PR rtl-optimization/82913
[official-gcc.git] / gcc / optabs-query.c
blob440462bc242340e46d04f86d41d3961817522096
1 /* IR-agnostic target query functions relating to optabs
2 Copyright (C) 1987-2017 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
21 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "target.h"
25 #include "insn-codes.h"
26 #include "optabs-query.h"
27 #include "optabs-libfuncs.h"
28 #include "insn-config.h"
29 #include "rtl.h"
30 #include "recog.h"
32 struct target_optabs default_target_optabs;
33 struct target_optabs *this_fn_optabs = &default_target_optabs;
34 #if SWITCHABLE_TARGET
35 struct target_optabs *this_target_optabs = &default_target_optabs;
36 #endif
38 /* Return the insn used to perform conversion OP from mode FROM_MODE
39 to mode TO_MODE; return CODE_FOR_nothing if the target does not have
40 such an insn, or if it is unsuitable for optimization type OPT_TYPE. */
42 insn_code
43 convert_optab_handler (convert_optab optab, machine_mode to_mode,
44 machine_mode from_mode, optimization_type opt_type)
46 insn_code icode = convert_optab_handler (optab, to_mode, from_mode);
47 if (icode == CODE_FOR_nothing
48 || !targetm.optab_supported_p (optab, to_mode, from_mode, opt_type))
49 return CODE_FOR_nothing;
50 return icode;
53 /* Return the insn used to implement mode MODE of OP; return
54 CODE_FOR_nothing if the target does not have such an insn,
55 or if it is unsuitable for optimization type OPT_TYPE. */
57 insn_code
58 direct_optab_handler (convert_optab optab, machine_mode mode,
59 optimization_type opt_type)
61 insn_code icode = direct_optab_handler (optab, mode);
62 if (icode == CODE_FOR_nothing
63 || !targetm.optab_supported_p (optab, mode, mode, opt_type))
64 return CODE_FOR_nothing;
65 return icode;
68 /* Enumerates the possible types of structure operand to an
69 extraction_insn. */
70 enum extraction_type { ET_unaligned_mem, ET_reg };
72 /* Check whether insv, extv or extzv pattern ICODE can be used for an
73 insertion or extraction of type TYPE on a structure of mode MODE.
74 Return true if so and fill in *INSN accordingly. STRUCT_OP is the
75 operand number of the structure (the first sign_extract or zero_extract
76 operand) and FIELD_OP is the operand number of the field (the other
77 side of the set from the sign_extract or zero_extract). */
79 static bool
80 get_traditional_extraction_insn (extraction_insn *insn,
81 enum extraction_type type,
82 machine_mode mode,
83 enum insn_code icode,
84 int struct_op, int field_op)
86 const struct insn_data_d *data = &insn_data[icode];
88 machine_mode struct_mode = data->operand[struct_op].mode;
89 if (struct_mode == VOIDmode)
90 struct_mode = word_mode;
91 if (mode != struct_mode)
92 return false;
94 machine_mode field_mode = data->operand[field_op].mode;
95 if (field_mode == VOIDmode)
96 field_mode = word_mode;
98 machine_mode pos_mode = data->operand[struct_op + 2].mode;
99 if (pos_mode == VOIDmode)
100 pos_mode = word_mode;
102 insn->icode = icode;
103 insn->field_mode = as_a <scalar_int_mode> (field_mode);
104 if (type == ET_unaligned_mem)
105 insn->struct_mode = byte_mode;
106 else if (struct_mode == BLKmode)
107 insn->struct_mode = opt_scalar_int_mode ();
108 else
109 insn->struct_mode = as_a <scalar_int_mode> (struct_mode);
110 insn->pos_mode = as_a <scalar_int_mode> (pos_mode);
111 return true;
114 /* Return true if an optab exists to perform an insertion or extraction
115 of type TYPE in mode MODE. Describe the instruction in *INSN if so.
