Branch data Line data Source code
1 : : /* Analysis Utilities for Loop Vectorization.
2 : : Copyright (C) 2006-2025 Free Software Foundation, Inc.
3 : : Contributed by Dorit Nuzman <dorit@il.ibm.com>
4 : :
5 : : This file is part of GCC.
6 : :
7 : : GCC is free software; you can redistribute it and/or modify it under
8 : : the terms of the GNU General Public License as published by the Free
9 : : Software Foundation; either version 3, or (at your option) any later
10 : : version.
11 : :
12 : : GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 : : WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 : : FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 : : for more details.
16 : :
17 : : You should have received a copy of the GNU General Public License
18 : : along with GCC; see the file COPYING3. If not see
19 : : <http://www.gnu.org/licenses/>. */
20 : :
21 : : #include "config.h"
22 : : #include "system.h"
23 : : #include "coretypes.h"
24 : : #include "backend.h"
25 : : #include "rtl.h"
26 : : #include "tree.h"
27 : : #include "gimple.h"
28 : : #include "gimple-iterator.h"
29 : : #include "gimple-fold.h"
30 : : #include "ssa.h"
31 : : #include "expmed.h"
32 : : #include "optabs-tree.h"
33 : : #include "insn-config.h"
34 : : #include "recog.h" /* FIXME: for insn_data */
35 : : #include "fold-const.h"
36 : : #include "stor-layout.h"
37 : : #include "tree-eh.h"
38 : : #include "gimplify.h"
39 : : #include "gimple-iterator.h"
40 : : #include "gimple-fold.h"
41 : : #include "gimplify-me.h"
42 : : #include "cfgloop.h"
43 : : #include "tree-vectorizer.h"
44 : : #include "dumpfile.h"
45 : : #include "builtins.h"
46 : : #include "internal-fn.h"
47 : : #include "case-cfn-macros.h"
48 : : #include "fold-const-call.h"
49 : : #include "attribs.h"
50 : : #include "cgraph.h"
51 : : #include "omp-simd-clone.h"
52 : : #include "predict.h"
53 : : #include "tree-vector-builder.h"
54 : : #include "tree-ssa-loop-ivopts.h"
55 : : #include "vec-perm-indices.h"
56 : : #include "gimple-range.h"
57 : : #include "alias.h"
58 : :
59 : :
60 : : /* TODO: Note the vectorizer still builds COND_EXPRs with GENERIC compares
61 : : in the first operand. Disentangling this is future work, the
62 : : IL is properly transfered to VEC_COND_EXPRs with separate compares. */
63 : :
64 : :
65 : : /* Return true if we have a useful VR_RANGE range for VAR, storing it
66 : : in *MIN_VALUE and *MAX_VALUE if so. Note the range in the dump files. */
67 : :
68 : : bool
69 : 11451512 : vect_get_range_info (tree var, wide_int *min_value, wide_int *max_value)
70 : : {
71 : 11451512 : int_range_max vr;
72 : 11451512 : tree vr_min, vr_max;
73 : 22903024 : get_range_query (cfun)->range_of_expr (vr, var);
74 : 11451512 : if (vr.undefined_p ())
75 : 67 : vr.set_varying (TREE_TYPE (var));
76 : 11451512 : value_range_kind vr_type = get_legacy_range (vr, vr_min, vr_max);
77 : 11451512 : *min_value = wi::to_wide (vr_min);
78 : 11451512 : *max_value = wi::to_wide (vr_max);
79 : 11451512 : wide_int nonzero = get_nonzero_bits (var);
80 : 11451512 : signop sgn = TYPE_SIGN (TREE_TYPE (var));
81 : 11451512 : if (intersect_range_with_nonzero_bits (vr_type, min_value, max_value,
82 : : nonzero, sgn) == VR_RANGE)
83 : : {
84 : 5438268 : if (dump_enabled_p ())
85 : : {
86 : 72686 : dump_generic_expr_loc (MSG_NOTE, vect_location, TDF_SLIM, var);
87 : 72686 : dump_printf (MSG_NOTE, " has range [");
88 : 72686 : dump_hex (MSG_NOTE, *min_value);
89 : 72686 : dump_printf (MSG_NOTE, ", ");
90 : 72686 : dump_hex (MSG_NOTE, *max_value);
91 : 72686 : dump_printf (MSG_NOTE, "]\n");
92 : : }
93 : 5438268 : return true;
94 : : }
95 : : else
96 : : {
97 : 6013244 : if (dump_enabled_p ())
98 : : {
99 : 85187 : dump_generic_expr_loc (MSG_NOTE, vect_location, TDF_SLIM, var);
100 : 85187 : dump_printf (MSG_NOTE, " has no range info\n");
101 : : }
102 : 6013244 : return false;
103 : : }
104 : 11451512 : }
105 : :
106 : : /* Report that we've found an instance of pattern PATTERN in
107 : : statement STMT. */
108 : :
109 : : static void
110 : 1057984 : vect_pattern_detected (const char *name, gimple *stmt)
111 : : {
112 : 1057984 : if (dump_enabled_p ())
113 : 25155 : dump_printf_loc (MSG_NOTE, vect_location, "%s: detected: %G", name, stmt);
114 : 1057984 : }
115 : :
116 : : /* Associate pattern statement PATTERN_STMT with ORIG_STMT_INFO and
117 : : return the pattern statement's stmt_vec_info. Set its vector type to
118 : : VECTYPE if it doesn't have one already. */
119 : :
120 : : static stmt_vec_info
121 : 2004514 : vect_init_pattern_stmt (vec_info *vinfo, gimple *pattern_stmt,
122 : : stmt_vec_info orig_stmt_info, tree vectype)
123 : : {
124 : 2004514 : stmt_vec_info pattern_stmt_info = vinfo->lookup_stmt (pattern_stmt);
125 : 2004514 : if (pattern_stmt_info == NULL)
126 : 1202400 : pattern_stmt_info = vinfo->add_stmt (pattern_stmt);
127 : 2004514 : gimple_set_bb (pattern_stmt, gimple_bb (orig_stmt_info->stmt));
128 : :
129 : 2004514 : pattern_stmt_info->pattern_stmt_p = true;
130 : 2004514 : STMT_VINFO_RELATED_STMT (pattern_stmt_info) = orig_stmt_info;
131 : 2004514 : STMT_VINFO_DEF_TYPE (pattern_stmt_info)
132 : 2004514 : = STMT_VINFO_DEF_TYPE (orig_stmt_info);
133 : 2004514 : STMT_VINFO_TYPE (pattern_stmt_info) = STMT_VINFO_TYPE (orig_stmt_info);
134 : 2004514 : if (!STMT_VINFO_VECTYPE (pattern_stmt_info))
135 : : {
136 : 2134797 : gcc_assert (!vectype
137 : : || is_a <gcond *> (pattern_stmt)
138 : : || (VECTOR_BOOLEAN_TYPE_P (vectype)
139 : : == vect_use_mask_type_p (orig_stmt_info)));
140 : 1210737 : STMT_VINFO_VECTYPE (pattern_stmt_info) = vectype;
141 : 1210737 : pattern_stmt_info->mask_precision = orig_stmt_info->mask_precision;
142 : : }
143 : 2004514 : return pattern_stmt_info;
144 : : }
145 : :
146 : : /* Set the pattern statement of ORIG_STMT_INFO to PATTERN_STMT.
147 : : Also set the vector type of PATTERN_STMT to VECTYPE, if it doesn't
148 : : have one already. */
149 : :
150 : : static void
151 : 844056 : vect_set_pattern_stmt (vec_info *vinfo, gimple *pattern_stmt,
152 : : stmt_vec_info orig_stmt_info, tree vectype)
153 : : {
154 : 844056 : STMT_VINFO_IN_PATTERN_P (orig_stmt_info) = true;
155 : 844056 : STMT_VINFO_RELATED_STMT (orig_stmt_info)
156 : 0 : = vect_init_pattern_stmt (vinfo, pattern_stmt, orig_stmt_info, vectype);
157 : 814049 : }
158 : :
159 : : /* Add NEW_STMT to STMT_INFO's pattern definition statements. If VECTYPE
160 : : is nonnull, record that NEW_STMT's vector type is VECTYPE, which might
161 : : be different from the vector type of the final pattern statement.
162 : : If VECTYPE is a mask type, SCALAR_TYPE_FOR_MASK is the scalar type
163 : : from which it was derived. */
164 : :
165 : : static inline void
166 : 1117030 : append_pattern_def_seq (vec_info *vinfo,
167 : : stmt_vec_info stmt_info, gimple *new_stmt,
168 : : tree vectype = NULL_TREE,
169 : : tree scalar_type_for_mask = NULL_TREE)
170 : : {
171 : 1816932 : gcc_assert (!scalar_type_for_mask
172 : : == (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)));
173 : 1117030 : if (vectype)
174 : : {
175 : 793777 : stmt_vec_info new_stmt_info = vinfo->add_stmt (new_stmt);
176 : 793777 : STMT_VINFO_VECTYPE (new_stmt_info) = vectype;
177 : 793777 : if (scalar_type_for_mask)
178 : 417128 : new_stmt_info->mask_precision
179 : 834256 : = GET_MODE_BITSIZE (SCALAR_TYPE_MODE (scalar_type_for_mask));
180 : : }
181 : 1117030 : gimple_seq_add_stmt_without_update (&STMT_VINFO_PATTERN_DEF_SEQ (stmt_info),
182 : : new_stmt);
183 : 1117030 : }
184 : :
185 : :
186 : : /* Add NEW_STMT to VINFO's invariant pattern definition statements. These
187 : : statements are not vectorized but are materialized as scalar in the loop
188 : : preheader. */
189 : :
190 : : static inline void
191 : 1229 : append_inv_pattern_def_seq (vec_info *vinfo, gimple *new_stmt)
192 : : {
193 : 1229 : gimple_seq_add_stmt_without_update (&vinfo->inv_pattern_def_seq, new_stmt);
194 : : }
195 : :
196 : : /* The caller wants to perform new operations on vect_external variable
197 : : VAR, so that the result of the operations would also be vect_external.
198 : : Return the edge on which the operations can be performed, if one exists.
199 : : Return null if the operations should instead be treated as part of
200 : : the pattern that needs them. */
201 : :
202 : : static edge
203 : 8506 : vect_get_external_def_edge (vec_info *vinfo, tree var)
204 : : {
205 : 8506 : edge e = NULL;
206 : 8506 : if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
207 : : {
208 : 525 : e = loop_preheader_edge (loop_vinfo->loop);
209 : 525 : if (!SSA_NAME_IS_DEFAULT_DEF (var))
210 : : {
211 : 426 : basic_block bb = gimple_bb (SSA_NAME_DEF_STMT (var));
212 : 426 : if (bb == NULL
213 : 426 : || !dominated_by_p (CDI_DOMINATORS, e->dest, bb))
214 : : e = NULL;
215 : : }
216 : : }
217 : 8506 : return e;
218 : : }
219 : :
220 : : /* Return true if the target supports a vector version of CODE,
221 : : where CODE is known to map to a direct optab with the given SUBTYPE.
222 : : ITYPE specifies the type of (some of) the scalar inputs and OTYPE
223 : : specifies the type of the scalar result.
224 : :
225 : : If CODE allows the inputs and outputs to have different type
226 : : (such as for WIDEN_SUM_EXPR), it is the input mode rather
227 : : than the output mode that determines the appropriate target pattern.
228 : : Operand 0 of the target pattern then specifies the mode that the output
229 : : must have.
230 : :
231 : : When returning true, set *VECOTYPE_OUT to the vector version of OTYPE.
232 : : Also set *VECITYPE_OUT to the vector version of ITYPE if VECITYPE_OUT
233 : : is nonnull. */
234 : :
235 : : static bool
236 : 2137 : vect_supportable_direct_optab_p (vec_info *vinfo, tree otype, tree_code code,
237 : : tree itype, tree *vecotype_out,
238 : : tree *vecitype_out = NULL,
239 : : enum optab_subtype subtype = optab_default)
240 : : {
241 : 2137 : tree vecitype = get_vectype_for_scalar_type (vinfo, itype);
242 : 2137 : if (!vecitype)
243 : : return false;
244 : :
245 : 2137 : tree vecotype = get_vectype_for_scalar_type (vinfo, otype);
246 : 2137 : if (!vecotype)
247 : : return false;
248 : :
249 : 2030 : optab optab = optab_for_tree_code (code, vecitype, subtype);
250 : 2030 : if (!optab)
251 : : return false;
252 : :
253 : 2030 : insn_code icode = optab_handler (optab, TYPE_MODE (vecitype));
254 : 2030 : if (icode == CODE_FOR_nothing
255 : 2030 : || insn_data[icode].operand[0].mode != TYPE_MODE (vecotype))
256 : 1804 : return false;
257 : :
258 : 226 : *vecotype_out = vecotype;
259 : 226 : if (vecitype_out)
260 : 226 : *vecitype_out = vecitype;
261 : : return true;
262 : : }
263 : :
264 : : /* Return true if the target supports a vector version of CODE,
265 : : where CODE is known to map to a conversion optab with the given SUBTYPE.
266 : : ITYPE specifies the type of (some of) the scalar inputs and OTYPE
267 : : specifies the type of the scalar result.
268 : :
269 : : When returning true, set *VECOTYPE_OUT to the vector version of OTYPE.
270 : : Also set *VECITYPE_OUT to the vector version of ITYPE if VECITYPE_OUT
271 : : is nonnull. */
272 : :
273 : : static bool
274 : 966 : vect_supportable_conv_optab_p (vec_info *vinfo, tree otype, tree_code code,
275 : : tree itype, tree *vecotype_out,
276 : : tree *vecitype_out = NULL,
277 : : enum optab_subtype subtype = optab_default)
278 : : {
279 : 966 : tree vecitype = get_vectype_for_scalar_type (vinfo, itype);
280 : 966 : tree vecotype = get_vectype_for_scalar_type (vinfo, otype);
281 : 966 : if (!vecitype || !vecotype)
282 : : return false;
283 : :
284 : 849 : if (!directly_supported_p (code, vecotype, vecitype, subtype))
285 : : return false;
286 : :
287 : 423 : *vecotype_out = vecotype;
288 : 423 : if (vecitype_out)
289 : 423 : *vecitype_out = vecitype;
290 : : return true;
291 : : }
292 : :
293 : : /* Round bit precision PRECISION up to a full element. */
294 : :
295 : : static unsigned int
296 : 2794893 : vect_element_precision (unsigned int precision)
297 : : {
298 : 0 : precision = 1 << ceil_log2 (precision);
299 : 4079744 : return MAX (precision, BITS_PER_UNIT);
300 : : }
301 : :
302 : : /* If OP is defined by a statement that's being considered for vectorization,
303 : : return information about that statement, otherwise return NULL. */
304 : :
305 : : static stmt_vec_info
306 : 1552943 : vect_get_internal_def (vec_info *vinfo, tree op)
307 : : {
308 : 1552943 : stmt_vec_info def_stmt_info = vinfo->lookup_def (op);
309 : 1552943 : if (def_stmt_info
310 : 1494694 : && STMT_VINFO_DEF_TYPE (def_stmt_info) == vect_internal_def)
311 : 1484424 : return vect_stmt_to_vectorize (def_stmt_info);
312 : : return NULL;
313 : : }
314 : :
315 : : /* Holds information about an input operand after some sign changes
316 : : and type promotions have been peeled away. */
317 : : class vect_unpromoted_value {
318 : : public:
319 : : vect_unpromoted_value ();
320 : :
321 : : void set_op (tree, vect_def_type, stmt_vec_info = NULL);
322 : :
323 : : /* The value obtained after peeling away zero or more casts. */
324 : : tree op;
325 : :
326 : : /* The type of OP. */
327 : : tree type;
328 : :
329 : : /* The definition type of OP. */
330 : : vect_def_type dt;
331 : :
332 : : /* If OP is the result of peeling at least one cast, and if the cast
333 : : of OP itself is a vectorizable statement, CASTER identifies that
334 : : statement, otherwise it is null. */
335 : : stmt_vec_info caster;
336 : : };
337 : :
338 : 283257979 : inline vect_unpromoted_value::vect_unpromoted_value ()
339 : 283257979 : : op (NULL_TREE),
340 : 283257979 : type (NULL_TREE),
341 : 283257979 : dt (vect_uninitialized_def),
342 : 2559239 : caster (NULL)
343 : : {
344 : : }
345 : :
346 : : /* Set the operand to OP_IN, its definition type to DT_IN, and the
347 : : statement that casts it to CASTER_IN. */
348 : :
349 : : inline void
350 : 10157635 : vect_unpromoted_value::set_op (tree op_in, vect_def_type dt_in,
351 : : stmt_vec_info caster_in)
352 : : {
353 : 10157635 : op = op_in;
354 : 10157635 : type = TREE_TYPE (op);
355 : 10157635 : dt = dt_in;
356 : 10157635 : caster = caster_in;
357 : 10157635 : }
358 : :
359 : : /* If OP is a vectorizable SSA name, strip a sequence of integer conversions
360 : : to reach some vectorizable inner operand OP', continuing as long as it
361 : : is possible to convert OP' back to OP using a possible sign change
362 : : followed by a possible promotion P. Return this OP', or null if OP is
363 : : not a vectorizable SSA name. If there is a promotion P, describe its
364 : : input in UNPROM, otherwise describe OP' in UNPROM. If SINGLE_USE_P
365 : : is nonnull, set *SINGLE_USE_P to false if any of the SSA names involved
366 : : have more than one user.
367 : :
368 : : A successful return means that it is possible to go from OP' to OP
369 : : via UNPROM. The cast from OP' to UNPROM is at most a sign change,
370 : : whereas the cast from UNPROM to OP might be a promotion, a sign
371 : : change, or a nop.
372 : :
373 : : E.g. say we have:
374 : :
375 : : signed short *ptr = ...;
376 : : signed short C = *ptr;
377 : : unsigned short B = (unsigned short) C; // sign change
378 : : signed int A = (signed int) B; // unsigned promotion
379 : : ...possible other uses of A...
380 : : unsigned int OP = (unsigned int) A; // sign change
381 : :
382 : : In this case it's possible to go directly from C to OP using:
383 : :
384 : : OP = (unsigned int) (unsigned short) C;
385 : : +------------+ +--------------+
386 : : promotion sign change
387 : :
388 : : so OP' would be C. The input to the promotion is B, so UNPROM
389 : : would describe B. */
390 : :
391 : : static tree
392 : 7583548 : vect_look_through_possible_promotion (vec_info *vinfo, tree op,
393 : : vect_unpromoted_value *unprom,
394 : : bool *single_use_p = NULL)
395 : : {
396 : 7583548 : tree op_type = TREE_TYPE (op);
397 : 7583548 : if (!INTEGRAL_TYPE_P (op_type))
398 : : return NULL_TREE;
399 : :
400 : 7559162 : tree res = NULL_TREE;
401 : 7559162 : unsigned int orig_precision = TYPE_PRECISION (op_type);
402 : 7559162 : unsigned int min_precision = orig_precision;
403 : 7559162 : stmt_vec_info caster = NULL;
404 : 9115304 : while (TREE_CODE (op) == SSA_NAME && INTEGRAL_TYPE_P (op_type))
405 : : {
406 : : /* See whether OP is simple enough to vectorize. */
407 : 8913211 : stmt_vec_info def_stmt_info;
408 : 8913211 : gimple *def_stmt;
409 : 8913211 : vect_def_type dt;
410 : 8913211 : if (!vect_is_simple_use (op, vinfo, &dt, &def_stmt_info, &def_stmt))
411 : : break;
412 : :
413 : : /* If OP is the input of a demotion, skip over it to see whether
414 : : OP is itself the result of a promotion. If so, the combined
415 : : effect of the promotion and the demotion might fit the required
416 : : pattern, otherwise neither operation fits.
417 : :
418 : : This copes with cases such as the result of an arithmetic
419 : : operation being truncated before being stored, and where that
420 : : arithmetic operation has been recognized as an over-widened one. */
421 : 8908243 : if (TYPE_PRECISION (op_type) <= min_precision)
422 : : {
423 : : /* Use OP as the UNPROM described above if we haven't yet
424 : : found a promotion, or if using the new input preserves the
425 : : sign of the previous promotion. */
426 : 8764035 : if (!res
427 : 1312607 : || TYPE_PRECISION (unprom->type) == orig_precision
428 : 40047 : || TYPE_SIGN (unprom->type) == TYPE_SIGN (op_type)
429 : 8801597 : || (TYPE_UNSIGNED (op_type)
430 : 26158 : && TYPE_PRECISION (op_type) < TYPE_PRECISION (unprom->type)))
431 : : {
432 : 8727592 : unprom->set_op (op, dt, caster);
433 : 8727592 : min_precision = TYPE_PRECISION (op_type);
434 : : }
435 : : /* Stop if we've already seen a promotion and if this
436 : : conversion does more than change the sign. */
437 : 36443 : else if (TYPE_PRECISION (op_type)
438 : 36443 : != TYPE_PRECISION (unprom->type))
439 : : break;
440 : :
441 : : /* The sequence now extends to OP. */
442 : : res = op;
443 : : }
444 : :
445 : : /* See whether OP is defined by a cast. Record it as CASTER if
446 : : the cast is potentially vectorizable. */
447 : 8908202 : if (!def_stmt)
448 : : break;
449 : 8714321 : caster = def_stmt_info;
450 : :
451 : : /* Ignore pattern statements, since we don't link uses for them. */
452 : 8714321 : if (caster
453 : 8714321 : && single_use_p
454 : 1623131 : && !STMT_VINFO_RELATED_STMT (caster)
455 : 10174357 : && !has_single_use (res))
456 : 962629 : *single_use_p = false;
457 : :
458 : 16071390 : gassign *assign = dyn_cast <gassign *> (def_stmt);
459 : 5686695 : if (!assign || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (def_stmt)))
460 : : break;
461 : :
462 : : /* Continue with the input to the cast. */
463 : 1556142 : op = gimple_assign_rhs1 (def_stmt);
464 : 1556142 : op_type = TREE_TYPE (op);
465 : : }
466 : : return res;
467 : : }
468 : :
469 : : /* OP is an integer operand to an operation that returns TYPE, and we
470 : : want to treat the operation as a widening one. So far we can treat
471 : : it as widening from *COMMON_TYPE.
472 : :
473 : : Return true if OP is suitable for such a widening operation,
474 : : either widening from *COMMON_TYPE or from some supertype of it.
475 : : Update *COMMON_TYPE to the supertype in the latter case.
476 : :
477 : : SHIFT_P is true if OP is a shift amount. */
478 : :
479 : : static bool
480 : 284992 : vect_joust_widened_integer (tree type, bool shift_p, tree op,
481 : : tree *common_type)
482 : : {
483 : : /* Calculate the minimum precision required by OP, without changing
484 : : the sign of either operand. */
485 : 284992 : unsigned int precision;
486 : 284992 : if (shift_p)
487 : : {
488 : 13328 : if (!wi::leu_p (wi::to_widest (op), TYPE_PRECISION (type) / 2))
489 : : return false;
490 : 10789 : precision = TREE_INT_CST_LOW (op);
491 : : }
492 : : else
493 : : {
494 : 271664 : precision = wi::min_precision (wi::to_widest (op),
495 : 271664 : TYPE_SIGN (*common_type));
496 : 271664 : if (precision * 2 > TYPE_PRECISION (type))
497 : : return false;
498 : : }
499 : :
500 : : /* If OP requires a wider type, switch to that type. The checks
501 : : above ensure that this is still narrower than the result. */
502 : 267859 : precision = vect_element_precision (precision);
503 : 267859 : if (TYPE_PRECISION (*common_type) < precision)
504 : 7423 : *common_type = build_nonstandard_integer_type
505 : 7423 : (precision, TYPE_UNSIGNED (*common_type));
506 : : return true;
507 : : }
508 : :
509 : : /* Return true if the common supertype of NEW_TYPE and *COMMON_TYPE
510 : : is narrower than type, storing the supertype in *COMMON_TYPE if so. */
511 : :
512 : : static bool
513 : 40543 : vect_joust_widened_type (tree type, tree new_type, tree *common_type)
514 : : {
515 : 40543 : if (types_compatible_p (*common_type, new_type))
516 : : return true;
517 : :
518 : : /* See if *COMMON_TYPE can hold all values of NEW_TYPE. */
519 : 7117 : if ((TYPE_PRECISION (new_type) < TYPE_PRECISION (*common_type))
520 : 7117 : && (TYPE_UNSIGNED (new_type) || !TYPE_UNSIGNED (*common_type)))
521 : : return true;
522 : :
523 : : /* See if NEW_TYPE can hold all values of *COMMON_TYPE. */
524 : 6593 : if (TYPE_PRECISION (*common_type) < TYPE_PRECISION (new_type)
525 : 6593 : && (TYPE_UNSIGNED (*common_type) || !TYPE_UNSIGNED (new_type)))
526 : : {
527 : 363 : *common_type = new_type;
528 : 363 : return true;
529 : : }
530 : :
531 : : /* We have mismatched signs, with the signed type being
532 : : no wider than the unsigned type. In this case we need
533 : : a wider signed type. */
534 : 6230 : unsigned int precision = MAX (TYPE_PRECISION (*common_type),
535 : : TYPE_PRECISION (new_type));
536 : 6230 : precision *= 2;
537 : :
538 : 6230 : if (precision * 2 > TYPE_PRECISION (type))
539 : : return false;
540 : :
541 : 34 : *common_type = build_nonstandard_integer_type (precision, false);
542 : 34 : return true;
543 : : }
544 : :
545 : : /* Check whether STMT_INFO can be viewed as a tree of integer operations
546 : : in which each node either performs CODE or WIDENED_CODE, and where
547 : : each leaf operand is narrower than the result of STMT_INFO. MAX_NOPS
548 : : specifies the maximum number of leaf operands. SHIFT_P says whether
549 : : CODE and WIDENED_CODE are some sort of shift.
550 : :
551 : : If STMT_INFO is such a tree, return the number of leaf operands
552 : : and describe them in UNPROM[0] onwards. Also set *COMMON_TYPE
553 : : to a type that (a) is narrower than the result of STMT_INFO and
554 : : (b) can hold all leaf operand values.
555 : :
556 : : If SUBTYPE then allow that the signs of the operands
557 : : may differ in signs but not in precision. SUBTYPE is updated to reflect
558 : : this.
559 : :
560 : : Return 0 if STMT_INFO isn't such a tree, or if no such COMMON_TYPE
561 : : exists. */
562 : :
563 : : static unsigned int
564 : 119671360 : vect_widened_op_tree (vec_info *vinfo, stmt_vec_info stmt_info, tree_code code,
565 : : code_helper widened_code, bool shift_p,
566 : : unsigned int max_nops,
567 : : vect_unpromoted_value *unprom, tree *common_type,
568 : : enum optab_subtype *subtype = NULL)
569 : : {
570 : : /* Check for an integer operation with the right code. */
571 : 119671360 : gimple* stmt = stmt_info->stmt;
572 : 119671360 : if (!(is_gimple_assign (stmt) || is_gimple_call (stmt)))
573 : : return 0;
574 : :
575 : 96361093 : code_helper rhs_code;
576 : 96361093 : if (is_gimple_assign (stmt))
577 : 82488479 : rhs_code = gimple_assign_rhs_code (stmt);
578 : 13872614 : else if (is_gimple_call (stmt))
579 : 13872614 : rhs_code = gimple_call_combined_fn (stmt);
580 : : else
581 : : return 0;
582 : :
583 : 96361093 : if (rhs_code != code
584 : 96361093 : && rhs_code != widened_code)
585 : : return 0;
586 : :
587 : 5870441 : tree lhs = gimple_get_lhs (stmt);
588 : 5870441 : tree type = TREE_TYPE (lhs);
589 : 5870441 : if (!INTEGRAL_TYPE_P (type))
590 : : return 0;
591 : :
592 : : /* Assume that both operands will be leaf operands. */
593 : 5321085 : max_nops -= 2;
594 : :
595 : : /* Check the operands. */
596 : 5321085 : unsigned int next_op = 0;
597 : 6024170 : for (unsigned int i = 0; i < 2; ++i)
598 : : {
599 : 5721598 : vect_unpromoted_value *this_unprom = &unprom[next_op];
600 : 5721598 : unsigned int nops = 1;
601 : 5721598 : tree op = gimple_arg (stmt, i);
602 : 5721598 : if (i == 1 && TREE_CODE (op) == INTEGER_CST)
603 : : {
604 : : /* We already have a common type from earlier operands.
605 : : Update it to account for OP. */
606 : 284992 : this_unprom->set_op (op, vect_constant_def);
607 : 284992 : if (!vect_joust_widened_integer (type, shift_p, op, common_type))
608 : : return 0;
609 : : }
610 : : else
611 : : {
612 : : /* Only allow shifts by constants. */
613 : 5436606 : if (shift_p && i == 1)
614 : : return 0;
615 : :
616 : 5431811 : if (rhs_code != code)
617 : : {
618 : : /* If rhs_code is widened_code, don't look through further
619 : : possible promotions, there is a promotion already embedded
620 : : in the WIDEN_*_EXPR. */
621 : 3109 : if (TREE_CODE (op) != SSA_NAME
622 : 3109 : || !INTEGRAL_TYPE_P (TREE_TYPE (op)))
623 : 0 : return 0;
624 : :
625 : 3109 : stmt_vec_info def_stmt_info;
626 : 3109 : gimple *def_stmt;
627 : 3109 : vect_def_type dt;
628 : 3109 : if (!vect_is_simple_use (op, vinfo, &dt, &def_stmt_info,
629 : : &def_stmt))
630 : : return 0;
631 : 3109 : this_unprom->set_op (op, dt, NULL);
632 : : }
633 : 5428702 : else if (!vect_look_through_possible_promotion (vinfo, op,
634 : : this_unprom))
635 : : return 0;
636 : :
637 : 5326020 : if (TYPE_PRECISION (this_unprom->type) == TYPE_PRECISION (type))
638 : : {
639 : : /* The operand isn't widened. If STMT_INFO has the code
640 : : for an unwidened operation, recursively check whether
641 : : this operand is a node of the tree. */
642 : 4880699 : if (rhs_code != code
643 : 4880699 : || max_nops == 0
644 : 4881143 : || this_unprom->dt != vect_internal_def)
645 : : return 0;
646 : :
647 : : /* Give back the leaf slot allocated above now that we're
648 : : not treating this as a leaf operand. */
649 : 444 : max_nops += 1;
650 : :
651 : : /* Recursively process the definition of the operand. */
652 : 444 : stmt_vec_info def_stmt_info
653 : 444 : = vect_get_internal_def (vinfo, this_unprom->op);
654 : :
655 : 444 : nops = vect_widened_op_tree (vinfo, def_stmt_info, code,
656 : : widened_code, shift_p, max_nops,
657 : : this_unprom, common_type,
658 : : subtype);
659 : 444 : if (nops == 0)
660 : : return 0;
661 : :
662 : 299 : max_nops -= nops;
663 : : }
664 : : else
665 : : {
666 : : /* Make sure that the operand is narrower than the result. */
667 : 445321 : if (TYPE_PRECISION (this_unprom->type) * 2
668 : 445321 : > TYPE_PRECISION (type))
669 : : return 0;
670 : :
671 : : /* Update COMMON_TYPE for the new operand. */
672 : 440913 : if (i == 0)
673 : 400370 : *common_type = this_unprom->type;
674 : 40543 : else if (!vect_joust_widened_type (type, this_unprom->type,
675 : : common_type))
676 : : {
677 : 6196 : if (subtype)
678 : : {
679 : : /* See if we can sign extend the smaller type. */
680 : 210 : if (TYPE_PRECISION (this_unprom->type)
681 : 210 : > TYPE_PRECISION (*common_type))
682 : 36 : *common_type = this_unprom->type;
683 : 210 : *subtype = optab_vector_mixed_sign;
684 : : }
685 : : else
686 : : return 0;
687 : : }
688 : : }
689 : : }
690 : 703085 : next_op += nops;
691 : : }
692 : : return next_op;
693 : : }
694 : :
695 : : /* Helper to return a new temporary for pattern of TYPE for STMT. If STMT
696 : : is NULL, the caller must set SSA_NAME_DEF_STMT for the returned SSA var. */
697 : :
698 : : static tree
699 : 1737439 : vect_recog_temp_ssa_var (tree type, gimple *stmt = NULL)
700 : : {
701 : 0 : return make_temp_ssa_name (type, stmt, "patt");
702 : : }
703 : :
704 : : /* STMT2_INFO describes a type conversion that could be split into STMT1
705 : : followed by a version of STMT2_INFO that takes NEW_RHS as its first
706 : : input. Try to do this using pattern statements, returning true on
707 : : success. */
708 : :
709 : : static bool
710 : 31021 : vect_split_statement (vec_info *vinfo, stmt_vec_info stmt2_info, tree new_rhs,
711 : : gimple *stmt1, tree vectype)
712 : : {
713 : 31021 : if (is_pattern_stmt_p (stmt2_info))
714 : : {
715 : : /* STMT2_INFO is part of a pattern. Get the statement to which
716 : : the pattern is attached. */
717 : 1014 : stmt_vec_info orig_stmt2_info = STMT_VINFO_RELATED_STMT (stmt2_info);
718 : 1014 : vect_init_pattern_stmt (vinfo, stmt1, orig_stmt2_info, vectype);
719 : :
720 : 1014 : if (dump_enabled_p ())
721 : 22 : dump_printf_loc (MSG_NOTE, vect_location,
722 : : "Splitting pattern statement: %G", stmt2_info->stmt);
723 : :
724 : : /* Since STMT2_INFO is a pattern statement, we can change it
725 : : in-situ without worrying about changing the code for the
726 : : containing block. */
727 : 1014 : gimple_assign_set_rhs1 (stmt2_info->stmt, new_rhs);
728 : :
729 : 1014 : if (dump_enabled_p ())
730 : : {
731 : 22 : dump_printf_loc (MSG_NOTE, vect_location, "into: %G", stmt1);
732 : 22 : dump_printf_loc (MSG_NOTE, vect_location, "and: %G",
733 : : stmt2_info->stmt);
734 : : }
735 : :
736 : 1014 : gimple_seq *def_seq = &STMT_VINFO_PATTERN_DEF_SEQ (orig_stmt2_info);
737 : 1014 : if (STMT_VINFO_RELATED_STMT (orig_stmt2_info) == stmt2_info)
738 : : /* STMT2_INFO is the actual pattern statement. Add STMT1
739 : : to the end of the definition sequence. */
740 : 1011 : gimple_seq_add_stmt_without_update (def_seq, stmt1);
741 : : else
742 : : {
743 : : /* STMT2_INFO belongs to the definition sequence. Insert STMT1
744 : : before it. */
745 : 3 : gimple_stmt_iterator gsi = gsi_for_stmt (stmt2_info->stmt, def_seq);
746 : 3 : gsi_insert_before_without_update (&gsi, stmt1, GSI_SAME_STMT);
747 : : }
748 : 1014 : return true;
749 : : }
750 : : else
751 : : {
752 : : /* STMT2_INFO doesn't yet have a pattern. Try to create a
753 : : two-statement pattern now. */
754 : 30007 : gcc_assert (!STMT_VINFO_RELATED_STMT (stmt2_info));
755 : 30007 : tree lhs_type = TREE_TYPE (gimple_get_lhs (stmt2_info->stmt));
756 : 30007 : tree lhs_vectype = get_vectype_for_scalar_type (vinfo, lhs_type);
757 : 30007 : if (!lhs_vectype)
758 : : return false;
759 : :
760 : 30007 : if (dump_enabled_p ())
761 : 2152 : dump_printf_loc (MSG_NOTE, vect_location,
762 : : "Splitting statement: %G", stmt2_info->stmt);
763 : :
764 : : /* Add STMT1 as a singleton pattern definition sequence. */
765 : 30007 : gimple_seq *def_seq = &STMT_VINFO_PATTERN_DEF_SEQ (stmt2_info);
766 : 30007 : vect_init_pattern_stmt (vinfo, stmt1, stmt2_info, vectype);
767 : 30007 : gimple_seq_add_stmt_without_update (def_seq, stmt1);
768 : :
769 : : /* Build the second of the two pattern statements. */
770 : 30007 : tree new_lhs = vect_recog_temp_ssa_var (lhs_type, NULL);
771 : 30007 : gassign *new_stmt2 = gimple_build_assign (new_lhs, NOP_EXPR, new_rhs);
772 : 30007 : vect_set_pattern_stmt (vinfo, new_stmt2, stmt2_info, lhs_vectype);
773 : :
774 : 30007 : if (dump_enabled_p ())
775 : : {
776 : 2152 : dump_printf_loc (MSG_NOTE, vect_location,
777 : : "into pattern statements: %G", stmt1);
778 : 2152 : dump_printf_loc (MSG_NOTE, vect_location, "and: %G",
779 : : (gimple *) new_stmt2);
780 : : }
781 : :
782 : 30007 : return true;
783 : : }
784 : : }
785 : :
786 : : /* Look for the following pattern
787 : : X = x[i]
788 : : Y = y[i]
789 : : DIFF = X - Y
790 : : DAD = ABS_EXPR<DIFF>
791 : :
792 : : ABS_STMT should point to a statement of code ABS_EXPR or ABSU_EXPR.
