]> git.ipfire.org Git - thirdparty/gcc.git/blob - gcc/optabs.c
gcc/
[thirdparty/gcc.git] / gcc / optabs.c
1 /* Expand the basic unary and binary arithmetic operations, for GNU compiler.
2 Copyright (C) 1987, 1988, 1992, 1993, 1994, 1995, 1996, 1997, 1998,
3 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010,
4 2011 Free Software Foundation, Inc.
5
6 This file is part of GCC.
7
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
12
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 for more details.
17
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
21
22
23 #include "config.h"
24 #include "system.h"
25 #include "coretypes.h"
26 #include "tm.h"
27 #include "diagnostic-core.h"
28
29 /* Include insn-config.h before expr.h so that HAVE_conditional_move
30 is properly defined. */
31 #include "insn-config.h"
32 #include "rtl.h"
33 #include "tree.h"
34 #include "tm_p.h"
35 #include "flags.h"
36 #include "function.h"
37 #include "except.h"
38 #include "expr.h"
39 #include "optabs.h"
40 #include "libfuncs.h"
41 #include "recog.h"
42 #include "reload.h"
43 #include "ggc.h"
44 #include "basic-block.h"
45 #include "target.h"
46
47 struct target_optabs default_target_optabs;
48 struct target_libfuncs default_target_libfuncs;
49 #if SWITCHABLE_TARGET
50 struct target_optabs *this_target_optabs = &default_target_optabs;
51 struct target_libfuncs *this_target_libfuncs = &default_target_libfuncs;
52 #endif
53
54 #define libfunc_hash \
55 (this_target_libfuncs->x_libfunc_hash)
56
57 /* Contains the optab used for each rtx code. */
58 optab code_to_optab[NUM_RTX_CODE + 1];
59
60 static void prepare_float_lib_cmp (rtx, rtx, enum rtx_code, rtx *,
61 enum machine_mode *);
62 static rtx expand_unop_direct (enum machine_mode, optab, rtx, rtx, int);
63
64 /* Debug facility for use in GDB. */
65 void debug_optab_libfuncs (void);
66
67 /* Prefixes for the current version of decimal floating point (BID vs. DPD) */
68 #if ENABLE_DECIMAL_BID_FORMAT
69 #define DECIMAL_PREFIX "bid_"
70 #else
71 #define DECIMAL_PREFIX "dpd_"
72 #endif
73 \f
74 /* Used for libfunc_hash. */
75
76 static hashval_t
77 hash_libfunc (const void *p)
78 {
79 const struct libfunc_entry *const e = (const struct libfunc_entry *) p;
80
81 return (((int) e->mode1 + (int) e->mode2 * NUM_MACHINE_MODES)
82 ^ e->optab);
83 }
84
85 /* Used for libfunc_hash. */
86
87 static int
88 eq_libfunc (const void *p, const void *q)
89 {
90 const struct libfunc_entry *const e1 = (const struct libfunc_entry *) p;
91 const struct libfunc_entry *const e2 = (const struct libfunc_entry *) q;
92
93 return (e1->optab == e2->optab
94 && e1->mode1 == e2->mode1
95 && e1->mode2 == e2->mode2);
96 }
97
98 /* Return libfunc corresponding operation defined by OPTAB converting
99 from MODE2 to MODE1. Trigger lazy initialization if needed, return NULL
100 if no libfunc is available. */
101 rtx
102 convert_optab_libfunc (convert_optab optab, enum machine_mode mode1,
103 enum machine_mode mode2)
104 {
105 struct libfunc_entry e;
106 struct libfunc_entry **slot;
107
108 e.optab = (size_t) (optab - &convert_optab_table[0]);
109 e.mode1 = mode1;
110 e.mode2 = mode2;
111 slot = (struct libfunc_entry **) htab_find_slot (libfunc_hash, &e, NO_INSERT);
112 if (!slot)
113 {
114 if (optab->libcall_gen)
115 {
116 optab->libcall_gen (optab, optab->libcall_basename, mode1, mode2);
117 slot = (struct libfunc_entry **) htab_find_slot (libfunc_hash, &e, NO_INSERT);
118 if (slot)
119 return (*slot)->libfunc;
120 else
121 return NULL;
122 }
123 return NULL;
124 }
125 return (*slot)->libfunc;
126 }
127
128 /* Return libfunc corresponding operation defined by OPTAB in MODE.
129 Trigger lazy initialization if needed, return NULL if no libfunc is
130 available. */
131 rtx
132 optab_libfunc (optab optab, enum machine_mode mode)
133 {
134 struct libfunc_entry e;
135 struct libfunc_entry **slot;
136
137 e.optab = (size_t) (optab - &optab_table[0]);
138 e.mode1 = mode;
139 e.mode2 = VOIDmode;
140 slot = (struct libfunc_entry **) htab_find_slot (libfunc_hash, &e, NO_INSERT);
141 if (!slot)
142 {
143 if (optab->libcall_gen)
144 {
145 optab->libcall_gen (optab, optab->libcall_basename,
146 optab->libcall_suffix, mode);
147 slot = (struct libfunc_entry **) htab_find_slot (libfunc_hash,
148 &e, NO_INSERT);
149 if (slot)
150 return (*slot)->libfunc;
151 else
152 return NULL;
153 }
154 return NULL;
155 }
156 return (*slot)->libfunc;
157 }
158
159 \f
160 /* Add a REG_EQUAL note to the last insn in INSNS. TARGET is being set to
161 the result of operation CODE applied to OP0 (and OP1 if it is a binary
162 operation).
163
164 If the last insn does not set TARGET, don't do anything, but return 1.
165
166 If a previous insn sets TARGET and TARGET is one of OP0 or OP1,
167 don't add the REG_EQUAL note but return 0. Our caller can then try
168 again, ensuring that TARGET is not one of the operands. */
169
170 static int
171 add_equal_note (rtx insns, rtx target, enum rtx_code code, rtx op0, rtx op1)
172 {
173 rtx last_insn, insn, set;
174 rtx note;
175
176 gcc_assert (insns && INSN_P (insns) && NEXT_INSN (insns));
177
178 if (GET_RTX_CLASS (code) != RTX_COMM_ARITH
179 && GET_RTX_CLASS (code) != RTX_BIN_ARITH
180 && GET_RTX_CLASS (code) != RTX_COMM_COMPARE
181 && GET_RTX_CLASS (code) != RTX_COMPARE
182 && GET_RTX_CLASS (code) != RTX_UNARY)
183 return 1;
184
185 if (GET_CODE (target) == ZERO_EXTRACT)
186 return 1;
187
188 for (last_insn = insns;
189 NEXT_INSN (last_insn) != NULL_RTX;
190 last_insn = NEXT_INSN (last_insn))
191 ;
192
193 set = single_set (last_insn);
194 if (set == NULL_RTX)
195 return 1;
196
197 if (! rtx_equal_p (SET_DEST (set), target)
198 /* For a STRICT_LOW_PART, the REG_NOTE applies to what is inside it. */
199 && (GET_CODE (SET_DEST (set)) != STRICT_LOW_PART
200 || ! rtx_equal_p (XEXP (SET_DEST (set), 0), target)))
201 return 1;
202
203 /* If TARGET is in OP0 or OP1, check if anything in SEQ sets TARGET
204 besides the last insn. */
205 if (reg_overlap_mentioned_p (target, op0)
206 || (op1 && reg_overlap_mentioned_p (target, op1)))
207 {
208 insn = PREV_INSN (last_insn);
209 while (insn != NULL_RTX)
210 {
211 if (reg_set_p (target, insn))
212 return 0;
213
214 insn = PREV_INSN (insn);
215 }
216 }
217
218 if (GET_RTX_CLASS (code) == RTX_UNARY)
219 switch (code)
220 {
221 case FFS:
222 case CLZ:
223 case CTZ:
224 case CLRSB:
225 case POPCOUNT:
226 case PARITY:
227 case BSWAP:
228 if (GET_MODE (op0) != VOIDmode && GET_MODE (target) != GET_MODE (op0))
229 {
230 note = gen_rtx_fmt_e (code, GET_MODE (op0), copy_rtx (op0));
231 if (GET_MODE_SIZE (GET_MODE (op0))
232 > GET_MODE_SIZE (GET_MODE (target)))
233 note = simplify_gen_unary (TRUNCATE, GET_MODE (target),
234 note, GET_MODE (op0));
235 else
236 note = simplify_gen_unary (ZERO_EXTEND, GET_MODE (target),
237 note, GET_MODE (op0));
238 break;
239 }
240 /* FALLTHRU */
241 default:
242 note = gen_rtx_fmt_e (code, GET_MODE (target), copy_rtx (op0));
243 break;
244 }
245 else
246 note = gen_rtx_fmt_ee (code, GET_MODE (target), copy_rtx (op0), copy_rtx (op1));
247
248 set_unique_reg_note (last_insn, REG_EQUAL, note);
249
250 return 1;
251 }
252 \f
253 /* Given two input operands, OP0 and OP1, determine what the correct from_mode
254 for a widening operation would be. In most cases this would be OP0, but if
255 that's a constant it'll be VOIDmode, which isn't useful. */
256
257 static enum machine_mode
258 widened_mode (enum machine_mode to_mode, rtx op0, rtx op1)
259 {
260 enum machine_mode m0 = GET_MODE (op0);
261 enum machine_mode m1 = GET_MODE (op1);
262 enum machine_mode result;
263
264 if (m0 == VOIDmode && m1 == VOIDmode)
265 return to_mode;
266 else if (m0 == VOIDmode || GET_MODE_SIZE (m0) < GET_MODE_SIZE (m1))
267 result = m1;
268 else
269 result = m0;
270
271 if (GET_MODE_SIZE (result) > GET_MODE_SIZE (to_mode))
272 return to_mode;
273
274 return result;
275 }
276 \f
277 /* Find a widening optab even if it doesn't widen as much as we want.
278 E.g. if from_mode is HImode, and to_mode is DImode, and there is no
279 direct HI->SI insn, then return SI->DI, if that exists.
280 If PERMIT_NON_WIDENING is non-zero then this can be used with
281 non-widening optabs also. */
282
283 enum insn_code
284 find_widening_optab_handler_and_mode (optab op, enum machine_mode to_mode,
285 enum machine_mode from_mode,
286 int permit_non_widening,
287 enum machine_mode *found_mode)
288 {
289 for (; (permit_non_widening || from_mode != to_mode)
290 && GET_MODE_SIZE (from_mode) <= GET_MODE_SIZE (to_mode)
291 && from_mode != VOIDmode;
292 from_mode = GET_MODE_WIDER_MODE (from_mode))
293 {
294 enum insn_code handler = widening_optab_handler (op, to_mode,
295 from_mode);
296
297 if (handler != CODE_FOR_nothing)
298 {
299 if (found_mode)
300 *found_mode = from_mode;
301 return handler;
302 }
303 }
304
305 return CODE_FOR_nothing;
306 }
307 \f
308 /* Widen OP to MODE and return the rtx for the widened operand. UNSIGNEDP
309 says whether OP is signed or unsigned. NO_EXTEND is nonzero if we need
310 not actually do a sign-extend or zero-extend, but can leave the
311 higher-order bits of the result rtx undefined, for example, in the case
312 of logical operations, but not right shifts. */
313
314 static rtx
315 widen_operand (rtx op, enum machine_mode mode, enum machine_mode oldmode,
316 int unsignedp, int no_extend)
317 {
318 rtx result;
319
320 /* If we don't have to extend and this is a constant, return it. */
321 if (no_extend && GET_MODE (op) == VOIDmode)
322 return op;
323
324 /* If we must extend do so. If OP is a SUBREG for a promoted object, also
325 extend since it will be more efficient to do so unless the signedness of
326 a promoted object differs from our extension. */
327 if (! no_extend
328 || (GET_CODE (op) == SUBREG && SUBREG_PROMOTED_VAR_P (op)
329 && SUBREG_PROMOTED_UNSIGNED_P (op) == unsignedp))
330 return convert_modes (mode, oldmode, op, unsignedp);
331
332 /* If MODE is no wider than a single word, we return a paradoxical
333 SUBREG. */
334 if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD)
335 return gen_rtx_SUBREG (mode, force_reg (GET_MODE (op), op), 0);
336
337 /* Otherwise, get an object of MODE, clobber it, and set the low-order
338 part to OP. */
339
340 result = gen_reg_rtx (mode);
341 emit_clobber (result);
342 emit_move_insn (gen_lowpart (GET_MODE (op), result), op);
343 return result;
344 }
345 \f
346 /* Return the optab used for computing the operation given by the tree code,
347 CODE and the tree EXP. This function is not always usable (for example, it
348 cannot give complete results for multiplication or division) but probably
349 ought to be relied on more widely throughout the expander. */
350 optab
351 optab_for_tree_code (enum tree_code code, const_tree type,
352 enum optab_subtype subtype)
353 {
354 bool trapv;
355 switch (code)
356 {
357 case BIT_AND_EXPR:
358 return and_optab;
359
360 case BIT_IOR_EXPR:
361 return ior_optab;
362
363 case BIT_NOT_EXPR:
364 return one_cmpl_optab;
365
366 case BIT_XOR_EXPR:
367 return xor_optab;
368
369 case TRUNC_MOD_EXPR:
370 case CEIL_MOD_EXPR:
371 case FLOOR_MOD_EXPR:
372 case ROUND_MOD_EXPR:
373 return TYPE_UNSIGNED (type) ? umod_optab : smod_optab;
374
375 case RDIV_EXPR:
376 case TRUNC_DIV_EXPR:
377 case CEIL_DIV_EXPR:
378 case FLOOR_DIV_EXPR:
379 case ROUND_DIV_EXPR:
380 case EXACT_DIV_EXPR:
381 if (TYPE_SATURATING(type))
382 return TYPE_UNSIGNED(type) ? usdiv_optab : ssdiv_optab;
383 return TYPE_UNSIGNED (type) ? udiv_optab : sdiv_optab;
384
385 case LSHIFT_EXPR:
386 if (TREE_CODE (type) == VECTOR_TYPE)
387 {
388 if (subtype == optab_vector)
389 return TYPE_SATURATING (type) ? NULL : vashl_optab;
390
391 gcc_assert (subtype == optab_scalar);
392 }
393 if (TYPE_SATURATING(type))
394 return TYPE_UNSIGNED(type) ? usashl_optab : ssashl_optab;
395 return ashl_optab;
396
397 case RSHIFT_EXPR:
398 if (TREE_CODE (type) == VECTOR_TYPE)
399 {
400 if (subtype == optab_vector)
401 return TYPE_UNSIGNED (type) ? vlshr_optab : vashr_optab;
402
403 gcc_assert (subtype == optab_scalar);
404 }
405 return TYPE_UNSIGNED (type) ? lshr_optab : ashr_optab;
406
407 case LROTATE_EXPR:
408 if (TREE_CODE (type) == VECTOR_TYPE)
409 {
410 if (subtype == optab_vector)
411 return vrotl_optab;
412
413 gcc_assert (subtype == optab_scalar);
414 }
415 return rotl_optab;
416
417 case RROTATE_EXPR:
418 if (TREE_CODE (type) == VECTOR_TYPE)
419 {
420 if (subtype == optab_vector)
421 return vrotr_optab;
422
423 gcc_assert (subtype == optab_scalar);
424 }
425 return rotr_optab;
426
427 case MAX_EXPR:
428 return TYPE_UNSIGNED (type) ? umax_optab : smax_optab;
429
430 case MIN_EXPR:
431 return TYPE_UNSIGNED (type) ? umin_optab : smin_optab;
432
433 case REALIGN_LOAD_EXPR:
434 return vec_realign_load_optab;
435
436 case WIDEN_SUM_EXPR:
437 return TYPE_UNSIGNED (type) ? usum_widen_optab : ssum_widen_optab;
438
439 case DOT_PROD_EXPR:
440 return TYPE_UNSIGNED (type) ? udot_prod_optab : sdot_prod_optab;
441
442 case WIDEN_MULT_PLUS_EXPR:
443 return (TYPE_UNSIGNED (type)
444 ? (TYPE_SATURATING (type)
445 ? usmadd_widen_optab : umadd_widen_optab)
446 : (TYPE_SATURATING (type)
447 ? ssmadd_widen_optab : smadd_widen_optab));
448
449 case WIDEN_MULT_MINUS_EXPR:
450 return (TYPE_UNSIGNED (type)
451 ? (TYPE_SATURATING (type)
452 ? usmsub_widen_optab : umsub_widen_optab)
453 : (TYPE_SATURATING (type)
454 ? ssmsub_widen_optab : smsub_widen_optab));
455
456 case FMA_EXPR:
457 return fma_optab;
458
459 case REDUC_MAX_EXPR:
460 return TYPE_UNSIGNED (type) ? reduc_umax_optab : reduc_smax_optab;
461
462 case REDUC_MIN_EXPR:
463 return TYPE_UNSIGNED (type) ? reduc_umin_optab : reduc_smin_optab;
464
465 case REDUC_PLUS_EXPR:
466 return TYPE_UNSIGNED (type) ? reduc_uplus_optab : reduc_splus_optab;
467
468 case VEC_LSHIFT_EXPR:
469 return vec_shl_optab;
470
471 case VEC_RSHIFT_EXPR:
472 return vec_shr_optab;
473
474 case VEC_WIDEN_MULT_HI_EXPR:
475 return TYPE_UNSIGNED (type) ?
476 vec_widen_umult_hi_optab : vec_widen_smult_hi_optab;
477
478 case VEC_WIDEN_MULT_LO_EXPR:
479 return TYPE_UNSIGNED (type) ?
480 vec_widen_umult_lo_optab : vec_widen_smult_lo_optab;
481
482 case VEC_WIDEN_LSHIFT_HI_EXPR:
483 return TYPE_UNSIGNED (type) ?
484 vec_widen_ushiftl_hi_optab : vec_widen_sshiftl_hi_optab;
485
486 case VEC_WIDEN_LSHIFT_LO_EXPR:
487 return TYPE_UNSIGNED (type) ?
488 vec_widen_ushiftl_lo_optab : vec_widen_sshiftl_lo_optab;
489
490 case VEC_UNPACK_HI_EXPR:
491 return TYPE_UNSIGNED (type) ?
492 vec_unpacku_hi_optab : vec_unpacks_hi_optab;
493
494 case VEC_UNPACK_LO_EXPR:
495 return TYPE_UNSIGNED (type) ?
496 vec_unpacku_lo_optab : vec_unpacks_lo_optab;
497
498 case VEC_UNPACK_FLOAT_HI_EXPR:
499 /* The signedness is determined from input operand. */
500 return TYPE_UNSIGNED (type) ?
501 vec_unpacku_float_hi_optab : vec_unpacks_float_hi_optab;
502
503 case VEC_UNPACK_FLOAT_LO_EXPR:
504 /* The signedness is determined from input operand. */
505 return TYPE_UNSIGNED (type) ?
506 vec_unpacku_float_lo_optab : vec_unpacks_float_lo_optab;
507
508 case VEC_PACK_TRUNC_EXPR:
509 return vec_pack_trunc_optab;
510
511 case VEC_PACK_SAT_EXPR:
512 return TYPE_UNSIGNED (type) ? vec_pack_usat_optab : vec_pack_ssat_optab;
513
514 case VEC_PACK_FIX_TRUNC_EXPR:
515 /* The signedness is determined from output operand. */
516 return TYPE_UNSIGNED (type) ?
517 vec_pack_ufix_trunc_optab : vec_pack_sfix_trunc_optab;
518
519 default:
520 break;
521 }
522
523 trapv = INTEGRAL_TYPE_P (type) && TYPE_OVERFLOW_TRAPS (type);
524 switch (code)
525 {
526 case POINTER_PLUS_EXPR:
527 case PLUS_EXPR:
528 if (TYPE_SATURATING(type))
529 return TYPE_UNSIGNED(type) ? usadd_optab : ssadd_optab;
530 return trapv ? addv_optab : add_optab;
531
532 case MINUS_EXPR:
533 if (TYPE_SATURATING(type))
534 return TYPE_UNSIGNED(type) ? ussub_optab : sssub_optab;
535 return trapv ? subv_optab : sub_optab;
536
537 case MULT_EXPR:
538 if (TYPE_SATURATING(type))
539 return TYPE_UNSIGNED(type) ? usmul_optab : ssmul_optab;
540 return trapv ? smulv_optab : smul_optab;
541
542 case NEGATE_EXPR:
543 if (TYPE_SATURATING(type))
544 return TYPE_UNSIGNED(type) ? usneg_optab : ssneg_optab;
545 return trapv ? negv_optab : neg_optab;
546
547 case ABS_EXPR:
548 return trapv ? absv_optab : abs_optab;
549
550 case VEC_EXTRACT_EVEN_EXPR:
551 return vec_extract_even_optab;
552
553 case VEC_EXTRACT_ODD_EXPR:
554 return vec_extract_odd_optab;
555
556 case VEC_INTERLEAVE_HIGH_EXPR:
557 return vec_interleave_high_optab;
558
559 case VEC_INTERLEAVE_LOW_EXPR:
560 return vec_interleave_low_optab;
561
562 default:
563 return NULL;
564 }
565 }
566 \f
567
568 /* Expand vector widening operations.
569
570 There are two different classes of operations handled here:
571 1) Operations whose result is wider than all the arguments to the operation.
572 Examples: VEC_UNPACK_HI/LO_EXPR, VEC_WIDEN_MULT_HI/LO_EXPR
573 In this case OP0 and optionally OP1 would be initialized,
574 but WIDE_OP wouldn't (not relevant for this case).
575 2) Operations whose result is of the same size as the last argument to the
576 operation, but wider than all the other arguments to the operation.
577 Examples: WIDEN_SUM_EXPR, VEC_DOT_PROD_EXPR.
578 In the case WIDE_OP, OP0 and optionally OP1 would be initialized.
579
580 E.g, when called to expand the following operations, this is how
581 the arguments will be initialized:
582 nops OP0 OP1 WIDE_OP
583 widening-sum 2 oprnd0 - oprnd1
584 widening-dot-product 3 oprnd0 oprnd1 oprnd2
585 widening-mult 2 oprnd0 oprnd1 -
586 type-promotion (vec-unpack) 1 oprnd0 - - */
587
588 rtx
589 expand_widen_pattern_expr (sepops ops, rtx op0, rtx op1, rtx wide_op,
590 rtx target, int unsignedp)
591 {
592 struct expand_operand eops[4];
593 tree oprnd0, oprnd1, oprnd2;
594 enum machine_mode wmode = VOIDmode, tmode0, tmode1 = VOIDmode;
595 optab widen_pattern_optab;
596 enum insn_code icode;
597 int nops = TREE_CODE_LENGTH (ops->code);
598 int op;
599
600 oprnd0 = ops->op0;
601 tmode0 = TYPE_MODE (TREE_TYPE (oprnd0));
602 widen_pattern_optab =
603 optab_for_tree_code (ops->code, TREE_TYPE (oprnd0), optab_default);
604 if (ops->code == WIDEN_MULT_PLUS_EXPR
605 || ops->code == WIDEN_MULT_MINUS_EXPR)
606 icode = find_widening_optab_handler (widen_pattern_optab,
607 TYPE_MODE (TREE_TYPE (ops->op2)),
608 tmode0, 0);
609 else
610 icode = optab_handler (widen_pattern_optab, tmode0);
611 gcc_assert (icode != CODE_FOR_nothing);
612
613 if (nops >= 2)
614 {
615 oprnd1 = ops->op1;
616 tmode1 = TYPE_MODE (TREE_TYPE (oprnd1));
617 }
618
619 /* The last operand is of a wider mode than the rest of the operands. */
620 if (nops == 2)
621 wmode = tmode1;
622 else if (nops == 3)
623 {
624 gcc_assert (tmode1 == tmode0);
625 gcc_assert (op1);
626 oprnd2 = ops->op2;
627 wmode = TYPE_MODE (TREE_TYPE (oprnd2));
628 }
629
630 op = 0;
631 create_output_operand (&eops[op++], target, TYPE_MODE (ops->type));
632 create_convert_operand_from (&eops[op++], op0, tmode0, unsignedp);
633 if (op1)
634 create_convert_operand_from (&eops[op++], op1, tmode1, unsignedp);
635 if (wide_op)
636 create_convert_operand_from (&eops[op++], wide_op, wmode, unsignedp);
637 expand_insn (icode, op, eops);
638 return eops[0].value;
639 }
640
641 /* Generate code to perform an operation specified by TERNARY_OPTAB
642 on operands OP0, OP1 and OP2, with result having machine-mode MODE.
643
644 UNSIGNEDP is for the case where we have to widen the operands
645 to perform the operation. It says to use zero-extension.
646
647 If TARGET is nonzero, the value
648 is generated there, if it is convenient to do so.
649 In all cases an rtx is returned for the locus of the value;
650 this may or may not be TARGET. */
651
652 rtx
653 expand_ternary_op (enum machine_mode mode, optab ternary_optab, rtx op0,
654 rtx op1, rtx op2, rtx target, int unsignedp)
655 {
656 struct expand_operand ops[4];
657 enum insn_code icode = optab_handler (ternary_optab, mode);
658
659 gcc_assert (optab_handler (ternary_optab, mode) != CODE_FOR_nothing);
660
661 create_output_operand (&ops[0], target, mode);
662 create_convert_operand_from (&ops[1], op0, mode, unsignedp);
663 create_convert_operand_from (&ops[2], op1, mode, unsignedp);
664 create_convert_operand_from (&ops[3], op2, mode, unsignedp);
665 expand_insn (icode, 4, ops);
666 return ops[0].value;
667 }
668
669
670 /* Like expand_binop, but return a constant rtx if the result can be
671 calculated at compile time. The arguments and return value are
672 otherwise the same as for expand_binop. */
673
674 static rtx
675 simplify_expand_binop (enum machine_mode mode, optab binoptab,
676 rtx op0, rtx op1, rtx target, int unsignedp,
677 enum optab_methods methods)
678 {
679 if (CONSTANT_P (op0) && CONSTANT_P (op1))
680 {
681 rtx x = simplify_binary_operation (binoptab->code, mode, op0, op1);
682
683 if (x)
684 return x;
685 }
686
687 return expand_binop (mode, binoptab, op0, op1, target, unsignedp, methods);
688 }
689
690 /* Like simplify_expand_binop, but always put the result in TARGET.
691 Return true if the expansion succeeded. */
692
693 bool
694 force_expand_binop (enum machine_mode mode, optab binoptab,
695 rtx op0, rtx op1, rtx target, int unsignedp,
696 enum optab_methods methods)
697 {
698 rtx x = simplify_expand_binop (mode, binoptab, op0, op1,
699 target, unsignedp, methods);
700 if (x == 0)
701 return false;
702 if (x != target)
703 emit_move_insn (target, x);
704 return true;
705 }
706
707 /* Generate insns for VEC_LSHIFT_EXPR, VEC_RSHIFT_EXPR. */
708
709 rtx
710 expand_vec_shift_expr (sepops ops, rtx target)
711 {
712 struct expand_operand eops[3];
713 enum insn_code icode;
714 rtx rtx_op1, rtx_op2;
715 enum machine_mode mode = TYPE_MODE (ops->type);
716 tree vec_oprnd = ops->op0;
717 tree shift_oprnd = ops->op1;
718 optab shift_optab;
719
720 switch (ops->code)
721 {
722 case VEC_RSHIFT_EXPR:
723 shift_optab = vec_shr_optab;
724 break;
725 case VEC_LSHIFT_EXPR:
726 shift_optab = vec_shl_optab;
727 break;
728 default:
729 gcc_unreachable ();
730 }
731
732 icode = optab_handler (shift_optab, mode);
733 gcc_assert (icode != CODE_FOR_nothing);
734
735 rtx_op1 = expand_normal (vec_oprnd);
736 rtx_op2 = expand_normal (shift_oprnd);
737
738 create_output_operand (&eops[0], target, mode);
739 create_input_operand (&eops[1], rtx_op1, GET_MODE (rtx_op1));
740 create_convert_operand_from_type (&eops[2], rtx_op2, TREE_TYPE (shift_oprnd));
741 expand_insn (icode, 3, eops);
742
743 return eops[0].value;
744 }
745
746 /* Create a new vector value in VMODE with all elements set to OP. The
747 mode of OP must be the element mode of VMODE. If OP is a constant,
748 then the return value will be a constant. */
749
750 static rtx
751 expand_vector_broadcast (enum machine_mode vmode, rtx op)
752 {
753 enum insn_code icode;
754 rtvec vec;
755 rtx ret;
756 int i, n;
757
758 gcc_checking_assert (VECTOR_MODE_P (vmode));
759
760 n = GET_MODE_NUNITS (vmode);
761 vec = rtvec_alloc (n);
762 for (i = 0; i < n; ++i)
763 RTVEC_ELT (vec, i) = op;
764
765 if (CONSTANT_P (op))
766 return gen_rtx_CONST_VECTOR (vmode, vec);
767
768 /* ??? If the target doesn't have a vec_init, then we have no easy way
769 of performing this operation. Most of this sort of generic support
770 is hidden away in the vector lowering support in gimple. */
771 icode = optab_handler (vec_init_optab, vmode);
772 if (icode == CODE_FOR_nothing)
773 return NULL;
774
775 ret = gen_reg_rtx (vmode);
776 emit_insn (GEN_FCN (icode) (ret, gen_rtx_PARALLEL (vmode, vec)));
777
778 return ret;
779 }
780
781 /* This subroutine of expand_doubleword_shift handles the cases in which
782 the effective shift value is >= BITS_PER_WORD. The arguments and return
783 value are the same as for the parent routine, except that SUPERWORD_OP1
784 is the shift count to use when shifting OUTOF_INPUT into INTO_TARGET.
785 INTO_TARGET may be null if the caller has decided to calculate it. */
786
787 static bool
788 expand_superword_shift (optab binoptab, rtx outof_input, rtx superword_op1,
789 rtx outof_target, rtx into_target,
790 int unsignedp, enum optab_methods methods)
791 {
792 if (into_target != 0)
793 if (!force_expand_binop (word_mode, binoptab, outof_input, superword_op1,
794 into_target, unsignedp, methods))
795 return false;
796
797 if (outof_target != 0)
798 {
799 /* For a signed right shift, we must fill OUTOF_TARGET with copies
800 of the sign bit, otherwise we must fill it with zeros. */
801 if (binoptab != ashr_optab)
802 emit_move_insn (outof_target, CONST0_RTX (word_mode));
803 else
804 if (!force_expand_binop (word_mode, binoptab,
805 outof_input, GEN_INT (BITS_PER_WORD - 1),
806 outof_target, unsignedp, methods))
807 return false;
808 }
809 return true;
810 }
811
812 /* This subroutine of expand_doubleword_shift handles the cases in which
813 the effective shift value is < BITS_PER_WORD. The arguments and return
814 value are the same as for the parent routine. */
815
816 static bool
817 expand_subword_shift (enum machine_mode op1_mode, optab binoptab,
818 rtx outof_input, rtx into_input, rtx op1,
819 rtx outof_target, rtx into_target,
820 int unsignedp, enum optab_methods methods,
821 unsigned HOST_WIDE_INT shift_mask)
822 {
823 optab reverse_unsigned_shift, unsigned_shift;
824 rtx tmp, carries;
825
826 reverse_unsigned_shift = (binoptab == ashl_optab ? lshr_optab : ashl_optab);
827 unsigned_shift = (binoptab == ashl_optab ? ashl_optab : lshr_optab);
828
829 /* The low OP1 bits of INTO_TARGET come from the high bits of OUTOF_INPUT.
830 We therefore need to shift OUTOF_INPUT by (BITS_PER_WORD - OP1) bits in
831 the opposite direction to BINOPTAB. */
832 if (CONSTANT_P (op1) || shift_mask >= BITS_PER_WORD)
833 {
834 carries = outof_input;
835 tmp = immed_double_const (BITS_PER_WORD, 0, op1_mode);
836 tmp = simplify_expand_binop (op1_mode, sub_optab, tmp, op1,
837 0, true, methods);
838 }
839 else
840 {
841 /* We must avoid shifting by BITS_PER_WORD bits since that is either
842 the same as a zero shift (if shift_mask == BITS_PER_WORD - 1) or
843 has unknown behavior. Do a single shift first, then shift by the
844 remainder. It's OK to use ~OP1 as the remainder if shift counts
845 are truncated to the mode size. */
846 carries = expand_binop (word_mode, reverse_unsigned_shift,
847 outof_input, const1_rtx, 0, unsignedp, methods);
848 if (shift_mask == BITS_PER_WORD - 1)
849 {
850 tmp = immed_double_const (-1, -1, op1_mode);
851 tmp = simplify_expand_binop (op1_mode, xor_optab, op1, tmp,
852 0, true, methods);
853 }
854 else
855 {
856 tmp = immed_double_const (BITS_PER_WORD - 1, 0, op1_mode);
857 tmp = simplify_expand_binop (op1_mode, sub_optab, tmp, op1,
858 0, true, methods);
859 }
860 }
861 if (tmp == 0 || carries == 0)
862 return false;
863 carries = expand_binop (word_mode, reverse_unsigned_shift,
864 carries, tmp, 0, unsignedp, methods);
865 if (carries == 0)
866 return false;
867
868 /* Shift INTO_INPUT logically by OP1. This is the last use of INTO_INPUT
869 so the result can go directly into INTO_TARGET if convenient. */
870 tmp = expand_binop (word_mode, unsigned_shift, into_input, op1,
871 into_target, unsignedp, methods);
872 if (tmp == 0)
873 return false;
874
875 /* Now OR in the bits carried over from OUTOF_INPUT. */
876 if (!force_expand_binop (word_mode, ior_optab, tmp, carries,
877 into_target, unsignedp, methods))
878 return false;
879
880 /* Use a standard word_mode shift for the out-of half. */
881 if (outof_target != 0)
882 if (!force_expand_binop (word_mode, binoptab, outof_input, op1,
883 outof_target, unsignedp, methods))
884 return false;
885
886 return true;
887 }
888
889
890 #ifdef HAVE_conditional_move
891 /* Try implementing expand_doubleword_shift using conditional moves.
892 The shift is by < BITS_PER_WORD if (CMP_CODE CMP1 CMP2) is true,
893 otherwise it is by >= BITS_PER_WORD. SUBWORD_OP1 and SUPERWORD_OP1
894 are the shift counts to use in the former and latter case. All other
895 arguments are the same as the parent routine. */
896
897 static bool
898 expand_doubleword_shift_condmove (enum machine_mode op1_mode, optab binoptab,
899 enum rtx_code cmp_code, rtx cmp1, rtx cmp2,
900 rtx outof_input, rtx into_input,
901 rtx subword_op1, rtx superword_op1,
902 rtx outof_target, rtx into_target,
903 int unsignedp, enum optab_methods methods,
904 unsigned HOST_WIDE_INT shift_mask)
905 {
906 rtx outof_superword, into_superword;
907
908 /* Put the superword version of the output into OUTOF_SUPERWORD and
909 INTO_SUPERWORD. */
910 outof_superword = outof_target != 0 ? gen_reg_rtx (word_mode) : 0;
911 if (outof_target != 0 && subword_op1 == superword_op1)
912 {
913 /* The value INTO_TARGET >> SUBWORD_OP1, which we later store in
914 OUTOF_TARGET, is the same as the value of INTO_SUPERWORD. */
915 into_superword = outof_target;
916 if (!expand_superword_shift (binoptab, outof_input, superword_op1,
917 outof_superword, 0, unsignedp, methods))
918 return false;
919 }
920 else
921 {
922 into_superword = gen_reg_rtx (word_mode);
923 if (!expand_superword_shift (binoptab, outof_input, superword_op1,
924 outof_superword, into_superword,
925 unsignedp, methods))
926 return false;
927 }
928
929 /* Put the subword version directly in OUTOF_TARGET and INTO_TARGET. */
930 if (!expand_subword_shift (op1_mode, binoptab,
931 outof_input, into_input, subword_op1,
932 outof_target, into_target,
933 unsignedp, methods, shift_mask))
934 return false;
935
936 /* Select between them. Do the INTO half first because INTO_SUPERWORD
937 might be the current value of OUTOF_TARGET. */
938 if (!emit_conditional_move (into_target, cmp_code, cmp1, cmp2, op1_mode,
939 into_target, into_superword, word_mode, false))
940 return false;
941
942 if (outof_target != 0)
943 if (!emit_conditional_move (outof_target, cmp_code, cmp1, cmp2, op1_mode,
944 outof_target, outof_superword,
945 word_mode, false))
946 return false;
947
948 return true;
949 }
950 #endif
951
952 /* Expand a doubleword shift (ashl, ashr or lshr) using word-mode shifts.
953 OUTOF_INPUT and INTO_INPUT are the two word-sized halves of the first
954 input operand; the shift moves bits in the direction OUTOF_INPUT->
955 INTO_TARGET. OUTOF_TARGET and INTO_TARGET are the equivalent words
956 of the target. OP1 is the shift count and OP1_MODE is its mode.
957 If OP1 is constant, it will have been truncated as appropriate
958 and is known to be nonzero.
959
960 If SHIFT_MASK is zero, the result of word shifts is undefined when the
961 shift count is outside the range [0, BITS_PER_WORD). This routine must
962 avoid generating such shifts for OP1s in the range [0, BITS_PER_WORD * 2).
963
964 If SHIFT_MASK is nonzero, all word-mode shift counts are effectively
965 masked by it and shifts in the range [BITS_PER_WORD, SHIFT_MASK) will
966 fill with zeros or sign bits as appropriate.
967
968 If SHIFT_MASK is BITS_PER_WORD - 1, this routine will synthesize
969 a doubleword shift whose equivalent mask is BITS_PER_WORD * 2 - 1.
970 Doing this preserves semantics required by SHIFT_COUNT_TRUNCATED.
971 In all other cases, shifts by values outside [0, BITS_PER_UNIT * 2)
972 are undefined.
973
974 BINOPTAB, UNSIGNEDP and METHODS are as for expand_binop. This function
975 may not use INTO_INPUT after modifying INTO_TARGET, and similarly for
976 OUTOF_INPUT and OUTOF_TARGET. OUTOF_TARGET can be null if the parent
977 function wants to calculate it itself.
978
979 Return true if the shift could be successfully synthesized. */
980
981 static bool
982 expand_doubleword_shift (enum machine_mode op1_mode, optab binoptab,
983 rtx outof_input, rtx into_input, rtx op1,
984 rtx outof_target, rtx into_target,
985 int unsignedp, enum optab_methods methods,
986 unsigned HOST_WIDE_INT shift_mask)
987 {
988 rtx superword_op1, tmp, cmp1, cmp2;
989 rtx subword_label, done_label;
990 enum rtx_code cmp_code;
991
992 /* See if word-mode shifts by BITS_PER_WORD...BITS_PER_WORD * 2 - 1 will
993 fill the result with sign or zero bits as appropriate. If so, the value
994 of OUTOF_TARGET will always be (SHIFT OUTOF_INPUT OP1). Recursively call
995 this routine to calculate INTO_TARGET (which depends on both OUTOF_INPUT
996 and INTO_INPUT), then emit code to set up OUTOF_TARGET.
997
998 This isn't worthwhile for constant shifts since the optimizers will
999 cope better with in-range shift counts. */
1000 if (shift_mask >= BITS_PER_WORD
1001 && outof_target != 0
1002 && !CONSTANT_P (op1))
1003 {
1004 if (!expand_doubleword_shift (op1_mode, binoptab,
1005 outof_input, into_input, op1,
1006 0, into_target,
1007 unsignedp, methods, shift_mask))
1008 return false;
1009 if (!force_expand_binop (word_mode, binoptab, outof_input, op1,
1010 outof_target, unsignedp, methods))
1011 return false;
1012 return true;
1013 }
1014
1015 /* Set CMP_CODE, CMP1 and CMP2 so that the rtx (CMP_CODE CMP1 CMP2)
1016 is true when the effective shift value is less than BITS_PER_WORD.
1017 Set SUPERWORD_OP1 to the shift count that should be used to shift
1018 OUTOF_INPUT into INTO_TARGET when the condition is false. */
1019 tmp = immed_double_const (BITS_PER_WORD, 0, op1_mode);
1020 if (!CONSTANT_P (op1) && shift_mask == BITS_PER_WORD - 1)
1021 {
1022 /* Set CMP1 to OP1 & BITS_PER_WORD. The result is zero iff OP1
1023 is a subword shift count. */
1024 cmp1 = simplify_expand_binop (op1_mode, and_optab, op1, tmp,
1025 0, true, methods);
1026 cmp2 = CONST0_RTX (op1_mode);
1027 cmp_code = EQ;
1028 superword_op1 = op1;
1029 }
1030 else
1031 {
1032 /* Set CMP1 to OP1 - BITS_PER_WORD. */
1033 cmp1 = simplify_expand_binop (op1_mode, sub_optab, op1, tmp,
1034 0, true, methods);
1035 cmp2 = CONST0_RTX (op1_mode);
1036 cmp_code = LT;
1037 superword_op1 = cmp1;
1038 }
1039 if (cmp1 == 0)
1040 return false;
1041
1042 /* If we can compute the condition at compile time, pick the
1043 appropriate subroutine. */
1044 tmp = simplify_relational_operation (cmp_code, SImode, op1_mode, cmp1, cmp2);
1045 if (tmp != 0 && CONST_INT_P (tmp))
1046 {
1047 if (tmp == const0_rtx)
1048 return expand_superword_shift (binoptab, outof_input, superword_op1,
1049 outof_target, into_target,
1050 unsignedp, methods);
1051 else
1052 return expand_subword_shift (op1_mode, binoptab,
1053 outof_input, into_input, op1,
1054 outof_target, into_target,
1055 unsignedp, methods, shift_mask);
1056 }
1057
1058 #ifdef HAVE_conditional_move
1059 /* Try using conditional moves to generate straight-line code. */
1060 {
1061 rtx start = get_last_insn ();
1062 if (expand_doubleword_shift_condmove (op1_mode, binoptab,
1063 cmp_code, cmp1, cmp2,
1064 outof_input, into_input,
1065 op1, superword_op1,
1066 outof_target, into_target,
1067 unsignedp, methods, shift_mask))
1068 return true;
1069 delete_insns_since (start);
1070 }
1071 #endif
1072
1073 /* As a last resort, use branches to select the correct alternative. */
1074 subword_label = gen_label_rtx ();
1075 done_label = gen_label_rtx ();
1076
1077 NO_DEFER_POP;
1078 do_compare_rtx_and_jump (cmp1, cmp2, cmp_code, false, op1_mode,
1079 0, 0, subword_label, -1);
1080 OK_DEFER_POP;
1081
1082 if (!expand_superword_shift (binoptab, outof_input, superword_op1,
1083 outof_target, into_target,
1084 unsignedp, methods))
1085 return false;
1086
1087 emit_jump_insn (gen_jump (done_label));
1088 emit_barrier ();
1089 emit_label (subword_label);
1090
1091 if (!expand_subword_shift (op1_mode, binoptab,
1092 outof_input, into_input, op1,
1093 outof_target, into_target,
1094 unsignedp, methods, shift_mask))
1095 return false;
1096
1097 emit_label (done_label);
1098 return true;
1099 }
1100 \f
1101 /* Subroutine of expand_binop. Perform a double word multiplication of
1102 operands OP0 and OP1 both of mode MODE, which is exactly twice as wide
1103 as the target's word_mode. This function return NULL_RTX if anything
1104 goes wrong, in which case it may have already emitted instructions
1105 which need to be deleted.