117 REG_OPTAB is the optab to use for register structures and
118 MISALIGN_OPTAB is the optab to use for misaligned memory structures.
119 POS_OP is the operand number of the bit position. */
121 static bool
122 get_optab_extraction_insn (struct extraction_insn *insn,
123 enum extraction_type type,
124 machine_mode mode, direct_optab reg_optab,
125 direct_optab misalign_optab, int pos_op)
127 direct_optab optab = (type == ET_unaligned_mem ? misalign_optab : reg_optab);
128 enum insn_code icode = direct_optab_handler (optab, mode);
129 if (icode == CODE_FOR_nothing)
130 return false;
132 const struct insn_data_d *data = &insn_data[icode];
134 machine_mode pos_mode = data->operand[pos_op].mode;
135 if (pos_mode == VOIDmode)
136 pos_mode = word_mode;
138 insn->icode = icode;
139 insn->field_mode = as_a <scalar_int_mode> (mode);
140 if (type == ET_unaligned_mem)
141 insn->struct_mode = opt_scalar_int_mode ();
142 else
143 insn->struct_mode = insn->field_mode;
144 insn->pos_mode = as_a <scalar_int_mode> (pos_mode);
145 return true;
148 /* Return true if an instruction exists to perform an insertion or
149 extraction (PATTERN says which) of type TYPE in mode MODE.
150 Describe the instruction in *INSN if so. */
152 static bool
153 get_extraction_insn (extraction_insn *insn,
154 enum extraction_pattern pattern,
155 enum extraction_type type,
156 machine_mode mode)
158 switch (pattern)
160 case EP_insv:
161 if (targetm.have_insv ()
162 && get_traditional_extraction_insn (insn, type, mode,
163 targetm.code_for_insv, 0, 3))
164 return true;
165 return get_optab_extraction_insn (insn, type, mode, insv_optab,
166 insvmisalign_optab, 2);
168 case EP_extv:
169 if (targetm.have_extv ()
170 && get_traditional_extraction_insn (insn, type, mode,
171 targetm.code_for_extv, 1, 0))
172 return true;
173 return get_optab_extraction_insn (insn, type, mode, extv_optab,
174 extvmisalign_optab, 3);
176 case EP_extzv:
177 if (targetm.have_extzv ()
178 && get_traditional_extraction_insn (insn, type, mode,
179 targetm.code_for_extzv, 1, 0))
180 return true;
181 return get_optab_extraction_insn (insn, type, mode, extzv_optab,
182 extzvmisalign_optab, 3);
184 default:
185 gcc_unreachable ();
189 /* Return true if an instruction exists to access a field of mode
190 FIELDMODE in a structure that has STRUCT_BITS significant bits.
191 Describe the "best" such instruction in *INSN if so. PATTERN and
192 TYPE describe the type of insertion or extraction we want to perform.
194 For an insertion, the number of significant structure bits includes
195 all bits of the target. For an extraction, it need only include the
196 most significant bit of the field. Larger widths are acceptable
197 in both cases. */
199 static bool
200 get_best_extraction_insn (extraction_insn *insn,
201 enum extraction_pattern pattern,
202 enum extraction_type type,
203 unsigned HOST_WIDE_INT struct_bits,
204 machine_mode field_mode)
206 opt_scalar_int_mode mode_iter;
207 FOR_EACH_MODE_FROM (mode_iter, smallest_int_mode_for_size (struct_bits))
209 scalar_int_mode mode = mode_iter.require ();
210 if (get_extraction_insn (insn, pattern, type, mode))
212 FOR_EACH_MODE_FROM (mode_iter, mode)
214 mode = mode_iter.require ();
215 if (GET_MODE_SIZE (mode) > GET_MODE_SIZE (field_mode)
216 || TRULY_NOOP_TRUNCATION_MODES_P (insn->field_mode,
217 field_mode))
218 break;
219 get_extraction_insn (insn, pattern, type, mode);
221 return true;
224 return false;
227 /* Return true if an instruction exists to access a field of mode
228 FIELDMODE in a register structure that has STRUCT_BITS significant bits.
229 Describe the "best" such instruction in *INSN if so. PATTERN describes
230 the type of insertion or extraction we want to perform.
232 For an insertion, the number of significant structure bits includes
233 all bits of the target. For an extraction, it need only include the
234 most significant bit of the field. Larger widths are acceptable
235 in both cases. */
237 bool
238 get_best_reg_extraction_insn (extraction_insn *insn,
239 enum extraction_pattern pattern,
240 unsigned HOST_WIDE_INT struct_bits,
241 machine_mode field_mode)
243 return get_best_extraction_insn (insn, pattern, ET_reg, struct_bits,
244 field_mode);
247 /* Return true if an instruction exists to access a field of BITSIZE
248 bits starting BITNUM bits into a memory structure. Describe the
249 "best" such instruction in *INSN if so. PATTERN describes the type
250 of insertion or extraction we want to perform and FIELDMODE is the
251 natural mode of the extracted field.