793 : : HALF_TYPE and UNPROM will be set should the statement be found to
794 : : be a widened operation.
795 : : DIFF_STMT will be set to the MINUS_EXPR
796 : : statement that precedes the ABS_STMT if it is a MINUS_EXPR..
797 : : */
798 : : static bool
799 : 20314119 : vect_recog_absolute_difference (vec_info *vinfo, gassign *abs_stmt,
800 : : tree *half_type,
801 : : vect_unpromoted_value unprom[2],
802 : : gassign **diff_stmt)
803 : : {
804 : 20314119 : if (!abs_stmt)
805 : : return false;
806 : :
807 : : /* FORNOW. Can continue analyzing the def-use chain when this stmt in a phi
808 : : inside the loop (in case we are analyzing an outer-loop). */
809 : 20314119 : enum tree_code code = gimple_assign_rhs_code (abs_stmt);
810 : 20314119 : if (code != ABS_EXPR && code != ABSU_EXPR)
811 : : return false;
812 : :
813 : 24200 : tree abs_oprnd = gimple_assign_rhs1 (abs_stmt);
814 : 24200 : tree abs_type = TREE_TYPE (abs_oprnd);
815 : 24200 : if (!abs_oprnd)
816 : : return false;
817 : 17413 : if (!ANY_INTEGRAL_TYPE_P (abs_type)
818 : 7020 : || TYPE_OVERFLOW_WRAPS (abs_type)
819 : 31093 : || TYPE_UNSIGNED (abs_type))
820 : : return false;
821 : :
822 : : /* Peel off conversions from the ABS input. This can involve sign
823 : : changes (e.g. from an unsigned subtraction to a signed ABS input)
824 : : or signed promotion, but it can't include unsigned promotion.
825 : : (Note that ABS of an unsigned promotion should have been folded
826 : : away before now anyway.) */
827 : 6893 : vect_unpromoted_value unprom_diff;
828 : 6893 : abs_oprnd = vect_look_through_possible_promotion (vinfo, abs_oprnd,
829 : : &unprom_diff);
830 : 6893 : if (!abs_oprnd)
831 : : return false;
832 : 6660 : if (TYPE_PRECISION (unprom_diff.type) != TYPE_PRECISION (abs_type)
833 : 6660 : && TYPE_UNSIGNED (unprom_diff.type))
834 : : return false;
835 : :
836 : : /* We then detect if the operand of abs_expr is defined by a minus_expr. */
837 : 6660 : stmt_vec_info diff_stmt_vinfo = vect_get_internal_def (vinfo, abs_oprnd);
838 : 6660 : if (!diff_stmt_vinfo)
839 : : return false;
840 : :
841 : 6166 : gassign *diff = dyn_cast <gassign *> (STMT_VINFO_STMT (diff_stmt_vinfo));
842 : 6166 : if (diff_stmt && diff
843 : 5254 : && gimple_assign_rhs_code (diff) == MINUS_EXPR
844 : 8462 : && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (abs_oprnd)))
845 : 1291 : *diff_stmt = diff;
846 : :
847 : : /* FORNOW. Can continue analyzing the def-use chain when this stmt in a phi
848 : : inside the loop (in case we are analyzing an outer-loop). */
849 : 6166 : if (vect_widened_op_tree (vinfo, diff_stmt_vinfo,
850 : 6166 : MINUS_EXPR, IFN_VEC_WIDEN_MINUS,
851 : : false, 2, unprom, half_type))
852 : : return true;
853 : :
854 : : return false;
855 : : }
856 : :
857 : : /* Convert UNPROM to TYPE and return the result, adding new statements
858 : : to STMT_INFO's pattern definition statements if no better way is
859 : : available. VECTYPE is the vector form of TYPE.
860 : :
861 : : If SUBTYPE then convert the type based on the subtype. */
862 : :
863 : : static tree
864 : 471025 : vect_convert_input (vec_info *vinfo, stmt_vec_info stmt_info, tree type,
865 : : vect_unpromoted_value *unprom, tree vectype,
866 : : enum optab_subtype subtype = optab_default)
867 : : {
868 : : /* Update the type if the signs differ. */
869 : 471025 : if (subtype == optab_vector_mixed_sign)
870 : : {
871 : 164 : gcc_assert (!TYPE_UNSIGNED (type));
872 : 164 : if (TYPE_UNSIGNED (TREE_TYPE (unprom->op)))
873 : : {
874 : 82 : type = unsigned_type_for (type);
875 : 82 : vectype = unsigned_type_for (vectype);
876 : : }
877 : : }
878 : :
879 : : /* Check for a no-op conversion. */
880 : 471025 : if (types_compatible_p (type, TREE_TYPE (unprom->op)))
881 : 157215 : return unprom->op;
882 : :
883 : : /* Allow the caller to create constant vect_unpromoted_values. */
884 : 313810 : if (TREE_CODE (unprom->op) == INTEGER_CST)
885 : 188204 : return wide_int_to_tree (type, wi::to_widest (unprom->op));
886 : :
887 : 125606 : tree input = unprom->op;
888 : 125606 : if (unprom->caster)
889 : : {
890 : 62620 : tree lhs = gimple_get_lhs (unprom->caster->stmt);
891 : 62620 : tree lhs_type = TREE_TYPE (lhs);
892 : :
893 : : /* If the result of the existing cast is the right width, use it
894 : : instead of the source of the cast. */
895 : 62620 : if (TYPE_PRECISION (lhs_type) == TYPE_PRECISION (type))
896 : : input = lhs;
897 : : /* If the precision we want is between the source and result
898 : : precisions of the existing cast, try splitting the cast into
899 : : two and tapping into a mid-way point. */
900 : 60759 : else if (TYPE_PRECISION (lhs_type) > TYPE_PRECISION (type)
901 : 60759 : && TYPE_PRECISION (type) > TYPE_PRECISION (unprom->type))
902 : : {
903 : : /* In order to preserve the semantics of the original cast,
904 : : give the mid-way point the same signedness as the input value.
905 : :
906 : : It would be possible to use a signed type here instead if
907 : : TYPE is signed and UNPROM->TYPE is unsigned, but that would
908 : : make the sign of the midtype sensitive to the order in
909 : : which we process the statements, since the signedness of
910 : : TYPE is the signedness required by just one of possibly
911 : : many users. Also, unsigned promotions are usually as cheap
912 : : as or cheaper than signed ones, so it's better to keep an
913 : : unsigned promotion. */
914 : 31021 : tree midtype = build_nonstandard_integer_type
915 : 31021 : (TYPE_PRECISION (type), TYPE_UNSIGNED (unprom->type));
916 : 31021 : tree vec_midtype = get_vectype_for_scalar_type (vinfo, midtype);
917 : 31021 : if (vec_midtype)
918 : : {
919 : 31021 : input = vect_recog_temp_ssa_var (midtype, NULL);
920 : 31021 : gassign *new_stmt = gimple_build_assign (input, NOP_EXPR,
921 : : unprom->op);
922 : 31021 : if (!vect_split_statement (vinfo, unprom->caster, input, new_stmt,
923 : : vec_midtype))
924 : 0 : append_pattern_def_seq (vinfo, stmt_info,
925 : : new_stmt, vec_midtype);
926 : : }
927 : : }
928 : :
929 : : /* See if we can reuse an existing result. */
930 : 62620 : if (types_compatible_p (type, TREE_TYPE (input)))
931 : : return input;
932 : : }
933 : :
934 : : /* We need a new conversion statement. */
935 : 102580 : tree new_op = vect_recog_temp_ssa_var (type, NULL);
936 : 102580 : gassign *new_stmt = gimple_build_assign (new_op, NOP_EXPR, input);
937 : :
938 : : /* If OP is an external value, see if we can insert the new statement
939 : : on an incoming edge. */
940 : 102580 : if (input == unprom->op && unprom->dt == vect_external_def)
941 : 8493 : if (edge e = vect_get_external_def_edge (vinfo, input))
942 : : {
943 : 512 : basic_block new_bb = gsi_insert_on_edge_immediate (e, new_stmt);
944 : 512 : gcc_assert (!new_bb);
945 : : return new_op;
946 : : }
947 : :
948 : : /* As a (common) last resort, add the statement to the pattern itself. */
949 : 102068 : append_pattern_def_seq (vinfo, stmt_info, new_stmt, vectype);
950 : 102068 : return new_op;
951 : : }
952 : :
953 : : /* Invoke vect_convert_input for N elements of UNPROM and store the
954 : : result in the corresponding elements of RESULT.
955 : :
956 : : If SUBTYPE then convert the type based on the subtype. */
957 : :
958 : : static void
959 : 238878 : vect_convert_inputs (vec_info *vinfo, stmt_vec_info stmt_info, unsigned int n,
960 : : tree *result, tree type, vect_unpromoted_value *unprom,
961 : : tree vectype, enum optab_subtype subtype = optab_default)
962 : : {
963 : 709675 : for (unsigned int i = 0; i < n; ++i)
964 : : {
965 : : unsigned int j;
966 : 702438 : for (j = 0; j < i; ++j)
967 : 231919 : if (unprom[j].op == unprom[i].op)
968 : : break;
969 : :
970 : 470797 : if (j < i)
971 : 278 : result[i] = result[j];
972 : : else
973 : 470519 : result[i] = vect_convert_input (vinfo, stmt_info,
974 : 470519 : type, &unprom[i], vectype, subtype);
975 : : }
976 : 238878 : }
977 : :
978 : : /* The caller has created a (possibly empty) sequence of pattern definition
979 : : statements followed by a single statement PATTERN_STMT. Cast the result
980 : : of this final statement to TYPE. If a new statement is needed, add
981 : : PATTERN_STMT to the end of STMT_INFO's pattern definition statements
982 : : and return the new statement, otherwise return PATTERN_STMT as-is.
983 : : VECITYPE is the vector form of PATTERN_STMT's result type. */
984 : :
985 : : static gimple *
986 : 263960 : vect_convert_output (vec_info *vinfo, stmt_vec_info stmt_info, tree type,
987 : : gimple *pattern_stmt, tree vecitype)
988 : : {
989 : 263960 : tree lhs = gimple_get_lhs (pattern_stmt);
990 : 263960 : if (!types_compatible_p (type, TREE_TYPE (lhs)))
991 : : {
992 : 237616 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt, vecitype);
993 : 237616 : tree cast_var = vect_recog_temp_ssa_var (type, NULL);
994 : 237616 : pattern_stmt = gimple_build_assign (cast_var, NOP_EXPR, lhs);
995 : : }
996 : 263960 : return pattern_stmt;
997 : : }
998 : :
999 : : /* Return true if STMT_VINFO describes a reduction for which reassociation
1000 : : is allowed. If STMT_INFO is part of a group, assume that it's part of
1001 : : a reduction chain and optimistically assume that all statements
1002 : : except the last allow reassociation.
1003 : : Also require it to have code CODE and to be a reduction
1004 : : in the outermost loop. When returning true, store the operands in
1005 : : *OP0_OUT and *OP1_OUT. */
1006 : :
1007 : : static bool
1008 : 89328755 : vect_reassociating_reduction_p (vec_info *vinfo,
1009 : : stmt_vec_info stmt_info, tree_code code,
1010 : : tree *op0_out, tree *op1_out)
1011 : : {
1012 : 89328755 : loop_vec_info loop_info = dyn_cast <loop_vec_info> (vinfo);
1013 : 9964193 : if (!loop_info)
1014 : : return false;
1015 : :
1016 : 9964193 : gassign *assign = dyn_cast <gassign *> (stmt_info->stmt);
1017 : 10795238 : if (!assign || gimple_assign_rhs_code (assign) != code)
1018 : : return false;
1019 : :
1020 : : /* We don't allow changing the order of the computation in the inner-loop
1021 : : when doing outer-loop vectorization. */
1022 : 2095271 : class loop *loop = LOOP_VINFO_LOOP (loop_info);
1023 : 91323986 : if (loop && nested_in_vect_loop_p (loop, stmt_info))
1024 : : return false;
1025 : :
1026 : 2047028 : if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)
1027 : : {
1028 : 102829 : if (needs_fold_left_reduction_p (TREE_TYPE (gimple_assign_lhs (assign)),
1029 : 102829 : code))
1030 : : return false;
1031 : : }
1032 : 1944199 : else if (REDUC_GROUP_FIRST_ELEMENT (stmt_info) == NULL)
1033 : : return false;
1034 : :
1035 : 100040 : *op0_out = gimple_assign_rhs1 (assign);
1036 : 100040 : *op1_out = gimple_assign_rhs2 (assign);
1037 : 100040 : if (commutative_tree_code (code) && STMT_VINFO_REDUC_IDX (stmt_info) == 0)
1038 : 31487 : std::swap (*op0_out, *op1_out);
1039 : : return true;
1040 : : }
1041 : :
1042 : : /* match.pd function to match
1043 : : (cond (cmp@3 a b) (convert@1 c) (convert@2 d))
1044 : : with conditions:
1045 : : 1) @1, @2, c, d, a, b are all integral type.
1046 : : 2) There's single_use for both @1 and @2.
1047 : : 3) a, c have same precision.
1048 : : 4) c and @1 have different precision.
1049 : : 5) c, d are the same type or they can differ in sign when convert is
1050 : : truncation.
1051 : :
1052 : : record a and c and d and @3. */
1053 : :
1054 : : extern bool gimple_cond_expr_convert_p (tree, tree*, tree (*)(tree));
1055 : :
1056 : : /* Function vect_recog_cond_expr_convert
1057 : :
1058 : : Try to find the following pattern:
1059 : :
1060 : : TYPE_AB A,B;
1061 : : TYPE_CD C,D;
1062 : : TYPE_E E;
1063 : : TYPE_E op_true = (TYPE_E) A;
1064 : : TYPE_E op_false = (TYPE_E) B;
1065 : :
1066 : : E = C cmp D ? op_true : op_false;
1067 : :
1068 : : where
1069 : : TYPE_PRECISION (TYPE_E) != TYPE_PRECISION (TYPE_CD);
1070 : : TYPE_PRECISION (TYPE_AB) == TYPE_PRECISION (TYPE_CD);
1071 : : single_use of op_true and op_false.
1072 : : TYPE_AB could differ in sign when (TYPE_E) A is a truncation.
1073 : :
1074 : : Input:
1075 : :
1076 : : * STMT_VINFO: The stmt from which the pattern search begins.
1077 : : here it starts with E = c cmp D ? op_true : op_false;
1078 : :
1079 : : Output:
1080 : :
1081 : : TYPE1 E' = C cmp D ? A : B;
1082 : : TYPE3 E = (TYPE3) E';
1083 : :
1084 : : There may extra nop_convert for A or B to handle different signness.
1085 : :
1086 : : * TYPE_OUT: The vector type of the output of this pattern.
1087 : :
1088 : : * Return value: A new stmt that will be used to replace the sequence of
1089 : : stmts that constitute the pattern. In this case it will be:
1090 : : E = (TYPE3)E';
1091 : : E' = C cmp D ? A : B; is recorded in pattern definition statements; */
1092 : :
1093 : : static gimple *
1094 : 29848350 : vect_recog_cond_expr_convert_pattern (vec_info *vinfo,
1095 : : stmt_vec_info stmt_vinfo, tree *type_out)
1096 : : {
1097 : 29848350 : gassign *last_stmt = dyn_cast <gassign *> (stmt_vinfo->stmt);
1098 : 20419159 : tree lhs, match[4], temp, type, new_lhs, op2, op1;
1099 : 20419159 : gimple *cond_stmt;
1100 : 20419159 : gimple *pattern_stmt;
1101 : 29848265 : enum tree_code code = NOP_EXPR;
1102 : :
1103 : 20419159 : if (!last_stmt)
1104 : : return NULL;
1105 : :
1106 : 20419159 : lhs = gimple_assign_lhs (last_stmt);
1107 : :
1108 : : /* Find E = C cmp D ? (TYPE3) A ? (TYPE3) B;
1109 : : TYPE_PRECISION (A) == TYPE_PRECISION (C). */
1110 : 20419159 : if (!gimple_cond_expr_convert_p (lhs, &match[0], NULL))
1111 : : return NULL;
1112 : :
1113 : 85 : if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (lhs)))
1114 : 24 : code = INTEGRAL_TYPE_P (TREE_TYPE (match[1])) ? FLOAT_EXPR : CONVERT_EXPR;
1115 : 61 : else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (match[1])))
1116 : 0 : code = FIX_TRUNC_EXPR;
1117 : :
1118 : 85 : op1 = match[1];
1119 : 85 : op2 = match[2];
1120 : 85 : type = TREE_TYPE (op1);
1121 : : /* When op1/op2 is REAL_CST, the conversion must be CONVERT_EXPR from
1122 : : SCALAR_FLOAT_TYPE_P which is restricted in gimple_cond_expr_convert_p.
1123 : : Otherwise, the conversion could be FLOAT_EXPR, FIX_TRUNC_EXPR
1124 : : or CONVERT_EXPR. */
1125 : 85 : if (TREE_CODE (op1) == REAL_CST)
1126 : : {
1127 : 12 : op1 = const_unop (CONVERT_EXPR, TREE_TYPE (op2), op1);
1128 : 12 : type = TREE_TYPE (op2);
1129 : 12 : if (op1 == NULL_TREE)
1130 : : return NULL;
1131 : : }
1132 : 73 : else if (TREE_CODE (op2) == REAL_CST)
1133 : : {
1134 : 0 : op2 = const_unop (FLOAT_EXPR, TREE_TYPE (op1), op2);
1135 : 0 : if (op2 == NULL_TREE)
1136 : : return NULL;
1137 : : }
1138 : 73 : else if (code == NOP_EXPR)
1139 : : {
1140 : 61 : if (TYPE_SIGN (type) != TYPE_SIGN (TREE_TYPE (match[2])))
1141 : : {
1142 : 24 : op2 = vect_recog_temp_ssa_var (type, NULL);
1143 : 24 : gimple* nop_stmt = gimple_build_assign (op2, NOP_EXPR, match[2]);
1144 : 24 : append_pattern_def_seq (vinfo, stmt_vinfo, nop_stmt,
1145 : : get_vectype_for_scalar_type (vinfo, type));
1146 : : }
1147 : : }
1148 : :
1149 : 85 : vect_pattern_detected ("vect_recog_cond_expr_convert_pattern", last_stmt);
1150 : :
1151 : 85 : temp = vect_recog_temp_ssa_var (type, NULL);
1152 : 85 : cond_stmt = gimple_build_assign (temp, build3 (COND_EXPR, type, match[3],
1153 : : op1, op2));
1154 : 85 : append_pattern_def_seq (vinfo, stmt_vinfo, cond_stmt,
1155 : : get_vectype_for_scalar_type (vinfo, type));
1156 : 85 : new_lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL);
1157 : 85 : pattern_stmt = gimple_build_assign (new_lhs, code, temp);
1158 : 85 : *type_out = STMT_VINFO_VECTYPE (stmt_vinfo);
1159 : :
1160 : 85 : if (dump_enabled_p ())
1161 : 27 : dump_printf_loc (MSG_NOTE, vect_location,
1162 : : "created pattern stmt: %G", pattern_stmt);
1163 : : return pattern_stmt;
1164 : : }
1165 : :
1166 : : /* Function vect_recog_dot_prod_pattern
1167 : :
1168 : : Try to find the following pattern:
1169 : :
1170 : : type1a x_t
1171 : : type1b y_t;
1172 : : TYPE1 prod;
1173 : : TYPE2 sum = init;
1174 : : loop:
1175 : : sum_0 = phi <init, sum_1>
1176 : : S1 x_t = ...
1177 : : S2 y_t = ...
1178 : : S3 x_T = (TYPE1) x_t;
1179 : : S4 y_T = (TYPE1) y_t;
1180 : : S5 prod = x_T * y_T;
1181 : : [S6 prod = (TYPE2) prod; #optional]
1182 : : S7 sum_1 = prod + sum_0;
1183 : :
1184 : : where 'TYPE1' is exactly double the size of type 'type1a' and 'type1b',
1185 : : the sign of 'TYPE1' must be one of 'type1a' or 'type1b' but the sign of
1186 : : 'type1a' and 'type1b' can differ.
1187 : :
1188 : : Input:
1189 : :
1190 : : * STMT_VINFO: The stmt from which the pattern search begins. In the
1191 : : example, when this function is called with S7, the pattern {S3,S4,S5,S6,S7}
1192 : : will be detected.
1193 : :
1194 : : Output:
1195 : :
1196 : : * TYPE_OUT: The type of the output of this pattern.
1197 : :
1198 : : * Return value: A new stmt that will be used to replace the sequence of
1199 : : stmts that constitute the pattern. In this case it will be:
1200 : : WIDEN_DOT_PRODUCT <x_t, y_t, sum_0>
1201 : :
1202 : : Note: The dot-prod idiom is a widening reduction pattern that is
1203 : : vectorized without preserving all the intermediate results. It
1204 : : produces only N/2 (widened) results (by summing up pairs of
1205 : : intermediate results) rather than all N results. Therefore, we
1206 : : cannot allow this pattern when we want to get all the results and in
1207 : : the correct order (as is the case when this computation is in an
1208 : : inner-loop nested in an outer-loop that us being vectorized). */
1209 : :
1210 : : static gimple *
1211 : 29776605 : vect_recog_dot_prod_pattern (vec_info *vinfo,
1212 : : stmt_vec_info stmt_vinfo, tree *type_out)
1213 : : {
1214 : 29776605 : tree oprnd0, oprnd1;
1215 : 29776605 : gimple *last_stmt = stmt_vinfo->stmt;
1216 : 29776605 : tree type, half_type;
1217 : 29776605 : gimple *pattern_stmt;
1218 : 29776605 : tree var;
1219 : :
1220 : : /* Look for the following pattern
1221 : : DX = (TYPE1) X;
1222 : : DY = (TYPE1) Y;
1223 : : DPROD = DX * DY;
1224 : : DDPROD = (TYPE2) DPROD;
1225 : : sum_1 = DDPROD + sum_0;
1226 : : In which
1227 : : - DX is double the size of X
1228 : : - DY is double the size of Y
1229 : : - DX, DY, DPROD all have the same type but the sign
1230 : : between X, Y and DPROD can differ.
1231 : : - sum is the same size of DPROD or bigger
1232 : : - sum has been recognized as a reduction variable.
1233 : :
1234 : : This is equivalent to:
1235 : : DPROD = X w* Y; #widen mult
1236 : : sum_1 = DPROD w+ sum_0; #widen summation
1237 : : or
1238 : : DPROD = X w* Y; #widen mult
1239 : : sum_1 = DPROD + sum_0; #summation
1240 : : */
1241 : :
1242 : : /* Starting from LAST_STMT, follow the defs of its uses in search
1243 : : of the above pattern. */
1244 : :
1245 : 29776605 : if (!vect_reassociating_reduction_p (vinfo, stmt_vinfo, PLUS_EXPR,
1246 : : &oprnd0, &oprnd1))
1247 : : return NULL;
1248 : :
1249 : 33704 : type = TREE_TYPE (gimple_get_lhs (last_stmt));
1250 : :
1251 : 33704 : vect_unpromoted_value unprom_mult;
1252 : 33704 : oprnd0 = vect_look_through_possible_promotion (vinfo, oprnd0, &unprom_mult);
1253 : :
1254 : : /* So far so good. Since last_stmt was detected as a (summation) reduction,
1255 : : we know that oprnd1 is the reduction variable (defined by a loop-header
1256 : : phi), and oprnd0 is an ssa-name defined by a stmt in the loop body.
1257 : : Left to check that oprnd0 is defined by a (widen_)mult_expr */
1258 : 33704 : if (!oprnd0)
1259 : : return NULL;
1260 : :
1261 : 25462 : stmt_vec_info mult_vinfo = vect_get_internal_def (vinfo, oprnd0);
1262 : 25462 : if (!mult_vinfo)
1263 : : return NULL;
1264 : :
1265 : : /* FORNOW. Can continue analyzing the def-use chain when this stmt in a phi
1266 : : inside the loop (in case we are analyzing an outer-loop). */
1267 : 74241 : vect_unpromoted_value unprom0[2];
1268 : 24747 : enum optab_subtype subtype = optab_vector;
1269 : 24747 : if (!vect_widened_op_tree (vinfo, mult_vinfo, MULT_EXPR, WIDEN_MULT_EXPR,
1270 : : false, 2, unprom0, &half_type, &subtype))
1271 : : return NULL;
1272 : :
1273 : : /* If there are two widening operations, make sure they agree on the sign
1274 : : of the extension. The result of an optab_vector_mixed_sign operation
1275 : : is signed; otherwise, the result has the same sign as the operands. */
1276 : 1000 : if (TYPE_PRECISION (unprom_mult.type) != TYPE_PRECISION (type)
1277 : 1614 : && (subtype == optab_vector_mixed_sign
1278 : 614 : ? TYPE_UNSIGNED (unprom_mult.type)
1279 : 426 : : TYPE_SIGN (unprom_mult.type) != TYPE_SIGN (half_type)))
1280 : : return NULL;
1281 : :
1282 : 892 : vect_pattern_detected ("vect_recog_dot_prod_pattern", last_stmt);
1283 : :
1284 : : /* If the inputs have mixed signs, canonicalize on using the signed
1285 : : input type for analysis. This also helps when emulating mixed-sign
1286 : : operations using signed operations. */
1287 : 892 : if (subtype == optab_vector_mixed_sign)
1288 : 150 : half_type = signed_type_for (half_type);
1289 : :
1290 : 892 : tree half_vectype;
1291 : 892 : if (!vect_supportable_conv_optab_p (vinfo, type, DOT_PROD_EXPR, half_type,
1292 : : type_out, &half_vectype, subtype))
1293 : : {
1294 : : /* We can emulate a mixed-sign dot-product using a sequence of
1295 : : signed dot-products; see vect_emulate_mixed_dot_prod for details. */
1296 : 475 : if (subtype != optab_vector_mixed_sign
1297 : 475 : || !vect_supportable_conv_optab_p (vinfo, signed_type_for (type),
1298 : : DOT_PROD_EXPR, half_type,
1299 : : type_out, &half_vectype,
1300 : : optab_vector))
1301 : 469 : return NULL;
1302 : :
1303 : 6 : *type_out = signed_or_unsigned_type_for (TYPE_UNSIGNED (type),
1304 : : *type_out);
1305 : : }
1306 : :
1307 : : /* Get the inputs in the appropriate types. */
1308 : 423 : tree mult_oprnd[2];
1309 : 423 : vect_convert_inputs (vinfo, stmt_vinfo, 2, mult_oprnd, half_type,
1310 : : unprom0, half_vectype, subtype);
1311 : :
1312 : 423 : var = vect_recog_temp_ssa_var (type, NULL);
1313 : 423 : pattern_stmt = gimple_build_assign (var, DOT_PROD_EXPR,
1314 : : mult_oprnd[0], mult_oprnd[1], oprnd1);
1315 : :
1316 : 423 : return pattern_stmt;
1317 : : }
1318 : :
1319 : :
1320 : : /* Function vect_recog_sad_pattern
1321 : :
1322 : : Try to find the following Sum of Absolute Difference (SAD) pattern:
1323 : :
1324 : : type x_t, y_t;
1325 : : signed TYPE1 diff, abs_diff;
1326 : : TYPE2 sum = init;
1327 : : loop:
1328 : : sum_0 = phi <init, sum_1>
1329 : : S1 x_t = ...
1330 : : S2 y_t = ...
1331 : : S3 x_T = (TYPE1) x_t;
1332 : : S4 y_T = (TYPE1) y_t;
1333 : : S5 diff = x_T - y_T;
1334 : : S6 abs_diff = ABS_EXPR <diff>;
1335 : : [S7 abs_diff = (TYPE2) abs_diff; #optional]
1336 : : S8 sum_1 = abs_diff + sum_0;
1337 : :
1338 : : where 'TYPE1' is at least double the size of type 'type', and 'TYPE2' is the
1339 : : same size of 'TYPE1' or bigger. This is a special case of a reduction
1340 : : computation.
1341 : :
1342 : : Input:
1343 : :
1344 : : * STMT_VINFO: The stmt from which the pattern search begins. In the
1345 : : example, when this function is called with S8, the pattern
1346 : : {S3,S4,S5,S6,S7,S8} will be detected.
1347 : :
1348 : : Output:
1349 : :
1350 : : * TYPE_OUT: The type of the output of this pattern.
1351 : :
1352 : : * Return value: A new stmt that will be used to replace the sequence of
1353 : : stmts that constitute the pattern. In this case it will be:
1354 : : SAD_EXPR <x_t, y_t, sum_0>
1355 : : */
1356 : :
1357 : : static gimple *
1358 : 29776188 : vect_recog_sad_pattern (vec_info *vinfo,
1359 : : stmt_vec_info stmt_vinfo, tree *type_out)
1360 : : {
1361 : 29776188 : gimple *last_stmt = stmt_vinfo->stmt;
1362 : 29776188 : tree half_type;
1363 : :
1364 : : /* Look for the following pattern
1365 : : DX = (TYPE1) X;
1366 : : DY = (TYPE1) Y;
1367 : : DDIFF = DX - DY;
1368 : : DAD = ABS_EXPR <DDIFF>;
1369 : : DDPROD = (TYPE2) DPROD;
1370 : : sum_1 = DAD + sum_0;
1371 : : In which
1372 : : - DX is at least double the size of X
1373 : : - DY is at least double the size of Y
1374 : : - DX, DY, DDIFF, DAD all have the same type
1375 : : - sum is the same size of DAD or bigger
1376 : : - sum has been recognized as a reduction variable.
1377 : :
1378 : : This is equivalent to:
1379 : : DDIFF = X w- Y; #widen sub
1380 : : DAD = ABS_EXPR <DDIFF>;
1381 : : sum_1 = DAD w+ sum_0; #widen summation
1382 : : or
1383 : : DDIFF = X w- Y; #widen sub
1384 : : DAD = ABS_EXPR <DDIFF>;
1385 : : sum_1 = DAD + sum_0; #summation
1386 : : */
1387 : :
1388 : : /* Starting from LAST_STMT, follow the defs of its uses in search
1389 : : of the above pattern. */
1390 : :
1391 : 29776188 : tree plus_oprnd0, plus_oprnd1;
1392 : 29776188 : if (!vect_reassociating_reduction_p (vinfo, stmt_vinfo, PLUS_EXPR,
1393 : : &plus_oprnd0, &plus_oprnd1))
1394 : : return NULL;
1395 : :
1396 : 33281 : tree sum_type = TREE_TYPE (gimple_get_lhs (last_stmt));
1397 : :
1398 : : /* Any non-truncating sequence of conversions is OK here, since
1399 : : with a successful match, the result of the ABS(U) is known to fit
1400 : : within the nonnegative range of the result type. (It cannot be the
1401 : : negative of the minimum signed value due to the range of the widening
1402 : : MINUS_EXPR.) */
1403 : 33281 : vect_unpromoted_value unprom_abs;
1404 : 33281 : plus_oprnd0 = vect_look_through_possible_promotion (vinfo, plus_oprnd0,
1405 : : &unprom_abs);
1406 : :
1407 : : /* So far so good. Since last_stmt was detected as a (summation) reduction,
1408 : : we know that plus_oprnd1 is the reduction variable (defined by a loop-header
1409 : : phi), and plus_oprnd0 is an ssa-name defined by a stmt in the loop body.