1106
1107 If we want to multiply two two-word values and have normal and widening
1108 multiplies of single-word values, we can do this with three smaller
1109 multiplications.
1110
1111 The multiplication proceeds as follows:
1112 _______________________
1113 [__op0_high_|__op0_low__]
1114 _______________________
1115 * [__op1_high_|__op1_low__]
1116 _______________________________________________
1117 _______________________
1118 (1) [__op0_low__*__op1_low__]
1119 _______________________
1120 (2a) [__op0_low__*__op1_high_]
1121 _______________________
1122 (2b) [__op0_high_*__op1_low__]
1123 _______________________
1124 (3) [__op0_high_*__op1_high_]
1125
1126
1127 This gives a 4-word result. Since we are only interested in the
1128 lower 2 words, partial result (3) and the upper words of (2a) and
1129 (2b) don't need to be calculated. Hence (2a) and (2b) can be
1130 calculated using non-widening multiplication.
1131
1132 (1), however, needs to be calculated with an unsigned widening
1133 multiplication. If this operation is not directly supported we
1134 try using a signed widening multiplication and adjust the result.
1135 This adjustment works as follows:
1136
1137 If both operands are positive then no adjustment is needed.
1138
1139 If the operands have different signs, for example op0_low < 0 and
1140 op1_low >= 0, the instruction treats the most significant bit of
1141 op0_low as a sign bit instead of a bit with significance
1142 2**(BITS_PER_WORD-1), i.e. the instruction multiplies op1_low
1143 with 2**BITS_PER_WORD - op0_low, and two's complements the
1144 result. Conclusion: We need to add op1_low * 2**BITS_PER_WORD to
1145 the result.
1146
1147 Similarly, if both operands are negative, we need to add
1148 (op0_low + op1_low) * 2**BITS_PER_WORD.
1149
1150 We use a trick to adjust quickly. We logically shift op0_low right
1151 (op1_low) BITS_PER_WORD-1 steps to get 0 or 1, and add this to
1152 op0_high (op1_high) before it is used to calculate 2b (2a). If no
1153 logical shift exists, we do an arithmetic right shift and subtract
1154 the 0 or -1. */
1155
1156 static rtx
1157 expand_doubleword_mult (enum machine_mode mode, rtx op0, rtx op1, rtx target,
1158 bool umulp, enum optab_methods methods)
1159 {
1160 int low = (WORDS_BIG_ENDIAN ? 1 : 0);
1161 int high = (WORDS_BIG_ENDIAN ? 0 : 1);
1162 rtx wordm1 = umulp ? NULL_RTX : GEN_INT (BITS_PER_WORD - 1);
1163 rtx product, adjust, product_high, temp;
1164
1165 rtx op0_high = operand_subword_force (op0, high, mode);
1166 rtx op0_low = operand_subword_force (op0, low, mode);
1167 rtx op1_high = operand_subword_force (op1, high, mode);
1168 rtx op1_low = operand_subword_force (op1, low, mode);
1169
1170 /* If we're using an unsigned multiply to directly compute the product
1171 of the low-order words of the operands and perform any required
1172 adjustments of the operands, we begin by trying two more multiplications
1173 and then computing the appropriate sum.
1174
1175 We have checked above that the required addition is provided.
1176 Full-word addition will normally always succeed, especially if
1177 it is provided at all, so we don't worry about its failure. The
1178 multiplication may well fail, however, so we do handle that. */
1179
1180 if (!umulp)
1181 {
1182 /* ??? This could be done with emit_store_flag where available. */
1183 temp = expand_binop (word_mode, lshr_optab, op0_low, wordm1,
1184 NULL_RTX, 1, methods);
1185 if (temp)
1186 op0_high = expand_binop (word_mode, add_optab, op0_high, temp,
1187 NULL_RTX, 0, OPTAB_DIRECT);
1188 else
1189 {
1190 temp = expand_binop (word_mode, ashr_optab, op0_low, wordm1,
1191 NULL_RTX, 0, methods);
1192 if (!temp)
1193 return NULL_RTX;
1194 op0_high = expand_binop (word_mode, sub_optab, op0_high, temp,
1195 NULL_RTX, 0, OPTAB_DIRECT);
1196 }
1197
1198 if (!op0_high)
1199 return NULL_RTX;
1200 }
1201
1202 adjust = expand_binop (word_mode, smul_optab, op0_high, op1_low,
1203 NULL_RTX, 0, OPTAB_DIRECT);
1204 if (!adjust)
1205 return NULL_RTX;
1206
1207 /* OP0_HIGH should now be dead. */
1208
1209 if (!umulp)
1210 {
1211 /* ??? This could be done with emit_store_flag where available. */
1212 temp = expand_binop (word_mode, lshr_optab, op1_low, wordm1,
1213 NULL_RTX, 1, methods);
1214 if (temp)
1215 op1_high = expand_binop (word_mode, add_optab, op1_high, temp,
1216 NULL_RTX, 0, OPTAB_DIRECT);
1217 else
1218 {
1219 temp = expand_binop (word_mode, ashr_optab, op1_low, wordm1,
1220 NULL_RTX, 0, methods);
1221 if (!temp)
1222 return NULL_RTX;
1223 op1_high = expand_binop (word_mode, sub_optab, op1_high, temp,
1224 NULL_RTX, 0, OPTAB_DIRECT);
1225 }
1226
1227 if (!op1_high)
1228 return NULL_RTX;
1229 }
1230
1231 temp = expand_binop (word_mode, smul_optab, op1_high, op0_low,
1232 NULL_RTX, 0, OPTAB_DIRECT);
1233 if (!temp)
1234 return NULL_RTX;
1235
1236 /* OP1_HIGH should now be dead. */
1237
1238 adjust = expand_binop (word_mode, add_optab, adjust, temp,
1239 NULL_RTX, 0, OPTAB_DIRECT);
1240
1241 if (target && !REG_P (target))
1242 target = NULL_RTX;
1243
1244 if (umulp)
1245 product = expand_binop (mode, umul_widen_optab, op0_low, op1_low,
1246 target, 1, OPTAB_DIRECT);
1247 else
1248 product = expand_binop (mode, smul_widen_optab, op0_low, op1_low,
1249 target, 1, OPTAB_DIRECT);
1250
1251 if (!product)
1252 return NULL_RTX;
1253
1254 product_high = operand_subword (product, high, 1, mode);
1255 adjust = expand_binop (word_mode, add_optab, product_high, adjust,
1256 NULL_RTX, 0, OPTAB_DIRECT);
1257 emit_move_insn (product_high, adjust);
1258 return product;
1259 }
1260 \f
1261 /* Wrapper around expand_binop which takes an rtx code to specify
1262 the operation to perform, not an optab pointer. All other
1263 arguments are the same. */
1264 rtx
1265 expand_simple_binop (enum machine_mode mode, enum rtx_code code, rtx op0,
1266 rtx op1, rtx target, int unsignedp,
1267 enum optab_methods methods)
1268 {
1269 optab binop = code_to_optab[(int) code];
1270 gcc_assert (binop);
1271
1272 return expand_binop (mode, binop, op0, op1, target, unsignedp, methods);
1273 }
1274
1275 /* Return whether OP0 and OP1 should be swapped when expanding a commutative
1276 binop. Order them according to commutative_operand_precedence and, if
1277 possible, try to put TARGET or a pseudo first. */
1278 static bool
1279 swap_commutative_operands_with_target (rtx target, rtx op0, rtx op1)
1280 {
1281 int op0_prec = commutative_operand_precedence (op0);
1282 int op1_prec = commutative_operand_precedence (op1);
1283
1284 if (op0_prec < op1_prec)
1285 return true;
1286
1287 if (op0_prec > op1_prec)
1288 return false;
1289
1290 /* With equal precedence, both orders are ok, but it is better if the
1291 first operand is TARGET, or if both TARGET and OP0 are pseudos. */
1292 if (target == 0 || REG_P (target))
1293 return (REG_P (op1) && !REG_P (op0)) || target == op1;
1294 else
1295 return rtx_equal_p (op1, target);
1296 }
1297
1298 /* Return true if BINOPTAB implements a shift operation. */
1299
1300 static bool
1301 shift_optab_p (optab binoptab)
1302 {
1303 switch (binoptab->code)
1304 {
1305 case ASHIFT:
1306 case SS_ASHIFT:
1307 case US_ASHIFT:
1308 case ASHIFTRT:
1309 case LSHIFTRT:
1310 case ROTATE:
1311 case ROTATERT:
1312 return true;
1313
1314 default:
1315 return false;
1316 }
1317 }
1318
1319 /* Return true if BINOPTAB implements a commutative binary operation. */
1320
1321 static bool
1322 commutative_optab_p (optab binoptab)
1323 {
1324 return (GET_RTX_CLASS (binoptab->code) == RTX_COMM_ARITH
1325 || binoptab == smul_widen_optab
1326 || binoptab == umul_widen_optab
1327 || binoptab == smul_highpart_optab
1328 || binoptab == umul_highpart_optab);
1329 }
1330
1331 /* X is to be used in mode MODE as operand OPN to BINOPTAB. If we're
1332 optimizing, and if the operand is a constant that costs more than
1333 1 instruction, force the constant into a register and return that
1334 register. Return X otherwise. UNSIGNEDP says whether X is unsigned. */
1335
1336 static rtx
1337 avoid_expensive_constant (enum machine_mode mode, optab binoptab,
1338 int opn, rtx x, bool unsignedp)
1339 {
1340 bool speed = optimize_insn_for_speed_p ();
1341
1342 if (mode != VOIDmode
1343 && optimize
1344 && CONSTANT_P (x)
1345 && rtx_cost (x, binoptab->code, opn, speed) > set_src_cost (x, speed))
1346 {
1347 if (CONST_INT_P (x))
1348 {
1349 HOST_WIDE_INT intval = trunc_int_for_mode (INTVAL (x), mode);
1350 if (intval != INTVAL (x))
1351 x = GEN_INT (intval);
1352 }
1353 else
1354 x = convert_modes (mode, VOIDmode, x, unsignedp);
1355 x = force_reg (mode, x);
1356 }
1357 return x;
1358 }
1359
1360 /* Helper function for expand_binop: handle the case where there
1361 is an insn that directly implements the indicated operation.
1362 Returns null if this is not possible. */
1363 static rtx
1364 expand_binop_directly (enum machine_mode mode, optab binoptab,
1365 rtx op0, rtx op1,
1366 rtx target, int unsignedp, enum optab_methods methods,
1367 rtx last)
1368 {
1369 enum machine_mode from_mode = widened_mode (mode, op0, op1);
1370 enum insn_code icode = find_widening_optab_handler (binoptab, mode,
1371 from_mode, 1);
1372 enum machine_mode xmode0 = insn_data[(int) icode].operand[1].mode;
1373 enum machine_mode xmode1 = insn_data[(int) icode].operand[2].mode;
1374 enum machine_mode mode0, mode1, tmp_mode;
1375 struct expand_operand ops[3];
1376 bool commutative_p;
1377 rtx pat;
1378 rtx xop0 = op0, xop1 = op1;
1379 rtx swap;
1380
1381 /* If it is a commutative operator and the modes would match
1382 if we would swap the operands, we can save the conversions. */
1383 commutative_p = commutative_optab_p (binoptab);
1384 if (commutative_p
1385 && GET_MODE (xop0) != xmode0 && GET_MODE (xop1) != xmode1
1386 && GET_MODE (xop0) == xmode1 && GET_MODE (xop1) == xmode1)
1387 {
1388 swap = xop0;
1389 xop0 = xop1;
1390 xop1 = swap;
1391 }
1392
1393 /* If we are optimizing, force expensive constants into a register. */
1394 xop0 = avoid_expensive_constant (xmode0, binoptab, 0, xop0, unsignedp);
1395 if (!shift_optab_p (binoptab))
1396 xop1 = avoid_expensive_constant (xmode1, binoptab, 1, xop1, unsignedp);
1397
1398 /* In case the insn wants input operands in modes different from
1399 those of the actual operands, convert the operands. It would
1400 seem that we don't need to convert CONST_INTs, but we do, so
1401 that they're properly zero-extended, sign-extended or truncated
1402 for their mode. */
1403
1404 mode0 = GET_MODE (xop0) != VOIDmode ? GET_MODE (xop0) : mode;
1405 if (xmode0 != VOIDmode && xmode0 != mode0)
1406 {
1407 xop0 = convert_modes (xmode0, mode0, xop0, unsignedp);
1408 mode0 = xmode0;
1409 }
1410
1411 mode1 = GET_MODE (xop1) != VOIDmode ? GET_MODE (xop1) : mode;
1412 if (xmode1 != VOIDmode && xmode1 != mode1)
1413 {
1414 xop1 = convert_modes (xmode1, mode1, xop1, unsignedp);
1415 mode1 = xmode1;
1416 }
1417
1418 /* If operation is commutative,
1419 try to make the first operand a register.
1420 Even better, try to make it the same as the target.
1421 Also try to make the last operand a constant. */
1422 if (commutative_p
1423 && swap_commutative_operands_with_target (target, xop0, xop1))
1424 {
1425 swap = xop1;
1426 xop1 = xop0;
1427 xop0 = swap;
1428 }
1429
1430 /* Now, if insn's predicates don't allow our operands, put them into
1431 pseudo regs. */
1432
1433 if (binoptab == vec_pack_trunc_optab
1434 || binoptab == vec_pack_usat_optab
1435 || binoptab == vec_pack_ssat_optab
1436 || binoptab == vec_pack_ufix_trunc_optab
1437 || binoptab == vec_pack_sfix_trunc_optab)
1438 {
1439 /* The mode of the result is different then the mode of the
1440 arguments. */
1441 tmp_mode = insn_data[(int) icode].operand[0].mode;
1442 if (GET_MODE_NUNITS (tmp_mode) != 2 * GET_MODE_NUNITS (mode))
1443 {
1444 delete_insns_since (last);
1445 return NULL_RTX;
1446 }
1447 }
1448 else
1449 tmp_mode = mode;
1450
1451 create_output_operand (&ops[0], target, tmp_mode);
1452 create_input_operand (&ops[1], xop0, mode0);
1453 create_input_operand (&ops[2], xop1, mode1);
1454 pat = maybe_gen_insn (icode, 3, ops);
1455 if (pat)
1456 {
1457 /* If PAT is composed of more than one insn, try to add an appropriate
1458 REG_EQUAL note to it. If we can't because TEMP conflicts with an
1459 operand, call expand_binop again, this time without a target. */
1460 if (INSN_P (pat) && NEXT_INSN (pat) != NULL_RTX
1461 && ! add_equal_note (pat, ops[0].value, binoptab->code,
1462 ops[1].value, ops[2].value))
1463 {
1464 delete_insns_since (last);
1465 return expand_binop (mode, binoptab, op0, op1, NULL_RTX,
1466 unsignedp, methods);
1467 }
1468
1469 emit_insn (pat);
1470 return ops[0].value;
1471 }
1472 delete_insns_since (last);
1473 return NULL_RTX;
1474 }
1475
1476 /* Generate code to perform an operation specified by BINOPTAB
1477 on operands OP0 and OP1, with result having machine-mode MODE.
1478
1479 UNSIGNEDP is for the case where we have to widen the operands
1480 to perform the operation. It says to use zero-extension.
1481
1482 If TARGET is nonzero, the value
1483 is generated there, if it is convenient to do so.
1484 In all cases an rtx is returned for the locus of the value;
1485 this may or may not be TARGET. */
1486
1487 rtx
1488 expand_binop (enum machine_mode mode, optab binoptab, rtx op0, rtx op1,
1489 rtx target, int unsignedp, enum optab_methods methods)
1490 {
1491 enum optab_methods next_methods
1492 = (methods == OPTAB_LIB || methods == OPTAB_LIB_WIDEN
1493 ? OPTAB_WIDEN : methods);
1494 enum mode_class mclass;
1495 enum machine_mode wider_mode;
1496 rtx libfunc;
1497 rtx temp;
1498 rtx entry_last = get_last_insn ();
1499 rtx last;
1500
1501 mclass = GET_MODE_CLASS (mode);
1502
1503 /* If subtracting an integer constant, convert this into an addition of
1504 the negated constant. */
1505
1506 if (binoptab == sub_optab && CONST_INT_P (op1))
1507 {
1508 op1 = negate_rtx (mode, op1);
1509 binoptab = add_optab;
1510 }
1511
1512 /* Record where to delete back to if we backtrack. */
1513 last = get_last_insn ();
1514
1515 /* If we can do it with a three-operand insn, do so. */
1516
1517 if (methods != OPTAB_MUST_WIDEN
1518 && find_widening_optab_handler (binoptab, mode,
1519 widened_mode (mode, op0, op1), 1)
1520 != CODE_FOR_nothing)
1521 {
1522 temp = expand_binop_directly (mode, binoptab, op0, op1, target,
1523 unsignedp, methods, last);
1524 if (temp)
1525 return temp;
1526 }
1527
1528 /* If we were trying to rotate, and that didn't work, try rotating
1529 the other direction before falling back to shifts and bitwise-or. */
1530 if (((binoptab == rotl_optab
1531 && optab_handler (rotr_optab, mode) != CODE_FOR_nothing)
1532 || (binoptab == rotr_optab
1533 && optab_handler (rotl_optab, mode) != CODE_FOR_nothing))
1534 && mclass == MODE_INT)
1535 {
1536 optab otheroptab = (binoptab == rotl_optab ? rotr_optab : rotl_optab);
1537 rtx newop1;
1538 unsigned int bits = GET_MODE_PRECISION (mode);
1539
1540 if (CONST_INT_P (op1))
1541 newop1 = GEN_INT (bits - INTVAL (op1));
1542 else if (targetm.shift_truncation_mask (mode) == bits - 1)
1543 newop1 = negate_rtx (GET_MODE (op1), op1);
1544 else
1545 newop1 = expand_binop (GET_MODE (op1), sub_optab,
1546 GEN_INT (bits), op1,
1547 NULL_RTX, unsignedp, OPTAB_DIRECT);
1548
1549 temp = expand_binop_directly (mode, otheroptab, op0, newop1,
1550 target, unsignedp, methods, last);
1551 if (temp)
1552 return temp;
1553 }
1554
1555 /* If this is a multiply, see if we can do a widening operation that
1556 takes operands of this mode and makes a wider mode. */
1557
1558 if (binoptab == smul_optab
1559 && GET_MODE_2XWIDER_MODE (mode) != VOIDmode
1560 && (widening_optab_handler ((unsignedp ? umul_widen_optab
1561 : smul_widen_optab),
1562 GET_MODE_2XWIDER_MODE (mode), mode)
1563 != CODE_FOR_nothing))
1564 {
1565 temp = expand_binop (GET_MODE_2XWIDER_MODE (mode),
1566 unsignedp ? umul_widen_optab : smul_widen_optab,
1567 op0, op1, NULL_RTX, unsignedp, OPTAB_DIRECT);
1568
1569 if (temp != 0)
1570 {
1571 if (GET_MODE_CLASS (mode) == MODE_INT
1572 && TRULY_NOOP_TRUNCATION_MODES_P (mode, GET_MODE (temp)))
1573 return gen_lowpart (mode, temp);
1574 else
1575 return convert_to_mode (mode, temp, unsignedp);
1576 }
1577 }
1578
1579 /* If this is a vector shift by a scalar, see if we can do a vector
1580 shift by a vector. If so, broadcast the scalar into a vector. */
1581 if (mclass == MODE_VECTOR_INT)
1582 {
1583 optab otheroptab = NULL;
1584
1585 if (binoptab == ashl_optab)
1586 otheroptab = vashl_optab;
1587 else if (binoptab == ashr_optab)
1588 otheroptab = vashr_optab;
1589 else if (binoptab == lshr_optab)
1590 otheroptab = vlshr_optab;
1591 else if (binoptab == rotl_optab)
1592 otheroptab = vrotl_optab;
1593 else if (binoptab == rotr_optab)
1594 otheroptab = vrotr_optab;
1595
1596 if (otheroptab && optab_handler (otheroptab, mode) != CODE_FOR_nothing)
1597 {
1598 rtx vop1 = expand_vector_broadcast (mode, op1);
1599 if (vop1)
1600 {
1601 temp = expand_binop_directly (mode, otheroptab, op0, vop1,
1602 target, unsignedp, methods, last);
1603 if (temp)
1604 return temp;
1605 }
1606 }
1607 }
1608
1609 /* Certain vector operations can be implemented with vector permutation. */
1610 if (VECTOR_MODE_P (mode))
1611 {
1612 enum tree_code tcode = ERROR_MARK;
1613 rtx sel;
1614
1615 if (binoptab == vec_interleave_high_optab)
1616 tcode = VEC_INTERLEAVE_HIGH_EXPR;
1617 else if (binoptab == vec_interleave_low_optab)
1618 tcode = VEC_INTERLEAVE_LOW_EXPR;
1619 else if (binoptab == vec_extract_even_optab)
1620 tcode = VEC_EXTRACT_EVEN_EXPR;
1621 else if (binoptab == vec_extract_odd_optab)
1622 tcode = VEC_EXTRACT_ODD_EXPR;
1623
1624 if (tcode != ERROR_MARK
1625 && can_vec_perm_for_code_p (tcode, mode, &sel))
1626 {
1627 temp = expand_vec_perm (mode, op0, op1, sel, target);
1628 gcc_assert (temp != NULL);
1629 return temp;
1630 }
1631 }
1632
1633 /* Look for a wider mode of the same class for which we think we
1634 can open-code the operation. Check for a widening multiply at the
1635 wider mode as well. */
1636
1637 if (CLASS_HAS_WIDER_MODES_P (mclass)
1638 && methods != OPTAB_DIRECT && methods != OPTAB_LIB)
1639 for (wider_mode = GET_MODE_WIDER_MODE (mode);
1640 wider_mode != VOIDmode;
1641 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
1642 {
1643 if (optab_handler (binoptab, wider_mode) != CODE_FOR_nothing
1644 || (binoptab == smul_optab
1645 && GET_MODE_WIDER_MODE (wider_mode) != VOIDmode
1646 && (find_widening_optab_handler ((unsignedp
1647 ? umul_widen_optab
1648 : smul_widen_optab),
1649 GET_MODE_WIDER_MODE (wider_mode),
1650 mode, 0)
1651 != CODE_FOR_nothing)))
1652 {
1653 rtx xop0 = op0, xop1 = op1;
1654 int no_extend = 0;
1655
1656 /* For certain integer operations, we need not actually extend
1657 the narrow operands, as long as we will truncate
1658 the results to the same narrowness. */
1659
1660 if ((binoptab == ior_optab || binoptab == and_optab
1661 || binoptab == xor_optab
1662 || binoptab == add_optab || binoptab == sub_optab
1663 || binoptab == smul_optab || binoptab == ashl_optab)
1664 && mclass == MODE_INT)
1665 {
1666 no_extend = 1;
1667 xop0 = avoid_expensive_constant (mode, binoptab, 0,
1668 xop0, unsignedp);
1669 if (binoptab != ashl_optab)
1670 xop1 = avoid_expensive_constant (mode, binoptab, 1,
1671 xop1, unsignedp);
1672 }
1673
1674 xop0 = widen_operand (xop0, wider_mode, mode, unsignedp, no_extend);
1675
1676 /* The second operand of a shift must always be extended. */
1677 xop1 = widen_operand (xop1, wider_mode, mode, unsignedp,
1678 no_extend && binoptab != ashl_optab);
1679
1680 temp = expand_binop (wider_mode, binoptab, xop0, xop1, NULL_RTX,
1681 unsignedp, OPTAB_DIRECT);
1682 if (temp)
1683 {
1684 if (mclass != MODE_INT
1685 || !TRULY_NOOP_TRUNCATION_MODES_P (mode, wider_mode))
1686 {
1687 if (target == 0)
1688 target = gen_reg_rtx (mode);
1689 convert_move (target, temp, 0);
1690 return target;
1691 }
1692 else
1693 return gen_lowpart (mode, temp);
1694 }
1695 else
1696 delete_insns_since (last);
1697 }
1698 }
1699
1700 /* If operation is commutative,
1701 try to make the first operand a register.
1702 Even better, try to make it the same as the target.
1703 Also try to make the last operand a constant. */
1704 if (commutative_optab_p (binoptab)
1705 && swap_commutative_operands_with_target (target, op0, op1))
1706 {
1707 temp = op1;
1708 op1 = op0;
1709 op0 = temp;
1710 }
1711
1712 /* These can be done a word at a time. */
1713 if ((binoptab == and_optab || binoptab == ior_optab || binoptab == xor_optab)
1714 && mclass == MODE_INT
1715 && GET_MODE_SIZE (mode) > UNITS_PER_WORD
1716 && optab_handler (binoptab, word_mode) != CODE_FOR_nothing)
1717 {
1718 int i;
1719 rtx insns;
1720
1721 /* If TARGET is the same as one of the operands, the REG_EQUAL note
1722 won't be accurate, so use a new target. */
1723 if (target == 0
1724 || target == op0
1725 || target == op1
1726 || !valid_multiword_target_p (target))
1727 target = gen_reg_rtx (mode);
1728
1729 start_sequence ();
1730
1731 /* Do the actual arithmetic. */
1732 for (i = 0; i < GET_MODE_BITSIZE (mode) / BITS_PER_WORD; i++)
1733 {
1734 rtx target_piece = operand_subword (target, i, 1, mode);
1735 rtx x = expand_binop (word_mode, binoptab,
1736 operand_subword_force (op0, i, mode),
1737 operand_subword_force (op1, i, mode),
1738 target_piece, unsignedp, next_methods);
1739
1740 if (x == 0)
1741 break;
1742
1743 if (target_piece != x)
1744 emit_move_insn (target_piece, x);
1745 }
1746
1747 insns = get_insns ();
1748 end_sequence ();
1749
1750 if (i == GET_MODE_BITSIZE (mode) / BITS_PER_WORD)
1751 {
1752 emit_insn (insns);
1753 return target;
1754 }
1755 }
1756
1757 /* Synthesize double word shifts from single word shifts. */
1758 if ((binoptab == lshr_optab || binoptab == ashl_optab
1759 || binoptab == ashr_optab)
1760 && mclass == MODE_INT
1761 && (CONST_INT_P (op1) || optimize_insn_for_speed_p ())
1762 && GET_MODE_SIZE (mode) == 2 * UNITS_PER_WORD
1763 && GET_MODE_PRECISION (mode) == GET_MODE_BITSIZE (mode)
1764 && optab_handler (binoptab, word_mode) != CODE_FOR_nothing
1765 && optab_handler (ashl_optab, word_mode) != CODE_FOR_nothing
1766 && optab_handler (lshr_optab, word_mode) != CODE_FOR_nothing)
1767 {
1768 unsigned HOST_WIDE_INT shift_mask, double_shift_mask;
1769 enum machine_mode op1_mode;
1770
1771 double_shift_mask = targetm.shift_truncation_mask (mode);
1772 shift_mask = targetm.shift_truncation_mask (word_mode);
1773 op1_mode = GET_MODE (op1) != VOIDmode ? GET_MODE (op1) : word_mode;
1774
1775 /* Apply the truncation to constant shifts. */
1776 if (double_shift_mask > 0 && CONST_INT_P (op1))
1777 op1 = GEN_INT (INTVAL (op1) & double_shift_mask);
1778
1779 if (op1 == CONST0_RTX (op1_mode))
1780 return op0;
1781
1782 /* Make sure that this is a combination that expand_doubleword_shift
1783 can handle. See the comments there for details. */
1784 if (double_shift_mask == 0
1785 || (shift_mask == BITS_PER_WORD - 1
1786 && double_shift_mask == BITS_PER_WORD * 2 - 1))
1787 {
1788 rtx insns;
1789 rtx into_target, outof_target;
1790 rtx into_input, outof_input;
1791 int left_shift, outof_word;
1792
1793 /* If TARGET is the same as one of the operands, the REG_EQUAL note
1794 won't be accurate, so use a new target. */
1795 if (target == 0
1796 || target == op0
1797 || target == op1
1798 || !valid_multiword_target_p (target))
1799 target = gen_reg_rtx (mode);
1800
1801 start_sequence ();
1802
1803 /* OUTOF_* is the word we are shifting bits away from, and
1804 INTO_* is the word that we are shifting bits towards, thus
1805 they differ depending on the direction of the shift and
1806 WORDS_BIG_ENDIAN. */
1807
1808 left_shift = binoptab == ashl_optab;
1809 outof_word = left_shift ^ ! WORDS_BIG_ENDIAN;
1810
1811 outof_target = operand_subword (target, outof_word, 1, mode);
1812 into_target = operand_subword (target, 1 - outof_word, 1, mode);
1813
1814 outof_input = operand_subword_force (op0, outof_word, mode);
1815 into_input = operand_subword_force (op0, 1 - outof_word, mode);
1816
1817 if (expand_doubleword_shift (op1_mode, binoptab,
1818 outof_input, into_input, op1,
1819 outof_target, into_target,
1820 unsignedp, next_methods, shift_mask))
1821 {
1822 insns = get_insns ();
1823 end_sequence ();
1824
1825 emit_insn (insns);
1826 return target;
1827 }
1828 end_sequence ();
1829 }
1830 }
1831
1832 /* Synthesize double word rotates from single word shifts. */
1833 if ((binoptab == rotl_optab || binoptab == rotr_optab)
1834 && mclass == MODE_INT
1835 && CONST_INT_P (op1)
1836 && GET_MODE_PRECISION (mode) == 2 * BITS_PER_WORD
1837 && optab_handler (ashl_optab, word_mode) != CODE_FOR_nothing
1838 && optab_handler (lshr_optab, word_mode) != CODE_FOR_nothing)
1839 {
1840 rtx insns;
1841 rtx into_target, outof_target;
1842 rtx into_input, outof_input;
1843 rtx inter;
1844 int shift_count, left_shift, outof_word;
1845
1846 /* If TARGET is the same as one of the operands, the REG_EQUAL note
1847 won't be accurate, so use a new target. Do this also if target is not
1848 a REG, first because having a register instead may open optimization
1849 opportunities, and second because if target and op0 happen to be MEMs
1850 designating the same location, we would risk clobbering it too early
1851 in the code sequence we generate below. */
1852 if (target == 0
1853 || target == op0
1854 || target == op1
1855 || !REG_P (target)
1856 || !valid_multiword_target_p (target))
1857 target = gen_reg_rtx (mode);
1858
1859 start_sequence ();
1860
1861 shift_count = INTVAL (op1);
1862
1863 /* OUTOF_* is the word we are shifting bits away from, and
1864 INTO_* is the word that we are shifting bits towards, thus
1865 they differ depending on the direction of the shift and
1866 WORDS_BIG_ENDIAN. */
1867
1868 left_shift = (binoptab == rotl_optab);
1869 outof_word = left_shift ^ ! WORDS_BIG_ENDIAN;
1870
1871 outof_target = operand_subword (target, outof_word, 1, mode);
1872 into_target = operand_subword (target, 1 - outof_word, 1, mode);
1873
1874 outof_input = operand_subword_force (op0, outof_word, mode);
1875 into_input = operand_subword_force (op0, 1 - outof_word, mode);
1876
1877 if (shift_count == BITS_PER_WORD)
1878 {
1879 /* This is just a word swap. */
1880 emit_move_insn (outof_target, into_input);
1881 emit_move_insn (into_target, outof_input);
1882 inter = const0_rtx;
1883 }
1884 else
1885 {
1886 rtx into_temp1, into_temp2, outof_temp1, outof_temp2;
1887 rtx first_shift_count, second_shift_count;
1888 optab reverse_unsigned_shift, unsigned_shift;
1889
1890 reverse_unsigned_shift = (left_shift ^ (shift_count < BITS_PER_WORD)
1891 ? lshr_optab : ashl_optab);
1892
1893 unsigned_shift = (left_shift ^ (shift_count < BITS_PER_WORD)
1894 ? ashl_optab : lshr_optab);
1895
1896 if (shift_count > BITS_PER_WORD)
1897 {
1898 first_shift_count = GEN_INT (shift_count - BITS_PER_WORD);
1899 second_shift_count = GEN_INT (2 * BITS_PER_WORD - shift_count);
1900 }
1901 else
1902 {
1903 first_shift_count = GEN_INT (BITS_PER_WORD - shift_count);
1904 second_shift_count = GEN_INT (shift_count);
1905 }
1906
1907 into_temp1 = expand_binop (word_mode, unsigned_shift,
1908 outof_input, first_shift_count,
1909 NULL_RTX, unsignedp, next_methods);
1910 into_temp2 = expand_binop (word_mode, reverse_unsigned_shift,
1911 into_input, second_shift_count,
1912 NULL_RTX, unsignedp, next_methods);
1913
1914 if (into_temp1 != 0 && into_temp2 != 0)
1915 inter = expand_binop (word_mode, ior_optab, into_temp1, into_temp2,
1916 into_target, unsignedp, next_methods);
1917 else
1918 inter = 0;
1919
1920 if (inter != 0 && inter != into_target)
1921 emit_move_insn (into_target, inter);
1922
1923 outof_temp1 = expand_binop (word_mode, unsigned_shift,
1924 into_input, first_shift_count,
1925 NULL_RTX, unsignedp, next_methods);
1926 outof_temp2 = expand_binop (word_mode, reverse_unsigned_shift,
1927 outof_input, second_shift_count,
1928 NULL_RTX, unsignedp, next_methods);
1929
1930 if (inter != 0 && outof_temp1 != 0 && outof_temp2 != 0)
1931 inter = expand_binop (word_mode, ior_optab,
1932 outof_temp1, outof_temp2,
1933 outof_target, unsignedp, next_methods);
1934
1935 if (inter != 0 && inter != outof_target)
1936 emit_move_insn (outof_target, inter);
1937 }
1938
1939 insns = get_insns ();
1940 end_sequence ();
1941
1942 if (inter != 0)
1943 {
1944 emit_insn (insns);
1945 return target;
1946 }
1947 }
1948
1949 /* These can be done a word at a time by propagating carries. */
1950 if ((binoptab == add_optab || binoptab == sub_optab)
1951 && mclass == MODE_INT
1952 && GET_MODE_SIZE (mode) >= 2 * UNITS_PER_WORD
1953 && optab_handler (binoptab, word_mode) != CODE_FOR_nothing)
1954 {
1955 unsigned int i;
1956 optab otheroptab = binoptab == add_optab ? sub_optab : add_optab;
1957 const unsigned int nwords = GET_MODE_BITSIZE (mode) / BITS_PER_WORD;
1958 rtx carry_in = NULL_RTX, carry_out = NULL_RTX;
1959 rtx xop0, xop1, xtarget;
1960
1961 /* We can handle either a 1 or -1 value for the carry. If STORE_FLAG
1962 value is one of those, use it. Otherwise, use 1 since it is the
1963 one easiest to get. */
1964 #if STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1
1965 int normalizep = STORE_FLAG_VALUE;
1966 #else
1967 int normalizep = 1;
1968 #endif
1969
1970 /* Prepare the operands. */
1971 xop0 = force_reg (mode, op0);
1972 xop1 = force_reg (mode, op1);
1973
1974 xtarget = gen_reg_rtx (mode);
1975
1976 if (target == 0 || !REG_P (target) || !valid_multiword_target_p (target))
1977 target = xtarget;
1978
1979 /* Indicate for flow that the entire target reg is being set. */
1980 if (REG_P (target))
1981 emit_clobber (xtarget);
1982
1983 /* Do the actual arithmetic. */
1984 for (i = 0; i < nwords; i++)
1985 {
1986 int index = (WORDS_BIG_ENDIAN ? nwords - i - 1 : i);
1987 rtx target_piece = operand_subword (xtarget, index, 1, mode);
1988 rtx op0_piece = operand_subword_force (xop0, index, mode);
1989 rtx op1_piece = operand_subword_force (xop1, index, mode);
1990 rtx x;
1991
1992 /* Main add/subtract of the input operands. */
1993 x = expand_binop (word_mode, binoptab,
1994 op0_piece, op1_piece,
1995 target_piece, unsignedp, next_methods);
1996 if (x == 0)
1997 break;
1998
1999 if (i + 1 < nwords)
2000 {
2001 /* Store carry from main add/subtract. */
2002 carry_out = gen_reg_rtx (word_mode);
2003 carry_out = emit_store_flag_force (carry_out,
2004 (binoptab == add_optab
2005 ? LT : GT),
2006 x, op0_piece,
2007 word_mode, 1, normalizep);
2008 }
2009
2010 if (i > 0)
2011 {
2012 rtx newx;
2013
2014 /* Add/subtract previous carry to main result. */
2015 newx = expand_binop (word_mode,
2016 normalizep == 1 ? binoptab : otheroptab,
2017 x, carry_in,
2018 NULL_RTX, 1, next_methods);
2019
2020 if (i + 1 < nwords)
2021 {
2022 /* Get out carry from adding/subtracting carry in. */
2023 rtx carry_tmp = gen_reg_rtx (word_mode);
2024 carry_tmp = emit_store_flag_force (carry_tmp,
2025 (binoptab == add_optab
2026 ? LT : GT),
2027 newx, x,
2028 word_mode, 1, normalizep);
2029
2030 /* Logical-ior the two poss. carry together. */
2031 carry_out = expand_binop (word_mode, ior_optab,
2032 carry_out, carry_tmp,
2033 carry_out, 0, next_methods);
2034 if (carry_out == 0)
2035 break;
2036 }
2037 emit_move_insn (target_piece, newx);
2038 }
2039 else
2040 {
2041 if (x != target_piece)
2042 emit_move_insn (target_piece, x);
2043 }
2044
2045 carry_in = carry_out;
2046 }
2047
2048 if (i == GET_MODE_BITSIZE (mode) / (unsigned) BITS_PER_WORD)
2049 {
2050 if (optab_handler (mov_optab, mode) != CODE_FOR_nothing
2051 || ! rtx_equal_p (target, xtarget))
2052 {
2053 rtx temp = emit_move_insn (target, xtarget);
2054
2055 set_unique_reg_note (temp,
2056 REG_EQUAL,
2057 gen_rtx_fmt_ee (binoptab->code, mode,
2058 copy_rtx (xop0),
2059 copy_rtx (xop1)));
2060 }
2061 else
2062 target = xtarget;
2063
2064 return target;
2065 }
2066
2067 else
2068 delete_insns_since (last);
2069 }
2070
2071 /* Attempt to synthesize double word multiplies using a sequence of word
2072 mode multiplications. We first attempt to generate a sequence using a
2073 more efficient unsigned widening multiply, and if that fails we then
2074 try using a signed widening multiply. */
2075
2076 if (binoptab == smul_optab
2077 && mclass == MODE_INT
2078 && GET_MODE_SIZE (mode) == 2 * UNITS_PER_WORD
2079 && optab_handler (smul_optab, word_mode) != CODE_FOR_nothing
2080 && optab_handler (add_optab, word_mode) != CODE_FOR_nothing)
2081 {
2082 rtx product = NULL_RTX;
2083 if (widening_optab_handler (umul_widen_optab, mode, word_mode)
2084 != CODE_FOR_nothing)
2085 {
2086 product = expand_doubleword_mult (mode, op0, op1, target,
2087 true, methods);
2088 if (!product)
2089 delete_insns_since (last);
2090 }
2091
2092 if (product == NULL_RTX
2093 && widening_optab_handler (smul_widen_optab, mode, word_mode)
2094 != CODE_FOR_nothing)
2095 {
2096 product = expand_doubleword_mult (mode, op0, op1, target,
2097 false, methods);
2098 if (!product)
2099 delete_insns_since (last);
2100 }
2101
2102 if (product != NULL_RTX)
2103 {
2104 if (optab_handler (mov_optab, mode) != CODE_FOR_nothing)
2105 {
2106 temp = emit_move_insn (target ? target : product, product);
2107 set_unique_reg_note (temp,
2108 REG_EQUAL,
2109 gen_rtx_fmt_ee (MULT, mode,
2110 copy_rtx (op0),
2111 copy_rtx (op1)));
2112 }
2113 return product;
2114 }
2115 }
2116
2117 /* It can't be open-coded in this mode.
