111
|
1 /* IR-agnostic target query functions relating to optabs
|
|
2 Copyright (C) 1987-2017 Free Software Foundation, Inc.
|
|
3
|
|
4 This file is part of GCC.
|
|
5
|
|
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.
|
|
10
|
|
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.
|
|
15
|
|
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/>. */
|
|
19
|
|
20
|
|
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"
|
|
31
|
|
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
|
|
37
|
|
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. */
|
|
41
|
|
42 insn_code
|
|
43 convert_optab_handler (convert_optab optab, machine_mode to_mode,
|
|
44 machine_mode from_mode, optimization_type opt_type)
|
|
45 {
|
|
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;
|
|
51 }
|
|
52
|
|
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. */
|
|
56
|
|
57 insn_code
|
|
58 direct_optab_handler (convert_optab optab, machine_mode mode,
|
|
59 optimization_type opt_type)
|
|
60 {
|
|
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;
|
|
66 }
|
|
67
|
|
68 /* Enumerates the possible types of structure operand to an
|
|
69 extraction_insn. */
|
|
70 enum extraction_type { ET_unaligned_mem, ET_reg };
|
|
71
|
|
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). */
|
|
78
|
|
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)
|
|
85 {
|
|
86 const struct insn_data_d *data = &insn_data[icode];
|
|
87
|
|
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;
|
|
93
|
|
94 machine_mode field_mode = data->operand[field_op].mode;
|
|
95 if (field_mode == VOIDmode)
|
|
96 field_mode = word_mode;
|
|
97
|
|
98 machine_mode pos_mode = data->operand[struct_op + 2].mode;
|
|
99 if (pos_mode == VOIDmode)
|
|
100 pos_mode = word_mode;
|
|
101
|
|
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;
|
|
112 }
|
|
113
|
|
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.
|
|
116
|
|
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. */
|
|
120
|
|
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)
|
|
126 {
|
|
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;
|
|
131
|
|
132 const struct insn_data_d *data = &insn_data[icode];
|
|
133
|
|
134 machine_mode pos_mode = data->operand[pos_op].mode;
|
|
135 if (pos_mode == VOIDmode)
|
|
136 pos_mode = word_mode;
|
|
137
|
|
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;
|
|
146 }
|
|
147
|
|
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. */
|
|
151
|
|
152 static bool
|
|
153 get_extraction_insn (extraction_insn *insn,
|
|
154 enum extraction_pattern pattern,
|
|
155 enum extraction_type type,
|
|
156 machine_mode mode)
|
|
157 {
|
|
158 switch (pattern)
|
|
159 {
|
|
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);
|
|
167
|
|
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);
|
|
175
|
|
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);
|
|
183
|
|
184 default:
|
|
185 gcc_unreachable ();
|
|
186 }
|
|
187 }
|
|
188
|
|
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.
|
|
193
|
|
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. */
|
|
198
|
|
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)
|
|
205 {
|
|
206 opt_scalar_int_mode mode_iter;
|
|
207 FOR_EACH_MODE_FROM (mode_iter, smallest_int_mode_for_size (struct_bits))
|
|
208 {
|
|
209 scalar_int_mode mode = mode_iter.require ();
|
|
210 if (get_extraction_insn (insn, pattern, type, mode))
|
|
211 {
|
|
212 FOR_EACH_MODE_FROM (mode_iter, mode)
|
|
213 {
|
|
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);
|
|
220 }
|
|
221 return true;
|
|
222 }
|
|
223 }
|
|
224 return false;
|
|
225 }
|
|
226
|
|
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.
|
|
231
|
|
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. */
|
|
236
|
|
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)
|
|
242 {
|
|
243 return get_best_extraction_insn (insn, pattern, ET_reg, struct_bits,
|
|
244 field_mode);
|
|
245 }
|
|
246
|
|
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.
|
|
252
|
|
253 The instructions considered here only access bytes that overlap
|
|
254 the bitfield; they do not touch any surrounding bytes. */
|
|
255
|
|
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)
|
|
261 {
|
|
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);
|
|
268 }
|
|
269
|
|
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. */
|
|
273
|
|
274 enum insn_code
|
|
275 can_extend_p (machine_mode to_mode, machine_mode from_mode,
|
|
276 int unsignedp)
|
|
277 {
|
|
278 if (unsignedp < 0 && targetm.have_ptr_extend ())
|
|
279 return targetm.code_for_ptr_extend;
|
|
280
|
|
281 convert_optab tab = unsignedp ? zext_optab : sext_optab;
|
|
282 return convert_optab_handler (tab, to_mode, from_mode);
|
|
283 }
|
|
284
|
|
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. */
|
|
288
|
|
289 enum insn_code
|
|
290 can_float_p (machine_mode fltmode, machine_mode fixmode,
|
|
291 int unsignedp)
|
|
292 {
|
|
293 convert_optab tab = unsignedp ? ufloat_optab : sfloat_optab;
|
|
294 return convert_optab_handler (tab, fltmode, fixmode);
|
|
295 }
|
|
296
|
|
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.
