle TARGET_EXPAND_BUILTIN target hook.
9779 Expand builtin functions for sparc intrinsics. */
9782 sparc_expand_builtin (tree exp, rtx target,
9783 rtx subtarget ATTRIBUTE_UNUSED,
9784 enum machine_mode tmode ATTRIBUTE_UNUSED,
9785 int ignore ATTRIBUTE_UNUSED)
9788 call_expr_arg_iterator iter;
9789 tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0);
9790 unsigned int icode = DECL_FUNCTION_CODE (fndecl);
9795 nonvoid = TREE_TYPE (TREE_TYPE (fndecl)) != void_type_node;
9799 enum machine_mode tmode = insn_data[icode].operand[0].mode;
9801 || GET_MODE (target) != tmode
9802 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
9803 op[0] = gen_reg_rtx (tmode);
9807 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
9809 const struct insn_operand_data *insn_op;
9812 if (arg == error_mark_node)
9816 idx = arg_count - !nonvoid;
9817 insn_op = &insn_data[icode].operand[idx];
9818 op[arg_count] = expand_normal (arg);
9820 if (insn_op->mode == V1DImode
9821 && GET_MODE (op[arg_count]) == DImode)
9822 op[arg_count] = gen_lowpart (V1DImode, op[arg_count]);
9823 else if (insn_op->mode == V1SImode
9824 && GET_MODE (op[arg_count]) == SImode)
9825 op[arg_count] = gen_lowpart (V1SImode, op[arg_count]);
9827 if (! (*insn_data[icode].operand[idx].predicate) (op[arg_count],
9829 op[arg_count] = copy_to_mode_reg (insn_op->mode, op[arg_count]);
9835 pat = GEN_FCN (icode) (op[0]);
9839 pat = GEN_FCN (icode) (op[0], op[1]);
9841 pat = GEN_FCN (icode) (op[1]);
9844 pat = GEN_FCN (icode) (op[0], op[1], op[2]);
9847 pat = GEN_FCN (icode) (op[0], op[1], op[2], op[3]);
9865 sparc_vis_mul8x16 (int e8, int e16)
9867 return (e8 * e16 + 128) / 256;
9870 /* Multiply the vector elements in ELTS0 to the elements in ELTS1 as specified
9871 by FNCODE. All of the elements in ELTS0 and ELTS1 lists must be integer
9872 constants. A tree list with the results of the multiplications is returned,
9873 and each element in the list is of INNER_TYPE. */
9876 sparc_handle_vis_mul8x16 (int fncode, tree inner_type, tree elts0, tree elts1)
9878 tree n_elts = NULL_TREE;
9883 case CODE_FOR_fmul8x16_vis:
9884 for (; elts0 && elts1;
9885 elts0 = TREE_CHAIN (elts0), elts1 = TREE_CHAIN (elts1))
9888 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (TREE_VALUE (elts0)),
9889 TREE_INT_CST_LOW (TREE_VALUE (elts1)));
9890 n_elts = tree_cons (NULL_TREE,
9891 build_int_cst (inner_type, val),
9896 case CODE_FOR_fmul8x16au_vis:
9897 scale = TREE_INT_CST_LOW (TREE_VALUE (elts1));
9899 for (; elts0; elts0 = TREE_CHAIN (elts0))
9902 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (TREE_VALUE (elts0)),
9904 n_elts = tree_cons (NULL_TREE,
9905 build_int_cst (inner_type, val),
9910 case CODE_FOR_fmul8x16al_vis:
9911 scale = TREE_INT_CST_LOW (TREE_VALUE (TREE_CHAIN (elts1)));
9913 for (; elts0; elts0 = TREE_CHAIN (elts0))
9916 = sparc_vis_mul8x16 (TREE_INT_CST_LOW (TREE_VALUE (elts0)),
9918 n_elts = tree_cons (NULL_TREE,
9919 build_int_cst (inner_type, val),
9928 return nreverse (n_elts);
9931 /* Handle TARGET_FOLD_BUILTIN target hook.
9932 Fold builtin functions for SPARC intrinsics. If IGNORE is true the
9933 result of the function call is ignored. NULL_TREE is returned if the
9934 function could not be folded. */
9937 sparc_fold_builtin (tree fndecl, int n_args ATTRIBUTE_UNUSED,
9938 tree *args, bool ignore)
9940 tree arg0, arg1, arg2;
9941 tree rtype = TREE_TYPE (TREE_TYPE (fndecl));
9942 enum insn_code icode = (enum insn_code) DECL_FUNCTION_CODE (fndecl);
9946 /* Note that a switch statement instead of the sequence of tests would
9947 be incorrect as many of the CODE_FOR values could be CODE_FOR_nothing
9948 and that would yield multiple alternatives with identical values. */
9949 if (icode == CODE_FOR_alignaddrsi_vis
9950 || icode == CODE_FOR_alignaddrdi_vis
9951 || icode == CODE_FOR_wrgsr_vis
9952 || icode == CODE_FOR_bmasksi_vis
9953 || icode == CODE_FOR_bmaskdi_vis
9954 || icode == CODE_FOR_cmask8si_vis
9955 || icode == CODE_FOR_cmask8di_vis
9956 || icode == CODE_FOR_cmask16si_vis
9957 || icode == CODE_FOR_cmask16di_vis
9958 || icode == CODE_FOR_cmask32si_vis
9959 || icode == CODE_FOR_cmask32di_vis)
9962 return build_zero_cst (rtype);
9967 case CODE_FOR_fexpand_vis:
9971 if (TREE_CODE (arg0) == VECTOR_CST)
9973 tree inner_type = TREE_TYPE (rtype);
9974 tree elts = TREE_VECTOR_CST_ELTS (arg0);
9975 tree n_elts = NULL_TREE;
9977 for (; elts; elts = TREE_CHAIN (elts))
9979 unsigned int val = TREE_INT_CST_LOW (TREE_VALUE (elts)) << 4;
9980 n_elts = tree_cons (NULL_TREE,
9981 build_int_cst (inner_type, val),
9984 return build_vector (rtype, nreverse (n_elts));
9988 case CODE_FOR_fmul8x16_vis:
9989 case CODE_FOR_fmul8x16au_vis:
9990 case CODE_FOR_fmul8x16al_vis:
9996 if (TREE_CODE (arg0) == VECTOR_CST && TREE_CODE (arg1) == VECTOR_CST)
9998 tree inner_type = TREE_TYPE (rtype);
9999 tree elts0 = TREE_VECTOR_CST_ELTS (arg0);
10000 tree elts1 = TREE_VECTOR_CST_ELTS (arg1);
10001 tree n_elts = sparc_handle_vis_mul8x16 (icode, inner_type, elts0,
10004 return build_vector (rtype, n_elts);
10008 case CODE_FOR_fpmerge_vis:
10014 if (TREE_CODE (arg0) == VECTOR_CST && TREE_CODE (arg1) == VECTOR_CST)
10016 tree elts0 = TREE_VECTOR_CST_ELTS (arg0);
10017 tree elts1 = TREE_VECTOR_CST_ELTS (arg1);
10018 tree n_elts = NULL_TREE;
10020 for (; elts0 && elts1;
10021 elts0 = TREE_CHAIN (elts0), elts1 = TREE_CHAIN (elts1))
10023 n_elts = tree_cons (NULL_TREE, TREE_VALUE (elts0), n_elts);
10024 n_elts = tree_cons (NULL_TREE, TREE_VALUE (elts1), n_elts);