1410 : : Then check that plus_oprnd0 is defined by an abs_expr. */
1411 : :
1412 : 33281 : if (!plus_oprnd0)
1413 : : return NULL;
1414 : :
1415 : 25039 : stmt_vec_info abs_stmt_vinfo = vect_get_internal_def (vinfo, plus_oprnd0);
1416 : 25039 : if (!abs_stmt_vinfo)
1417 : : return NULL;
1418 : :
1419 : : /* FORNOW. Can continue analyzing the def-use chain when this stmt in a phi
1420 : : inside the loop (in case we are analyzing an outer-loop). */
1421 : 24324 : gassign *abs_stmt = dyn_cast <gassign *> (abs_stmt_vinfo->stmt);
1422 : 72972 : vect_unpromoted_value unprom[2];
1423 : :
1424 : 24324 : if (!abs_stmt)
1425 : : {
1426 : 29776234 : gcall *abd_stmt = dyn_cast <gcall *> (abs_stmt_vinfo->stmt);
1427 : 272 : if (!abd_stmt
1428 : 272 : || !gimple_call_internal_p (abd_stmt)
1429 : 0 : || gimple_call_num_args (abd_stmt) != 2)
1430 : : return NULL;
1431 : :
1432 : 0 : tree abd_oprnd0 = gimple_call_arg (abd_stmt, 0);
1433 : 0 : tree abd_oprnd1 = gimple_call_arg (abd_stmt, 1);
1434 : :
1435 : 0 : if (gimple_call_internal_fn (abd_stmt) == IFN_ABD
1436 : 0 : || gimple_call_internal_fn (abd_stmt) == IFN_VEC_WIDEN_ABD)
1437 : : {
1438 : 0 : unprom[0].op = abd_oprnd0;
1439 : 0 : unprom[0].type = TREE_TYPE (abd_oprnd0);
1440 : 0 : unprom[1].op = abd_oprnd1;
1441 : 0 : unprom[1].type = TREE_TYPE (abd_oprnd1);
1442 : : }
1443 : : else
1444 : : return NULL;
1445 : :
1446 : 0 : half_type = unprom[0].type;
1447 : : }
1448 : 24001 : else if (!vect_recog_absolute_difference (vinfo, abs_stmt, &half_type,
1449 : : unprom, NULL))
1450 : : return NULL;
1451 : :
1452 : 377 : vect_pattern_detected ("vect_recog_sad_pattern", last_stmt);
1453 : :
1454 : 377 : tree half_vectype;
1455 : 377 : if (!vect_supportable_direct_optab_p (vinfo, sum_type, SAD_EXPR, half_type,
1456 : : type_out, &half_vectype))
1457 : : return NULL;
1458 : :
1459 : : /* Get the inputs to the SAD_EXPR in the appropriate types. */
1460 : 226 : tree sad_oprnd[2];
1461 : 226 : vect_convert_inputs (vinfo, stmt_vinfo, 2, sad_oprnd, half_type,
1462 : : unprom, half_vectype);
1463 : :
1464 : 226 : tree var = vect_recog_temp_ssa_var (sum_type, NULL);
1465 : 226 : gimple *pattern_stmt = gimple_build_assign (var, SAD_EXPR, sad_oprnd[0],
1466 : : sad_oprnd[1], plus_oprnd1);
1467 : :
1468 : 226 : return pattern_stmt;
1469 : : }
1470 : :
1471 : : /* Function vect_recog_abd_pattern
1472 : :
1473 : : Try to find the following ABsolute Difference (ABD) or
1474 : : widening ABD (WIDEN_ABD) pattern:
1475 : :
1476 : : TYPE1 x;
1477 : : TYPE2 y;
1478 : : TYPE3 x_cast = (TYPE3) x; // widening or no-op
1479 : : TYPE3 y_cast = (TYPE3) y; // widening or no-op
1480 : : TYPE3 diff = x_cast - y_cast;
1481 : : TYPE4 diff_cast = (TYPE4) diff; // widening or no-op
1482 : : TYPE5 abs = ABS(U)_EXPR <diff_cast>;
1483 : :
1484 : : WIDEN_ABD exists to optimize the case where TYPE4 is at least
1485 : : twice as wide as TYPE3.
1486 : :
1487 : : Input:
1488 : :
1489 : : * STMT_VINFO: The stmt from which the pattern search begins
1490 : :
1491 : : Output:
1492 : :
1493 : : * TYPE_OUT: The type of the output of this pattern
1494 : :
1495 : : * Return value: A new stmt that will be used to replace the sequence of
1496 : : stmts that constitute the pattern, principally:
1497 : : out = IFN_ABD (x, y)
1498 : : out = IFN_WIDEN_ABD (x, y)
1499 : : */
1500 : :
1501 : : static gimple *
1502 : 29719175 : vect_recog_abd_pattern (vec_info *vinfo,
1503 : : stmt_vec_info stmt_vinfo, tree *type_out)
1504 : : {
1505 : 50009293 : gassign *last_stmt = dyn_cast <gassign *> (STMT_VINFO_STMT (stmt_vinfo));
1506 : 20290118 : if (!last_stmt)
1507 : : return NULL;
1508 : :
1509 : 20290118 : tree out_type = TREE_TYPE (gimple_assign_lhs (last_stmt));
1510 : :
1511 : 60870354 : vect_unpromoted_value unprom[2];
1512 : 20290118 : gassign *diff_stmt = NULL;
1513 : 20290118 : tree abd_in_type;
1514 : 20290118 : if (!vect_recog_absolute_difference (vinfo, last_stmt, &abd_in_type,
1515 : : unprom, &diff_stmt))
1516 : : {
1517 : : /* We cannot try further without having a non-widening MINUS. */
1518 : 20289219 : if (!diff_stmt)
1519 : : return NULL;
1520 : :
1521 : 1291 : unprom[0].op = gimple_assign_rhs1 (diff_stmt);
1522 : 1291 : unprom[1].op = gimple_assign_rhs2 (diff_stmt);
1523 : 1291 : abd_in_type = signed_type_for (out_type);
1524 : : }
1525 : :
1526 : 2190 : tree abd_out_type = abd_in_type;
1527 : :
1528 : 2190 : tree vectype_in = get_vectype_for_scalar_type (vinfo, abd_in_type);
1529 : 2190 : if (!vectype_in)
1530 : : return NULL;
1531 : :
1532 : 1908 : internal_fn ifn = IFN_ABD;
1533 : 1908 : tree vectype_out = vectype_in;
1534 : :
1535 : 1908 : if (TYPE_PRECISION (out_type) >= TYPE_PRECISION (abd_in_type) * 2
1536 : 1908 : && stmt_vinfo->min_output_precision >= TYPE_PRECISION (abd_in_type) * 2)
1537 : : {
1538 : 792 : tree mid_type
1539 : 792 : = build_nonstandard_integer_type (TYPE_PRECISION (abd_in_type) * 2,
1540 : 792 : TYPE_UNSIGNED (abd_in_type));
1541 : 792 : tree mid_vectype = get_vectype_for_scalar_type (vinfo, mid_type);
1542 : :
1543 : 792 : code_helper dummy_code;
1544 : 792 : int dummy_int;
1545 : 792 : auto_vec<tree> dummy_vec;
1546 : 792 : if (mid_vectype
1547 : 792 : && supportable_widening_operation (vinfo, IFN_VEC_WIDEN_ABD,
1548 : : stmt_vinfo, mid_vectype,
1549 : : vectype_in,
1550 : : &dummy_code, &dummy_code,
1551 : : &dummy_int, &dummy_vec))
1552 : : {
1553 : 0 : ifn = IFN_VEC_WIDEN_ABD;
1554 : 0 : abd_out_type = mid_type;
1555 : 0 : vectype_out = mid_vectype;
1556 : : }
1557 : 792 : }
1558 : :
1559 : 792 : if (ifn == IFN_ABD
1560 : 1908 : && !direct_internal_fn_supported_p (ifn, vectype_in,
1561 : : OPTIMIZE_FOR_SPEED))
1562 : : return NULL;
1563 : :
1564 : 0 : vect_pattern_detected ("vect_recog_abd_pattern", last_stmt);
1565 : :
1566 : 0 : tree abd_oprnds[2];
1567 : 0 : vect_convert_inputs (vinfo, stmt_vinfo, 2, abd_oprnds,
1568 : : abd_in_type, unprom, vectype_in);
1569 : :
1570 : 0 : *type_out = get_vectype_for_scalar_type (vinfo, out_type);
1571 : :
1572 : 0 : tree abd_result = vect_recog_temp_ssa_var (abd_out_type, NULL);
1573 : 0 : gcall *abd_stmt = gimple_build_call_internal (ifn, 2,
1574 : : abd_oprnds[0], abd_oprnds[1]);
1575 : 0 : gimple_call_set_lhs (abd_stmt, abd_result);
1576 : 0 : gimple_set_location (abd_stmt, gimple_location (last_stmt));
1577 : :
1578 : 0 : gimple *stmt = abd_stmt;
1579 : 0 : if (TYPE_PRECISION (abd_in_type) == TYPE_PRECISION (abd_out_type)
1580 : 0 : && TYPE_PRECISION (abd_out_type) < TYPE_PRECISION (out_type)
1581 : 0 : && !TYPE_UNSIGNED (abd_out_type))
1582 : : {
1583 : 0 : tree unsign = unsigned_type_for (abd_out_type);
1584 : 0 : stmt = vect_convert_output (vinfo, stmt_vinfo, unsign, stmt, vectype_out);
1585 : 0 : vectype_out = get_vectype_for_scalar_type (vinfo, unsign);
1586 : : }
1587 : :
1588 : 0 : return vect_convert_output (vinfo, stmt_vinfo, out_type, stmt, vectype_out);
1589 : : }
1590 : :
1591 : : /* Recognize an operation that performs ORIG_CODE on widened inputs,
1592 : : so that it can be treated as though it had the form:
1593 : :
1594 : : A_TYPE a;
1595 : : B_TYPE b;
1596 : : HALF_TYPE a_cast = (HALF_TYPE) a; // possible no-op
1597 : : HALF_TYPE b_cast = (HALF_TYPE) b; // possible no-op
1598 : : | RES_TYPE a_extend = (RES_TYPE) a_cast; // promotion from HALF_TYPE
1599 : : | RES_TYPE b_extend = (RES_TYPE) b_cast; // promotion from HALF_TYPE
1600 : : | RES_TYPE res = a_extend ORIG_CODE b_extend;
1601 : :
1602 : : Try to replace the pattern with:
1603 : :
1604 : : A_TYPE a;
1605 : : B_TYPE b;
1606 : : HALF_TYPE a_cast = (HALF_TYPE) a; // possible no-op
1607 : : HALF_TYPE b_cast = (HALF_TYPE) b; // possible no-op
1608 : : | EXT_TYPE ext = a_cast WIDE_CODE b_cast;
1609 : : | RES_TYPE res = (EXT_TYPE) ext; // possible no-op
1610 : :
1611 : : where EXT_TYPE is wider than HALF_TYPE but has the same signedness.
1612 : :
1613 : : SHIFT_P is true if ORIG_CODE and WIDE_CODE are shifts. NAME is the
1614 : : name of the pattern being matched, for dump purposes. */
1615 : :
1616 : : static gimple *
1617 : 119612938 : vect_recog_widen_op_pattern (vec_info *vinfo,
1618 : : stmt_vec_info last_stmt_info, tree *type_out,
1619 : : tree_code orig_code, code_helper wide_code,
1620 : : bool shift_p, const char *name)
1621 : : {
1622 : 119612938 : gimple *last_stmt = last_stmt_info->stmt;
1623 : :
1624 : 358838814 : vect_unpromoted_value unprom[2];
1625 : 119612938 : tree half_type;
1626 : 119612938 : if (!vect_widened_op_tree (vinfo, last_stmt_info, orig_code, orig_code,
1627 : : shift_p, 2, unprom, &half_type))
1628 : :
1629 : : return NULL;
1630 : :
1631 : : /* Pattern detected. */
1632 : 294607 : vect_pattern_detected (name, last_stmt);
1633 : :
1634 : 294607 : tree type = TREE_TYPE (gimple_get_lhs (last_stmt));
1635 : 294607 : tree itype = type;
1636 : 294607 : if (TYPE_PRECISION (type) != TYPE_PRECISION (half_type) * 2
1637 : 294607 : || TYPE_UNSIGNED (type) != TYPE_UNSIGNED (half_type))
1638 : 204181 : itype = build_nonstandard_integer_type (TYPE_PRECISION (half_type) * 2,
1639 : 204181 : TYPE_UNSIGNED (half_type));
1640 : :
1641 : : /* Check target support */
1642 : 294607 : tree vectype = get_vectype_for_scalar_type (vinfo, half_type);
1643 : 294607 : tree vecitype = get_vectype_for_scalar_type (vinfo, itype);
1644 : 294607 : tree ctype = itype;
1645 : 294607 : tree vecctype = vecitype;
1646 : 294607 : if (orig_code == MINUS_EXPR
1647 : 7920 : && TYPE_UNSIGNED (itype)
1648 : 298253 : && TYPE_PRECISION (type) > TYPE_PRECISION (itype))
1649 : : {
1650 : : /* Subtraction is special, even if half_type is unsigned and no matter
1651 : : whether type is signed or unsigned, if type is wider than itype,
1652 : : we need to sign-extend from the widening operation result to the
1653 : : result type.
1654 : : Consider half_type unsigned char, operand 1 0xfe, operand 2 0xff,
1655 : : itype unsigned short and type either int or unsigned int.
1656 : : Widened (unsigned short) 0xfe - (unsigned short) 0xff is
1657 : : (unsigned short) 0xffff, but for type int we want the result -1
1658 : : and for type unsigned int 0xffffffff rather than 0xffff. */
1659 : 590 : ctype = build_nonstandard_integer_type (TYPE_PRECISION (itype), 0);
1660 : 590 : vecctype = get_vectype_for_scalar_type (vinfo, ctype);
1661 : : }
1662 : :
1663 : 294607 : code_helper dummy_code;
1664 : 294607 : int dummy_int;
1665 : 294607 : auto_vec<tree> dummy_vec;
1666 : 294607 : if (!vectype
1667 : 294607 : || !vecitype
1668 : 231879 : || !vecctype
1669 : 526486 : || !supportable_widening_operation (vinfo, wide_code, last_stmt_info,
1670 : : vecitype, vectype,
1671 : : &dummy_code, &dummy_code,
1672 : : &dummy_int, &dummy_vec))
1673 : 199565 : return NULL;
1674 : :
1675 : 95042 : *type_out = get_vectype_for_scalar_type (vinfo, type);
1676 : 95042 : if (!*type_out)
1677 : : return NULL;
1678 : :
1679 : 95042 : tree oprnd[2];
1680 : 95042 : vect_convert_inputs (vinfo, last_stmt_info,
1681 : : 2, oprnd, half_type, unprom, vectype);
1682 : :
1683 : 95042 : tree var = vect_recog_temp_ssa_var (itype, NULL);
1684 : 95042 : gimple *pattern_stmt = vect_gimple_build (var, wide_code, oprnd[0], oprnd[1]);
1685 : :
1686 : 95042 : if (vecctype != vecitype)
1687 : 0 : pattern_stmt = vect_convert_output (vinfo, last_stmt_info, ctype,
1688 : : pattern_stmt, vecitype);
1689 : :
1690 : 95042 : return vect_convert_output (vinfo, last_stmt_info,
1691 : 95042 : type, pattern_stmt, vecctype);
1692 : 294607 : }
1693 : :
1694 : : /* Try to detect multiplication on widened inputs, converting MULT_EXPR
1695 : : to WIDEN_MULT_EXPR. See vect_recog_widen_op_pattern for details. */
1696 : :
1697 : : static gimple *
1698 : 29799617 : vect_recog_widen_mult_pattern (vec_info *vinfo, stmt_vec_info last_stmt_info,
1699 : : tree *type_out)
1700 : : {
1701 : 29799617 : return vect_recog_widen_op_pattern (vinfo, last_stmt_info, type_out,
1702 : 29799617 : MULT_EXPR, WIDEN_MULT_EXPR, false,
1703 : 29799617 : "vect_recog_widen_mult_pattern");
1704 : : }
1705 : :
1706 : : /* Try to detect addition on widened inputs, converting PLUS_EXPR
1707 : : to IFN_VEC_WIDEN_PLUS. See vect_recog_widen_op_pattern for details. */
1708 : :
1709 : : static gimple *
1710 : 30018576 : vect_recog_widen_plus_pattern (vec_info *vinfo, stmt_vec_info last_stmt_info,
1711 : : tree *type_out)
1712 : : {
1713 : 30018576 : return vect_recog_widen_op_pattern (vinfo, last_stmt_info, type_out,
1714 : 30018576 : PLUS_EXPR, IFN_VEC_WIDEN_PLUS,
1715 : 30018576 : false, "vect_recog_widen_plus_pattern");
1716 : : }
1717 : :
1718 : : /* Try to detect subtraction on widened inputs, converting MINUS_EXPR
1719 : : to IFN_VEC_WIDEN_MINUS. See vect_recog_widen_op_pattern for details. */
1720 : : static gimple *
1721 : 30018576 : vect_recog_widen_minus_pattern (vec_info *vinfo, stmt_vec_info last_stmt_info,
1722 : : tree *type_out)
1723 : : {
1724 : 30018576 : return vect_recog_widen_op_pattern (vinfo, last_stmt_info, type_out,
1725 : 30018576 : MINUS_EXPR, IFN_VEC_WIDEN_MINUS,
1726 : 30018576 : false, "vect_recog_widen_minus_pattern");
1727 : : }
1728 : :
1729 : : /* Try to detect abd on widened inputs, converting IFN_ABD
1730 : : to IFN_VEC_WIDEN_ABD. */
1731 : : static gimple *
1732 : 30018576 : vect_recog_widen_abd_pattern (vec_info *vinfo, stmt_vec_info stmt_vinfo,
1733 : : tree *type_out)
1734 : : {
1735 : 30018576 : gassign *last_stmt = dyn_cast <gassign *> (STMT_VINFO_STMT (stmt_vinfo));
1736 : 27879595 : if (!last_stmt || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (last_stmt)))
1737 : : return NULL;
1738 : :
1739 : 2816301 : tree last_rhs = gimple_assign_rhs1 (last_stmt);
1740 : :
1741 : 2816301 : tree in_type = TREE_TYPE (last_rhs);
1742 : 2816301 : tree out_type = TREE_TYPE (gimple_assign_lhs (last_stmt));
1743 : 2816301 : if (!INTEGRAL_TYPE_P (in_type)
1744 : 2527075 : || !INTEGRAL_TYPE_P (out_type)
1745 : 2432523 : || TYPE_PRECISION (in_type) * 2 != TYPE_PRECISION (out_type)
1746 : 3393758 : || !TYPE_UNSIGNED (in_type))
1747 : : return NULL;
1748 : :
1749 : 204873 : vect_unpromoted_value unprom;
1750 : 204873 : tree op = vect_look_through_possible_promotion (vinfo, last_rhs, &unprom);
1751 : 204873 : if (!op || TYPE_PRECISION (TREE_TYPE (op)) != TYPE_PRECISION (in_type))
1752 : : return NULL;
1753 : :
1754 : 204298 : stmt_vec_info abd_pattern_vinfo = vect_get_internal_def (vinfo, op);
1755 : 204298 : if (!abd_pattern_vinfo)
1756 : : return NULL;
1757 : :
1758 : 30027529 : gcall *abd_stmt = dyn_cast <gcall *> (STMT_VINFO_STMT (abd_pattern_vinfo));
1759 : 8953 : if (!abd_stmt
1760 : 8953 : || !gimple_call_internal_p (abd_stmt)
1761 : 246 : || gimple_call_internal_fn (abd_stmt) != IFN_ABD)
1762 : : return NULL;
1763 : :
1764 : 0 : tree vectype_in = get_vectype_for_scalar_type (vinfo, in_type);
1765 : 0 : tree vectype_out = get_vectype_for_scalar_type (vinfo, out_type);
1766 : :
1767 : 0 : code_helper dummy_code;
1768 : 0 : int dummy_int;
1769 : 0 : auto_vec<tree> dummy_vec;
1770 : 0 : if (!supportable_widening_operation (vinfo, IFN_VEC_WIDEN_ABD, stmt_vinfo,
1771 : : vectype_out, vectype_in,
1772 : : &dummy_code, &dummy_code,
1773 : : &dummy_int, &dummy_vec))
1774 : : return NULL;
1775 : :
1776 : 0 : vect_pattern_detected ("vect_recog_widen_abd_pattern", last_stmt);
1777 : :
1778 : 0 : *type_out = vectype_out;
1779 : :
1780 : 0 : tree abd_oprnd0 = gimple_call_arg (abd_stmt, 0);
1781 : 0 : tree abd_oprnd1 = gimple_call_arg (abd_stmt, 1);
1782 : 0 : tree widen_abd_result = vect_recog_temp_ssa_var (out_type, NULL);
1783 : 0 : gcall *widen_abd_stmt = gimple_build_call_internal (IFN_VEC_WIDEN_ABD, 2,
1784 : : abd_oprnd0, abd_oprnd1);
1785 : 0 : gimple_call_set_lhs (widen_abd_stmt, widen_abd_result);
1786 : 0 : gimple_set_location (widen_abd_stmt, gimple_location (last_stmt));
1787 : 0 : return widen_abd_stmt;
1788 : 0 : }
1789 : :
1790 : : /* Function vect_recog_ctz_ffs_pattern
1791 : :
1792 : : Try to find the following pattern:
1793 : :
1794 : : TYPE1 A;
1795 : : TYPE1 B;
1796 : :
1797 : : B = __builtin_ctz{,l,ll} (A);
1798 : :
1799 : : or
1800 : :
1801 : : B = __builtin_ffs{,l,ll} (A);
1802 : :
1803 : : Input:
1804 : :
1805 : : * STMT_VINFO: The stmt from which the pattern search begins.
1806 : : here it starts with B = __builtin_* (A);
1807 : :
1808 : : Output:
1809 : :
1810 : : * TYPE_OUT: The vector type of the output of this pattern.
1811 : :
1812 : : * Return value: A new stmt that will be used to replace the sequence of
1813 : : stmts that constitute the pattern, using clz or popcount builtins. */
1814 : :
1815 : : static gimple *
1816 : 29776073 : vect_recog_ctz_ffs_pattern (vec_info *vinfo, stmt_vec_info stmt_vinfo,
1817 : : tree *type_out)
1818 : : {
1819 : 29776073 : gimple *call_stmt = stmt_vinfo->stmt;
1820 : 29776073 : gimple *pattern_stmt;
1821 : 29776073 : tree rhs_oprnd, rhs_type, lhs_oprnd, lhs_type, vec_type, vec_rhs_type;
1822 : 29776073 : tree new_var;
1823 : 29776073 : internal_fn ifn = IFN_LAST, ifnnew = IFN_LAST;
1824 : 29776073 : bool defined_at_zero = true, defined_at_zero_new = false;
1825 : 29776073 : int val = 0, val_new = 0, val_cmp = 0;
1826 : 29776073 : int prec;
1827 : 29776073 : int sub = 0, add = 0;
1828 : 29776073 : location_t loc;
1829 : :
1830 : 29776073 : if (!is_gimple_call (call_stmt))
1831 : : return NULL;
1832 : :
1833 : 3470090 : if (gimple_call_num_args (call_stmt) != 1
1834 : 3470090 : && gimple_call_num_args (call_stmt) != 2)
1835 : : return NULL;
1836 : :
1837 : 1899192 : rhs_oprnd = gimple_call_arg (call_stmt, 0);
1838 : 1899192 : rhs_type = TREE_TYPE (rhs_oprnd);
1839 : 1899192 : lhs_oprnd = gimple_call_lhs (call_stmt);
1840 : 1899192 : if (!lhs_oprnd)
1841 : : return NULL;
1842 : 943496 : lhs_type = TREE_TYPE (lhs_oprnd);
1843 : 943496 : if (!INTEGRAL_TYPE_P (lhs_type)
1844 : 326817 : || !INTEGRAL_TYPE_P (rhs_type)
1845 : 51161 : || !type_has_mode_precision_p (rhs_type)
1846 : 993083 : || TREE_CODE (rhs_oprnd) != SSA_NAME)
1847 : 905019 : return NULL;
1848 : :
1849 : 38477 : switch (gimple_call_combined_fn (call_stmt))
1850 : : {
1851 : 1121 : CASE_CFN_CTZ:
1852 : 1121 : ifn = IFN_CTZ;
1853 : 1121 : if (!gimple_call_internal_p (call_stmt)
1854 : 1121 : || gimple_call_num_args (call_stmt) != 2)
1855 : : defined_at_zero = false;
1856 : : else
1857 : 48 : val = tree_to_shwi (gimple_call_arg (call_stmt, 1));
1858 : : break;
1859 : : CASE_CFN_FFS:
1860 : : ifn = IFN_FFS;
1861 : : break;
1862 : : default:
1863 : : return NULL;
1864 : : }
1865 : :
1866 : 1281 : prec = TYPE_PRECISION (rhs_type);
1867 : 1281 : loc = gimple_location (call_stmt);
1868 : :
1869 : 1281 : vec_type = get_vectype_for_scalar_type (vinfo, lhs_type);
1870 : 1281 : if (!vec_type)
1871 : : return NULL;
1872 : :
1873 : 1275 : vec_rhs_type = get_vectype_for_scalar_type (vinfo, rhs_type);
1874 : 1275 : if (!vec_rhs_type)
1875 : : return NULL;
1876 : :
1877 : : /* Do it only if the backend doesn't have ctz<vector_mode>2 or
1878 : : ffs<vector_mode>2 pattern but does have clz<vector_mode>2 or
1879 : : popcount<vector_mode>2. */
1880 : 1047 : if (!vec_type
1881 : 1047 : || direct_internal_fn_supported_p (ifn, vec_rhs_type,
1882 : : OPTIMIZE_FOR_SPEED))
1883 : : return NULL;
1884 : :
1885 : 1047 : if (ifn == IFN_FFS
1886 : 1047 : && direct_internal_fn_supported_p (IFN_CTZ, vec_rhs_type,
1887 : : OPTIMIZE_FOR_SPEED))
1888 : : {
1889 : 0 : ifnnew = IFN_CTZ;
1890 : 0 : defined_at_zero_new
1891 : 0 : = CTZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (rhs_type),
1892 : : val_new) == 2;
1893 : : }
1894 : 1047 : else if (direct_internal_fn_supported_p (IFN_CLZ, vec_rhs_type,
1895 : : OPTIMIZE_FOR_SPEED))
1896 : : {
1897 : 88 : ifnnew = IFN_CLZ;
1898 : 88 : defined_at_zero_new
1899 : 88 : = CLZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (rhs_type),
1900 : : val_new) == 2;
1901 : : }
1902 : 88 : if ((ifnnew == IFN_LAST
1903 : 88 : || (defined_at_zero && !defined_at_zero_new))
1904 : 959 : && direct_internal_fn_supported_p (IFN_POPCOUNT, vec_rhs_type,
1905 : : OPTIMIZE_FOR_SPEED))
1906 : : {
1907 : : ifnnew = IFN_POPCOUNT;
1908 : : defined_at_zero_new = true;
1909 : : val_new = prec;
1910 : : }
1911 : 1011 : if (ifnnew == IFN_LAST)
1912 : : return NULL;
1913 : :
1914 : 124 : vect_pattern_detected ("vec_recog_ctz_ffs_pattern", call_stmt);
1915 : :
1916 : 124 : val_cmp = val_new;
1917 : 124 : if ((ifnnew == IFN_CLZ
1918 : 124 : && defined_at_zero
1919 : 60 : && defined_at_zero_new
1920 : 60 : && val == prec
1921 : 31 : && val_new == prec)
1922 : 93 : || (ifnnew == IFN_POPCOUNT && ifn == IFN_CTZ))
1923 : : {
1924 : : /* .CTZ (X) = PREC - .CLZ ((X - 1) & ~X)
1925 : : .CTZ (X) = .POPCOUNT ((X - 1) & ~X). */
1926 : : if (ifnnew == IFN_CLZ)
1927 : : sub = prec;
1928 : 56 : val_cmp = prec;
1929 : :
1930 : 56 : if (!TYPE_UNSIGNED (rhs_type))
1931 : : {
1932 : 12 : rhs_type = unsigned_type_for (rhs_type);
1933 : 12 : vec_rhs_type = get_vectype_for_scalar_type (vinfo, rhs_type);
1934 : 12 : new_var = vect_recog_temp_ssa_var (rhs_type, NULL);
1935 : 12 : pattern_stmt = gimple_build_assign (new_var, NOP_EXPR, rhs_oprnd);
1936 : 12 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt,
1937 : : vec_rhs_type);
1938 : 12 : rhs_oprnd = new_var;
1939 : : }
1940 : :
1941 : 56 : tree m1 = vect_recog_temp_ssa_var (rhs_type, NULL);
1942 : 56 : pattern_stmt = gimple_build_assign (m1, PLUS_EXPR, rhs_oprnd,
1943 : : build_int_cst (rhs_type, -1));
1944 : 56 : gimple_set_location (pattern_stmt, loc);
1945 : 56 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_rhs_type);
1946 : :
1947 : 56 : new_var = vect_recog_temp_ssa_var (rhs_type, NULL);
1948 : 56 : pattern_stmt = gimple_build_assign (new_var, BIT_NOT_EXPR, rhs_oprnd);
1949 : 56 : gimple_set_location (pattern_stmt, loc);
1950 : 56 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_rhs_type);
1951 : 56 : rhs_oprnd = new_var;
1952 : :
1953 : 56 : new_var = vect_recog_temp_ssa_var (rhs_type, NULL);
1954 : 56 : pattern_stmt = gimple_build_assign (new_var, BIT_AND_EXPR,
1955 : : m1, rhs_oprnd);
1956 : 56 : gimple_set_location (pattern_stmt, loc);
1957 : 56 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_rhs_type);
1958 : 56 : rhs_oprnd = new_var;
1959 : 56 : }
1960 : 68 : else if (ifnnew == IFN_CLZ)
1961 : : {
1962 : : /* .CTZ (X) = (PREC - 1) - .CLZ (X & -X)
1963 : : .FFS (X) = PREC - .CLZ (X & -X). */
1964 : 57 : sub = prec - (ifn == IFN_CTZ);
1965 : 57 : val_cmp = sub - val_new;
1966 : :
1967 : 57 : tree neg = vect_recog_temp_ssa_var (rhs_type, NULL);
1968 : 57 : pattern_stmt = gimple_build_assign (neg, NEGATE_EXPR, rhs_oprnd);
1969 : 57 : gimple_set_location (pattern_stmt, loc);
1970 : 57 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_rhs_type);
1971 : :
1972 : 57 : new_var = vect_recog_temp_ssa_var (rhs_type, NULL);
1973 : 57 : pattern_stmt = gimple_build_assign (new_var, BIT_AND_EXPR,
1974 : : rhs_oprnd, neg);
1975 : 57 : gimple_set_location (pattern_stmt, loc);
1976 : 57 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_rhs_type);
1977 : 57 : rhs_oprnd = new_var;
1978 : : }
1979 : 11 : else if (ifnnew == IFN_POPCOUNT)
1980 : : {
1981 : : /* .CTZ (X) = PREC - .POPCOUNT (X | -X)
1982 : : .FFS (X) = (PREC + 1) - .POPCOUNT (X | -X). */
1983 : 11 : sub = prec + (ifn == IFN_FFS);
1984 : 11 : val_cmp = sub;
1985 : :
1986 : 11 : tree neg = vect_recog_temp_ssa_var (rhs_type, NULL);
1987 : 11 : pattern_stmt = gimple_build_assign (neg, NEGATE_EXPR, rhs_oprnd);
1988 : 11 : gimple_set_location (pattern_stmt, loc);
1989 : 11 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_rhs_type);
1990 : :
1991 : 11 : new_var = vect_recog_temp_ssa_var (rhs_type, NULL);
1992 : 11 : pattern_stmt = gimple_build_assign (new_var, BIT_IOR_EXPR,
1993 : : rhs_oprnd, neg);
1994 : 11 : gimple_set_location (pattern_stmt, loc);
1995 : 11 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_rhs_type);
1996 : 11 : rhs_oprnd = new_var;
1997 : : }
1998 : 0 : else if (ifnnew == IFN_CTZ)
1999 : : {
2000 : : /* .FFS (X) = .CTZ (X) + 1. */
2001 : 0 : add = 1;
2002 : 0 : val_cmp++;
2003 : : }
2004 : :
2005 : : /* Create B = .IFNNEW (A). */
2006 : 124 : new_var = vect_recog_temp_ssa_var (lhs_type, NULL);
2007 : 124 : if ((ifnnew == IFN_CLZ || ifnnew == IFN_CTZ) && defined_at_zero_new)
2008 : 88 : pattern_stmt
2009 : 88 : = gimple_build_call_internal (ifnnew, 2, rhs_oprnd,
2010 : : build_int_cst (integer_type_node,
2011 : 88 : val_new));
2012 : : else
2013 : 36 : pattern_stmt = gimple_build_call_internal (ifnnew, 1, rhs_oprnd);
2014 : 124 : gimple_call_set_lhs (pattern_stmt, new_var);
2015 : 124 : gimple_set_location (pattern_stmt, loc);
2016 : 124 : *type_out = vec_type;
2017 : :
2018 : 124 : if (sub)
2019 : : {
2020 : 99 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_type);
2021 : 99 : tree ret_var = vect_recog_temp_ssa_var (lhs_type, NULL);
2022 : 99 : pattern_stmt = gimple_build_assign (ret_var, MINUS_EXPR,
2023 : 99 : build_int_cst (lhs_type, sub),
2024 : : new_var);
2025 : 99 : gimple_set_location (pattern_stmt, loc);
2026 : 99 : new_var = ret_var;
2027 : : }
2028 : 25 : else if (add)
2029 : : {
2030 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_type);
2031 : 0 : tree ret_var = vect_recog_temp_ssa_var (lhs_type, NULL);
2032 : 0 : pattern_stmt = gimple_build_assign (ret_var, PLUS_EXPR, new_var,
2033 : 0 : build_int_cst (lhs_type, add));
2034 : 0 : gimple_set_location (pattern_stmt, loc);
2035 : 0 : new_var = ret_var;
2036 : : }
2037 : :
2038 : 124 : if (defined_at_zero
2039 : 88 : && (!defined_at_zero_new || val != val_cmp))
2040 : : {
2041 : 11 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_type);
2042 : 11 : tree ret_var = vect_recog_temp_ssa_var (lhs_type, NULL);
2043 : 11 : rhs_oprnd = gimple_call_arg (call_stmt, 0);
2044 : 11 : rhs_type = TREE_TYPE (rhs_oprnd);
2045 : 11 : tree cmp = vect_recog_temp_ssa_var (boolean_type_node, NULL);
2046 : 11 : pattern_stmt = gimple_build_assign (cmp, NE_EXPR, rhs_oprnd,
2047 : : build_zero_cst (rhs_type));
2048 : 11 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt,
2049 : : truth_type_for (vec_type), rhs_type);
2050 : 11 : pattern_stmt = gimple_build_assign (ret_var, COND_EXPR, cmp,
2051 : : new_var,
2052 : 11 : build_int_cst (lhs_type, val));
2053 : : }
2054 : :
2055 : 124 : if (dump_enabled_p ())
2056 : 36 : dump_printf_loc (MSG_NOTE, vect_location,
2057 : : "created pattern stmt: %G", pattern_stmt);
2058 : :
2059 : : return pattern_stmt;
2060 : : }
2061 : :
2062 : : /* Function vect_recog_popcount_clz_ctz_ffs_pattern
2063 : :
2064 : : Try to find the following pattern:
2065 : :
2066 : : UTYPE1 A;
2067 : : TYPE1 B;
2068 : : UTYPE2 temp_in;
2069 : : TYPE3 temp_out;
2070 : : temp_in = (UTYPE2)A;
2071 : :
2072 : : temp_out = __builtin_popcount{,l,ll} (temp_in);
2073 : : B = (TYPE1) temp_out;
2074 : :
2075 : : TYPE2 may or may not be equal to TYPE3.
2076 : : i.e. TYPE2 is equal to TYPE3 for __builtin_popcount
2077 : : i.e. TYPE2 is not equal to TYPE3 for __builtin_popcountll
2078 : :
2079 : : Input:
2080 : :
2081 : : * STMT_VINFO: The stmt from which the pattern search begins.
2082 : : here it starts with B = (TYPE1) temp_out;
2083 : :
2084 : : Output:
2085 : :
2086 : : * TYPE_OUT: The vector type of the output of this pattern.
2087 : :
2088 : : * Return value: A new stmt that will be used to replace the sequence of
2089 : : stmts that constitute the pattern. In this case it will be:
2090 : : B = .POPCOUNT (A);
2091 : :
2092 : : Similarly for clz, ctz and ffs.