253 The instructions considered here only access bytes that overlap
254 the bitfield; they do not touch any surrounding bytes. */
256 bool
257 get_best_mem_extraction_insn (extraction_insn *insn,
258 enum extraction_pattern pattern,
259 HOST_WIDE_INT bitsize, HOST_WIDE_INT bitnum,
260 machine_mode field_mode)
262 unsigned HOST_WIDE_INT struct_bits = (bitnum % BITS_PER_UNIT
263 + bitsize
264 + BITS_PER_UNIT - 1);
265 struct_bits -= struct_bits % BITS_PER_UNIT;
266 return get_best_extraction_insn (insn, pattern, ET_unaligned_mem,
267 struct_bits, field_mode);
270 /* Return the insn code used to extend FROM_MODE to TO_MODE.
271 UNSIGNEDP specifies zero-extension instead of sign-extension. If
272 no such operation exists, CODE_FOR_nothing will be returned. */
274 enum insn_code
275 can_extend_p (machine_mode to_mode, machine_mode from_mode,
276 int unsignedp)
278 if (unsignedp < 0 && targetm.have_ptr_extend ())
279 return targetm.code_for_ptr_extend;
281 convert_optab tab = unsignedp ? zext_optab : sext_optab;
282 return convert_optab_handler (tab, to_mode, from_mode);
285 /* Return the insn code to convert fixed-point mode FIXMODE to floating-point
286 mode FLTMODE, or CODE_FOR_nothing if no such instruction exists.
287 UNSIGNEDP specifies whether FIXMODE is unsigned. */
289 enum insn_code
290 can_float_p (machine_mode fltmode, machine_mode fixmode,
291 int unsignedp)
293 convert_optab tab = unsignedp ? ufloat_optab : sfloat_optab;
294 return convert_optab_handler (tab, fltmode, fixmode);
297 /* Return the insn code to convert floating-point mode FLTMODE to fixed-point
298 mode FIXMODE, or CODE_FOR_nothing if no such instruction exists.
299 UNSIGNEDP specifies whether FIXMODE is unsigned.
301 On a successful return, set *TRUNCP_PTR to true if it is necessary to
302 output an explicit FTRUNC before the instruction. */
304 enum insn_code
305 can_fix_p (machine_mode fixmode, machine_mode fltmode,
306 int unsignedp, bool *truncp_ptr)
308 convert_optab tab;
309 enum insn_code icode;
311 tab = unsignedp ? ufixtrunc_optab : sfixtrunc_optab;
312 icode = convert_optab_handler (tab, fixmode, fltmode);
313 if (icode != CODE_FOR_nothing)
315 *truncp_ptr = false;
316 return icode;
319 /* FIXME: This requires a port to define both FIX and FTRUNC pattern
320 for this to work. We need to rework the fix* and ftrunc* patterns
321 and documentation. */
322 tab = unsignedp ? ufix_optab : sfix_optab;
323 icode = convert_optab_handler (tab, fixmode, fltmode);
324 if (icode != CODE_FOR_nothing
325 && optab_handler (ftrunc_optab, fltmode) != CODE_FOR_nothing)
327 *truncp_ptr = true;
328 return icode;
331 return CODE_FOR_nothing;
334 /* Return nonzero if a conditional move of mode MODE is supported.
336 This function is for combine so it can tell whether an insn that looks
337 like a conditional move is actually supported by the hardware. If we
338 guess wrong we lose a bit on optimization, but that's it. */
339 /* ??? sparc64 supports conditionally moving integers values based on fp
340 comparisons, and vice versa. How do we handle them? */
342 bool
343 can_conditionally_move_p (machine_mode mode)
345 return direct_optab_handler (movcc_optab, mode) != CODE_FOR_nothing;
348 /* Return true if VEC_PERM_EXPR of arbitrary input vectors can be
349 expanded using SIMD extensions of the CPU. SEL may be NULL, which
350 stands for an unknown constant. Note that additional permutations
351 representing whole-vector shifts may also be handled via the vec_shr
352 optab, but only where the second input vector is entirely constant
353 zeroes; this case is not dealt with here. */
355 bool
356 can_vec_perm_p (machine_mode mode, bool variable, vec_perm_indices *sel)
358 machine_mode qimode;
360 /* If the target doesn't implement a vector mode for the vector type,
361 then no operations are supported. */
362 if (!VECTOR_MODE_P (mode))
363 return false;
365 if (!variable)
367 if (direct_optab_handler (vec_perm_const_optab, mode) != CODE_FOR_nothing
368 && (sel == NULL
369 || targetm.vectorize.vec_perm_const_ok == NULL
370 || targetm.vectorize.vec_perm_const_ok (mode, *sel)))
371 return true;
374 if (direct_optab_handler (vec_perm_optab, mode) != CODE_FOR_nothing)
375 return true;
377 /* We allow fallback to a QI vector mode, and adjust the mask. */
378 if (GET_MODE_INNER (mode) == QImode
379 || !mode_for_vector (QImode, GET_MODE_SIZE (mode)).exists (&qimode)
380 || !VECTOR_MODE_P (qimode))
381 return false;
383 /* ??? For completeness, we ought to check the QImode version of
384 vec_perm_const_optab. But all users of this implicit lowering
385 feature implement the variable vec_perm_optab. */
386 if (direct_optab_handler (vec_perm_optab, qimode) == CODE_FOR_nothing)
387 return false;
389 /* In order to support the lowering of variable permutations,
390 we need to support shifts and adds. */
391 if (variable)
393 if (GET_MODE_UNIT_SIZE (mode) > 2
394 && optab_handler (ashl_optab, mode) == CODE_FOR_nothing
395 && optab_handler (vashl_optab, mode) == CODE_FOR_nothing)
396 return false;
397 if (optab_handler (add_optab, qimode) == CODE_FOR_nothing)
398 return false;
401 return true;
404 /* Find a widening optab even if it doesn't widen as much as we want.