10027 return build_vector (rtype, nreverse (n_elts));
10031 case CODE_FOR_pdist_vis:
10039 if (TREE_CODE (arg0) == VECTOR_CST
10040 && TREE_CODE (arg1) == VECTOR_CST
10041 && TREE_CODE (arg2) == INTEGER_CST)
10044 unsigned HOST_WIDE_INT low = TREE_INT_CST_LOW (arg2);
10045 HOST_WIDE_INT high = TREE_INT_CST_HIGH (arg2);
10046 tree elts0 = TREE_VECTOR_CST_ELTS (arg0);
10047 tree elts1 = TREE_VECTOR_CST_ELTS (arg1);
10049 for (; elts0 && elts1;
10050 elts0 = TREE_CHAIN (elts0), elts1 = TREE_CHAIN (elts1))
10052 unsigned HOST_WIDE_INT
10053 low0 = TREE_INT_CST_LOW (TREE_VALUE (elts0)),
10054 low1 = TREE_INT_CST_LOW (TREE_VALUE (elts1));
10055 HOST_WIDE_INT high0 = TREE_INT_CST_HIGH (TREE_VALUE (elts0));
10056 HOST_WIDE_INT high1 = TREE_INT_CST_HIGH (TREE_VALUE (elts1));
10058 unsigned HOST_WIDE_INT l;
10061 overflow |= neg_double (low1, high1, &l, &h);
10062 overflow |= add_double (low0, high0, l, h, &l, &h);
10064 overflow |= neg_double (l, h, &l, &h);
10066 overflow |= add_double (low, high, l, h, &low, &high);
10069 gcc_assert (overflow == 0);
10071 return build_int_cst_wide (rtype, low, high);
10081 /* ??? This duplicates information provided to the compiler by the
10082 ??? scheduler description. Some day, teach genautomata to output
10083 ??? the latencies and then CSE will just use that. */
10086 sparc_rtx_costs (rtx x, int code, int outer_code, int opno ATTRIBUTE_UNUSED,
10087 int *total, bool speed ATTRIBUTE_UNUSED)
10089 enum machine_mode mode = GET_MODE (x);
10090 bool float_mode_p = FLOAT_MODE_P (mode);
10095 if (INTVAL (x) < 0x1000 && INTVAL (x) >= -0x1000)
10113 if (GET_MODE (x) == VOIDmode
10114 && ((CONST_DOUBLE_HIGH (x) == 0
10115 && CONST_DOUBLE_LOW (x) < 0x1000)
10116 || (CONST_DOUBLE_HIGH (x) == -1
10117 && CONST_DOUBLE_LOW (x) < 0
10118 && CONST_DOUBLE_LOW (x) >= -0x1000)))
10125 /* If outer-code was a sign or zero extension, a cost
10126 of COSTS_N_INSNS (1) was already added in. This is
10127 why we are subtracting it back out. */
10128 if (outer_code == ZERO_EXTEND)
10130 *total = sparc_costs->int_zload - COSTS_N_INSNS (1);
10132 else if (outer_code == SIGN_EXTEND)
10134 *total = sparc_costs->int_sload - COSTS_N_INSNS (1);
10136 else if (float_mode_p)
10138 *total = sparc_costs->float_load;
10142 *total = sparc_costs->int_load;
10150 *total = sparc_costs->float_plusminus;
10152 *total = COSTS_N_INSNS (1);
10159 gcc_assert (float_mode_p);
10160 *total = sparc_costs->float_mul;
10163 if (GET_CODE (sub) == NEG)
10164 sub = XEXP (sub, 0);
10165 *total += rtx_cost (sub, FMA, 0, speed);
10168 if (GET_CODE (sub) == NEG)
10169 sub = XEXP (sub, 0);
10170 *total += rtx_cost (sub, FMA, 2, speed);
10176 *total = sparc_costs->float_mul;
10177 else if (! TARGET_HARD_MUL)
10178 *total = COSTS_N_INSNS (25);
10184 if (sparc_costs->int_mul_bit_factor)
10188 if (GET_CODE (XEXP (x, 1)) == CONST_INT)
10190 unsigned HOST_WIDE_INT value = INTVAL (XEXP (x, 1));
10191 for (nbits = 0; value != 0; value &= value - 1)
10194 else if (GET_CODE (XEXP (x, 1)) == CONST_DOUBLE
10195 && GET_MODE (XEXP (x, 1)) == VOIDmode)
10197 rtx x1 = XEXP (x, 1);
10198 unsigned HOST_WIDE_INT value1 = CONST_DOUBLE_LOW (x1);
10199 unsigned HOST_WIDE_INT value2 = CONST_DOUBLE_HIGH (x1);
10201 for (nbits = 0; value1 != 0; value1 &= value1 - 1)
10203 for (; value2 != 0; value2 &= value2 - 1)
10211 bit_cost = (nbits - 3) / sparc_costs->int_mul_bit_factor;
10212 bit_cost = COSTS_N_INSNS (bit_cost);
10215 if (mode == DImode)
10216 *total = sparc_costs->int_mulX + bit_cost;
10218 *total = sparc_costs->int_mul + bit_cost;
10225 *total = COSTS_N_INSNS (1) + sparc_costs->shift_penalty;
10234 if (mode == DFmode)
10235 *total = sparc_costs->float_div_df;
10237 *total = sparc_costs->float_div_sf;
10241 if (mode == DImode)
10242 *total = sparc_costs->int_divX;
10244 *total = sparc_costs->int_div;
10249 if (! float_mode_p)
10251 *total = COSTS_N_INSNS (1);
10258 case UNSIGNED_FLOAT:
10262 case FLOAT_TRUNCATE:
10263 *total = sparc_costs->float_move;
10267 if (mode == DFmode)
10268 *total = sparc_costs->float_sqrt_df;
10270 *total = sparc_costs->float_sqrt_sf;
10275 *total = sparc_costs->float_cmp;
10277 *total = COSTS_N_INSNS (1);
10282 *total = sparc_costs->float_cmove;
10284 *total = sparc_costs->int_cmove;
10288 /* Handle the NAND vector patterns. */
10289 if (sparc_vector_mode_supported_p (GET_MODE (x))
10290 && GET_CODE (XEXP (x, 0)) == NOT
10291 && GET_CODE (XEXP (x, 1)) == NOT)
10293 *total = COSTS_N_INSNS (1);
10304 /* Return true if CLASS is either GENERAL_REGS or I64_REGS. */
10307 general_or_i64_p (reg_class_t rclass)
10309 return (rclass == GENERAL_REGS || rclass == I64_REGS);
10312 /* Implement TARGET_REGISTER_MOVE_COST. */
10315 sparc_register_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED,
10316 reg_class_t from, reg_class_t to)
10318 bool need_memory = false;
10320 if (from == FPCC_REGS || to == FPCC_REGS)
10321 need_memory = true;
10322 else if ((FP_REG_CLASS_P (from) && general_or_i64_p (to))
10323 || (general_or_i64_p (from) && FP_REG_CLASS_P (to)))
10327 int size = GET_MODE_SIZE (mode);
10328 if (size == 8 || size == 4)
10330 if (! TARGET_ARCH32 || size == 4)
10336 need_memory = true;
10341 if (sparc_cpu == PROCESSOR_ULTRASPARC
10342 || sparc_cpu == PROCESSOR_ULTRASPARC3
10343 || sparc_cpu == PROCESSOR_NIAGARA
10344 || sparc_cpu == PROCESSOR_NIAGARA2
10345 || sparc_cpu == PROCESSOR_NIAGARA3
10346 || sparc_cpu == PROCESSOR_NIAGARA4)
10355 /* Emit the sequence of insns SEQ while preserving the registers REG and REG2.