|
|
300
|
|
301 On a successful return, set *TRUNCP_PTR to true if it is necessary to
|
|
302 output an explicit FTRUNC before the instruction. */
|
|
303
|
|
304 enum insn_code
|
|
305 can_fix_p (machine_mode fixmode, machine_mode fltmode,
|
|
306 int unsignedp, bool *truncp_ptr)
|
|
307 {
|
|
308 convert_optab tab;
|
|
309 enum insn_code icode;
|
|
310
|
|
311 tab = unsignedp ? ufixtrunc_optab : sfixtrunc_optab;
|
|
312 icode = convert_optab_handler (tab, fixmode, fltmode);
|
|
313 if (icode != CODE_FOR_nothing)
|
|
314 {
|
|
315 *truncp_ptr = false;
|
|
316 return icode;
|
|
317 }
|
|
318
|
|
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)
|
|
326 {
|
|
327 *truncp_ptr = true;
|
|
328 return icode;
|
|
329 }
|
|
330
|
|
331 return CODE_FOR_nothing;
|
|
332 }
|
|
333
|
|
334 /* Return nonzero if a conditional move of mode MODE is supported.
|
|
335
|
|
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? */
|
|
341
|
|
342 bool
|
|
343 can_conditionally_move_p (machine_mode mode)
|
|
344 {
|
|
345 return direct_optab_handler (movcc_optab, mode) != CODE_FOR_nothing;
|
|
346 }
|
|
347
|
|
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. */
|
|
354
|
|
355 bool
|
|
356 can_vec_perm_p (machine_mode mode, bool variable, vec_perm_indices *sel)
|
|
357 {
|
|
358 machine_mode qimode;
|
|
359
|
|
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;
|
|
364
|
|
365 if (!variable)
|
|
366 {
|
|
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;
|
|
372 }
|
|
373
|
|
374 if (direct_optab_handler (vec_perm_optab, mode) != CODE_FOR_nothing)
|
|
375 return true;
|
|
376
|
|
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;
|
|
382
|
|
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;
|
|
388
|
|
389 /* In order to support the lowering of variable permutations,
|
|
390 we need to support shifts and adds. */
|
|
391 if (variable)
|
|
392 {
|
|
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;
|
|
399 }
|
|
400
|
|
401 return true;
|
|
402 }
|
|
403
|
|
404 /* Like optab_handler, but for widening_operations that have a
|
|
405 TO_MODE and a FROM_MODE. */
|
|
406
|
|
407 enum insn_code
|
|
408 widening_optab_handler (optab op, machine_mode to_mode,
|
|
409 machine_mode from_mode)
|
|
410 {
|
|
411 unsigned scode = (op << 16) | to_mode;
|
|
412 if (to_mode != from_mode && from_mode != VOIDmode)
|
|
413 {
|
|
414 /* ??? Why does find_widening_optab_handler_and_mode attempt to
|
|
415 widen things that can't be widened? E.g. add_optab... */
|
|
416 if (op > LAST_CONV_OPTAB)
|
|
417 return CODE_FOR_nothing;
|
|
418 scode |= from_mode << 8;
|
|
419 }
|
|
420 return raw_optab_handler (scode);
|
|
421 }
|
|
422
|
|
423 /* Find a widening optab even if it doesn't widen as much as we want.
|
|
424 E.g. if from_mode is HImode, and to_mode is DImode, and there is no
|
|
425 direct HI->SI insn, then return SI->DI, if that exists.
|
|
426 If PERMIT_NON_WIDENING is non-zero then this can be used with
|
|
427 non-widening optabs also. */
|
|
428
|
|
429 enum insn_code
|
|
430 find_widening_optab_handler_and_mode (optab op, machine_mode to_mode,
|
|
431 machine_mode from_mode,
|
|
432 int permit_non_widening,
|
|
433 machine_mode *found_mode)
|
|
434 {
|
|
435 for (; (permit_non_widening || from_mode != to_mode)
|
|
436 && GET_MODE_SIZE (from_mode) <= GET_MODE_SIZE (to_mode)
|
|
437 && from_mode != VOIDmode;
|
|
438 from_mode = GET_MODE_WIDER_MODE (from_mode).else_void ())
|
|
439 {
|
|
440 enum insn_code handler = widening_optab_handler (op, to_mode,
|
|
441 from_mode);
|
|
442
|
|
443 if (handler != CODE_FOR_nothing)
|
|
444 {
|
|
445 if (found_mode)
|
|
446 *found_mode = from_mode;
|
|
447 return handler;
|
|
448 }
|
|
449 }
|
|
450
|
|
451 return CODE_FOR_nothing;
|
|
452 }
|
|
453
|
|
454 /* Return non-zero if a highpart multiply is supported of can be synthisized.