2118 Use a library call if one is available and caller says that's ok. */
2119
2120 libfunc = optab_libfunc (binoptab, mode);
2121 if (libfunc
2122 && (methods == OPTAB_LIB || methods == OPTAB_LIB_WIDEN))
2123 {
2124 rtx insns;
2125 rtx op1x = op1;
2126 enum machine_mode op1_mode = mode;
2127 rtx value;
2128
2129 start_sequence ();
2130
2131 if (shift_optab_p (binoptab))
2132 {
2133 op1_mode = targetm.libgcc_shift_count_mode ();
2134 /* Specify unsigned here,
2135 since negative shift counts are meaningless. */
2136 op1x = convert_to_mode (op1_mode, op1, 1);
2137 }
2138
2139 if (GET_MODE (op0) != VOIDmode
2140 && GET_MODE (op0) != mode)
2141 op0 = convert_to_mode (mode, op0, unsignedp);
2142
2143 /* Pass 1 for NO_QUEUE so we don't lose any increments
2144 if the libcall is cse'd or moved. */
2145 value = emit_library_call_value (libfunc,
2146 NULL_RTX, LCT_CONST, mode, 2,
2147 op0, mode, op1x, op1_mode);
2148
2149 insns = get_insns ();
2150 end_sequence ();
2151
2152 target = gen_reg_rtx (mode);
2153 emit_libcall_block (insns, target, value,
2154 gen_rtx_fmt_ee (binoptab->code, mode, op0, op1));
2155
2156 return target;
2157 }
2158
2159 delete_insns_since (last);
2160
2161 /* It can't be done in this mode. Can we do it in a wider mode? */
2162
2163 if (! (methods == OPTAB_WIDEN || methods == OPTAB_LIB_WIDEN
2164 || methods == OPTAB_MUST_WIDEN))
2165 {
2166 /* Caller says, don't even try. */
2167 delete_insns_since (entry_last);
2168 return 0;
2169 }
2170
2171 /* Compute the value of METHODS to pass to recursive calls.
2172 Don't allow widening to be tried recursively. */
2173
2174 methods = (methods == OPTAB_LIB_WIDEN ? OPTAB_LIB : OPTAB_DIRECT);
2175
2176 /* Look for a wider mode of the same class for which it appears we can do
2177 the operation. */
2178
2179 if (CLASS_HAS_WIDER_MODES_P (mclass))
2180 {
2181 for (wider_mode = GET_MODE_WIDER_MODE (mode);
2182 wider_mode != VOIDmode;
2183 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
2184 {
2185 if (find_widening_optab_handler (binoptab, wider_mode, mode, 1)
2186 != CODE_FOR_nothing
2187 || (methods == OPTAB_LIB
2188 && optab_libfunc (binoptab, wider_mode)))
2189 {
2190 rtx xop0 = op0, xop1 = op1;
2191 int no_extend = 0;
2192
2193 /* For certain integer operations, we need not actually extend
2194 the narrow operands, as long as we will truncate
2195 the results to the same narrowness. */
2196
2197 if ((binoptab == ior_optab || binoptab == and_optab
2198 || binoptab == xor_optab
2199 || binoptab == add_optab || binoptab == sub_optab
2200 || binoptab == smul_optab || binoptab == ashl_optab)
2201 && mclass == MODE_INT)
2202 no_extend = 1;
2203
2204 xop0 = widen_operand (xop0, wider_mode, mode,
2205 unsignedp, no_extend);
2206
2207 /* The second operand of a shift must always be extended. */
2208 xop1 = widen_operand (xop1, wider_mode, mode, unsignedp,
2209 no_extend && binoptab != ashl_optab);
2210
2211 temp = expand_binop (wider_mode, binoptab, xop0, xop1, NULL_RTX,
2212 unsignedp, methods);
2213 if (temp)
2214 {
2215 if (mclass != MODE_INT
2216 || !TRULY_NOOP_TRUNCATION_MODES_P (mode, wider_mode))
2217 {
2218 if (target == 0)
2219 target = gen_reg_rtx (mode);
2220 convert_move (target, temp, 0);
2221 return target;
2222 }
2223 else
2224 return gen_lowpart (mode, temp);
2225 }
2226 else
2227 delete_insns_since (last);
2228 }
2229 }
2230 }
2231
2232 delete_insns_since (entry_last);
2233 return 0;
2234 }
2235 \f
2236 /* Expand a binary operator which has both signed and unsigned forms.
2237 UOPTAB is the optab for unsigned operations, and SOPTAB is for
2238 signed operations.
2239
2240 If we widen unsigned operands, we may use a signed wider operation instead
2241 of an unsigned wider operation, since the result would be the same. */
2242
2243 rtx
2244 sign_expand_binop (enum machine_mode mode, optab uoptab, optab soptab,
2245 rtx op0, rtx op1, rtx target, int unsignedp,
2246 enum optab_methods methods)
2247 {
2248 rtx temp;
2249 optab direct_optab = unsignedp ? uoptab : soptab;
2250 struct optab_d wide_soptab;
2251
2252 /* Do it without widening, if possible. */
2253 temp = expand_binop (mode, direct_optab, op0, op1, target,
2254 unsignedp, OPTAB_DIRECT);
2255 if (temp || methods == OPTAB_DIRECT)
2256 return temp;
2257
2258 /* Try widening to a signed int. Make a fake signed optab that
2259 hides any signed insn for direct use. */
2260 wide_soptab = *soptab;
2261 set_optab_handler (&wide_soptab, mode, CODE_FOR_nothing);
2262 /* We don't want to generate new hash table entries from this fake
2263 optab. */
2264 wide_soptab.libcall_gen = NULL;
2265
2266 temp = expand_binop (mode, &wide_soptab, op0, op1, target,
2267 unsignedp, OPTAB_WIDEN);
2268
2269 /* For unsigned operands, try widening to an unsigned int. */
2270 if (temp == 0 && unsignedp)
2271 temp = expand_binop (mode, uoptab, op0, op1, target,
2272 unsignedp, OPTAB_WIDEN);
2273 if (temp || methods == OPTAB_WIDEN)
2274 return temp;
2275
2276 /* Use the right width libcall if that exists. */
2277 temp = expand_binop (mode, direct_optab, op0, op1, target, unsignedp, OPTAB_LIB);
2278 if (temp || methods == OPTAB_LIB)
2279 return temp;
2280
2281 /* Must widen and use a libcall, use either signed or unsigned. */
2282 temp = expand_binop (mode, &wide_soptab, op0, op1, target,
2283 unsignedp, methods);
2284 if (temp != 0)
2285 return temp;
2286 if (unsignedp)
2287 return expand_binop (mode, uoptab, op0, op1, target,
2288 unsignedp, methods);
2289 return 0;
2290 }
2291 \f
2292 /* Generate code to perform an operation specified by UNOPPTAB
2293 on operand OP0, with two results to TARG0 and TARG1.
2294 We assume that the order of the operands for the instruction
2295 is TARG0, TARG1, OP0.
2296
2297 Either TARG0 or TARG1 may be zero, but what that means is that
2298 the result is not actually wanted. We will generate it into
2299 a dummy pseudo-reg and discard it. They may not both be zero.
2300
2301 Returns 1 if this operation can be performed; 0 if not. */
2302
2303 int
2304 expand_twoval_unop (optab unoptab, rtx op0, rtx targ0, rtx targ1,
2305 int unsignedp)
2306 {
2307 enum machine_mode mode = GET_MODE (targ0 ? targ0 : targ1);
2308 enum mode_class mclass;
2309 enum machine_mode wider_mode;
2310 rtx entry_last = get_last_insn ();
2311 rtx last;
2312
2313 mclass = GET_MODE_CLASS (mode);
2314
2315 if (!targ0)
2316 targ0 = gen_reg_rtx (mode);
2317 if (!targ1)
2318 targ1 = gen_reg_rtx (mode);
2319
2320 /* Record where to go back to if we fail. */
2321 last = get_last_insn ();
2322
2323 if (optab_handler (unoptab, mode) != CODE_FOR_nothing)
2324 {
2325 struct expand_operand ops[3];
2326 enum insn_code icode = optab_handler (unoptab, mode);
2327
2328 create_fixed_operand (&ops[0], targ0);
2329 create_fixed_operand (&ops[1], targ1);
2330 create_convert_operand_from (&ops[2], op0, mode, unsignedp);
2331 if (maybe_expand_insn (icode, 3, ops))
2332 return 1;
2333 }
2334
2335 /* It can't be done in this mode. Can we do it in a wider mode? */
2336
2337 if (CLASS_HAS_WIDER_MODES_P (mclass))
2338 {
2339 for (wider_mode = GET_MODE_WIDER_MODE (mode);
2340 wider_mode != VOIDmode;
2341 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
2342 {
2343 if (optab_handler (unoptab, wider_mode) != CODE_FOR_nothing)
2344 {
2345 rtx t0 = gen_reg_rtx (wider_mode);
2346 rtx t1 = gen_reg_rtx (wider_mode);
2347 rtx cop0 = convert_modes (wider_mode, mode, op0, unsignedp);
2348
2349 if (expand_twoval_unop (unoptab, cop0, t0, t1, unsignedp))
2350 {
2351 convert_move (targ0, t0, unsignedp);
2352 convert_move (targ1, t1, unsignedp);
2353 return 1;
2354 }
2355 else
2356 delete_insns_since (last);
2357 }
2358 }
2359 }
2360
2361 delete_insns_since (entry_last);
2362 return 0;
2363 }
2364 \f
2365 /* Generate code to perform an operation specified by BINOPTAB
2366 on operands OP0 and OP1, with two results to TARG1 and TARG2.
2367 We assume that the order of the operands for the instruction
2368 is TARG0, OP0, OP1, TARG1, which would fit a pattern like
2369 [(set TARG0 (operate OP0 OP1)) (set TARG1 (operate ...))].
2370
2371 Either TARG0 or TARG1 may be zero, but what that means is that
2372 the result is not actually wanted. We will generate it into
2373 a dummy pseudo-reg and discard it. They may not both be zero.
2374
2375 Returns 1 if this operation can be performed; 0 if not. */
2376
2377 int
2378 expand_twoval_binop (optab binoptab, rtx op0, rtx op1, rtx targ0, rtx targ1,
2379 int unsignedp)
2380 {
2381 enum machine_mode mode = GET_MODE (targ0 ? targ0 : targ1);
2382 enum mode_class mclass;
2383 enum machine_mode wider_mode;
2384 rtx entry_last = get_last_insn ();
2385 rtx last;
2386
2387 mclass = GET_MODE_CLASS (mode);
2388
2389 if (!targ0)
2390 targ0 = gen_reg_rtx (mode);
2391 if (!targ1)
2392 targ1 = gen_reg_rtx (mode);
2393
2394 /* Record where to go back to if we fail. */
2395 last = get_last_insn ();
2396
2397 if (optab_handler (binoptab, mode) != CODE_FOR_nothing)
2398 {
2399 struct expand_operand ops[4];
2400 enum insn_code icode = optab_handler (binoptab, mode);
2401 enum machine_mode mode0 = insn_data[icode].operand[1].mode;
2402 enum machine_mode mode1 = insn_data[icode].operand[2].mode;
2403 rtx xop0 = op0, xop1 = op1;
2404
2405 /* If we are optimizing, force expensive constants into a register. */
2406 xop0 = avoid_expensive_constant (mode0, binoptab, 0, xop0, unsignedp);
2407 xop1 = avoid_expensive_constant (mode1, binoptab, 1, xop1, unsignedp);
2408
2409 create_fixed_operand (&ops[0], targ0);
2410 create_convert_operand_from (&ops[1], op0, mode, unsignedp);
2411 create_convert_operand_from (&ops[2], op1, mode, unsignedp);
2412 create_fixed_operand (&ops[3], targ1);
2413 if (maybe_expand_insn (icode, 4, ops))
2414 return 1;
2415 delete_insns_since (last);
2416 }
2417
2418 /* It can't be done in this mode. Can we do it in a wider mode? */
2419
2420 if (CLASS_HAS_WIDER_MODES_P (mclass))
2421 {
2422 for (wider_mode = GET_MODE_WIDER_MODE (mode);
2423 wider_mode != VOIDmode;
2424 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
2425 {
2426 if (optab_handler (binoptab, wider_mode) != CODE_FOR_nothing)
2427 {
2428 rtx t0 = gen_reg_rtx (wider_mode);
2429 rtx t1 = gen_reg_rtx (wider_mode);
2430 rtx cop0 = convert_modes (wider_mode, mode, op0, unsignedp);
2431 rtx cop1 = convert_modes (wider_mode, mode, op1, unsignedp);
2432
2433 if (expand_twoval_binop (binoptab, cop0, cop1,
2434 t0, t1, unsignedp))
2435 {
2436 convert_move (targ0, t0, unsignedp);
2437 convert_move (targ1, t1, unsignedp);
2438 return 1;
2439 }
2440 else
2441 delete_insns_since (last);
2442 }
2443 }
2444 }
2445
2446 delete_insns_since (entry_last);
2447 return 0;
2448 }
2449
2450 /* Expand the two-valued library call indicated by BINOPTAB, but
2451 preserve only one of the values. If TARG0 is non-NULL, the first
2452 value is placed into TARG0; otherwise the second value is placed
2453 into TARG1. Exactly one of TARG0 and TARG1 must be non-NULL. The
2454 value stored into TARG0 or TARG1 is equivalent to (CODE OP0 OP1).
2455 This routine assumes that the value returned by the library call is
2456 as if the return value was of an integral mode twice as wide as the
2457 mode of OP0. Returns 1 if the call was successful. */
2458
2459 bool
2460 expand_twoval_binop_libfunc (optab binoptab, rtx op0, rtx op1,
2461 rtx targ0, rtx targ1, enum rtx_code code)
2462 {
2463 enum machine_mode mode;
2464 enum machine_mode libval_mode;
2465 rtx libval;
2466 rtx insns;
2467 rtx libfunc;
2468
2469 /* Exactly one of TARG0 or TARG1 should be non-NULL. */
2470 gcc_assert (!targ0 != !targ1);
2471
2472 mode = GET_MODE (op0);
2473 libfunc = optab_libfunc (binoptab, mode);
2474 if (!libfunc)
2475 return false;
2476
2477 /* The value returned by the library function will have twice as
2478 many bits as the nominal MODE. */
2479 libval_mode = smallest_mode_for_size (2 * GET_MODE_BITSIZE (mode),
2480 MODE_INT);
2481 start_sequence ();
2482 libval = emit_library_call_value (libfunc, NULL_RTX, LCT_CONST,
2483 libval_mode, 2,
2484 op0, mode,
2485 op1, mode);
2486 /* Get the part of VAL containing the value that we want. */
2487 libval = simplify_gen_subreg (mode, libval, libval_mode,
2488 targ0 ? 0 : GET_MODE_SIZE (mode));
2489 insns = get_insns ();
2490 end_sequence ();
2491 /* Move the into the desired location. */
2492 emit_libcall_block (insns, targ0 ? targ0 : targ1, libval,
2493 gen_rtx_fmt_ee (code, mode, op0, op1));
2494
2495 return true;
2496 }
2497
2498 \f
2499 /* Wrapper around expand_unop which takes an rtx code to specify
2500 the operation to perform, not an optab pointer. All other
2501 arguments are the same. */
2502 rtx
2503 expand_simple_unop (enum machine_mode mode, enum rtx_code code, rtx op0,
2504 rtx target, int unsignedp)
2505 {
2506 optab unop = code_to_optab[(int) code];
2507 gcc_assert (unop);
2508
2509 return expand_unop (mode, unop, op0, target, unsignedp);
2510 }
2511
2512 /* Try calculating
2513 (clz:narrow x)
2514 as
2515 (clz:wide (zero_extend:wide x)) - ((width wide) - (width narrow)).
2516
2517 A similar operation can be used for clrsb. UNOPTAB says which operation
2518 we are trying to expand. */
2519 static rtx
2520 widen_leading (enum machine_mode mode, rtx op0, rtx target, optab unoptab)
2521 {
2522 enum mode_class mclass = GET_MODE_CLASS (mode);
2523 if (CLASS_HAS_WIDER_MODES_P (mclass))
2524 {
2525 enum machine_mode wider_mode;
2526 for (wider_mode = GET_MODE_WIDER_MODE (mode);
2527 wider_mode != VOIDmode;
2528 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
2529 {
2530 if (optab_handler (unoptab, wider_mode) != CODE_FOR_nothing)
2531 {
2532 rtx xop0, temp, last;
2533
2534 last = get_last_insn ();
2535
2536 if (target == 0)
2537 target = gen_reg_rtx (mode);
2538 xop0 = widen_operand (op0, wider_mode, mode,
2539 unoptab != clrsb_optab, false);
2540 temp = expand_unop (wider_mode, unoptab, xop0, NULL_RTX,
2541 unoptab != clrsb_optab);
2542 if (temp != 0)
2543 temp = expand_binop (wider_mode, sub_optab, temp,
2544 GEN_INT (GET_MODE_PRECISION (wider_mode)
2545 - GET_MODE_PRECISION (mode)),
2546 target, true, OPTAB_DIRECT);
2547 if (temp == 0)
2548 delete_insns_since (last);
2549
2550 return temp;
2551 }
2552 }
2553 }
2554 return 0;
2555 }
2556
2557 /* Try calculating clz of a double-word quantity as two clz's of word-sized
2558 quantities, choosing which based on whether the high word is nonzero. */
2559 static rtx
2560 expand_doubleword_clz (enum machine_mode mode, rtx op0, rtx target)
2561 {
2562 rtx xop0 = force_reg (mode, op0);
2563 rtx subhi = gen_highpart (word_mode, xop0);
2564 rtx sublo = gen_lowpart (word_mode, xop0);
2565 rtx hi0_label = gen_label_rtx ();
2566 rtx after_label = gen_label_rtx ();
2567 rtx seq, temp, result;
2568
2569 /* If we were not given a target, use a word_mode register, not a
2570 'mode' register. The result will fit, and nobody is expecting
2571 anything bigger (the return type of __builtin_clz* is int). */
2572 if (!target)
2573 target = gen_reg_rtx (word_mode);
2574
2575 /* In any case, write to a word_mode scratch in both branches of the
2576 conditional, so we can ensure there is a single move insn setting
2577 'target' to tag a REG_EQUAL note on. */
2578 result = gen_reg_rtx (word_mode);
2579
2580 start_sequence ();
2581
2582 /* If the high word is not equal to zero,
2583 then clz of the full value is clz of the high word. */
2584 emit_cmp_and_jump_insns (subhi, CONST0_RTX (word_mode), EQ, 0,
2585 word_mode, true, hi0_label);
2586
2587 temp = expand_unop_direct (word_mode, clz_optab, subhi, result, true);
2588 if (!temp)
2589 goto fail;
2590
2591 if (temp != result)
2592 convert_move (result, temp, true);
2593
2594 emit_jump_insn (gen_jump (after_label));
2595 emit_barrier ();
2596
2597 /* Else clz of the full value is clz of the low word plus the number
2598 of bits in the high word. */
2599 emit_label (hi0_label);
2600
2601 temp = expand_unop_direct (word_mode, clz_optab, sublo, 0, true);
2602 if (!temp)
2603 goto fail;
2604 temp = expand_binop (word_mode, add_optab, temp,
2605 GEN_INT (GET_MODE_BITSIZE (word_mode)),
2606 result, true, OPTAB_DIRECT);
2607 if (!temp)
2608 goto fail;
2609 if (temp != result)
2610 convert_move (result, temp, true);
2611
2612 emit_label (after_label);
2613 convert_move (target, result, true);
2614
2615 seq = get_insns ();
2616 end_sequence ();
2617
2618 add_equal_note (seq, target, CLZ, xop0, 0);
2619 emit_insn (seq);
2620 return target;
2621
2622 fail:
2623 end_sequence ();
2624 return 0;
2625 }
2626
2627 /* Try calculating
2628 (bswap:narrow x)
2629 as
2630 (lshiftrt:wide (bswap:wide x) ((width wide) - (width narrow))). */
2631 static rtx
2632 widen_bswap (enum machine_mode mode, rtx op0, rtx target)
2633 {
2634 enum mode_class mclass = GET_MODE_CLASS (mode);
2635 enum machine_mode wider_mode;
2636 rtx x, last;
2637
2638 if (!CLASS_HAS_WIDER_MODES_P (mclass))
2639 return NULL_RTX;
2640
2641 for (wider_mode = GET_MODE_WIDER_MODE (mode);
2642 wider_mode != VOIDmode;
2643 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
2644 if (optab_handler (bswap_optab, wider_mode) != CODE_FOR_nothing)
2645 goto found;
2646 return NULL_RTX;
2647
2648 found:
2649 last = get_last_insn ();
2650
2651 x = widen_operand (op0, wider_mode, mode, true, true);
2652 x = expand_unop (wider_mode, bswap_optab, x, NULL_RTX, true);
2653
2654 gcc_assert (GET_MODE_PRECISION (wider_mode) == GET_MODE_BITSIZE (wider_mode)
2655 && GET_MODE_PRECISION (mode) == GET_MODE_BITSIZE (mode));
2656 if (x != 0)
2657 x = expand_shift (RSHIFT_EXPR, wider_mode, x,
2658 GET_MODE_BITSIZE (wider_mode)
2659 - GET_MODE_BITSIZE (mode),
2660 NULL_RTX, true);
2661
2662 if (x != 0)
2663 {
2664 if (target == 0)
2665 target = gen_reg_rtx (mode);
2666 emit_move_insn (target, gen_lowpart (mode, x));
2667 }
2668 else
2669 delete_insns_since (last);
2670
2671 return target;
2672 }
2673
2674 /* Try calculating bswap as two bswaps of two word-sized operands. */
2675
2676 static rtx
2677 expand_doubleword_bswap (enum machine_mode mode, rtx op, rtx target)
2678 {
2679 rtx t0, t1;
2680
2681 t1 = expand_unop (word_mode, bswap_optab,
2682 operand_subword_force (op, 0, mode), NULL_RTX, true);
2683 t0 = expand_unop (word_mode, bswap_optab,
2684 operand_subword_force (op, 1, mode), NULL_RTX, true);
2685
2686 if (target == 0 || !valid_multiword_target_p (target))
2687 target = gen_reg_rtx (mode);
2688 if (REG_P (target))
2689 emit_clobber (target);
2690 emit_move_insn (operand_subword (target, 0, 1, mode), t0);
2691 emit_move_insn (operand_subword (target, 1, 1, mode), t1);
2692
2693 return target;
2694 }
2695
2696 /* Try calculating (parity x) as (and (popcount x) 1), where
2697 popcount can also be done in a wider mode. */
2698 static rtx
2699 expand_parity (enum machine_mode mode, rtx op0, rtx target)
2700 {
2701 enum mode_class mclass = GET_MODE_CLASS (mode);
2702 if (CLASS_HAS_WIDER_MODES_P (mclass))
2703 {
2704 enum machine_mode wider_mode;
2705 for (wider_mode = mode; wider_mode != VOIDmode;
2706 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
2707 {
2708 if (optab_handler (popcount_optab, wider_mode) != CODE_FOR_nothing)
2709 {
2710 rtx xop0, temp, last;
2711
2712 last = get_last_insn ();
2713
2714 if (target == 0)
2715 target = gen_reg_rtx (mode);
2716 xop0 = widen_operand (op0, wider_mode, mode, true, false);
2717 temp = expand_unop (wider_mode, popcount_optab, xop0, NULL_RTX,
2718 true);
2719 if (temp != 0)
2720 temp = expand_binop (wider_mode, and_optab, temp, const1_rtx,
2721 target, true, OPTAB_DIRECT);
2722 if (temp == 0)
2723 delete_insns_since (last);
2724
2725 return temp;
2726 }
2727 }
2728 }
2729 return 0;
2730 }
2731
2732 /* Try calculating ctz(x) as K - clz(x & -x) ,
2733 where K is GET_MODE_PRECISION(mode) - 1.
2734
2735 Both __builtin_ctz and __builtin_clz are undefined at zero, so we
2736 don't have to worry about what the hardware does in that case. (If
2737 the clz instruction produces the usual value at 0, which is K, the
2738 result of this code sequence will be -1; expand_ffs, below, relies
2739 on this. It might be nice to have it be K instead, for consistency
2740 with the (very few) processors that provide a ctz with a defined
2741 value, but that would take one more instruction, and it would be
2742 less convenient for expand_ffs anyway. */
2743
2744 static rtx
2745 expand_ctz (enum machine_mode mode, rtx op0, rtx target)
2746 {
2747 rtx seq, temp;
2748
2749 if (optab_handler (clz_optab, mode) == CODE_FOR_nothing)
2750 return 0;
2751
2752 start_sequence ();
2753
2754 temp = expand_unop_direct (mode, neg_optab, op0, NULL_RTX, true);
2755 if (temp)
2756 temp = expand_binop (mode, and_optab, op0, temp, NULL_RTX,
2757 true, OPTAB_DIRECT);
2758 if (temp)
2759 temp = expand_unop_direct (mode, clz_optab, temp, NULL_RTX, true);
2760 if (temp)
2761 temp = expand_binop (mode, sub_optab, GEN_INT (GET_MODE_PRECISION (mode) - 1),
2762 temp, target,
2763 true, OPTAB_DIRECT);
2764 if (temp == 0)
2765 {
2766 end_sequence ();
2767 return 0;
2768 }
2769
2770 seq = get_insns ();
2771 end_sequence ();
2772
2773 add_equal_note (seq, temp, CTZ, op0, 0);
2774 emit_insn (seq);
2775 return temp;
2776 }
2777
2778
2779 /* Try calculating ffs(x) using ctz(x) if we have that instruction, or
2780 else with the sequence used by expand_clz.
2781
2782 The ffs builtin promises to return zero for a zero value and ctz/clz
2783 may have an undefined value in that case. If they do not give us a
2784 convenient value, we have to generate a test and branch. */
2785 static rtx
2786 expand_ffs (enum machine_mode mode, rtx op0, rtx target)
2787 {
2788 HOST_WIDE_INT val = 0;
2789 bool defined_at_zero = false;
2790 rtx temp, seq;
2791
2792 if (optab_handler (ctz_optab, mode) != CODE_FOR_nothing)
2793 {
2794 start_sequence ();
2795
2796 temp = expand_unop_direct (mode, ctz_optab, op0, 0, true);
2797 if (!temp)
2798 goto fail;
2799
2800 defined_at_zero = (CTZ_DEFINED_VALUE_AT_ZERO (mode, val) == 2);
2801 }
2802 else if (optab_handler (clz_optab, mode) != CODE_FOR_nothing)
2803 {
2804 start_sequence ();
2805 temp = expand_ctz (mode, op0, 0);
2806 if (!temp)
2807 goto fail;
2808
2809 if (CLZ_DEFINED_VALUE_AT_ZERO (mode, val) == 2)
2810 {
2811 defined_at_zero = true;
2812 val = (GET_MODE_PRECISION (mode) - 1) - val;
2813 }
2814 }
2815 else
2816 return 0;
2817
2818 if (defined_at_zero && val == -1)
2819 /* No correction needed at zero. */;
2820 else
2821 {
2822 /* We don't try to do anything clever with the situation found
2823 on some processors (eg Alpha) where ctz(0:mode) ==
2824 bitsize(mode). If someone can think of a way to send N to -1
2825 and leave alone all values in the range 0..N-1 (where N is a
2826 power of two), cheaper than this test-and-branch, please add it.
2827
2828 The test-and-branch is done after the operation itself, in case
2829 the operation sets condition codes that can be recycled for this.
2830 (This is true on i386, for instance.) */
2831
2832 rtx nonzero_label = gen_label_rtx ();
2833 emit_cmp_and_jump_insns (op0, CONST0_RTX (mode), NE, 0,
2834 mode, true, nonzero_label);
2835
2836 convert_move (temp, GEN_INT (-1), false);
2837 emit_label (nonzero_label);
2838 }
2839
2840 /* temp now has a value in the range -1..bitsize-1. ffs is supposed
2841 to produce a value in the range 0..bitsize. */
2842 temp = expand_binop (mode, add_optab, temp, GEN_INT (1),
2843 target, false, OPTAB_DIRECT);
2844 if (!temp)
2845 goto fail;
2846
2847 seq = get_insns ();
2848 end_sequence ();
2849
2850 add_equal_note (seq, temp, FFS, op0, 0);
2851 emit_insn (seq);
2852 return temp;
2853
2854 fail:
2855 end_sequence ();
2856 return 0;
2857 }
2858
2859 /* Extract the OMODE lowpart from VAL, which has IMODE. Under certain
2860 conditions, VAL may already be a SUBREG against which we cannot generate
2861 a further SUBREG. In this case, we expect forcing the value into a
2862 register will work around the situation. */
2863
2864 static rtx
2865 lowpart_subreg_maybe_copy (enum machine_mode omode, rtx val,
2866 enum machine_mode imode)
2867 {
2868 rtx ret;
2869 ret = lowpart_subreg (omode, val, imode);
2870 if (ret == NULL)
2871 {
2872 val = force_reg (imode, val);
2873 ret = lowpart_subreg (omode, val, imode);
2874 gcc_assert (ret != NULL);
2875 }
2876 return ret;
2877 }
2878
2879 /* Expand a floating point absolute value or negation operation via a
2880 logical operation on the sign bit. */
2881
2882 static rtx
2883 expand_absneg_bit (enum rtx_code code, enum machine_mode mode,
2884 rtx op0, rtx target)
2885 {
2886 const struct real_format *fmt;
2887 int bitpos, word, nwords, i;
2888 enum machine_mode imode;
2889 double_int mask;
2890 rtx temp, insns;
2891
2892 /* The format has to have a simple sign bit. */
2893 fmt = REAL_MODE_FORMAT (mode);
2894 if (fmt == NULL)
2895 return NULL_RTX;
2896
2897 bitpos = fmt->signbit_rw;
2898 if (bitpos < 0)
2899 return NULL_RTX;
2900
2901 /* Don't create negative zeros if the format doesn't support them. */
2902 if (code == NEG && !fmt->has_signed_zero)
2903 return NULL_RTX;
2904
2905 if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD)
2906 {
2907 imode = int_mode_for_mode (mode);
2908 if (imode == BLKmode)
2909 return NULL_RTX;
2910 word = 0;
2911 nwords = 1;
2912 }
2913 else
2914 {
2915 imode = word_mode;
2916
2917 if (FLOAT_WORDS_BIG_ENDIAN)
2918 word = (GET_MODE_BITSIZE (mode) - bitpos) / BITS_PER_WORD;
2919 else
2920 word = bitpos / BITS_PER_WORD;
2921 bitpos = bitpos % BITS_PER_WORD;
2922 nwords = (GET_MODE_BITSIZE (mode) + BITS_PER_WORD - 1) / BITS_PER_WORD;
2923 }
2924
2925 mask = double_int_setbit (double_int_zero, bitpos);
2926 if (code == ABS)
2927 mask = double_int_not (mask);
2928
2929 if (target == 0
2930 || target == op0
2931 || (nwords > 1 && !valid_multiword_target_p (target)))
2932 target = gen_reg_rtx (mode);
2933
2934 if (nwords > 1)
2935 {
2936 start_sequence ();
2937
2938 for (i = 0; i < nwords; ++i)
2939 {
2940 rtx targ_piece = operand_subword (target, i, 1, mode);
2941 rtx op0_piece = operand_subword_force (op0, i, mode);
2942
2943 if (i == word)
2944 {
2945 temp = expand_binop (imode, code == ABS ? and_optab : xor_optab,
2946 op0_piece,
2947 immed_double_int_const (mask, imode),
2948 targ_piece, 1, OPTAB_LIB_WIDEN);
2949 if (temp != targ_piece)
2950 emit_move_insn (targ_piece, temp);
2951 }
2952 else
2953 emit_move_insn (targ_piece, op0_piece);
2954 }
2955
2956 insns = get_insns ();
2957 end_sequence ();
2958
2959 emit_insn (insns);
2960 }
2961 else
2962 {
2963 temp = expand_binop (imode, code == ABS ? and_optab : xor_optab,
2964 gen_lowpart (imode, op0),
2965 immed_double_int_const (mask, imode),
2966 gen_lowpart (imode, target), 1, OPTAB_LIB_WIDEN);
2967 target = lowpart_subreg_maybe_copy (mode, temp, imode);
2968
2969 set_unique_reg_note (get_last_insn (), REG_EQUAL,
2970 gen_rtx_fmt_e (code, mode, copy_rtx (op0)));
2971 }
2972
2973 return target;
2974 }
2975
2976 /* As expand_unop, but will fail rather than attempt the operation in a
2977 different mode or with a libcall. */
2978 static rtx
2979 expand_unop_direct (enum machine_mode mode, optab unoptab, rtx op0, rtx target,
2980 int unsignedp)
2981 {
2982 if (optab_handler (unoptab, mode) != CODE_FOR_nothing)
2983 {
2984 struct expand_operand ops[2];
2985 enum insn_code icode = optab_handler (unoptab, mode);
2986 rtx last = get_last_insn ();
2987 rtx pat;
2988
2989 create_output_operand (&ops[0], target, mode);
2990 create_convert_operand_from (&ops[1], op0, mode, unsignedp);
2991 pat = maybe_gen_insn (icode, 2, ops);
2992 if (pat)
2993 {
2994 if (INSN_P (pat) && NEXT_INSN (pat) != NULL_RTX
2995 && ! add_equal_note (pat, ops[0].value, unoptab->code,
2996 ops[1].value, NULL_RTX))
2997 {
2998 delete_insns_since (last);
2999 return expand_unop (mode, unoptab, op0, NULL_RTX, unsignedp);
3000 }
3001
3002 emit_insn (pat);
3003
3004 return ops[0].value;
3005 }
3006 }
3007 return 0;
3008 }
3009
3010 /* Generate code to perform an operation specified by UNOPTAB
3011 on operand OP0, with result having machine-mode MODE.
3012
3013 UNSIGNEDP is for the case where we have to widen the operands
3014 to perform the operation. It says to use zero-extension.
3015
3016 If TARGET is nonzero, the value
3017 is generated there, if it is convenient to do so.
3018 In all cases an rtx is returned for the locus of the value;
3019 this may or may not be TARGET. */
3020
3021 rtx
3022 expand_unop (enum machine_mode mode, optab unoptab, rtx op0, rtx target,
3023 int unsignedp)
3024 {
3025 enum mode_class mclass = GET_MODE_CLASS (mode);
3026 enum machine_mode wider_mode;
3027 rtx temp;
3028 rtx libfunc;
3029
3030 temp = expand_unop_direct (mode, unoptab, op0, target, unsignedp);
3031 if (temp)
3032 return temp;
3033
3034 /* It can't be done in this mode. Can we open-code it in a wider mode? */
3035
3036 /* Widening (or narrowing) clz needs special treatment. */
3037 if (unoptab == clz_optab)
3038 {
3039 temp = widen_leading (mode, op0, target, unoptab);
3040 if (temp)
3041 return temp;
3042
3043 if (GET_MODE_SIZE (mode) == 2 * UNITS_PER_WORD
3044 && optab_handler (unoptab, word_mode) != CODE_FOR_nothing)
3045 {
3046 temp = expand_doubleword_clz (mode, op0, target);
3047 if (temp)
3048 return temp;
3049 }
3050
3051 goto try_libcall;
3052 }
3053
3054 if (unoptab == clrsb_optab)
3055 {
3056 temp = widen_leading (mode, op0, target, unoptab);
3057 if (temp)
3058 return temp;
3059 goto try_libcall;
3060 }
3061
3062 /* Widening (or narrowing) bswap needs special treatment. */
3063 if (unoptab == bswap_optab)
3064 {
3065 temp = widen_bswap (mode, op0, target);
3066 if (temp)
3067 return temp;
3068
3069 if (GET_MODE_SIZE (mode) == 2 * UNITS_PER_WORD
3070 && optab_handler (unoptab, word_mode) != CODE_FOR_nothing)
3071 {
3072 temp = expand_doubleword_bswap (mode, op0, target);
3073 if (temp)
3074 return temp;
3075 }
3076
3077 goto try_libcall;
3078 }
3079
3080 if (CLASS_HAS_WIDER_MODES_P (mclass))
3081 for (wider_mode = GET_MODE_WIDER_MODE (mode);
3082 wider_mode != VOIDmode;
3083 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
3084 {
3085 if (optab_handler (unoptab, wider_mode) != CODE_FOR_nothing)
3086 {
3087 rtx xop0 = op0;
3088 rtx last = get_last_insn ();
3089
3090 /* For certain operations, we need not actually extend
3091 the narrow operand, as long as we will truncate the
3092 results to the same narrowness. */
3093
3094 xop0 = widen_operand (xop0, wider_mode, mode, unsignedp,
3095 (unoptab == neg_optab
3096 || unoptab == one_cmpl_optab)
3097 && mclass == MODE_INT);
3098
3099 temp = expand_unop (wider_mode, unoptab, xop0, NULL_RTX,
3100 unsignedp);
3101
3102 if (temp)
3103 {
3104 if (mclass != MODE_INT
3105 || !TRULY_NOOP_TRUNCATION_MODES_P (mode, wider_mode))
3106 {
3107 if (target == 0)
3108 target = gen_reg_rtx (mode);
3109 convert_move (target, temp, 0);
3110 return target;
3111 }
3112 else
3113 return gen_lowpart (mode, temp);
3114 }
3115 else
3116 delete_insns_since (last);
3117 }
3118 }
3119
3120 /* These can be done a word at a time. */
3121 if (unoptab == one_cmpl_optab
3122 && mclass == MODE_INT
3123 && GET_MODE_SIZE (mode) > UNITS_PER_WORD
3124 && optab_handler (unoptab, word_mode) != CODE_FOR_nothing)
3125 {
3126 int i;
3127 rtx insns;
3128
3129 if (target == 0 || target == op0 || !valid_multiword_target_p (target))
3130 target = gen_reg_rtx (mode);
3131
3132 start_sequence ();
3133
3134 /* Do the actual arithmetic. */
3135 for (i = 0; i < GET_MODE_BITSIZE (mode) / BITS_PER_WORD; i++)
3136 {
3137 rtx target_piece = operand_subword (target, i, 1, mode);
3138 rtx x = expand_unop (word_mode, unoptab,
3139 operand_subword_force (op0, i, mode),
3140 target_piece, unsignedp);
3141
3142 if (target_piece != x)
3143 emit_move_insn (target_piece, x);
3144 }
3145
3146 insns = get_insns ();
3147 end_sequence ();
3148
3149 emit_insn (insns);
3150 return target;
3151 }
3152
3153 if (unoptab->code == NEG)
3154 {
3155 /* Try negating floating point values by flipping the sign bit. */
3156 if (SCALAR_FLOAT_MODE_P (mode))
3157 {
3158 temp = expand_absneg_bit (NEG, mode, op0, target);
3159 if (temp)
3160 return temp;
3161 }
3162
3163 /* If there is no negation pattern, and we have no negative zero,
3164 try subtracting from zero. */
3165 if (!HONOR_SIGNED_ZEROS (mode))
3166 {
3167 temp = expand_binop (mode, (unoptab == negv_optab
3168 ? subv_optab : sub_optab),
3169 CONST0_RTX (mode), op0, target,
3170 unsignedp, OPTAB_DIRECT);
3171 if (temp)
3172 return temp;
3173 }
3174 }
3175
3176 /* Try calculating parity (x) as popcount (x) % 2. */
3177 if (unoptab == parity_optab)
3178 {
3179 temp = expand_parity (mode, op0, target);
3180 if (temp)
3181 return temp;
3182 }
3183
3184 /* Try implementing ffs (x) in terms of clz (x). */
3185 if (unoptab == ffs_optab)
3186 {
3187 temp = expand_ffs (mode, op0, target);
3188 if (temp)
3189 return temp;
3190 }
3191
3192 /* Try implementing ctz (x) in terms of clz (x). */
3193 if (unoptab == ctz_optab)
3194 {
3195 temp = expand_ctz (mode, op0, target);
3196 if (temp)
3197 return temp;
3198 }
3199
3200 try_libcall:
3201 /* Now try a library call in this mode. */
3202 libfunc = optab_libfunc (unoptab, mode);
3203 if (libfunc)
3204 {
3205 rtx insns;
3206 rtx value;
3207 rtx eq_value;
3208 enum machine_mode outmode = mode;
3209
3210 /* All of these functions return small values. Thus we choose to
3211 have them return something that isn't a double-word. */
3212 if (unoptab == ffs_optab || unoptab == clz_optab || unoptab == ctz_optab
3213 || unoptab == clrsb_optab || unoptab == popcount_optab
3214 || unoptab == parity_optab)
3215 outmode
3216 = GET_MODE (hard_libcall_value (TYPE_MODE (integer_type_node),
3217 optab_libfunc (unoptab, mode)));
3218
3219 start_sequence ();
3220
3221 /* Pass 1 for NO_QUEUE so we don't lose any increments
3222 if the libcall is cse'd or moved. */
3223 value = emit_library_call_value (libfunc, NULL_RTX, LCT_CONST, outmode,
3224 1, op0, mode);
3225 insns = get_insns ();
3226 end_sequence ();
3227
3228 target = gen_reg_rtx (outmode);
3229 eq_value = gen_rtx_fmt_e (unoptab->code, mode, op0);
3230 if (GET_MODE_SIZE (outmode) < GET_MODE_SIZE (mode))
3231 eq_value = simplify_gen_unary (TRUNCATE, outmode, eq_value, mode);
3232 else if (GET_MODE_SIZE (outmode) > GET_MODE_SIZE (mode))
3233 eq_value = simplify_gen_unary (ZERO_EXTEND, outmode, eq_value, mode);
3234 emit_libcall_block (insns, target, value, eq_value);
3235
3236 return target;
3237 }
3238
3239 /* It can't be done in this mode. Can we do it in a wider mode? */
3240
3241 if (CLASS_HAS_WIDER_MODES_P (mclass))
3242 {
3243 for (wider_mode = GET_MODE_WIDER_MODE (mode);
3244 wider_mode != VOIDmode;
3245 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
3246 {
3247 if (optab_handler (unoptab, wider_mode) != CODE_FOR_nothing
3248 || optab_libfunc (unoptab, wider_mode))
3249 {
3250 rtx xop0 = op0;
3251 rtx last = get_last_insn ();
3252
3253 /* For certain operations, we need not actually extend
3254 the narrow operand, as long as we will truncate the
3255 results to the same narrowness. */
3256
3257 xop0 = widen_operand (xop0, wider_mode, mode, unsignedp,
3258 (unoptab == neg_optab
3259 || unoptab == one_cmpl_optab)
3260 && mclass == MODE_INT);
3261
3262 temp = expand_unop (wider_mode, unoptab, xop0, NULL_RTX,
3263 unsignedp);
3264
3265 /* If we are generating clz using wider mode, adjust the
3266 result. Similarly for clrsb. */
3267 if ((unoptab == clz_optab || unoptab == clrsb_optab)
3268 && temp != 0)
3269 temp = expand_binop (wider_mode, sub_optab, temp,
3270 GEN_INT (GET_MODE_PRECISION (wider_mode)
3271 - GET_MODE_PRECISION (mode)),
3272 target, true, OPTAB_DIRECT);
3273
3274 if (temp)
3275 {
3276 if (mclass != MODE_INT)
3277 {
3278 if (target == 0)
3279 target = gen_reg_rtx (mode);
3280 convert_move (target, temp, 0);
3281 return target;
3282 }
3283 else
3284 return gen_lowpart (mode, temp);
3285 }
3286 else
3287 delete_insns_since (last);
3288 }
3289 }
3290 }
3291
3292 /* One final attempt at implementing negation via subtraction,
3293 this time allowing widening of the operand. */
3294 if (unoptab->code == NEG && !HONOR_SIGNED_ZEROS (mode))
3295 {
3296 rtx temp;
3297 temp = expand_binop (mode,
3298 unoptab == negv_optab ? subv_optab : sub_optab,
3299 CONST0_RTX (mode), op0,
3300 target, unsignedp, OPTAB_LIB_WIDEN);
3301 if (temp)
3302 return temp;
3303 }
3304
3305 return 0;
3306 }
3307 \f
3308 /* Emit code to compute the absolute value of OP0, with result to
3309 TARGET if convenient. (TARGET may be 0.) The return value says
3310 where the result actually is to be found.