10356 This is achieved by means of a manual dynamic stack space allocation in
10357 the current frame. We make the assumption that SEQ doesn't contain any
10358 function calls, with the possible exception of calls to the GOT helper. */
10361 emit_and_preserve (rtx seq, rtx reg, rtx reg2)
10363 /* We must preserve the lowest 16 words for the register save area. */
10364 HOST_WIDE_INT offset = 16*UNITS_PER_WORD;
10365 /* We really need only 2 words of fresh stack space. */
10366 HOST_WIDE_INT size = SPARC_STACK_ALIGN (offset + 2*UNITS_PER_WORD);
10369 = gen_rtx_MEM (word_mode, plus_constant (stack_pointer_rtx,
10370 SPARC_STACK_BIAS + offset));
10372 emit_insn (gen_stack_pointer_dec (GEN_INT (size)));
10373 emit_insn (gen_rtx_SET (VOIDmode, slot, reg));
10375 emit_insn (gen_rtx_SET (VOIDmode,
10376 adjust_address (slot, word_mode, UNITS_PER_WORD),
10380 emit_insn (gen_rtx_SET (VOIDmode,
10382 adjust_address (slot, word_mode, UNITS_PER_WORD)));
10383 emit_insn (gen_rtx_SET (VOIDmode, reg, slot));
10384 emit_insn (gen_stack_pointer_inc (GEN_INT (size)));
10387 /* Output the assembler code for a thunk function. THUNK_DECL is the
10388 declaration for the thunk function itself, FUNCTION is the decl for
10389 the target function. DELTA is an immediate constant offset to be
10390 added to THIS. If VCALL_OFFSET is nonzero, the word at address
10391 (*THIS + VCALL_OFFSET) should be additionally added to THIS. */
10394 sparc_output_mi_thunk (FILE *file, tree thunk_fndecl ATTRIBUTE_UNUSED,
10395 HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset,
10398 rtx this_rtx, insn, funexp;
10399 unsigned int int_arg_first;
10401 reload_completed = 1;
10402 epilogue_completed = 1;
10404 emit_note (NOTE_INSN_PROLOGUE_END);
10408 sparc_leaf_function_p = 1;
10410 int_arg_first = SPARC_OUTGOING_INT_ARG_FIRST;
10412 else if (flag_delayed_branch)
10414 /* We will emit a regular sibcall below, so we need to instruct
10415 output_sibcall that we are in a leaf function. */
10416 sparc_leaf_function_p = current_function_uses_only_leaf_regs = 1;
10418 /* This will cause final.c to invoke leaf_renumber_regs so we
10419 must behave as if we were in a not-yet-leafified function. */
10420 int_arg_first = SPARC_INCOMING_INT_ARG_FIRST;
10424 /* We will emit the sibcall manually below, so we will need to
10425 manually spill non-leaf registers. */
10426 sparc_leaf_function_p = current_function_uses_only_leaf_regs = 0;
10428 /* We really are in a leaf function. */
10429 int_arg_first = SPARC_OUTGOING_INT_ARG_FIRST;
10432 /* Find the "this" pointer. Normally in %o0, but in ARCH64 if the function
10433 returns a structure, the structure return pointer is there instead. */
10435 && aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function))
10436 this_rtx = gen_rtx_REG (Pmode, int_arg_first + 1);
10438 this_rtx = gen_rtx_REG (Pmode, int_arg_first);
10440 /* Add DELTA. When possible use a plain add, otherwise load it into
10441 a register first. */
10444 rtx delta_rtx = GEN_INT (delta);
10446 if (! SPARC_SIMM13_P (delta))
10448 rtx scratch = gen_rtx_REG (Pmode, 1);
10449 emit_move_insn (scratch, delta_rtx);
10450 delta_rtx = scratch;
10453 /* THIS_RTX += DELTA. */
10454 emit_insn (gen_add2_insn (this_rtx, delta_rtx));
10457 /* Add the word at address (*THIS_RTX + VCALL_OFFSET). */
10460 rtx vcall_offset_rtx = GEN_INT (vcall_offset);
10461 rtx scratch = gen_rtx_REG (Pmode, 1);
10463 gcc_assert (vcall_offset < 0);
10465 /* SCRATCH = *THIS_RTX. */
10466 emit_move_insn (scratch, gen_rtx_MEM (Pmode, this_rtx));
10468 /* Prepare for adding VCALL_OFFSET. The difficulty is that we
10469 may not have any available scratch register at this point. */
10470 if (SPARC_SIMM13_P (vcall_offset))
10472 /* This is the case if ARCH64 (unless -ffixed-g5 is passed). */
10473 else if (! fixed_regs[5]
10474 /* The below sequence is made up of at least 2 insns,
10475 while the default method may need only one. */
10476 && vcall_offset < -8192)
10478 rtx scratch2 = gen_rtx_REG (Pmode, 5);
10479 emit_move_insn (scratch2, vcall_offset_rtx);
10480 vcall_offset_rtx = scratch2;
10484 rtx increment = GEN_INT (-4096);
10486 /* VCALL_OFFSET is a negative number whose typical range can be
10487 estimated as -32768..0 in 32-bit mode. In almost all cases
10488 it is therefore cheaper to emit multiple add insns than
10489 spilling and loading the constant into a register (at least
10491 while (! SPARC_SIMM13_P (vcall_offset))
10493 emit_insn (gen_add2_insn (scratch, increment));
10494 vcall_offset += 4096;
10496 vcall_offset_rtx = GEN_INT (vcall_offset); /* cannot be 0 */
10499 /* SCRATCH = *(*THIS_RTX + VCALL_OFFSET). */
10500 emit_move_insn (scratch, gen_rtx_MEM (Pmode,
10501 gen_rtx_PLUS (Pmode,
10503 vcall_offset_rtx)));
10505 /* THIS_RTX += *(*THIS_RTX + VCALL_OFFSET). */
10506 emit_insn (gen_add2_insn (this_rtx, scratch));
10509 /* Generate a tail call to the target function. */
10510 if (! TREE_USED (function))
10512 assemble_external (function);
10513 TREE_USED (function) = 1;
10515 funexp = XEXP (DECL_RTL (function), 0);
10517 if (flag_delayed_branch)
10519 funexp = gen_rtx_MEM (FUNCTION_MODE, funexp);
10520 insn = emit_call_insn (gen_sibcall (funexp));
10521 SIBLING_CALL_P (insn) = 1;
10525 /* The hoops we have to jump through in order to generate a sibcall
10526 without using delay slots... */
10527 rtx spill_reg, seq, scratch = gen_rtx_REG (Pmode, 1);
10531 spill_reg = gen_rtx_REG (word_mode, 15); /* %o7 */
10533 load_got_register (); /* clobbers %o7 */
10534 scratch = sparc_legitimize_pic_address (funexp, scratch);
10535 seq = get_insns ();
10537 emit_and_preserve (seq, spill_reg, pic_offset_table_rtx);
10539 else if (TARGET_ARCH32)
10541 emit_insn (gen_rtx_SET (VOIDmode,
10543 gen_rtx_HIGH (SImode, funexp)));
10544 emit_insn (gen_rtx_SET (VOIDmode,
10546 gen_rtx_LO_SUM (SImode, scratch, funexp)));
10548 else /* TARGET_ARCH64 */
10550 switch (sparc_cmodel)
10554 /* The destination can serve as a temporary. */
10555 sparc_emit_set_symbolic_const64 (scratch, funexp, scratch);
10560 /* The destination cannot serve as a temporary. */
10561 spill_reg = gen_rtx_REG (DImode, 15); /* %o7 */
10563 sparc_emit_set_symbolic_const64 (scratch, funexp, spill_reg);
10564 seq = get_insns ();
10566 emit_and_preserve (seq, spill_reg, 0);
10570 gcc_unreachable ();
10574 emit_jump_insn (gen_indirect_jump (scratch));
10579 /* Run just enough of rest_of_compilation to get the insns emitted.