|
|
455 For the benefit of expand_mult_highpart, the return value is 1 for direct,
|
|
456 2 for even/odd widening, and 3 for hi/lo widening. */
|
|
457
|
|
458 int
|
|
459 can_mult_highpart_p (machine_mode mode, bool uns_p)
|
|
460 {
|
|
461 optab op;
|
|
462 unsigned i, nunits;
|
|
463
|
|
464 op = uns_p ? umul_highpart_optab : smul_highpart_optab;
|
|
465 if (optab_handler (op, mode) != CODE_FOR_nothing)
|
|
466 return 1;
|
|
467
|
|
468 /* If the mode is an integral vector, synth from widening operations. */
|
|
469 if (GET_MODE_CLASS (mode) != MODE_VECTOR_INT)
|
|
470 return 0;
|
|
471
|
|
472 nunits = GET_MODE_NUNITS (mode);
|
|
473
|
|
474 op = uns_p ? vec_widen_umult_even_optab : vec_widen_smult_even_optab;
|
|
475 if (optab_handler (op, mode) != CODE_FOR_nothing)
|
|
476 {
|
|
477 op = uns_p ? vec_widen_umult_odd_optab : vec_widen_smult_odd_optab;
|
|
478 if (optab_handler (op, mode) != CODE_FOR_nothing)
|
|
479 {
|
|
480 auto_vec_perm_indices sel (nunits);
|
|
481 for (i = 0; i < nunits; ++i)
|
|
482 sel.quick_push (!BYTES_BIG_ENDIAN
|
|
483 + (i & ~1)
|
|
484 + ((i & 1) ? nunits : 0));
|
|
485 if (can_vec_perm_p (mode, false, &sel))
|
|
486 return 2;
|
|
487 }
|
|
488 }
|
|
489
|
|
490 op = uns_p ? vec_widen_umult_hi_optab : vec_widen_smult_hi_optab;
|
|
491 if (optab_handler (op, mode) != CODE_FOR_nothing)
|
|
492 {
|
|
493 op = uns_p ? vec_widen_umult_lo_optab : vec_widen_smult_lo_optab;
|
|
494 if (optab_handler (op, mode) != CODE_FOR_nothing)
|
|
495 {
|
|
496 auto_vec_perm_indices sel (nunits);
|
|
497 for (i = 0; i < nunits; ++i)
|
|
498 sel.quick_push (2 * i + (BYTES_BIG_ENDIAN ? 0 : 1));
|
|
499 if (can_vec_perm_p (mode, false, &sel))
|
|
500 return 3;
|
|
501 }
|
|
502 }
|
|
503
|
|
504 return 0;
|
|
505 }
|
|
506
|
|
507 /* Return true if target supports vector masked load/store for mode. */
|
|
508
|
|
509 bool
|
|
510 can_vec_mask_load_store_p (machine_mode mode,
|
|
511 machine_mode mask_mode,
|
|
512 bool is_load)
|
|
513 {
|
|
514 optab op = is_load ? maskload_optab : maskstore_optab;
|
|
515 machine_mode vmode;
|
|
516 unsigned int vector_sizes;
|
|
517
|
|
518 /* If mode is vector mode, check it directly. */
|
|
519 if (VECTOR_MODE_P (mode))
|
|
520 return convert_optab_handler (op, mode, mask_mode) != CODE_FOR_nothing;
|
|
521
|
|
522 /* Otherwise, return true if there is some vector mode with
|
|
523 the mask load/store supported. */
|
|
524
|
|
525 /* See if there is any chance the mask load or store might be
|
|
526 vectorized. If not, punt. */
|
|
527 scalar_mode smode;
|
|
528 if (!is_a <scalar_mode> (mode, &smode))
|
|
529 return false;
|
|
530
|
|
531 vmode = targetm.vectorize.preferred_simd_mode (smode);
|
|
532 if (!VECTOR_MODE_P (vmode))
|
|
533 return false;
|
|
534
|
|
535 if ((targetm.vectorize.get_mask_mode
|
|
536 (GET_MODE_NUNITS (vmode), GET_MODE_SIZE (vmode)).exists (&mask_mode))
|
|
537 && convert_optab_handler (op, vmode, mask_mode) != CODE_FOR_nothing)
|
|
538 return true;
|
|
539
|
|
540 vector_sizes = targetm.vectorize.autovectorize_vector_sizes ();
|
|
541 while (vector_sizes != 0)
|
|
542 {
|
|
543 unsigned int cur = 1 << floor_log2 (vector_sizes);
|
|
544 vector_sizes &= ~cur;
|
|
545 if (cur <= GET_MODE_SIZE (smode))
|
|
546 continue;
|
|
547 unsigned int nunits = cur / GET_MODE_SIZE (smode);
|
|
548 if (mode_for_vector (smode, nunits).exists (&vmode)