2093 : : */
2094 : :
2095 : : static gimple *
2096 : 29775953 : vect_recog_popcount_clz_ctz_ffs_pattern (vec_info *vinfo,
2097 : : stmt_vec_info stmt_vinfo,
2098 : : tree *type_out)
2099 : : {
2100 : 29775953 : gassign *last_stmt = dyn_cast <gassign *> (stmt_vinfo->stmt);
2101 : 20346686 : gimple *call_stmt, *pattern_stmt;
2102 : 20346686 : tree rhs_oprnd, rhs_origin, lhs_oprnd, lhs_type, vec_type, new_var;
2103 : 50122480 : internal_fn ifn = IFN_LAST;
2104 : 29775794 : int addend = 0;
2105 : :
2106 : : /* Find B = (TYPE1) temp_out. */
2107 : 20346686 : if (!last_stmt)
2108 : : return NULL;
2109 : 20346686 : tree_code code = gimple_assign_rhs_code (last_stmt);
2110 : 20346686 : if (!CONVERT_EXPR_CODE_P (code))
2111 : : return NULL;
2112 : :
2113 : 2780360 : lhs_oprnd = gimple_assign_lhs (last_stmt);
2114 : 2780360 : lhs_type = TREE_TYPE (lhs_oprnd);
2115 : 2780360 : if (!INTEGRAL_TYPE_P (lhs_type))
2116 : : return NULL;
2117 : :
2118 : 2625509 : rhs_oprnd = gimple_assign_rhs1 (last_stmt);
2119 : 2625509 : if (TREE_CODE (rhs_oprnd) != SSA_NAME
2120 : 2625509 : || !has_single_use (rhs_oprnd))
2121 : : return NULL;
2122 : 1369222 : call_stmt = SSA_NAME_DEF_STMT (rhs_oprnd);
2123 : :
2124 : : /* Find temp_out = __builtin_popcount{,l,ll} (temp_in); */
2125 : 1369222 : if (!is_gimple_call (call_stmt))
2126 : : return NULL;
2127 : 100760 : switch (gimple_call_combined_fn (call_stmt))
2128 : : {
2129 : : int val;
2130 : : CASE_CFN_POPCOUNT:
2131 : : ifn = IFN_POPCOUNT;
2132 : : break;
2133 : 3348 : CASE_CFN_CLZ:
2134 : 3348 : ifn = IFN_CLZ;
2135 : : /* Punt if call result is unsigned and defined value at zero
2136 : : is negative, as the negative value doesn't extend correctly. */
2137 : 3348 : if (TYPE_UNSIGNED (TREE_TYPE (rhs_oprnd))
2138 : 0 : && gimple_call_internal_p (call_stmt)
2139 : 3348 : && CLZ_DEFINED_VALUE_AT_ZERO
2140 : : (SCALAR_INT_TYPE_MODE (TREE_TYPE (rhs_oprnd)), val) == 2
2141 : 3348 : && val < 0)
2142 : : return NULL;
2143 : : break;
2144 : 561 : CASE_CFN_CTZ:
2145 : 561 : ifn = IFN_CTZ;
2146 : : /* Punt if call result is unsigned and defined value at zero
2147 : : is negative, as the negative value doesn't extend correctly. */
2148 : 561 : if (TYPE_UNSIGNED (TREE_TYPE (rhs_oprnd))
2149 : 0 : && gimple_call_internal_p (call_stmt)
2150 : 561 : && CTZ_DEFINED_VALUE_AT_ZERO
2151 : : (SCALAR_INT_TYPE_MODE (TREE_TYPE (rhs_oprnd)), val) == 2
2152 : 561 : && val < 0)
2153 : : return NULL;
2154 : : break;
2155 : 18 : CASE_CFN_FFS:
2156 : 18 : ifn = IFN_FFS;
2157 : 18 : break;
2158 : : default:
2159 : : return NULL;
2160 : : }
2161 : :
2162 : 4165 : if (gimple_call_num_args (call_stmt) != 1
2163 : 4165 : && gimple_call_num_args (call_stmt) != 2)
2164 : : return NULL;
2165 : :
2166 : 4165 : rhs_oprnd = gimple_call_arg (call_stmt, 0);
2167 : 4165 : vect_unpromoted_value unprom_diff;
2168 : 4165 : rhs_origin
2169 : 4165 : = vect_look_through_possible_promotion (vinfo, rhs_oprnd, &unprom_diff);
2170 : :
2171 : 4165 : if (!rhs_origin)
2172 : : return NULL;
2173 : :
2174 : : /* Input and output of .POPCOUNT should be same-precision integer. */
2175 : 4165 : if (TYPE_PRECISION (unprom_diff.type) != TYPE_PRECISION (lhs_type))
2176 : : return NULL;
2177 : :
2178 : : /* Also A should be unsigned or same precision as temp_in, otherwise
2179 : : different builtins/internal functions have different behaviors. */
2180 : 1640 : if (TYPE_PRECISION (unprom_diff.type)
2181 : 1640 : != TYPE_PRECISION (TREE_TYPE (rhs_oprnd)))
2182 : 158 : switch (ifn)
2183 : : {
2184 : 79 : case IFN_POPCOUNT:
2185 : : /* For popcount require zero extension, which doesn't add any
2186 : : further bits to the count. */
2187 : 79 : if (!TYPE_UNSIGNED (unprom_diff.type))
2188 : : return NULL;
2189 : : break;
2190 : 61 : case IFN_CLZ:
2191 : : /* clzll (x) == clz (x) + 32 for unsigned x != 0, so ok
2192 : : if it is undefined at zero or if it matches also for the
2193 : : defined value there. */
2194 : 61 : if (!TYPE_UNSIGNED (unprom_diff.type))
2195 : : return NULL;
2196 : 61 : if (!type_has_mode_precision_p (lhs_type)
2197 : 61 : || !type_has_mode_precision_p (TREE_TYPE (rhs_oprnd)))
2198 : 0 : return NULL;
2199 : 61 : addend = (TYPE_PRECISION (TREE_TYPE (rhs_oprnd))
2200 : 61 : - TYPE_PRECISION (lhs_type));
2201 : 61 : if (gimple_call_internal_p (call_stmt)
2202 : 61 : && gimple_call_num_args (call_stmt) == 2)
2203 : : {
2204 : 0 : int val1, val2;
2205 : 0 : val1 = tree_to_shwi (gimple_call_arg (call_stmt, 1));
2206 : 0 : int d2
2207 : 0 : = CLZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (lhs_type),
2208 : : val2);
2209 : 0 : if (d2 != 2 || val1 != val2 + addend)
2210 : : return NULL;
2211 : : }
2212 : : break;
2213 : 13 : case IFN_CTZ:
2214 : : /* ctzll (x) == ctz (x) for unsigned or signed x != 0, so ok
2215 : : if it is undefined at zero or if it matches also for the
2216 : : defined value there. */
2217 : 13 : if (gimple_call_internal_p (call_stmt)
2218 : 13 : && gimple_call_num_args (call_stmt) == 2)
2219 : : {
2220 : 0 : int val1, val2;
2221 : 0 : val1 = tree_to_shwi (gimple_call_arg (call_stmt, 1));
2222 : 0 : int d2
2223 : 0 : = CTZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (lhs_type),
2224 : : val2);
2225 : 0 : if (d2 != 2 || val1 != val2)
2226 : : return NULL;
2227 : : }
2228 : : break;
2229 : : case IFN_FFS:
2230 : : /* ffsll (x) == ffs (x) for unsigned or signed x. */
2231 : : break;
2232 : 0 : default:
2233 : 0 : gcc_unreachable ();
2234 : : }
2235 : :
2236 : 1640 : vec_type = get_vectype_for_scalar_type (vinfo, lhs_type);
2237 : : /* Do it only if the backend has popcount<vector_mode>2 etc. pattern. */
2238 : 1640 : if (!vec_type)
2239 : : return NULL;
2240 : :
2241 : 1517 : bool supported
2242 : 1517 : = direct_internal_fn_supported_p (ifn, vec_type, OPTIMIZE_FOR_SPEED);
2243 : 1517 : if (!supported)
2244 : 1414 : switch (ifn)
2245 : : {
2246 : : case IFN_POPCOUNT:
2247 : : case IFN_CLZ:
2248 : : return NULL;
2249 : 18 : case IFN_FFS:
2250 : : /* vect_recog_ctz_ffs_pattern can implement ffs using ctz. */
2251 : 18 : if (direct_internal_fn_supported_p (IFN_CTZ, vec_type,
2252 : : OPTIMIZE_FOR_SPEED))
2253 : : break;
2254 : : /* FALLTHRU */
2255 : 359 : case IFN_CTZ:
2256 : : /* vect_recog_ctz_ffs_pattern can implement ffs or ctz using
2257 : : clz or popcount. */
2258 : 359 : if (direct_internal_fn_supported_p (IFN_CLZ, vec_type,
2259 : : OPTIMIZE_FOR_SPEED))
2260 : : break;
2261 : 325 : if (direct_internal_fn_supported_p (IFN_POPCOUNT, vec_type,
2262 : : OPTIMIZE_FOR_SPEED))
2263 : : break;
2264 : : return NULL;
2265 : 0 : default:
2266 : 0 : gcc_unreachable ();
2267 : : }
2268 : :
2269 : 159 : vect_pattern_detected ("vec_recog_popcount_clz_ctz_ffs_pattern",
2270 : : call_stmt);
2271 : :
2272 : : /* Create B = .POPCOUNT (A). */
2273 : 159 : new_var = vect_recog_temp_ssa_var (lhs_type, NULL);
2274 : 159 : tree arg2 = NULL_TREE;
2275 : 159 : int val;
2276 : 159 : if (ifn == IFN_CLZ
2277 : 191 : && CLZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (lhs_type),
2278 : : val) == 2)
2279 : 30 : arg2 = build_int_cst (integer_type_node, val);
2280 : 129 : else if (ifn == IFN_CTZ
2281 : 167 : && CTZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (lhs_type),
2282 : : val) == 2)
2283 : 38 : arg2 = build_int_cst (integer_type_node, val);
2284 : 159 : if (arg2)
2285 : 68 : pattern_stmt = gimple_build_call_internal (ifn, 2, unprom_diff.op, arg2);
2286 : : else
2287 : 91 : pattern_stmt = gimple_build_call_internal (ifn, 1, unprom_diff.op);
2288 : 159 : gimple_call_set_lhs (pattern_stmt, new_var);
2289 : 159 : gimple_set_location (pattern_stmt, gimple_location (last_stmt));
2290 : 159 : *type_out = vec_type;
2291 : :
2292 : 159 : if (dump_enabled_p ())
2293 : 24 : dump_printf_loc (MSG_NOTE, vect_location,
2294 : : "created pattern stmt: %G", pattern_stmt);
2295 : :
2296 : 159 : if (addend)
2297 : : {
2298 : 6 : gcc_assert (supported);
2299 : 6 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_type);
2300 : 6 : tree ret_var = vect_recog_temp_ssa_var (lhs_type, NULL);
2301 : 6 : pattern_stmt = gimple_build_assign (ret_var, PLUS_EXPR, new_var,
2302 : 6 : build_int_cst (lhs_type, addend));
2303 : : }
2304 : 153 : else if (!supported)
2305 : : {
2306 : 56 : stmt_vec_info new_stmt_info = vinfo->add_stmt (pattern_stmt);
2307 : 56 : STMT_VINFO_VECTYPE (new_stmt_info) = vec_type;
2308 : 56 : pattern_stmt
2309 : 56 : = vect_recog_ctz_ffs_pattern (vinfo, new_stmt_info, type_out);
2310 : 56 : if (pattern_stmt == NULL)
2311 : : return NULL;
2312 : 56 : if (gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (new_stmt_info))
2313 : : {
2314 : 56 : gimple_seq *pseq = &STMT_VINFO_PATTERN_DEF_SEQ (stmt_vinfo);
2315 : 56 : gimple_seq_add_seq_without_update (pseq, seq);
2316 : : }
2317 : : }
2318 : : return pattern_stmt;
2319 : : }
2320 : :
2321 : : /* Function vect_recog_pow_pattern
2322 : :
2323 : : Try to find the following pattern:
2324 : :
2325 : : x = POW (y, N);
2326 : :
2327 : : with POW being one of pow, powf, powi, powif and N being
2328 : : either 2 or 0.5.
2329 : :
2330 : : Input:
2331 : :
2332 : : * STMT_VINFO: The stmt from which the pattern search begins.
2333 : :
2334 : : Output:
2335 : :
2336 : : * TYPE_OUT: The type of the output of this pattern.
2337 : :
2338 : : * Return value: A new stmt that will be used to replace the sequence of
2339 : : stmts that constitute the pattern. In this case it will be:
2340 : : x = x * x
2341 : : or
2342 : : x = sqrt (x)
2343 : : */
2344 : :
2345 : : static gimple *
2346 : 29775962 : vect_recog_pow_pattern (vec_info *vinfo,
2347 : : stmt_vec_info stmt_vinfo, tree *type_out)
2348 : : {
2349 : 29775962 : gimple *last_stmt = stmt_vinfo->stmt;
2350 : 29775962 : tree base, exp;
2351 : 29775962 : gimple *stmt;
2352 : 29775962 : tree var;
2353 : :
2354 : 29775962 : if (!is_gimple_call (last_stmt) || gimple_call_lhs (last_stmt) == NULL)
2355 : : return NULL;
2356 : :
2357 : 1480971 : switch (gimple_call_combined_fn (last_stmt))
2358 : : {
2359 : 248 : CASE_CFN_POW:
2360 : 248 : CASE_CFN_POWI:
2361 : 248 : break;
2362 : :
2363 : : default:
2364 : : return NULL;
2365 : : }
2366 : :
2367 : 248 : base = gimple_call_arg (last_stmt, 0);
2368 : 248 : exp = gimple_call_arg (last_stmt, 1);
2369 : 248 : if (TREE_CODE (exp) != REAL_CST
2370 : 226 : && TREE_CODE (exp) != INTEGER_CST)
2371 : : {
2372 : 226 : if (flag_unsafe_math_optimizations
2373 : 26 : && TREE_CODE (base) == REAL_CST
2374 : 228 : && gimple_call_builtin_p (last_stmt, BUILT_IN_NORMAL))
2375 : : {
2376 : 2 : combined_fn log_cfn;
2377 : 2 : built_in_function exp_bfn;
2378 : 2 : switch (DECL_FUNCTION_CODE (gimple_call_fndecl (last_stmt)))
2379 : : {
2380 : : case BUILT_IN_POW:
2381 : : log_cfn = CFN_BUILT_IN_LOG;
2382 : : exp_bfn = BUILT_IN_EXP;
2383 : : break;
2384 : 0 : case BUILT_IN_POWF:
2385 : 0 : log_cfn = CFN_BUILT_IN_LOGF;
2386 : 0 : exp_bfn = BUILT_IN_EXPF;
2387 : 0 : break;
2388 : 0 : case BUILT_IN_POWL:
2389 : 0 : log_cfn = CFN_BUILT_IN_LOGL;
2390 : 0 : exp_bfn = BUILT_IN_EXPL;
2391 : 0 : break;
2392 : : default:
2393 : : return NULL;
2394 : : }
2395 : 2 : tree logc = fold_const_call (log_cfn, TREE_TYPE (base), base);
2396 : 2 : tree exp_decl = builtin_decl_implicit (exp_bfn);
2397 : : /* Optimize pow (C, x) as exp (log (C) * x). Normally match.pd
2398 : : does that, but if C is a power of 2, we want to use
2399 : : exp2 (log2 (C) * x) in the non-vectorized version, but for
2400 : : vectorization we don't have vectorized exp2. */
2401 : 2 : if (logc
2402 : 2 : && TREE_CODE (logc) == REAL_CST
2403 : 2 : && exp_decl
2404 : 4 : && lookup_attribute ("omp declare simd",
2405 : 2 : DECL_ATTRIBUTES (exp_decl)))
2406 : : {
2407 : 2 : cgraph_node *node = cgraph_node::get_create (exp_decl);
2408 : 2 : if (node->simd_clones == NULL)
2409 : : {
2410 : 2 : if (targetm.simd_clone.compute_vecsize_and_simdlen == NULL
2411 : 2 : || node->definition)
2412 : : return NULL;
2413 : 2 : expand_simd_clones (node);
2414 : 2 : if (node->simd_clones == NULL)
2415 : : return NULL;
2416 : : }
2417 : 2 : *type_out = get_vectype_for_scalar_type (vinfo, TREE_TYPE (base));
2418 : 2 : if (!*type_out)
2419 : : return NULL;
2420 : 2 : tree def = vect_recog_temp_ssa_var (TREE_TYPE (base), NULL);
2421 : 2 : gimple *g = gimple_build_assign (def, MULT_EXPR, exp, logc);
2422 : 2 : append_pattern_def_seq (vinfo, stmt_vinfo, g);
2423 : 2 : tree res = vect_recog_temp_ssa_var (TREE_TYPE (base), NULL);
2424 : 2 : g = gimple_build_call (exp_decl, 1, def);
2425 : 2 : gimple_call_set_lhs (g, res);
2426 : 2 : return g;
2427 : : }
2428 : : }
2429 : :
2430 : 224 : return NULL;
2431 : : }
2432 : :
2433 : : /* We now have a pow or powi builtin function call with a constant
2434 : : exponent. */
2435 : :
2436 : : /* Catch squaring. */
2437 : 22 : if ((tree_fits_shwi_p (exp)
2438 : 0 : && tree_to_shwi (exp) == 2)
2439 : 22 : || (TREE_CODE (exp) == REAL_CST
2440 : 22 : && real_equal (&TREE_REAL_CST (exp), &dconst2)))
2441 : : {
2442 : 0 : if (!vect_supportable_direct_optab_p (vinfo, TREE_TYPE (base), MULT_EXPR,
2443 : 0 : TREE_TYPE (base), type_out))
2444 : : return NULL;
2445 : :
2446 : 0 : var = vect_recog_temp_ssa_var (TREE_TYPE (base), NULL);
2447 : 0 : stmt = gimple_build_assign (var, MULT_EXPR, base, base);
2448 : 0 : return stmt;
2449 : : }
2450 : :
2451 : : /* Catch square root. */
2452 : 22 : if (TREE_CODE (exp) == REAL_CST
2453 : 22 : && real_equal (&TREE_REAL_CST (exp), &dconsthalf))
2454 : : {
2455 : 11 : *type_out = get_vectype_for_scalar_type (vinfo, TREE_TYPE (base));
2456 : 11 : if (*type_out
2457 : 11 : && direct_internal_fn_supported_p (IFN_SQRT, *type_out,
2458 : : OPTIMIZE_FOR_SPEED))
2459 : : {
2460 : 9 : gcall *stmt = gimple_build_call_internal (IFN_SQRT, 1, base);
2461 : 9 : var = vect_recog_temp_ssa_var (TREE_TYPE (base), stmt);
2462 : 9 : gimple_call_set_lhs (stmt, var);
2463 : 9 : gimple_call_set_nothrow (stmt, true);
2464 : 9 : return stmt;
2465 : : }
2466 : : }
2467 : :
2468 : : return NULL;
2469 : : }
2470 : :
2471 : :
2472 : : /* Function vect_recog_widen_sum_pattern
2473 : :
2474 : : Try to find the following pattern:
2475 : :
2476 : : type x_t;
2477 : : TYPE x_T, sum = init;
2478 : : loop:
2479 : : sum_0 = phi <init, sum_1>
2480 : : S1 x_t = *p;
2481 : : S2 x_T = (TYPE) x_t;
2482 : : S3 sum_1 = x_T + sum_0;
2483 : :
2484 : : where type 'TYPE' is at least double the size of type 'type', i.e - we're
2485 : : summing elements of type 'type' into an accumulator of type 'TYPE'. This is
2486 : : a special case of a reduction computation.
2487 : :
2488 : : Input:
2489 : :
2490 : : * STMT_VINFO: The stmt from which the pattern search begins. In the example,
2491 : : when this function is called with S3, the pattern {S2,S3} will be detected.
2492 : :
2493 : : Output:
2494 : :
2495 : : * TYPE_OUT: The type of the output of this pattern.
2496 : :
2497 : : * Return value: A new stmt that will be used to replace the sequence of
2498 : : stmts that constitute the pattern. In this case it will be:
2499 : : WIDEN_SUM <x_t, sum_0>
2500 : :
2501 : : Note: The widening-sum idiom is a widening reduction pattern that is
2502 : : vectorized without preserving all the intermediate results. It
2503 : : produces only N/2 (widened) results (by summing up pairs of
2504 : : intermediate results) rather than all N results. Therefore, we
2505 : : cannot allow this pattern when we want to get all the results and in
2506 : : the correct order (as is the case when this computation is in an
2507 : : inner-loop nested in an outer-loop that us being vectorized). */
2508 : :
2509 : : static gimple *
2510 : 29775962 : vect_recog_widen_sum_pattern (vec_info *vinfo,
2511 : : stmt_vec_info stmt_vinfo, tree *type_out)
2512 : : {
2513 : 29775962 : gimple *last_stmt = stmt_vinfo->stmt;
2514 : 29775962 : tree oprnd0, oprnd1;
2515 : 29775962 : tree type;
2516 : 29775962 : gimple *pattern_stmt;
2517 : 29775962 : tree var;
2518 : :
2519 : : /* Look for the following pattern
2520 : : DX = (TYPE) X;
2521 : : sum_1 = DX + sum_0;
2522 : : In which DX is at least double the size of X, and sum_1 has been
2523 : : recognized as a reduction variable.
2524 : : */
2525 : :
2526 : : /* Starting from LAST_STMT, follow the defs of its uses in search
2527 : : of the above pattern. */
2528 : :
2529 : 29775962 : if (!vect_reassociating_reduction_p (vinfo, stmt_vinfo, PLUS_EXPR,
2530 : : &oprnd0, &oprnd1)
2531 : 33055 : || TREE_CODE (oprnd0) != SSA_NAME
2532 : 29808835 : || !vinfo->lookup_def (oprnd0))
2533 : 29743161 : return NULL;
2534 : :
2535 : 32801 : type = TREE_TYPE (gimple_get_lhs (last_stmt));
2536 : :
2537 : : /* So far so good. Since last_stmt was detected as a (summation) reduction,
2538 : : we know that oprnd1 is the reduction variable (defined by a loop-header
2539 : : phi), and oprnd0 is an ssa-name defined by a stmt in the loop body.
2540 : : Left to check that oprnd0 is defined by a cast from type 'type' to type
2541 : : 'TYPE'. */
2542 : :
2543 : 32801 : vect_unpromoted_value unprom0;
2544 : 32801 : if (!vect_look_through_possible_promotion (vinfo, oprnd0, &unprom0)
2545 : 32801 : || TYPE_PRECISION (unprom0.type) * 2 > TYPE_PRECISION (type))
2546 : : return NULL;
2547 : :
2548 : 1760 : vect_pattern_detected ("vect_recog_widen_sum_pattern", last_stmt);
2549 : :
2550 : 1760 : if (!vect_supportable_direct_optab_p (vinfo, type, WIDEN_SUM_EXPR,
2551 : : unprom0.type, type_out))
2552 : : return NULL;
2553 : :
2554 : 0 : var = vect_recog_temp_ssa_var (type, NULL);
2555 : 0 : pattern_stmt = gimple_build_assign (var, WIDEN_SUM_EXPR, unprom0.op, oprnd1);
2556 : :
2557 : 0 : return pattern_stmt;
2558 : : }
2559 : :
2560 : : /* Function vect_recog_bitfield_ref_pattern
2561 : :
2562 : : Try to find the following pattern:
2563 : :
2564 : : bf_value = BIT_FIELD_REF (container, bitsize, bitpos);
2565 : : result = (type_out) bf_value;
2566 : :
2567 : : or
2568 : :
2569 : : if (BIT_FIELD_REF (container, bitsize, bitpos) `cmp` <constant>)
2570 : :
2571 : : where type_out is a non-bitfield type, that is to say, it's precision matches
2572 : : 2^(TYPE_SIZE(type_out) - (TYPE_UNSIGNED (type_out) ? 1 : 2)).
2573 : :
2574 : : Input:
2575 : :
2576 : : * STMT_VINFO: The stmt from which the pattern search begins.
2577 : : here it starts with:
2578 : : result = (type_out) bf_value;
2579 : :
2580 : : or
2581 : :
2582 : : if (BIT_FIELD_REF (container, bitsize, bitpos) `cmp` <constant>)
2583 : :
2584 : : Output:
2585 : :
2586 : : * TYPE_OUT: The vector type of the output of this pattern.
2587 : :
2588 : : * Return value: A new stmt that will be used to replace the sequence of
2589 : : stmts that constitute the pattern. If the precision of type_out is bigger
2590 : : than the precision type of _1 we perform the widening before the shifting,
2591 : : since the new precision will be large enough to shift the value and moving
2592 : : widening operations up the statement chain enables the generation of
2593 : : widening loads. If we are widening and the operation after the pattern is
2594 : : an addition then we mask first and shift later, to enable the generation of
2595 : : shifting adds. In the case of narrowing we will always mask first, shift
2596 : : last and then perform a narrowing operation. This will enable the
2597 : : generation of narrowing shifts.
2598 : :
2599 : : Widening with mask first, shift later:
2600 : : container = (type_out) container;
2601 : : masked = container & (((1 << bitsize) - 1) << bitpos);
2602 : : result = masked >> bitpos;
2603 : :
2604 : : Widening with shift first, mask last:
2605 : : container = (type_out) container;
2606 : : shifted = container >> bitpos;
2607 : : result = shifted & ((1 << bitsize) - 1);
2608 : :
2609 : : Narrowing:
2610 : : masked = container & (((1 << bitsize) - 1) << bitpos);
2611 : : result = masked >> bitpos;
2612 : : result = (type_out) result;
2613 : :
2614 : : If the bitfield is signed and it's wider than type_out, we need to
2615 : : keep the result sign-extended:
2616 : : container = (type) container;
2617 : : masked = container << (prec - bitsize - bitpos);
2618 : : result = (type_out) (masked >> (prec - bitsize));
2619 : :
2620 : : Here type is the signed variant of the wider of type_out and the type
2621 : : of container.
2622 : :
2623 : : The shifting is always optional depending on whether bitpos != 0.
2624 : :
2625 : : When the original bitfield was inside a gcond then an new gcond is also
2626 : : generated with the newly `result` as the operand to the comparison.
2627 : :
2628 : : */
2629 : :
2630 : : static gimple *
2631 : 29715804 : vect_recog_bitfield_ref_pattern (vec_info *vinfo, stmt_vec_info stmt_info,
2632 : : tree *type_out)
2633 : : {
2634 : 29715804 : gimple *bf_stmt = NULL;
2635 : 29715804 : tree lhs = NULL_TREE;
2636 : 29715804 : tree ret_type = NULL_TREE;
2637 : 29715804 : gimple *stmt = STMT_VINFO_STMT (stmt_info);
2638 : 29715804 : if (gcond *cond_stmt = dyn_cast <gcond *> (stmt))
2639 : : {
2640 : 4954678 : tree op = gimple_cond_lhs (cond_stmt);
2641 : 4954678 : if (TREE_CODE (op) != SSA_NAME)
2642 : : return NULL;
2643 : 4954448 : bf_stmt = dyn_cast <gassign *> (SSA_NAME_DEF_STMT (op));
2644 : 4954448 : if (TREE_CODE (gimple_cond_rhs (cond_stmt)) != INTEGER_CST)
2645 : : return NULL;
2646 : : }
2647 : 24761126 : else if (is_gimple_assign (stmt)
2648 : 20286152 : && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
2649 : 27461654 : && TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME)
2650 : : {
2651 : 2656293 : gimple *second_stmt = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt));
2652 : 2656293 : bf_stmt = dyn_cast <gassign *> (second_stmt);
2653 : 2656293 : lhs = gimple_assign_lhs (stmt);
2654 : 2656293 : ret_type = TREE_TYPE (lhs);
2655 : : }
2656 : :
2657 : 5930095 : if (!bf_stmt
2658 : 5930095 : || gimple_assign_rhs_code (bf_stmt) != BIT_FIELD_REF)
2659 : : return NULL;
2660 : :
2661 : 14118 : tree bf_ref = gimple_assign_rhs1 (bf_stmt);
2662 : 14118 : tree container = TREE_OPERAND (bf_ref, 0);
2663 : 14118 : ret_type = ret_type ? ret_type : TREE_TYPE (container);
2664 : :
2665 : 14118 : if (!bit_field_offset (bf_ref).is_constant ()
2666 : 14118 : || !bit_field_size (bf_ref).is_constant ()
2667 : 14118 : || !tree_fits_uhwi_p (TYPE_SIZE (TREE_TYPE (container))))
2668 : : return NULL;
2669 : :
2670 : 27858 : if (!INTEGRAL_TYPE_P (TREE_TYPE (bf_ref))
2671 : 14113 : || !INTEGRAL_TYPE_P (TREE_TYPE (container))
2672 : 16116 : || TYPE_MODE (TREE_TYPE (container)) == E_BLKmode)
2673 : 12120 : return NULL;
2674 : :
2675 : 1998 : gimple *use_stmt, *pattern_stmt;
2676 : 1998 : use_operand_p use_p;
2677 : 1998 : bool shift_first = true;
2678 : 1998 : tree container_type = TREE_TYPE (container);
2679 : 1998 : tree vectype = get_vectype_for_scalar_type (vinfo, container_type);
2680 : :
2681 : : /* Calculate shift_n before the adjustments for widening loads, otherwise
2682 : : the container may change and we have to consider offset change for
2683 : : widening loads on big endianness. The shift_n calculated here can be
2684 : : independent of widening. */
2685 : 1998 : unsigned HOST_WIDE_INT shift_n = bit_field_offset (bf_ref).to_constant ();
2686 : 1998 : unsigned HOST_WIDE_INT mask_width = bit_field_size (bf_ref).to_constant ();
2687 : 1998 : unsigned HOST_WIDE_INT prec = tree_to_uhwi (TYPE_SIZE (container_type));
2688 : 1998 : if (BYTES_BIG_ENDIAN)
2689 : : shift_n = prec - shift_n - mask_width;
2690 : :
2691 : 1998 : bool ref_sext = (!TYPE_UNSIGNED (TREE_TYPE (bf_ref)) &&
2692 : 1353 : TYPE_PRECISION (ret_type) > mask_width);
2693 : 1998 : bool load_widen = (TYPE_PRECISION (TREE_TYPE (container)) <
2694 : 1998 : TYPE_PRECISION (ret_type));
2695 : :
2696 : : /* We move the conversion earlier if the loaded type is smaller than the
2697 : : return type to enable the use of widening loads. And if we need a
2698 : : sign extension, we need to convert the loaded value early to a signed
2699 : : type as well. */
2700 : 1998 : if (ref_sext || load_widen)
2701 : : {
2702 : 936 : tree type = load_widen ? ret_type : container_type;
2703 : 936 : if (ref_sext)
2704 : 891 : type = gimple_signed_type (type);
2705 : 936 : pattern_stmt = gimple_build_assign (vect_recog_temp_ssa_var (type),
2706 : : NOP_EXPR, container);
2707 : 936 : container = gimple_get_lhs (pattern_stmt);
2708 : 936 : container_type = TREE_TYPE (container);
2709 : 936 : prec = tree_to_uhwi (TYPE_SIZE (container_type));
2710 : 936 : vectype = get_vectype_for_scalar_type (vinfo, container_type);
2711 : 936 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt, vectype);
2712 : : }
2713 : 1062 : else if (!useless_type_conversion_p (TREE_TYPE (container), ret_type))
2714 : : /* If we are doing the conversion last then also delay the shift as we may
2715 : : be able to combine the shift and conversion in certain cases. */
2716 : : shift_first = false;
2717 : :
2718 : : /* If the only use of the result of this BIT_FIELD_REF + CONVERT is a
2719 : : PLUS_EXPR then do the shift last as some targets can combine the shift and
2720 : : add into a single instruction. */
2721 : 1295 : if (lhs && !is_pattern_stmt_p (stmt_info)
2722 : 3293 : && single_imm_use (lhs, &use_p, &use_stmt))
2723 : : {
2724 : 940 : if (gimple_code (use_stmt) == GIMPLE_ASSIGN
2725 : 940 : && gimple_assign_rhs_code (use_stmt) == PLUS_EXPR)
2726 : : shift_first = false;
2727 : : }
2728 : :
2729 : : /* If we don't have to shift we only generate the mask, so just fix the
2730 : : code-path to shift_first. */
2731 : 1998 : if (shift_n == 0)
2732 : 707 : shift_first = true;
2733 : :
2734 : 1998 : tree result;
2735 : 1998 : if (shift_first && !ref_sext)
2736 : : {
2737 : 446 : tree shifted = container;
2738 : 446 : if (shift_n)
2739 : : {
2740 : 53 : pattern_stmt
2741 : 53 : = gimple_build_assign (vect_recog_temp_ssa_var (container_type),
2742 : : RSHIFT_EXPR, container,
2743 : 53 : build_int_cst (sizetype, shift_n));
2744 : 53 : shifted = gimple_assign_lhs (pattern_stmt);
2745 : 53 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt, vectype);
2746 : : }
2747 : :
2748 : 446 : tree mask = wide_int_to_tree (container_type,
2749 : 446 : wi::mask (mask_width, false, prec));
2750 : :
2751 : 446 : pattern_stmt
2752 : 446 : = gimple_build_assign (vect_recog_temp_ssa_var (container_type),
2753 : : BIT_AND_EXPR, shifted, mask);
2754 : 446 : result = gimple_assign_lhs (pattern_stmt);
2755 : : }
2756 : : else
2757 : : {
2758 : 1552 : tree temp = vect_recog_temp_ssa_var (container_type);
2759 : 1552 : if (!ref_sext)
2760 : : {
2761 : 661 : tree mask = wide_int_to_tree (container_type,
2762 : 661 : wi::shifted_mask (shift_n,
2763 : : mask_width,
2764 : : false, prec));
2765 : 661 : pattern_stmt = gimple_build_assign (temp, BIT_AND_EXPR,
2766 : : container, mask);
2767 : : }
2768 : : else
2769 : : {
2770 : 891 : HOST_WIDE_INT shl = prec - shift_n - mask_width;
2771 : 891 : shift_n += shl;
2772 : 891 : pattern_stmt = gimple_build_assign (temp, LSHIFT_EXPR,
2773 : : container,
2774 : : build_int_cst (sizetype,
2775 : 891 : shl));
2776 : : }
2777 : :
2778 : 1552 : tree masked = gimple_assign_lhs (pattern_stmt);
2779 : 1552 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt, vectype);
2780 : 1552 : pattern_stmt
2781 : 1552 : = gimple_build_assign (vect_recog_temp_ssa_var (container_type),
2782 : : RSHIFT_EXPR, masked,
2783 : 1552 : build_int_cst (sizetype, shift_n));
2784 : 1552 : result = gimple_assign_lhs (pattern_stmt);
2785 : : }
2786 : :
2787 : 1998 : if (!useless_type_conversion_p (TREE_TYPE (result), ret_type))
2788 : : {
2789 : 1299 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt, vectype);
2790 : 1299 : pattern_stmt
2791 : 1299 : = gimple_build_assign (vect_recog_temp_ssa_var (ret_type),
2792 : : NOP_EXPR, result);
2793 : : }
2794 : :
2795 : 1998 : if (!lhs)
2796 : : {
2797 : 703 : if (!vectype)
2798 : : return NULL;
2799 : :
2800 : 595 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt, vectype);
2801 : 595 : vectype = truth_type_for (vectype);
2802 : :
2803 : : /* FIXME: This part extracts the boolean value out of the bitfield in the
2804 : : same way as vect_recog_gcond_pattern does. However because
2805 : : patterns cannot match the same root twice, when we handle and
2806 : : lower the bitfield in the gcond, vect_recog_gcond_pattern can't
2807 : : apply anymore. We should really fix it so that we don't need to
2808 : : duplicate transformations like these. */
2809 : 595 : tree new_lhs = vect_recog_temp_ssa_var (boolean_type_node, NULL);
2810 : 595 : gcond *cond_stmt = dyn_cast <gcond *> (stmt_info->stmt);
2811 : 595 : tree cond_cst = gimple_cond_rhs (cond_stmt);
2812 : 595 : gimple *new_stmt
2813 : 595 : = gimple_build_assign (new_lhs, gimple_cond_code (cond_stmt),
2814 : : gimple_get_lhs (pattern_stmt),
2815 : : fold_convert (container_type, cond_cst));
2816 : 595 : append_pattern_def_seq (vinfo, stmt_info, new_stmt, vectype, container_type);
2817 : 595 : pattern_stmt
2818 : 595 : = gimple_build_cond (NE_EXPR, new_lhs,
2819 : 595 : build_zero_cst (TREE_TYPE (new_lhs)),
2820 : : NULL_TREE, NULL_TREE);
2821 : : }
2822 : :
2823 : 1890 : *type_out = STMT_VINFO_VECTYPE (stmt_info);
2824 : 1890 : vect_pattern_detected ("bitfield_ref pattern", stmt_info->stmt);
2825 : :
2826 : 1890 : return pattern_stmt;
2827 : : }
2828 : :
2829 : : /* Function vect_recog_bit_insert_pattern
2830 : :
2831 : : Try to find the following pattern:
2832 : :
2833 : : written = BIT_INSERT_EXPR (container, value, bitpos);
2834 : :
2835 : : Input:
2836 : :
2837 : : * STMT_VINFO: The stmt we want to replace.