10580 There's not really enough bulk here to make other passes such as
10581 instruction scheduling worth while. Note that use_thunk calls
10582 assemble_start_function and assemble_end_function. */
10583 insn = get_insns ();
10584 insn_locators_alloc ();
10585 shorten_branches (insn);
10586 final_start_function (insn, file, 1);
10587 final (insn, file, 1);
10588 final_end_function ();
10590 reload_completed = 0;
10591 epilogue_completed = 0;
10594 /* Return true if sparc_output_mi_thunk would be able to output the
10595 assembler code for the thunk function specified by the arguments
10596 it is passed, and false otherwise. */
10598 sparc_can_output_mi_thunk (const_tree thunk_fndecl ATTRIBUTE_UNUSED,
10599 HOST_WIDE_INT delta ATTRIBUTE_UNUSED,
10600 HOST_WIDE_INT vcall_offset,
10601 const_tree function ATTRIBUTE_UNUSED)
10603 /* Bound the loop used in the default method above. */
10604 return (vcall_offset >= -32768 || ! fixed_regs[5]);
10607 /* We use the machine specific reorg pass to enable workarounds for errata. */
10614 /* The only erratum we handle for now is that of the AT697F processor. */
10615 if (!sparc_fix_at697f)
10618 /* We need to have the (essentially) final form of the insn stream in order
10619 to properly detect the various hazards. Run delay slot scheduling. */
10620 if (optimize > 0 && flag_delayed_branch)
10621 dbr_schedule (get_insns ());
10623 /* Now look for specific patterns in the insn stream. */
10624 for (insn = get_insns (); insn; insn = next)
10626 bool insert_nop = false;
10629 /* Look for a single-word load into an odd-numbered FP register. */
10630 if (NONJUMP_INSN_P (insn)
10631 && (set = single_set (insn)) != NULL_RTX
10632 && GET_MODE_SIZE (GET_MODE (SET_SRC (set))) == 4
10633 && MEM_P (SET_SRC (set))
10634 && REG_P (SET_DEST (set))
10635 && REGNO (SET_DEST (set)) > 31
10636 && REGNO (SET_DEST (set)) % 2 != 0)
10638 /* The wrong dependency is on the enclosing double register. */
10639 unsigned int x = REGNO (SET_DEST (set)) - 1;
10640 unsigned int src1, src2, dest;
10643 /* If the insn has a delay slot, then it cannot be problematic. */
10644 next = next_active_insn (insn);
10645 if (NONJUMP_INSN_P (next) && GET_CODE (PATTERN (next)) == SEQUENCE)
10649 extract_insn (next);
10650 code = INSN_CODE (next);
10655 case CODE_FOR_adddf3:
10656 case CODE_FOR_subdf3:
10657 case CODE_FOR_muldf3:
10658 case CODE_FOR_divdf3:
10659 dest = REGNO (recog_data.operand[0]);
10660 src1 = REGNO (recog_data.operand[1]);
10661 src2 = REGNO (recog_data.operand[2]);
10665 ld [address], %fx+1
10666 FPOPd %f{x,y}, %f{y,x}, %f{x,y} */
10667 if ((src1 == x || src2 == x)
10668 && (dest == src1 || dest == src2))
10674 ld [address], %fx+1
10675 FPOPd %fx, %fx, %fx */
10678 && (code == CODE_FOR_adddf3 || code == CODE_FOR_muldf3))
10683 case CODE_FOR_sqrtdf2:
10684 dest = REGNO (recog_data.operand[0]);
10685 src1 = REGNO (recog_data.operand[1]);
10687 ld [address], %fx+1
10689 if (src1 == x && dest == src1)
10698 next = NEXT_INSN (insn);
10701 emit_insn_after (gen_nop (), insn);
10705 /* How to allocate a 'struct machine_function'. */
10707 static struct machine_function *
10708 sparc_init_machine_status (void)
10710 return ggc_alloc_cleared_machine_function ();
10713 /* Locate some local-dynamic symbol still in use by this function
10714 so that we can print its name in local-dynamic base patterns. */
10716 static const char *
10717 get_some_local_dynamic_name (void)
10721 if (cfun->machine->some_ld_name)
10722 return cfun->machine->some_ld_name;
10724 for (insn = get_insns (); insn ; insn = NEXT_INSN (insn))
10726 && for_each_rtx (&PATTERN (insn), get_some_local_dynamic_name_1, 0))
10727 return cfun->machine->some_ld_name;
10729 gcc_unreachable ();
10733 get_some_local_dynamic_name_1 (rtx *px, void *data ATTRIBUTE_UNUSED)
10738 && GET_CODE (x) == SYMBOL_REF
10739 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_DYNAMIC)
10741 cfun->machine->some_ld_name = XSTR (x, 0);
10748 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
10749 We need to emit DTP-relative relocations. */
10752 sparc_output_dwarf_dtprel (FILE *file, int size, rtx x)
10757 fputs ("\t.word\t%r_tls_dtpoff32(", file);
10760 fputs ("\t.xword\t%r_tls_dtpoff64(", file);
10763 gcc_unreachable ();
10765 output_addr_const (file, x);
10769 /* Do whatever processing is required at the end of a file. */
10772 sparc_file_end (void)
10774 /* If we need to emit the special GOT helper function, do so now. */
10775 if (got_helper_rtx)
10777 const char *name = XSTR (got_helper_rtx, 0);
10778 const char *reg_name = reg_names[GLOBAL_OFFSET_TABLE_REGNUM];
10779 #ifdef DWARF2_UNWIND_INFO
10783 if (USE_HIDDEN_LINKONCE)
10785 tree decl = build_decl (BUILTINS_LOCATION, FUNCTION_DECL,
10786 get_identifier (name),
10787 build_function_type_list (void_type_node,
10789 DECL_RESULT (decl) = build_decl (BUILTINS_LOCATION, RESULT_DECL,
10790 NULL_TREE, void_type_node);
10791 TREE_STATIC (decl) = 1;
10792 make_decl_one_only (decl, DECL_ASSEMBLER_NAME (decl));
10793 DECL_VISIBILITY (decl) = VISIBILITY_HIDDEN;
10794 DECL_VISIBILITY_SPECIFIED (decl) = 1;
10795 resolve_unique_section (decl, 0, flag_function_sections);
10796 allocate_struct_function (decl, true);
10797 cfun->is_thunk = 1;
10798 current_function_decl = decl;
10799 init_varasm_status ();
10800 assemble_start_function (decl, name);
10804 const int align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT);
10805 switch_to_section (text_section);
10807 ASM_OUTPUT_ALIGN (asm_out_file, align);
10808 ASM_OUTPUT_LABEL (asm_out_file, name);
10811 #ifdef DWARF2_UNWIND_INFO
10812 do_cfi = dwarf2out_do_cfi_asm ();
10814 fprintf (asm_out_file, "\t.cfi_startproc\n");
10816 if (flag_delayed_branch)
10817 fprintf (asm_out_file, "\tjmp\t%%o7+8\n\t add\t%%o7, %s, %s\n",
10818 reg_name, reg_name);
10820 fprintf (asm_out_file, "\tadd\t%%o7, %s, %s\n\tjmp\t%%o7+8\n\t nop\n",
10821 reg_name, reg_name);
10822 #ifdef DWARF2_UNWIND_INFO
10824 fprintf (asm_out_file, "\t.cfi_endproc\n");
10828 if (NEED_INDICATE_EXEC_STACK)
10829 file_end_indicate_exec_stack ();
10831 #ifdef TARGET_SOLARIS
10832 solaris_file_end ();
10836 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
10837 /* Implement TARGET_MANGLE_TYPE. */
10839 static const char *
10840 sparc_mangle_type (const_tree type)
10843 && TYPE_MAIN_VARIANT (type) == long_double_type_node
10844 && TARGET_LONG_DOUBLE_128)
10847 /* For all other types, use normal C++ mangling. */
10852 /* Expand a membar instruction for various use cases. Both the LOAD_STORE
10853 and BEFORE_AFTER arguments of the form X_Y. They are two-bit masks where
10854 bit 0 indicates that X is true, and bit 1 indicates Y is true. */
10857 sparc_emit_membar_for_model (enum memmodel model,
10858 int load_store, int before_after)
10860 /* Bits for the MEMBAR mmask field. */
10861 const int LoadLoad = 1;
10862 const int StoreLoad = 2;
10863 const int LoadStore = 4;
10864 const int StoreStore = 8;
10868 if (before_after & 1)
10870 if (model == MEMMODEL_ACQUIRE
10871 || model == MEMMODEL_ACQ_REL
10872 || model == MEMMODEL_SEQ_CST)
10874 if (load_store & 1)
10875 mm |= LoadLoad | LoadStore;
10876 if (load_store & 2)
10877 mm |= StoreLoad | StoreStore;
10880 if (before_after & 2)
10882 if (model == MEMMODEL_RELEASE
10883 || model == MEMMODEL_ACQ_REL
10884 || model == MEMMODEL_SEQ_CST)
10886 if (load_store & 1)
10887 mm |= LoadLoad | StoreLoad;
10888 if (load_store & 2)
10889 mm |= LoadStore | StoreStore;
10893 /* For raw barriers (before+after), always emit a barrier.