|
|
549 && VECTOR_MODE_P (vmode)
|
|
550 && targetm.vectorize.get_mask_mode (nunits, cur).exists (&mask_mode)
|
|
551 && convert_optab_handler (op, vmode, mask_mode) != CODE_FOR_nothing)
|
|
552 return true;
|
|
553 }
|
|
554 return false;
|
|
555 }
|
|
556
|
|
557 /* Return true if there is a compare_and_swap pattern. */
|
|
558
|
|
559 bool
|
|
560 can_compare_and_swap_p (machine_mode mode, bool allow_libcall)
|
|
561 {
|
|
562 enum insn_code icode;
|
|
563
|
|
564 /* Check for __atomic_compare_and_swap. */
|
|
565 icode = direct_optab_handler (atomic_compare_and_swap_optab, mode);
|
|
566 if (icode != CODE_FOR_nothing)
|
|
567 return true;
|
|
568
|
|
569 /* Check for __sync_compare_and_swap. */
|
|
570 icode = optab_handler (sync_compare_and_swap_optab, mode);
|
|
571 if (icode != CODE_FOR_nothing)
|
|
572 return true;
|
|
573 if (allow_libcall && optab_libfunc (sync_compare_and_swap_optab, mode))
|
|
574 return true;
|
|
575
|
|
576 /* No inline compare and swap. */
|
|
577 return false;
|
|
578 }
|
|
579
|
|
580 /* Return true if an atomic exchange can be performed. */
|
|
581
|
|
582 bool
|
|
583 can_atomic_exchange_p (machine_mode mode, bool allow_libcall)
|
|
584 {
|
|
585 enum insn_code icode;
|
|
586
|
|
587 /* Check for __atomic_exchange. */
|
|
588 icode = direct_optab_handler (atomic_exchange_optab, mode);
|
|
589 if (icode != CODE_FOR_nothing)
|
|
590 return true;
|
|
591
|
|
592 /* Don't check __sync_test_and_set, as on some platforms that
|
|
593 has reduced functionality. Targets that really do support
|
|
594 a proper exchange should simply be updated to the __atomics. */
|
|
595
|
|
596 return can_compare_and_swap_p (mode, allow_libcall);
|
|
597 }
|
|
598
|
|
599 /* Return true if an atomic load can be performed without falling back to
|
|
600 a compare-and-swap. */
|
|
601
|
|
602 bool
|
|
603 can_atomic_load_p (machine_mode mode)
|
|
604 {
|
|
605 enum insn_code icode;
|
|
606
|
|
607 /* Does the target supports the load directly? */
|
|
608 icode = direct_optab_handler (atomic_load_optab, mode);
|
|
609 if (icode != CODE_FOR_nothing)
|
|
610 return true;
|
|
611
|
|
612 /* If the size of the object is greater than word size on this target,
|
|
613 then we assume that a load will not be atomic. Also see
|
|
614 expand_atomic_load. */
|
|
615 return GET_MODE_PRECISION (mode) <= BITS_PER_WORD;
|
|
616 }
|
|
617
|
|
618 /* Determine whether "1 << x" is relatively cheap in word_mode. */
|
|
619
|
|
620 bool
|
|
621 lshift_cheap_p (bool speed_p)
|
|
622 {
|
|
623 /* FIXME: This should be made target dependent via this "this_target"
|
|
624 mechanism, similar to e.g. can_copy_init_p in gcse.c. */
|
|
625 static bool init[2] = { false, false };
|
|
626 static bool cheap[2] = { true, true };
|
|
627
|
|
628 /* If the targer has no lshift in word_mode, the operation will most
|
|
629 probably not be cheap. ??? Does GCC even work for such targets? */
|
|
630 if (optab_handler (ashl_optab, word_mode) == CODE_FOR_nothing)
|
|
631 return false;
|
|
632
|
|
633 if (!init[speed_p])
|
|
634 {
|
|
635 rtx reg = gen_raw_REG (word_mode, 10000);
|
|
636 int cost = set_src_cost (gen_rtx_ASHIFT (word_mode, const1_rtx, reg),
|
|
637 word_mode, speed_p);
|
|
638 cheap[speed_p] = cost < COSTS_N_INSNS (3);
|
|
639 init[speed_p] = true;
|
|
640 }
|
|
641
|
|
642 return cheap[speed_p];
|
|
643 }
|