2838 : :
2839 : : Output:
2840 : :
2841 : : * TYPE_OUT: The vector type of the output of this pattern.
2842 : :
2843 : : * Return value: A new stmt that will be used to replace the sequence of
2844 : : stmts that constitute the pattern. In this case it will be:
2845 : : value = (container_type) value; // Make sure
2846 : : shifted = value << bitpos; // Shift value into place
2847 : : masked = shifted & (mask << bitpos); // Mask off the non-relevant bits in
2848 : : // the 'to-write value'.
2849 : : cleared = container & ~(mask << bitpos); // Clearing the bits we want to
2850 : : // write to from the value we want
2851 : : // to write to.
2852 : : written = cleared | masked; // Write bits.
2853 : :
2854 : :
2855 : : where mask = ((1 << TYPE_PRECISION (value)) - 1), a mask to keep the number of
2856 : : bits corresponding to the real size of the bitfield value we are writing to.
2857 : : The shifting is always optional depending on whether bitpos != 0.
2858 : :
2859 : : */
2860 : :
2861 : : static gimple *
2862 : 29718735 : vect_recog_bit_insert_pattern (vec_info *vinfo, stmt_vec_info stmt_info,
2863 : : tree *type_out)
2864 : : {
2865 : 29718735 : gassign *bf_stmt = dyn_cast <gassign *> (stmt_info->stmt);
2866 : 27317632 : if (!bf_stmt || gimple_assign_rhs_code (bf_stmt) != BIT_INSERT_EXPR)
2867 : : return NULL;
2868 : :
2869 : 610 : tree container = gimple_assign_rhs1 (bf_stmt);
2870 : 610 : tree value = gimple_assign_rhs2 (bf_stmt);
2871 : 610 : tree shift = gimple_assign_rhs3 (bf_stmt);
2872 : :
2873 : 610 : tree bf_type = TREE_TYPE (value);
2874 : 610 : tree container_type = TREE_TYPE (container);
2875 : :
2876 : 610 : if (!INTEGRAL_TYPE_P (container_type)
2877 : 610 : || !tree_fits_uhwi_p (TYPE_SIZE (container_type)))
2878 : : return NULL;
2879 : :
2880 : 506 : gimple *pattern_stmt;
2881 : :
2882 : 506 : vect_unpromoted_value unprom;
2883 : 506 : unprom.set_op (value, vect_internal_def);
2884 : 506 : value = vect_convert_input (vinfo, stmt_info, container_type, &unprom,
2885 : : get_vectype_for_scalar_type (vinfo,
2886 : : container_type));
2887 : :
2888 : 506 : unsigned HOST_WIDE_INT mask_width = TYPE_PRECISION (bf_type);
2889 : 506 : unsigned HOST_WIDE_INT prec = tree_to_uhwi (TYPE_SIZE (container_type));
2890 : 506 : unsigned HOST_WIDE_INT shift_n = tree_to_uhwi (shift);
2891 : 506 : if (BYTES_BIG_ENDIAN)
2892 : : {
2893 : : shift_n = prec - shift_n - mask_width;
2894 : : shift = build_int_cst (TREE_TYPE (shift), shift_n);
2895 : : }
2896 : :
2897 : 506 : if (!useless_type_conversion_p (TREE_TYPE (value), container_type))
2898 : : {
2899 : 0 : pattern_stmt =
2900 : 0 : gimple_build_assign (vect_recog_temp_ssa_var (container_type),
2901 : : NOP_EXPR, value);
2902 : 0 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt);
2903 : 0 : value = gimple_get_lhs (pattern_stmt);
2904 : : }
2905 : :
2906 : : /* Shift VALUE into place. */
2907 : 506 : tree shifted = value;
2908 : 506 : if (shift_n)
2909 : : {
2910 : 261 : gimple_seq stmts = NULL;
2911 : 261 : shifted
2912 : 261 : = gimple_build (&stmts, LSHIFT_EXPR, container_type, value, shift);
2913 : 261 : if (!gimple_seq_empty_p (stmts))
2914 : 112 : append_pattern_def_seq (vinfo, stmt_info,
2915 : : gimple_seq_first_stmt (stmts));
2916 : : }
2917 : :
2918 : 506 : tree mask_t
2919 : 506 : = wide_int_to_tree (container_type,
2920 : 506 : wi::shifted_mask (shift_n, mask_width, false, prec));
2921 : :
2922 : : /* Clear bits we don't want to write back from SHIFTED. */
2923 : 506 : gimple_seq stmts = NULL;
2924 : 506 : tree masked = gimple_build (&stmts, BIT_AND_EXPR, container_type, shifted,
2925 : : mask_t);
2926 : 506 : if (!gimple_seq_empty_p (stmts))
2927 : : {
2928 : 110 : pattern_stmt = gimple_seq_first_stmt (stmts);
2929 : 110 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt);
2930 : : }
2931 : :
2932 : : /* Mask off the bits in the container that we are to write to. */
2933 : 506 : mask_t = wide_int_to_tree (container_type,
2934 : 506 : wi::shifted_mask (shift_n, mask_width, true, prec));
2935 : 506 : tree cleared = vect_recog_temp_ssa_var (container_type);
2936 : 506 : pattern_stmt = gimple_build_assign (cleared, BIT_AND_EXPR, container, mask_t);
2937 : 506 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt);
2938 : :
2939 : : /* Write MASKED into CLEARED. */
2940 : 506 : pattern_stmt
2941 : 506 : = gimple_build_assign (vect_recog_temp_ssa_var (container_type),
2942 : : BIT_IOR_EXPR, cleared, masked);
2943 : :
2944 : 506 : *type_out = STMT_VINFO_VECTYPE (stmt_info);
2945 : 506 : vect_pattern_detected ("bit_insert pattern", stmt_info->stmt);
2946 : :
2947 : 506 : return pattern_stmt;
2948 : : }
2949 : :
2950 : :
2951 : : /* Recognize cases in which an operation is performed in one type WTYPE
2952 : : but could be done more efficiently in a narrower type NTYPE. For example,
2953 : : if we have:
2954 : :
2955 : : ATYPE a; // narrower than NTYPE
2956 : : BTYPE b; // narrower than NTYPE
2957 : : WTYPE aw = (WTYPE) a;
2958 : : WTYPE bw = (WTYPE) b;
2959 : : WTYPE res = aw + bw; // only uses of aw and bw
2960 : :
2961 : : then it would be more efficient to do:
2962 : :
2963 : : NTYPE an = (NTYPE) a;
2964 : : NTYPE bn = (NTYPE) b;
2965 : : NTYPE resn = an + bn;
2966 : : WTYPE res = (WTYPE) resn;
2967 : :
2968 : : Other situations include things like:
2969 : :
2970 : : ATYPE a; // NTYPE or narrower
2971 : : WTYPE aw = (WTYPE) a;
2972 : : WTYPE res = aw + b;
2973 : :
2974 : : when only "(NTYPE) res" is significant. In that case it's more efficient
2975 : : to truncate "b" and do the operation on NTYPE instead:
2976 : :
2977 : : NTYPE an = (NTYPE) a;
2978 : : NTYPE bn = (NTYPE) b; // truncation
2979 : : NTYPE resn = an + bn;
2980 : : WTYPE res = (WTYPE) resn;
2981 : :
2982 : : All users of "res" should then use "resn" instead, making the final
2983 : : statement dead (not marked as relevant). The final statement is still
2984 : : needed to maintain the type correctness of the IR.
2985 : :
2986 : : vect_determine_precisions has already determined the minimum
2987 : : precison of the operation and the minimum precision required
2988 : : by users of the result. */
2989 : :
2990 : : static gimple *
2991 : 29719175 : vect_recog_over_widening_pattern (vec_info *vinfo,
2992 : : stmt_vec_info last_stmt_info, tree *type_out)
2993 : : {
2994 : 29719175 : gassign *last_stmt = dyn_cast <gassign *> (last_stmt_info->stmt);
2995 : 20290118 : if (!last_stmt)
2996 : : return NULL;
2997 : :
2998 : : /* See whether we have found that this operation can be done on a
2999 : : narrower type without changing its semantics. */
3000 : 20290118 : unsigned int new_precision = last_stmt_info->operation_precision;
3001 : 20290118 : if (!new_precision)
3002 : : return NULL;
3003 : :
3004 : 1289664 : tree lhs = gimple_assign_lhs (last_stmt);
3005 : 1289664 : tree type = TREE_TYPE (lhs);
3006 : 1289664 : tree_code code = gimple_assign_rhs_code (last_stmt);
3007 : :
3008 : : /* Punt for reductions where we don't handle the type conversions. */
3009 : 1289664 : if (STMT_VINFO_DEF_TYPE (last_stmt_info) == vect_reduction_def)
3010 : : return NULL;
3011 : :
3012 : : /* Keep the first operand of a COND_EXPR as-is: only the other two
3013 : : operands are interesting. */
3014 : 1285512 : unsigned int first_op = (code == COND_EXPR ? 2 : 1);
3015 : :
3016 : : /* Check the operands. */
3017 : 1285512 : unsigned int nops = gimple_num_ops (last_stmt) - first_op;
3018 : 1285512 : auto_vec <vect_unpromoted_value, 3> unprom (nops);
3019 : 1285512 : unprom.quick_grow_cleared (nops);
3020 : 1285512 : unsigned int min_precision = 0;
3021 : 1285512 : bool single_use_p = false;
3022 : 3843613 : for (unsigned int i = 0; i < nops; ++i)
3023 : : {
3024 : 2558762 : tree op = gimple_op (last_stmt, first_op + i);
3025 : 2558762 : if (TREE_CODE (op) == INTEGER_CST)
3026 : 1141436 : unprom[i].set_op (op, vect_constant_def);
3027 : 1417326 : else if (TREE_CODE (op) == SSA_NAME)
3028 : : {
3029 : 1417326 : bool op_single_use_p = true;
3030 : 1417326 : if (!vect_look_through_possible_promotion (vinfo, op, &unprom[i],
3031 : : &op_single_use_p))
3032 : 661 : return NULL;
3033 : : /* If:
3034 : :
3035 : : (1) N bits of the result are needed;
3036 : : (2) all inputs are widened from M<N bits; and
3037 : : (3) one operand OP is a single-use SSA name
3038 : :
3039 : : we can shift the M->N widening from OP to the output
3040 : : without changing the number or type of extensions involved.
3041 : : This then reduces the number of copies of STMT_INFO.
3042 : :
3043 : : If instead of (3) more than one operand is a single-use SSA name,
3044 : : shifting the extension to the output is even more of a win.
3045 : :
3046 : : If instead:
3047 : :
3048 : : (1) N bits of the result are needed;
3049 : : (2) one operand OP2 is widened from M2<N bits;
3050 : : (3) another operand OP1 is widened from M1<M2 bits; and
3051 : : (4) both OP1 and OP2 are single-use
3052 : :
3053 : : the choice is between:
3054 : :
3055 : : (a) truncating OP2 to M1, doing the operation on M1,
3056 : : and then widening the result to N
3057 : :
3058 : : (b) widening OP1 to M2, doing the operation on M2, and then
3059 : : widening the result to N
3060 : :
3061 : : Both shift the M2->N widening of the inputs to the output.
3062 : : (a) additionally shifts the M1->M2 widening to the output;
3063 : : it requires fewer copies of STMT_INFO but requires an extra
3064 : : M2->M1 truncation.
3065 : :
3066 : : Which is better will depend on the complexity and cost of
3067 : : STMT_INFO, which is hard to predict at this stage. However,
3068 : : a clear tie-breaker in favor of (b) is the fact that the
3069 : : truncation in (a) increases the length of the operation chain.
3070 : :
3071 : : If instead of (4) only one of OP1 or OP2 is single-use,
3072 : : (b) is still a win over doing the operation in N bits:
3073 : : it still shifts the M2->N widening on the single-use operand
3074 : : to the output and reduces the number of STMT_INFO copies.
3075 : :
3076 : : If neither operand is single-use then operating on fewer than
3077 : : N bits might lead to more extensions overall. Whether it does
3078 : : or not depends on global information about the vectorization
3079 : : region, and whether that's a good trade-off would again
3080 : : depend on the complexity and cost of the statements involved,
3081 : : as well as things like register pressure that are not normally
3082 : : modelled at this stage. We therefore ignore these cases
3083 : : and just optimize the clear single-use wins above.
3084 : :
3085 : : Thus we take the maximum precision of the unpromoted operands
3086 : : and record whether any operand is single-use. */
3087 : 1416665 : if (unprom[i].dt == vect_internal_def)
3088 : : {
3089 : 1017963 : min_precision = MAX (min_precision,
3090 : : TYPE_PRECISION (unprom[i].type));
3091 : 1017963 : single_use_p |= op_single_use_p;
3092 : : }
3093 : : }
3094 : : else
3095 : : return NULL;
3096 : : }
3097 : :
3098 : : /* Although the operation could be done in operation_precision, we have
3099 : : to balance that against introducing extra truncations or extensions.
3100 : : Calculate the minimum precision that can be handled efficiently.
3101 : :
3102 : : The loop above determined that the operation could be handled
3103 : : efficiently in MIN_PRECISION if SINGLE_USE_P; this would shift an
3104 : : extension from the inputs to the output without introducing more
3105 : : instructions, and would reduce the number of instructions required
3106 : : for STMT_INFO itself.
3107 : :
3108 : : vect_determine_precisions has also determined that the result only
3109 : : needs min_output_precision bits. Truncating by a factor of N times
3110 : : requires a tree of N - 1 instructions, so if TYPE is N times wider
3111 : : than min_output_precision, doing the operation in TYPE and truncating
3112 : : the result requires N + (N - 1) = 2N - 1 instructions per output vector.
3113 : : In contrast:
3114 : :
3115 : : - truncating the input to a unary operation and doing the operation
3116 : : in the new type requires at most N - 1 + 1 = N instructions per
3117 : : output vector
3118 : :
3119 : : - doing the same for a binary operation requires at most
3120 : : (N - 1) * 2 + 1 = 2N - 1 instructions per output vector
3121 : :
3122 : : Both unary and binary operations require fewer instructions than
3123 : : this if the operands were extended from a suitable truncated form.
3124 : : Thus there is usually nothing to lose by doing operations in
3125 : : min_output_precision bits, but there can be something to gain. */
3126 : 1284851 : if (!single_use_p)
3127 : 936088 : min_precision = last_stmt_info->min_output_precision;
3128 : : else
3129 : 348763 : min_precision = MIN (min_precision, last_stmt_info->min_output_precision);
3130 : :
3131 : : /* Apply the minimum efficient precision we just calculated. */
3132 : 1284851 : if (new_precision < min_precision)
3133 : : new_precision = min_precision;
3134 : 1284851 : new_precision = vect_element_precision (new_precision);
3135 : 1284851 : if (new_precision >= TYPE_PRECISION (type))
3136 : : return NULL;
3137 : :
3138 : 159801 : vect_pattern_detected ("vect_recog_over_widening_pattern", last_stmt);
3139 : :
3140 : 159801 : *type_out = get_vectype_for_scalar_type (vinfo, type);
3141 : 159801 : if (!*type_out)
3142 : : return NULL;
3143 : :
3144 : : /* We've found a viable pattern. Get the new type of the operation. */
3145 : 142700 : bool unsigned_p = (last_stmt_info->operation_sign == UNSIGNED);
3146 : 142700 : tree new_type = build_nonstandard_integer_type (new_precision, unsigned_p);
3147 : :
3148 : : /* If we're truncating an operation, we need to make sure that we
3149 : : don't introduce new undefined overflow. The codes tested here are
3150 : : a subset of those accepted by vect_truncatable_operation_p. */
3151 : 142700 : tree op_type = new_type;
3152 : 142700 : if (TYPE_OVERFLOW_UNDEFINED (new_type)
3153 : 180035 : && (code == PLUS_EXPR || code == MINUS_EXPR || code == MULT_EXPR))
3154 : 25731 : op_type = build_nonstandard_integer_type (new_precision, true);
3155 : :
3156 : : /* We specifically don't check here whether the target supports the
3157 : : new operation, since it might be something that a later pattern
3158 : : wants to rewrite anyway. If targets have a minimum element size
3159 : : for some optabs, we should pattern-match smaller ops to larger ops
3160 : : where beneficial. */
3161 : 142700 : tree new_vectype = get_vectype_for_scalar_type (vinfo, new_type);
3162 : 142700 : tree op_vectype = get_vectype_for_scalar_type (vinfo, op_type);
3163 : 142700 : if (!new_vectype || !op_vectype)
3164 : : return NULL;
3165 : :
3166 : 142700 : if (dump_enabled_p ())
3167 : 4586 : dump_printf_loc (MSG_NOTE, vect_location, "demoting %T to %T\n",
3168 : : type, new_type);
3169 : :
3170 : : /* Calculate the rhs operands for an operation on OP_TYPE. */
3171 : 142700 : tree ops[3] = {};
3172 : 142910 : for (unsigned int i = 1; i < first_op; ++i)
3173 : 210 : ops[i - 1] = gimple_op (last_stmt, i);
3174 : 142700 : vect_convert_inputs (vinfo, last_stmt_info, nops, &ops[first_op - 1],
3175 : 142700 : op_type, &unprom[0], op_vectype);
3176 : :
3177 : : /* Use the operation to produce a result of type OP_TYPE. */
3178 : 142700 : tree new_var = vect_recog_temp_ssa_var (op_type, NULL);
3179 : 142700 : gimple *pattern_stmt = gimple_build_assign (new_var, code,
3180 : : ops[0], ops[1], ops[2]);
3181 : 142700 : gimple_set_location (pattern_stmt, gimple_location (last_stmt));
3182 : :
3183 : 142700 : if (dump_enabled_p ())
3184 : 4586 : dump_printf_loc (MSG_NOTE, vect_location,
3185 : : "created pattern stmt: %G", pattern_stmt);
3186 : :
3187 : : /* Convert back to the original signedness, if OP_TYPE is different
3188 : : from NEW_TYPE. */
3189 : 142700 : if (op_type != new_type)
3190 : 25731 : pattern_stmt = vect_convert_output (vinfo, last_stmt_info, new_type,
3191 : : pattern_stmt, op_vectype);
3192 : :
3193 : : /* Promote the result to the original type. */
3194 : 142700 : pattern_stmt = vect_convert_output (vinfo, last_stmt_info, type,
3195 : : pattern_stmt, new_vectype);
3196 : :
3197 : 142700 : return pattern_stmt;
3198 : 1285512 : }
3199 : :
3200 : : /* Recognize the following patterns:
3201 : :
3202 : : ATYPE a; // narrower than TYPE
3203 : : BTYPE b; // narrower than TYPE
3204 : :
3205 : : 1) Multiply high with scaling
3206 : : TYPE res = ((TYPE) a * (TYPE) b) >> c;
3207 : : Here, c is bitsize (TYPE) / 2 - 1.
3208 : :
3209 : : 2) ... or also with rounding
3210 : : TYPE res = (((TYPE) a * (TYPE) b) >> d + 1) >> 1;
3211 : : Here, d is bitsize (TYPE) / 2 - 2.
3212 : :
3213 : : 3) Normal multiply high
3214 : : TYPE res = ((TYPE) a * (TYPE) b) >> e;
3215 : : Here, e is bitsize (TYPE) / 2.
3216 : :
3217 : : where only the bottom half of res is used. */
3218 : :
3219 : : static gimple *
3220 : 29848394 : vect_recog_mulhs_pattern (vec_info *vinfo,
3221 : : stmt_vec_info last_stmt_info, tree *type_out)
3222 : : {
3223 : : /* Check for a right shift. */
3224 : 29848394 : gassign *last_stmt = dyn_cast <gassign *> (last_stmt_info->stmt);
3225 : 20419203 : if (!last_stmt
3226 : 20419203 : || gimple_assign_rhs_code (last_stmt) != RSHIFT_EXPR)
3227 : : return NULL;
3228 : :
3229 : : /* Check that the shift result is wider than the users of the
3230 : : result need (i.e. that narrowing would be a natural choice). */
3231 : 344435 : tree lhs_type = TREE_TYPE (gimple_assign_lhs (last_stmt));
3232 : 344435 : unsigned int target_precision
3233 : 344435 : = vect_element_precision (last_stmt_info->min_output_precision);
3234 : 344435 : if (!INTEGRAL_TYPE_P (lhs_type)
3235 : 344435 : || target_precision >= TYPE_PRECISION (lhs_type))
3236 : : return NULL;
3237 : :
3238 : : /* Look through any change in sign on the outer shift input. */
3239 : 46820 : vect_unpromoted_value unprom_rshift_input;
3240 : 46820 : tree rshift_input = vect_look_through_possible_promotion
3241 : 46820 : (vinfo, gimple_assign_rhs1 (last_stmt), &unprom_rshift_input);
3242 : 46820 : if (!rshift_input
3243 : 46820 : || TYPE_PRECISION (TREE_TYPE (rshift_input))
3244 : 46212 : != TYPE_PRECISION (lhs_type))
3245 : : return NULL;
3246 : :
3247 : : /* Get the definition of the shift input. */
3248 : 44249 : stmt_vec_info rshift_input_stmt_info
3249 : 44249 : = vect_get_internal_def (vinfo, rshift_input);
3250 : 44249 : if (!rshift_input_stmt_info)
3251 : : return NULL;
3252 : 39835 : gassign *rshift_input_stmt
3253 : 29882816 : = dyn_cast <gassign *> (rshift_input_stmt_info->stmt);
3254 : 34509 : if (!rshift_input_stmt)
3255 : : return NULL;
3256 : :
3257 : 34509 : stmt_vec_info mulh_stmt_info;
3258 : 34509 : tree scale_term;
3259 : 34509 : bool rounding_p = false;
3260 : :
3261 : : /* Check for the presence of the rounding term. */
3262 : 41444 : if (gimple_assign_rhs_code (rshift_input_stmt) == PLUS_EXPR)
3263 : : {
3264 : : /* Check that the outer shift was by 1. */
3265 : 17704 : if (!integer_onep (gimple_assign_rhs2 (last_stmt)))
3266 : 8798 : return NULL;
3267 : :
3268 : : /* Check that the second operand of the PLUS_EXPR is 1. */
3269 : 1338 : if (!integer_onep (gimple_assign_rhs2 (rshift_input_stmt)))
3270 : : return NULL;
3271 : :
3272 : : /* Look through any change in sign on the addition input. */
3273 : 97 : vect_unpromoted_value unprom_plus_input;
3274 : 97 : tree plus_input = vect_look_through_possible_promotion
3275 : 97 : (vinfo, gimple_assign_rhs1 (rshift_input_stmt), &unprom_plus_input);
3276 : 97 : if (!plus_input
3277 : 97 : || TYPE_PRECISION (TREE_TYPE (plus_input))
3278 : 97 : != TYPE_PRECISION (TREE_TYPE (rshift_input)))
3279 : : return NULL;
3280 : :
3281 : : /* Get the definition of the multiply-high-scale part. */
3282 : 97 : stmt_vec_info plus_input_stmt_info
3283 : 97 : = vect_get_internal_def (vinfo, plus_input);
3284 : 97 : if (!plus_input_stmt_info)
3285 : : return NULL;
3286 : 97 : gassign *plus_input_stmt
3287 : 8895 : = dyn_cast <gassign *> (plus_input_stmt_info->stmt);
3288 : 97 : if (!plus_input_stmt
3289 : 97 : || gimple_assign_rhs_code (plus_input_stmt) != RSHIFT_EXPR)
3290 : : return NULL;
3291 : :
3292 : : /* Look through any change in sign on the scaling input. */
3293 : 54 : vect_unpromoted_value unprom_scale_input;
3294 : 54 : tree scale_input = vect_look_through_possible_promotion
3295 : 54 : (vinfo, gimple_assign_rhs1 (plus_input_stmt), &unprom_scale_input);
3296 : 54 : if (!scale_input
3297 : 54 : || TYPE_PRECISION (TREE_TYPE (scale_input))
3298 : 54 : != TYPE_PRECISION (TREE_TYPE (plus_input)))
3299 : : return NULL;
3300 : :
3301 : : /* Get the definition of the multiply-high part. */
3302 : 54 : mulh_stmt_info = vect_get_internal_def (vinfo, scale_input);
3303 : 54 : if (!mulh_stmt_info)
3304 : : return NULL;
3305 : :
3306 : : /* Get the scaling term. */
3307 : 54 : scale_term = gimple_assign_rhs2 (plus_input_stmt);
3308 : 54 : rounding_p = true;
3309 : : }
3310 : : else
3311 : : {
3312 : 25657 : mulh_stmt_info = rshift_input_stmt_info;
3313 : 25657 : scale_term = gimple_assign_rhs2 (last_stmt);
3314 : : }
3315 : :
3316 : : /* Check that the scaling factor is constant. */
3317 : 25711 : if (TREE_CODE (scale_term) != INTEGER_CST)
3318 : : return NULL;
3319 : :
3320 : : /* Check whether the scaling input term can be seen as two widened
3321 : : inputs multiplied together. */
3322 : 74205 : vect_unpromoted_value unprom_mult[2];
3323 : 24735 : tree new_type;
3324 : 24735 : unsigned int nops
3325 : 24735 : = vect_widened_op_tree (vinfo, mulh_stmt_info, MULT_EXPR, WIDEN_MULT_EXPR,
3326 : : false, 2, unprom_mult, &new_type);
3327 : 24735 : if (nops != 2)
3328 : : return NULL;
3329 : :
3330 : : /* Adjust output precision. */
3331 : 4466 : if (TYPE_PRECISION (new_type) < target_precision)
3332 : 0 : new_type = build_nonstandard_integer_type
3333 : 0 : (target_precision, TYPE_UNSIGNED (new_type));
3334 : :
3335 : 4466 : unsigned mult_precision = TYPE_PRECISION (new_type);
3336 : 4466 : internal_fn ifn;
3337 : : /* Check that the scaling factor is expected. Instead of
3338 : : target_precision, we should use the one that we actually
3339 : : use for internal function. */
3340 : 4466 : if (rounding_p)
3341 : : {
3342 : : /* Check pattern 2). */
3343 : 108 : if (wi::to_widest (scale_term) + mult_precision + 2
3344 : 162 : != TYPE_PRECISION (lhs_type))
3345 : : return NULL;
3346 : :
3347 : : ifn = IFN_MULHRS;
3348 : : }
3349 : : else
3350 : : {
3351 : : /* Check for pattern 1). */
3352 : 8824 : if (wi::to_widest (scale_term) + mult_precision + 1
3353 : 13236 : == TYPE_PRECISION (lhs_type))
3354 : : ifn = IFN_MULHS;
3355 : : /* Check for pattern 3). */
3356 : 4368 : else if (wi::to_widest (scale_term) + mult_precision
3357 : 8736 : == TYPE_PRECISION (lhs_type))
3358 : : ifn = IFN_MULH;
3359 : : else
3360 : : return NULL;
3361 : : }
3362 : :
3363 : 4394 : vect_pattern_detected ("vect_recog_mulhs_pattern", last_stmt);
3364 : :
3365 : : /* Check for target support. */
3366 : 4394 : tree new_vectype = get_vectype_for_scalar_type (vinfo, new_type);
3367 : 4394 : if (!new_vectype
3368 : 8774 : || !direct_internal_fn_supported_p
3369 : 4380 : (ifn, new_vectype, OPTIMIZE_FOR_SPEED))
3370 : 4307 : return NULL;
3371 : :
3372 : : /* The IR requires a valid vector type for the cast result, even though
3373 : : it's likely to be discarded. */
3374 : 87 : *type_out = get_vectype_for_scalar_type (vinfo, lhs_type);
3375 : 87 : if (!*type_out)
3376 : : return NULL;
3377 : :
3378 : : /* Generate the IFN_MULHRS call. */
3379 : 87 : tree new_var = vect_recog_temp_ssa_var (new_type, NULL);
3380 : 87 : tree new_ops[2];
3381 : 87 : vect_convert_inputs (vinfo, last_stmt_info, 2, new_ops, new_type,
3382 : : unprom_mult, new_vectype);
3383 : 87 : gcall *mulhrs_stmt
3384 : 87 : = gimple_build_call_internal (ifn, 2, new_ops[0], new_ops[1]);
3385 : 87 : gimple_call_set_lhs (mulhrs_stmt, new_var);
3386 : 87 : gimple_set_location (mulhrs_stmt, gimple_location (last_stmt));
3387 : :
3388 : 87 : if (dump_enabled_p ())
3389 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
3390 : : "created pattern stmt: %G", (gimple *) mulhrs_stmt);
3391 : :
3392 : 87 : return vect_convert_output (vinfo, last_stmt_info, lhs_type,
3393 : 87 : mulhrs_stmt, new_vectype);
3394 : : }
3395 : :
3396 : : /* Recognize the patterns:
3397 : :
3398 : : ATYPE a; // narrower than TYPE
3399 : : BTYPE b; // narrower than TYPE
3400 : : (1) TYPE avg = ((TYPE) a + (TYPE) b) >> 1;
3401 : : or (2) TYPE avg = ((TYPE) a + (TYPE) b + 1) >> 1;
3402 : :
3403 : : where only the bottom half of avg is used. Try to transform them into:
3404 : :
3405 : : (1) NTYPE avg' = .AVG_FLOOR ((NTYPE) a, (NTYPE) b);
3406 : : or (2) NTYPE avg' = .AVG_CEIL ((NTYPE) a, (NTYPE) b);
3407 : :
3408 : : followed by:
3409 : :
3410 : : TYPE avg = (TYPE) avg';
3411 : :
3412 : : where NTYPE is no wider than half of TYPE. Since only the bottom half
3413 : : of avg is used, all or part of the cast of avg' should become redundant.
3414 : :
3415 : : If there is no target support available, generate code to distribute rshift
3416 : : over plus and add a carry. */
3417 : :
3418 : : static gimple *
3419 : 29846657 : vect_recog_average_pattern (vec_info *vinfo,
3420 : : stmt_vec_info last_stmt_info, tree *type_out)
3421 : : {
3422 : : /* Check for a shift right by one bit. */
3423 : 29846657 : gassign *last_stmt = dyn_cast <gassign *> (last_stmt_info->stmt);
3424 : 20417600 : if (!last_stmt
3425 : 20417600 : || gimple_assign_rhs_code (last_stmt) != RSHIFT_EXPR
3426 : 344303 : || !integer_onep (gimple_assign_rhs2 (last_stmt)))
3427 : 29793657 : return NULL;
3428 : :
3429 : : /* Check that the shift result is wider than the users of the
3430 : : result need (i.e. that narrowing would be a natural choice). */
3431 : 53000 : tree lhs = gimple_assign_lhs (last_stmt);
3432 : 53000 : tree type = TREE_TYPE (lhs);
3433 : 53000 : unsigned int target_precision
3434 : 53000 : = vect_element_precision (last_stmt_info->min_output_precision);
3435 : 53000 : if (!INTEGRAL_TYPE_P (type) || target_precision >= TYPE_PRECISION (type))
3436 : : return NULL;
3437 : :
3438 : : /* Look through any change in sign on the shift input. */
3439 : 2358 : tree rshift_rhs = gimple_assign_rhs1 (last_stmt);
3440 : 2358 : vect_unpromoted_value unprom_plus;
3441 : 2358 : rshift_rhs = vect_look_through_possible_promotion (vinfo, rshift_rhs,
3442 : : &unprom_plus);
3443 : 2358 : if (!rshift_rhs
3444 : 2358 : || TYPE_PRECISION (TREE_TYPE (rshift_rhs)) != TYPE_PRECISION (type))
3445 : : return NULL;
3446 : :
3447 : : /* Get the definition of the shift input. */
3448 : 2348 : stmt_vec_info plus_stmt_info = vect_get_internal_def (vinfo, rshift_rhs);
3449 : 2348 : if (!plus_stmt_info)
3450 : : return NULL;
3451 : :
3452 : : /* Check whether the shift input can be seen as a tree of additions on
3453 : : 2 or 3 widened inputs.