3311
3312 MODE is the mode of the operand; the mode of the result is
3313 different but can be deduced from MODE.
3314
3315 */
3316
3317 rtx
3318 expand_abs_nojump (enum machine_mode mode, rtx op0, rtx target,
3319 int result_unsignedp)
3320 {
3321 rtx temp;
3322
3323 if (! flag_trapv)
3324 result_unsignedp = 1;
3325
3326 /* First try to do it with a special abs instruction. */
3327 temp = expand_unop (mode, result_unsignedp ? abs_optab : absv_optab,
3328 op0, target, 0);
3329 if (temp != 0)
3330 return temp;
3331
3332 /* For floating point modes, try clearing the sign bit. */
3333 if (SCALAR_FLOAT_MODE_P (mode))
3334 {
3335 temp = expand_absneg_bit (ABS, mode, op0, target);
3336 if (temp)
3337 return temp;
3338 }
3339
3340 /* If we have a MAX insn, we can do this as MAX (x, -x). */
3341 if (optab_handler (smax_optab, mode) != CODE_FOR_nothing
3342 && !HONOR_SIGNED_ZEROS (mode))
3343 {
3344 rtx last = get_last_insn ();
3345
3346 temp = expand_unop (mode, neg_optab, op0, NULL_RTX, 0);
3347 if (temp != 0)
3348 temp = expand_binop (mode, smax_optab, op0, temp, target, 0,
3349 OPTAB_WIDEN);
3350
3351 if (temp != 0)
3352 return temp;
3353
3354 delete_insns_since (last);
3355 }
3356
3357 /* If this machine has expensive jumps, we can do integer absolute
3358 value of X as (((signed) x >> (W-1)) ^ x) - ((signed) x >> (W-1)),
3359 where W is the width of MODE. */
3360
3361 if (GET_MODE_CLASS (mode) == MODE_INT
3362 && BRANCH_COST (optimize_insn_for_speed_p (),
3363 false) >= 2)
3364 {
3365 rtx extended = expand_shift (RSHIFT_EXPR, mode, op0,
3366 GET_MODE_PRECISION (mode) - 1,
3367 NULL_RTX, 0);
3368
3369 temp = expand_binop (mode, xor_optab, extended, op0, target, 0,
3370 OPTAB_LIB_WIDEN);
3371 if (temp != 0)
3372 temp = expand_binop (mode, result_unsignedp ? sub_optab : subv_optab,
3373 temp, extended, target, 0, OPTAB_LIB_WIDEN);
3374
3375 if (temp != 0)
3376 return temp;
3377 }
3378
3379 return NULL_RTX;
3380 }
3381
3382 rtx
3383 expand_abs (enum machine_mode mode, rtx op0, rtx target,
3384 int result_unsignedp, int safe)
3385 {
3386 rtx temp, op1;
3387
3388 if (! flag_trapv)
3389 result_unsignedp = 1;
3390
3391 temp = expand_abs_nojump (mode, op0, target, result_unsignedp);
3392 if (temp != 0)
3393 return temp;
3394
3395 /* If that does not win, use conditional jump and negate. */
3396
3397 /* It is safe to use the target if it is the same
3398 as the source if this is also a pseudo register */
3399 if (op0 == target && REG_P (op0)
3400 && REGNO (op0) >= FIRST_PSEUDO_REGISTER)
3401 safe = 1;
3402
3403 op1 = gen_label_rtx ();
3404 if (target == 0 || ! safe
3405 || GET_MODE (target) != mode
3406 || (MEM_P (target) && MEM_VOLATILE_P (target))
3407 || (REG_P (target)
3408 && REGNO (target) < FIRST_PSEUDO_REGISTER))
3409 target = gen_reg_rtx (mode);
3410
3411 emit_move_insn (target, op0);
3412 NO_DEFER_POP;
3413
3414 do_compare_rtx_and_jump (target, CONST0_RTX (mode), GE, 0, mode,
3415 NULL_RTX, NULL_RTX, op1, -1);
3416
3417 op0 = expand_unop (mode, result_unsignedp ? neg_optab : negv_optab,
3418 target, target, 0);
3419 if (op0 != target)
3420 emit_move_insn (target, op0);
3421 emit_label (op1);
3422 OK_DEFER_POP;
3423 return target;
3424 }
3425
3426 /* Emit code to compute the one's complement absolute value of OP0
3427 (if (OP0 < 0) OP0 = ~OP0), with result to TARGET if convenient.
3428 (TARGET may be NULL_RTX.) The return value says where the result
3429 actually is to be found.
3430
3431 MODE is the mode of the operand; the mode of the result is
3432 different but can be deduced from MODE. */
3433
3434 rtx
3435 expand_one_cmpl_abs_nojump (enum machine_mode mode, rtx op0, rtx target)
3436 {
3437 rtx temp;
3438
3439 /* Not applicable for floating point modes. */
3440 if (FLOAT_MODE_P (mode))
3441 return NULL_RTX;
3442
3443 /* If we have a MAX insn, we can do this as MAX (x, ~x). */
3444 if (optab_handler (smax_optab, mode) != CODE_FOR_nothing)
3445 {
3446 rtx last = get_last_insn ();
3447
3448 temp = expand_unop (mode, one_cmpl_optab, op0, NULL_RTX, 0);
3449 if (temp != 0)
3450 temp = expand_binop (mode, smax_optab, op0, temp, target, 0,
3451 OPTAB_WIDEN);
3452
3453 if (temp != 0)
3454 return temp;
3455
3456 delete_insns_since (last);
3457 }
3458
3459 /* If this machine has expensive jumps, we can do one's complement
3460 absolute value of X as (((signed) x >> (W-1)) ^ x). */
3461
3462 if (GET_MODE_CLASS (mode) == MODE_INT
3463 && BRANCH_COST (optimize_insn_for_speed_p (),
3464 false) >= 2)
3465 {
3466 rtx extended = expand_shift (RSHIFT_EXPR, mode, op0,
3467 GET_MODE_PRECISION (mode) - 1,
3468 NULL_RTX, 0);
3469
3470 temp = expand_binop (mode, xor_optab, extended, op0, target, 0,
3471 OPTAB_LIB_WIDEN);
3472
3473 if (temp != 0)
3474 return temp;
3475 }
3476
3477 return NULL_RTX;
3478 }
3479
3480 /* A subroutine of expand_copysign, perform the copysign operation using the
3481 abs and neg primitives advertised to exist on the target. The assumption
3482 is that we have a split register file, and leaving op0 in fp registers,
3483 and not playing with subregs so much, will help the register allocator. */
3484
3485 static rtx
3486 expand_copysign_absneg (enum machine_mode mode, rtx op0, rtx op1, rtx target,
3487 int bitpos, bool op0_is_abs)
3488 {
3489 enum machine_mode imode;
3490 enum insn_code icode;
3491 rtx sign, label;
3492
3493 if (target == op1)
3494 target = NULL_RTX;
3495
3496 /* Check if the back end provides an insn that handles signbit for the
3497 argument's mode. */
3498 icode = optab_handler (signbit_optab, mode);
3499 if (icode != CODE_FOR_nothing)
3500 {
3501 imode = insn_data[(int) icode].operand[0].mode;
3502 sign = gen_reg_rtx (imode);
3503 emit_unop_insn (icode, sign, op1, UNKNOWN);
3504 }
3505 else
3506 {
3507 double_int mask;
3508
3509 if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD)
3510 {
3511 imode = int_mode_for_mode (mode);
3512 if (imode == BLKmode)
3513 return NULL_RTX;
3514 op1 = gen_lowpart (imode, op1);
3515 }
3516 else
3517 {
3518 int word;
3519
3520 imode = word_mode;
3521 if (FLOAT_WORDS_BIG_ENDIAN)
3522 word = (GET_MODE_BITSIZE (mode) - bitpos) / BITS_PER_WORD;
3523 else
3524 word = bitpos / BITS_PER_WORD;
3525 bitpos = bitpos % BITS_PER_WORD;
3526 op1 = operand_subword_force (op1, word, mode);
3527 }
3528
3529 mask = double_int_setbit (double_int_zero, bitpos);
3530
3531 sign = expand_binop (imode, and_optab, op1,
3532 immed_double_int_const (mask, imode),
3533 NULL_RTX, 1, OPTAB_LIB_WIDEN);
3534 }
3535
3536 if (!op0_is_abs)
3537 {
3538 op0 = expand_unop (mode, abs_optab, op0, target, 0);
3539 if (op0 == NULL)
3540 return NULL_RTX;
3541 target = op0;
3542 }
3543 else
3544 {
3545 if (target == NULL_RTX)
3546 target = copy_to_reg (op0);
3547 else
3548 emit_move_insn (target, op0);
3549 }
3550
3551 label = gen_label_rtx ();
3552 emit_cmp_and_jump_insns (sign, const0_rtx, EQ, NULL_RTX, imode, 1, label);
3553
3554 if (GET_CODE (op0) == CONST_DOUBLE)
3555 op0 = simplify_unary_operation (NEG, mode, op0, mode);
3556 else
3557 op0 = expand_unop (mode, neg_optab, op0, target, 0);
3558 if (op0 != target)
3559 emit_move_insn (target, op0);
3560
3561 emit_label (label);
3562
3563 return target;
3564 }
3565
3566
3567 /* A subroutine of expand_copysign, perform the entire copysign operation
3568 with integer bitmasks. BITPOS is the position of the sign bit; OP0_IS_ABS
3569 is true if op0 is known to have its sign bit clear. */
3570
3571 static rtx
3572 expand_copysign_bit (enum machine_mode mode, rtx op0, rtx op1, rtx target,
3573 int bitpos, bool op0_is_abs)
3574 {
3575 enum machine_mode imode;
3576 double_int mask;
3577 int word, nwords, i;
3578 rtx temp, insns;
3579
3580 if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD)
3581 {
3582 imode = int_mode_for_mode (mode);
3583 if (imode == BLKmode)
3584 return NULL_RTX;
3585 word = 0;
3586 nwords = 1;
3587 }
3588 else
3589 {
3590 imode = word_mode;
3591
3592 if (FLOAT_WORDS_BIG_ENDIAN)
3593 word = (GET_MODE_BITSIZE (mode) - bitpos) / BITS_PER_WORD;
3594 else
3595 word = bitpos / BITS_PER_WORD;
3596 bitpos = bitpos % BITS_PER_WORD;
3597 nwords = (GET_MODE_BITSIZE (mode) + BITS_PER_WORD - 1) / BITS_PER_WORD;
3598 }
3599
3600 mask = double_int_setbit (double_int_zero, bitpos);
3601
3602 if (target == 0
3603 || target == op0
3604 || target == op1
3605 || (nwords > 1 && !valid_multiword_target_p (target)))
3606 target = gen_reg_rtx (mode);
3607
3608 if (nwords > 1)
3609 {
3610 start_sequence ();
3611
3612 for (i = 0; i < nwords; ++i)
3613 {
3614 rtx targ_piece = operand_subword (target, i, 1, mode);
3615 rtx op0_piece = operand_subword_force (op0, i, mode);
3616
3617 if (i == word)
3618 {
3619 if (!op0_is_abs)
3620 op0_piece
3621 = expand_binop (imode, and_optab, op0_piece,
3622 immed_double_int_const (double_int_not (mask),
3623 imode),
3624 NULL_RTX, 1, OPTAB_LIB_WIDEN);
3625
3626 op1 = expand_binop (imode, and_optab,
3627 operand_subword_force (op1, i, mode),
3628 immed_double_int_const (mask, imode),
3629 NULL_RTX, 1, OPTAB_LIB_WIDEN);
3630
3631 temp = expand_binop (imode, ior_optab, op0_piece, op1,
3632 targ_piece, 1, OPTAB_LIB_WIDEN);
3633 if (temp != targ_piece)
3634 emit_move_insn (targ_piece, temp);
3635 }
3636 else
3637 emit_move_insn (targ_piece, op0_piece);
3638 }
3639
3640 insns = get_insns ();
3641 end_sequence ();
3642
3643 emit_insn (insns);
3644 }
3645 else
3646 {
3647 op1 = expand_binop (imode, and_optab, gen_lowpart (imode, op1),
3648 immed_double_int_const (mask, imode),
3649 NULL_RTX, 1, OPTAB_LIB_WIDEN);
3650
3651 op0 = gen_lowpart (imode, op0);
3652 if (!op0_is_abs)
3653 op0 = expand_binop (imode, and_optab, op0,
3654 immed_double_int_const (double_int_not (mask),
3655 imode),
3656 NULL_RTX, 1, OPTAB_LIB_WIDEN);
3657
3658 temp = expand_binop (imode, ior_optab, op0, op1,
3659 gen_lowpart (imode, target), 1, OPTAB_LIB_WIDEN);
3660 target = lowpart_subreg_maybe_copy (mode, temp, imode);
3661 }
3662
3663 return target;
3664 }
3665
3666 /* Expand the C99 copysign operation. OP0 and OP1 must be the same
3667 scalar floating point mode. Return NULL if we do not know how to
3668 expand the operation inline. */
3669
3670 rtx
3671 expand_copysign (rtx op0, rtx op1, rtx target)
3672 {
3673 enum machine_mode mode = GET_MODE (op0);
3674 const struct real_format *fmt;
3675 bool op0_is_abs;
3676 rtx temp;
3677
3678 gcc_assert (SCALAR_FLOAT_MODE_P (mode));
3679 gcc_assert (GET_MODE (op1) == mode);
3680
3681 /* First try to do it with a special instruction. */
3682 temp = expand_binop (mode, copysign_optab, op0, op1,
3683 target, 0, OPTAB_DIRECT);
3684 if (temp)
3685 return temp;
3686
3687 fmt = REAL_MODE_FORMAT (mode);
3688 if (fmt == NULL || !fmt->has_signed_zero)
3689 return NULL_RTX;
3690
3691 op0_is_abs = false;
3692 if (GET_CODE (op0) == CONST_DOUBLE)
3693 {
3694 if (real_isneg (CONST_DOUBLE_REAL_VALUE (op0)))
3695 op0 = simplify_unary_operation (ABS, mode, op0, mode);
3696 op0_is_abs = true;
3697 }
3698
3699 if (fmt->signbit_ro >= 0
3700 && (GET_CODE (op0) == CONST_DOUBLE
3701 || (optab_handler (neg_optab, mode) != CODE_FOR_nothing
3702 && optab_handler (abs_optab, mode) != CODE_FOR_nothing)))
3703 {
3704 temp = expand_copysign_absneg (mode, op0, op1, target,
3705 fmt->signbit_ro, op0_is_abs);
3706 if (temp)
3707 return temp;
3708 }
3709
3710 if (fmt->signbit_rw < 0)
3711 return NULL_RTX;
3712 return expand_copysign_bit (mode, op0, op1, target,
3713 fmt->signbit_rw, op0_is_abs);
3714 }
3715 \f
3716 /* Generate an instruction whose insn-code is INSN_CODE,
3717 with two operands: an output TARGET and an input OP0.
3718 TARGET *must* be nonzero, and the output is always stored there.
3719 CODE is an rtx code such that (CODE OP0) is an rtx that describes
3720 the value that is stored into TARGET.
3721
3722 Return false if expansion failed. */
3723
3724 bool
3725 maybe_emit_unop_insn (enum insn_code icode, rtx target, rtx op0,
3726 enum rtx_code code)
3727 {
3728 struct expand_operand ops[2];
3729 rtx pat;
3730
3731 create_output_operand (&ops[0], target, GET_MODE (target));
3732 create_input_operand (&ops[1], op0, GET_MODE (op0));
3733 pat = maybe_gen_insn (icode, 2, ops);
3734 if (!pat)
3735 return false;
3736
3737 if (INSN_P (pat) && NEXT_INSN (pat) != NULL_RTX && code != UNKNOWN)
3738 add_equal_note (pat, ops[0].value, code, ops[1].value, NULL_RTX);
3739
3740 emit_insn (pat);
3741
3742 if (ops[0].value != target)
3743 emit_move_insn (target, ops[0].value);
3744 return true;
3745 }
3746 /* Generate an instruction whose insn-code is INSN_CODE,
3747 with two operands: an output TARGET and an input OP0.
3748 TARGET *must* be nonzero, and the output is always stored there.
3749 CODE is an rtx code such that (CODE OP0) is an rtx that describes
3750 the value that is stored into TARGET. */
3751
3752 void
3753 emit_unop_insn (enum insn_code icode, rtx target, rtx op0, enum rtx_code code)
3754 {
3755 bool ok = maybe_emit_unop_insn (icode, target, op0, code);
3756 gcc_assert (ok);
3757 }
3758 \f
3759 struct no_conflict_data
3760 {
3761 rtx target, first, insn;
3762 bool must_stay;
3763 };
3764
3765 /* Called via note_stores by emit_libcall_block. Set P->must_stay if
3766 the currently examined clobber / store has to stay in the list of
3767 insns that constitute the actual libcall block. */
3768 static void
3769 no_conflict_move_test (rtx dest, const_rtx set, void *p0)
3770 {
3771 struct no_conflict_data *p= (struct no_conflict_data *) p0;
3772
3773 /* If this inns directly contributes to setting the target, it must stay. */
3774 if (reg_overlap_mentioned_p (p->target, dest))
3775 p->must_stay = true;
3776 /* If we haven't committed to keeping any other insns in the list yet,
3777 there is nothing more to check. */
3778 else if (p->insn == p->first)
3779 return;
3780 /* If this insn sets / clobbers a register that feeds one of the insns
3781 already in the list, this insn has to stay too. */
3782 else if (reg_overlap_mentioned_p (dest, PATTERN (p->first))
3783 || (CALL_P (p->first) && (find_reg_fusage (p->first, USE, dest)))
3784 || reg_used_between_p (dest, p->first, p->insn)
3785 /* Likewise if this insn depends on a register set by a previous
3786 insn in the list, or if it sets a result (presumably a hard
3787 register) that is set or clobbered by a previous insn.
3788 N.B. the modified_*_p (SET_DEST...) tests applied to a MEM
3789 SET_DEST perform the former check on the address, and the latter
3790 check on the MEM. */
3791 || (GET_CODE (set) == SET
3792 && (modified_in_p (SET_SRC (set), p->first)
3793 || modified_in_p (SET_DEST (set), p->first)
3794 || modified_between_p (SET_SRC (set), p->first, p->insn)
3795 || modified_between_p (SET_DEST (set), p->first, p->insn))))
3796 p->must_stay = true;
3797 }
3798
3799 \f
3800 /* Emit code to make a call to a constant function or a library call.
3801
3802 INSNS is a list containing all insns emitted in the call.
3803 These insns leave the result in RESULT. Our block is to copy RESULT
3804 to TARGET, which is logically equivalent to EQUIV.
3805
3806 We first emit any insns that set a pseudo on the assumption that these are
3807 loading constants into registers; doing so allows them to be safely cse'ed
3808 between blocks. Then we emit all the other insns in the block, followed by
3809 an insn to move RESULT to TARGET. This last insn will have a REQ_EQUAL
3810 note with an operand of EQUIV. */
3811
3812 void
3813 emit_libcall_block (rtx insns, rtx target, rtx result, rtx equiv)
3814 {
3815 rtx final_dest = target;
3816 rtx next, last, insn;
3817
3818 /* If this is a reg with REG_USERVAR_P set, then it could possibly turn
3819 into a MEM later. Protect the libcall block from this change. */
3820 if (! REG_P (target) || REG_USERVAR_P (target))
3821 target = gen_reg_rtx (GET_MODE (target));
3822
3823 /* If we're using non-call exceptions, a libcall corresponding to an
3824 operation that may trap may also trap. */
3825 /* ??? See the comment in front of make_reg_eh_region_note. */
3826 if (cfun->can_throw_non_call_exceptions && may_trap_p (equiv))
3827 {
3828 for (insn = insns; insn; insn = NEXT_INSN (insn))
3829 if (CALL_P (insn))
3830 {
3831 rtx note = find_reg_note (insn, REG_EH_REGION, NULL_RTX);
3832 if (note)
3833 {
3834 int lp_nr = INTVAL (XEXP (note, 0));
3835 if (lp_nr == 0 || lp_nr == INT_MIN)
3836 remove_note (insn, note);
3837 }
3838 }
3839 }
3840 else
3841 {
3842 /* Look for any CALL_INSNs in this sequence, and attach a REG_EH_REGION
3843 reg note to indicate that this call cannot throw or execute a nonlocal
3844 goto (unless there is already a REG_EH_REGION note, in which case
3845 we update it). */
3846 for (insn = insns; insn; insn = NEXT_INSN (insn))
3847 if (CALL_P (insn))
3848 make_reg_eh_region_note_nothrow_nononlocal (insn);
3849 }
3850
3851 /* First emit all insns that set pseudos. Remove them from the list as
3852 we go. Avoid insns that set pseudos which were referenced in previous
3853 insns. These can be generated by move_by_pieces, for example,
3854 to update an address. Similarly, avoid insns that reference things
3855 set in previous insns. */
3856
3857 for (insn = insns; insn; insn = next)
3858 {
3859 rtx set = single_set (insn);
3860
3861 next = NEXT_INSN (insn);
3862
3863 if (set != 0 && REG_P (SET_DEST (set))
3864 && REGNO (SET_DEST (set)) >= FIRST_PSEUDO_REGISTER)
3865 {
3866 struct no_conflict_data data;
3867
3868 data.target = const0_rtx;
3869 data.first = insns;
3870 data.insn = insn;
3871 data.must_stay = 0;
3872 note_stores (PATTERN (insn), no_conflict_move_test, &data);
3873 if (! data.must_stay)
3874 {
3875 if (PREV_INSN (insn))
3876 NEXT_INSN (PREV_INSN (insn)) = next;
3877 else
3878 insns = next;
3879
3880 if (next)
3881 PREV_INSN (next) = PREV_INSN (insn);
3882
3883 add_insn (insn);
3884 }
3885 }
3886
3887 /* Some ports use a loop to copy large arguments onto the stack.
3888 Don't move anything outside such a loop. */
3889 if (LABEL_P (insn))
3890 break;
3891 }
3892
3893 /* Write the remaining insns followed by the final copy. */
3894 for (insn = insns; insn; insn = next)
3895 {
3896 next = NEXT_INSN (insn);
3897
3898 add_insn (insn);
3899 }
3900
3901 last = emit_move_insn (target, result);
3902 if (optab_handler (mov_optab, GET_MODE (target)) != CODE_FOR_nothing)
3903 set_unique_reg_note (last, REG_EQUAL, copy_rtx (equiv));
3904
3905 if (final_dest != target)
3906 emit_move_insn (final_dest, target);
3907 }
3908 \f
3909 /* Nonzero if we can perform a comparison of mode MODE straightforwardly.
3910 PURPOSE describes how this comparison will be used. CODE is the rtx
3911 comparison code we will be using.
3912
3913 ??? Actually, CODE is slightly weaker than that. A target is still
3914 required to implement all of the normal bcc operations, but not
3915 required to implement all (or any) of the unordered bcc operations. */
3916
3917 int
3918 can_compare_p (enum rtx_code code, enum machine_mode mode,
3919 enum can_compare_purpose purpose)
3920 {
3921 rtx test;
3922 test = gen_rtx_fmt_ee (code, mode, const0_rtx, const0_rtx);
3923 do
3924 {
3925 enum insn_code icode;
3926
3927 if (purpose == ccp_jump
3928 && (icode = optab_handler (cbranch_optab, mode)) != CODE_FOR_nothing
3929 && insn_operand_matches (icode, 0, test))
3930 return 1;
3931 if (purpose == ccp_store_flag
3932 && (icode = optab_handler (cstore_optab, mode)) != CODE_FOR_nothing
3933 && insn_operand_matches (icode, 1, test))
3934 return 1;
3935 if (purpose == ccp_cmov
3936 && optab_handler (cmov_optab, mode) != CODE_FOR_nothing)
3937 return 1;
3938
3939 mode = GET_MODE_WIDER_MODE (mode);
3940 PUT_MODE (test, mode);
3941 }
3942 while (mode != VOIDmode);
3943
3944 return 0;
3945 }
3946
3947 /* This function is called when we are going to emit a compare instruction that
3948 compares the values found in *PX and *PY, using the rtl operator COMPARISON.
3949
3950 *PMODE is the mode of the inputs (in case they are const_int).
3951 *PUNSIGNEDP nonzero says that the operands are unsigned;
3952 this matters if they need to be widened (as given by METHODS).
3953
3954 If they have mode BLKmode, then SIZE specifies the size of both operands.
3955
3956 This function performs all the setup necessary so that the caller only has
3957 to emit a single comparison insn. This setup can involve doing a BLKmode
3958 comparison or emitting a library call to perform the comparison if no insn
3959 is available to handle it.
3960 The values which are passed in through pointers can be modified; the caller
3961 should perform the comparison on the modified values. Constant
3962 comparisons must have already been folded. */
3963
3964 static void
3965 prepare_cmp_insn (rtx x, rtx y, enum rtx_code comparison, rtx size,
3966 int unsignedp, enum optab_methods methods,
3967 rtx *ptest, enum machine_mode *pmode)
3968 {
3969 enum machine_mode mode = *pmode;
3970 rtx libfunc, test;
3971 enum machine_mode cmp_mode;
3972 enum mode_class mclass;
3973
3974 /* The other methods are not needed. */
3975 gcc_assert (methods == OPTAB_DIRECT || methods == OPTAB_WIDEN
3976 || methods == OPTAB_LIB_WIDEN);
3977
3978 /* If we are optimizing, force expensive constants into a register. */
3979 if (CONSTANT_P (x) && optimize
3980 && (rtx_cost (x, COMPARE, 0, optimize_insn_for_speed_p ())
3981 > COSTS_N_INSNS (1)))
3982 x = force_reg (mode, x);
3983
3984 if (CONSTANT_P (y) && optimize
3985 && (rtx_cost (y, COMPARE, 1, optimize_insn_for_speed_p ())
3986 > COSTS_N_INSNS (1)))
3987 y = force_reg (mode, y);
3988
3989 #ifdef HAVE_cc0
3990 /* Make sure if we have a canonical comparison. The RTL
3991 documentation states that canonical comparisons are required only
3992 for targets which have cc0. */
3993 gcc_assert (!CONSTANT_P (x) || CONSTANT_P (y));
3994 #endif
3995
3996 /* Don't let both operands fail to indicate the mode. */
3997 if (GET_MODE (x) == VOIDmode && GET_MODE (y) == VOIDmode)
3998 x = force_reg (mode, x);
3999 if (mode == VOIDmode)
4000 mode = GET_MODE (x) != VOIDmode ? GET_MODE (x) : GET_MODE (y);
4001
4002 /* Handle all BLKmode compares. */
4003
4004 if (mode == BLKmode)
4005 {
4006 enum machine_mode result_mode;
4007 enum insn_code cmp_code;
4008 tree length_type;
4009 rtx libfunc;
4010 rtx result;
4011 rtx opalign
4012 = GEN_INT (MIN (MEM_ALIGN (x), MEM_ALIGN (y)) / BITS_PER_UNIT);
4013
4014 gcc_assert (size);
4015
4016 /* Try to use a memory block compare insn - either cmpstr
4017 or cmpmem will do. */
4018 for (cmp_mode = GET_CLASS_NARROWEST_MODE (MODE_INT);
4019 cmp_mode != VOIDmode;
4020 cmp_mode = GET_MODE_WIDER_MODE (cmp_mode))
4021 {
4022 cmp_code = direct_optab_handler (cmpmem_optab, cmp_mode);
4023 if (cmp_code == CODE_FOR_nothing)
4024 cmp_code = direct_optab_handler (cmpstr_optab, cmp_mode);
4025 if (cmp_code == CODE_FOR_nothing)
4026 cmp_code = direct_optab_handler (cmpstrn_optab, cmp_mode);
4027 if (cmp_code == CODE_FOR_nothing)
4028 continue;
4029
4030 /* Must make sure the size fits the insn's mode. */
4031 if ((CONST_INT_P (size)
4032 && INTVAL (size) >= (1 << GET_MODE_BITSIZE (cmp_mode)))
4033 || (GET_MODE_BITSIZE (GET_MODE (size))
4034 > GET_MODE_BITSIZE (cmp_mode)))
4035 continue;
4036
4037 result_mode = insn_data[cmp_code].operand[0].mode;
4038 result = gen_reg_rtx (result_mode);
4039 size = convert_to_mode (cmp_mode, size, 1);
4040 emit_insn (GEN_FCN (cmp_code) (result, x, y, size, opalign));
4041
4042 *ptest = gen_rtx_fmt_ee (comparison, VOIDmode, result, const0_rtx);
4043 *pmode = result_mode;
4044 return;
4045 }
4046
4047 if (methods != OPTAB_LIB && methods != OPTAB_LIB_WIDEN)
4048 goto fail;
4049
4050 /* Otherwise call a library function, memcmp. */
4051 libfunc = memcmp_libfunc;
4052 length_type = sizetype;
4053 result_mode = TYPE_MODE (integer_type_node);
4054 cmp_mode = TYPE_MODE (length_type);
4055 size = convert_to_mode (TYPE_MODE (length_type), size,
4056 TYPE_UNSIGNED (length_type));
4057
4058 result = emit_library_call_value (libfunc, 0, LCT_PURE,
4059 result_mode, 3,
4060 XEXP (x, 0), Pmode,
4061 XEXP (y, 0), Pmode,
4062 size, cmp_mode);
4063
4064 *ptest = gen_rtx_fmt_ee (comparison, VOIDmode, result, const0_rtx);
4065 *pmode = result_mode;
4066 return;
4067 }
4068
4069 /* Don't allow operands to the compare to trap, as that can put the
4070 compare and branch in different basic blocks. */
4071 if (cfun->can_throw_non_call_exceptions)
4072 {
4073 if (may_trap_p (x))
4074 x = force_reg (mode, x);
4075 if (may_trap_p (y))
4076 y = force_reg (mode, y);
4077 }
4078
4079 if (GET_MODE_CLASS (mode) == MODE_CC)
4080 {
4081 gcc_assert (can_compare_p (comparison, CCmode, ccp_jump));
4082 *ptest = gen_rtx_fmt_ee (comparison, VOIDmode, x, y);
4083 return;
4084 }
4085
4086 mclass = GET_MODE_CLASS (mode);
4087 test = gen_rtx_fmt_ee (comparison, VOIDmode, x, y);
4088 cmp_mode = mode;
4089 do
4090 {
4091 enum insn_code icode;
4092 icode = optab_handler (cbranch_optab, cmp_mode);
4093 if (icode != CODE_FOR_nothing
4094 && insn_operand_matches (icode, 0, test))
4095 {
4096 rtx last = get_last_insn ();
4097 rtx op0 = prepare_operand (icode, x, 1, mode, cmp_mode, unsignedp);
4098 rtx op1 = prepare_operand (icode, y, 2, mode, cmp_mode, unsignedp);
4099 if (op0 && op1
4100 && insn_operand_matches (icode, 1, op0)
4101 && insn_operand_matches (icode, 2, op1))
4102 {
4103 XEXP (test, 0) = op0;
4104 XEXP (test, 1) = op1;
4105 *ptest = test;
4106 *pmode = cmp_mode;
4107 return;
4108 }
4109 delete_insns_since (last);
4110 }
4111
4112 if (methods == OPTAB_DIRECT || !CLASS_HAS_WIDER_MODES_P (mclass))
4113 break;
4114 cmp_mode = GET_MODE_WIDER_MODE (cmp_mode);
4115 }
4116 while (cmp_mode != VOIDmode);
4117
4118 if (methods != OPTAB_LIB_WIDEN)
4119 goto fail;
4120
4121 if (!SCALAR_FLOAT_MODE_P (mode))
4122 {
4123 rtx result;
4124
4125 /* Handle a libcall just for the mode we are using. */
4126 libfunc = optab_libfunc (cmp_optab, mode);
4127 gcc_assert (libfunc);
4128
4129 /* If we want unsigned, and this mode has a distinct unsigned
4130 comparison routine, use that. */
4131 if (unsignedp)
4132 {
4133 rtx ulibfunc = optab_libfunc (ucmp_optab, mode);
4134 if (ulibfunc)
4135 libfunc = ulibfunc;
4136 }
4137
4138 result = emit_library_call_value (libfunc, NULL_RTX, LCT_CONST,
4139 targetm.libgcc_cmp_return_mode (),
4140 2, x, mode, y, mode);
4141
4142 /* There are two kinds of comparison routines. Biased routines
4143 return 0/1/2, and unbiased routines return -1/0/1. Other parts
4144 of gcc expect that the comparison operation is equivalent
4145 to the modified comparison. For signed comparisons compare the
4146 result against 1 in the biased case, and zero in the unbiased
4147 case. For unsigned comparisons always compare against 1 after
4148 biasing the unbiased result by adding 1. This gives us a way to
4149 represent LTU.
4150 The comparisons in the fixed-point helper library are always
4151 biased. */
4152 x = result;
4153 y = const1_rtx;
4154
4155 if (!TARGET_LIB_INT_CMP_BIASED && !ALL_FIXED_POINT_MODE_P (mode))
4156 {
4157 if (unsignedp)
4158 x = plus_constant (result, 1);
4159 else
4160 y = const0_rtx;
4161 }
4162
4163 *pmode = word_mode;
4164 prepare_cmp_insn (x, y, comparison, NULL_RTX, unsignedp, methods,
4165 ptest, pmode);
4166 }
4167 else
4168 prepare_float_lib_cmp (x, y, comparison, ptest, pmode);
4169
4170 return;
4171
4172 fail:
4173 *ptest = NULL_RTX;
4174 }
4175
4176 /* Before emitting an insn with code ICODE, make sure that X, which is going
4177 to be used for operand OPNUM of the insn, is converted from mode MODE to
4178 WIDER_MODE (UNSIGNEDP determines whether it is an unsigned conversion), and
4179 that it is accepted by the operand predicate. Return the new value. */
4180
4181 rtx
4182 prepare_operand (enum insn_code icode, rtx x, int opnum, enum machine_mode mode,
4183 enum machine_mode wider_mode, int unsignedp)
4184 {
4185 if (mode != wider_mode)
4186 x = convert_modes (wider_mode, mode, x, unsignedp);
4187
4188 if (!insn_operand_matches (icode, opnum, x))
4189 {
4190 if (reload_completed)
4191 return NULL_RTX;
4192 x = copy_to_mode_reg (insn_data[(int) icode].operand[opnum].mode, x);
4193 }
4194
4195 return x;
4196 }
4197
4198 /* Subroutine of emit_cmp_and_jump_insns; this function is called when we know
4199 we can do the branch. */
4200
4201 static void
4202 emit_cmp_and_jump_insn_1 (rtx test, enum machine_mode mode, rtx label)
4203 {
4204 enum machine_mode optab_mode;
4205 enum mode_class mclass;
4206 enum insn_code icode;
4207
4208 mclass = GET_MODE_CLASS (mode);
4209 optab_mode = (mclass == MODE_CC) ? CCmode : mode;
4210 icode = optab_handler (cbranch_optab, optab_mode);
4211
4212 gcc_assert (icode != CODE_FOR_nothing);
4213 gcc_assert (insn_operand_matches (icode, 0, test));
4214 emit_jump_insn (GEN_FCN (icode) (test, XEXP (test, 0), XEXP (test, 1), label));
4215 }
4216
4217 /* Generate code to compare X with Y so that the condition codes are
4218 set and to jump to LABEL if the condition is true. If X is a
4219 constant and Y is not a constant, then the comparison is swapped to
4220 ensure that the comparison RTL has the canonical form.
4221
4222 UNSIGNEDP nonzero says that X and Y are unsigned; this matters if they
4223 need to be widened. UNSIGNEDP is also used to select the proper
4224 branch condition code.
4225
4226 If X and Y have mode BLKmode, then SIZE specifies the size of both X and Y.
4227
4228 MODE is the mode of the inputs (in case they are const_int).
4229
4230 COMPARISON is the rtl operator to compare with (EQ, NE, GT, etc.).
4231 It will be potentially converted into an unsigned variant based on
4232 UNSIGNEDP to select a proper jump instruction. */
4233
4234 void
4235 emit_cmp_and_jump_insns (rtx x, rtx y, enum rtx_code comparison, rtx size,
4236 enum machine_mode mode, int unsignedp, rtx label)
4237 {
4238 rtx op0 = x, op1 = y;
4239 rtx test;
4240
4241 /* Swap operands and condition to ensure canonical RTL. */
4242 if (swap_commutative_operands_p (x, y)
4243 && can_compare_p (swap_condition (comparison), mode, ccp_jump))
4244 {
4245 op0 = y, op1 = x;
4246 comparison = swap_condition (comparison);
4247 }
4248
4249 /* If OP0 is still a constant, then both X and Y must be constants
4250 or the opposite comparison is not supported. Force X into a register
4251 to create canonical RTL. */
4252 if (CONSTANT_P (op0))
4253 op0 = force_reg (mode, op0);
4254
4255 if (unsignedp)
4256 comparison = unsigned_condition (comparison);
4257
4258 prepare_cmp_insn (op0, op1, comparison, size, unsignedp, OPTAB_LIB_WIDEN,
4259 &test, &mode);
4260 emit_cmp_and_jump_insn_1 (test, mode, label);
4261 }
4262
4263 \f
4264 /* Emit a library call comparison between floating point X and Y.
4265 COMPARISON is the rtl operator to compare with (EQ, NE, GT, etc.). */
4266
4267 static void
4268 prepare_float_lib_cmp (rtx x, rtx y, enum rtx_code comparison,
4269 rtx *ptest, enum machine_mode *pmode)
4270 {
4271 enum rtx_code swapped = swap_condition (comparison);
4272 enum rtx_code reversed = reverse_condition_maybe_unordered (comparison);
4273 enum machine_mode orig_mode = GET_MODE (x);
4274 enum machine_mode mode, cmp_mode;
4275 rtx true_rtx, false_rtx;
4276 rtx value, target, insns, equiv;
4277 rtx libfunc = 0;
4278 bool reversed_p = false;
4279 cmp_mode = targetm.libgcc_cmp_return_mode ();
4280
4281 for (mode = orig_mode;
4282 mode != VOIDmode;
4283 mode = GET_MODE_WIDER_MODE (mode))
4284 {
4285 if (code_to_optab[comparison]
4286 && (libfunc = optab_libfunc (code_to_optab[comparison], mode)))
4287 break;
4288
4289 if (code_to_optab[swapped]
4290 && (libfunc = optab_libfunc (code_to_optab[swapped], mode)))
4291 {
4292 rtx tmp;
4293 tmp = x; x = y; y = tmp;
4294 comparison = swapped;
4295 break;
4296 }
4297
4298 if (code_to_optab[reversed]
4299 && (libfunc = optab_libfunc (code_to_optab[reversed], mode)))
4300 {
4301 comparison = reversed;
4302 reversed_p = true;
4303 break;
4304 }
4305 }
4306
4307 gcc_assert (mode != VOIDmode);
4308
4309 if (mode != orig_mode)
4310 {
4311 x = convert_to_mode (mode, x, 0);
4312 y = convert_to_mode (mode, y, 0);
4313 }
4314
4315 /* Attach a REG_EQUAL note describing the semantics of the libcall to
4316 the RTL. The allows the RTL optimizers to delete the libcall if the
4317 condition can be determined at compile-time. */
4318 if (comparison == UNORDERED
4319 || FLOAT_LIB_COMPARE_RETURNS_BOOL (mode, comparison))
4320 {
4321 true_rtx = const_true_rtx;
4322 false_rtx = const0_rtx;
4323 }
4324 else
4325 {
4326 switch (comparison)
4327 {
4328 case EQ:
4329 true_rtx = const0_rtx;
4330 false_rtx = const_true_rtx;
4331 break;
4332
4333 case NE:
4334 true_rtx = const_true_rtx;
4335 false_rtx = const0_rtx;
4336 break;
4337
4338 case GT:
4339 true_rtx = const1_rtx;
4340 false_rtx = const0_rtx;
4341 break;
4342
4343 case GE:
4344 true_rtx = const0_rtx;
4345 false_rtx = constm1_rtx;
4346 break;
4347
4348 case LT:
4349 true_rtx = constm1_rtx;
4350 false_rtx = const0_rtx;
4351 break;
4352
4353 case LE:
4354 true_rtx = const0_rtx;
4355 false_rtx = const1_rtx;
4356 break;
4357
4358 default:
4359 gcc_unreachable ();
4360 }
4361 }
4362
4363 if (comparison == UNORDERED)
4364 {
4365 rtx temp = simplify_gen_relational (NE, cmp_mode, mode, x, x);
4366 equiv = simplify_gen_relational (NE, cmp_mode, mode, y, y);
4367 equiv = simplify_gen_ternary (IF_THEN_ELSE, cmp_mode, cmp_mode,
4368 temp, const_true_rtx, equiv);
4369 }
4370 else
4371 {
4372 equiv = simplify_gen_relational (comparison, cmp_mode, mode, x, y);
4373 if (! FLOAT_LIB_COMPARE_RETURNS_BOOL (mode, comparison))
4374 equiv = simplify_gen_ternary (IF_THEN_ELSE, cmp_mode, cmp_mode,
4375 equiv, true_rtx, false_rtx);
4376 }
4377
4378 start_sequence ();
4379 value = emit_library_call_value (libfunc, NULL_RTX, LCT_CONST,
4380 cmp_mode, 2, x, mode, y, mode);
4381 insns = get_insns ();
4382 end_sequence ();
4383
4384 target = gen_reg_rtx (cmp_mode);
4385 emit_libcall_block (insns, target, value, equiv);
4386
4387 if (comparison == UNORDERED
4388 || FLOAT_LIB_COMPARE_RETURNS_BOOL (mode, comparison)
4389 || reversed_p)
4390 *ptest = gen_rtx_fmt_ee (reversed_p ? EQ : NE, VOIDmode, target, false_rtx);
4391 else
4392 *ptest = gen_rtx_fmt_ee (comparison, VOIDmode, target, const0_rtx);
4393
4394 *pmode = cmp_mode;
4395 }
4396 \f
4397 /* Generate code to indirectly jump to a location given in the rtx LOC. */
4398
4399 void
4400 emit_indirect_jump (rtx loc)
4401 {
4402 struct expand_operand ops[1];
4403
4404 create_address_operand (&ops[0], loc);
4405 expand_jump_insn (CODE_FOR_indirect_jump, 1, ops);
4406 emit_barrier ();
4407 }
4408 \f
4409 #ifdef HAVE_conditional_move
4410
4411 /* Emit a conditional move instruction if the machine supports one for that
4412 condition and machine mode.