10894 This will become a compile-time barrier if needed. */
10895 if (mm || before_after == 3)
10896 emit_insn (gen_membar (GEN_INT (mm)));
10899 /* Expand code to perform a 8 or 16-bit compare and swap by doing 32-bit
10900 compare and swap on the word containing the byte or half-word. */
10903 sparc_expand_compare_and_swap_12 (rtx result, rtx mem, rtx oldval, rtx newval)
10905 rtx addr1 = force_reg (Pmode, XEXP (mem, 0));
10906 rtx addr = gen_reg_rtx (Pmode);
10907 rtx off = gen_reg_rtx (SImode);
10908 rtx oldv = gen_reg_rtx (SImode);
10909 rtx newv = gen_reg_rtx (SImode);
10910 rtx oldvalue = gen_reg_rtx (SImode);
10911 rtx newvalue = gen_reg_rtx (SImode);
10912 rtx res = gen_reg_rtx (SImode);
10913 rtx resv = gen_reg_rtx (SImode);
10914 rtx memsi, val, mask, end_label, loop_label, cc;
10916 emit_insn (gen_rtx_SET (VOIDmode, addr,
10917 gen_rtx_AND (Pmode, addr1, GEN_INT (-4))));
10919 if (Pmode != SImode)
10920 addr1 = gen_lowpart (SImode, addr1);
10921 emit_insn (gen_rtx_SET (VOIDmode, off,
10922 gen_rtx_AND (SImode, addr1, GEN_INT (3))));
10924 memsi = gen_rtx_MEM (SImode, addr);
10925 set_mem_alias_set (memsi, ALIAS_SET_MEMORY_BARRIER);
10926 MEM_VOLATILE_P (memsi) = MEM_VOLATILE_P (mem);
10928 val = force_reg (SImode, memsi);
10930 emit_insn (gen_rtx_SET (VOIDmode, off,
10931 gen_rtx_XOR (SImode, off,
10932 GEN_INT (GET_MODE (mem) == QImode
10935 emit_insn (gen_rtx_SET (VOIDmode, off,
10936 gen_rtx_ASHIFT (SImode, off, GEN_INT (3))));
10938 if (GET_MODE (mem) == QImode)
10939 mask = force_reg (SImode, GEN_INT (0xff));
10941 mask = force_reg (SImode, GEN_INT (0xffff));
10943 emit_insn (gen_rtx_SET (VOIDmode, mask,
10944 gen_rtx_ASHIFT (SImode, mask, off)));
10946 emit_insn (gen_rtx_SET (VOIDmode, val,
10947 gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask),
10950 oldval = gen_lowpart (SImode, oldval);
10951 emit_insn (gen_rtx_SET (VOIDmode, oldv,
10952 gen_rtx_ASHIFT (SImode, oldval, off)));
10954 newval = gen_lowpart_common (SImode, newval);
10955 emit_insn (gen_rtx_SET (VOIDmode, newv,
10956 gen_rtx_ASHIFT (SImode, newval, off)));
10958 emit_insn (gen_rtx_SET (VOIDmode, oldv,
10959 gen_rtx_AND (SImode, oldv, mask)));
10961 emit_insn (gen_rtx_SET (VOIDmode, newv,
10962 gen_rtx_AND (SImode, newv, mask)));
10964 end_label = gen_label_rtx ();
10965 loop_label = gen_label_rtx ();
10966 emit_label (loop_label);
10968 emit_insn (gen_rtx_SET (VOIDmode, oldvalue,
10969 gen_rtx_IOR (SImode, oldv, val)));
10971 emit_insn (gen_rtx_SET (VOIDmode, newvalue,
10972 gen_rtx_IOR (SImode, newv, val)));
10974 emit_insn (gen_sync_compare_and_swapsi (res, memsi, oldvalue, newvalue));
10976 emit_cmp_and_jump_insns (res, oldvalue, EQ, NULL, SImode, 0, end_label);
10978 emit_insn (gen_rtx_SET (VOIDmode, resv,
10979 gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask),
10982 cc = gen_compare_reg_1 (NE, resv, val);
10983 emit_insn (gen_rtx_SET (VOIDmode, val, resv));
10985 /* Use cbranchcc4 to separate the compare and branch! */
10986 emit_jump_insn (gen_cbranchcc4 (gen_rtx_NE (VOIDmode, cc, const0_rtx),
10987 cc, const0_rtx, loop_label));
10989 emit_label (end_label);
10991 emit_insn (gen_rtx_SET (VOIDmode, res,
10992 gen_rtx_AND (SImode, res, mask)));
10994 emit_insn (gen_rtx_SET (VOIDmode, res,
10995 gen_rtx_LSHIFTRT (SImode, res, off)));
10997 emit_move_insn (result, gen_lowpart (GET_MODE (result), res));
11001 sparc_expand_vec_perm_bmask (enum machine_mode vmode, rtx sel)
11005 sel = gen_lowpart (DImode, sel);
11009 /* inp = xxxxxxxAxxxxxxxB */
11010 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (16),
11011 NULL_RTX, 1, OPTAB_DIRECT);
11012 /* t_1 = ....xxxxxxxAxxx. */
11013 sel = expand_simple_binop (SImode, AND, gen_lowpart (SImode, sel),
11014 GEN_INT (3), NULL_RTX, 1, OPTAB_DIRECT);
11015 t_1 = expand_simple_binop (SImode, AND, gen_lowpart (SImode, t_1),
11016 GEN_INT (0x30000), NULL_RTX, 1, OPTAB_DIRECT);
11017 /* sel = .......B */
11018 /* t_1 = ...A.... */
11019 sel = expand_simple_binop (SImode, IOR, sel, t_1, sel, 1, OPTAB_DIRECT);
11020 /* sel = ...A...B */
11021 sel = expand_mult (SImode, sel, GEN_INT (0x4444), sel, 1);
11022 /* sel = AAAABBBB * 4 */
11023 t_1 = force_reg (SImode, GEN_INT (0x01230123));
11024 /* sel = { A*4, A*4+1, A*4+2, ... } */
11028 /* inp = xxxAxxxBxxxCxxxD */
11029 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (8),
11030 NULL_RTX, 1, OPTAB_DIRECT);
11031 t_2 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (16),
11032 NULL_RTX, 1, OPTAB_DIRECT);
11033 t_3 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (24),
11034 NULL_RTX, 1, OPTAB_DIRECT);
11035 /* t_1 = ..xxxAxxxBxxxCxx */
11036 /* t_2 = ....xxxAxxxBxxxC */
11037 /* t_3 = ......xxxAxxxBxx */
11038 sel = expand_simple_binop (SImode, AND, gen_lowpart (SImode, sel),
11040 NULL_RTX, 1, OPTAB_DIRECT);
11041 t_1 = expand_simple_binop (SImode, AND, gen_lowpart (SImode, t_1),
11043 NULL_RTX, 1, OPTAB_DIRECT);
11044 t_2 = expand_simple_binop (SImode, AND, gen_lowpart (SImode, t_2),
11045 GEN_INT (0x070000),
11046 NULL_RTX, 1, OPTAB_DIRECT);
11047 t_3 = expand_simple_binop (SImode, AND, gen_lowpart (SImode, t_3),
11048 GEN_INT (0x07000000),