3454 : :
3455 : : Note that the pattern should be a win even if the result of one or
3456 : : more additions is reused elsewhere: if the pattern matches, we'd be
3457 : : replacing 2N RSHIFT_EXPRs and N VEC_PACK_*s with N IFN_AVG_*s. */
3458 : 9320 : internal_fn ifn = IFN_AVG_FLOOR;
3459 : 9320 : vect_unpromoted_value unprom[3];
3460 : 2330 : tree new_type;
3461 : 2330 : unsigned int nops = vect_widened_op_tree (vinfo, plus_stmt_info, PLUS_EXPR,
3462 : 2330 : IFN_VEC_WIDEN_PLUS, false, 3,
3463 : : unprom, &new_type);
3464 : 2330 : if (nops == 0)
3465 : : return NULL;
3466 : 924 : if (nops == 3)
3467 : : {
3468 : : /* Check that one operand is 1. */
3469 : : unsigned int i;
3470 : 957 : for (i = 0; i < 3; ++i)
3471 : 897 : if (integer_onep (unprom[i].op))
3472 : : break;
3473 : 299 : if (i == 3)
3474 : : return NULL;
3475 : : /* Throw away the 1 operand and keep the other two. */
3476 : 239 : if (i < 2)
3477 : 0 : unprom[i] = unprom[2];
3478 : : ifn = IFN_AVG_CEIL;
3479 : : }
3480 : :
3481 : 864 : vect_pattern_detected ("vect_recog_average_pattern", last_stmt);
3482 : :
3483 : : /* We know that:
3484 : :
3485 : : (a) the operation can be viewed as:
3486 : :
3487 : : TYPE widened0 = (TYPE) UNPROM[0];
3488 : : TYPE widened1 = (TYPE) UNPROM[1];
3489 : : TYPE tmp1 = widened0 + widened1 {+ 1};
3490 : : TYPE tmp2 = tmp1 >> 1; // LAST_STMT_INFO
3491 : :
3492 : : (b) the first two statements are equivalent to:
3493 : :
3494 : : TYPE widened0 = (TYPE) (NEW_TYPE) UNPROM[0];
3495 : : TYPE widened1 = (TYPE) (NEW_TYPE) UNPROM[1];
3496 : :
3497 : : (c) vect_recog_over_widening_pattern has already tried to narrow TYPE
3498 : : where sensible;
3499 : :
3500 : : (d) all the operations can be performed correctly at twice the width of
3501 : : NEW_TYPE, due to the nature of the average operation; and
3502 : :
3503 : : (e) users of the result of the right shift need only TARGET_PRECISION
3504 : : bits, where TARGET_PRECISION is no more than half of TYPE's
3505 : : precision.
3506 : :
3507 : : Under these circumstances, the only situation in which NEW_TYPE
3508 : : could be narrower than TARGET_PRECISION is if widened0, widened1
3509 : : and an addition result are all used more than once. Thus we can
3510 : : treat any widening of UNPROM[0] and UNPROM[1] to TARGET_PRECISION
3511 : : as "free", whereas widening the result of the average instruction
3512 : : from NEW_TYPE to TARGET_PRECISION would be a new operation. It's
3513 : : therefore better not to go narrower than TARGET_PRECISION. */
3514 : 864 : if (TYPE_PRECISION (new_type) < target_precision)
3515 : 8 : new_type = build_nonstandard_integer_type (target_precision,
3516 : 8 : TYPE_UNSIGNED (new_type));
3517 : :
3518 : : /* Check for target support. */
3519 : 864 : tree new_vectype = get_vectype_for_scalar_type (vinfo, new_type);
3520 : 864 : if (!new_vectype)
3521 : : return NULL;
3522 : :
3523 : 864 : bool fallback_p = false;
3524 : :
3525 : 864 : if (direct_internal_fn_supported_p (ifn, new_vectype, OPTIMIZE_FOR_SPEED))
3526 : : ;
3527 : 730 : else if (TYPE_UNSIGNED (new_type)
3528 : 267 : && optab_for_tree_code (RSHIFT_EXPR, new_vectype, optab_scalar)
3529 : 267 : && optab_for_tree_code (PLUS_EXPR, new_vectype, optab_default)
3530 : 267 : && optab_for_tree_code (BIT_IOR_EXPR, new_vectype, optab_default)
3531 : 997 : && optab_for_tree_code (BIT_AND_EXPR, new_vectype, optab_default))
3532 : : fallback_p = true;
3533 : : else
3534 : 463 : return NULL;
3535 : :
3536 : : /* The IR requires a valid vector type for the cast result, even though
3537 : : it's likely to be discarded. */
3538 : 401 : *type_out = get_vectype_for_scalar_type (vinfo, type);
3539 : 401 : if (!*type_out)
3540 : : return NULL;
3541 : :
3542 : 400 : tree new_var = vect_recog_temp_ssa_var (new_type, NULL);
3543 : 400 : tree new_ops[2];
3544 : 400 : vect_convert_inputs (vinfo, last_stmt_info, 2, new_ops, new_type,
3545 : : unprom, new_vectype);
3546 : :
3547 : 400 : if (fallback_p)
3548 : : {
3549 : : /* As a fallback, generate code for following sequence:
3550 : :
3551 : : shifted_op0 = new_ops[0] >> 1;
3552 : : shifted_op1 = new_ops[1] >> 1;
3553 : : sum_of_shifted = shifted_op0 + shifted_op1;
3554 : : unmasked_carry = new_ops[0] and/or new_ops[1];
3555 : : carry = unmasked_carry & 1;
3556 : : new_var = sum_of_shifted + carry;
3557 : : */
3558 : :
3559 : 266 : tree one_cst = build_one_cst (new_type);
3560 : 266 : gassign *g;
3561 : :
3562 : 266 : tree shifted_op0 = vect_recog_temp_ssa_var (new_type, NULL);
3563 : 266 : g = gimple_build_assign (shifted_op0, RSHIFT_EXPR, new_ops[0], one_cst);
3564 : 266 : append_pattern_def_seq (vinfo, last_stmt_info, g, new_vectype);
3565 : :
3566 : 266 : tree shifted_op1 = vect_recog_temp_ssa_var (new_type, NULL);
3567 : 266 : g = gimple_build_assign (shifted_op1, RSHIFT_EXPR, new_ops[1], one_cst);
3568 : 266 : append_pattern_def_seq (vinfo, last_stmt_info, g, new_vectype);
3569 : :
3570 : 266 : tree sum_of_shifted = vect_recog_temp_ssa_var (new_type, NULL);
3571 : 266 : g = gimple_build_assign (sum_of_shifted, PLUS_EXPR,
3572 : : shifted_op0, shifted_op1);
3573 : 266 : append_pattern_def_seq (vinfo, last_stmt_info, g, new_vectype);
3574 : :
3575 : 266 : tree unmasked_carry = vect_recog_temp_ssa_var (new_type, NULL);
3576 : 266 : tree_code c = (ifn == IFN_AVG_CEIL) ? BIT_IOR_EXPR : BIT_AND_EXPR;
3577 : 266 : g = gimple_build_assign (unmasked_carry, c, new_ops[0], new_ops[1]);
3578 : 266 : append_pattern_def_seq (vinfo, last_stmt_info, g, new_vectype);
3579 : :
3580 : 266 : tree carry = vect_recog_temp_ssa_var (new_type, NULL);
3581 : 266 : g = gimple_build_assign (carry, BIT_AND_EXPR, unmasked_carry, one_cst);
3582 : 266 : append_pattern_def_seq (vinfo, last_stmt_info, g, new_vectype);
3583 : :
3584 : 266 : g = gimple_build_assign (new_var, PLUS_EXPR, sum_of_shifted, carry);
3585 : 266 : return vect_convert_output (vinfo, last_stmt_info, type, g, new_vectype);
3586 : : }
3587 : :
3588 : : /* Generate the IFN_AVG* call. */
3589 : 134 : gcall *average_stmt = gimple_build_call_internal (ifn, 2, new_ops[0],
3590 : : new_ops[1]);
3591 : 134 : gimple_call_set_lhs (average_stmt, new_var);
3592 : 134 : gimple_set_location (average_stmt, gimple_location (last_stmt));
3593 : :
3594 : 134 : if (dump_enabled_p ())
3595 : 42 : dump_printf_loc (MSG_NOTE, vect_location,
3596 : : "created pattern stmt: %G", (gimple *) average_stmt);
3597 : :
3598 : 134 : return vect_convert_output (vinfo, last_stmt_info,
3599 : 134 : type, average_stmt, new_vectype);
3600 : : }
3601 : :
3602 : : /* Recognize cases in which the input to a cast is wider than its
3603 : : output, and the input is fed by a widening operation. Fold this
3604 : : by removing the unnecessary intermediate widening. E.g.:
3605 : :
3606 : : unsigned char a;
3607 : : unsigned int b = (unsigned int) a;
3608 : : unsigned short c = (unsigned short) b;
3609 : :
3610 : : -->
3611 : :
3612 : : unsigned short c = (unsigned short) a;
3613 : :
3614 : : Although this is rare in input IR, it is an expected side-effect
3615 : : of the over-widening pattern above.
3616 : :
3617 : : This is beneficial also for integer-to-float conversions, if the
3618 : : widened integer has more bits than the float, and if the unwidened
3619 : : input doesn't. */
3620 : :
3621 : : static gimple *
3622 : 29848394 : vect_recog_cast_forwprop_pattern (vec_info *vinfo,
3623 : : stmt_vec_info last_stmt_info, tree *type_out)
3624 : : {
3625 : : /* Check for a cast, including an integer-to-float conversion. */
3626 : 50218733 : gassign *last_stmt = dyn_cast <gassign *> (last_stmt_info->stmt);
3627 : 20419116 : if (!last_stmt)
3628 : : return NULL;
3629 : 20419116 : tree_code code = gimple_assign_rhs_code (last_stmt);
3630 : 20419116 : if (!CONVERT_EXPR_CODE_P (code) && code != FLOAT_EXPR)
3631 : : return NULL;
3632 : :
3633 : : /* Make sure that the rhs is a scalar with a natural bitsize. */
3634 : 2886066 : tree lhs = gimple_assign_lhs (last_stmt);
3635 : 2886066 : if (!lhs)
3636 : : return NULL;
3637 : 2886066 : tree lhs_type = TREE_TYPE (lhs);
3638 : 2886066 : scalar_mode lhs_mode;
3639 : 2867653 : if (VECT_SCALAR_BOOLEAN_TYPE_P (lhs_type)
3640 : 5751989 : || !is_a <scalar_mode> (TYPE_MODE (lhs_type), &lhs_mode))
3641 : 23695 : return NULL;
3642 : :
3643 : : /* Check for a narrowing operation (from a vector point of view). */
3644 : 2862371 : tree rhs = gimple_assign_rhs1 (last_stmt);
3645 : 2862371 : tree rhs_type = TREE_TYPE (rhs);
3646 : 2862371 : if (!INTEGRAL_TYPE_P (rhs_type)
3647 : 2575251 : || VECT_SCALAR_BOOLEAN_TYPE_P (rhs_type)
3648 : 7834097 : || TYPE_PRECISION (rhs_type) <= GET_MODE_BITSIZE (lhs_mode))
3649 : : return NULL;
3650 : :
3651 : : /* Try to find an unpromoted input. */
3652 : 372474 : vect_unpromoted_value unprom;
3653 : 372474 : if (!vect_look_through_possible_promotion (vinfo, rhs, &unprom)
3654 : 372474 : || TYPE_PRECISION (unprom.type) >= TYPE_PRECISION (rhs_type))
3655 : : return NULL;
3656 : :
3657 : : /* If the bits above RHS_TYPE matter, make sure that they're the
3658 : : same when extending from UNPROM as they are when extending from RHS. */
3659 : 48907 : if (!INTEGRAL_TYPE_P (lhs_type)
3660 : 48907 : && TYPE_SIGN (rhs_type) != TYPE_SIGN (unprom.type))
3661 : : return NULL;
3662 : :
3663 : : /* We can get the same result by casting UNPROM directly, to avoid
3664 : : the unnecessary widening and narrowing. */
3665 : 48777 : vect_pattern_detected ("vect_recog_cast_forwprop_pattern", last_stmt);
3666 : :
3667 : 48777 : *type_out = get_vectype_for_scalar_type (vinfo, lhs_type);
3668 : 48777 : if (!*type_out)
3669 : : return NULL;
3670 : :
3671 : 48777 : tree new_var = vect_recog_temp_ssa_var (lhs_type, NULL);
3672 : 48777 : gimple *pattern_stmt = gimple_build_assign (new_var, code, unprom.op);
3673 : 48777 : gimple_set_location (pattern_stmt, gimple_location (last_stmt));
3674 : :
3675 : 48777 : return pattern_stmt;
3676 : : }
3677 : :
3678 : : /* Try to detect a shift left of a widened input, converting LSHIFT_EXPR
3679 : : to WIDEN_LSHIFT_EXPR. See vect_recog_widen_op_pattern for details. */
3680 : :
3681 : : static gimple *
3682 : 29776169 : vect_recog_widen_shift_pattern (vec_info *vinfo,
3683 : : stmt_vec_info last_stmt_info, tree *type_out)
3684 : : {
3685 : 29776169 : return vect_recog_widen_op_pattern (vinfo, last_stmt_info, type_out,
3686 : 29776169 : LSHIFT_EXPR, WIDEN_LSHIFT_EXPR, true,
3687 : 29776169 : "vect_recog_widen_shift_pattern");
3688 : : }
3689 : :
3690 : : /* Detect a rotate pattern wouldn't be otherwise vectorized:
3691 : :
3692 : : type a_t, b_t, c_t;
3693 : :
3694 : : S0 a_t = b_t r<< c_t;
3695 : :
3696 : : Input/Output:
3697 : :
3698 : : * STMT_VINFO: The stmt from which the pattern search begins,
3699 : : i.e. the shift/rotate stmt. The original stmt (S0) is replaced
3700 : : with a sequence:
3701 : :
3702 : : S1 d_t = -c_t;
3703 : : S2 e_t = d_t & (B - 1);
3704 : : S3 f_t = b_t << c_t;
3705 : : S4 g_t = b_t >> e_t;
3706 : : S0 a_t = f_t | g_t;
3707 : :
3708 : : where B is element bitsize of type.
3709 : :
3710 : : Output:
3711 : :
3712 : : * TYPE_OUT: The type of the output of this pattern.
3713 : :
3714 : : * Return value: A new stmt that will be used to replace the rotate
3715 : : S0 stmt. */
3716 : :
3717 : : static gimple *
3718 : 29776169 : vect_recog_rotate_pattern (vec_info *vinfo,
3719 : : stmt_vec_info stmt_vinfo, tree *type_out)
3720 : : {
3721 : 29776169 : gimple *last_stmt = stmt_vinfo->stmt;
3722 : 29776169 : tree oprnd0, oprnd1, lhs, var, var1, var2, vectype, type, stype, def, def2;
3723 : 29776169 : gimple *pattern_stmt, *def_stmt;
3724 : 29776169 : enum tree_code rhs_code;
3725 : 29776169 : enum vect_def_type dt;
3726 : 29776169 : optab optab1, optab2;
3727 : 29776169 : edge ext_def = NULL;
3728 : 29776169 : bool bswap16_p = false;
3729 : :
3730 : 29776169 : if (is_gimple_assign (last_stmt))
3731 : : {
3732 : 20346865 : rhs_code = gimple_assign_rhs_code (last_stmt);
3733 : 20346865 : switch (rhs_code)
3734 : : {
3735 : 5627 : case LROTATE_EXPR:
3736 : 5627 : case RROTATE_EXPR:
3737 : 5627 : break;
3738 : : default:
3739 : : return NULL;
3740 : : }
3741 : :
3742 : 5627 : lhs = gimple_assign_lhs (last_stmt);
3743 : 5627 : oprnd0 = gimple_assign_rhs1 (last_stmt);
3744 : 5627 : type = TREE_TYPE (oprnd0);
3745 : 5627 : oprnd1 = gimple_assign_rhs2 (last_stmt);
3746 : : }
3747 : 9429304 : else if (gimple_call_builtin_p (last_stmt, BUILT_IN_BSWAP16))
3748 : : {
3749 : : /* __builtin_bswap16 (x) is another form of x r>> 8.
3750 : : The vectorizer has bswap support, but only if the argument isn't
3751 : : promoted. */
3752 : 150 : lhs = gimple_call_lhs (last_stmt);
3753 : 150 : oprnd0 = gimple_call_arg (last_stmt, 0);
3754 : 150 : type = TREE_TYPE (oprnd0);
3755 : 150 : if (!lhs
3756 : 150 : || TYPE_PRECISION (TREE_TYPE (lhs)) != 16
3757 : 150 : || TYPE_PRECISION (type) <= 16
3758 : 0 : || TREE_CODE (oprnd0) != SSA_NAME
3759 : 150 : || BITS_PER_UNIT != 8)
3760 : 150 : return NULL;
3761 : :
3762 : 0 : stmt_vec_info def_stmt_info;
3763 : 0 : if (!vect_is_simple_use (oprnd0, vinfo, &dt, &def_stmt_info, &def_stmt))
3764 : : return NULL;
3765 : :
3766 : 0 : if (dt != vect_internal_def)
3767 : : return NULL;
3768 : :
3769 : 0 : if (gimple_assign_cast_p (def_stmt))
3770 : : {
3771 : 0 : def = gimple_assign_rhs1 (def_stmt);
3772 : 0 : if (INTEGRAL_TYPE_P (TREE_TYPE (def))
3773 : 0 : && TYPE_PRECISION (TREE_TYPE (def)) == 16)
3774 : : oprnd0 = def;
3775 : : }
3776 : :
3777 : 0 : type = TREE_TYPE (lhs);
3778 : 0 : vectype = get_vectype_for_scalar_type (vinfo, type);
3779 : 0 : if (vectype == NULL_TREE)
3780 : : return NULL;
3781 : :
3782 : 0 : if (tree char_vectype = get_same_sized_vectype (char_type_node, vectype))
3783 : : {
3784 : : /* The encoding uses one stepped pattern for each byte in the
3785 : : 16-bit word. */
3786 : 0 : vec_perm_builder elts (TYPE_VECTOR_SUBPARTS (char_vectype), 2, 3);
3787 : 0 : for (unsigned i = 0; i < 3; ++i)
3788 : 0 : for (unsigned j = 0; j < 2; ++j)
3789 : 0 : elts.quick_push ((i + 1) * 2 - j - 1);
3790 : :
3791 : 0 : vec_perm_indices indices (elts, 1,
3792 : 0 : TYPE_VECTOR_SUBPARTS (char_vectype));
3793 : 0 : machine_mode vmode = TYPE_MODE (char_vectype);
3794 : 0 : if (can_vec_perm_const_p (vmode, vmode, indices))
3795 : : {
3796 : : /* vectorizable_bswap can handle the __builtin_bswap16 if we
3797 : : undo the argument promotion. */
3798 : 0 : if (!useless_type_conversion_p (type, TREE_TYPE (oprnd0)))
3799 : : {
3800 : 0 : def = vect_recog_temp_ssa_var (type, NULL);
3801 : 0 : def_stmt = gimple_build_assign (def, NOP_EXPR, oprnd0);
3802 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
3803 : 0 : oprnd0 = def;
3804 : : }
3805 : :
3806 : : /* Pattern detected. */
3807 : 0 : vect_pattern_detected ("vect_recog_rotate_pattern", last_stmt);
3808 : :
3809 : 0 : *type_out = vectype;
3810 : :
3811 : : /* Pattern supported. Create a stmt to be used to replace the
3812 : : pattern, with the unpromoted argument. */
3813 : 0 : var = vect_recog_temp_ssa_var (type, NULL);
3814 : 0 : pattern_stmt = gimple_build_call (gimple_call_fndecl (last_stmt),
3815 : : 1, oprnd0);
3816 : 0 : gimple_call_set_lhs (pattern_stmt, var);
3817 : 0 : gimple_call_set_fntype (as_a <gcall *> (pattern_stmt),
3818 : : gimple_call_fntype (last_stmt));
3819 : 0 : return pattern_stmt;
3820 : : }
3821 : 0 : }
3822 : :
3823 : 0 : oprnd1 = build_int_cst (integer_type_node, 8);
3824 : 0 : rhs_code = LROTATE_EXPR;
3825 : 0 : bswap16_p = true;
3826 : : }
3827 : : else
3828 : : return NULL;
3829 : :
3830 : 5627 : if (TREE_CODE (oprnd0) != SSA_NAME
3831 : 5507 : || !INTEGRAL_TYPE_P (type)
3832 : 11061 : || TYPE_PRECISION (TREE_TYPE (lhs)) != TYPE_PRECISION (type))
3833 : : return NULL;
3834 : :
3835 : 5434 : stmt_vec_info def_stmt_info;
3836 : 5434 : if (!vect_is_simple_use (oprnd1, vinfo, &dt, &def_stmt_info, &def_stmt))
3837 : : return NULL;
3838 : :
3839 : 5434 : if (dt != vect_internal_def
3840 : 5228 : && dt != vect_constant_def
3841 : 21 : && dt != vect_external_def)
3842 : : return NULL;
3843 : :
3844 : 5428 : vectype = get_vectype_for_scalar_type (vinfo, type);
3845 : 5428 : if (vectype == NULL_TREE)
3846 : : return NULL;
3847 : :
3848 : : /* If vector/vector or vector/scalar rotate is supported by the target,
3849 : : don't do anything here. */
3850 : 5201 : optab1 = optab_for_tree_code (rhs_code, vectype, optab_vector);
3851 : 5201 : if (optab1
3852 : 5201 : && can_implement_p (optab1, TYPE_MODE (vectype)))
3853 : : {
3854 : 36 : use_rotate:
3855 : 36 : if (bswap16_p)
3856 : : {
3857 : 0 : if (!useless_type_conversion_p (type, TREE_TYPE (oprnd0)))
3858 : : {
3859 : 0 : def = vect_recog_temp_ssa_var (type, NULL);
3860 : 0 : def_stmt = gimple_build_assign (def, NOP_EXPR, oprnd0);
3861 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
3862 : 0 : oprnd0 = def;
3863 : : }
3864 : :
3865 : : /* Pattern detected. */
3866 : 0 : vect_pattern_detected ("vect_recog_rotate_pattern", last_stmt);
3867 : :
3868 : 0 : *type_out = vectype;
3869 : :
3870 : : /* Pattern supported. Create a stmt to be used to replace the
3871 : : pattern. */
3872 : 0 : var = vect_recog_temp_ssa_var (type, NULL);
3873 : 0 : pattern_stmt = gimple_build_assign (var, LROTATE_EXPR, oprnd0,
3874 : : oprnd1);
3875 : 0 : return pattern_stmt;
3876 : : }
3877 : : return NULL;
3878 : : }
3879 : :
3880 : 5165 : if (is_a <bb_vec_info> (vinfo) || dt != vect_internal_def)
3881 : : {
3882 : 5117 : optab2 = optab_for_tree_code (rhs_code, vectype, optab_scalar);
3883 : 5117 : if (optab2
3884 : 5117 : && can_implement_p (optab2, TYPE_MODE (vectype)))
3885 : 0 : goto use_rotate;
3886 : : }
3887 : :
3888 : 5165 : tree utype = unsigned_type_for (type);
3889 : 5165 : tree uvectype = get_vectype_for_scalar_type (vinfo, utype);
3890 : 5165 : if (!uvectype)
3891 : : return NULL;
3892 : :
3893 : : /* If vector/vector or vector/scalar shifts aren't supported by the target,
3894 : : don't do anything here either. */
3895 : 5165 : optab1 = optab_for_tree_code (LSHIFT_EXPR, uvectype, optab_vector);
3896 : 5165 : optab2 = optab_for_tree_code (RSHIFT_EXPR, uvectype, optab_vector);
3897 : 5165 : if (!optab1
3898 : 5165 : || !can_implement_p (optab1, TYPE_MODE (uvectype))
3899 : 519 : || !optab2
3900 : 5684 : || !can_implement_p (optab2, TYPE_MODE (uvectype)))
3901 : : {
3902 : 4646 : if (! is_a <bb_vec_info> (vinfo) && dt == vect_internal_def)
3903 : : return NULL;
3904 : 4611 : optab1 = optab_for_tree_code (LSHIFT_EXPR, uvectype, optab_scalar);
3905 : 4611 : optab2 = optab_for_tree_code (RSHIFT_EXPR, uvectype, optab_scalar);
3906 : 4611 : if (!optab1
3907 : 4611 : || !can_implement_p (optab1, TYPE_MODE (uvectype))
3908 : 3410 : || !optab2
3909 : 8021 : || !can_implement_p (optab2, TYPE_MODE (uvectype)))
3910 : 1201 : return NULL;
3911 : : }
3912 : :
3913 : 3929 : *type_out = vectype;
3914 : :
3915 : 3929 : if (!useless_type_conversion_p (utype, TREE_TYPE (oprnd0)))
3916 : : {
3917 : 49 : def = vect_recog_temp_ssa_var (utype, NULL);
3918 : 49 : def_stmt = gimple_build_assign (def, NOP_EXPR, oprnd0);
3919 : 49 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, uvectype);
3920 : 49 : oprnd0 = def;
3921 : : }
3922 : :
3923 : 3929 : if (dt == vect_external_def && TREE_CODE (oprnd1) == SSA_NAME)
3924 : 13 : ext_def = vect_get_external_def_edge (vinfo, oprnd1);
3925 : :
3926 : 3929 : def = NULL_TREE;
3927 : 3929 : scalar_int_mode mode = SCALAR_INT_TYPE_MODE (utype);
3928 : 3929 : if (dt != vect_internal_def || TYPE_MODE (TREE_TYPE (oprnd1)) == mode)
3929 : : def = oprnd1;
3930 : 28 : else if (def_stmt && gimple_assign_cast_p (def_stmt))
3931 : : {
3932 : 0 : tree rhs1 = gimple_assign_rhs1 (def_stmt);
3933 : 0 : if (TYPE_MODE (TREE_TYPE (rhs1)) == mode
3934 : 0 : && TYPE_PRECISION (TREE_TYPE (rhs1))
3935 : 0 : == TYPE_PRECISION (type))
3936 : : def = rhs1;
3937 : : }
3938 : :
3939 : 3901 : if (def == NULL_TREE)
3940 : : {
3941 : 28 : def = vect_recog_temp_ssa_var (utype, NULL);
3942 : 28 : def_stmt = gimple_build_assign (def, NOP_EXPR, oprnd1);
3943 : 28 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, uvectype);
3944 : : }
3945 : 3929 : stype = TREE_TYPE (def);
3946 : :
3947 : 3929 : if (TREE_CODE (def) == INTEGER_CST)
3948 : : {
3949 : 3817 : if (!tree_fits_uhwi_p (def)
3950 : 3817 : || tree_to_uhwi (def) >= GET_MODE_PRECISION (mode)
3951 : 7634 : || integer_zerop (def))
3952 : 0 : return NULL;
3953 : 3817 : def2 = build_int_cst (stype,
3954 : 3817 : GET_MODE_PRECISION (mode) - tree_to_uhwi (def));
3955 : : }
3956 : : else
3957 : : {
3958 : 112 : tree vecstype = get_vectype_for_scalar_type (vinfo, stype);
3959 : :
3960 : 112 : if (vecstype == NULL_TREE)
3961 : : return NULL;
3962 : 112 : def2 = vect_recog_temp_ssa_var (stype, NULL);
3963 : 112 : def_stmt = gimple_build_assign (def2, NEGATE_EXPR, def);
3964 : 112 : if (ext_def)
3965 : : {
3966 : 13 : basic_block new_bb
3967 : 13 : = gsi_insert_on_edge_immediate (ext_def, def_stmt);
3968 : 13 : gcc_assert (!new_bb);
3969 : : }
3970 : : else
3971 : 99 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, vecstype);
3972 : :
3973 : 112 : def2 = vect_recog_temp_ssa_var (stype, NULL);
3974 : 112 : tree mask = build_int_cst (stype, GET_MODE_PRECISION (mode) - 1);
3975 : 112 : def_stmt = gimple_build_assign (def2, BIT_AND_EXPR,
3976 : : gimple_assign_lhs (def_stmt), mask);
3977 : 112 : if (ext_def)
3978 : : {
3979 : 13 : basic_block new_bb
3980 : 13 : = gsi_insert_on_edge_immediate (ext_def, def_stmt);
3981 : 13 : gcc_assert (!new_bb);
3982 : : }
3983 : : else
3984 : 99 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, vecstype);
3985 : : }
3986 : :
3987 : 3929 : var1 = vect_recog_temp_ssa_var (utype, NULL);
3988 : 7771 : def_stmt = gimple_build_assign (var1, rhs_code == LROTATE_EXPR
3989 : : ? LSHIFT_EXPR : RSHIFT_EXPR,
3990 : : oprnd0, def);
3991 : 3929 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, uvectype);
3992 : :
3993 : 3929 : var2 = vect_recog_temp_ssa_var (utype, NULL);
3994 : 7771 : def_stmt = gimple_build_assign (var2, rhs_code == LROTATE_EXPR
3995 : : ? RSHIFT_EXPR : LSHIFT_EXPR,
3996 : : oprnd0, def2);
3997 : 3929 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, uvectype);
3998 : :
3999 : : /* Pattern detected. */
4000 : 3929 : vect_pattern_detected ("vect_recog_rotate_pattern", last_stmt);
4001 : :
4002 : : /* Pattern supported. Create a stmt to be used to replace the pattern. */
4003 : 3929 : var = vect_recog_temp_ssa_var (utype, NULL);
4004 : 3929 : pattern_stmt = gimple_build_assign (var, BIT_IOR_EXPR, var1, var2);
4005 : :
4006 : 3929 : if (!useless_type_conversion_p (type, utype))
4007 : : {
4008 : 49 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, uvectype);
4009 : 49 : tree result = vect_recog_temp_ssa_var (type, NULL);
4010 : 49 : pattern_stmt = gimple_build_assign (result, NOP_EXPR, var);
4011 : : }
4012 : : return pattern_stmt;
4013 : : }
4014 : :
4015 : : /* Detect a vector by vector shift pattern that wouldn't be otherwise
4016 : : vectorized:
4017 : :
4018 : : type a_t;
4019 : : TYPE b_T, res_T;
4020 : :
4021 : : S1 a_t = ;
4022 : : S2 b_T = ;
4023 : : S3 res_T = b_T op a_t;
4024 : :
4025 : : where type 'TYPE' is a type with different size than 'type',
4026 : : and op is <<, >> or rotate.
4027 : :
4028 : : Also detect cases:
4029 : :
4030 : : type a_t;
4031 : : TYPE b_T, c_T, res_T;
4032 : :
4033 : : S0 c_T = ;
4034 : : S1 a_t = (type) c_T;
4035 : : S2 b_T = ;
4036 : : S3 res_T = b_T op a_t;
4037 : :
4038 : : Input/Output:
4039 : :
4040 : : * STMT_VINFO: The stmt from which the pattern search begins,
4041 : : i.e. the shift/rotate stmt. The original stmt (S3) is replaced
4042 : : with a shift/rotate which has same type on both operands, in the
4043 : : second case just b_T op c_T, in the first case with added cast
4044 : : from a_t to c_T in STMT_VINFO_PATTERN_DEF_SEQ.
4045 : :
4046 : : Output:
4047 : :
4048 : : * TYPE_OUT: The type of the output of this pattern.