4413
4414 OP0 and OP1 are the operands that should be compared using CODE. CMODE is
4415 the mode to use should they be constants. If it is VOIDmode, they cannot
4416 both be constants.
4417
4418 OP2 should be stored in TARGET if the comparison is true, otherwise OP3
4419 should be stored there. MODE is the mode to use should they be constants.
4420 If it is VOIDmode, they cannot both be constants.
4421
4422 The result is either TARGET (perhaps modified) or NULL_RTX if the operation
4423 is not supported. */
4424
4425 rtx
4426 emit_conditional_move (rtx target, enum rtx_code code, rtx op0, rtx op1,
4427 enum machine_mode cmode, rtx op2, rtx op3,
4428 enum machine_mode mode, int unsignedp)
4429 {
4430 rtx tem, comparison, last;
4431 enum insn_code icode;
4432 enum rtx_code reversed;
4433
4434 /* If one operand is constant, make it the second one. Only do this
4435 if the other operand is not constant as well. */
4436
4437 if (swap_commutative_operands_p (op0, op1))
4438 {
4439 tem = op0;
4440 op0 = op1;
4441 op1 = tem;
4442 code = swap_condition (code);
4443 }
4444
4445 /* get_condition will prefer to generate LT and GT even if the old
4446 comparison was against zero, so undo that canonicalization here since
4447 comparisons against zero are cheaper. */
4448 if (code == LT && op1 == const1_rtx)
4449 code = LE, op1 = const0_rtx;
4450 else if (code == GT && op1 == constm1_rtx)
4451 code = GE, op1 = const0_rtx;
4452
4453 if (cmode == VOIDmode)
4454 cmode = GET_MODE (op0);
4455
4456 if (swap_commutative_operands_p (op2, op3)
4457 && ((reversed = reversed_comparison_code_parts (code, op0, op1, NULL))
4458 != UNKNOWN))
4459 {
4460 tem = op2;
4461 op2 = op3;
4462 op3 = tem;
4463 code = reversed;
4464 }
4465
4466 if (mode == VOIDmode)
4467 mode = GET_MODE (op2);
4468
4469 icode = direct_optab_handler (movcc_optab, mode);
4470
4471 if (icode == CODE_FOR_nothing)
4472 return 0;
4473
4474 if (!target)
4475 target = gen_reg_rtx (mode);
4476
4477 code = unsignedp ? unsigned_condition (code) : code;
4478 comparison = simplify_gen_relational (code, VOIDmode, cmode, op0, op1);
4479
4480 /* We can get const0_rtx or const_true_rtx in some circumstances. Just
4481 return NULL and let the caller figure out how best to deal with this
4482 situation. */
4483 if (!COMPARISON_P (comparison))
4484 return NULL_RTX;
4485
4486 do_pending_stack_adjust ();
4487 last = get_last_insn ();
4488 prepare_cmp_insn (XEXP (comparison, 0), XEXP (comparison, 1),
4489 GET_CODE (comparison), NULL_RTX, unsignedp, OPTAB_WIDEN,
4490 &comparison, &cmode);
4491 if (comparison)
4492 {
4493 struct expand_operand ops[4];
4494
4495 create_output_operand (&ops[0], target, mode);
4496 create_fixed_operand (&ops[1], comparison);
4497 create_input_operand (&ops[2], op2, mode);
4498 create_input_operand (&ops[3], op3, mode);
4499 if (maybe_expand_insn (icode, 4, ops))
4500 {
4501 if (ops[0].value != target)
4502 convert_move (target, ops[0].value, false);
4503 return target;
4504 }
4505 }
4506 delete_insns_since (last);
4507 return NULL_RTX;
4508 }
4509
4510 /* Return nonzero if a conditional move of mode MODE is supported.
4511
4512 This function is for combine so it can tell whether an insn that looks
4513 like a conditional move is actually supported by the hardware. If we
4514 guess wrong we lose a bit on optimization, but that's it. */
4515 /* ??? sparc64 supports conditionally moving integers values based on fp
4516 comparisons, and vice versa. How do we handle them? */
4517
4518 int
4519 can_conditionally_move_p (enum machine_mode mode)
4520 {
4521 if (direct_optab_handler (movcc_optab, mode) != CODE_FOR_nothing)
4522 return 1;
4523
4524 return 0;
4525 }
4526
4527 #endif /* HAVE_conditional_move */
4528
4529 /* Emit a conditional addition instruction if the machine supports one for that
4530 condition and machine mode.
4531
4532 OP0 and OP1 are the operands that should be compared using CODE. CMODE is
4533 the mode to use should they be constants. If it is VOIDmode, they cannot
4534 both be constants.
4535
4536 OP2 should be stored in TARGET if the comparison is true, otherwise OP2+OP3
4537 should be stored there. MODE is the mode to use should they be constants.
4538 If it is VOIDmode, they cannot both be constants.
4539
4540 The result is either TARGET (perhaps modified) or NULL_RTX if the operation
4541 is not supported. */
4542
4543 rtx
4544 emit_conditional_add (rtx target, enum rtx_code code, rtx op0, rtx op1,
4545 enum machine_mode cmode, rtx op2, rtx op3,
4546 enum machine_mode mode, int unsignedp)
4547 {
4548 rtx tem, comparison, last;
4549 enum insn_code icode;
4550 enum rtx_code reversed;
4551
4552 /* If one operand is constant, make it the second one. Only do this
4553 if the other operand is not constant as well. */
4554
4555 if (swap_commutative_operands_p (op0, op1))
4556 {
4557 tem = op0;
4558 op0 = op1;
4559 op1 = tem;
4560 code = swap_condition (code);
4561 }
4562
4563 /* get_condition will prefer to generate LT and GT even if the old
4564 comparison was against zero, so undo that canonicalization here since
4565 comparisons against zero are cheaper. */
4566 if (code == LT && op1 == const1_rtx)
4567 code = LE, op1 = const0_rtx;
4568 else if (code == GT && op1 == constm1_rtx)
4569 code = GE, op1 = const0_rtx;
4570
4571 if (cmode == VOIDmode)
4572 cmode = GET_MODE (op0);
4573
4574 if (swap_commutative_operands_p (op2, op3)
4575 && ((reversed = reversed_comparison_code_parts (code, op0, op1, NULL))
4576 != UNKNOWN))
4577 {
4578 tem = op2;
4579 op2 = op3;
4580 op3 = tem;
4581 code = reversed;
4582 }
4583
4584 if (mode == VOIDmode)
4585 mode = GET_MODE (op2);
4586
4587 icode = optab_handler (addcc_optab, mode);
4588
4589 if (icode == CODE_FOR_nothing)
4590 return 0;
4591
4592 if (!target)
4593 target = gen_reg_rtx (mode);
4594
4595 code = unsignedp ? unsigned_condition (code) : code;
4596 comparison = simplify_gen_relational (code, VOIDmode, cmode, op0, op1);
4597
4598 /* We can get const0_rtx or const_true_rtx in some circumstances. Just
4599 return NULL and let the caller figure out how best to deal with this
4600 situation. */
4601 if (!COMPARISON_P (comparison))
4602 return NULL_RTX;
4603
4604 do_pending_stack_adjust ();
4605 last = get_last_insn ();
4606 prepare_cmp_insn (XEXP (comparison, 0), XEXP (comparison, 1),
4607 GET_CODE (comparison), NULL_RTX, unsignedp, OPTAB_WIDEN,
4608 &comparison, &cmode);
4609 if (comparison)
4610 {
4611 struct expand_operand ops[4];
4612
4613 create_output_operand (&ops[0], target, mode);
4614 create_fixed_operand (&ops[1], comparison);
4615 create_input_operand (&ops[2], op2, mode);
4616 create_input_operand (&ops[3], op3, mode);
4617 if (maybe_expand_insn (icode, 4, ops))
4618 {
4619 if (ops[0].value != target)
4620 convert_move (target, ops[0].value, false);
4621 return target;
4622 }
4623 }
4624 delete_insns_since (last);
4625 return NULL_RTX;
4626 }
4627 \f
4628 /* These functions attempt to generate an insn body, rather than
4629 emitting the insn, but if the gen function already emits them, we
4630 make no attempt to turn them back into naked patterns. */
4631
4632 /* Generate and return an insn body to add Y to X. */
4633
4634 rtx
4635 gen_add2_insn (rtx x, rtx y)
4636 {
4637 enum insn_code icode = optab_handler (add_optab, GET_MODE (x));
4638
4639 gcc_assert (insn_operand_matches (icode, 0, x));
4640 gcc_assert (insn_operand_matches (icode, 1, x));
4641 gcc_assert (insn_operand_matches (icode, 2, y));
4642
4643 return GEN_FCN (icode) (x, x, y);
4644 }
4645
4646 /* Generate and return an insn body to add r1 and c,
4647 storing the result in r0. */
4648
4649 rtx
4650 gen_add3_insn (rtx r0, rtx r1, rtx c)
4651 {
4652 enum insn_code icode = optab_handler (add_optab, GET_MODE (r0));
4653
4654 if (icode == CODE_FOR_nothing
4655 || !insn_operand_matches (icode, 0, r0)
4656 || !insn_operand_matches (icode, 1, r1)
4657 || !insn_operand_matches (icode, 2, c))
4658 return NULL_RTX;
4659
4660 return GEN_FCN (icode) (r0, r1, c);
4661 }
4662
4663 int
4664 have_add2_insn (rtx x, rtx y)
4665 {
4666 enum insn_code icode;
4667
4668 gcc_assert (GET_MODE (x) != VOIDmode);
4669
4670 icode = optab_handler (add_optab, GET_MODE (x));
4671
4672 if (icode == CODE_FOR_nothing)
4673 return 0;
4674
4675 if (!insn_operand_matches (icode, 0, x)
4676 || !insn_operand_matches (icode, 1, x)
4677 || !insn_operand_matches (icode, 2, y))
4678 return 0;
4679
4680 return 1;
4681 }
4682
4683 /* Generate and return an insn body to subtract Y from X. */
4684
4685 rtx
4686 gen_sub2_insn (rtx x, rtx y)
4687 {
4688 enum insn_code icode = optab_handler (sub_optab, GET_MODE (x));
4689
4690 gcc_assert (insn_operand_matches (icode, 0, x));
4691 gcc_assert (insn_operand_matches (icode, 1, x));
4692 gcc_assert (insn_operand_matches (icode, 2, y));
4693
4694 return GEN_FCN (icode) (x, x, y);
4695 }
4696
4697 /* Generate and return an insn body to subtract r1 and c,
4698 storing the result in r0. */
4699
4700 rtx
4701 gen_sub3_insn (rtx r0, rtx r1, rtx c)
4702 {
4703 enum insn_code icode = optab_handler (sub_optab, GET_MODE (r0));
4704
4705 if (icode == CODE_FOR_nothing
4706 || !insn_operand_matches (icode, 0, r0)
4707 || !insn_operand_matches (icode, 1, r1)
4708 || !insn_operand_matches (icode, 2, c))
4709 return NULL_RTX;
4710
4711 return GEN_FCN (icode) (r0, r1, c);
4712 }
4713
4714 int
4715 have_sub2_insn (rtx x, rtx y)
4716 {
4717 enum insn_code icode;
4718
4719 gcc_assert (GET_MODE (x) != VOIDmode);
4720
4721 icode = optab_handler (sub_optab, GET_MODE (x));
4722
4723 if (icode == CODE_FOR_nothing)
4724 return 0;
4725
4726 if (!insn_operand_matches (icode, 0, x)
4727 || !insn_operand_matches (icode, 1, x)
4728 || !insn_operand_matches (icode, 2, y))
4729 return 0;
4730
4731 return 1;
4732 }
4733
4734 /* Generate the body of an instruction to copy Y into X.
4735 It may be a list of insns, if one insn isn't enough. */
4736
4737 rtx
4738 gen_move_insn (rtx x, rtx y)
4739 {
4740 rtx seq;
4741
4742 start_sequence ();
4743 emit_move_insn_1 (x, y);
4744 seq = get_insns ();
4745 end_sequence ();
4746 return seq;
4747 }
4748 \f
4749 /* Return the insn code used to extend FROM_MODE to TO_MODE.
4750 UNSIGNEDP specifies zero-extension instead of sign-extension. If
4751 no such operation exists, CODE_FOR_nothing will be returned. */
4752
4753 enum insn_code
4754 can_extend_p (enum machine_mode to_mode, enum machine_mode from_mode,
4755 int unsignedp)
4756 {
4757 convert_optab tab;
4758 #ifdef HAVE_ptr_extend
4759 if (unsignedp < 0)
4760 return CODE_FOR_ptr_extend;
4761 #endif
4762
4763 tab = unsignedp ? zext_optab : sext_optab;
4764 return convert_optab_handler (tab, to_mode, from_mode);
4765 }
4766
4767 /* Generate the body of an insn to extend Y (with mode MFROM)
4768 into X (with mode MTO). Do zero-extension if UNSIGNEDP is nonzero. */
4769
4770 rtx
4771 gen_extend_insn (rtx x, rtx y, enum machine_mode mto,
4772 enum machine_mode mfrom, int unsignedp)
4773 {
4774 enum insn_code icode = can_extend_p (mto, mfrom, unsignedp);
4775 return GEN_FCN (icode) (x, y);
4776 }
4777 \f
4778 /* can_fix_p and can_float_p say whether the target machine
4779 can directly convert a given fixed point type to
4780 a given floating point type, or vice versa.
4781 The returned value is the CODE_FOR_... value to use,
4782 or CODE_FOR_nothing if these modes cannot be directly converted.
4783
4784 *TRUNCP_PTR is set to 1 if it is necessary to output
4785 an explicit FTRUNC insn before the fix insn; otherwise 0. */
4786
4787 static enum insn_code
4788 can_fix_p (enum machine_mode fixmode, enum machine_mode fltmode,
4789 int unsignedp, int *truncp_ptr)
4790 {
4791 convert_optab tab;
4792 enum insn_code icode;
4793
4794 tab = unsignedp ? ufixtrunc_optab : sfixtrunc_optab;
4795 icode = convert_optab_handler (tab, fixmode, fltmode);
4796 if (icode != CODE_FOR_nothing)
4797 {
4798 *truncp_ptr = 0;
4799 return icode;
4800 }
4801
4802 /* FIXME: This requires a port to define both FIX and FTRUNC pattern
4803 for this to work. We need to rework the fix* and ftrunc* patterns
4804 and documentation. */
4805 tab = unsignedp ? ufix_optab : sfix_optab;
4806 icode = convert_optab_handler (tab, fixmode, fltmode);
4807 if (icode != CODE_FOR_nothing
4808 && optab_handler (ftrunc_optab, fltmode) != CODE_FOR_nothing)
4809 {
4810 *truncp_ptr = 1;
4811 return icode;
4812 }
4813
4814 *truncp_ptr = 0;
4815 return CODE_FOR_nothing;
4816 }
4817
4818 enum insn_code
4819 can_float_p (enum machine_mode fltmode, enum machine_mode fixmode,
4820 int unsignedp)
4821 {
4822 convert_optab tab;
4823
4824 tab = unsignedp ? ufloat_optab : sfloat_optab;
4825 return convert_optab_handler (tab, fltmode, fixmode);
4826 }
4827
4828 /* Function supportable_convert_operation
4829
4830 Check whether an operation represented by the code CODE is a
4831 convert operation that is supported by the target platform in
4832 vector form (i.e., when operating on arguments of type VECTYPE_IN
4833 producing a result of type VECTYPE_OUT).
4834
4835 Convert operations we currently support directly are FIX_TRUNC and FLOAT.
4836 This function checks if these operations are supported
4837 by the target platform either directly (via vector tree-codes), or via
4838 target builtins.
4839
4840 Output:
4841 - CODE1 is code of vector operation to be used when
4842 vectorizing the operation, if available.
4843 - DECL is decl of target builtin functions to be used
4844 when vectorizing the operation, if available. In this case,
4845 CODE1 is CALL_EXPR. */
4846
4847 bool
4848 supportable_convert_operation (enum tree_code code,
4849 tree vectype_out, tree vectype_in,
4850 tree *decl, enum tree_code *code1)
4851 {
4852 enum machine_mode m1,m2;
4853 int truncp;
4854
4855 m1 = TYPE_MODE (vectype_out);
4856 m2 = TYPE_MODE (vectype_in);
4857
4858 /* First check if we can done conversion directly. */
4859 if ((code == FIX_TRUNC_EXPR
4860 && can_fix_p (m1,m2,TYPE_UNSIGNED (vectype_out), &truncp)
4861 != CODE_FOR_nothing)
4862 || (code == FLOAT_EXPR
4863 && can_float_p (m1,m2,TYPE_UNSIGNED (vectype_in))
4864 != CODE_FOR_nothing))
4865 {
4866 *code1 = code;
4867 return true;
4868 }
4869
4870 /* Now check for builtin. */
4871 if (targetm.vectorize.builtin_conversion
4872 && targetm.vectorize.builtin_conversion (code, vectype_out, vectype_in))
4873 {
4874 *code1 = CALL_EXPR;
4875 *decl = targetm.vectorize.builtin_conversion (code, vectype_out, vectype_in);
4876 return true;
4877 }
4878 return false;
4879 }
4880
4881 \f
4882 /* Generate code to convert FROM to floating point
4883 and store in TO. FROM must be fixed point and not VOIDmode.
4884 UNSIGNEDP nonzero means regard FROM as unsigned.
4885 Normally this is done by correcting the final value
4886 if it is negative. */
4887
4888 void
4889 expand_float (rtx to, rtx from, int unsignedp)
4890 {
4891 enum insn_code icode;
4892 rtx target = to;
4893 enum machine_mode fmode, imode;
4894 bool can_do_signed = false;
4895
4896 /* Crash now, because we won't be able to decide which mode to use. */
4897 gcc_assert (GET_MODE (from) != VOIDmode);
4898
4899 /* Look for an insn to do the conversion. Do it in the specified
4900 modes if possible; otherwise convert either input, output or both to
4901 wider mode. If the integer mode is wider than the mode of FROM,
4902 we can do the conversion signed even if the input is unsigned. */
4903
4904 for (fmode = GET_MODE (to); fmode != VOIDmode;
4905 fmode = GET_MODE_WIDER_MODE (fmode))
4906 for (imode = GET_MODE (from); imode != VOIDmode;
4907 imode = GET_MODE_WIDER_MODE (imode))
4908 {
4909 int doing_unsigned = unsignedp;
4910
4911 if (fmode != GET_MODE (to)
4912 && significand_size (fmode) < GET_MODE_PRECISION (GET_MODE (from)))
4913 continue;
4914
4915 icode = can_float_p (fmode, imode, unsignedp);
4916 if (icode == CODE_FOR_nothing && unsignedp)
4917 {
4918 enum insn_code scode = can_float_p (fmode, imode, 0);
4919 if (scode != CODE_FOR_nothing)
4920 can_do_signed = true;
4921 if (imode != GET_MODE (from))
4922 icode = scode, doing_unsigned = 0;
4923 }
4924
4925 if (icode != CODE_FOR_nothing)
4926 {
4927 if (imode != GET_MODE (from))
4928 from = convert_to_mode (imode, from, unsignedp);
4929
4930 if (fmode != GET_MODE (to))
4931 target = gen_reg_rtx (fmode);
4932
4933 emit_unop_insn (icode, target, from,
4934 doing_unsigned ? UNSIGNED_FLOAT : FLOAT);
4935
4936 if (target != to)
4937 convert_move (to, target, 0);
4938 return;
4939 }
4940 }
4941
4942 /* Unsigned integer, and no way to convert directly. Convert as signed,
4943 then unconditionally adjust the result. */
4944 if (unsignedp && can_do_signed)
4945 {
4946 rtx label = gen_label_rtx ();
4947 rtx temp;
4948 REAL_VALUE_TYPE offset;
4949
4950 /* Look for a usable floating mode FMODE wider than the source and at
4951 least as wide as the target. Using FMODE will avoid rounding woes
4952 with unsigned values greater than the signed maximum value. */
4953
4954 for (fmode = GET_MODE (to); fmode != VOIDmode;
4955 fmode = GET_MODE_WIDER_MODE (fmode))
4956 if (GET_MODE_PRECISION (GET_MODE (from)) < GET_MODE_BITSIZE (fmode)
4957 && can_float_p (fmode, GET_MODE (from), 0) != CODE_FOR_nothing)
4958 break;
4959
4960 if (fmode == VOIDmode)
4961 {
4962 /* There is no such mode. Pretend the target is wide enough. */
4963 fmode = GET_MODE (to);
4964
4965 /* Avoid double-rounding when TO is narrower than FROM. */
4966 if ((significand_size (fmode) + 1)
4967 < GET_MODE_PRECISION (GET_MODE (from)))
4968 {
4969 rtx temp1;
4970 rtx neglabel = gen_label_rtx ();
4971
4972 /* Don't use TARGET if it isn't a register, is a hard register,
4973 or is the wrong mode. */
4974 if (!REG_P (target)
4975 || REGNO (target) < FIRST_PSEUDO_REGISTER
4976 || GET_MODE (target) != fmode)
4977 target = gen_reg_rtx (fmode);
4978
4979 imode = GET_MODE (from);
4980 do_pending_stack_adjust ();
4981
4982 /* Test whether the sign bit is set. */
4983 emit_cmp_and_jump_insns (from, const0_rtx, LT, NULL_RTX, imode,
4984 0, neglabel);
4985
4986 /* The sign bit is not set. Convert as signed. */
4987 expand_float (target, from, 0);
4988 emit_jump_insn (gen_jump (label));
4989 emit_barrier ();
4990
4991 /* The sign bit is set.
4992 Convert to a usable (positive signed) value by shifting right
4993 one bit, while remembering if a nonzero bit was shifted
4994 out; i.e., compute (from & 1) | (from >> 1). */
4995
4996 emit_label (neglabel);
4997 temp = expand_binop (imode, and_optab, from, const1_rtx,
4998 NULL_RTX, 1, OPTAB_LIB_WIDEN);
4999 temp1 = expand_shift (RSHIFT_EXPR, imode, from, 1, NULL_RTX, 1);
5000 temp = expand_binop (imode, ior_optab, temp, temp1, temp, 1,
5001 OPTAB_LIB_WIDEN);
5002 expand_float (target, temp, 0);
5003
5004 /* Multiply by 2 to undo the shift above. */
5005 temp = expand_binop (fmode, add_optab, target, target,
5006 target, 0, OPTAB_LIB_WIDEN);
5007 if (temp != target)
5008 emit_move_insn (target, temp);
5009
5010 do_pending_stack_adjust ();
5011 emit_label (label);
5012 goto done;
5013 }
5014 }
5015
5016 /* If we are about to do some arithmetic to correct for an
5017 unsigned operand, do it in a pseudo-register. */
5018
5019 if (GET_MODE (to) != fmode
5020 || !REG_P (to) || REGNO (to) < FIRST_PSEUDO_REGISTER)
5021 target = gen_reg_rtx (fmode);
5022
5023 /* Convert as signed integer to floating. */
5024 expand_float (target, from, 0);
5025
5026 /* If FROM is negative (and therefore TO is negative),
5027 correct its value by 2**bitwidth. */
5028
5029 do_pending_stack_adjust ();
5030 emit_cmp_and_jump_insns (from, const0_rtx, GE, NULL_RTX, GET_MODE (from),
5031 0, label);
5032
5033
5034 real_2expN (&offset, GET_MODE_PRECISION (GET_MODE (from)), fmode);
5035 temp = expand_binop (fmode, add_optab, target,
5036 CONST_DOUBLE_FROM_REAL_VALUE (offset, fmode),
5037 target, 0, OPTAB_LIB_WIDEN);
5038 if (temp != target)
5039 emit_move_insn (target, temp);
5040
5041 do_pending_stack_adjust ();
5042 emit_label (label);
5043 goto done;
5044 }
5045
5046 /* No hardware instruction available; call a library routine. */
5047 {
5048 rtx libfunc;
5049 rtx insns;
5050 rtx value;
5051 convert_optab tab = unsignedp ? ufloat_optab : sfloat_optab;
5052
5053 if (GET_MODE_SIZE (GET_MODE (from)) < GET_MODE_SIZE (SImode))
5054 from = convert_to_mode (SImode, from, unsignedp);
5055
5056 libfunc = convert_optab_libfunc (tab, GET_MODE (to), GET_MODE (from));
5057 gcc_assert (libfunc);
5058
5059 start_sequence ();
5060
5061 value = emit_library_call_value (libfunc, NULL_RTX, LCT_CONST,
5062 GET_MODE (to), 1, from,
5063 GET_MODE (from));
5064 insns = get_insns ();
5065 end_sequence ();
5066
5067 emit_libcall_block (insns, target, value,
5068 gen_rtx_fmt_e (unsignedp ? UNSIGNED_FLOAT : FLOAT,
5069 GET_MODE (to), from));
5070 }
5071
5072 done:
5073
5074 /* Copy result to requested destination
5075 if we have been computing in a temp location. */
5076
5077 if (target != to)
5078 {
5079 if (GET_MODE (target) == GET_MODE (to))
5080 emit_move_insn (to, target);
5081 else
5082 convert_move (to, target, 0);
5083 }
5084 }
5085 \f
5086 /* Generate code to convert FROM to fixed point and store in TO. FROM
5087 must be floating point. */
5088
5089 void
5090 expand_fix (rtx to, rtx from, int unsignedp)
5091 {
5092 enum insn_code icode;
5093 rtx target = to;
5094 enum machine_mode fmode, imode;
5095 int must_trunc = 0;
5096
5097 /* We first try to find a pair of modes, one real and one integer, at
5098 least as wide as FROM and TO, respectively, in which we can open-code
5099 this conversion. If the integer mode is wider than the mode of TO,
5100 we can do the conversion either signed or unsigned. */
5101
5102 for (fmode = GET_MODE (from); fmode != VOIDmode;
5103 fmode = GET_MODE_WIDER_MODE (fmode))
5104 for (imode = GET_MODE (to); imode != VOIDmode;
5105 imode = GET_MODE_WIDER_MODE (imode))
5106 {
5107 int doing_unsigned = unsignedp;
5108
5109 icode = can_fix_p (imode, fmode, unsignedp, &must_trunc);
5110 if (icode == CODE_FOR_nothing && imode != GET_MODE (to) && unsignedp)
5111 icode = can_fix_p (imode, fmode, 0, &must_trunc), doing_unsigned = 0;
5112
5113 if (icode != CODE_FOR_nothing)
5114 {
5115 rtx last = get_last_insn ();
5116 if (fmode != GET_MODE (from))
5117 from = convert_to_mode (fmode, from, 0);
5118
5119 if (must_trunc)
5120 {
5121 rtx temp = gen_reg_rtx (GET_MODE (from));
5122 from = expand_unop (GET_MODE (from), ftrunc_optab, from,
5123 temp, 0);
5124 }
5125
5126 if (imode != GET_MODE (to))
5127 target = gen_reg_rtx (imode);
5128
5129 if (maybe_emit_unop_insn (icode, target, from,
5130 doing_unsigned ? UNSIGNED_FIX : FIX))
5131 {
5132 if (target != to)
5133 convert_move (to, target, unsignedp);
5134 return;
5135 }
5136 delete_insns_since (last);
5137 }
5138 }
5139
5140 /* For an unsigned conversion, there is one more way to do it.
5141 If we have a signed conversion, we generate code that compares
5142 the real value to the largest representable positive number. If if
5143 is smaller, the conversion is done normally. Otherwise, subtract
5144 one plus the highest signed number, convert, and add it back.
5145
5146 We only need to check all real modes, since we know we didn't find
5147 anything with a wider integer mode.
5148
5149 This code used to extend FP value into mode wider than the destination.
5150 This is needed for decimal float modes which cannot accurately
5151 represent one plus the highest signed number of the same size, but
5152 not for binary modes. Consider, for instance conversion from SFmode
5153 into DImode.
5154
5155 The hot path through the code is dealing with inputs smaller than 2^63
5156 and doing just the conversion, so there is no bits to lose.
5157
5158 In the other path we know the value is positive in the range 2^63..2^64-1
5159 inclusive. (as for other input overflow happens and result is undefined)
5160 So we know that the most important bit set in mantissa corresponds to
5161 2^63. The subtraction of 2^63 should not generate any rounding as it
5162 simply clears out that bit. The rest is trivial. */
5163
5164 if (unsignedp && GET_MODE_PRECISION (GET_MODE (to)) <= HOST_BITS_PER_WIDE_INT)
5165 for (fmode = GET_MODE (from); fmode != VOIDmode;
5166 fmode = GET_MODE_WIDER_MODE (fmode))
5167 if (CODE_FOR_nothing != can_fix_p (GET_MODE (to), fmode, 0, &must_trunc)
5168 && (!DECIMAL_FLOAT_MODE_P (fmode)
5169 || GET_MODE_BITSIZE (fmode) > GET_MODE_PRECISION (GET_MODE (to))))
5170 {
5171 int bitsize;
5172 REAL_VALUE_TYPE offset;
5173 rtx limit, lab1, lab2, insn;
5174
5175 bitsize = GET_MODE_PRECISION (GET_MODE (to));
5176 real_2expN (&offset, bitsize - 1, fmode);
5177 limit = CONST_DOUBLE_FROM_REAL_VALUE (offset, fmode);
5178 lab1 = gen_label_rtx ();
5179 lab2 = gen_label_rtx ();
5180
5181 if (fmode != GET_MODE (from))
5182 from = convert_to_mode (fmode, from, 0);
5183
5184 /* See if we need to do the subtraction. */
5185 do_pending_stack_adjust ();
5186 emit_cmp_and_jump_insns (from, limit, GE, NULL_RTX, GET_MODE (from),
5187 0, lab1);
5188
5189 /* If not, do the signed "fix" and branch around fixup code. */
5190 expand_fix (to, from, 0);
5191 emit_jump_insn (gen_jump (lab2));
5192 emit_barrier ();
5193
5194 /* Otherwise, subtract 2**(N-1), convert to signed number,
5195 then add 2**(N-1). Do the addition using XOR since this
5196 will often generate better code. */
5197 emit_label (lab1);
5198 target = expand_binop (GET_MODE (from), sub_optab, from, limit,
5199 NULL_RTX, 0, OPTAB_LIB_WIDEN);
5200 expand_fix (to, target, 0);
5201 target = expand_binop (GET_MODE (to), xor_optab, to,
5202 gen_int_mode
5203 ((HOST_WIDE_INT) 1 << (bitsize - 1),
5204 GET_MODE (to)),
5205 to, 1, OPTAB_LIB_WIDEN);
5206
5207 if (target != to)
5208 emit_move_insn (to, target);
5209
5210 emit_label (lab2);
5211
5212 if (optab_handler (mov_optab, GET_MODE (to)) != CODE_FOR_nothing)
5213 {
5214 /* Make a place for a REG_NOTE and add it. */
5215 insn = emit_move_insn (to, to);
5216 set_unique_reg_note (insn,
5217 REG_EQUAL,
5218 gen_rtx_fmt_e (UNSIGNED_FIX,
5219 GET_MODE (to),
5220 copy_rtx (from)));
5221 }
5222
5223 return;
5224 }
5225
5226 /* We can't do it with an insn, so use a library call. But first ensure
5227 that the mode of TO is at least as wide as SImode, since those are the
5228 only library calls we know about. */
5229
5230 if (GET_MODE_SIZE (GET_MODE (to)) < GET_MODE_SIZE (SImode))
5231 {
5232 target = gen_reg_rtx (SImode);
5233
5234 expand_fix (target, from, unsignedp);
5235 }
5236 else
5237 {
5238 rtx insns;
5239 rtx value;
5240 rtx libfunc;
5241
5242 convert_optab tab = unsignedp ? ufix_optab : sfix_optab;
5243 libfunc = convert_optab_libfunc (tab, GET_MODE (to), GET_MODE (from));
5244 gcc_assert (libfunc);
5245
5246 start_sequence ();
5247
5248 value = emit_library_call_value (libfunc, NULL_RTX, LCT_CONST,
5249 GET_MODE (to), 1, from,
5250 GET_MODE (from));
5251 insns = get_insns ();
5252 end_sequence ();
5253
5254 emit_libcall_block (insns, target, value,
5255 gen_rtx_fmt_e (unsignedp ? UNSIGNED_FIX : FIX,
5256 GET_MODE (to), from));
5257 }
5258
5259 if (target != to)
5260 {
5261 if (GET_MODE (to) == GET_MODE (target))
5262 emit_move_insn (to, target);
5263 else
5264 convert_move (to, target, 0);
5265 }
5266 }
5267
5268 /* Generate code to convert FROM or TO a fixed-point.
5269 If UINTP is true, either TO or FROM is an unsigned integer.
5270 If SATP is true, we need to saturate the result. */
5271
5272 void
5273 expand_fixed_convert (rtx to, rtx from, int uintp, int satp)
5274 {
5275 enum machine_mode to_mode = GET_MODE (to);
5276 enum machine_mode from_mode = GET_MODE (from);
5277 convert_optab tab;
5278 enum rtx_code this_code;
5279 enum insn_code code;
5280 rtx insns, value;
5281 rtx libfunc;
5282
5283 if (to_mode == from_mode)
5284 {
5285 emit_move_insn (to, from);
5286 return;
5287 }
5288
5289 if (uintp)
5290 {
5291 tab = satp ? satfractuns_optab : fractuns_optab;
5292 this_code = satp ? UNSIGNED_SAT_FRACT : UNSIGNED_FRACT_CONVERT;
5293 }
5294 else
5295 {
5296 tab = satp ? satfract_optab : fract_optab;
5297 this_code = satp ? SAT_FRACT : FRACT_CONVERT;
5298 }
5299 code = convert_optab_handler (tab, to_mode, from_mode);
5300 if (code != CODE_FOR_nothing)
5301 {
5302 emit_unop_insn (code, to, from, this_code);
5303 return;
5304 }
5305
5306 libfunc = convert_optab_libfunc (tab, to_mode, from_mode);
5307 gcc_assert (libfunc);
5308
5309 start_sequence ();
5310 value = emit_library_call_value (libfunc, NULL_RTX, LCT_CONST, to_mode,
5311 1, from, from_mode);
5312 insns = get_insns ();
5313 end_sequence ();
5314
5315 emit_libcall_block (insns, to, value,
5316 gen_rtx_fmt_e (tab->code, to_mode, from));
5317 }
5318
5319 /* Generate code to convert FROM to fixed point and store in TO. FROM
5320 must be floating point, TO must be signed. Use the conversion optab
5321 TAB to do the conversion. */
5322
5323 bool
5324 expand_sfix_optab (rtx to, rtx from, convert_optab tab)
5325 {
5326 enum insn_code icode;
5327 rtx target = to;
5328 enum machine_mode fmode, imode;
5329
5330 /* We first try to find a pair of modes, one real and one integer, at
5331 least as wide as FROM and TO, respectively, in which we can open-code
5332 this conversion. If the integer mode is wider than the mode of TO,
5333 we can do the conversion either signed or unsigned. */
5334
5335 for (fmode = GET_MODE (from); fmode != VOIDmode;
5336 fmode = GET_MODE_WIDER_MODE (fmode))
5337 for (imode = GET_MODE (to); imode != VOIDmode;
5338 imode = GET_MODE_WIDER_MODE (imode))
5339 {
5340 icode = convert_optab_handler (tab, imode, fmode);
5341 if (icode != CODE_FOR_nothing)
5342 {
5343 rtx last = get_last_insn ();
5344 if (fmode != GET_MODE (from))
5345 from = convert_to_mode (fmode, from, 0);
5346
5347 if (imode != GET_MODE (to))
5348 target = gen_reg_rtx (imode);
5349
5350 if (!maybe_emit_unop_insn (icode, target, from, UNKNOWN))
5351 {
5352 delete_insns_since (last);
5353 continue;
5354 }
5355 if (target != to)
5356 convert_move (to, target, 0);
5357 return true;
5358 }
5359 }
5360
5361 return false;
5362 }
5363 \f
5364 /* Report whether we have an instruction to perform the operation
5365 specified by CODE on operands of mode MODE. */
5366 int
5367 have_insn_for (enum rtx_code code, enum machine_mode mode)
5368 {
5369 return (code_to_optab[(int) code] != 0
5370 && (optab_handler (code_to_optab[(int) code], mode)
5371 != CODE_FOR_nothing));
5372 }
5373
5374 /* Set all insn_code fields to CODE_FOR_nothing. */
5375
5376 static void
5377 init_insn_codes (void)
5378 {
5379 memset (optab_table, 0, sizeof (optab_table));
5380 memset (convert_optab_table, 0, sizeof (convert_optab_table));
5381 memset (direct_optab_table, 0, sizeof (direct_optab_table));
5382 }
5383
5384 /* Initialize OP's code to CODE, and write it into the code_to_optab table. */
5385 static inline void
5386 init_optab (optab op, enum rtx_code code)
5387 {
5388 op->code = code;
5389 code_to_optab[(int) code] = op;
5390 }
5391
5392 /* Same, but fill in its code as CODE, and do _not_ write it into
5393 the code_to_optab table. */
5394 static inline void
5395 init_optabv (optab op, enum rtx_code code)
5396 {
5397 op->code = code;
5398 }
5399
5400 /* Conversion optabs never go in the code_to_optab table. */
5401 static void
5402 init_convert_optab (convert_optab op, enum rtx_code code)
5403 {
5404 op->code = code;
5405 }
5406
5407 /* Initialize the libfunc fields of an entire group of entries in some
5408 optab. Each entry is set equal to a string consisting of a leading
5409 pair of underscores followed by a generic operation name followed by
5410 a mode name (downshifted to lowercase) followed by a single character
5411 representing the number of operands for the given operation (which is
5412 usually one of the characters '2', '3', or '4').
5413
5414 OPTABLE is the table in which libfunc fields are to be initialized.
5415 OPNAME is the generic (string) name of the operation.
5416 SUFFIX is the character which specifies the number of operands for
5417 the given generic operation.