11049 NULL_RTX, 1, OPTAB_DIRECT);
11050 /* sel = .......D */
11051 /* t_1 = .....C.. */
11052 /* t_2 = ...B.... */
11053 /* t_3 = .A...... */
11054 sel = expand_simple_binop (SImode, IOR, sel, t_1, sel, 1, OPTAB_DIRECT);
11055 t_2 = expand_simple_binop (SImode, IOR, t_2, t_3, t_2, 1, OPTAB_DIRECT);
11056 sel = expand_simple_binop (SImode, IOR, sel, t_2, sel, 1, OPTAB_DIRECT);
11057 /* sel = .A.B.C.D */
11058 sel = expand_mult (SImode, sel, GEN_INT (0x22), sel, 1);
11059 /* sel = AABBCCDD * 2 */
11060 t_1 = force_reg (SImode, GEN_INT (0x01010101));
11061 /* sel = { A*2, A*2+1, B*2, B*2+1, ... } */
11065 /* input = xAxBxCxDxExFxGxH */
11066 sel = expand_simple_binop (DImode, AND, sel,
11067 GEN_INT ((HOST_WIDE_INT)0x0f0f0f0f << 32
11069 NULL_RTX, 1, OPTAB_DIRECT);
11070 /* sel = .A.B.C.D.E.F.G.H */
11071 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (4),
11072 NULL_RTX, 1, OPTAB_DIRECT);
11073 /* t_1 = ..A.B.C.D.E.F.G. */
11074 sel = expand_simple_binop (DImode, IOR, sel, t_1,
11075 NULL_RTX, 1, OPTAB_DIRECT);
11076 /* sel = .AABBCCDDEEFFGGH */
11077 sel = expand_simple_binop (DImode, AND, sel,
11078 GEN_INT ((HOST_WIDE_INT)0xff00ff << 32
11080 NULL_RTX, 1, OPTAB_DIRECT);
11081 /* sel = ..AB..CD..EF..GH */
11082 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (8),
11083 NULL_RTX, 1, OPTAB_DIRECT);
11084 /* t_1 = ....AB..CD..EF.. */
11085 sel = expand_simple_binop (DImode, IOR, sel, t_1,
11086 NULL_RTX, 1, OPTAB_DIRECT);
11087 /* sel = ..ABABCDCDEFEFGH */
11088 sel = expand_simple_binop (DImode, AND, sel,
11089 GEN_INT ((HOST_WIDE_INT)0xffff << 32 | 0xffff),
11090 NULL_RTX, 1, OPTAB_DIRECT);
11091 /* sel = ....ABCD....EFGH */
11092 t_1 = expand_simple_binop (DImode, LSHIFTRT, sel, GEN_INT (16),
11093 NULL_RTX, 1, OPTAB_DIRECT);
11094 /* t_1 = ........ABCD.... */
11095 sel = gen_lowpart (SImode, sel);
11096 t_1 = gen_lowpart (SImode, t_1);
11100 gcc_unreachable ();
11103 /* Always perform the final addition/merge within the bmask insn. */
11104 emit_insn (gen_bmasksi_vis (gen_reg_rtx (SImode), sel, t_1));
11107 /* Implement TARGET_FRAME_POINTER_REQUIRED. */
11110 sparc_frame_pointer_required (void)
11112 /* If the stack pointer is dynamically modified in the function, it cannot
11113 serve as the frame pointer. */
11114 if (cfun->calls_alloca)
11117 /* If the function receives nonlocal gotos, it needs to save the frame
11118 pointer in the nonlocal_goto_save_area object. */
11119 if (cfun->has_nonlocal_label)
11122 /* In flat mode, that's it. */
11126 /* Otherwise, the frame pointer is required if the function isn't leaf. */
11127 return !(current_function_is_leaf && only_leaf_regs_used ());
11130 /* The way this is structured, we can't eliminate SFP in favor of SP
11131 if the frame pointer is required: we want to use the SFP->HFP elimination
11132 in that case. But the test in update_eliminables doesn't know we are
11133 assuming below that we only do the former elimination. */
11136 sparc_can_eliminate (const int from ATTRIBUTE_UNUSED, const int to)
11138 return to == HARD_FRAME_POINTER_REGNUM || !sparc_frame_pointer_required ();
11141 /* Return the hard frame pointer directly to bypass the stack bias. */
11144 sparc_builtin_setjmp_frame_value (void)
11146 return hard_frame_pointer_rtx;
11149 /* If !TARGET_FPU, then make the fp registers and fp cc regs fixed so that
11150 they won't be allocated. */
11153 sparc_conditional_register_usage (void)
11155 if (PIC_OFFSET_TABLE_REGNUM != INVALID_REGNUM)
11157 fixed_regs[PIC_OFFSET_TABLE_REGNUM] = 1;
11158 call_used_regs[PIC_OFFSET_TABLE_REGNUM] = 1;
11160 /* If the user has passed -f{fixed,call-{used,saved}}-g5 */
11161 /* then honor it. */
11162 if (TARGET_ARCH32 && fixed_regs[5])
11164 else if (TARGET_ARCH64 && fixed_regs[5] == 2)
11169 for (regno = SPARC_FIRST_V9_FP_REG;
11170 regno <= SPARC_LAST_V9_FP_REG;
11172 fixed_regs[regno] = 1;
11173 /* %fcc0 is used by v8 and v9. */
11174 for (regno = SPARC_FIRST_V9_FCC_REG + 1;
11175 regno <= SPARC_LAST_V9_FCC_REG;
11177 fixed_regs[regno] = 1;
11182 for (regno = 32; regno < SPARC_LAST_V9_FCC_REG; regno++)
11183 fixed_regs[regno] = 1;
11185 /* If the user has passed -f{fixed,call-{used,saved}}-g2 */
11186 /* then honor it. Likewise with g3 and g4. */
11187 if (fixed_regs[2] == 2)
11188 fixed_regs[2] = ! TARGET_APP_REGS;
11189 if (fixed_regs[3] == 2)
11190 fixed_regs[3] = ! TARGET_APP_REGS;
11191 if (TARGET_ARCH32 && fixed_regs[4] == 2)
11192 fixed_regs[4] = ! TARGET_APP_REGS;
11193 else if (TARGET_CM_EMBMEDANY)
11195 else if (fixed_regs[4] == 2)
11200 /* Disable leaf functions. */
11201 memset (sparc_leaf_regs, 0, FIRST_PSEUDO_REGISTER);
11202 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
11203 leaf_reg_remap [regno] = regno;
11206 global_regs[SPARC_GSR_REG] = 1;
11209 /* Implement TARGET_PREFERRED_RELOAD_CLASS:
11211 - We can't load constants into FP registers.
11212 - We can't load FP constants into integer registers when soft-float,
11213 because there is no soft-float pattern with a r/F constraint.
11214 - We can't load FP constants into integer registers for TFmode unless
11215 it is 0.0L, because there is no movtf pattern with a r/F constraint.