405 E.g. if from_mode is HImode, and to_mode is DImode, and there is no
406 direct HI->SI insn, then return SI->DI, if that exists. */
408 enum insn_code
409 find_widening_optab_handler_and_mode (optab op, machine_mode to_mode,
410 machine_mode from_mode,
411 machine_mode *found_mode)
413 gcc_checking_assert (GET_MODE_CLASS (from_mode) == GET_MODE_CLASS (to_mode));
414 gcc_checking_assert (from_mode < to_mode);
415 FOR_EACH_MODE (from_mode, from_mode, to_mode)
417 enum insn_code handler = convert_optab_handler (op, to_mode, from_mode);
419 if (handler != CODE_FOR_nothing)
421 if (found_mode)
422 *found_mode = from_mode;
423 return handler;
427 return CODE_FOR_nothing;
430 /* Return non-zero if a highpart multiply is supported of can be synthisized.
431 For the benefit of expand_mult_highpart, the return value is 1 for direct,
432 2 for even/odd widening, and 3 for hi/lo widening. */
435 can_mult_highpart_p (machine_mode mode, bool uns_p)
437 optab op;
438 unsigned i, nunits;
440 op = uns_p ? umul_highpart_optab : smul_highpart_optab;
441 if (optab_handler (op, mode) != CODE_FOR_nothing)
442 return 1;
444 /* If the mode is an integral vector, synth from widening operations. */
445 if (GET_MODE_CLASS (mode) != MODE_VECTOR_INT)
446 return 0;
448 nunits = GET_MODE_NUNITS (mode);
450 op = uns_p ? vec_widen_umult_even_optab : vec_widen_smult_even_optab;
451 if (optab_handler (op, mode) != CODE_FOR_nothing)
453 op = uns_p ? vec_widen_umult_odd_optab : vec_widen_smult_odd_optab;
454 if (optab_handler (op, mode) != CODE_FOR_nothing)
456 auto_vec_perm_indices sel (nunits);
457 for (i = 0; i < nunits; ++i)
458 sel.quick_push (!BYTES_BIG_ENDIAN
459 + (i & ~1)
460 + ((i & 1) ? nunits : 0));
461 if (can_vec_perm_p (mode, false, &sel))
462 return 2;
466 op = uns_p ? vec_widen_umult_hi_optab : vec_widen_smult_hi_optab;
467 if (optab_handler (op, mode) != CODE_FOR_nothing)
469 op = uns_p ? vec_widen_umult_lo_optab : vec_widen_smult_lo_optab;
470 if (optab_handler (op, mode) != CODE_FOR_nothing)
472 auto_vec_perm_indices sel (nunits);
473 for (i = 0; i < nunits; ++i)
474 sel.quick_push (2 * i + (BYTES_BIG_ENDIAN ? 0 : 1));
475 if (can_vec_perm_p (mode, false, &sel))
476 return 3;
480 return 0;
483 /* Return true if target supports vector masked load/store for mode. */
485 bool
486 can_vec_mask_load_store_p (machine_mode mode,
487 machine_mode mask_mode,
488 bool is_load)
490 optab op = is_load ? maskload_optab : maskstore_optab;
491 machine_mode vmode;
492 unsigned int vector_sizes;
494 /* If mode is vector mode, check it directly. */
495 if (VECTOR_MODE_P (mode))
496 return convert_optab_handler (op, mode, mask_mode) != CODE_FOR_nothing;
498 /* Otherwise, return true if there is some vector mode with
499 the mask load/store supported. */
501 /* See if there is any chance the mask load or store might be
502 vectorized. If not, punt. */
503 scalar_mode smode;
504 if (!is_a <scalar_mode> (mode, &smode))
505 return false;
507 vmode = targetm.vectorize.preferred_simd_mode (smode);
508 if (!VECTOR_MODE_P (vmode))
509 return false;
511 if ((targetm.vectorize.get_mask_mode
512 (GET_MODE_NUNITS (vmode), GET_MODE_SIZE (vmode)).exists (&mask_mode))
513 && convert_optab_handler (op, vmode, mask_mode) != CODE_FOR_nothing)