5418 MODE is the mode to generate for.
5419 */
5420
5421 static void
5422 gen_libfunc (optab optable, const char *opname, int suffix, enum machine_mode mode)
5423 {
5424 unsigned opname_len = strlen (opname);
5425 const char *mname = GET_MODE_NAME (mode);
5426 unsigned mname_len = strlen (mname);
5427 int prefix_len = targetm.libfunc_gnu_prefix ? 6 : 2;
5428 int len = prefix_len + opname_len + mname_len + 1 + 1;
5429 char *libfunc_name = XALLOCAVEC (char, len);
5430 char *p;
5431 const char *q;
5432
5433 p = libfunc_name;
5434 *p++ = '_';
5435 *p++ = '_';
5436 if (targetm.libfunc_gnu_prefix)
5437 {
5438 *p++ = 'g';
5439 *p++ = 'n';
5440 *p++ = 'u';
5441 *p++ = '_';
5442 }
5443 for (q = opname; *q; )
5444 *p++ = *q++;
5445 for (q = mname; *q; q++)
5446 *p++ = TOLOWER (*q);
5447 *p++ = suffix;
5448 *p = '\0';
5449
5450 set_optab_libfunc (optable, mode,
5451 ggc_alloc_string (libfunc_name, p - libfunc_name));
5452 }
5453
5454 /* Like gen_libfunc, but verify that integer operation is involved. */
5455
5456 static void
5457 gen_int_libfunc (optab optable, const char *opname, char suffix,
5458 enum machine_mode mode)
5459 {
5460 int maxsize = 2 * BITS_PER_WORD;
5461
5462 if (GET_MODE_CLASS (mode) != MODE_INT)
5463 return;
5464 if (maxsize < LONG_LONG_TYPE_SIZE)
5465 maxsize = LONG_LONG_TYPE_SIZE;
5466 if (GET_MODE_CLASS (mode) != MODE_INT
5467 || mode < word_mode || GET_MODE_BITSIZE (mode) > maxsize)
5468 return;
5469 gen_libfunc (optable, opname, suffix, mode);
5470 }
5471
5472 /* Like gen_libfunc, but verify that FP and set decimal prefix if needed. */
5473
5474 static void
5475 gen_fp_libfunc (optab optable, const char *opname, char suffix,
5476 enum machine_mode mode)
5477 {
5478 char *dec_opname;
5479
5480 if (GET_MODE_CLASS (mode) == MODE_FLOAT)
5481 gen_libfunc (optable, opname, suffix, mode);
5482 if (DECIMAL_FLOAT_MODE_P (mode))
5483 {
5484 dec_opname = XALLOCAVEC (char, sizeof (DECIMAL_PREFIX) + strlen (opname));
5485 /* For BID support, change the name to have either a bid_ or dpd_ prefix
5486 depending on the low level floating format used. */
5487 memcpy (dec_opname, DECIMAL_PREFIX, sizeof (DECIMAL_PREFIX) - 1);
5488 strcpy (dec_opname + sizeof (DECIMAL_PREFIX) - 1, opname);
5489 gen_libfunc (optable, dec_opname, suffix, mode);
5490 }
5491 }
5492
5493 /* Like gen_libfunc, but verify that fixed-point operation is involved. */
5494
5495 static void
5496 gen_fixed_libfunc (optab optable, const char *opname, char suffix,
5497 enum machine_mode mode)
5498 {
5499 if (!ALL_FIXED_POINT_MODE_P (mode))
5500 return;
5501 gen_libfunc (optable, opname, suffix, mode);
5502 }
5503
5504 /* Like gen_libfunc, but verify that signed fixed-point operation is
5505 involved. */
5506
5507 static void
5508 gen_signed_fixed_libfunc (optab optable, const char *opname, char suffix,
5509 enum machine_mode mode)
5510 {
5511 if (!SIGNED_FIXED_POINT_MODE_P (mode))
5512 return;
5513 gen_libfunc (optable, opname, suffix, mode);
5514 }
5515
5516 /* Like gen_libfunc, but verify that unsigned fixed-point operation is
5517 involved. */
5518
5519 static void
5520 gen_unsigned_fixed_libfunc (optab optable, const char *opname, char suffix,
5521 enum machine_mode mode)
5522 {
5523 if (!UNSIGNED_FIXED_POINT_MODE_P (mode))
5524 return;
5525 gen_libfunc (optable, opname, suffix, mode);
5526 }
5527
5528 /* Like gen_libfunc, but verify that FP or INT operation is involved. */
5529
5530 static void
5531 gen_int_fp_libfunc (optab optable, const char *name, char suffix,
5532 enum machine_mode mode)
5533 {
5534 if (DECIMAL_FLOAT_MODE_P (mode) || GET_MODE_CLASS (mode) == MODE_FLOAT)
5535 gen_fp_libfunc (optable, name, suffix, mode);
5536 if (INTEGRAL_MODE_P (mode))
5537 gen_int_libfunc (optable, name, suffix, mode);
5538 }
5539
5540 /* Like gen_libfunc, but verify that FP or INT operation is involved
5541 and add 'v' suffix for integer operation. */
5542
5543 static void
5544 gen_intv_fp_libfunc (optab optable, const char *name, char suffix,
5545 enum machine_mode mode)
5546 {
5547 if (DECIMAL_FLOAT_MODE_P (mode) || GET_MODE_CLASS (mode) == MODE_FLOAT)
5548 gen_fp_libfunc (optable, name, suffix, mode);
5549 if (GET_MODE_CLASS (mode) == MODE_INT)
5550 {
5551 int len = strlen (name);
5552 char *v_name = XALLOCAVEC (char, len + 2);
5553 strcpy (v_name, name);
5554 v_name[len] = 'v';
5555 v_name[len + 1] = 0;
5556 gen_int_libfunc (optable, v_name, suffix, mode);
5557 }
5558 }
5559
5560 /* Like gen_libfunc, but verify that FP or INT or FIXED operation is
5561 involved. */
5562
5563 static void
5564 gen_int_fp_fixed_libfunc (optab optable, const char *name, char suffix,
5565 enum machine_mode mode)
5566 {
5567 if (DECIMAL_FLOAT_MODE_P (mode) || GET_MODE_CLASS (mode) == MODE_FLOAT)
5568 gen_fp_libfunc (optable, name, suffix, mode);
5569 if (INTEGRAL_MODE_P (mode))
5570 gen_int_libfunc (optable, name, suffix, mode);
5571 if (ALL_FIXED_POINT_MODE_P (mode))
5572 gen_fixed_libfunc (optable, name, suffix, mode);
5573 }
5574
5575 /* Like gen_libfunc, but verify that FP or INT or signed FIXED operation is
5576 involved. */
5577
5578 static void
5579 gen_int_fp_signed_fixed_libfunc (optab optable, const char *name, char suffix,
5580 enum machine_mode mode)
5581 {
5582 if (DECIMAL_FLOAT_MODE_P (mode) || GET_MODE_CLASS (mode) == MODE_FLOAT)
5583 gen_fp_libfunc (optable, name, suffix, mode);
5584 if (INTEGRAL_MODE_P (mode))
5585 gen_int_libfunc (optable, name, suffix, mode);
5586 if (SIGNED_FIXED_POINT_MODE_P (mode))
5587 gen_signed_fixed_libfunc (optable, name, suffix, mode);
5588 }
5589
5590 /* Like gen_libfunc, but verify that INT or FIXED operation is
5591 involved. */
5592
5593 static void
5594 gen_int_fixed_libfunc (optab optable, const char *name, char suffix,
5595 enum machine_mode mode)
5596 {
5597 if (INTEGRAL_MODE_P (mode))
5598 gen_int_libfunc (optable, name, suffix, mode);
5599 if (ALL_FIXED_POINT_MODE_P (mode))
5600 gen_fixed_libfunc (optable, name, suffix, mode);
5601 }
5602
5603 /* Like gen_libfunc, but verify that INT or signed FIXED operation is
5604 involved. */
5605
5606 static void
5607 gen_int_signed_fixed_libfunc (optab optable, const char *name, char suffix,
5608 enum machine_mode mode)
5609 {
5610 if (INTEGRAL_MODE_P (mode))
5611 gen_int_libfunc (optable, name, suffix, mode);
5612 if (SIGNED_FIXED_POINT_MODE_P (mode))
5613 gen_signed_fixed_libfunc (optable, name, suffix, mode);
5614 }
5615
5616 /* Like gen_libfunc, but verify that INT or unsigned FIXED operation is
5617 involved. */
5618
5619 static void
5620 gen_int_unsigned_fixed_libfunc (optab optable, const char *name, char suffix,
5621 enum machine_mode mode)
5622 {
5623 if (INTEGRAL_MODE_P (mode))
5624 gen_int_libfunc (optable, name, suffix, mode);
5625 if (UNSIGNED_FIXED_POINT_MODE_P (mode))
5626 gen_unsigned_fixed_libfunc (optable, name, suffix, mode);
5627 }
5628
5629 /* Initialize the libfunc fields of an entire group of entries of an
5630 inter-mode-class conversion optab. The string formation rules are
5631 similar to the ones for init_libfuncs, above, but instead of having
5632 a mode name and an operand count these functions have two mode names
5633 and no operand count. */
5634
5635 static void
5636 gen_interclass_conv_libfunc (convert_optab tab,
5637 const char *opname,
5638 enum machine_mode tmode,
5639 enum machine_mode fmode)
5640 {
5641 size_t opname_len = strlen (opname);
5642 size_t mname_len = 0;
5643
5644 const char *fname, *tname;
5645 const char *q;
5646 int prefix_len = targetm.libfunc_gnu_prefix ? 6 : 2;
5647 char *libfunc_name, *suffix;
5648 char *nondec_name, *dec_name, *nondec_suffix, *dec_suffix;
5649 char *p;
5650
5651 /* If this is a decimal conversion, add the current BID vs. DPD prefix that
5652 depends on which underlying decimal floating point format is used. */
5653 const size_t dec_len = sizeof (DECIMAL_PREFIX) - 1;
5654
5655 mname_len = strlen (GET_MODE_NAME (tmode)) + strlen (GET_MODE_NAME (fmode));
5656
5657 nondec_name = XALLOCAVEC (char, prefix_len + opname_len + mname_len + 1 + 1);
5658 nondec_name[0] = '_';
5659 nondec_name[1] = '_';
5660 if (targetm.libfunc_gnu_prefix)
5661 {
5662 nondec_name[2] = 'g';
5663 nondec_name[3] = 'n';
5664 nondec_name[4] = 'u';
5665 nondec_name[5] = '_';
5666 }
5667
5668 memcpy (&nondec_name[prefix_len], opname, opname_len);
5669 nondec_suffix = nondec_name + opname_len + prefix_len;
5670
5671 dec_name = XALLOCAVEC (char, 2 + dec_len + opname_len + mname_len + 1 + 1);
5672 dec_name[0] = '_';
5673 dec_name[1] = '_';
5674 memcpy (&dec_name[2], DECIMAL_PREFIX, dec_len);
5675 memcpy (&dec_name[2+dec_len], opname, opname_len);
5676 dec_suffix = dec_name + dec_len + opname_len + 2;
5677
5678 fname = GET_MODE_NAME (fmode);
5679 tname = GET_MODE_NAME (tmode);
5680
5681 if (DECIMAL_FLOAT_MODE_P(fmode) || DECIMAL_FLOAT_MODE_P(tmode))
5682 {
5683 libfunc_name = dec_name;
5684 suffix = dec_suffix;
5685 }
5686 else
5687 {
5688 libfunc_name = nondec_name;
5689 suffix = nondec_suffix;
5690 }
5691
5692 p = suffix;
5693 for (q = fname; *q; p++, q++)
5694 *p = TOLOWER (*q);
5695 for (q = tname; *q; p++, q++)
5696 *p = TOLOWER (*q);
5697
5698 *p = '\0';
5699
5700 set_conv_libfunc (tab, tmode, fmode,
5701 ggc_alloc_string (libfunc_name, p - libfunc_name));
5702 }
5703
5704 /* Same as gen_interclass_conv_libfunc but verify that we are producing
5705 int->fp conversion. */
5706
5707 static void
5708 gen_int_to_fp_conv_libfunc (convert_optab tab,
5709 const char *opname,
5710 enum machine_mode tmode,
5711 enum machine_mode fmode)
5712 {
5713 if (GET_MODE_CLASS (fmode) != MODE_INT)
5714 return;
5715 if (GET_MODE_CLASS (tmode) != MODE_FLOAT && !DECIMAL_FLOAT_MODE_P (tmode))
5716 return;
5717 gen_interclass_conv_libfunc (tab, opname, tmode, fmode);
5718 }
5719
5720 /* ufloat_optab is special by using floatun for FP and floatuns decimal fp
5721 naming scheme. */
5722
5723 static void
5724 gen_ufloat_conv_libfunc (convert_optab tab,
5725 const char *opname ATTRIBUTE_UNUSED,
5726 enum machine_mode tmode,
5727 enum machine_mode fmode)
5728 {
5729 if (DECIMAL_FLOAT_MODE_P (tmode))
5730 gen_int_to_fp_conv_libfunc (tab, "floatuns", tmode, fmode);
5731 else
5732 gen_int_to_fp_conv_libfunc (tab, "floatun", tmode, fmode);
5733 }
5734
5735 /* Same as gen_interclass_conv_libfunc but verify that we are producing
5736 fp->int conversion. */
5737
5738 static void
5739 gen_int_to_fp_nondecimal_conv_libfunc (convert_optab tab,
5740 const char *opname,
5741 enum machine_mode tmode,
5742 enum machine_mode fmode)
5743 {
5744 if (GET_MODE_CLASS (fmode) != MODE_INT)
5745 return;
5746 if (GET_MODE_CLASS (tmode) != MODE_FLOAT)
5747 return;
5748 gen_interclass_conv_libfunc (tab, opname, tmode, fmode);
5749 }
5750
5751 /* Same as gen_interclass_conv_libfunc but verify that we are producing
5752 fp->int conversion with no decimal floating point involved. */
5753
5754 static void
5755 gen_fp_to_int_conv_libfunc (convert_optab tab,
5756 const char *opname,
5757 enum machine_mode tmode,
5758 enum machine_mode fmode)
5759 {
5760 if (GET_MODE_CLASS (fmode) != MODE_FLOAT && !DECIMAL_FLOAT_MODE_P (fmode))
5761 return;
5762 if (GET_MODE_CLASS (tmode) != MODE_INT)
5763 return;
5764 gen_interclass_conv_libfunc (tab, opname, tmode, fmode);
5765 }
5766
5767 /* Initialize the libfunc fields of an of an intra-mode-class conversion optab.
5768 The string formation rules are
5769 similar to the ones for init_libfunc, above. */
5770
5771 static void
5772 gen_intraclass_conv_libfunc (convert_optab tab, const char *opname,
5773 enum machine_mode tmode, enum machine_mode fmode)
5774 {
5775 size_t opname_len = strlen (opname);
5776 size_t mname_len = 0;
5777
5778 const char *fname, *tname;
5779 const char *q;
5780 int prefix_len = targetm.libfunc_gnu_prefix ? 6 : 2;
5781 char *nondec_name, *dec_name, *nondec_suffix, *dec_suffix;
5782 char *libfunc_name, *suffix;
5783 char *p;
5784
5785 /* If this is a decimal conversion, add the current BID vs. DPD prefix that
5786 depends on which underlying decimal floating point format is used. */
5787 const size_t dec_len = sizeof (DECIMAL_PREFIX) - 1;
5788
5789 mname_len = strlen (GET_MODE_NAME (tmode)) + strlen (GET_MODE_NAME (fmode));
5790
5791 nondec_name = XALLOCAVEC (char, 2 + opname_len + mname_len + 1 + 1);
5792 nondec_name[0] = '_';
5793 nondec_name[1] = '_';
5794 if (targetm.libfunc_gnu_prefix)
5795 {
5796 nondec_name[2] = 'g';
5797 nondec_name[3] = 'n';
5798 nondec_name[4] = 'u';
5799 nondec_name[5] = '_';
5800 }
5801 memcpy (&nondec_name[prefix_len], opname, opname_len);
5802 nondec_suffix = nondec_name + opname_len + prefix_len;
5803
5804 dec_name = XALLOCAVEC (char, 2 + dec_len + opname_len + mname_len + 1 + 1);
5805 dec_name[0] = '_';
5806 dec_name[1] = '_';
5807 memcpy (&dec_name[2], DECIMAL_PREFIX, dec_len);
5808 memcpy (&dec_name[2 + dec_len], opname, opname_len);
5809 dec_suffix = dec_name + dec_len + opname_len + 2;
5810
5811 fname = GET_MODE_NAME (fmode);
5812 tname = GET_MODE_NAME (tmode);
5813
5814 if (DECIMAL_FLOAT_MODE_P(fmode) || DECIMAL_FLOAT_MODE_P(tmode))
5815 {
5816 libfunc_name = dec_name;
5817 suffix = dec_suffix;
5818 }
5819 else
5820 {
5821 libfunc_name = nondec_name;
5822 suffix = nondec_suffix;
5823 }
5824
5825 p = suffix;
5826 for (q = fname; *q; p++, q++)
5827 *p = TOLOWER (*q);
5828 for (q = tname; *q; p++, q++)
5829 *p = TOLOWER (*q);
5830
5831 *p++ = '2';
5832 *p = '\0';
5833
5834 set_conv_libfunc (tab, tmode, fmode,
5835 ggc_alloc_string (libfunc_name, p - libfunc_name));
5836 }
5837
5838 /* Pick proper libcall for trunc_optab. We need to chose if we do
5839 truncation or extension and interclass or intraclass. */
5840
5841 static void
5842 gen_trunc_conv_libfunc (convert_optab tab,
5843 const char *opname,
5844 enum machine_mode tmode,
5845 enum machine_mode fmode)
5846 {
5847 if (GET_MODE_CLASS (tmode) != MODE_FLOAT && !DECIMAL_FLOAT_MODE_P (tmode))
5848 return;
5849 if (GET_MODE_CLASS (fmode) != MODE_FLOAT && !DECIMAL_FLOAT_MODE_P (fmode))
5850 return;
5851 if (tmode == fmode)
5852 return;
5853
5854 if ((GET_MODE_CLASS (tmode) == MODE_FLOAT && DECIMAL_FLOAT_MODE_P (fmode))
5855 || (GET_MODE_CLASS (fmode) == MODE_FLOAT && DECIMAL_FLOAT_MODE_P (tmode)))
5856 gen_interclass_conv_libfunc (tab, opname, tmode, fmode);
5857
5858 if (GET_MODE_PRECISION (fmode) <= GET_MODE_PRECISION (tmode))
5859 return;
5860
5861 if ((GET_MODE_CLASS (tmode) == MODE_FLOAT
5862 && GET_MODE_CLASS (fmode) == MODE_FLOAT)
5863 || (DECIMAL_FLOAT_MODE_P (fmode) && DECIMAL_FLOAT_MODE_P (tmode)))
5864 gen_intraclass_conv_libfunc (tab, opname, tmode, fmode);
5865 }
5866
5867 /* Pick proper libcall for extend_optab. We need to chose if we do
5868 truncation or extension and interclass or intraclass. */
5869
5870 static void
5871 gen_extend_conv_libfunc (convert_optab tab,
5872 const char *opname ATTRIBUTE_UNUSED,
5873 enum machine_mode tmode,
5874 enum machine_mode fmode)
5875 {
5876 if (GET_MODE_CLASS (tmode) != MODE_FLOAT && !DECIMAL_FLOAT_MODE_P (tmode))
5877 return;
5878 if (GET_MODE_CLASS (fmode) != MODE_FLOAT && !DECIMAL_FLOAT_MODE_P (fmode))
5879 return;
5880 if (tmode == fmode)
5881 return;
5882
5883 if ((GET_MODE_CLASS (tmode) == MODE_FLOAT && DECIMAL_FLOAT_MODE_P (fmode))
5884 || (GET_MODE_CLASS (fmode) == MODE_FLOAT && DECIMAL_FLOAT_MODE_P (tmode)))
5885 gen_interclass_conv_libfunc (tab, opname, tmode, fmode);
5886
5887 if (GET_MODE_PRECISION (fmode) > GET_MODE_PRECISION (tmode))
5888 return;
5889
5890 if ((GET_MODE_CLASS (tmode) == MODE_FLOAT
5891 && GET_MODE_CLASS (fmode) == MODE_FLOAT)
5892 || (DECIMAL_FLOAT_MODE_P (fmode) && DECIMAL_FLOAT_MODE_P (tmode)))
5893 gen_intraclass_conv_libfunc (tab, opname, tmode, fmode);
5894 }
5895
5896 /* Pick proper libcall for fract_optab. We need to chose if we do
5897 interclass or intraclass. */
5898
5899 static void
5900 gen_fract_conv_libfunc (convert_optab tab,
5901 const char *opname,
5902 enum machine_mode tmode,
5903 enum machine_mode fmode)
5904 {
5905 if (tmode == fmode)
5906 return;
5907 if (!(ALL_FIXED_POINT_MODE_P (tmode) || ALL_FIXED_POINT_MODE_P (fmode)))
5908 return;
5909
5910 if (GET_MODE_CLASS (tmode) == GET_MODE_CLASS (fmode))
5911 gen_intraclass_conv_libfunc (tab, opname, tmode, fmode);
5912 else
5913 gen_interclass_conv_libfunc (tab, opname, tmode, fmode);
5914 }
5915
5916 /* Pick proper libcall for fractuns_optab. */
5917
5918 static void
5919 gen_fractuns_conv_libfunc (convert_optab tab,
5920 const char *opname,
5921 enum machine_mode tmode,
5922 enum machine_mode fmode)
5923 {
5924 if (tmode == fmode)
5925 return;
5926 /* One mode must be a fixed-point mode, and the other must be an integer
5927 mode. */
5928 if (!((ALL_FIXED_POINT_MODE_P (tmode) && GET_MODE_CLASS (fmode) == MODE_INT)
5929 || (ALL_FIXED_POINT_MODE_P (fmode)
5930 && GET_MODE_CLASS (tmode) == MODE_INT)))
5931 return;
5932
5933 gen_interclass_conv_libfunc (tab, opname, tmode, fmode);
5934 }
5935
5936 /* Pick proper libcall for satfract_optab. We need to chose if we do
5937 interclass or intraclass. */
5938
5939 static void
5940 gen_satfract_conv_libfunc (convert_optab tab,
5941 const char *opname,
5942 enum machine_mode tmode,
5943 enum machine_mode fmode)
5944 {
5945 if (tmode == fmode)
5946 return;
5947 /* TMODE must be a fixed-point mode. */
5948 if (!ALL_FIXED_POINT_MODE_P (tmode))
5949 return;
5950
5951 if (GET_MODE_CLASS (tmode) == GET_MODE_CLASS (fmode))
5952 gen_intraclass_conv_libfunc (tab, opname, tmode, fmode);
5953 else
5954 gen_interclass_conv_libfunc (tab, opname, tmode, fmode);
5955 }
5956
5957 /* Pick proper libcall for satfractuns_optab. */
5958
5959 static void
5960 gen_satfractuns_conv_libfunc (convert_optab tab,
5961 const char *opname,
5962 enum machine_mode tmode,
5963 enum machine_mode fmode)
5964 {
5965 if (tmode == fmode)
5966 return;
5967 /* TMODE must be a fixed-point mode, and FMODE must be an integer mode. */
5968 if (!(ALL_FIXED_POINT_MODE_P (tmode) && GET_MODE_CLASS (fmode) == MODE_INT))
5969 return;
5970
5971 gen_interclass_conv_libfunc (tab, opname, tmode, fmode);
5972 }
5973
5974 /* A table of previously-created libfuncs, hashed by name. */
5975 static GTY ((param_is (union tree_node))) htab_t libfunc_decls;
5976
5977 /* Hashtable callbacks for libfunc_decls. */
5978
5979 static hashval_t
5980 libfunc_decl_hash (const void *entry)
5981 {
5982 return IDENTIFIER_HASH_VALUE (DECL_NAME ((const_tree) entry));
5983 }
5984
5985 static int
5986 libfunc_decl_eq (const void *entry1, const void *entry2)
5987 {
5988 return DECL_NAME ((const_tree) entry1) == (const_tree) entry2;
5989 }
5990
5991 /* Build a decl for a libfunc named NAME. */
5992
5993 tree
5994 build_libfunc_function (const char *name)
5995 {
5996 tree decl = build_decl (UNKNOWN_LOCATION, FUNCTION_DECL,
5997 get_identifier (name),
5998 build_function_type (integer_type_node, NULL_TREE));
5999 /* ??? We don't have any type information except for this is
6000 a function. Pretend this is "int foo()". */
6001 DECL_ARTIFICIAL (decl) = 1;
6002 DECL_EXTERNAL (decl) = 1;
6003 TREE_PUBLIC (decl) = 1;
6004 gcc_assert (DECL_ASSEMBLER_NAME (decl));
6005
6006 /* Zap the nonsensical SYMBOL_REF_DECL for this. What we're left with
6007 are the flags assigned by targetm.encode_section_info. */
6008 SET_SYMBOL_REF_DECL (XEXP (DECL_RTL (decl), 0), NULL);
6009
6010 return decl;
6011 }
6012
6013 rtx
6014 init_one_libfunc (const char *name)
6015 {
6016 tree id, decl;
6017 void **slot;
6018 hashval_t hash;
6019
6020 if (libfunc_decls == NULL)
6021 libfunc_decls = htab_create_ggc (37, libfunc_decl_hash,
6022 libfunc_decl_eq, NULL);
6023
6024 /* See if we have already created a libfunc decl for this function. */
6025 id = get_identifier (name);
6026 hash = IDENTIFIER_HASH_VALUE (id);
6027 slot = htab_find_slot_with_hash (libfunc_decls, id, hash, INSERT);
6028 decl = (tree) *slot;
6029 if (decl == NULL)
6030 {
6031 /* Create a new decl, so that it can be passed to
6032 targetm.encode_section_info. */
6033 decl = build_libfunc_function (name);
6034 *slot = decl;
6035 }
6036 return XEXP (DECL_RTL (decl), 0);
6037 }
6038
6039 /* Adjust the assembler name of libfunc NAME to ASMSPEC. */
6040
6041 rtx
6042 set_user_assembler_libfunc (const char *name, const char *asmspec)
6043 {
6044 tree id, decl;
6045 void **slot;
6046 hashval_t hash;
6047
6048 id = get_identifier (name);
6049 hash = IDENTIFIER_HASH_VALUE (id);
6050 slot = htab_find_slot_with_hash (libfunc_decls, id, hash, NO_INSERT);
6051 gcc_assert (slot);
6052 decl = (tree) *slot;
6053 set_user_assembler_name (decl, asmspec);
6054 return XEXP (DECL_RTL (decl), 0);
6055 }
6056
6057 /* Call this to reset the function entry for one optab (OPTABLE) in mode
6058 MODE to NAME, which should be either 0 or a string constant. */
6059 void
6060 set_optab_libfunc (optab optable, enum machine_mode mode, const char *name)
6061 {
6062 rtx val;
6063 struct libfunc_entry e;
6064 struct libfunc_entry **slot;
6065 e.optab = (size_t) (optable - &optab_table[0]);
6066 e.mode1 = mode;
6067 e.mode2 = VOIDmode;
6068
6069 if (name)
6070 val = init_one_libfunc (name);
6071 else
6072 val = 0;
6073 slot = (struct libfunc_entry **) htab_find_slot (libfunc_hash, &e, INSERT);
6074 if (*slot == NULL)
6075 *slot = ggc_alloc_libfunc_entry ();
6076 (*slot)->optab = (size_t) (optable - &optab_table[0]);
6077 (*slot)->mode1 = mode;
6078 (*slot)->mode2 = VOIDmode;
6079 (*slot)->libfunc = val;
6080 }
6081
6082 /* Call this to reset the function entry for one conversion optab
6083 (OPTABLE) from mode FMODE to mode TMODE to NAME, which should be
6084 either 0 or a string constant. */
6085 void
6086 set_conv_libfunc (convert_optab optable, enum machine_mode tmode,
6087 enum machine_mode fmode, const char *name)
6088 {
6089 rtx val;
6090 struct libfunc_entry e;
6091 struct libfunc_entry **slot;
6092 e.optab = (size_t) (optable - &convert_optab_table[0]);
6093 e.mode1 = tmode;
6094 e.mode2 = fmode;
6095
6096 if (name)
6097 val = init_one_libfunc (name);
6098 else
6099 val = 0;
6100 slot = (struct libfunc_entry **) htab_find_slot (libfunc_hash, &e, INSERT);
6101 if (*slot == NULL)
6102 *slot = ggc_alloc_libfunc_entry ();
6103 (*slot)->optab = (size_t) (optable - &convert_optab_table[0]);
6104 (*slot)->mode1 = tmode;
6105 (*slot)->mode2 = fmode;
6106 (*slot)->libfunc = val;
6107 }
6108
6109 /* Call this to initialize the contents of the optabs
6110 appropriately for the current target machine. */
6111
6112 void
6113 init_optabs (void)
6114 {
6115 if (libfunc_hash)
6116 {
6117 htab_empty (libfunc_hash);
6118 /* We statically initialize the insn_codes with the equivalent of
6119 CODE_FOR_nothing. Repeat the process if reinitialising. */
6120 init_insn_codes ();
6121 }
6122 else
6123 libfunc_hash = htab_create_ggc (10, hash_libfunc, eq_libfunc, NULL);
6124
6125 init_optab (add_optab, PLUS);
6126 init_optabv (addv_optab, PLUS);
6127 init_optab (sub_optab, MINUS);
6128 init_optabv (subv_optab, MINUS);
6129 init_optab (ssadd_optab, SS_PLUS);
6130 init_optab (usadd_optab, US_PLUS);
6131 init_optab (sssub_optab, SS_MINUS);
6132 init_optab (ussub_optab, US_MINUS);
6133 init_optab (smul_optab, MULT);
6134 init_optab (ssmul_optab, SS_MULT);
6135 init_optab (usmul_optab, US_MULT);
6136 init_optabv (smulv_optab, MULT);
6137 init_optab (smul_highpart_optab, UNKNOWN);
6138 init_optab (umul_highpart_optab, UNKNOWN);
6139 init_optab (smul_widen_optab, UNKNOWN);
6140 init_optab (umul_widen_optab, UNKNOWN);
6141 init_optab (usmul_widen_optab, UNKNOWN);
6142 init_optab (smadd_widen_optab, UNKNOWN);
6143 init_optab (umadd_widen_optab, UNKNOWN);
6144 init_optab (ssmadd_widen_optab, UNKNOWN);
6145 init_optab (usmadd_widen_optab, UNKNOWN);
6146 init_optab (smsub_widen_optab, UNKNOWN);
6147 init_optab (umsub_widen_optab, UNKNOWN);
6148 init_optab (ssmsub_widen_optab, UNKNOWN);
6149 init_optab (usmsub_widen_optab, UNKNOWN);
6150 init_optab (sdiv_optab, DIV);
6151 init_optab (ssdiv_optab, SS_DIV);
6152 init_optab (usdiv_optab, US_DIV);
6153 init_optabv (sdivv_optab, DIV);
6154 init_optab (sdivmod_optab, UNKNOWN);
6155 init_optab (udiv_optab, UDIV);
6156 init_optab (udivmod_optab, UNKNOWN);
6157 init_optab (smod_optab, MOD);
6158 init_optab (umod_optab, UMOD);
6159 init_optab (fmod_optab, UNKNOWN);
6160 init_optab (remainder_optab, UNKNOWN);
6161 init_optab (ftrunc_optab, UNKNOWN);
6162 init_optab (and_optab, AND);
6163 init_optab (ior_optab, IOR);
6164 init_optab (xor_optab, XOR);
6165 init_optab (ashl_optab, ASHIFT);
6166 init_optab (ssashl_optab, SS_ASHIFT);
6167 init_optab (usashl_optab, US_ASHIFT);
6168 init_optab (ashr_optab, ASHIFTRT);
6169 init_optab (lshr_optab, LSHIFTRT);
6170 init_optabv (vashl_optab, ASHIFT);
6171 init_optabv (vashr_optab, ASHIFTRT);
6172 init_optabv (vlshr_optab, LSHIFTRT);
6173 init_optab (rotl_optab, ROTATE);
6174 init_optab (rotr_optab, ROTATERT);
6175 init_optab (smin_optab, SMIN);
6176 init_optab (smax_optab, SMAX);
6177 init_optab (umin_optab, UMIN);
6178 init_optab (umax_optab, UMAX);
6179 init_optab (pow_optab, UNKNOWN);
6180 init_optab (atan2_optab, UNKNOWN);
6181 init_optab (fma_optab, FMA);
6182 init_optab (fms_optab, UNKNOWN);
6183 init_optab (fnma_optab, UNKNOWN);
6184 init_optab (fnms_optab, UNKNOWN);
6185
6186 /* These three have codes assigned exclusively for the sake of
6187 have_insn_for. */
6188 init_optab (mov_optab, SET);
6189 init_optab (movstrict_optab, STRICT_LOW_PART);
6190 init_optab (cbranch_optab, COMPARE);
6191
6192 init_optab (cmov_optab, UNKNOWN);
6193 init_optab (cstore_optab, UNKNOWN);
6194 init_optab (ctrap_optab, UNKNOWN);
6195
6196 init_optab (storent_optab, UNKNOWN);
6197
6198 init_optab (cmp_optab, UNKNOWN);
6199 init_optab (ucmp_optab, UNKNOWN);
6200
6201 init_optab (eq_optab, EQ);
6202 init_optab (ne_optab, NE);
6203 init_optab (gt_optab, GT);
6204 init_optab (ge_optab, GE);
6205 init_optab (lt_optab, LT);
6206 init_optab (le_optab, LE);
6207 init_optab (unord_optab, UNORDERED);
6208
6209 init_optab (neg_optab, NEG);
6210 init_optab (ssneg_optab, SS_NEG);
6211 init_optab (usneg_optab, US_NEG);
6212 init_optabv (negv_optab, NEG);
6213 init_optab (abs_optab, ABS);
6214 init_optabv (absv_optab, ABS);
6215 init_optab (addcc_optab, UNKNOWN);
6216 init_optab (one_cmpl_optab, NOT);
6217 init_optab (bswap_optab, BSWAP);
6218 init_optab (ffs_optab, FFS);
6219 init_optab (clz_optab, CLZ);
6220 init_optab (ctz_optab, CTZ);
6221 init_optab (clrsb_optab, CLRSB);
6222 init_optab (popcount_optab, POPCOUNT);
6223 init_optab (parity_optab, PARITY);
6224 init_optab (sqrt_optab, SQRT);
6225 init_optab (floor_optab, UNKNOWN);
6226 init_optab (ceil_optab, UNKNOWN);
6227 init_optab (round_optab, UNKNOWN);
6228 init_optab (btrunc_optab, UNKNOWN);
6229 init_optab (nearbyint_optab, UNKNOWN);
6230 init_optab (rint_optab, UNKNOWN);
6231 init_optab (sincos_optab, UNKNOWN);
6232 init_optab (sin_optab, UNKNOWN);
6233 init_optab (asin_optab, UNKNOWN);
6234 init_optab (cos_optab, UNKNOWN);
6235 init_optab (acos_optab, UNKNOWN);
6236 init_optab (exp_optab, UNKNOWN);
6237 init_optab (exp10_optab, UNKNOWN);
6238 init_optab (exp2_optab, UNKNOWN);
6239 init_optab (expm1_optab, UNKNOWN);
6240 init_optab (ldexp_optab, UNKNOWN);
6241 init_optab (scalb_optab, UNKNOWN);
6242 init_optab (significand_optab, UNKNOWN);
6243 init_optab (logb_optab, UNKNOWN);
6244 init_optab (ilogb_optab, UNKNOWN);
6245 init_optab (log_optab, UNKNOWN);
6246 init_optab (log10_optab, UNKNOWN);
6247 init_optab (log2_optab, UNKNOWN);
6248 init_optab (log1p_optab, UNKNOWN);
6249 init_optab (tan_optab, UNKNOWN);
6250 init_optab (atan_optab, UNKNOWN);
6251 init_optab (copysign_optab, UNKNOWN);
6252 init_optab (signbit_optab, UNKNOWN);
6253
6254 init_optab (isinf_optab, UNKNOWN);
6255
6256 init_optab (strlen_optab, UNKNOWN);
6257 init_optab (push_optab, UNKNOWN);
6258
6259 init_optab (reduc_smax_optab, UNKNOWN);
6260 init_optab (reduc_umax_optab, UNKNOWN);
6261 init_optab (reduc_smin_optab, UNKNOWN);
6262 init_optab (reduc_umin_optab, UNKNOWN);
6263 init_optab (reduc_splus_optab, UNKNOWN);
6264 init_optab (reduc_uplus_optab, UNKNOWN);
6265
6266 init_optab (ssum_widen_optab, UNKNOWN);
6267 init_optab (usum_widen_optab, UNKNOWN);
6268 init_optab (sdot_prod_optab, UNKNOWN);
6269 init_optab (udot_prod_optab, UNKNOWN);
6270
6271 init_optab (vec_extract_optab, UNKNOWN);
6272 init_optab (vec_extract_even_optab, UNKNOWN);
6273 init_optab (vec_extract_odd_optab, UNKNOWN);
6274 init_optab (vec_interleave_high_optab, UNKNOWN);
6275 init_optab (vec_interleave_low_optab, UNKNOWN);
6276 init_optab (vec_set_optab, UNKNOWN);
6277 init_optab (vec_init_optab, UNKNOWN);
6278 init_optab (vec_shl_optab, UNKNOWN);
6279 init_optab (vec_shr_optab, UNKNOWN);
6280 init_optab (vec_realign_load_optab, UNKNOWN);
6281 init_optab (movmisalign_optab, UNKNOWN);
6282 init_optab (vec_widen_umult_hi_optab, UNKNOWN);
6283 init_optab (vec_widen_umult_lo_optab, UNKNOWN);
6284 init_optab (vec_widen_smult_hi_optab, UNKNOWN);
6285 init_optab (vec_widen_smult_lo_optab, UNKNOWN);
6286 init_optab (vec_widen_ushiftl_hi_optab, UNKNOWN);
6287 init_optab (vec_widen_ushiftl_lo_optab, UNKNOWN);
6288 init_optab (vec_widen_sshiftl_hi_optab, UNKNOWN);
6289 init_optab (vec_widen_sshiftl_lo_optab, UNKNOWN);
6290 init_optab (vec_unpacks_hi_optab, UNKNOWN);
6291 init_optab (vec_unpacks_lo_optab, UNKNOWN);
6292 init_optab (vec_unpacku_hi_optab, UNKNOWN);
6293 init_optab (vec_unpacku_lo_optab, UNKNOWN);
6294 init_optab (vec_unpacks_float_hi_optab, UNKNOWN);
6295 init_optab (vec_unpacks_float_lo_optab, UNKNOWN);
6296 init_optab (vec_unpacku_float_hi_optab, UNKNOWN);
6297 init_optab (vec_unpacku_float_lo_optab, UNKNOWN);
6298 init_optab (vec_pack_trunc_optab, UNKNOWN);
6299 init_optab (vec_pack_usat_optab, UNKNOWN);
6300 init_optab (vec_pack_ssat_optab, UNKNOWN);
6301 init_optab (vec_pack_ufix_trunc_optab, UNKNOWN);
6302 init_optab (vec_pack_sfix_trunc_optab, UNKNOWN);
6303
6304 init_optab (powi_optab, UNKNOWN);
6305
6306 /* Conversions. */
6307 init_convert_optab (sext_optab, SIGN_EXTEND);
6308 init_convert_optab (zext_optab, ZERO_EXTEND);
6309 init_convert_optab (trunc_optab, TRUNCATE);
6310 init_convert_optab (sfix_optab, FIX);
6311 init_convert_optab (ufix_optab, UNSIGNED_FIX);
6312 init_convert_optab (sfixtrunc_optab, UNKNOWN);
6313 init_convert_optab (ufixtrunc_optab, UNKNOWN);
6314 init_convert_optab (sfloat_optab, FLOAT);
6315 init_convert_optab (ufloat_optab, UNSIGNED_FLOAT);
6316 init_convert_optab (lrint_optab, UNKNOWN);
6317 init_convert_optab (lround_optab, UNKNOWN);
6318 init_convert_optab (lfloor_optab, UNKNOWN);
6319 init_convert_optab (lceil_optab, UNKNOWN);
6320
6321 init_convert_optab (fract_optab, FRACT_CONVERT);
6322 init_convert_optab (fractuns_optab, UNSIGNED_FRACT_CONVERT);
6323 init_convert_optab (satfract_optab, SAT_FRACT);
6324 init_convert_optab (satfractuns_optab, UNSIGNED_SAT_FRACT);
6325
6326 /* Fill in the optabs with the insns we support. */
6327 init_all_optabs ();
6328
6329 /* Initialize the optabs with the names of the library functions. */
6330 add_optab->libcall_basename = "add";
6331 add_optab->libcall_suffix = '3';
6332 add_optab->libcall_gen = gen_int_fp_fixed_libfunc;
6333 addv_optab->libcall_basename = "add";
6334 addv_optab->libcall_suffix = '3';
6335 addv_optab->libcall_gen = gen_intv_fp_libfunc;
6336 ssadd_optab->libcall_basename = "ssadd";
6337 ssadd_optab->libcall_suffix = '3';
6338 ssadd_optab->libcall_gen = gen_signed_fixed_libfunc;
6339 usadd_optab->libcall_basename = "usadd";
6340 usadd_optab->libcall_suffix = '3';
6341 usadd_optab->libcall_gen = gen_unsigned_fixed_libfunc;
6342 sub_optab->libcall_basename = "sub";
6343 sub_optab->libcall_suffix = '3';
6344 sub_optab->libcall_gen = gen_int_fp_fixed_libfunc;
6345 subv_optab->libcall_basename = "sub";
6346 subv_optab->libcall_suffix = '3';
6347 subv_optab->libcall_gen = gen_intv_fp_libfunc;
6348 sssub_optab->libcall_basename = "sssub";
6349 sssub_optab->libcall_suffix = '3';
6350 sssub_optab->libcall_gen = gen_signed_fixed_libfunc;
6351 ussub_optab->libcall_basename = "ussub";
6352 ussub_optab->libcall_suffix = '3';
6353 ussub_optab->libcall_gen = gen_unsigned_fixed_libfunc;
6354 smul_optab->libcall_basename = "mul";
6355 smul_optab->libcall_suffix = '3';
6356 smul_optab->libcall_gen = gen_int_fp_fixed_libfunc;
6357 smulv_optab->libcall_basename = "mul";
6358 smulv_optab->libcall_suffix = '3';
6359 smulv_optab->libcall_gen = gen_intv_fp_libfunc;
6360 ssmul_optab->libcall_basename = "ssmul";
6361 ssmul_optab->libcall_suffix = '3';
6362 ssmul_optab->libcall_gen = gen_signed_fixed_libfunc;
6363 usmul_optab->libcall_basename = "usmul";
6364 usmul_optab->libcall_suffix = '3';
6365 usmul_optab->libcall_gen = gen_unsigned_fixed_libfunc;
6366 sdiv_optab->libcall_basename = "div";
6367 sdiv_optab->libcall_suffix = '3';
6368 sdiv_optab->libcall_gen = gen_int_fp_signed_fixed_libfunc;
6369 sdivv_optab->libcall_basename = "divv";
6370 sdivv_optab->libcall_suffix = '3';
6371 sdivv_optab->libcall_gen = gen_int_libfunc;
6372 ssdiv_optab->libcall_basename = "ssdiv";
6373 ssdiv_optab->libcall_suffix = '3';
6374 ssdiv_optab->libcall_gen = gen_signed_fixed_libfunc;
6375 udiv_optab->libcall_basename = "udiv";
6376 udiv_optab->libcall_suffix = '3';
6377 udiv_optab->libcall_gen = gen_int_unsigned_fixed_libfunc;
6378 usdiv_optab->libcall_basename = "usdiv";
6379 usdiv_optab->libcall_suffix = '3';
6380 usdiv_optab->libcall_gen = gen_unsigned_fixed_libfunc;
6381 sdivmod_optab->libcall_basename = "divmod";
6382 sdivmod_optab->libcall_suffix = '4';
6383 sdivmod_optab->libcall_gen = gen_int_libfunc;
6384 udivmod_optab->libcall_basename = "udivmod";
6385 udivmod_optab->libcall_suffix = '4';
6386 udivmod_optab->libcall_gen = gen_int_libfunc;
6387 smod_optab->libcall_basename = "mod";
6388 smod_optab->libcall_suffix = '3';
6389 smod_optab->libcall_gen = gen_int_libfunc;
6390 umod_optab->libcall_basename = "umod";
6391 umod_optab->libcall_suffix = '3';
6392 umod_optab->libcall_gen = gen_int_libfunc;
6393 ftrunc_optab->libcall_basename = "ftrunc";
6394 ftrunc_optab->libcall_suffix = '2';
6395 ftrunc_optab->libcall_gen = gen_fp_libfunc;
6396 and_optab->libcall_basename = "and";
6397 and_optab->libcall_suffix = '3';
6398 and_optab->libcall_gen = gen_int_libfunc;
6399 ior_optab->libcall_basename = "ior";
6400 ior_optab->libcall_suffix = '3';
6401 ior_optab->libcall_gen = gen_int_libfunc;
6402 xor_optab->libcall_basename = "xor";
6403 xor_optab->libcall_suffix = '3';
6404 xor_optab->libcall_gen = gen_int_libfunc;
6405 ashl_optab->libcall_basename = "ashl";
6406 ashl_optab->libcall_suffix = '3';
6407 ashl_optab->libcall_gen = gen_int_fixed_libfunc;
6408 ssashl_optab->libcall_basename = "ssashl";
6409 ssashl_optab->libcall_suffix = '3';
6410 ssashl_optab->libcall_gen = gen_signed_fixed_libfunc;
6411 usashl_optab->libcall_basename = "usashl";
6412 usashl_optab->libcall_suffix = '3';
6413 usashl_optab->libcall_gen = gen_unsigned_fixed_libfunc;
6414 ashr_optab->libcall_basename = "ashr";
6415 ashr_optab->libcall_suffix = '3';
6416 ashr_optab->libcall_gen = gen_int_signed_fixed_libfunc;
6417 lshr_optab->libcall_basename = "lshr";
6418 lshr_optab->libcall_suffix = '3';
6419 lshr_optab->libcall_gen = gen_int_unsigned_fixed_libfunc;
6420 smin_optab->libcall_basename = "min";
6421 smin_optab->libcall_suffix = '3';
6422 smin_optab->libcall_gen = gen_int_fp_libfunc;
6423 smax_optab->libcall_basename = "max";
6424 smax_optab->libcall_suffix = '3';
6425 smax_optab->libcall_gen = gen_int_fp_libfunc;
6426 umin_optab->libcall_basename = "umin";
6427 umin_optab->libcall_suffix = '3';
6428 umin_optab->libcall_gen = gen_int_libfunc;
6429 umax_optab->libcall_basename = "umax";
6430 umax_optab->libcall_suffix = '3';
6431 umax_optab->libcall_gen = gen_int_libfunc;
6432 neg_optab->libcall_basename = "neg";
6433 neg_optab->libcall_suffix = '2';
6434 neg_optab->libcall_gen = gen_int_fp_fixed_libfunc;
6435 ssneg_optab->libcall_basename = "ssneg";
6436 ssneg_optab->libcall_suffix = '2';
6437 ssneg_optab->libcall_gen = gen_signed_fixed_libfunc;
6438 usneg_optab->libcall_basename = "usneg";
6439 usneg_optab->libcall_suffix = '2';
6440 usneg_optab->libcall_gen = gen_unsigned_fixed_libfunc;
6441 negv_optab->libcall_basename = "neg";
6442 negv_optab->libcall_suffix = '2';
6443 negv_optab->libcall_gen = gen_intv_fp_libfunc;
6444 one_cmpl_optab->libcall_basename = "one_cmpl";
6445 one_cmpl_optab->libcall_suffix = '2';
6446 one_cmpl_optab->libcall_gen = gen_int_libfunc;
6447 ffs_optab->libcall_basename = "ffs";
6448 ffs_optab->libcall_suffix = '2';
6449 ffs_optab->libcall_gen = gen_int_libfunc;
6450 clz_optab->libcall_basename = "clz";
6451 clz_optab->libcall_suffix = '2';
6452 clz_optab->libcall_gen = gen_int_libfunc;
6453 ctz_optab->libcall_basename = "ctz";
6454 ctz_optab->libcall_suffix = '2';
6455 ctz_optab->libcall_gen = gen_int_libfunc;
6456 clrsb_optab->libcall_basename = "clrsb";
6457 clrsb_optab->libcall_suffix = '2';
6458 clrsb_optab->libcall_gen = gen_int_libfunc;
6459 popcount_optab->libcall_basename = "popcount";
6460 popcount_optab->libcall_suffix = '2';
6461 popcount_optab->libcall_gen = gen_int_libfunc;
6462 parity_optab->libcall_basename = "parity";
6463 parity_optab->libcall_suffix = '2';
6464 parity_optab->libcall_gen = gen_int_libfunc;
6465
6466 /* Comparison libcalls for integers MUST come in pairs,
6467 signed/unsigned. */
6468 cmp_optab->libcall_basename = "cmp";
6469 cmp_optab->libcall_suffix = '2';
6470 cmp_optab->libcall_gen = gen_int_fp_fixed_libfunc;
6471 ucmp_optab->libcall_basename = "ucmp";
6472 ucmp_optab->libcall_suffix = '2';
6473 ucmp_optab->libcall_gen = gen_int_libfunc;
6474
6475 /* EQ etc are floating point only. */
6476 eq_optab->libcall_basename = "eq";
6477 eq_optab->libcall_suffix = '2';
6478 eq_optab->libcall_gen = gen_fp_libfunc;
6479 ne_optab->libcall_basename = "ne";
6480 ne_optab->libcall_suffix = '2';
6481 ne_optab->libcall_gen = gen_fp_libfunc;
6482 gt_optab->libcall_basename = "gt";
6483 gt_optab->libcall_suffix = '2';
6484 gt_optab->libcall_gen = gen_fp_libfunc;
6485 ge_optab->libcall_basename = "ge";
6486 ge_optab->libcall_suffix = '2';
6487 ge_optab->libcall_gen = gen_fp_libfunc;
6488 lt_optab->libcall_basename = "lt";
6489 lt_optab->libcall_suffix = '2';
6490 lt_optab->libcall_gen = gen_fp_libfunc;
6491 le_optab->libcall_basename = "le";
6492 le_optab->libcall_suffix = '2';
6493 le_optab->libcall_gen = gen_fp_libfunc;
6494 unord_optab->libcall_basename = "unord";
6495 unord_optab->libcall_suffix = '2';
6496 unord_optab->libcall_gen = gen_fp_libfunc;
6497
6498 powi_optab->libcall_basename = "powi";
6499 powi_optab->libcall_suffix = '2';
6500 powi_optab->libcall_gen = gen_fp_libfunc;
6501
6502 /* Conversions. */
6503 sfloat_optab->libcall_basename = "float";
6504 sfloat_optab->libcall_gen = gen_int_to_fp_conv_libfunc;
6505 ufloat_optab->libcall_gen = gen_ufloat_conv_libfunc;
6506 sfix_optab->libcall_basename = "fix";
6507 sfix_optab->libcall_gen = gen_fp_to_int_conv_libfunc;
6508 ufix_optab->libcall_basename = "fixuns";
6509 ufix_optab->libcall_gen = gen_fp_to_int_conv_libfunc;
6510 lrint_optab->libcall_basename = "lrint";
6511 lrint_optab->libcall_gen = gen_int_to_fp_nondecimal_conv_libfunc;
6512 lround_optab->libcall_basename = "lround";
6513 lround_optab->libcall_gen = gen_int_to_fp_nondecimal_conv_libfunc;
6514 lfloor_optab->libcall_basename = "lfloor";
6515 lfloor_optab->libcall_gen = gen_int_to_fp_nondecimal_conv_libfunc;
6516 lceil_optab->libcall_basename = "lceil";
6517 lceil_optab->libcall_gen = gen_int_to_fp_nondecimal_conv_libfunc;
6518
6519 /* trunc_optab is also used for FLOAT_EXTEND. */
6520 sext_optab->libcall_basename = "extend";
6521 sext_optab->libcall_gen = gen_extend_conv_libfunc;
6522 trunc_optab->libcall_basename = "trunc";
6523 trunc_optab->libcall_gen = gen_trunc_conv_libfunc;
6524
6525 /* Conversions for fixed-point modes and other modes. */
6526 fract_optab->libcall_basename = "fract";
6527 fract_optab->libcall_gen = gen_fract_conv_libfunc;
6528 satfract_optab->libcall_basename = "satfract";
6529 satfract_optab->libcall_gen = gen_satfract_conv_libfunc;
6530 fractuns_optab->libcall_basename = "fractuns";
6531 fractuns_optab->libcall_gen = gen_fractuns_conv_libfunc;
6532 satfractuns_optab->libcall_basename = "satfractuns";
6533 satfractuns_optab->libcall_gen = gen_satfractuns_conv_libfunc;
6534
6535 /* The ffs function operates on `int'. Fall back on it if we do not
6536 have a libgcc2 function for that width. */
6537 if (INT_TYPE_SIZE < BITS_PER_WORD)
6538 set_optab_libfunc (ffs_optab, mode_for_size (INT_TYPE_SIZE, MODE_INT, 0),
6539 "ffs");
6540
6541 /* Explicitly initialize the bswap libfuncs since we need them to be
6542 valid for things other than word_mode. */
6543 if (targetm.libfunc_gnu_prefix)
6544 {
6545 set_optab_libfunc (bswap_optab, SImode, "__gnu_bswapsi2");
6546 set_optab_libfunc (bswap_optab, DImode, "__gnu_bswapdi2");
6547 }
6548 else
6549 {
6550 set_optab_libfunc (bswap_optab, SImode, "__bswapsi2");
6551 set_optab_libfunc (bswap_optab, DImode, "__bswapdi2");
6552 }
6553
6554 /* Use cabs for double complex abs, since systems generally have cabs.