11216 - Try and reload integer constants (symbolic or otherwise) back into
11217 registers directly, rather than having them dumped to memory. */
11220 sparc_preferred_reload_class (rtx x, reg_class_t rclass)
11222 enum machine_mode mode = GET_MODE (x);
11223 if (CONSTANT_P (x))
11225 if (FP_REG_CLASS_P (rclass)
11226 || rclass == GENERAL_OR_FP_REGS
11227 || rclass == GENERAL_OR_EXTRA_FP_REGS
11228 || (GET_MODE_CLASS (mode) == MODE_FLOAT && ! TARGET_FPU)
11229 || (mode == TFmode && ! const_zero_operand (x, mode)))
11232 if (GET_MODE_CLASS (mode) == MODE_INT)
11233 return GENERAL_REGS;
11235 if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT)
11237 if (! FP_REG_CLASS_P (rclass)
11238 || !(const_zero_operand (x, mode)
11239 || const_all_ones_operand (x, mode)))
11246 && (rclass == EXTRA_FP_REGS
11247 || rclass == GENERAL_OR_EXTRA_FP_REGS))
11249 int regno = true_regnum (x);
11251 if (SPARC_INT_REG_P (regno))
11252 return (rclass == EXTRA_FP_REGS
11253 ? FP_REGS : GENERAL_OR_FP_REGS);
11259 /* Output a wide multiply instruction in V8+ mode. INSN is the instruction,
11260 OPERANDS are its operands and OPCODE is the mnemonic to be used. */
11263 output_v8plus_mult (rtx insn, rtx *operands, const char *opcode)
11267 gcc_assert (! TARGET_ARCH64);
11269 if (sparc_check_64 (operands[1], insn) <= 0)
11270 output_asm_insn ("srl\t%L1, 0, %L1", operands);
11271 if (which_alternative == 1)
11272 output_asm_insn ("sllx\t%H1, 32, %H1", operands);
11273 if (GET_CODE (operands[2]) == CONST_INT)
11275 if (which_alternative == 1)
11277 output_asm_insn ("or\t%L1, %H1, %H1", operands);
11278 sprintf (mulstr, "%s\t%%H1, %%2, %%L0", opcode);
11279 output_asm_insn (mulstr, operands);
11280 return "srlx\t%L0, 32, %H0";
11284 output_asm_insn ("sllx\t%H1, 32, %3", operands);
11285 output_asm_insn ("or\t%L1, %3, %3", operands);
11286 sprintf (mulstr, "%s\t%%3, %%2, %%3", opcode);
11287 output_asm_insn (mulstr, operands);
11288 output_asm_insn ("srlx\t%3, 32, %H0", operands);
11289 return "mov\t%3, %L0";
11292 else if (rtx_equal_p (operands[1], operands[2]))
11294 if (which_alternative == 1)
11296 output_asm_insn ("or\t%L1, %H1, %H1", operands);
11297 sprintf (mulstr, "%s\t%%H1, %%H1, %%L0", opcode);
11298 output_asm_insn (mulstr, operands);
11299 return "srlx\t%L0, 32, %H0";
11303 output_asm_insn ("sllx\t%H1, 32, %3", operands);
11304 output_asm_insn ("or\t%L1, %3, %3", operands);
11305 sprintf (mulstr, "%s\t%%3, %%3, %%3", opcode);
11306 output_asm_insn (mulstr, operands);
11307 output_asm_insn ("srlx\t%3, 32, %H0", operands);
11308 return "mov\t%3, %L0";
11311 if (sparc_check_64 (operands[2], insn) <= 0)
11312 output_asm_insn ("srl\t%L2, 0, %L2", operands);
11313 if (which_alternative == 1)
11315 output_asm_insn ("or\t%L1, %H1, %H1", operands);
11316 output_asm_insn ("sllx\t%H2, 32, %L1", operands);
11317 output_asm_insn ("or\t%L2, %L1, %L1", operands);
11318 sprintf (mulstr, "%s\t%%H1, %%L1, %%L0", opcode);
11319 output_asm_insn (mulstr, operands);
11320 return "srlx\t%L0, 32, %H0";
11324 output_asm_insn ("sllx\t%H1, 32, %3", operands);
11325 output_asm_insn ("sllx\t%H2, 32, %4", operands);
11326 output_asm_insn ("or\t%L1, %3, %3", operands);
11327 output_asm_insn ("or\t%L2, %4, %4", operands);
11328 sprintf (mulstr, "%s\t%%3, %%4, %%3", opcode);
11329 output_asm_insn (mulstr, operands);
11330 output_asm_insn ("srlx\t%3, 32, %H0", operands);
11331 return "mov\t%3, %L0";
11335 /* Subroutine of sparc_expand_vector_init. Emit code to initialize
11336 all fields of TARGET to ELT by means of VIS2 BSHUFFLE insn. MODE
11337 and INNER_MODE are the modes describing TARGET. */
11340 vector_init_bshuffle (rtx target, rtx elt, enum machine_mode mode,
11341 enum machine_mode inner_mode)
11343 rtx t1, final_insn;
11346 t1 = gen_reg_rtx (mode);
11348 elt = convert_modes (SImode, inner_mode, elt, true);
11349 emit_move_insn (gen_lowpart(SImode, t1), elt);
11354 final_insn = gen_bshufflev2si_vis (target, t1, t1);
11355 bmask = 0x45674567;
11358 final_insn = gen_bshufflev4hi_vis (target, t1, t1);
11359 bmask = 0x67676767;
11362 final_insn = gen_bshufflev8qi_vis (target, t1, t1);
11363 bmask = 0x77777777;
11366 gcc_unreachable ();
11369 emit_insn (gen_bmasksi_vis (gen_reg_rtx (SImode), CONST0_RTX (SImode),
11370 force_reg (SImode, GEN_INT (bmask))));
11371 emit_insn (final_insn);
11375 vector_init_fpmerge (rtx target, rtx elt, enum machine_mode inner_mode)
11377 rtx t1, t2, t3, t3_low;
11379 t1 = gen_reg_rtx (V4QImode);
11380 elt = convert_modes (SImode, inner_mode, elt, true);
11381 emit_move_insn (gen_lowpart (SImode, t1), elt);
11383 t2 = gen_reg_rtx (V4QImode);
11384 emit_move_insn (t2, t1);
11386 t3 = gen_reg_rtx (V8QImode);
11387 t3_low = gen_lowpart (V4QImode, t3);
11389 emit_insn (gen_fpmerge_vis (t3, t1, t2));
11390 emit_move_insn (t1, t3_low);
11391 emit_move_insn (t2, t3_low);
11393 emit_insn (gen_fpmerge_vis (t3, t1, t2));
11394 emit_move_insn (t1, t3_low);
11395 emit_move_insn (t2, t3_low);
11397 emit_insn (gen_fpmerge_vis (gen_lowpart (V8QImode, target), t1, t2));
11401 vector_init_faligndata (rtx target, rtx elt, enum machine_mode inner_mode)
11403 rtx t1 = gen_reg_rtx (V4HImode);
11405 elt = convert_modes (SImode, inner_mode, elt, true);
11407 emit_move_insn (gen_lowpart (SImode, t1), elt);
11409 emit_insn (gen_alignaddrsi_vis (gen_reg_rtx (SImode),
11410 force_reg (SImode, GEN_INT (6)),
11411 CONST0_RTX (SImode)));
11413 emit_insn (gen_faligndatav4hi_vis (target, t1, target));
11414 emit_insn (gen_faligndatav4hi_vis (target, t1, target));
11415 emit_insn (gen_faligndatav4hi_vis (target, t1, target));
11416 emit_insn (gen_faligndatav4hi_vis (target, t1, target));
11419 /* Emit code to initialize TARGET to values for individual fields VALS. */
11422 sparc_expand_vector_init (rtx target, rtx vals)
11424 enum machine_mode mode = GET_MODE (target);
11425 enum machine_mode inner_mode = GET_MODE_INNER (mode);
11426 int n_elts = GET_MODE_NUNITS (mode);
11432 for (i = 0; i < n_elts; i++)
11434 rtx x = XVECEXP (vals, 0, i);
11435 if (!CONSTANT_P (x))
11438 if (i > 0 && !rtx_equal_p (x, XVECEXP (vals, 0, 0)))
11444 emit_move_insn (target, gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0)));
11448 if (GET_MODE_SIZE (inner_mode) == GET_MODE_SIZE (mode))
11450 if (GET_MODE_SIZE (inner_mode) == 4)
11452 emit_move_insn (gen_lowpart (SImode, target),
11453 gen_lowpart (SImode, XVECEXP (vals, 0, 0)));
11456 else if (GET_MODE_SIZE (inner_mode) == 8)
11458 emit_move_insn (gen_lowpart (DImode, target),
11459 gen_lowpart (DImode, XVECEXP (vals, 0, 0)));
11463 else if (GET_MODE_SIZE (inner_mode) == GET_MODE_SIZE (word_mode)
11464 && GET_MODE_SIZE (mode) == 2 * GET_MODE_SIZE (word_mode))
11466 emit_move_insn (gen_highpart (word_mode, target),
11467 gen_lowpart (word_mode, XVECEXP (vals, 0, 0)));
11468 emit_move_insn (gen_lowpart (word_mode, target),