4049 : :
4050 : : * Return value: A new stmt that will be used to replace the shift/rotate
4051 : : S3 stmt. */
4052 : :
4053 : : static gimple *
4054 : 29780422 : vect_recog_vector_vector_shift_pattern (vec_info *vinfo,
4055 : : stmt_vec_info stmt_vinfo,
4056 : : tree *type_out)
4057 : : {
4058 : 29780422 : gimple *last_stmt = stmt_vinfo->stmt;
4059 : 29780422 : tree oprnd0, oprnd1, lhs, var;
4060 : 29780422 : gimple *pattern_stmt;
4061 : 29780422 : enum tree_code rhs_code;
4062 : :
4063 : 29780422 : if (!is_gimple_assign (last_stmt))
4064 : : return NULL;
4065 : :
4066 : 20351118 : rhs_code = gimple_assign_rhs_code (last_stmt);
4067 : 20351118 : switch (rhs_code)
4068 : : {
4069 : 520544 : case LSHIFT_EXPR:
4070 : 520544 : case RSHIFT_EXPR:
4071 : 520544 : case LROTATE_EXPR:
4072 : 520544 : case RROTATE_EXPR:
4073 : 520544 : break;
4074 : : default:
4075 : : return NULL;
4076 : : }
4077 : :
4078 : 520544 : lhs = gimple_assign_lhs (last_stmt);
4079 : 520544 : oprnd0 = gimple_assign_rhs1 (last_stmt);
4080 : 520544 : oprnd1 = gimple_assign_rhs2 (last_stmt);
4081 : 520544 : if (TREE_CODE (oprnd0) != SSA_NAME
4082 : 469784 : || TREE_CODE (oprnd1) != SSA_NAME
4083 : 47518 : || TYPE_MODE (TREE_TYPE (oprnd0)) == TYPE_MODE (TREE_TYPE (oprnd1))
4084 : 18672 : || !INTEGRAL_TYPE_P (TREE_TYPE (oprnd0))
4085 : 18390 : || !type_has_mode_precision_p (TREE_TYPE (oprnd1))
4086 : 538934 : || TYPE_PRECISION (TREE_TYPE (lhs))
4087 : 18390 : != TYPE_PRECISION (TREE_TYPE (oprnd0)))
4088 : 502154 : return NULL;
4089 : :
4090 : 18390 : stmt_vec_info def_vinfo = vect_get_internal_def (vinfo, oprnd1);
4091 : 18390 : if (!def_vinfo)
4092 : : return NULL;
4093 : :
4094 : 16441 : *type_out = get_vectype_for_scalar_type (vinfo, TREE_TYPE (oprnd0));
4095 : 16441 : if (*type_out == NULL_TREE)
4096 : : return NULL;
4097 : :
4098 : 11084 : tree def = NULL_TREE;
4099 : 11084 : gassign *def_stmt = dyn_cast <gassign *> (def_vinfo->stmt);
4100 : 9311 : if (def_stmt && gimple_assign_cast_p (def_stmt))
4101 : : {
4102 : 1948 : tree rhs1 = gimple_assign_rhs1 (def_stmt);
4103 : 1948 : if (TYPE_MODE (TREE_TYPE (rhs1)) == TYPE_MODE (TREE_TYPE (oprnd0))
4104 : 1948 : && TYPE_PRECISION (TREE_TYPE (rhs1))
4105 : 593 : == TYPE_PRECISION (TREE_TYPE (oprnd0)))
4106 : : {
4107 : 593 : if (TYPE_PRECISION (TREE_TYPE (oprnd1))
4108 : 593 : >= TYPE_PRECISION (TREE_TYPE (rhs1)))
4109 : : def = rhs1;
4110 : : else
4111 : : {
4112 : 588 : tree mask
4113 : 588 : = build_low_bits_mask (TREE_TYPE (rhs1),
4114 : 588 : TYPE_PRECISION (TREE_TYPE (oprnd1)));
4115 : 588 : def = vect_recog_temp_ssa_var (TREE_TYPE (rhs1), NULL);
4116 : 588 : def_stmt = gimple_build_assign (def, BIT_AND_EXPR, rhs1, mask);
4117 : 588 : tree vecstype = get_vectype_for_scalar_type (vinfo,
4118 : 588 : TREE_TYPE (rhs1));
4119 : 588 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, vecstype);
4120 : : }
4121 : : }
4122 : : }
4123 : :
4124 : 593 : if (def == NULL_TREE)
4125 : : {
4126 : 10491 : def = vect_recog_temp_ssa_var (TREE_TYPE (oprnd0), NULL);
4127 : 10491 : def_stmt = gimple_build_assign (def, NOP_EXPR, oprnd1);
4128 : 10491 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
4129 : : }
4130 : :
4131 : : /* Pattern detected. */
4132 : 11084 : vect_pattern_detected ("vect_recog_vector_vector_shift_pattern", last_stmt);
4133 : :
4134 : : /* Pattern supported. Create a stmt to be used to replace the pattern. */
4135 : 11084 : var = vect_recog_temp_ssa_var (TREE_TYPE (oprnd0), NULL);
4136 : 11084 : pattern_stmt = gimple_build_assign (var, rhs_code, oprnd0, def);
4137 : :
4138 : 11084 : return pattern_stmt;
4139 : : }
4140 : :
4141 : : /* Return true iff the target has a vector optab implementing the operation
4142 : : CODE on type VECTYPE. */
4143 : :
4144 : : static bool
4145 : 583574 : target_has_vecop_for_code (tree_code code, tree vectype)
4146 : : {
4147 : 583574 : optab voptab = optab_for_tree_code (code, vectype, optab_vector);
4148 : 583574 : return voptab
4149 : 583574 : && can_implement_p (voptab, TYPE_MODE (vectype));
4150 : : }
4151 : :
4152 : : /* Verify that the target has optabs of VECTYPE to perform all the steps
4153 : : needed by the multiplication-by-immediate synthesis algorithm described by
4154 : : ALG and VAR. If SYNTH_SHIFT_P is true ensure that vector addition is
4155 : : present. Return true iff the target supports all the steps. */
4156 : :
4157 : : static bool
4158 : 258399 : target_supports_mult_synth_alg (struct algorithm *alg, mult_variant var,
4159 : : tree vectype, bool synth_shift_p)
4160 : : {
4161 : 258399 : if (alg->op[0] != alg_zero && alg->op[0] != alg_m)
4162 : : return false;
4163 : :
4164 : 258399 : bool supports_vminus = target_has_vecop_for_code (MINUS_EXPR, vectype);
4165 : 258399 : bool supports_vplus = target_has_vecop_for_code (PLUS_EXPR, vectype);
4166 : :
4167 : 258399 : if (var == negate_variant
4168 : 258399 : && !target_has_vecop_for_code (NEGATE_EXPR, vectype))
4169 : : return false;
4170 : :
4171 : : /* If we must synthesize shifts with additions make sure that vector
4172 : : addition is available. */
4173 : 257922 : if ((var == add_variant || synth_shift_p) && !supports_vplus)
4174 : : return false;
4175 : :
4176 : 141846 : for (int i = 1; i < alg->ops; i++)
4177 : : {
4178 : 110142 : switch (alg->op[i])
4179 : : {
4180 : : case alg_shift:
4181 : : break;
4182 : 28691 : case alg_add_t_m2:
4183 : 28691 : case alg_add_t2_m:
4184 : 28691 : case alg_add_factor:
4185 : 28691 : if (!supports_vplus)
4186 : : return false;
4187 : : break;
4188 : 17997 : case alg_sub_t_m2:
4189 : 17997 : case alg_sub_t2_m:
4190 : 17997 : case alg_sub_factor:
4191 : 17997 : if (!supports_vminus)
4192 : : return false;
4193 : : break;
4194 : : case alg_unknown:
4195 : : case alg_m:
4196 : : case alg_zero:
4197 : : case alg_impossible:
4198 : : return false;
4199 : 0 : default:
4200 : 0 : gcc_unreachable ();
4201 : : }
4202 : : }
4203 : :
4204 : : return true;
4205 : : }
4206 : :
4207 : : /* Synthesize a left shift of OP by AMNT bits using a series of additions and
4208 : : putting the final result in DEST. Append all statements but the last into
4209 : : VINFO. Return the last statement. */
4210 : :
4211 : : static gimple *
4212 : 0 : synth_lshift_by_additions (vec_info *vinfo,
4213 : : tree dest, tree op, HOST_WIDE_INT amnt,
4214 : : stmt_vec_info stmt_info)
4215 : : {
4216 : 0 : HOST_WIDE_INT i;
4217 : 0 : tree itype = TREE_TYPE (op);
4218 : 0 : tree prev_res = op;
4219 : 0 : gcc_assert (amnt >= 0);
4220 : 0 : for (i = 0; i < amnt; i++)
4221 : : {
4222 : 0 : tree tmp_var = (i < amnt - 1) ? vect_recog_temp_ssa_var (itype, NULL)
4223 : : : dest;
4224 : 0 : gimple *stmt
4225 : 0 : = gimple_build_assign (tmp_var, PLUS_EXPR, prev_res, prev_res);
4226 : 0 : prev_res = tmp_var;
4227 : 0 : if (i < amnt - 1)
4228 : 0 : append_pattern_def_seq (vinfo, stmt_info, stmt);
4229 : : else
4230 : 0 : return stmt;
4231 : : }
4232 : 0 : gcc_unreachable ();
4233 : : return NULL;
4234 : : }
4235 : :
4236 : : /* Helper for vect_synth_mult_by_constant. Apply a binary operation
4237 : : CODE to operands OP1 and OP2, creating a new temporary SSA var in
4238 : : the process if necessary. Append the resulting assignment statements
4239 : : to the sequence in STMT_VINFO. Return the SSA variable that holds the
4240 : : result of the binary operation. If SYNTH_SHIFT_P is true synthesize
4241 : : left shifts using additions. */
4242 : :
4243 : : static tree
4244 : 46595 : apply_binop_and_append_stmt (vec_info *vinfo,
4245 : : tree_code code, tree op1, tree op2,
4246 : : stmt_vec_info stmt_vinfo, bool synth_shift_p)
4247 : : {
4248 : 46595 : if (integer_zerop (op2)
4249 : 46595 : && (code == LSHIFT_EXPR
4250 : 40840 : || code == PLUS_EXPR))
4251 : : {
4252 : 40840 : gcc_assert (TREE_CODE (op1) == SSA_NAME);
4253 : : return op1;
4254 : : }
4255 : :
4256 : 5755 : gimple *stmt;
4257 : 5755 : tree itype = TREE_TYPE (op1);
4258 : 5755 : tree tmp_var = vect_recog_temp_ssa_var (itype, NULL);
4259 : :
4260 : 5755 : if (code == LSHIFT_EXPR
4261 : 5755 : && synth_shift_p)
4262 : : {
4263 : 0 : stmt = synth_lshift_by_additions (vinfo, tmp_var, op1,
4264 : 0 : TREE_INT_CST_LOW (op2), stmt_vinfo);
4265 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, stmt);
4266 : 0 : return tmp_var;
4267 : : }
4268 : :
4269 : 5755 : stmt = gimple_build_assign (tmp_var, code, op1, op2);
4270 : 5755 : append_pattern_def_seq (vinfo, stmt_vinfo, stmt);
4271 : 5755 : return tmp_var;
4272 : : }
4273 : :
4274 : : /* Synthesize a multiplication of OP by an INTEGER_CST VAL using shifts
4275 : : and simple arithmetic operations to be vectorized. Record the statements
4276 : : produced in STMT_VINFO and return the last statement in the sequence or
4277 : : NULL if it's not possible to synthesize such a multiplication.
4278 : : This function mirrors the behavior of expand_mult_const in expmed.cc but
4279 : : works on tree-ssa form. */
4280 : :
4281 : : static gimple *
4282 : 261073 : vect_synth_mult_by_constant (vec_info *vinfo, tree op, tree val,
4283 : : stmt_vec_info stmt_vinfo)
4284 : : {
4285 : 261073 : tree itype = TREE_TYPE (op);
4286 : 261073 : machine_mode mode = TYPE_MODE (itype);
4287 : 261073 : struct algorithm alg;
4288 : 261073 : mult_variant variant;
4289 : 261073 : if (!tree_fits_shwi_p (val))
4290 : : return NULL;
4291 : :
4292 : : /* Multiplication synthesis by shifts, adds and subs can introduce
4293 : : signed overflow where the original operation didn't. Perform the
4294 : : operations on an unsigned type and cast back to avoid this.
4295 : : In the future we may want to relax this for synthesis algorithms
4296 : : that we can prove do not cause unexpected overflow. */
4297 : 258399 : bool cast_to_unsigned_p = !TYPE_OVERFLOW_WRAPS (itype);
4298 : :
4299 : 54270 : tree multtype = cast_to_unsigned_p ? unsigned_type_for (itype) : itype;
4300 : 258399 : tree vectype = get_vectype_for_scalar_type (vinfo, multtype);
4301 : 258399 : if (!vectype)
4302 : : return NULL;
4303 : :
4304 : : /* Targets that don't support vector shifts but support vector additions
4305 : : can synthesize shifts that way. */
4306 : 258399 : bool synth_shift_p = !vect_supportable_shift (vinfo, LSHIFT_EXPR, multtype);
4307 : :
4308 : 258399 : HOST_WIDE_INT hwval = tree_to_shwi (val);
4309 : : /* Use MAX_COST here as we don't want to limit the sequence on rtx costs.
4310 : : The vectorizer's benefit analysis will decide whether it's beneficial
4311 : : to do this. */
4312 : 516798 : bool possible = choose_mult_variant (VECTOR_MODE_P (TYPE_MODE (vectype))
4313 : 258399 : ? TYPE_MODE (vectype) : mode,
4314 : : hwval, &alg, &variant, MAX_COST);
4315 : 258399 : if (!possible)
4316 : : return NULL;
4317 : :
4318 : 258399 : if (!target_supports_mult_synth_alg (&alg, variant, vectype, synth_shift_p))
4319 : : return NULL;
4320 : :
4321 : 31704 : tree accumulator;
4322 : :
4323 : : /* Clear out the sequence of statements so we can populate it below. */
4324 : 31704 : gimple *stmt = NULL;
4325 : :
4326 : 31704 : if (cast_to_unsigned_p)
4327 : : {
4328 : 11942 : tree tmp_op = vect_recog_temp_ssa_var (multtype, NULL);
4329 : 11942 : stmt = gimple_build_assign (tmp_op, CONVERT_EXPR, op);
4330 : 11942 : append_pattern_def_seq (vinfo, stmt_vinfo, stmt);
4331 : 11942 : op = tmp_op;
4332 : : }
4333 : :
4334 : 31704 : if (alg.op[0] == alg_zero)
4335 : 177 : accumulator = build_int_cst (multtype, 0);
4336 : : else
4337 : : accumulator = op;
4338 : :
4339 : 31704 : bool needs_fixup = (variant == negate_variant)
4340 : 31704 : || (variant == add_variant);
4341 : :
4342 : 141693 : for (int i = 1; i < alg.ops; i++)
4343 : : {
4344 : 109989 : tree shft_log = build_int_cst (multtype, alg.log[i]);
4345 : 109989 : tree accum_tmp = vect_recog_temp_ssa_var (multtype, NULL);
4346 : 109989 : tree tmp_var = NULL_TREE;
4347 : :
4348 : 109989 : switch (alg.op[i])
4349 : : {
4350 : 63394 : case alg_shift:
4351 : 63394 : if (synth_shift_p)
4352 : 0 : stmt
4353 : 0 : = synth_lshift_by_additions (vinfo, accum_tmp, accumulator,
4354 : 0 : alg.log[i], stmt_vinfo);
4355 : : else
4356 : 63394 : stmt = gimple_build_assign (accum_tmp, LSHIFT_EXPR, accumulator,
4357 : : shft_log);
4358 : : break;
4359 : 23752 : case alg_add_t_m2:
4360 : 23752 : tmp_var
4361 : 23752 : = apply_binop_and_append_stmt (vinfo, LSHIFT_EXPR, op, shft_log,
4362 : : stmt_vinfo, synth_shift_p);
4363 : 23752 : stmt = gimple_build_assign (accum_tmp, PLUS_EXPR, accumulator,
4364 : : tmp_var);
4365 : 23752 : break;
4366 : 17242 : case alg_sub_t_m2:
4367 : 17242 : tmp_var = apply_binop_and_append_stmt (vinfo, LSHIFT_EXPR, op,
4368 : : shft_log, stmt_vinfo,
4369 : : synth_shift_p);
4370 : : /* In some algorithms the first step involves zeroing the
4371 : : accumulator. If subtracting from such an accumulator
4372 : : just emit the negation directly. */
4373 : 17242 : if (integer_zerop (accumulator))
4374 : 177 : stmt = gimple_build_assign (accum_tmp, NEGATE_EXPR, tmp_var);
4375 : : else
4376 : 17065 : stmt = gimple_build_assign (accum_tmp, MINUS_EXPR, accumulator,
4377 : : tmp_var);
4378 : : break;
4379 : 0 : case alg_add_t2_m:
4380 : 0 : tmp_var
4381 : 0 : = apply_binop_and_append_stmt (vinfo, LSHIFT_EXPR, accumulator,
4382 : : shft_log, stmt_vinfo, synth_shift_p);
4383 : 0 : stmt = gimple_build_assign (accum_tmp, PLUS_EXPR, tmp_var, op);
4384 : 0 : break;
4385 : 0 : case alg_sub_t2_m:
4386 : 0 : tmp_var
4387 : 0 : = apply_binop_and_append_stmt (vinfo, LSHIFT_EXPR, accumulator,
4388 : : shft_log, stmt_vinfo, synth_shift_p);
4389 : 0 : stmt = gimple_build_assign (accum_tmp, MINUS_EXPR, tmp_var, op);
4390 : 0 : break;
4391 : 4881 : case alg_add_factor:
4392 : 4881 : tmp_var
4393 : 4881 : = apply_binop_and_append_stmt (vinfo, LSHIFT_EXPR, accumulator,
4394 : : shft_log, stmt_vinfo, synth_shift_p);
4395 : 4881 : stmt = gimple_build_assign (accum_tmp, PLUS_EXPR, accumulator,
4396 : : tmp_var);
4397 : 4881 : break;
4398 : 720 : case alg_sub_factor:
4399 : 720 : tmp_var
4400 : 720 : = apply_binop_and_append_stmt (vinfo, LSHIFT_EXPR, accumulator,
4401 : : shft_log, stmt_vinfo, synth_shift_p);
4402 : 720 : stmt = gimple_build_assign (accum_tmp, MINUS_EXPR, tmp_var,
4403 : : accumulator);
4404 : 720 : break;
4405 : 0 : default:
4406 : 0 : gcc_unreachable ();
4407 : : }
4408 : : /* We don't want to append the last stmt in the sequence to stmt_vinfo
4409 : : but rather return it directly. */
4410 : :
4411 : 109989 : if ((i < alg.ops - 1) || needs_fixup || cast_to_unsigned_p)
4412 : 90474 : append_pattern_def_seq (vinfo, stmt_vinfo, stmt);
4413 : 109989 : accumulator = accum_tmp;
4414 : : }
4415 : 31704 : if (variant == negate_variant)
4416 : : {
4417 : 361 : tree accum_tmp = vect_recog_temp_ssa_var (multtype, NULL);
4418 : 361 : stmt = gimple_build_assign (accum_tmp, NEGATE_EXPR, accumulator);
4419 : 361 : accumulator = accum_tmp;
4420 : 361 : if (cast_to_unsigned_p)
4421 : 124 : append_pattern_def_seq (vinfo, stmt_vinfo, stmt);
4422 : : }
4423 : 31343 : else if (variant == add_variant)
4424 : : {
4425 : 68 : tree accum_tmp = vect_recog_temp_ssa_var (multtype, NULL);
4426 : 68 : stmt = gimple_build_assign (accum_tmp, PLUS_EXPR, accumulator, op);
4427 : 68 : accumulator = accum_tmp;
4428 : 68 : if (cast_to_unsigned_p)
4429 : 58 : append_pattern_def_seq (vinfo, stmt_vinfo, stmt);
4430 : : }
4431 : : /* Move back to a signed if needed. */
4432 : 31457 : if (cast_to_unsigned_p)
4433 : : {
4434 : 11942 : tree accum_tmp = vect_recog_temp_ssa_var (itype, NULL);
4435 : 11942 : stmt = gimple_build_assign (accum_tmp, CONVERT_EXPR, accumulator);
4436 : : }
4437 : :
4438 : : return stmt;
4439 : : }
4440 : :
4441 : : /* Detect multiplication by constant and convert it into a sequence of
4442 : : shifts and additions, subtractions, negations. We reuse the
4443 : : choose_mult_variant algorithms from expmed.cc
4444 : :
4445 : : Input/Output:
4446 : :
4447 : : STMT_VINFO: The stmt from which the pattern search begins,
4448 : : i.e. the mult stmt.
4449 : :
4450 : : Output:
4451 : :
4452 : : * TYPE_OUT: The type of the output of this pattern.
4453 : :
4454 : : * Return value: A new stmt that will be used to replace
4455 : : the multiplication. */
4456 : :
4457 : : static gimple *
4458 : 29964791 : vect_recog_mult_pattern (vec_info *vinfo,
4459 : : stmt_vec_info stmt_vinfo, tree *type_out)
4460 : : {
4461 : 29964791 : gimple *last_stmt = stmt_vinfo->stmt;
4462 : 29964791 : tree oprnd0, oprnd1, vectype, itype;
4463 : 29964791 : gimple *pattern_stmt;
4464 : :
4465 : 29964791 : if (!is_gimple_assign (last_stmt))
4466 : : return NULL;
4467 : :
4468 : 20535487 : if (gimple_assign_rhs_code (last_stmt) != MULT_EXPR)
4469 : : return NULL;
4470 : :
4471 : 1317377 : oprnd0 = gimple_assign_rhs1 (last_stmt);
4472 : 1317377 : oprnd1 = gimple_assign_rhs2 (last_stmt);
4473 : 1317377 : itype = TREE_TYPE (oprnd0);
4474 : :
4475 : 1317377 : if (TREE_CODE (oprnd0) != SSA_NAME
4476 : 1317314 : || TREE_CODE (oprnd1) != INTEGER_CST
4477 : 808354 : || !INTEGRAL_TYPE_P (itype)
4478 : 2125731 : || !type_has_mode_precision_p (itype))
4479 : 509075 : return NULL;
4480 : :
4481 : 808302 : vectype = get_vectype_for_scalar_type (vinfo, itype);
4482 : 808302 : if (vectype == NULL_TREE)
4483 : : return NULL;
4484 : :
4485 : : /* If the target can handle vectorized multiplication natively,
4486 : : don't attempt to optimize this. */
4487 : 657127 : optab mul_optab = optab_for_tree_code (MULT_EXPR, vectype, optab_default);
4488 : 657127 : if (mul_optab != unknown_optab
4489 : 657127 : && can_implement_p (mul_optab, TYPE_MODE (vectype)))
4490 : : return NULL;
4491 : :
4492 : 261073 : pattern_stmt = vect_synth_mult_by_constant (vinfo,
4493 : : oprnd0, oprnd1, stmt_vinfo);
4494 : 261073 : if (!pattern_stmt)
4495 : : return NULL;
4496 : :
4497 : : /* Pattern detected. */
4498 : 31704 : vect_pattern_detected ("vect_recog_mult_pattern", last_stmt);
4499 : :
4500 : 31704 : *type_out = vectype;
4501 : :
4502 : 31704 : return pattern_stmt;
4503 : : }
4504 : :
4505 : : extern bool gimple_unsigned_integer_sat_add (tree, tree*, tree (*)(tree));
4506 : : extern bool gimple_unsigned_integer_sat_sub (tree, tree*, tree (*)(tree));
4507 : : extern bool gimple_unsigned_integer_sat_trunc (tree, tree*, tree (*)(tree));
4508 : :
4509 : : extern bool gimple_signed_integer_sat_add (tree, tree*, tree (*)(tree));
4510 : : extern bool gimple_signed_integer_sat_sub (tree, tree*, tree (*)(tree));
4511 : : extern bool gimple_signed_integer_sat_trunc (tree, tree*, tree (*)(tree));
4512 : :
4513 : : static gimple *
4514 : 257 : vect_recog_build_binary_gimple_stmt (vec_info *vinfo, stmt_vec_info stmt_info,
4515 : : internal_fn fn, tree *type_out,
4516 : : tree lhs, tree op_0, tree op_1)
4517 : : {
4518 : 257 : tree itype = TREE_TYPE (op_0);
4519 : 257 : tree otype = TREE_TYPE (lhs);
4520 : 257 : tree v_itype = get_vectype_for_scalar_type (vinfo, itype);
4521 : 257 : tree v_otype = get_vectype_for_scalar_type (vinfo, otype);
4522 : :
4523 : 257 : if (v_itype != NULL_TREE && v_otype != NULL_TREE
4524 : 257 : && direct_internal_fn_supported_p (fn, v_itype, OPTIMIZE_FOR_BOTH))
4525 : : {
4526 : 61 : gcall *call = gimple_build_call_internal (fn, 2, op_0, op_1);
4527 : 61 : tree in_ssa = vect_recog_temp_ssa_var (itype, NULL);
4528 : :
4529 : 61 : gimple_call_set_lhs (call, in_ssa);
4530 : 61 : gimple_call_set_nothrow (call, /* nothrow_p */ false);
4531 : 61 : gimple_set_location (call, gimple_location (STMT_VINFO_STMT (stmt_info)));
4532 : :
4533 : 61 : *type_out = v_otype;
4534 : :
4535 : 61 : if (types_compatible_p (itype, otype))
4536 : : return call;
4537 : : else
4538 : : {
4539 : 0 : append_pattern_def_seq (vinfo, stmt_info, call, v_itype);
4540 : 0 : tree out_ssa = vect_recog_temp_ssa_var (otype, NULL);
4541 : :
4542 : 0 : return gimple_build_assign (out_ssa, NOP_EXPR, in_ssa);
4543 : : }
4544 : : }
4545 : :
4546 : : return NULL;
4547 : : }
4548 : :
4549 : : /*
4550 : : * Try to detect saturation add pattern (SAT_ADD), aka below gimple:
4551 : : * _7 = _4 + _6;
4552 : : * _8 = _4 > _7;
4553 : : * _9 = (long unsigned int) _8;
4554 : : * _10 = -_9;
4555 : : * _12 = _7 | _10;
4556 : : *
4557 : : * And then simplied to
4558 : : * _12 = .SAT_ADD (_4, _6);
4559 : : */
4560 : :
4561 : : static gimple *
4562 : 30045813 : vect_recog_sat_add_pattern (vec_info *vinfo, stmt_vec_info stmt_vinfo,
4563 : : tree *type_out)
4564 : : {
4565 : 30045813 : gimple *last_stmt = STMT_VINFO_STMT (stmt_vinfo);
4566 : :
4567 : 30045813 : if (!is_gimple_assign (last_stmt))
4568 : : return NULL;
4569 : :
4570 : 20616509 : tree ops[2];
4571 : 20616509 : tree lhs = gimple_assign_lhs (last_stmt);
4572 : :
4573 : 20616509 : if (gimple_unsigned_integer_sat_add (lhs, ops, NULL)
4574 : 20616509 : || gimple_signed_integer_sat_add (lhs, ops, NULL))
4575 : : {
4576 : 50 : if (TREE_CODE (ops[1]) == INTEGER_CST)
4577 : 12 : ops[1] = fold_convert (TREE_TYPE (ops[0]), ops[1]);
4578 : :
4579 : 50 : gimple *stmt = vect_recog_build_binary_gimple_stmt (vinfo, stmt_vinfo,
4580 : : IFN_SAT_ADD, type_out,
4581 : : lhs, ops[0], ops[1]);
4582 : 50 : if (stmt)
4583 : : {
4584 : 32 : vect_pattern_detected ("vect_recog_sat_add_pattern", last_stmt);
4585 : 32 : return stmt;
4586 : : }
4587 : : }
4588 : :
4589 : : return NULL;
4590 : : }
4591 : :
4592 : : /*
4593 : : * Try to transform the truncation for .SAT_SUB pattern, mostly occurs in
4594 : : * the benchmark zip. Aka:
4595 : : *
4596 : : * unsigned int _1;
4597 : : * unsigned int _2;
4598 : : * unsigned short int _4;
4599 : : * _9 = (unsigned short int).SAT_SUB (_1, _2);
4600 : : *
4601 : : * if _1 is known to be in the range of unsigned short int. For example
4602 : : * there is a def _1 = (unsigned short int)_4. Then we can transform the
4603 : : * truncation to:
4604 : : *
4605 : : * _3 = (unsigned short int) MIN (65535, _2); // aka _3 = .SAT_TRUNC (_2);
4606 : : * _9 = .SAT_SUB (_4, _3);
4607 : : *
4608 : : * Then, we can better vectorized code and avoid the unnecessary narrowing
4609 : : * stmt during vectorization with below stmt(s).
4610 : : *
4611 : : * _3 = .SAT_TRUNC(_2); // SI => HI
4612 : : * _9 = .SAT_SUB (_4, _3);
4613 : : */
4614 : : static void
4615 : 207 : vect_recog_sat_sub_pattern_transform (vec_info *vinfo,
4616 : : stmt_vec_info stmt_vinfo,
4617 : : tree lhs, tree *ops)
4618 : : {
4619 : 207 : tree otype = TREE_TYPE (lhs);
4620 : 207 : tree itype = TREE_TYPE (ops[0]);
4621 : 207 : unsigned itype_prec = TYPE_PRECISION (itype);
4622 : 207 : unsigned otype_prec = TYPE_PRECISION (otype);
4623 : :
4624 : 207 : if (types_compatible_p (otype, itype) || otype_prec >= itype_prec)
4625 : 207 : return;
4626 : :
4627 : 0 : tree v_otype = get_vectype_for_scalar_type (vinfo, otype);
4628 : 0 : tree v_itype = get_vectype_for_scalar_type (vinfo, itype);
4629 : 0 : tree_pair v_pair = tree_pair (v_otype, v_itype);
4630 : :
4631 : 0 : if (v_otype == NULL_TREE || v_itype == NULL_TREE
4632 : 0 : || !direct_internal_fn_supported_p (IFN_SAT_TRUNC, v_pair,
4633 : : OPTIMIZE_FOR_BOTH))
4634 : 0 : return;
4635 : :
4636 : : /* 1. Find the _4 and update ops[0] as above example. */
4637 : 0 : vect_unpromoted_value unprom;
4638 : 0 : tree tmp = vect_look_through_possible_promotion (vinfo, ops[0], &unprom);
4639 : :
4640 : 0 : if (tmp == NULL_TREE || TYPE_PRECISION (unprom.type) != otype_prec)
4641 : : return;
4642 : :
4643 : 0 : ops[0] = tmp;
4644 : :
4645 : : /* 2. Generate _3 = .SAT_TRUNC (_2) and update ops[1] as above example. */
4646 : 0 : tree trunc_lhs_ssa = vect_recog_temp_ssa_var (otype, NULL);
4647 : 0 : gcall *call = gimple_build_call_internal (IFN_SAT_TRUNC, 1, ops[1]);
4648 : :
4649 : 0 : gimple_call_set_lhs (call, trunc_lhs_ssa);
4650 : 0 : gimple_call_set_nothrow (call, /* nothrow_p */ false);
4651 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, call, v_otype);
4652 : :
4653 : 0 : ops[1] = trunc_lhs_ssa;
4654 : : }
4655 : :
4656 : : /*
4657 : : * Try to detect saturation sub pattern (SAT_ADD), aka below gimple:
4658 : : * Unsigned:
4659 : : * _7 = _1 >= _2;
4660 : : * _8 = _1 - _2;
4661 : : * _10 = (long unsigned int) _7;
4662 : : * _9 = _8 * _10;
4663 : : *
4664 : : * And then simplied to
4665 : : * _9 = .SAT_SUB (_1, _2);
4666 : : *
4667 : : * Signed:
4668 : : * x.0_4 = (unsigned char) x_16;
4669 : : * y.1_5 = (unsigned char) y_18;
4670 : : * _6 = x.0_4 - y.1_5;
4671 : : * minus_19 = (int8_t) _6;
4672 : : * _7 = x_16 ^ y_18;
4673 : : * _8 = x_16 ^ minus_19;
4674 : : * _44 = _7 < 0;
4675 : : * _23 = x_16 < 0;
4676 : : * _24 = (signed char) _23;
4677 : : * _58 = (unsigned char) _24;
4678 : : * _59 = -_58;
4679 : : * _25 = (signed char) _59;
4680 : : * _26 = _25 ^ 127;
4681 : : * _42 = _8 < 0;
4682 : : * _41 = _42 & _44;
4683 : : * iftmp.2_11 = _41 ? _26 : minus_19;
4684 : : *
4685 : : * And then simplied to
4686 : : * iftmp.2_11 = .SAT_SUB (x_16, y_18);
4687 : : */
4688 : :
4689 : : static gimple *
4690 : 30045781 : vect_recog_sat_sub_pattern (vec_info *vinfo, stmt_vec_info stmt_vinfo,
4691 : : tree *type_out)
4692 : : {
4693 : 30045781 : gimple *last_stmt = STMT_VINFO_STMT (stmt_vinfo);
4694 : :
4695 : 30045781 : if (!is_gimple_assign (last_stmt))
4696 : : return NULL;
4697 : :
4698 : 20616477 : tree ops[2];
4699 : 20616477 : tree lhs = gimple_assign_lhs (last_stmt);
4700 : :
4701 : 20616477 : if (gimple_unsigned_integer_sat_sub (lhs, ops, NULL)
4702 : 20616477 : || gimple_signed_integer_sat_sub (lhs, ops, NULL))
4703 : : {
4704 : 207 : vect_recog_sat_sub_pattern_transform (vinfo, stmt_vinfo, lhs, ops);
4705 : 207 : gimple *stmt = vect_recog_build_binary_gimple_stmt (vinfo, stmt_vinfo,
4706 : : IFN_SAT_SUB, type_out,
4707 : : lhs, ops[0], ops[1]);
4708 : 207 : if (stmt)
4709 : : {
4710 : 29 : vect_pattern_detected ("vect_recog_sat_sub_pattern", last_stmt);
4711 : 29 : return stmt;
4712 : : }
4713 : : }
4714 : :
4715 : : return NULL;
4716 : : }
4717 : :
4718 : : /*
4719 : : * Try to detect saturation truncation pattern (SAT_TRUNC), aka below gimple:
4720 : : * overflow_5 = x_4(D) > 4294967295;
4721 : : * _1 = (unsigned int) x_4(D);
4722 : : * _2 = (unsigned int) overflow_5;
4723 : : * _3 = -_2;
4724 : : * _6 = _1 | _3;
4725 : : *
4726 : : * And then simplied to
4727 : : * _6 = .SAT_TRUNC (x_4(D));
4728 : : */
4729 : :
4730 : : static gimple *
4731 : 30045752 : vect_recog_sat_trunc_pattern (vec_info *vinfo, stmt_vec_info stmt_vinfo,
4732 : : tree *type_out)
4733 : : {
4734 : 30045752 : gimple *last_stmt = STMT_VINFO_STMT (stmt_vinfo);
4735 : :
4736 : 30045752 : if (!is_gimple_assign (last_stmt))
4737 : : return NULL;
4738 : :
4739 : 20616448 : tree ops[1];
4740 : 20616448 : tree lhs = gimple_assign_lhs (last_stmt);
4741 : 20616448 : tree otype = TREE_TYPE (lhs);
4742 : :
4743 : 20616448 : if ((gimple_unsigned_integer_sat_trunc (lhs, ops, NULL)
4744 : 20616177 : || gimple_signed_integer_sat_trunc (lhs, ops, NULL))
4745 : 20616448 : && type_has_mode_precision_p (otype))
4746 : : {
4747 : 259 : tree itype = TREE_TYPE (ops[0]);
4748 : 259 : tree v_itype = get_vectype_for_scalar_type (vinfo, itype);
4749 : 259 : tree v_otype = get_vectype_for_scalar_type (vinfo, otype);
4750 : 259 : internal_fn fn = IFN_SAT_TRUNC;
4751 : :
4752 : 253 : if (v_itype != NULL_TREE && v_otype != NULL_TREE
4753 : 512 : && direct_internal_fn_supported_p (fn, tree_pair (v_otype, v_itype),
4754 : : OPTIMIZE_FOR_BOTH))
4755 : : {
4756 : 0 : gcall *call = gimple_build_call_internal (fn, 1, ops[0]);
4757 : 0 : tree out_ssa = vect_recog_temp_ssa_var (otype, NULL);
4758 : :
4759 : 0 : gimple_call_set_lhs (call, out_ssa);
4760 : 0 : gimple_call_set_nothrow (call, /* nothrow_p */ false);
4761 : 0 : gimple_set_location (call, gimple_location (last_stmt));
4762 : :
4763 : 0 : *type_out = v_otype;
4764 : :
4765 : 0 : return call;
4766 : : }
4767 : : }
4768 : :
4769 : : return NULL;
4770 : : }
4771 : :
4772 : : /* Detect a signed division by a constant that wouldn't be
4773 : : otherwise vectorized:
4774 : :
4775 : : type a_t, b_t;
4776 : :
4777 : : S1 a_t = b_t / N;
4778 : :
4779 : : where type 'type' is an integral type and N is a constant.
4780 : :
4781 : : Similarly handle modulo by a constant:
4782 : :
4783 : : S4 a_t = b_t % N;
4784 : :
4785 : : Input/Output:
4786 : :
4787 : : * STMT_VINFO: The stmt from which the pattern search begins,
4788 : : i.e. the division stmt. S1 is replaced by if N is a power
4789 : : of two constant and type is signed:
4790 : : S3 y_t = b_t < 0 ? N - 1 : 0;
4791 : : S2 x_t = b_t + y_t;
4792 : : S1' a_t = x_t >> log2 (N);
4793 : :
4794 : : S4 is replaced if N is a power of two constant and
4795 : : type is signed by (where *_T temporaries have unsigned type):
4796 : : S9 y_T = b_t < 0 ? -1U : 0U;
4797 : : S8 z_T = y_T >> (sizeof (type_t) * CHAR_BIT - log2 (N));
4798 : : S7 z_t = (type) z_T;
4799 : : S6 w_t = b_t + z_t;
4800 : : S5 x_t = w_t & (N - 1);
4801 : : S4' a_t = x_t - z_t;
4802 : :
4803 : : Output:
4804 : :
4805 : : * TYPE_OUT: The type of the output of this pattern.