6555 Don't define any libcall for float complex, so that cabs will be used. */
6556 if (complex_double_type_node)
6557 set_optab_libfunc (abs_optab, TYPE_MODE (complex_double_type_node), "cabs");
6558
6559 abort_libfunc = init_one_libfunc ("abort");
6560 memcpy_libfunc = init_one_libfunc ("memcpy");
6561 memmove_libfunc = init_one_libfunc ("memmove");
6562 memcmp_libfunc = init_one_libfunc ("memcmp");
6563 memset_libfunc = init_one_libfunc ("memset");
6564 setbits_libfunc = init_one_libfunc ("__setbits");
6565
6566 #ifndef DONT_USE_BUILTIN_SETJMP
6567 setjmp_libfunc = init_one_libfunc ("__builtin_setjmp");
6568 longjmp_libfunc = init_one_libfunc ("__builtin_longjmp");
6569 #else
6570 setjmp_libfunc = init_one_libfunc ("setjmp");
6571 longjmp_libfunc = init_one_libfunc ("longjmp");
6572 #endif
6573 unwind_sjlj_register_libfunc = init_one_libfunc ("_Unwind_SjLj_Register");
6574 unwind_sjlj_unregister_libfunc
6575 = init_one_libfunc ("_Unwind_SjLj_Unregister");
6576
6577 /* For function entry/exit instrumentation. */
6578 profile_function_entry_libfunc
6579 = init_one_libfunc ("__cyg_profile_func_enter");
6580 profile_function_exit_libfunc
6581 = init_one_libfunc ("__cyg_profile_func_exit");
6582
6583 gcov_flush_libfunc = init_one_libfunc ("__gcov_flush");
6584
6585 /* Allow the target to add more libcalls or rename some, etc. */
6586 targetm.init_libfuncs ();
6587 }
6588
6589 /* A helper function for init_sync_libfuncs. Using the basename BASE,
6590 install libfuncs into TAB for BASE_N for 1 <= N <= MAX. */
6591
6592 static void
6593 init_sync_libfuncs_1 (optab tab, const char *base, int max)
6594 {
6595 enum machine_mode mode;
6596 char buf[64];
6597 size_t len = strlen (base);
6598 int i;
6599
6600 gcc_assert (max <= 8);
6601 gcc_assert (len + 3 < sizeof (buf));
6602
6603 memcpy (buf, base, len);
6604 buf[len] = '_';
6605 buf[len + 1] = '0';
6606 buf[len + 2] = '\0';
6607
6608 mode = QImode;
6609 for (i = 1; i <= max; i *= 2)
6610 {
6611 buf[len + 1] = '0' + i;
6612 set_optab_libfunc (tab, mode, buf);
6613 mode = GET_MODE_2XWIDER_MODE (mode);
6614 }
6615 }
6616
6617 void
6618 init_sync_libfuncs (int max)
6619 {
6620 init_sync_libfuncs_1 (sync_compare_and_swap_optab,
6621 "__sync_val_compare_and_swap", max);
6622 init_sync_libfuncs_1 (sync_lock_test_and_set_optab,
6623 "__sync_lock_test_and_set", max);
6624
6625 init_sync_libfuncs_1 (sync_old_add_optab, "__sync_fetch_and_add", max);
6626 init_sync_libfuncs_1 (sync_old_sub_optab, "__sync_fetch_and_sub", max);
6627 init_sync_libfuncs_1 (sync_old_ior_optab, "__sync_fetch_and_or", max);
6628 init_sync_libfuncs_1 (sync_old_and_optab, "__sync_fetch_and_and", max);
6629 init_sync_libfuncs_1 (sync_old_xor_optab, "__sync_fetch_and_xor", max);
6630 init_sync_libfuncs_1 (sync_old_nand_optab, "__sync_fetch_and_nand", max);
6631
6632 init_sync_libfuncs_1 (sync_new_add_optab, "__sync_add_and_fetch", max);
6633 init_sync_libfuncs_1 (sync_new_sub_optab, "__sync_sub_and_fetch", max);
6634 init_sync_libfuncs_1 (sync_new_ior_optab, "__sync_or_and_fetch", max);
6635 init_sync_libfuncs_1 (sync_new_and_optab, "__sync_and_and_fetch", max);
6636 init_sync_libfuncs_1 (sync_new_xor_optab, "__sync_xor_and_fetch", max);
6637 init_sync_libfuncs_1 (sync_new_nand_optab, "__sync_nand_and_fetch", max);
6638 }
6639
6640 /* Print information about the current contents of the optabs on
6641 STDERR. */
6642
6643 DEBUG_FUNCTION void
6644 debug_optab_libfuncs (void)
6645 {
6646 int i;
6647 int j;
6648 int k;
6649
6650 /* Dump the arithmetic optabs. */
6651 for (i = 0; i != (int) OTI_MAX; i++)
6652 for (j = 0; j < NUM_MACHINE_MODES; ++j)
6653 {
6654 optab o;
6655 rtx l;
6656
6657 o = &optab_table[i];
6658 l = optab_libfunc (o, (enum machine_mode) j);
6659 if (l)
6660 {
6661 gcc_assert (GET_CODE (l) == SYMBOL_REF);
6662 fprintf (stderr, "%s\t%s:\t%s\n",
6663 GET_RTX_NAME (o->code),
6664 GET_MODE_NAME (j),
6665 XSTR (l, 0));
6666 }
6667 }
6668
6669 /* Dump the conversion optabs. */
6670 for (i = 0; i < (int) COI_MAX; ++i)
6671 for (j = 0; j < NUM_MACHINE_MODES; ++j)
6672 for (k = 0; k < NUM_MACHINE_MODES; ++k)
6673 {
6674 convert_optab o;
6675 rtx l;
6676
6677 o = &convert_optab_table[i];
6678 l = convert_optab_libfunc (o, (enum machine_mode) j,
6679 (enum machine_mode) k);
6680 if (l)
6681 {
6682 gcc_assert (GET_CODE (l) == SYMBOL_REF);
6683 fprintf (stderr, "%s\t%s\t%s:\t%s\n",
6684 GET_RTX_NAME (o->code),
6685 GET_MODE_NAME (j),
6686 GET_MODE_NAME (k),
6687 XSTR (l, 0));
6688 }
6689 }
6690 }
6691
6692 \f
6693 /* Generate insns to trap with code TCODE if OP1 and OP2 satisfy condition
6694 CODE. Return 0 on failure. */
6695
6696 rtx
6697 gen_cond_trap (enum rtx_code code, rtx op1, rtx op2, rtx tcode)
6698 {
6699 enum machine_mode mode = GET_MODE (op1);
6700 enum insn_code icode;
6701 rtx insn;
6702 rtx trap_rtx;
6703
6704 if (mode == VOIDmode)
6705 return 0;
6706
6707 icode = optab_handler (ctrap_optab, mode);
6708 if (icode == CODE_FOR_nothing)
6709 return 0;
6710
6711 /* Some targets only accept a zero trap code. */
6712 if (!insn_operand_matches (icode, 3, tcode))
6713 return 0;
6714
6715 do_pending_stack_adjust ();
6716 start_sequence ();
6717 prepare_cmp_insn (op1, op2, code, NULL_RTX, false, OPTAB_DIRECT,
6718 &trap_rtx, &mode);
6719 if (!trap_rtx)
6720 insn = NULL_RTX;
6721 else
6722 insn = GEN_FCN (icode) (trap_rtx, XEXP (trap_rtx, 0), XEXP (trap_rtx, 1),
6723 tcode);
6724
6725 /* If that failed, then give up. */
6726 if (insn == 0)
6727 {
6728 end_sequence ();
6729 return 0;
6730 }
6731
6732 emit_insn (insn);
6733 insn = get_insns ();
6734 end_sequence ();
6735 return insn;
6736 }
6737
6738 /* Return rtx code for TCODE. Use UNSIGNEDP to select signed
6739 or unsigned operation code. */
6740
6741 static enum rtx_code
6742 get_rtx_code (enum tree_code tcode, bool unsignedp)
6743 {
6744 enum rtx_code code;
6745 switch (tcode)
6746 {
6747 case EQ_EXPR:
6748 code = EQ;
6749 break;
6750 case NE_EXPR:
6751 code = NE;
6752 break;
6753 case LT_EXPR:
6754 code = unsignedp ? LTU : LT;
6755 break;
6756 case LE_EXPR:
6757 code = unsignedp ? LEU : LE;
6758 break;
6759 case GT_EXPR:
6760 code = unsignedp ? GTU : GT;
6761 break;
6762 case GE_EXPR:
6763 code = unsignedp ? GEU : GE;
6764 break;
6765
6766 case UNORDERED_EXPR:
6767 code = UNORDERED;
6768 break;
6769 case ORDERED_EXPR:
6770 code = ORDERED;
6771 break;
6772 case UNLT_EXPR:
6773 code = UNLT;
6774 break;
6775 case UNLE_EXPR:
6776 code = UNLE;
6777 break;
6778 case UNGT_EXPR:
6779 code = UNGT;
6780 break;
6781 case UNGE_EXPR:
6782 code = UNGE;
6783 break;
6784 case UNEQ_EXPR:
6785 code = UNEQ;
6786 break;
6787 case LTGT_EXPR:
6788 code = LTGT;
6789 break;
6790
6791 default:
6792 gcc_unreachable ();
6793 }
6794 return code;
6795 }
6796
6797 /* Return comparison rtx for COND. Use UNSIGNEDP to select signed or
6798 unsigned operators. Do not generate compare instruction. */
6799
6800 static rtx
6801 vector_compare_rtx (tree cond, bool unsignedp, enum insn_code icode)
6802 {
6803 struct expand_operand ops[2];
6804 enum rtx_code rcode;
6805 tree t_op0, t_op1;
6806 rtx rtx_op0, rtx_op1;
6807
6808 /* This is unlikely. While generating VEC_COND_EXPR, auto vectorizer
6809 ensures that condition is a relational operation. */
6810 gcc_assert (COMPARISON_CLASS_P (cond));
6811
6812 rcode = get_rtx_code (TREE_CODE (cond), unsignedp);
6813 t_op0 = TREE_OPERAND (cond, 0);
6814 t_op1 = TREE_OPERAND (cond, 1);
6815
6816 /* Expand operands. */
6817 rtx_op0 = expand_expr (t_op0, NULL_RTX, TYPE_MODE (TREE_TYPE (t_op0)),
6818 EXPAND_STACK_PARM);
6819 rtx_op1 = expand_expr (t_op1, NULL_RTX, TYPE_MODE (TREE_TYPE (t_op1)),
6820 EXPAND_STACK_PARM);
6821
6822 create_input_operand (&ops[0], rtx_op0, GET_MODE (rtx_op0));
6823 create_input_operand (&ops[1], rtx_op1, GET_MODE (rtx_op1));
6824 if (!maybe_legitimize_operands (icode, 4, 2, ops))
6825 gcc_unreachable ();
6826 return gen_rtx_fmt_ee (rcode, VOIDmode, ops[0].value, ops[1].value);
6827 }
6828
6829 /* Return true if VEC_PERM_EXPR can be expanded using SIMD extensions
6830 of the CPU. SEL may be NULL, which stands for an unknown constant. */
6831
6832 bool
6833 can_vec_perm_p (enum machine_mode mode, bool variable,
6834 const unsigned char *sel)
6835 {
6836 enum machine_mode qimode;
6837
6838 /* If the target doesn't implement a vector mode for the vector type,
6839 then no operations are supported. */
6840 if (!VECTOR_MODE_P (mode))
6841 return false;
6842
6843 if (!variable)
6844 {
6845 if (direct_optab_handler (vec_perm_const_optab, mode) != CODE_FOR_nothing
6846 && (sel == NULL
6847 || targetm.vectorize.vec_perm_const_ok == NULL
6848 || targetm.vectorize.vec_perm_const_ok (mode, sel)))
6849 return true;
6850 }
6851
6852 if (direct_optab_handler (vec_perm_optab, mode) != CODE_FOR_nothing)
6853 return true;
6854
6855 /* We allow fallback to a QI vector mode, and adjust the mask. */
6856 if (GET_MODE_INNER (mode) == QImode)
6857 return false;
6858 qimode = mode_for_vector (QImode, GET_MODE_SIZE (mode));
6859 if (!VECTOR_MODE_P (qimode))
6860 return false;
6861
6862 /* ??? For completeness, we ought to check the QImode version of
6863 vec_perm_const_optab. But all users of this implicit lowering
6864 feature implement the variable vec_perm_optab. */
6865 if (direct_optab_handler (vec_perm_optab, qimode) == CODE_FOR_nothing)
6866 return false;
6867
6868 /* In order to support the lowering of variable permutations,
6869 we need to support shifts and adds. */
6870 if (variable)
6871 {
6872 if (GET_MODE_UNIT_SIZE (mode) > 2
6873 && optab_handler (ashl_optab, mode) == CODE_FOR_nothing
6874 && optab_handler (vashl_optab, mode) == CODE_FOR_nothing)
6875 return false;
6876 if (optab_handler (add_optab, qimode) == CODE_FOR_nothing)
6877 return false;
6878 }
6879
6880 return true;
6881 }
6882
6883 /* Return true if we can implement VEC_INTERLEAVE_{HIGH,LOW}_EXPR or
6884 VEC_EXTRACT_{EVEN,ODD}_EXPR with VEC_PERM_EXPR for this target.
6885 If PSEL is non-null, return the selector for the permutation. */
6886
6887 bool
6888 can_vec_perm_for_code_p (enum tree_code code, enum machine_mode mode,
6889 rtx *psel)
6890 {
6891 bool need_sel_test = false;
6892 enum insn_code icode;
6893
6894 /* If the target doesn't implement a vector mode for the vector type,
6895 then no operations are supported. */
6896 if (!VECTOR_MODE_P (mode))
6897 return false;
6898
6899 /* Do as many tests as possible without reqiring the selector. */
6900 icode = direct_optab_handler (vec_perm_optab, mode);
6901 if (icode == CODE_FOR_nothing && GET_MODE_INNER (mode) != QImode)
6902 {
6903 enum machine_mode qimode
6904 = mode_for_vector (QImode, GET_MODE_SIZE (mode));
6905 if (VECTOR_MODE_P (qimode))
6906 icode = direct_optab_handler (vec_perm_optab, qimode);
6907 }
6908 if (icode == CODE_FOR_nothing)
6909 {
6910 icode = direct_optab_handler (vec_perm_const_optab, mode);
6911 if (icode != CODE_FOR_nothing
6912 && targetm.vectorize.vec_perm_const_ok != NULL)
6913 need_sel_test = true;
6914 }
6915 if (icode == CODE_FOR_nothing)
6916 return false;
6917
6918 /* If the selector is required, or if we need to test it, build it. */
6919 if (psel || need_sel_test)
6920 {
6921 int i, nelt = GET_MODE_NUNITS (mode), alt = 0;
6922 unsigned char *data = XALLOCAVEC (unsigned char, nelt);
6923
6924 switch (code)
6925 {
6926 case VEC_EXTRACT_ODD_EXPR:
6927 alt = 1;
6928 /* FALLTHRU */
6929 case VEC_EXTRACT_EVEN_EXPR:
6930 for (i = 0; i < nelt; ++i)
6931 data[i] = i * 2 + alt;
6932 break;
6933
6934 case VEC_INTERLEAVE_HIGH_EXPR:
6935 case VEC_INTERLEAVE_LOW_EXPR:
6936 if ((BYTES_BIG_ENDIAN != 0) ^ (code == VEC_INTERLEAVE_HIGH_EXPR))
6937 alt = nelt / 2;
6938 for (i = 0; i < nelt / 2; ++i)
6939 {
6940 data[i * 2] = i + alt;
6941 data[i * 2 + 1] = i + nelt + alt;
6942 }
6943 break;
6944
6945 default:
6946 gcc_unreachable ();
6947 }
6948
6949 if (need_sel_test
6950 && !targetm.vectorize.vec_perm_const_ok (mode, data))
6951 return false;
6952
6953 if (psel)
6954 {
6955 rtvec vec = rtvec_alloc (nelt);
6956 enum machine_mode imode = mode;
6957
6958 for (i = 0; i < nelt; ++i)
6959 RTVEC_ELT (vec, i) = GEN_INT (data[i]);
6960
6961 if (GET_MODE_CLASS (mode) != MODE_VECTOR_INT)
6962 {
6963 imode = int_mode_for_mode (GET_MODE_INNER (mode));
6964 imode = mode_for_vector (imode, nelt);
6965 gcc_assert (GET_MODE_CLASS (imode) == MODE_VECTOR_INT);
6966 }
6967
6968 *psel = gen_rtx_CONST_VECTOR (imode, vec);
6969 }
6970 }
6971
6972 return true;
6973 }
6974
6975 /* A subroutine of expand_vec_perm for expanding one vec_perm insn. */
6976
6977 static rtx
6978 expand_vec_perm_1 (enum insn_code icode, rtx target,
6979 rtx v0, rtx v1, rtx sel)
6980 {
6981 enum machine_mode tmode = GET_MODE (target);
6982 enum machine_mode smode = GET_MODE (sel);
6983 struct expand_operand ops[4];
6984
6985 create_output_operand (&ops[0], target, tmode);
6986 create_input_operand (&ops[3], sel, smode);
6987
6988 /* Make an effort to preserve v0 == v1. The target expander is able to
6989 rely on this to determine if we're permuting a single input operand. */
6990 if (rtx_equal_p (v0, v1))
6991 {
6992 if (!insn_operand_matches (icode, 1, v0))
6993 v0 = force_reg (tmode, v0);
6994 gcc_checking_assert (insn_operand_matches (icode, 1, v0));
6995 gcc_checking_assert (insn_operand_matches (icode, 2, v0));
6996
6997 create_fixed_operand (&ops[1], v0);
6998 create_fixed_operand (&ops[2], v0);
6999 }
7000 else
7001 {
7002 create_input_operand (&ops[1], v0, tmode);
7003 create_input_operand (&ops[2], v1, tmode);
7004 }
7005
7006 if (maybe_expand_insn (icode, 4, ops))
7007 return ops[0].value;
7008 return NULL_RTX;
7009 }
7010
7011 /* Generate instructions for vec_perm optab given its mode
7012 and three operands. */
7013
7014 rtx
7015 expand_vec_perm (enum machine_mode mode, rtx v0, rtx v1, rtx sel, rtx target)
7016 {
7017 enum insn_code icode;
7018 enum machine_mode qimode;
7019 unsigned int i, w, e, u;
7020 rtx tmp, sel_qi = NULL;
7021 rtvec vec;
7022
7023 if (!target || GET_MODE (target) != mode)
7024 target = gen_reg_rtx (mode);
7025
7026 w = GET_MODE_SIZE (mode);
7027 e = GET_MODE_NUNITS (mode);
7028 u = GET_MODE_UNIT_SIZE (mode);
7029
7030 /* Set QIMODE to a different vector mode with byte elements.
7031 If no such mode, or if MODE already has byte elements, use VOIDmode. */
7032 qimode = VOIDmode;
7033 if (GET_MODE_INNER (mode) != QImode)
7034 {
7035 qimode = mode_for_vector (QImode, w);
7036 if (!VECTOR_MODE_P (qimode))
7037 qimode = VOIDmode;
7038 }
7039
7040 /* If the input is a constant, expand it specially. */
7041 if (CONSTANT_P (sel))
7042 {
7043 icode = direct_optab_handler (vec_perm_const_optab, mode);
7044 if (icode != CODE_FOR_nothing)
7045 {
7046 tmp = expand_vec_perm_1 (icode, target, v0, v1, sel);
7047 if (tmp)
7048 return tmp;
7049 }
7050
7051 /* Fall back to a constant byte-based permutation. */
7052 if (qimode != VOIDmode)
7053 {
7054 vec = rtvec_alloc (w);
7055 for (i = 0; i < e; ++i)
7056 {
7057 unsigned int j, this_e;
7058
7059 this_e = INTVAL (XVECEXP (sel, 0, i));
7060 this_e &= 2 * e - 1;
7061 this_e *= u;
7062
7063 for (j = 0; j < u; ++j)
7064 RTVEC_ELT (vec, i * u + j) = GEN_INT (this_e + j);
7065 }
7066 sel_qi = gen_rtx_CONST_VECTOR (qimode, vec);
7067
7068 icode = direct_optab_handler (vec_perm_const_optab, qimode);
7069 if (icode != CODE_FOR_nothing)
7070 {
7071 tmp = expand_vec_perm_1 (icode, gen_lowpart (qimode, target),
7072 gen_lowpart (qimode, v0),
7073 gen_lowpart (qimode, v1), sel_qi);
7074 if (tmp)
7075 return gen_lowpart (mode, tmp);
7076 }
7077 }
7078 }
7079
7080 /* Otherwise expand as a fully variable permuation. */
7081 icode = direct_optab_handler (vec_perm_optab, mode);
7082 if (icode != CODE_FOR_nothing)
7083 {
7084 tmp = expand_vec_perm_1 (icode, target, v0, v1, sel);
7085 if (tmp)
7086 return tmp;
7087 }
7088
7089 /* As a special case to aid several targets, lower the element-based
7090 permutation to a byte-based permutation and try again. */
7091 if (qimode == VOIDmode)
7092 return NULL_RTX;
7093 icode = direct_optab_handler (vec_perm_optab, qimode);
7094 if (icode == CODE_FOR_nothing)
7095 return NULL_RTX;
7096
7097 if (sel_qi == NULL)
7098 {
7099 /* Multiply each element by its byte size. */
7100 enum machine_mode selmode = GET_MODE (sel);
7101 if (u == 2)
7102 sel = expand_simple_binop (selmode, PLUS, sel, sel,
7103 sel, 0, OPTAB_DIRECT);
7104 else
7105 sel = expand_simple_binop (selmode, ASHIFT, sel,
7106 GEN_INT (exact_log2 (u)),
7107 sel, 0, OPTAB_DIRECT);
7108 gcc_assert (sel != NULL);
7109
7110 /* Broadcast the low byte each element into each of its bytes. */
7111 vec = rtvec_alloc (w);
7112 for (i = 0; i < w; ++i)
7113 {
7114 int this_e = i / u * u;
7115 if (BYTES_BIG_ENDIAN)
7116 this_e += u - 1;
7117 RTVEC_ELT (vec, i) = GEN_INT (this_e);
7118 }
7119 tmp = gen_rtx_CONST_VECTOR (qimode, vec);
7120 sel = gen_lowpart (qimode, sel);
7121 sel = expand_vec_perm (qimode, sel, sel, tmp, NULL);
7122 gcc_assert (sel != NULL);
7123
7124 /* Add the byte offset to each byte element. */
7125 /* Note that the definition of the indicies here is memory ordering,
7126 so there should be no difference between big and little endian. */
7127 vec = rtvec_alloc (w);
7128 for (i = 0; i < w; ++i)
7129 RTVEC_ELT (vec, i) = GEN_INT (i % u);
7130 tmp = gen_rtx_CONST_VECTOR (qimode, vec);
7131 sel_qi = expand_simple_binop (qimode, PLUS, sel, tmp,
7132 sel, 0, OPTAB_DIRECT);
7133 gcc_assert (sel_qi != NULL);
7134 }
7135
7136 tmp = expand_vec_perm_1 (icode, gen_lowpart (qimode, target),
7137 gen_lowpart (qimode, v0),
7138 gen_lowpart (qimode, v1), sel_qi);
7139 if (tmp)
7140 tmp = gen_lowpart (mode, tmp);
7141 return tmp;
7142 }
7143
7144 /* Return insn code for a conditional operator with a comparison in
7145 mode CMODE, unsigned if UNS is true, resulting in a value of mode VMODE. */
7146
7147 static inline enum insn_code
7148 get_vcond_icode (enum machine_mode vmode, enum machine_mode cmode, bool uns)
7149 {
7150 enum insn_code icode = CODE_FOR_nothing;
7151 if (uns)
7152 icode = convert_optab_handler (vcondu_optab, vmode, cmode);
7153 else
7154 icode = convert_optab_handler (vcond_optab, vmode, cmode);
7155 return icode;
7156 }
7157
7158 /* Return TRUE iff, appropriate vector insns are available
7159 for vector cond expr with vector type VALUE_TYPE and a comparison
7160 with operand vector types in CMP_OP_TYPE. */
7161
7162 bool
7163 expand_vec_cond_expr_p (tree value_type, tree cmp_op_type)
7164 {
7165 enum machine_mode value_mode = TYPE_MODE (value_type);
7166 enum machine_mode cmp_op_mode = TYPE_MODE (cmp_op_type);
7167 if (GET_MODE_SIZE (value_mode) != GET_MODE_SIZE (cmp_op_mode)
7168 || GET_MODE_NUNITS (value_mode) != GET_MODE_NUNITS (cmp_op_mode)
7169 || get_vcond_icode (TYPE_MODE (value_type), TYPE_MODE (cmp_op_type),
7170 TYPE_UNSIGNED (cmp_op_type)) == CODE_FOR_nothing)
7171 return false;
7172 return true;
7173 }
7174
7175 /* Generate insns for a VEC_COND_EXPR, given its TYPE and its
7176 three operands. */
7177
7178 rtx
7179 expand_vec_cond_expr (tree vec_cond_type, tree op0, tree op1, tree op2,
7180 rtx target)
7181 {
7182 struct expand_operand ops[6];
7183 enum insn_code icode;
7184 rtx comparison, rtx_op1, rtx_op2;
7185 enum machine_mode mode = TYPE_MODE (vec_cond_type);
7186 enum machine_mode cmp_op_mode;
7187 bool unsignedp;
7188
7189 gcc_assert (COMPARISON_CLASS_P (op0));
7190
7191 unsignedp = TYPE_UNSIGNED (TREE_TYPE (TREE_OPERAND (op0, 0)));
7192 cmp_op_mode = TYPE_MODE (TREE_TYPE (TREE_OPERAND (op0, 0)));
7193
7194 gcc_assert (GET_MODE_SIZE (mode) == GET_MODE_SIZE (cmp_op_mode)
7195 && GET_MODE_NUNITS (mode) == GET_MODE_NUNITS (cmp_op_mode));
7196
7197 icode = get_vcond_icode (mode, cmp_op_mode, unsignedp);
7198 if (icode == CODE_FOR_nothing)
7199 return 0;
7200
7201 comparison = vector_compare_rtx (op0, unsignedp, icode);
7202 rtx_op1 = expand_normal (op1);
7203 rtx_op2 = expand_normal (op2);
7204
7205 create_output_operand (&ops[0], target, mode);
7206 create_input_operand (&ops[1], rtx_op1, mode);
7207 create_input_operand (&ops[2], rtx_op2, mode);
7208 create_fixed_operand (&ops[3], comparison);
7209 create_fixed_operand (&ops[4], XEXP (comparison, 0));
7210 create_fixed_operand (&ops[5], XEXP (comparison, 1));
7211 expand_insn (icode, 6, ops);
7212 return ops[0].value;
7213 }
7214
7215 \f
7216 /* Return true if there is a compare_and_swap pattern. */
7217
7218 bool
7219 can_compare_and_swap_p (enum machine_mode mode, bool allow_libcall)
7220 {
7221 enum insn_code icode;
7222
7223 /* Check for __atomic_compare_and_swap. */
7224 icode = direct_optab_handler (atomic_compare_and_swap_optab, mode);
7225 if (icode != CODE_FOR_nothing)
7226 return true;
7227
7228 /* Check for __sync_compare_and_swap. */
7229 icode = optab_handler (sync_compare_and_swap_optab, mode);
7230 if (icode != CODE_FOR_nothing)
7231 return true;
7232 if (allow_libcall && optab_libfunc (sync_compare_and_swap_optab, mode))
7233 return true;
7234
7235 /* No inline compare and swap. */
7236 return false;
7237 }
7238
7239 /* Return true if an atomic exchange can be performed. */
7240
7241 bool
7242 can_atomic_exchange_p (enum machine_mode mode, bool allow_libcall)
7243 {
7244 enum insn_code icode;
7245
7246 /* Check for __atomic_exchange. */
7247 icode = direct_optab_handler (atomic_exchange_optab, mode);
7248 if (icode != CODE_FOR_nothing)
7249 return true;
7250
7251 /* Don't check __sync_test_and_set, as on some platforms that
7252 has reduced functionality. Targets that really do support
7253 a proper exchange should simply be updated to the __atomics. */
7254
7255 return can_compare_and_swap_p (mode, allow_libcall);
7256 }
7257
7258
7259 /* Helper function to find the MODE_CC set in a sync_compare_and_swap
7260 pattern. */
7261
7262 static void
7263 find_cc_set (rtx x, const_rtx pat, void *data)
7264 {
7265 if (REG_P (x) && GET_MODE_CLASS (GET_MODE (x)) == MODE_CC
7266 && GET_CODE (pat) == SET)
7267 {
7268 rtx *p_cc_reg = (rtx *) data;
7269 gcc_assert (!*p_cc_reg);
7270 *p_cc_reg = x;
7271 }
7272 }
7273
7274 /* This is a helper function for the other atomic operations. This function
7275 emits a loop that contains SEQ that iterates until a compare-and-swap
7276 operation at the end succeeds. MEM is the memory to be modified. SEQ is
7277 a set of instructions that takes a value from OLD_REG as an input and
7278 produces a value in NEW_REG as an output. Before SEQ, OLD_REG will be
7279 set to the current contents of MEM. After SEQ, a compare-and-swap will
7280 attempt to update MEM with NEW_REG. The function returns true when the
7281 loop was generated successfully. */
7282
7283 static bool
7284 expand_compare_and_swap_loop (rtx mem, rtx old_reg, rtx new_reg, rtx seq)
7285 {
7286 enum machine_mode mode = GET_MODE (mem);
7287 rtx label, cmp_reg, success, oldval;
7288
7289 /* The loop we want to generate looks like
7290
7291 cmp_reg = mem;
7292 label:
7293 old_reg = cmp_reg;
7294 seq;
7295 (success, cmp_reg) = compare-and-swap(mem, old_reg, new_reg)
7296 if (success)
7297 goto label;
7298
7299 Note that we only do the plain load from memory once. Subsequent
7300 iterations use the value loaded by the compare-and-swap pattern. */
7301
7302 label = gen_label_rtx ();
7303 cmp_reg = gen_reg_rtx (mode);
7304
7305 emit_move_insn (cmp_reg, mem);
7306 emit_label (label);
7307 emit_move_insn (old_reg, cmp_reg);
7308 if (seq)
7309 emit_insn (seq);
7310
7311 success = NULL_RTX;
7312 oldval = cmp_reg;
7313 if (!expand_atomic_compare_and_swap (&success, &oldval, mem, old_reg,
7314 new_reg, false, MEMMODEL_SEQ_CST,
7315 MEMMODEL_RELAXED))
7316 return false;
7317
7318 if (oldval != cmp_reg)
7319 emit_move_insn (cmp_reg, oldval);
7320
7321 /* ??? Mark this jump predicted not taken? */
7322 emit_cmp_and_jump_insns (success, const0_rtx, EQ, const0_rtx,
7323 GET_MODE (success), 1, label);
7324 return true;
7325 }
7326
7327
7328 /* This function tries to emit an atomic_exchange intruction. VAL is written
7329 to *MEM using memory model MODEL. The previous contents of *MEM are returned,
7330 using TARGET if possible. */
7331
7332 static rtx
7333 maybe_emit_atomic_exchange (rtx target, rtx mem, rtx val, enum memmodel model)
7334 {
7335 enum machine_mode mode = GET_MODE (mem);
7336 enum insn_code icode;
7337
7338 /* If the target supports the exchange directly, great. */
7339 icode = direct_optab_handler (atomic_exchange_optab, mode);
7340 if (icode != CODE_FOR_nothing)
7341 {
7342 struct expand_operand ops[4];
7343
7344 create_output_operand (&ops[0], target, mode);
7345 create_fixed_operand (&ops[1], mem);
7346 /* VAL may have been promoted to a wider mode. Shrink it if so. */
7347 create_convert_operand_to (&ops[2], val, mode, true);
7348 create_integer_operand (&ops[3], model);
7349 if (maybe_expand_insn (icode, 4, ops))
7350 return ops[0].value;
7351 }
7352
7353 return NULL_RTX;
7354 }
7355
7356 /* This function tries to implement an atomic exchange operation using
7357 __sync_lock_test_and_set. VAL is written to *MEM using memory model MODEL.