11469 gen_lowpart (word_mode, XVECEXP (vals, 0, 1)));
11473 if (all_same && GET_MODE_SIZE (mode) == 8)
11477 vector_init_bshuffle (target, XVECEXP (vals, 0, 0), mode, inner_mode);
11480 if (mode == V8QImode)
11482 vector_init_fpmerge (target, XVECEXP (vals, 0, 0), inner_mode);
11485 if (mode == V4HImode)
11487 vector_init_faligndata (target, XVECEXP (vals, 0, 0), inner_mode);
11492 mem = assign_stack_temp (mode, GET_MODE_SIZE (mode), 0);
11493 for (i = 0; i < n_elts; i++)
11494 emit_move_insn (adjust_address_nv (mem, inner_mode,
11495 i * GET_MODE_SIZE (inner_mode)),
11496 XVECEXP (vals, 0, i));
11497 emit_move_insn (target, mem);
11500 /* Implement TARGET_SECONDARY_RELOAD. */
11503 sparc_secondary_reload (bool in_p, rtx x, reg_class_t rclass_i,
11504 enum machine_mode mode, secondary_reload_info *sri)
11506 enum reg_class rclass = (enum reg_class) rclass_i;
11508 sri->icode = CODE_FOR_nothing;
11509 sri->extra_cost = 0;
11511 /* We need a temporary when loading/storing a HImode/QImode value
11512 between memory and the FPU registers. This can happen when combine puts
11513 a paradoxical subreg in a float/fix conversion insn. */
11514 if (FP_REG_CLASS_P (rclass)
11515 && (mode == HImode || mode == QImode)
11516 && (GET_CODE (x) == MEM
11517 || ((GET_CODE (x) == REG || GET_CODE (x) == SUBREG)
11518 && true_regnum (x) == -1)))
11519 return GENERAL_REGS;
11521 /* On 32-bit we need a temporary when loading/storing a DFmode value
11522 between unaligned memory and the upper FPU registers. */
11524 && rclass == EXTRA_FP_REGS
11526 && GET_CODE (x) == MEM
11527 && ! mem_min_alignment (x, 8))
11530 if (((TARGET_CM_MEDANY
11531 && symbolic_operand (x, mode))
11532 || (TARGET_CM_EMBMEDANY
11533 && text_segment_operand (x, mode)))
11537 sri->icode = direct_optab_handler (reload_in_optab, mode);
11539 sri->icode = direct_optab_handler (reload_out_optab, mode);
11543 if (TARGET_VIS3 && TARGET_ARCH32)
11545 int regno = true_regnum (x);
11547 /* When using VIS3 fp<-->int register moves, on 32-bit we have
11548 to move 8-byte values in 4-byte pieces. This only works via
11549 FP_REGS, and not via EXTRA_FP_REGS. Therefore if we try to
11550 move between EXTRA_FP_REGS and GENERAL_REGS, we will need
11551 an FP_REGS intermediate move. */
11552 if ((rclass == EXTRA_FP_REGS && SPARC_INT_REG_P (regno))
11553 || ((general_or_i64_p (rclass)
11554 || rclass == GENERAL_OR_FP_REGS)
11555 && SPARC_FP_REG_P (regno)))
11557 sri->extra_cost = 2;
11565 /* Emit code to conditionally move either OPERANDS[2] or OPERANDS[3] into
11566 OPERANDS[0] in MODE. OPERANDS[1] is the operator of the condition. */
11569 sparc_expand_conditional_move (enum machine_mode mode, rtx *operands)
11571 enum rtx_code rc = GET_CODE (operands[1]);
11572 enum machine_mode cmp_mode;
11573 rtx cc_reg, dst, cmp;
11576 if (GET_MODE (XEXP (cmp, 0)) == DImode && !TARGET_ARCH64)
11579 if (GET_MODE (XEXP (cmp, 0)) == TFmode && !TARGET_HARD_QUAD)
11580 cmp = sparc_emit_float_lib_cmp (XEXP (cmp, 0), XEXP (cmp, 1), rc);
11582 cmp_mode = GET_MODE (XEXP (cmp, 0));
11583 rc = GET_CODE (cmp);
11586 if (! rtx_equal_p (operands[2], dst)
11587 && ! rtx_equal_p (operands[3], dst))
11589 if (reg_overlap_mentioned_p (dst, cmp))
11590 dst = gen_reg_rtx (mode);
11592 emit_move_insn (dst, operands[3]);
11594 else if (operands[2] == dst)
11596 operands[2] = operands[3];
11598 if (GET_MODE_CLASS (cmp_mode) == MODE_FLOAT)
11599 rc = reverse_condition_maybe_unordered (rc);
11601 rc = reverse_condition (rc);
11604 if (XEXP (cmp, 1) == const0_rtx
11605 && GET_CODE (XEXP (cmp, 0)) == REG
11606 && cmp_mode == DImode
11607 && v9_regcmp_p (rc))
11608 cc_reg = XEXP (cmp, 0);
11610 cc_reg = gen_compare_reg_1 (rc, XEXP (cmp, 0), XEXP (cmp, 1));
11612 cmp = gen_rtx_fmt_ee (rc, GET_MODE (cc_reg), cc_reg, const0_rtx);
11614 emit_insn (gen_rtx_SET (VOIDmode, dst,
11615 gen_rtx_IF_THEN_ELSE (mode, cmp, operands[2], dst)));
11617 if (dst != operands[0])
11618 emit_move_insn (operands[0], dst);
11623 /* Emit code to conditionally move a combination of OPERANDS[1] and OPERANDS[2]
11624 into OPERANDS[0] in MODE, depending on the outcome of the comparison of
11625 OPERANDS[4] and OPERANDS[5]. OPERANDS[3] is the operator of the condition.
11626 FCODE is the machine code to be used for OPERANDS[3] and CCODE the machine
11627 code to be used for the condition mask. */
11630 sparc_expand_vcond (enum machine_mode mode, rtx *operands, int ccode, int fcode)
11632 rtx mask, cop0, cop1, fcmp, cmask, bshuf, gsr;
11633 enum rtx_code code = GET_CODE (operands[3]);
11635 mask = gen_reg_rtx (Pmode);
11636 cop0 = operands[4];
11637 cop1 = operands[5];
11638 if (code == LT || code == GE)
11642 code = swap_condition (code);
11643 t = cop0; cop0 = cop1; cop1 = t;
11646 gsr = gen_rtx_REG (DImode, SPARC_GSR_REG);
11648 fcmp = gen_rtx_UNSPEC (Pmode,
11649 gen_rtvec (1, gen_rtx_fmt_ee (code, mode, cop0, cop1)),
11652 cmask = gen_rtx_UNSPEC (DImode,
11653 gen_rtvec (2, mask, gsr),
11656 bshuf = gen_rtx_UNSPEC (mode,
11657 gen_rtvec (3, operands[1], operands[2], gsr),
11660 emit_insn (gen_rtx_SET (VOIDmode, mask, fcmp));
11661 emit_insn (gen_rtx_SET (VOIDmode, gsr, cmask));
11663 emit_insn (gen_rtx_SET (VOIDmode, operands[0], bshuf));
11666 /* On sparc, any mode which naturally allocates into the float
11667 registers should return 4 here. */
11670 sparc_regmode_natural_size (enum machine_mode mode)
11672 int size = UNITS_PER_WORD;
11676 enum mode_class mclass = GET_MODE_CLASS (mode);
11678 if (mclass == MODE_FLOAT || mclass == MODE_VECTOR_INT)
11685 /* Return TRUE if it is a good idea to tie two pseudo registers
11686 when one has mode MODE1 and one has mode MODE2.
11687 If HARD_REGNO_MODE_OK could produce different values for MODE1 and MODE2,
11688 for any hard reg, then this must be FALSE for correct output.
11690 For V9 we have to deal with the fact that only the lower 32 floating
11691 point registers are 32-bit addressable. */
11694 sparc_modes_tieable_p (enum machine_mode mode1, enum machine_mode mode2)
11696 enum mode_class mclass1, mclass2;
11697 unsigned short size1, size2;
11699 if (mode1 == mode2)
11702 mclass1 = GET_MODE_CLASS (mode1);
11703 mclass2 = GET_MODE_CLASS (mode2);
11704 if (mclass1 != mclass2)
11710 /* Classes are the same and we are V9 so we have to deal with upper
11711 vs. lower floating point registers. If one of the modes is a
11712 4-byte mode, and the other is not, we have to mark them as not
11713 tieable because only the lower 32 floating point register are
11714 addressable 32-bits at a time.
11716 We can't just test explicitly for SFmode, otherwise we won't
11717 cover the vector mode cases properly. */
11719 if (mclass1 != MODE_FLOAT && mclass1 != MODE_VECTOR_INT)
11722 size1 = GET_MODE_SIZE (mode1);
11723 size2 = GET_MODE_SIZE (mode2);
11724 if ((size1 > 4 && size2 == 4)
11725 || (size2 > 4 && size1 == 4))
11731 #include "gt-sparc.h"