514 return true;
516 vector_sizes = targetm.vectorize.autovectorize_vector_sizes ();
517 while (vector_sizes != 0)
519 unsigned int cur = 1 << floor_log2 (vector_sizes);
520 vector_sizes &= ~cur;
521 if (cur <= GET_MODE_SIZE (smode))
522 continue;
523 unsigned int nunits = cur / GET_MODE_SIZE (smode);
524 if (mode_for_vector (smode, nunits).exists (&vmode)
525 && VECTOR_MODE_P (vmode)
526 && targetm.vectorize.get_mask_mode (nunits, cur).exists (&mask_mode)
527 && convert_optab_handler (op, vmode, mask_mode) != CODE_FOR_nothing)
528 return true;
530 return false;
533 /* Return true if there is a compare_and_swap pattern. */
535 bool
536 can_compare_and_swap_p (machine_mode mode, bool allow_libcall)
538 enum insn_code icode;
540 /* Check for __atomic_compare_and_swap. */
541 icode = direct_optab_handler (atomic_compare_and_swap_optab, mode);
542 if (icode != CODE_FOR_nothing)
543 return true;
545 /* Check for __sync_compare_and_swap. */
546 icode = optab_handler (sync_compare_and_swap_optab, mode);
547 if (icode != CODE_FOR_nothing)
548 return true;
549 if (allow_libcall && optab_libfunc (sync_compare_and_swap_optab, mode))
550 return true;
552 /* No inline compare and swap. */
553 return false;
556 /* Return true if an atomic exchange can be performed. */
558 bool
559 can_atomic_exchange_p (machine_mode mode, bool allow_libcall)
561 enum insn_code icode;
563 /* Check for __atomic_exchange. */
564 icode = direct_optab_handler (atomic_exchange_optab, mode);
565 if (icode != CODE_FOR_nothing)
566 return true;
568 /* Don't check __sync_test_and_set, as on some platforms that
569 has reduced functionality. Targets that really do support
570 a proper exchange should simply be updated to the __atomics. */
572 return can_compare_and_swap_p (mode, allow_libcall);
575 /* Return true if an atomic load can be performed without falling back to
576 a compare-and-swap. */
578 bool
579 can_atomic_load_p (machine_mode mode)
581 enum insn_code icode;
583 /* Does the target supports the load directly? */
584 icode = direct_optab_handler (atomic_load_optab, mode);
585 if (icode != CODE_FOR_nothing)
586 return true;
588 /* If the size of the object is greater than word size on this target,
589 then we assume that a load will not be atomic. Also see
590 expand_atomic_load. */
591 return GET_MODE_PRECISION (mode) <= BITS_PER_WORD;
594 /* Determine whether "1 << x" is relatively cheap in word_mode. */
596 bool
597 lshift_cheap_p (bool speed_p)
599 /* FIXME: This should be made target dependent via this "this_target"
600 mechanism, similar to e.g. can_copy_init_p in gcse.c. */
601 static bool init[2] = { false, false };
602 static bool cheap[2] = { true, true };
604 /* If the targer has no lshift in word_mode, the operation will most
605 probably not be cheap. ??? Does GCC even work for such targets? */
606 if (optab_handler (ashl_optab, word_mode) == CODE_FOR_nothing)
607 return false;
609 if (!init[speed_p])
611 rtx reg = gen_raw_REG (word_mode, 10000);
612 int cost = set_src_cost (gen_rtx_ASHIFT (word_mode, const1_rtx, reg),
613 word_mode, speed_p);
614 cheap[speed_p] = cost < COSTS_N_INSNS (3);
615 init[speed_p] = true;
618 return cheap[speed_p];