7358 The previous contents of *MEM are returned, using TARGET if possible.
7359 Since this instructionn is an acquire barrier only, stronger memory
7360 models may require additional barriers to be emitted. */
7361
7362 static rtx
7363 maybe_emit_sync_lock_test_and_set (rtx target, rtx mem, rtx val,
7364 enum memmodel model)
7365 {
7366 enum machine_mode mode = GET_MODE (mem);
7367 enum insn_code icode;
7368 rtx last_insn = get_last_insn ();
7369
7370 icode = optab_handler (sync_lock_test_and_set_optab, mode);
7371
7372 /* Legacy sync_lock_test_and_set is an acquire barrier. If the pattern
7373 exists, and the memory model is stronger than acquire, add a release
7374 barrier before the instruction. */
7375
7376 if (model == MEMMODEL_SEQ_CST
7377 || model == MEMMODEL_RELEASE
7378 || model == MEMMODEL_ACQ_REL)
7379 expand_mem_thread_fence (model);
7380
7381 if (icode != CODE_FOR_nothing)
7382 {
7383 struct expand_operand ops[3];
7384 create_output_operand (&ops[0], target, mode);
7385 create_fixed_operand (&ops[1], mem);
7386 /* VAL may have been promoted to a wider mode. Shrink it if so. */
7387 create_convert_operand_to (&ops[2], val, mode, true);
7388 if (maybe_expand_insn (icode, 3, ops))
7389 return ops[0].value;
7390 }
7391
7392 /* If an external test-and-set libcall is provided, use that instead of
7393 any external compare-and-swap that we might get from the compare-and-
7394 swap-loop expansion later. */
7395 if (!can_compare_and_swap_p (mode, false))
7396 {
7397 rtx libfunc = optab_libfunc (sync_lock_test_and_set_optab, mode);
7398 if (libfunc != NULL)
7399 {
7400 rtx addr;
7401
7402 addr = convert_memory_address (ptr_mode, XEXP (mem, 0));
7403 return emit_library_call_value (libfunc, NULL_RTX, LCT_NORMAL,
7404 mode, 2, addr, ptr_mode,
7405 val, mode);
7406 }
7407 }
7408
7409 /* If the test_and_set can't be emitted, eliminate any barrier that might
7410 have been emitted. */
7411 delete_insns_since (last_insn);
7412 return NULL_RTX;
7413 }
7414
7415 /* This function tries to implement an atomic exchange operation using a
7416 compare_and_swap loop. VAL is written to *MEM. The previous contents of
7417 *MEM are returned, using TARGET if possible. No memory model is required
7418 since a compare_and_swap loop is seq-cst. */
7419
7420 static rtx
7421 maybe_emit_compare_and_swap_exchange_loop (rtx target, rtx mem, rtx val)
7422 {
7423 enum machine_mode mode = GET_MODE (mem);
7424
7425 if (can_compare_and_swap_p (mode, true))
7426 {
7427 if (!target || !register_operand (target, mode))
7428 target = gen_reg_rtx (mode);
7429 if (GET_MODE (val) != VOIDmode && GET_MODE (val) != mode)
7430 val = convert_modes (mode, GET_MODE (val), val, 1);
7431 if (expand_compare_and_swap_loop (mem, target, val, NULL_RTX))
7432 return target;
7433 }
7434
7435 return NULL_RTX;
7436 }
7437
7438 #ifndef HAVE_atomic_test_and_set
7439 #define HAVE_atomic_test_and_set 0
7440 #define gen_atomic_test_and_set(x,y,z) (gcc_unreachable (), NULL_RTX)
7441 #endif
7442
7443 /* This function expands the legacy _sync_lock test_and_set operation which is
7444 generally an atomic exchange. Some limited targets only allow the
7445 constant 1 to be stored. This is an ACQUIRE operation.
7446
7447 TARGET is an optional place to stick the return value.
7448 MEM is where VAL is stored. */
7449
7450 rtx
7451 expand_sync_lock_test_and_set (rtx target, rtx mem, rtx val)
7452 {
7453 rtx ret;
7454
7455 /* Try an atomic_exchange first. */
7456 ret = maybe_emit_atomic_exchange (target, mem, val, MEMMODEL_ACQUIRE);
7457
7458 if (!ret)
7459 ret = maybe_emit_sync_lock_test_and_set (target, mem, val,
7460 MEMMODEL_ACQUIRE);
7461 if (!ret)
7462 ret = maybe_emit_compare_and_swap_exchange_loop (target, mem, val);
7463
7464 /* If there are no other options, try atomic_test_and_set if the value
7465 being stored is 1. */
7466 if (!ret && val == const1_rtx && HAVE_atomic_test_and_set)
7467 {
7468 ret = gen_atomic_test_and_set (target, mem, GEN_INT (MEMMODEL_ACQUIRE));
7469 emit_insn (ret);
7470 }
7471
7472 return ret;
7473 }
7474
7475 /* This function expands the atomic test_and_set operation:
7476 atomically store a boolean TRUE into MEM and return the previous value.
7477
7478 MEMMODEL is the memory model variant to use.
7479 TARGET is an optional place to stick the return value. */
7480
7481 rtx
7482 expand_atomic_test_and_set (rtx target, rtx mem, enum memmodel model)
7483 {
7484 enum machine_mode mode = GET_MODE (mem);
7485 rtx ret = NULL_RTX;
7486
7487 if (target == NULL_RTX)
7488 target = gen_reg_rtx (mode);
7489
7490 if (HAVE_atomic_test_and_set)
7491 {
7492 ret = gen_atomic_test_and_set (target, mem, GEN_INT (MEMMODEL_ACQUIRE));
7493 emit_insn (ret);
7494 return ret;
7495 }
7496
7497 /* If there is no test and set, try exchange, then a compare_and_swap loop,
7498 then __sync_test_and_set. */
7499 ret = maybe_emit_atomic_exchange (target, mem, const1_rtx, model);
7500
7501 if (!ret)
7502 ret = maybe_emit_compare_and_swap_exchange_loop (target, mem, const1_rtx);
7503
7504 if (!ret)
7505 ret = maybe_emit_sync_lock_test_and_set (target, mem, const1_rtx, model);
7506
7507 if (ret)
7508 return ret;
7509
7510 /* Failing all else, assume a single threaded environment and simply perform
7511 the operation. */
7512 emit_move_insn (target, mem);
7513 emit_move_insn (mem, const1_rtx);
7514 return target;
7515 }
7516
7517 /* This function expands the atomic exchange operation:
7518 atomically store VAL in MEM and return the previous value in MEM.
7519
7520 MEMMODEL is the memory model variant to use.
7521 TARGET is an optional place to stick the return value. */
7522
7523 rtx
7524 expand_atomic_exchange (rtx target, rtx mem, rtx val, enum memmodel model)
7525 {
7526 rtx ret;
7527
7528 ret = maybe_emit_atomic_exchange (target, mem, val, model);
7529
7530 /* Next try a compare-and-swap loop for the exchange. */
7531 if (!ret)
7532 ret = maybe_emit_compare_and_swap_exchange_loop (target, mem, val);
7533
7534 return ret;
7535 }
7536
7537 /* This function expands the atomic compare exchange operation:
7538
7539 *PTARGET_BOOL is an optional place to store the boolean success/failure.
7540 *PTARGET_OVAL is an optional place to store the old value from memory.
7541 Both target parameters may be NULL to indicate that we do not care about
7542 that return value. Both target parameters are updated on success to
7543 the actual location of the corresponding result.
7544
7545 MEMMODEL is the memory model variant to use.
7546
7547 The return value of the function is true for success. */
7548
7549 bool
7550 expand_atomic_compare_and_swap (rtx *ptarget_bool, rtx *ptarget_oval,
7551 rtx mem, rtx expected, rtx desired,
7552 bool is_weak, enum memmodel succ_model,
7553 enum memmodel fail_model)
7554 {
7555 enum machine_mode mode = GET_MODE (mem);
7556 struct expand_operand ops[8];
7557 enum insn_code icode;
7558 rtx target_oval, target_bool = NULL_RTX;
7559 rtx libfunc;
7560
7561 /* Load expected into a register for the compare and swap. */
7562 if (MEM_P (expected))
7563 expected = copy_to_reg (expected);
7564
7565 /* Make sure we always have some place to put the return oldval.
7566 Further, make sure that place is distinct from the input expected,
7567 just in case we need that path down below. */
7568 if (ptarget_oval == NULL
7569 || (target_oval = *ptarget_oval) == NULL
7570 || reg_overlap_mentioned_p (expected, target_oval))
7571 target_oval = gen_reg_rtx (mode);
7572
7573 icode = direct_optab_handler (atomic_compare_and_swap_optab, mode);
7574 if (icode != CODE_FOR_nothing)
7575 {
7576 enum machine_mode bool_mode = insn_data[icode].operand[0].mode;
7577
7578 /* Make sure we always have a place for the bool operand. */
7579 if (ptarget_bool == NULL
7580 || (target_bool = *ptarget_bool) == NULL
7581 || GET_MODE (target_bool) != bool_mode)
7582 target_bool = gen_reg_rtx (bool_mode);
7583
7584 /* Emit the compare_and_swap. */
7585 create_output_operand (&ops[0], target_bool, bool_mode);
7586 create_output_operand (&ops[1], target_oval, mode);
7587 create_fixed_operand (&ops[2], mem);
7588 create_convert_operand_to (&ops[3], expected, mode, true);
7589 create_convert_operand_to (&ops[4], desired, mode, true);
7590 create_integer_operand (&ops[5], is_weak);
7591 create_integer_operand (&ops[6], succ_model);
7592 create_integer_operand (&ops[7], fail_model);
7593 expand_insn (icode, 8, ops);
7594
7595 /* Return success/failure. */
7596 target_bool = ops[0].value;
7597 target_oval = ops[1].value;
7598 goto success;
7599 }
7600
7601 /* Otherwise fall back to the original __sync_val_compare_and_swap
7602 which is always seq-cst. */
7603 icode = optab_handler (sync_compare_and_swap_optab, mode);
7604 if (icode != CODE_FOR_nothing)
7605 {
7606 rtx cc_reg;
7607
7608 create_output_operand (&ops[0], target_oval, mode);
7609 create_fixed_operand (&ops[1], mem);
7610 create_convert_operand_to (&ops[2], expected, mode, true);
7611 create_convert_operand_to (&ops[3], desired, mode, true);
7612 if (!maybe_expand_insn (icode, 4, ops))
7613 return false;
7614
7615 target_oval = ops[0].value;
7616
7617 /* If the caller isn't interested in the boolean return value,
7618 skip the computation of it. */
7619 if (ptarget_bool == NULL)
7620 goto success;
7621
7622 /* Otherwise, work out if the compare-and-swap succeeded. */
7623 cc_reg = NULL_RTX;
7624 if (have_insn_for (COMPARE, CCmode))
7625 note_stores (PATTERN (get_last_insn ()), find_cc_set, &cc_reg);
7626 if (cc_reg)
7627 {
7628 target_bool = emit_store_flag_force (target_bool, EQ, cc_reg,
7629 const0_rtx, VOIDmode, 0, 1);
7630 goto success;
7631 }
7632 goto success_bool_from_val;
7633 }
7634
7635 /* Also check for library support for __sync_val_compare_and_swap. */
7636 libfunc = optab_libfunc (sync_compare_and_swap_optab, mode);
7637 if (libfunc != NULL)
7638 {
7639 rtx addr = convert_memory_address (ptr_mode, XEXP (mem, 0));
7640 target_oval = emit_library_call_value (libfunc, NULL_RTX, LCT_NORMAL,
7641 mode, 3, addr, ptr_mode,
7642 expected, mode, desired, mode);
7643
7644 /* Compute the boolean return value only if requested. */
7645 if (ptarget_bool)
7646 goto success_bool_from_val;
7647 else
7648 goto success;
7649 }
7650
7651 /* Failure. */
7652 return false;
7653
7654 success_bool_from_val:
7655 target_bool = emit_store_flag_force (target_bool, EQ, target_oval,
7656 expected, VOIDmode, 1, 1);
7657 success:
7658 /* Make sure that the oval output winds up where the caller asked. */
7659 if (ptarget_oval)
7660 *ptarget_oval = target_oval;
7661 if (ptarget_bool)
7662 *ptarget_bool = target_bool;
7663 return true;
7664 }
7665
7666 /* Generate asm volatile("" : : : "memory") as the memory barrier. */
7667
7668 static void
7669 expand_asm_memory_barrier (void)
7670 {
7671 rtx asm_op, clob;
7672
7673 asm_op = gen_rtx_ASM_OPERANDS (VOIDmode, empty_string, empty_string, 0,
7674 rtvec_alloc (0), rtvec_alloc (0),
7675 rtvec_alloc (0), UNKNOWN_LOCATION);
7676 MEM_VOLATILE_P (asm_op) = 1;
7677
7678 clob = gen_rtx_SCRATCH (VOIDmode);
7679 clob = gen_rtx_MEM (BLKmode, clob);
7680 clob = gen_rtx_CLOBBER (VOIDmode, clob);
7681
7682 emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, asm_op, clob)));
7683 }
7684
7685 /* This routine will either emit the mem_thread_fence pattern or issue a
7686 sync_synchronize to generate a fence for memory model MEMMODEL. */
7687
7688 #ifndef HAVE_mem_thread_fence
7689 # define HAVE_mem_thread_fence 0
7690 # define gen_mem_thread_fence(x) (gcc_unreachable (), NULL_RTX)
7691 #endif
7692 #ifndef HAVE_memory_barrier
7693 # define HAVE_memory_barrier 0
7694 # define gen_memory_barrier() (gcc_unreachable (), NULL_RTX)
7695 #endif
7696
7697 void
7698 expand_mem_thread_fence (enum memmodel model)
7699 {
7700 if (HAVE_mem_thread_fence)
7701 emit_insn (gen_mem_thread_fence (GEN_INT (model)));
7702 else if (model != MEMMODEL_RELAXED)
7703 {
7704 if (HAVE_memory_barrier)
7705 emit_insn (gen_memory_barrier ());
7706 else if (synchronize_libfunc != NULL_RTX)
7707 emit_library_call (synchronize_libfunc, LCT_NORMAL, VOIDmode, 0);
7708 else
7709 expand_asm_memory_barrier ();
7710 }
7711 }
7712
7713 /* This routine will either emit the mem_signal_fence pattern or issue a
7714 sync_synchronize to generate a fence for memory model MEMMODEL. */
7715
7716 #ifndef HAVE_mem_signal_fence
7717 # define HAVE_mem_signal_fence 0
7718 # define gen_mem_signal_fence(x) (gcc_unreachable (), NULL_RTX)
7719 #endif
7720
7721 void
7722 expand_mem_signal_fence (enum memmodel model)
7723 {
7724 if (HAVE_mem_signal_fence)
7725 emit_insn (gen_mem_signal_fence (GEN_INT (model)));
7726 else if (model != MEMMODEL_RELAXED)
7727 {
7728 /* By default targets are coherent between a thread and the signal
7729 handler running on the same thread. Thus this really becomes a
7730 compiler barrier, in that stores must not be sunk past
7731 (or raised above) a given point. */
7732 expand_asm_memory_barrier ();
7733 }
7734 }
7735
7736 /* This function expands the atomic load operation:
7737 return the atomically loaded value in MEM.
7738
7739 MEMMODEL is the memory model variant to use.
7740 TARGET is an option place to stick the return value. */
7741
7742 rtx
7743 expand_atomic_load (rtx target, rtx mem, enum memmodel model)
7744 {
7745 enum machine_mode mode = GET_MODE (mem);
7746 enum insn_code icode;
7747
7748 /* If the target supports the load directly, great. */
7749 icode = direct_optab_handler (atomic_load_optab, mode);
7750 if (icode != CODE_FOR_nothing)
7751 {
7752 struct expand_operand ops[3];
7753
7754 create_output_operand (&ops[0], target, mode);
7755 create_fixed_operand (&ops[1], mem);
7756 create_integer_operand (&ops[2], model);
7757 if (maybe_expand_insn (icode, 3, ops))
7758 return ops[0].value;
7759 }
7760
7761 /* If the size of the object is greater than word size on this target,
7762 then we assume that a load will not be atomic. */
7763 if (GET_MODE_PRECISION (mode) > BITS_PER_WORD)
7764 {
7765 /* Issue val = compare_and_swap (mem, 0, 0).
7766 This may cause the occasional harmless store of 0 when the value is
7767 already 0, but it seems to be OK according to the standards guys. */
7768 expand_atomic_compare_and_swap (NULL, &target, mem, const0_rtx,
7769 const0_rtx, false, model, model);
7770 return target;
7771 }
7772
7773 /* Otherwise assume loads are atomic, and emit the proper barriers. */
7774 if (!target || target == const0_rtx)
7775 target = gen_reg_rtx (mode);
7776
7777 /* Emit the appropriate barrier before the load. */
7778 expand_mem_thread_fence (model);
7779
7780 emit_move_insn (target, mem);
7781
7782 /* For SEQ_CST, also emit a barrier after the load. */
7783 if (model == MEMMODEL_SEQ_CST)
7784 expand_mem_thread_fence (model);
7785
7786 return target;
7787 }
7788
7789 /* This function expands the atomic store operation:
7790 Atomically store VAL in MEM.
7791 MEMMODEL is the memory model variant to use.
7792 USE_RELEASE is true if __sync_lock_release can be used as a fall back.
7793 function returns const0_rtx if a pattern was emitted. */
7794
7795 rtx
7796 expand_atomic_store (rtx mem, rtx val, enum memmodel model, bool use_release)
7797 {
7798 enum machine_mode mode = GET_MODE (mem);
7799 enum insn_code icode;
7800 struct expand_operand ops[3];
7801
7802 /* If the target supports the store directly, great. */
7803 icode = direct_optab_handler (atomic_store_optab, mode);
7804 if (icode != CODE_FOR_nothing)
7805 {
7806 create_fixed_operand (&ops[0], mem);
7807 create_input_operand (&ops[1], val, mode);
7808 create_integer_operand (&ops[2], model);
7809 if (maybe_expand_insn (icode, 3, ops))
7810 return const0_rtx;
7811 }
7812
7813 /* If using __sync_lock_release is a viable alternative, try it. */
7814 if (use_release)
7815 {
7816 icode = direct_optab_handler (sync_lock_release_optab, mode);
7817 if (icode != CODE_FOR_nothing)
7818 {
7819 create_fixed_operand (&ops[0], mem);
7820 create_input_operand (&ops[1], const0_rtx, mode);
7821 if (maybe_expand_insn (icode, 2, ops))
7822 {
7823 /* lock_release is only a release barrier. */
7824 if (model == MEMMODEL_SEQ_CST)
7825 expand_mem_thread_fence (model);
7826 return const0_rtx;
7827 }
7828 }
7829 }
7830
7831 /* If the size of the object is greater than word size on this target,
7832 a default store will not be atomic, Try a mem_exchange and throw away
7833 the result. If that doesn't work, don't do anything. */
7834 if (GET_MODE_PRECISION(mode) > BITS_PER_WORD)
7835 {
7836 rtx target = maybe_emit_atomic_exchange (NULL_RTX, mem, val, model);
7837 if (!target)
7838 target = maybe_emit_compare_and_swap_exchange_loop (NULL_RTX, mem, val);
7839 if (target)
7840 return const0_rtx;
7841 else
7842 return NULL_RTX;
7843 }
7844
7845 /* If there is no mem_store, default to a move with barriers */
7846 if (model == MEMMODEL_SEQ_CST || model == MEMMODEL_RELEASE)
7847 expand_mem_thread_fence (model);
7848
7849 emit_move_insn (mem, val);
7850
7851 /* For SEQ_CST, also emit a barrier after the load. */
7852 if (model == MEMMODEL_SEQ_CST)
7853 expand_mem_thread_fence (model);
7854
7855 return const0_rtx;
7856 }
7857
7858
7859 /* Structure containing the pointers and values required to process the
7860 various forms of the atomic_fetch_op and atomic_op_fetch builtins. */
7861
7862 struct atomic_op_functions
7863 {
7864 direct_optab mem_fetch_before;
7865 direct_optab mem_fetch_after;
7866 direct_optab mem_no_result;
7867 optab fetch_before;
7868 optab fetch_after;
7869 direct_optab no_result;
7870 enum rtx_code reverse_code;
7871 };
7872
7873
7874 /* Fill in structure pointed to by OP with the various optab entries for an
7875 operation of type CODE. */
7876
7877 static void
7878 get_atomic_op_for_code (struct atomic_op_functions *op, enum rtx_code code)
7879 {
7880 gcc_assert (op!= NULL);
7881
7882 /* If SWITCHABLE_TARGET is defined, then subtargets can be switched
7883 in the source code during compilation, and the optab entries are not
7884 computable until runtime. Fill in the values at runtime. */
7885 switch (code)
7886 {
7887 case PLUS:
7888 op->mem_fetch_before = atomic_fetch_add_optab;
7889 op->mem_fetch_after = atomic_add_fetch_optab;
7890 op->mem_no_result = atomic_add_optab;
7891 op->fetch_before = sync_old_add_optab;
7892 op->fetch_after = sync_new_add_optab;
7893 op->no_result = sync_add_optab;
7894 op->reverse_code = MINUS;
7895 break;
7896 case MINUS:
7897 op->mem_fetch_before = atomic_fetch_sub_optab;
7898 op->mem_fetch_after = atomic_sub_fetch_optab;
7899 op->mem_no_result = atomic_sub_optab;
7900 op->fetch_before = sync_old_sub_optab;
7901 op->fetch_after = sync_new_sub_optab;
7902 op->no_result = sync_sub_optab;
7903 op->reverse_code = PLUS;
7904 break;
7905 case XOR:
7906 op->mem_fetch_before = atomic_fetch_xor_optab;
7907 op->mem_fetch_after = atomic_xor_fetch_optab;
7908 op->mem_no_result = atomic_xor_optab;
7909 op->fetch_before = sync_old_xor_optab;
7910 op->fetch_after = sync_new_xor_optab;
7911 op->no_result = sync_xor_optab;
7912 op->reverse_code = XOR;
7913 break;
7914 case AND:
7915 op->mem_fetch_before = atomic_fetch_and_optab;
7916 op->mem_fetch_after = atomic_and_fetch_optab;
7917 op->mem_no_result = atomic_and_optab;
7918 op->fetch_before = sync_old_and_optab;
7919 op->fetch_after = sync_new_and_optab;
7920 op->no_result = sync_and_optab;
7921 op->reverse_code = UNKNOWN;
7922 break;
7923 case IOR:
7924 op->mem_fetch_before = atomic_fetch_or_optab;
7925 op->mem_fetch_after = atomic_or_fetch_optab;
7926 op->mem_no_result = atomic_or_optab;
7927 op->fetch_before = sync_old_ior_optab;
7928 op->fetch_after = sync_new_ior_optab;
7929 op->no_result = sync_ior_optab;
7930 op->reverse_code = UNKNOWN;
7931 break;
7932 case NOT:
7933 op->mem_fetch_before = atomic_fetch_nand_optab;
7934 op->mem_fetch_after = atomic_nand_fetch_optab;
7935 op->mem_no_result = atomic_nand_optab;
7936 op->fetch_before = sync_old_nand_optab;
7937 op->fetch_after = sync_new_nand_optab;
7938 op->no_result = sync_nand_optab;
7939 op->reverse_code = UNKNOWN;
7940 break;
7941 default:
7942 gcc_unreachable ();
7943 }
7944 }
7945
7946 /* See if there is a more optimal way to implement the operation "*MEM CODE VAL"
7947 using memory order MODEL. If AFTER is true the operation needs to return
7948 the value of *MEM after the operation, otherwise the previous value.
7949 TARGET is an optional place to place the result. The result is unused if
7950 it is const0_rtx.
7951 Return the result if there is a better sequence, otherwise NULL_RTX. */
7952
7953 static rtx
7954 maybe_optimize_fetch_op (rtx target, rtx mem, rtx val, enum rtx_code code,
7955 enum memmodel model, bool after)
7956 {
7957 /* If the value is prefetched, or not used, it may be possible to replace
7958 the sequence with a native exchange operation. */
7959 if (!after || target == const0_rtx)
7960 {
7961 /* fetch_and (&x, 0, m) can be replaced with exchange (&x, 0, m). */
7962 if (code == AND && val == const0_rtx)
7963 {
7964 if (target == const0_rtx)
7965 target = gen_reg_rtx (GET_MODE (mem));
7966 return maybe_emit_atomic_exchange (target, mem, val, model);
7967 }
7968
7969 /* fetch_or (&x, -1, m) can be replaced with exchange (&x, -1, m). */
7970 if (code == IOR && val == constm1_rtx)
7971 {
7972 if (target == const0_rtx)
7973 target = gen_reg_rtx (GET_MODE (mem));
7974 return maybe_emit_atomic_exchange (target, mem, val, model);
7975 }
7976 }
7977
7978 return NULL_RTX;
7979 }
7980
7981 /* Try to emit an instruction for a specific operation varaition.
7982 OPTAB contains the OP functions.
7983 TARGET is an optional place to return the result. const0_rtx means unused.
7984 MEM is the memory location to operate on.
7985 VAL is the value to use in the operation.
7986 USE_MEMMODEL is TRUE if the variation with a memory model should be tried.
7987 MODEL is the memory model, if used.
7988 AFTER is true if the returned result is the value after the operation. */
7989
7990 static rtx
7991 maybe_emit_op (const struct atomic_op_functions *optab, rtx target, rtx mem,
7992 rtx val, bool use_memmodel, enum memmodel model, bool after)
7993 {
7994 enum machine_mode mode = GET_MODE (mem);
7995 struct expand_operand ops[4];
7996 enum insn_code icode;
7997 int op_counter = 0;
7998 int num_ops;
7999
8000 /* Check to see if there is a result returned. */
8001 if (target == const0_rtx)
8002 {
8003 if (use_memmodel)
8004 {
8005 icode = direct_optab_handler (optab->mem_no_result, mode);
8006 create_integer_operand (&ops[2], model);
8007 num_ops = 3;
8008 }
8009 else
8010 {
8011 icode = direct_optab_handler (optab->no_result, mode);
8012 num_ops = 2;
8013 }
8014 }
8015 /* Otherwise, we need to generate a result. */
8016 else
8017 {
8018 if (use_memmodel)
8019 {
8020 icode = direct_optab_handler (after ? optab->mem_fetch_after
8021 : optab->mem_fetch_before, mode);
8022 create_integer_operand (&ops[3], model);
8023 num_ops = 4;
8024 }
8025 else
8026 {
8027 icode = optab_handler (after ? optab->fetch_after
8028 : optab->fetch_before, mode);
8029 num_ops = 3;
8030 }
8031 create_output_operand (&ops[op_counter++], target, mode);
8032 }
8033 if (icode == CODE_FOR_nothing)
8034 return NULL_RTX;
8035
8036 create_fixed_operand (&ops[op_counter++], mem);
8037 /* VAL may have been promoted to a wider mode. Shrink it if so. */
8038 create_convert_operand_to (&ops[op_counter++], val, mode, true);
8039
8040 if (maybe_expand_insn (icode, num_ops, ops))
8041 return (target == const0_rtx ? const0_rtx : ops[0].value);
8042
8043 return NULL_RTX;
8044 }
8045
8046
8047 /* This function expands an atomic fetch_OP or OP_fetch operation:
8048 TARGET is an option place to stick the return value. const0_rtx indicates
8049 the result is unused.
8050 atomically fetch MEM, perform the operation with VAL and return it to MEM.
8051 CODE is the operation being performed (OP)
8052 MEMMODEL is the memory model variant to use.
8053 AFTER is true to return the result of the operation (OP_fetch).
8054 AFTER is false to return the value before the operation (fetch_OP). */
8055 rtx
8056 expand_atomic_fetch_op (rtx target, rtx mem, rtx val, enum rtx_code code,
8057 enum memmodel model, bool after)
8058 {
8059 enum machine_mode mode = GET_MODE (mem);
8060 struct atomic_op_functions optab;
8061 rtx result;
8062 bool unused_result = (target == const0_rtx);
8063
8064 get_atomic_op_for_code (&optab, code);
8065
8066 /* Check to see if there are any better instructions. */
8067 result = maybe_optimize_fetch_op (target, mem, val, code, model, after);
8068 if (result)
8069 return result;
8070
8071 /* Check for the case where the result isn't used and try those patterns. */
8072 if (unused_result)
8073 {
8074 /* Try the memory model variant first. */
8075 result = maybe_emit_op (&optab, target, mem, val, true, model, true);
8076 if (result)
8077 return result;
8078
8079 /* Next try the old style withuot a memory model. */
8080 result = maybe_emit_op (&optab, target, mem, val, false, model, true);
8081 if (result)
8082 return result;
8083
8084 /* There is no no-result pattern, so try patterns with a result. */
8085 target = NULL_RTX;
8086 }
8087
8088 /* Try the __atomic version. */
8089 result = maybe_emit_op (&optab, target, mem, val, true, model, after);
8090 if (result)
8091 return result;
8092
8093 /* Try the older __sync version. */
8094 result = maybe_emit_op (&optab, target, mem, val, false, model, after);
8095 if (result)
8096 return result;
8097
8098 /* If the fetch value can be calculated from the other variation of fetch,
8099 try that operation. */
8100 if (after || unused_result || optab.reverse_code != UNKNOWN)
8101 {
8102 /* Try the __atomic version, then the older __sync version. */
8103 result = maybe_emit_op (&optab, target, mem, val, true, model, !after);
8104 if (!result)
8105 result = maybe_emit_op (&optab, target, mem, val, false, model, !after);
8106
8107 if (result)
8108 {
8109 /* If the result isn't used, no need to do compensation code. */
8110 if (unused_result)
8111 return result;
8112
8113 /* Issue compensation code. Fetch_after == fetch_before OP val.
8114 Fetch_before == after REVERSE_OP val. */
8115 if (!after)
8116 code = optab.reverse_code;
8117 if (code == NOT)
8118 {
8119 result = expand_simple_binop (mode, AND, result, val, NULL_RTX,
8120 true, OPTAB_LIB_WIDEN);
8121 result = expand_simple_unop (mode, NOT, result, target, true);
8122 }
8123 else
8124 result = expand_simple_binop (mode, code, result, val, target,
8125 true, OPTAB_LIB_WIDEN);
8126 return result;
8127 }
8128 }
8129
8130 /* Try the __sync libcalls only if we can't do compare-and-swap inline. */
8131 if (!can_compare_and_swap_p (mode, false))
8132 {
8133 rtx libfunc;
8134 bool fixup = false;
8135
8136 libfunc = optab_libfunc (after ? optab.fetch_after
8137 : optab.fetch_before, mode);
8138 if (libfunc == NULL
8139 && (after || unused_result || optab.reverse_code != UNKNOWN))
8140 {
8141 fixup = true;
8142 if (!after)
8143 code = optab.reverse_code;
8144 libfunc = optab_libfunc (after ? optab.fetch_before
8145 : optab.fetch_after, mode);
8146 }
8147 if (libfunc != NULL)
8148 {
8149 rtx addr = convert_memory_address (ptr_mode, XEXP (mem, 0));
8150 result = emit_library_call_value (libfunc, NULL, LCT_NORMAL, mode,
8151 2, addr, ptr_mode, val, mode);
8152
8153 if (!unused_result && fixup)
8154 result = expand_simple_binop (mode, code, result, val, target,
8155 true, OPTAB_LIB_WIDEN);
8156 return result;
8157 }
8158 }
8159
8160 /* If nothing else has succeeded, default to a compare and swap loop. */
8161 if (can_compare_and_swap_p (mode, true))
8162 {
8163 rtx insn;
8164 rtx t0 = gen_reg_rtx (mode), t1;
8165
8166 start_sequence ();
8167
8168 /* If the result is used, get a register for it. */
8169 if (!unused_result)
8170 {
8171 if (!target || !register_operand (target, mode))
8172 target = gen_reg_rtx (mode);
8173 /* If fetch_before, copy the value now. */
8174 if (!after)
8175 emit_move_insn (target, t0);
8176 }
8177 else
8178 target = const0_rtx;
8179
8180 t1 = t0;
8181 if (code == NOT)
8182 {
8183 t1 = expand_simple_binop (mode, AND, t1, val, NULL_RTX,
8184 true, OPTAB_LIB_WIDEN);
8185 t1 = expand_simple_unop (mode, code, t1, NULL_RTX, true);
8186 }
8187 else
8188 t1 = expand_simple_binop (mode, code, t1, val, NULL_RTX, true,
8189 OPTAB_LIB_WIDEN);
8190
8191 /* For after, copy the value now. */
8192 if (!unused_result && after)
8193 emit_move_insn (target, t1);
8194 insn = get_insns ();
8195 end_sequence ();
8196
8197 if (t1 != NULL && expand_compare_and_swap_loop (mem, t0, t1, insn))
8198 return target;
8199 }
8200
8201 return NULL_RTX;
8202 }
8203 \f
8204 /* Return true if OPERAND is suitable for operand number OPNO of
8205 instruction ICODE. */
8206
8207 bool
8208 insn_operand_matches (enum insn_code icode, unsigned int opno, rtx operand)
8209 {
8210 return (!insn_data[(int) icode].operand[opno].predicate
8211 || (insn_data[(int) icode].operand[opno].predicate
8212 (operand, insn_data[(int) icode].operand[opno].mode)));
8213 }
8214 \f
8215 /* TARGET is a target of a multiword operation that we are going to
8216 implement as a series of word-mode operations. Return true if
8217 TARGET is suitable for this purpose. */
8218
8219 bool
8220 valid_multiword_target_p (rtx target)
8221 {
8222 enum machine_mode mode;
8223 int i;
8224
8225 mode = GET_MODE (target);
8226 for (i = 0; i < GET_MODE_SIZE (mode); i += UNITS_PER_WORD)
8227 if (!validate_subreg (word_mode, mode, target, i))
8228 return false;
8229 return true;
8230 }
8231
8232 /* Like maybe_legitimize_operand, but do not change the code of the
8233 current rtx value. */
8234
8235 static bool
8236 maybe_legitimize_operand_same_code (enum insn_code icode, unsigned int opno,
8237 struct expand_operand *op)
8238 {
8239 /* See if the operand matches in its current form. */
8240 if (insn_operand_matches (icode, opno, op->value))
8241 return true;
8242
8243 /* If the operand is a memory whose address has no side effects,
8244 try forcing the address into a register. The check for side
8245 effects is important because force_reg cannot handle things
8246 like auto-modified addresses. */
8247 if (insn_data[(int) icode].operand[opno].allows_mem
8248 && MEM_P (op->value)
8249 && !side_effects_p (XEXP (op->value, 0)))
8250 {
8251 rtx addr, mem, last;
8252
8253 last = get_last_insn ();
8254 addr = force_reg (Pmode, XEXP (op->value, 0));
8255 mem = replace_equiv_address (op->value, addr);
8256 if (insn_operand_matches (icode, opno, mem))
8257 {
8258 op->value = mem;
8259 return true;
8260 }
8261 delete_insns_since (last);
8262 }
8263
8264 return false;
8265 }
8266
8267 /* Try to make OP match operand OPNO of instruction ICODE. Return true
8268 on success, storing the new operand value back in OP. */
8269
8270 static bool
8271 maybe_legitimize_operand (enum insn_code icode, unsigned int opno,
8272 struct expand_operand *op)
8273 {
8274 enum machine_mode mode, imode;
8275 bool old_volatile_ok, result;
8276
8277 mode = op->mode;
8278 switch (op->type)
8279 {
8280 case EXPAND_FIXED:
8281 old_volatile_ok = volatile_ok;
8282 volatile_ok = true;
8283 result = maybe_legitimize_operand_same_code (icode, opno, op);
8284 volatile_ok = old_volatile_ok;
8285 return result;
8286
8287 case EXPAND_OUTPUT:
8288 gcc_assert (mode != VOIDmode);
8289 if (op->value
8290 && op->value != const0_rtx
8291 && GET_MODE (op->value) == mode
8292 && maybe_legitimize_operand_same_code (icode, opno, op))
8293 return true;
8294
8295 op->value = gen_reg_rtx (mode);
8296 break;
8297
8298 case EXPAND_INPUT:
8299 input:
8300 gcc_assert (mode != VOIDmode);
8301 gcc_assert (GET_MODE (op->value) == VOIDmode
8302 || GET_MODE (op->value) == mode);
8303 if (maybe_legitimize_operand_same_code (icode, opno, op))
8304 return true;
8305
8306 op->value = copy_to_mode_reg (mode, op->value);
8307 break;
8308
8309 case EXPAND_CONVERT_TO:
8310 gcc_assert (mode != VOIDmode);
8311 op->value = convert_to_mode (mode, op->value, op->unsigned_p);
8312 goto input;
8313
8314 case EXPAND_CONVERT_FROM:
8315 if (GET_MODE (op->value) != VOIDmode)
8316 mode = GET_MODE (op->value);
8317 else
8318 /* The caller must tell us what mode this value has. */
8319 gcc_assert (mode != VOIDmode);
8320
8321 imode = insn_data[(int) icode].operand[opno].mode;
8322 if (imode != VOIDmode && imode != mode)
8323 {
8324 op->value = convert_modes (imode, mode, op->value, op->unsigned_p);
8325 mode = imode;
8326 }
8327 goto input;
8328
8329 case EXPAND_ADDRESS:
8330 gcc_assert (mode != VOIDmode);
8331 op->value = convert_memory_address (mode, op->value);
8332 goto input;
8333
8334 case EXPAND_INTEGER:
8335 mode = insn_data[(int) icode].operand[opno].mode;
8336 if (mode != VOIDmode && const_int_operand (op->value, mode))
8337 goto input;
8338 break;
8339 }
8340 return insn_operand_matches (icode, opno, op->value);
8341 }
8342
8343 /* Make OP describe an input operand that should have the same value
8344 as VALUE, after any mode conversion that the target might request.
8345 TYPE is the type of VALUE. */
8346
8347 void
8348 create_convert_operand_from_type (struct expand_operand *op,
8349 rtx value, tree type)
8350 {
8351 create_convert_operand_from (op, value, TYPE_MODE (type),
8352 TYPE_UNSIGNED (type));
8353 }
8354
8355 /* Try to make operands [OPS, OPS + NOPS) match operands [OPNO, OPNO + NOPS)
8356 of instruction ICODE. Return true on success, leaving the new operand
8357 values in the OPS themselves. Emit no code on failure. */
8358
8359 bool
8360 maybe_legitimize_operands (enum insn_code icode, unsigned int opno,
8361 unsigned int nops, struct expand_operand *ops)
8362 {
8363 rtx last;
8364 unsigned int i;
8365
8366 last = get_last_insn ();
8367 for (i = 0; i < nops; i++)
8368 if (!maybe_legitimize_operand (icode, opno + i, &ops[i]))
8369 {
8370 delete_insns_since (last);
8371 return false;
8372 }
8373 return true;
8374 }
8375
8376 /* Try to generate instruction ICODE, using operands [OPS, OPS + NOPS)
8377 as its operands. Return the instruction pattern on success,
8378 and emit any necessary set-up code. Return null and emit no
8379 code on failure. */
8380
8381 rtx
8382 maybe_gen_insn (enum insn_code icode, unsigned int nops,
8383 struct expand_operand *ops)
8384 {
8385 gcc_assert (nops == (unsigned int) insn_data[(int) icode].n_generator_args);
8386 if (!maybe_legitimize_operands (icode, 0, nops, ops))
8387 return NULL_RTX;
8388
8389 switch (nops)
8390 {
8391 case 1:
8392 return GEN_FCN (icode) (ops[0].value);
8393 case 2:
8394 return GEN_FCN (icode) (ops[0].value, ops[1].value);
8395 case 3:
8396 return GEN_FCN (icode) (ops[0].value, ops[1].value, ops[2].value);
8397 case 4:
8398 return GEN_FCN (icode) (ops[0].value, ops[1].value, ops[2].value,
8399 ops[3].value);
8400 case 5:
8401 return GEN_FCN (icode) (ops[0].value, ops[1].value, ops[2].value,
8402 ops[3].value, ops[4].value);
8403 case 6:
8404 return GEN_FCN (icode) (ops[0].value, ops[1].value, ops[2].value,
8405 ops[3].value, ops[4].value, ops[5].value);
8406 case 7:
8407 return GEN_FCN (icode) (ops[0].value, ops[1].value, ops[2].value,
8408 ops[3].value, ops[4].value, ops[5].value,
8409 ops[6].value);
8410 case 8:
8411 return GEN_FCN (icode) (ops[0].value, ops[1].value, ops[2].value,
8412 ops[3].value, ops[4].value, ops[5].value,
8413 ops[6].value, ops[7].value);
8414 }
8415 gcc_unreachable ();
8416 }
8417
8418 /* Try to emit instruction ICODE, using operands [OPS, OPS + NOPS)
8419 as its operands. Return true on success and emit no code on failure. */
8420
8421 bool
8422 maybe_expand_insn (enum insn_code icode, unsigned int nops,
8423 struct expand_operand *ops)
8424 {
8425 rtx pat = maybe_gen_insn (icode, nops, ops);
8426 if (pat)
8427 {
8428 emit_insn (pat);
8429 return true;
8430 }
8431 return false;
8432 }
8433
8434 /* Like maybe_expand_insn, but for jumps. */
8435
8436 bool
8437 maybe_expand_jump_insn (enum insn_code icode, unsigned int nops,
8438 struct expand_operand *ops)
8439 {
8440 rtx pat = maybe_gen_insn (icode, nops, ops);
8441 if (pat)
8442 {
8443 emit_jump_insn (pat);
8444 return true;
8445 }
8446 return false;
8447 }
8448
8449 /* Emit instruction ICODE, using operands [OPS, OPS + NOPS)
8450 as its operands. */
8451
8452 void
8453 expand_insn (enum insn_code icode, unsigned int nops,
8454 struct expand_operand *ops)
8455 {
8456 if (!maybe_expand_insn (icode, nops, ops))
8457 gcc_unreachable ();
8458 }
8459
8460 /* Like expand_insn, but for jumps. */
8461
8462 void
8463 expand_jump_insn (enum insn_code icode, unsigned int nops,
8464 struct expand_operand *ops)
8465 {
8466 if (!maybe_expand_jump_insn (icode, nops, ops))
8467 gcc_unreachable ();
8468 }
8469
8470 #include "gt-optabs.h"