4806 : :
4807 : : * Return value: A new stmt that will be used to replace the division
4808 : : S1 or modulo S4 stmt. */
4809 : :
4810 : : static gimple *
4811 : 29780417 : vect_recog_divmod_pattern (vec_info *vinfo,
4812 : : stmt_vec_info stmt_vinfo, tree *type_out)
4813 : : {
4814 : 29780417 : gimple *last_stmt = stmt_vinfo->stmt;
4815 : 29780417 : tree oprnd0, oprnd1, vectype, itype, cond;
4816 : 29780417 : gimple *pattern_stmt, *def_stmt;
4817 : 29780417 : enum tree_code rhs_code;
4818 : 29780417 : optab optab;
4819 : 29780417 : tree q, cst;
4820 : 29780417 : int prec;
4821 : :
4822 : 29780417 : if (!is_gimple_assign (last_stmt))
4823 : : return NULL;
4824 : :
4825 : 20351113 : rhs_code = gimple_assign_rhs_code (last_stmt);
4826 : 20351113 : switch (rhs_code)
4827 : : {
4828 : 289519 : case TRUNC_DIV_EXPR:
4829 : 289519 : case EXACT_DIV_EXPR:
4830 : 289519 : case TRUNC_MOD_EXPR:
4831 : 289519 : break;
4832 : : default:
4833 : : return NULL;
4834 : : }
4835 : :
4836 : 289519 : oprnd0 = gimple_assign_rhs1 (last_stmt);
4837 : 289519 : oprnd1 = gimple_assign_rhs2 (last_stmt);
4838 : 289519 : itype = TREE_TYPE (oprnd0);
4839 : 289519 : if (TREE_CODE (oprnd0) != SSA_NAME
4840 : 272392 : || TREE_CODE (oprnd1) != INTEGER_CST
4841 : 183246 : || TREE_CODE (itype) != INTEGER_TYPE
4842 : 472765 : || !type_has_mode_precision_p (itype))
4843 : 106273 : return NULL;
4844 : :
4845 : 183246 : scalar_int_mode itype_mode = SCALAR_INT_TYPE_MODE (itype);
4846 : 183246 : vectype = get_vectype_for_scalar_type (vinfo, itype);
4847 : 183246 : if (vectype == NULL_TREE)
4848 : : return NULL;
4849 : :
4850 : 148284 : if (optimize_bb_for_size_p (gimple_bb (last_stmt)))
4851 : : {
4852 : : /* If the target can handle vectorized division or modulo natively,
4853 : : don't attempt to optimize this, since native division is likely
4854 : : to give smaller code. */
4855 : 1502 : optab = optab_for_tree_code (rhs_code, vectype, optab_default);
4856 : 1502 : if (optab != unknown_optab
4857 : 1502 : && can_implement_p (optab, TYPE_MODE (vectype)))
4858 : : return NULL;
4859 : : }
4860 : :
4861 : 148284 : prec = TYPE_PRECISION (itype);
4862 : 148284 : if (integer_pow2p (oprnd1))
4863 : : {
4864 : 72302 : if (TYPE_UNSIGNED (itype) || tree_int_cst_sgn (oprnd1) != 1)
4865 : 152 : return NULL;
4866 : :
4867 : : /* Pattern detected. */
4868 : 72150 : vect_pattern_detected ("vect_recog_divmod_pattern", last_stmt);
4869 : :
4870 : 72150 : *type_out = vectype;
4871 : :
4872 : : /* Check if the target supports this internal function. */
4873 : 72150 : internal_fn ifn = IFN_DIV_POW2;
4874 : 72150 : if (direct_internal_fn_supported_p (ifn, vectype, OPTIMIZE_FOR_SPEED))
4875 : : {
4876 : 0 : tree shift = build_int_cst (itype, tree_log2 (oprnd1));
4877 : :
4878 : 0 : tree var_div = vect_recog_temp_ssa_var (itype, NULL);
4879 : 0 : gimple *div_stmt = gimple_build_call_internal (ifn, 2, oprnd0, shift);
4880 : 0 : gimple_call_set_lhs (div_stmt, var_div);
4881 : :
4882 : 0 : if (rhs_code == TRUNC_MOD_EXPR)
4883 : : {
4884 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, div_stmt);
4885 : 0 : def_stmt
4886 : 0 : = gimple_build_assign (vect_recog_temp_ssa_var (itype, NULL),
4887 : : LSHIFT_EXPR, var_div, shift);
4888 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
4889 : 0 : pattern_stmt
4890 : 0 : = gimple_build_assign (vect_recog_temp_ssa_var (itype, NULL),
4891 : : MINUS_EXPR, oprnd0,
4892 : : gimple_assign_lhs (def_stmt));
4893 : : }
4894 : : else
4895 : : pattern_stmt = div_stmt;
4896 : 0 : gimple_set_location (pattern_stmt, gimple_location (last_stmt));
4897 : :
4898 : 0 : return pattern_stmt;
4899 : : }
4900 : :
4901 : 72150 : cond = vect_recog_temp_ssa_var (boolean_type_node, NULL);
4902 : 72150 : def_stmt = gimple_build_assign (cond, LT_EXPR, oprnd0,
4903 : : build_int_cst (itype, 0));
4904 : 72150 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt,
4905 : : truth_type_for (vectype), itype);
4906 : 72150 : if (rhs_code == TRUNC_DIV_EXPR
4907 : 72150 : || rhs_code == EXACT_DIV_EXPR)
4908 : : {
4909 : 69341 : tree var = vect_recog_temp_ssa_var (itype, NULL);
4910 : 69341 : tree shift;
4911 : 69341 : def_stmt
4912 : 69341 : = gimple_build_assign (var, COND_EXPR, cond,
4913 : : fold_build2 (MINUS_EXPR, itype, oprnd1,
4914 : : build_int_cst (itype, 1)),
4915 : : build_int_cst (itype, 0));
4916 : 69341 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
4917 : 69341 : var = vect_recog_temp_ssa_var (itype, NULL);
4918 : 69341 : def_stmt
4919 : 69341 : = gimple_build_assign (var, PLUS_EXPR, oprnd0,
4920 : : gimple_assign_lhs (def_stmt));
4921 : 69341 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
4922 : :
4923 : 69341 : shift = build_int_cst (itype, tree_log2 (oprnd1));
4924 : 69341 : pattern_stmt
4925 : 69341 : = gimple_build_assign (vect_recog_temp_ssa_var (itype, NULL),
4926 : : RSHIFT_EXPR, var, shift);
4927 : : }
4928 : : else
4929 : : {
4930 : 2809 : tree signmask;
4931 : 2809 : if (compare_tree_int (oprnd1, 2) == 0)
4932 : : {
4933 : 1558 : signmask = vect_recog_temp_ssa_var (itype, NULL);
4934 : 1558 : def_stmt = gimple_build_assign (signmask, COND_EXPR, cond,
4935 : : build_int_cst (itype, 1),
4936 : : build_int_cst (itype, 0));
4937 : 1558 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
4938 : : }
4939 : : else
4940 : : {
4941 : 1251 : tree utype
4942 : 1251 : = build_nonstandard_integer_type (prec, 1);
4943 : 1251 : tree vecutype = get_vectype_for_scalar_type (vinfo, utype);
4944 : 1251 : tree shift
4945 : 1251 : = build_int_cst (utype, GET_MODE_BITSIZE (itype_mode)
4946 : 1251 : - tree_log2 (oprnd1));
4947 : 1251 : tree var = vect_recog_temp_ssa_var (utype, NULL);
4948 : :
4949 : 1251 : def_stmt = gimple_build_assign (var, COND_EXPR, cond,
4950 : : build_int_cst (utype, -1),
4951 : : build_int_cst (utype, 0));
4952 : 1251 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, vecutype);
4953 : 1251 : var = vect_recog_temp_ssa_var (utype, NULL);
4954 : 1251 : def_stmt = gimple_build_assign (var, RSHIFT_EXPR,
4955 : : gimple_assign_lhs (def_stmt),
4956 : : shift);
4957 : 1251 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, vecutype);
4958 : 1251 : signmask = vect_recog_temp_ssa_var (itype, NULL);
4959 : 1251 : def_stmt
4960 : 1251 : = gimple_build_assign (signmask, NOP_EXPR, var);
4961 : 1251 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
4962 : : }
4963 : 2809 : def_stmt
4964 : 2809 : = gimple_build_assign (vect_recog_temp_ssa_var (itype, NULL),
4965 : : PLUS_EXPR, oprnd0, signmask);
4966 : 2809 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
4967 : 2809 : def_stmt
4968 : 2809 : = gimple_build_assign (vect_recog_temp_ssa_var (itype, NULL),
4969 : : BIT_AND_EXPR, gimple_assign_lhs (def_stmt),
4970 : : fold_build2 (MINUS_EXPR, itype, oprnd1,
4971 : : build_int_cst (itype, 1)));
4972 : 2809 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
4973 : :
4974 : 2809 : pattern_stmt
4975 : 2809 : = gimple_build_assign (vect_recog_temp_ssa_var (itype, NULL),
4976 : : MINUS_EXPR, gimple_assign_lhs (def_stmt),
4977 : : signmask);
4978 : : }
4979 : :
4980 : 72150 : return pattern_stmt;
4981 : : }
4982 : :
4983 : 75982 : if ((cst = uniform_integer_cst_p (oprnd1))
4984 : 75982 : && TYPE_UNSIGNED (itype)
4985 : : && rhs_code == TRUNC_DIV_EXPR
4986 : 47057 : && vectype
4987 : 106162 : && targetm.vectorize.preferred_div_as_shifts_over_mult (vectype))
4988 : : {
4989 : : /* We can use the relationship:
4990 : :
4991 : : x // N == ((x+N+2) // (N+1) + x) // (N+1) for 0 <= x < N(N+3)
4992 : :
4993 : : to optimize cases where N+1 is a power of 2, and where // (N+1)
4994 : : is therefore a shift right. When operating in modes that are
4995 : : multiples of a byte in size, there are two cases:
4996 : :
4997 : : (1) N(N+3) is not representable, in which case the question
4998 : : becomes whether the replacement expression overflows.
4999 : : It is enough to test that x+N+2 does not overflow,
5000 : : i.e. that x < MAX-(N+1).
5001 : :
5002 : : (2) N(N+3) is representable, in which case it is the (only)
5003 : : bound that we need to check.
5004 : :
5005 : : ??? For now we just handle the case where // (N+1) is a shift
5006 : : right by half the precision, since some architectures can
5007 : : optimize the associated addition and shift combinations
5008 : : into single instructions. */
5009 : :
5010 : 19083 : auto wcst = wi::to_wide (cst);
5011 : 19083 : int pow = wi::exact_log2 (wcst + 1);
5012 : 19083 : if (pow == prec / 2)
5013 : : {
5014 : 572 : gimple *stmt = SSA_NAME_DEF_STMT (oprnd0);
5015 : :
5016 : 572 : gimple_ranger ranger;
5017 : 572 : int_range_max r;
5018 : :
5019 : : /* Check that no overflow will occur. If we don't have range
5020 : : information we can't perform the optimization. */
5021 : :
5022 : 572 : if (ranger.range_of_expr (r, oprnd0, stmt) && !r.undefined_p ())
5023 : : {
5024 : 570 : wide_int max = r.upper_bound ();
5025 : 570 : wide_int one = wi::shwi (1, prec);
5026 : 570 : wide_int adder = wi::add (one, wi::lshift (one, pow));
5027 : 570 : wi::overflow_type ovf;
5028 : 570 : wi::add (max, adder, UNSIGNED, &ovf);
5029 : 570 : if (ovf == wi::OVF_NONE)
5030 : : {
5031 : 333 : *type_out = vectype;
5032 : 333 : tree tadder = wide_int_to_tree (itype, adder);
5033 : 333 : tree rshift = wide_int_to_tree (itype, pow);
5034 : :
5035 : 333 : tree new_lhs1 = vect_recog_temp_ssa_var (itype, NULL);
5036 : 333 : gassign *patt1
5037 : 333 : = gimple_build_assign (new_lhs1, PLUS_EXPR, oprnd0, tadder);
5038 : 333 : append_pattern_def_seq (vinfo, stmt_vinfo, patt1, vectype);
5039 : :
5040 : 333 : tree new_lhs2 = vect_recog_temp_ssa_var (itype, NULL);
5041 : 333 : patt1 = gimple_build_assign (new_lhs2, RSHIFT_EXPR, new_lhs1,
5042 : : rshift);
5043 : 333 : append_pattern_def_seq (vinfo, stmt_vinfo, patt1, vectype);
5044 : :
5045 : 333 : tree new_lhs3 = vect_recog_temp_ssa_var (itype, NULL);
5046 : 333 : patt1 = gimple_build_assign (new_lhs3, PLUS_EXPR, new_lhs2,
5047 : : oprnd0);
5048 : 333 : append_pattern_def_seq (vinfo, stmt_vinfo, patt1, vectype);
5049 : :
5050 : 333 : tree new_lhs4 = vect_recog_temp_ssa_var (itype, NULL);
5051 : 333 : pattern_stmt = gimple_build_assign (new_lhs4, RSHIFT_EXPR,
5052 : : new_lhs3, rshift);
5053 : :
5054 : 333 : return pattern_stmt;
5055 : : }
5056 : 570 : }
5057 : 572 : }
5058 : : }
5059 : :
5060 : 75649 : if (prec > HOST_BITS_PER_WIDE_INT
5061 : 75649 : || integer_zerop (oprnd1))
5062 : 572 : return NULL;
5063 : :
5064 : 75077 : if (!can_mult_highpart_p (TYPE_MODE (vectype), TYPE_UNSIGNED (itype)))
5065 : : return NULL;
5066 : :
5067 : 22385 : if (TYPE_UNSIGNED (itype))
5068 : : {
5069 : 15724 : unsigned HOST_WIDE_INT mh, ml;
5070 : 15724 : int pre_shift, post_shift;
5071 : 15724 : unsigned HOST_WIDE_INT d = (TREE_INT_CST_LOW (oprnd1)
5072 : 15724 : & GET_MODE_MASK (itype_mode));
5073 : 15724 : tree t1, t2, t3, t4;
5074 : :
5075 : 15724 : if (d >= (HOST_WIDE_INT_1U << (prec - 1)))
5076 : : /* FIXME: Can transform this into oprnd0 >= oprnd1 ? 1 : 0. */
5077 : 23 : return NULL;
5078 : :
5079 : : /* Find a suitable multiplier and right shift count instead of
5080 : : directly dividing by D. */
5081 : 15701 : mh = choose_multiplier (d, prec, prec, &ml, &post_shift);
5082 : :
5083 : : /* If the suggested multiplier is more than PREC bits, we can do better
5084 : : for even divisors, using an initial right shift. */
5085 : 15701 : if (mh != 0 && (d & 1) == 0)
5086 : : {
5087 : 448 : pre_shift = ctz_or_zero (d);
5088 : 448 : mh = choose_multiplier (d >> pre_shift, prec, prec - pre_shift,
5089 : : &ml, &post_shift);
5090 : 448 : gcc_assert (!mh);
5091 : : }
5092 : : else
5093 : : pre_shift = 0;
5094 : :
5095 : 932 : if (mh != 0)
5096 : : {
5097 : 932 : if (post_shift - 1 >= prec)
5098 : : return NULL;
5099 : :
5100 : : /* t1 = oprnd0 h* ml;
5101 : : t2 = oprnd0 - t1;
5102 : : t3 = t2 >> 1;
5103 : : t4 = t1 + t3;
5104 : : q = t4 >> (post_shift - 1); */
5105 : 932 : t1 = vect_recog_temp_ssa_var (itype, NULL);
5106 : 932 : def_stmt = gimple_build_assign (t1, MULT_HIGHPART_EXPR, oprnd0,
5107 : 932 : build_int_cst (itype, ml));
5108 : 932 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5109 : :
5110 : 932 : t2 = vect_recog_temp_ssa_var (itype, NULL);
5111 : 932 : def_stmt
5112 : 932 : = gimple_build_assign (t2, MINUS_EXPR, oprnd0, t1);
5113 : 932 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5114 : :
5115 : 932 : t3 = vect_recog_temp_ssa_var (itype, NULL);
5116 : 932 : def_stmt
5117 : 932 : = gimple_build_assign (t3, RSHIFT_EXPR, t2, integer_one_node);
5118 : 932 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5119 : :
5120 : 932 : t4 = vect_recog_temp_ssa_var (itype, NULL);
5121 : 932 : def_stmt
5122 : 932 : = gimple_build_assign (t4, PLUS_EXPR, t1, t3);
5123 : :
5124 : 932 : if (post_shift != 1)
5125 : : {
5126 : 932 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5127 : :
5128 : 932 : q = vect_recog_temp_ssa_var (itype, NULL);
5129 : 932 : pattern_stmt
5130 : 932 : = gimple_build_assign (q, RSHIFT_EXPR, t4,
5131 : 932 : build_int_cst (itype, post_shift - 1));
5132 : : }
5133 : : else
5134 : : {
5135 : : q = t4;
5136 : : pattern_stmt = def_stmt;
5137 : : }
5138 : : }
5139 : : else
5140 : : {
5141 : 14769 : if (pre_shift >= prec || post_shift >= prec)
5142 : : return NULL;
5143 : :
5144 : : /* t1 = oprnd0 >> pre_shift;
5145 : : t2 = t1 h* ml;
5146 : : q = t2 >> post_shift; */
5147 : 14769 : if (pre_shift)
5148 : : {
5149 : 448 : t1 = vect_recog_temp_ssa_var (itype, NULL);
5150 : 448 : def_stmt
5151 : 448 : = gimple_build_assign (t1, RSHIFT_EXPR, oprnd0,
5152 : 448 : build_int_cst (NULL, pre_shift));
5153 : 448 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5154 : : }
5155 : : else
5156 : : t1 = oprnd0;
5157 : :
5158 : 14769 : t2 = vect_recog_temp_ssa_var (itype, NULL);
5159 : 14769 : def_stmt = gimple_build_assign (t2, MULT_HIGHPART_EXPR, t1,
5160 : 14769 : build_int_cst (itype, ml));
5161 : :
5162 : 14769 : if (post_shift)
5163 : : {
5164 : 14763 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5165 : :
5166 : 14763 : q = vect_recog_temp_ssa_var (itype, NULL);
5167 : 14763 : def_stmt
5168 : 14763 : = gimple_build_assign (q, RSHIFT_EXPR, t2,
5169 : 14763 : build_int_cst (itype, post_shift));
5170 : : }
5171 : : else
5172 : : q = t2;
5173 : :
5174 : : pattern_stmt = def_stmt;
5175 : : }
5176 : : }
5177 : : else
5178 : : {
5179 : 6661 : unsigned HOST_WIDE_INT ml;
5180 : 6661 : int post_shift;
5181 : 6661 : HOST_WIDE_INT d = TREE_INT_CST_LOW (oprnd1);
5182 : 6661 : unsigned HOST_WIDE_INT abs_d;
5183 : 6661 : bool add = false;
5184 : 6661 : tree t1, t2, t3, t4;
5185 : :
5186 : : /* Give up for -1. */
5187 : 6661 : if (d == -1)
5188 : 0 : return NULL;
5189 : :
5190 : : /* Since d might be INT_MIN, we have to cast to
5191 : : unsigned HOST_WIDE_INT before negating to avoid
5192 : : undefined signed overflow. */
5193 : 6661 : abs_d = (d >= 0
5194 : 6661 : ? (unsigned HOST_WIDE_INT) d
5195 : : : - (unsigned HOST_WIDE_INT) d);
5196 : :
5197 : : /* n rem d = n rem -d */
5198 : 6661 : if (rhs_code == TRUNC_MOD_EXPR && d < 0)
5199 : : {
5200 : 0 : d = abs_d;
5201 : 0 : oprnd1 = build_int_cst (itype, abs_d);
5202 : : }
5203 : 6661 : if (HOST_BITS_PER_WIDE_INT >= prec
5204 : 6661 : && abs_d == HOST_WIDE_INT_1U << (prec - 1))
5205 : : /* This case is not handled correctly below. */
5206 : : return NULL;
5207 : :
5208 : 6661 : choose_multiplier (abs_d, prec, prec - 1, &ml, &post_shift);
5209 : 6661 : if (ml >= HOST_WIDE_INT_1U << (prec - 1))
5210 : : {
5211 : 1472 : add = true;
5212 : 1472 : ml |= HOST_WIDE_INT_M1U << (prec - 1);
5213 : : }
5214 : 6661 : if (post_shift >= prec)
5215 : : return NULL;
5216 : :
5217 : : /* t1 = oprnd0 h* ml; */
5218 : 6661 : t1 = vect_recog_temp_ssa_var (itype, NULL);
5219 : 6661 : def_stmt = gimple_build_assign (t1, MULT_HIGHPART_EXPR, oprnd0,
5220 : 6661 : build_int_cst (itype, ml));
5221 : :
5222 : 6661 : if (add)
5223 : : {
5224 : : /* t2 = t1 + oprnd0; */
5225 : 1472 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5226 : 1472 : t2 = vect_recog_temp_ssa_var (itype, NULL);
5227 : 1472 : def_stmt = gimple_build_assign (t2, PLUS_EXPR, t1, oprnd0);
5228 : : }
5229 : : else
5230 : : t2 = t1;
5231 : :
5232 : 6661 : if (post_shift)
5233 : : {
5234 : : /* t3 = t2 >> post_shift; */
5235 : 6038 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5236 : 6038 : t3 = vect_recog_temp_ssa_var (itype, NULL);
5237 : 6038 : def_stmt = gimple_build_assign (t3, RSHIFT_EXPR, t2,
5238 : 6038 : build_int_cst (itype, post_shift));
5239 : : }
5240 : : else
5241 : : t3 = t2;
5242 : :
5243 : 6661 : int msb = 1;
5244 : 6661 : int_range_max r;
5245 : 13322 : get_range_query (cfun)->range_of_expr (r, oprnd0);
5246 : 6661 : if (!r.varying_p () && !r.undefined_p ())
5247 : : {
5248 : 3922 : if (!wi::neg_p (r.lower_bound (), TYPE_SIGN (itype)))
5249 : : msb = 0;
5250 : 695 : else if (wi::neg_p (r.upper_bound (), TYPE_SIGN (itype)))
5251 : : msb = -1;
5252 : : }
5253 : :
5254 : 3227 : if (msb == 0 && d >= 0)
5255 : : {
5256 : : /* q = t3; */
5257 : : q = t3;
5258 : : pattern_stmt = def_stmt;
5259 : : }
5260 : : else
5261 : : {
5262 : : /* t4 = oprnd0 >> (prec - 1);
5263 : : or if we know from VRP that oprnd0 >= 0
5264 : : t4 = 0;
5265 : : or if we know from VRP that oprnd0 < 0
5266 : : t4 = -1; */
5267 : 3494 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5268 : 3494 : t4 = vect_recog_temp_ssa_var (itype, NULL);
5269 : 3494 : if (msb != 1)
5270 : 68 : def_stmt = gimple_build_assign (t4, INTEGER_CST,
5271 : 68 : build_int_cst (itype, msb));
5272 : : else
5273 : 3426 : def_stmt = gimple_build_assign (t4, RSHIFT_EXPR, oprnd0,
5274 : 3426 : build_int_cst (itype, prec - 1));
5275 : 3494 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5276 : :
5277 : : /* q = t3 - t4; or q = t4 - t3; */
5278 : 3494 : q = vect_recog_temp_ssa_var (itype, NULL);
5279 : 6823 : pattern_stmt = gimple_build_assign (q, MINUS_EXPR, d < 0 ? t4 : t3,
5280 : : d < 0 ? t3 : t4);
5281 : : }
5282 : 6661 : }
5283 : :
5284 : 22362 : if (rhs_code == TRUNC_MOD_EXPR)
5285 : : {
5286 : 9282 : tree r, t1;
5287 : :
5288 : : /* We divided. Now finish by:
5289 : : t1 = q * oprnd1;
5290 : : r = oprnd0 - t1; */
5291 : 9282 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt);
5292 : :
5293 : 9282 : t1 = vect_recog_temp_ssa_var (itype, NULL);
5294 : 9282 : def_stmt = gimple_build_assign (t1, MULT_EXPR, q, oprnd1);
5295 : 9282 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5296 : :
5297 : 9282 : r = vect_recog_temp_ssa_var (itype, NULL);
5298 : 9282 : pattern_stmt = gimple_build_assign (r, MINUS_EXPR, oprnd0, t1);
5299 : : }
5300 : :
5301 : : /* Pattern detected. */
5302 : 22362 : vect_pattern_detected ("vect_recog_divmod_pattern", last_stmt);
5303 : :
5304 : 22362 : *type_out = vectype;
5305 : 22362 : return pattern_stmt;
5306 : : }
5307 : :
5308 : : /* Detects pattern with a modulo operation (S1) where both arguments
5309 : : are variables of integral type.
5310 : : The statement is replaced by division, multiplication, and subtraction.
5311 : : The last statement (S4) is returned.
5312 : :
5313 : : Example:
5314 : : S1 c_t = a_t % b_t;
5315 : :
5316 : : is replaced by
5317 : : S2 x_t = a_t / b_t;
5318 : : S3 y_t = x_t * b_t;
5319 : : S4 z_t = a_t - y_t; */
5320 : :
5321 : : static gimple *
5322 : 29964791 : vect_recog_mod_var_pattern (vec_info *vinfo,
5323 : : stmt_vec_info stmt_vinfo, tree *type_out)
5324 : : {
5325 : 29964791 : gimple *last_stmt = STMT_VINFO_STMT (stmt_vinfo);
5326 : 29964791 : tree oprnd0, oprnd1, vectype, itype;
5327 : 29964791 : gimple *pattern_stmt, *def_stmt;
5328 : 29964791 : enum tree_code rhs_code;
5329 : :
5330 : 29964791 : if (!is_gimple_assign (last_stmt))
5331 : : return NULL;
5332 : :
5333 : 20535487 : rhs_code = gimple_assign_rhs_code (last_stmt);
5334 : 20535487 : if (rhs_code != TRUNC_MOD_EXPR)
5335 : : return NULL;
5336 : :
5337 : 71634 : oprnd0 = gimple_assign_rhs1 (last_stmt);
5338 : 71634 : oprnd1 = gimple_assign_rhs2 (last_stmt);
5339 : 71634 : itype = TREE_TYPE (oprnd0);
5340 : 71634 : if (TREE_CODE (oprnd0) != SSA_NAME
5341 : 63590 : || TREE_CODE (oprnd1) != SSA_NAME
5342 : 40812 : || TREE_CODE (itype) != INTEGER_TYPE)
5343 : : return NULL;
5344 : :
5345 : 40708 : vectype = get_vectype_for_scalar_type (vinfo, itype);
5346 : :
5347 : 40708 : if (!vectype
5348 : 32969 : || target_has_vecop_for_code (TRUNC_MOD_EXPR, vectype)
5349 : 32969 : || !target_has_vecop_for_code (TRUNC_DIV_EXPR, vectype)
5350 : 0 : || !target_has_vecop_for_code (MULT_EXPR, vectype)
5351 : 40708 : || !target_has_vecop_for_code (MINUS_EXPR, vectype))
5352 : 40708 : return NULL;
5353 : :
5354 : 0 : tree q, tmp, r;
5355 : 0 : q = vect_recog_temp_ssa_var (itype, NULL);
5356 : 0 : def_stmt = gimple_build_assign (q, TRUNC_DIV_EXPR, oprnd0, oprnd1);
5357 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, vectype);
5358 : :
5359 : 0 : tmp = vect_recog_temp_ssa_var (itype, NULL);
5360 : 0 : def_stmt = gimple_build_assign (tmp, MULT_EXPR, q, oprnd1);
5361 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, vectype);
5362 : :
5363 : 0 : r = vect_recog_temp_ssa_var (itype, NULL);
5364 : 0 : pattern_stmt = gimple_build_assign (r, MINUS_EXPR, oprnd0, tmp);
5365 : :
5366 : : /* Pattern detected. */
5367 : 0 : *type_out = vectype;
5368 : 0 : vect_pattern_detected ("vect_recog_mod_var_pattern", last_stmt);
5369 : :
5370 : 0 : return pattern_stmt;
5371 : : }
5372 : :
5373 : :
5374 : : /* Return the proper type for converting bool VAR into
5375 : : an integer value or NULL_TREE if no such type exists.
5376 : : The type is chosen so that the converted value has the
5377 : : same number of elements as VAR's vector type. */
5378 : :
5379 : : static tree
5380 : 3498862 : integer_type_for_mask (tree var, vec_info *vinfo)
5381 : : {
5382 : 3498862 : if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (var)))
5383 : : return NULL_TREE;
5384 : :
5385 : 1225902 : stmt_vec_info def_stmt_info = vect_get_internal_def (vinfo, var);
5386 : 1225902 : if (!def_stmt_info || !vect_use_mask_type_p (def_stmt_info))
5387 : : return NULL_TREE;
5388 : :
5389 : 676547 : return build_nonstandard_integer_type (def_stmt_info->mask_precision, 1);
5390 : : }
5391 : :
5392 : : /* Function vect_recog_gcond_pattern
5393 : :
5394 : : Try to find pattern like following:
5395 : :
5396 : : if (a op b)
5397 : :
5398 : : where operator 'op' is not != and convert it to an adjusted boolean pattern
5399 : :
5400 : : mask = a op b
5401 : : if (mask != 0)
5402 : :
5403 : : and set the mask type on MASK.
5404 : :
5405 : : Input:
5406 : :
5407 : : * STMT_VINFO: The stmt at the end from which the pattern
5408 : : search begins, i.e. cast of a bool to
5409 : : an integer type.
5410 : :
5411 : : Output:
5412 : :
5413 : : * TYPE_OUT: The type of the output of this pattern.
5414 : :
5415 : : * Return value: A new stmt that will be used to replace the pattern. */
5416 : :
5417 : : static gimple *
5418 : 30045752 : vect_recog_gcond_pattern (vec_info *vinfo,
5419 : : stmt_vec_info stmt_vinfo, tree *type_out)
5420 : : {
5421 : : /* Currently we only support this for loop vectorization and when multiple
5422 : : exits. */
5423 : 30045752 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
5424 : 3336372 : if (!loop_vinfo || !LOOP_VINFO_EARLY_BREAKS (loop_vinfo))
5425 : : return NULL;
5426 : :
5427 : 1081104 : gimple *last_stmt = STMT_VINFO_STMT (stmt_vinfo);
5428 : 1081104 : gcond* cond = NULL;
5429 : 30065373 : if (!(cond = dyn_cast <gcond *> (last_stmt)))
5430 : : return NULL;
5431 : :
5432 : 283317 : auto lhs = gimple_cond_lhs (cond);
5433 : 283317 : auto rhs = gimple_cond_rhs (cond);
5434 : 283317 : auto code = gimple_cond_code (cond);
5435 : :
5436 : 283317 : tree scalar_type = TREE_TYPE (lhs);
5437 : 283317 : if (VECTOR_TYPE_P (scalar_type))
5438 : : return NULL;
5439 : :
5440 : : /* If the input is a boolean then try to figure out the precision that the
5441 : : vector type should use. We cannot use the scalar precision as this would
5442 : : later mismatch. This is similar to what recog_bool does. */
5443 : 283317 : if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type))
5444 : : {
5445 : 8546 : if (tree stype = integer_type_for_mask (lhs, vinfo))
5446 : 283317 : scalar_type = stype;
5447 : : }
5448 : :
5449 : 283317 : tree vectype = get_mask_type_for_scalar_type (vinfo, scalar_type);
5450 : 283317 : if (vectype == NULL_TREE)
5451 : : return NULL;
5452 : :
5453 : 263696 : tree new_lhs = vect_recog_temp_ssa_var (boolean_type_node, NULL);
5454 : 263696 : gimple *new_stmt = gimple_build_assign (new_lhs, code, lhs, rhs);
5455 : 263696 : append_pattern_def_seq (vinfo, stmt_vinfo, new_stmt, vectype, scalar_type);
5456 : :
5457 : 263696 : gimple *pattern_stmt
5458 : 263696 : = gimple_build_cond (NE_EXPR, new_lhs,
5459 : 263696 : build_int_cst (TREE_TYPE (new_lhs), 0),
5460 : : NULL_TREE, NULL_TREE);
5461 : 263696 : *type_out = vectype;
5462 : 263696 : vect_pattern_detected ("vect_recog_gcond_pattern", last_stmt);
5463 : 263696 : return pattern_stmt;
5464 : : }
5465 : :
5466 : : /* Function vect_recog_bool_pattern
5467 : :
5468 : : Try to find pattern like following:
5469 : :
5470 : : bool a_b, b_b, c_b, d_b, e_b;
5471 : : TYPE f_T;
5472 : : loop:
5473 : : S1 a_b = x1 CMP1 y1;
5474 : : S2 b_b = x2 CMP2 y2;
5475 : : S3 c_b = a_b & b_b;
5476 : : S4 d_b = x3 CMP3 y3;
5477 : : S5 e_b = c_b | d_b;
5478 : : S6 f_T = (TYPE) e_b;
5479 : :
5480 : : where type 'TYPE' is an integral type. Or a similar pattern
5481 : : ending in
5482 : :
5483 : : S6 f_Y = e_b ? r_Y : s_Y;
5484 : :
5485 : : as results from if-conversion of a complex condition.
5486 : :
5487 : : Input:
5488 : :
5489 : : * STMT_VINFO: The stmt at the end from which the pattern
5490 : : search begins, i.e. cast of a bool to
5491 : : an integer type.
5492 : :
5493 : : Output:
5494 : :
5495 : : * TYPE_OUT: The type of the output of this pattern.
5496 : :
5497 : : * Return value: A new stmt that will be used to replace the pattern.
5498 : :
5499 : : Assuming size of TYPE is the same as size of all comparisons
5500 : : (otherwise some casts would be added where needed), the above
5501 : : sequence we create related pattern stmts:
5502 : : S1' a_T = x1 CMP1 y1 ? 1 : 0;
5503 : : S3' c_T = x2 CMP2 y2 ? a_T : 0;
5504 : : S4' d_T = x3 CMP3 y3 ? 1 : 0;
5505 : : S5' e_T = c_T | d_T;
5506 : : S6' f_T = e_T;
5507 : :
5508 : : Instead of the above S3' we could emit:
5509 : : S2' b_T = x2 CMP2 y2 ? 1 : 0;
5510 : : S3' c_T = a_T | b_T;
5511 : : but the above is more efficient. */
5512 : :
5513 : : static gimple *
5514 : 30045752 : vect_recog_bool_pattern (vec_info *vinfo,
5515 : : stmt_vec_info stmt_vinfo, tree *type_out)
5516 : : {
5517 : 30045752 : gimple *last_stmt = stmt_vinfo->stmt;
5518 : 30045752 : enum tree_code rhs_code;
5519 : 30045752 : tree var, lhs, rhs, vectype;
5520 : 30045752 : gimple *pattern_stmt;
5521 : :
5522 : 30045752 : if (!is_gimple_assign (last_stmt))
5523 : : return NULL;
5524 : :
5525 : 20880144 : var = gimple_assign_rhs1 (last_stmt);
5526 : 20880144 : lhs = gimple_assign_lhs (last_stmt);
5527 : 20880144 : rhs_code = gimple_assign_rhs_code (last_stmt);
5528 : :
5529 : 20880144 : if (rhs_code == VIEW_CONVERT_EXPR)
5530 : 170285 : var = TREE_OPERAND (var, 0);
5531 : :
5532 : 20880144 : if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (var)))
5533 : : return NULL;
5534 : :
5535 : 729447 : hash_set<gimple *> bool_stmts;
5536 : :
5537 : 729447 : if (CONVERT_EXPR_CODE_P (rhs_code)
5538 : 636747 : || rhs_code == VIEW_CONVERT_EXPR)
5539 : : {
5540 : 192762 : if (! INTEGRAL_TYPE_P (TREE_TYPE (lhs))
5541 : 192631 : || VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (lhs)))
5542 : : return NULL;
5543 : 92759 : vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (lhs));
5544 : :
5545 : 92759 : tree type = integer_type_for_mask (var, vinfo);
5546 : 92759 : tree cst0, cst1, tmp;
5547 : :
5548 : 92759 : if (!type)
5549 : : return NULL;
5550 : :
5551 : : /* We may directly use cond with narrowed type to avoid multiple cond
5552 : : exprs with following result packing and perform single cond with
5553 : : packed mask instead. In case of widening we better make cond first
5554 : : and then extract results. */
5555 : 45848 : if (TYPE_MODE (type) == TYPE_MODE (TREE_TYPE (lhs)))
5556 : 31246 : type = TREE_TYPE (lhs);
5557 : :
5558 : 45848 : cst0 = build_int_cst (type, 0);
5559 : 45848 : cst1 = build_int_cst (type, 1);
5560 : 45848 : tmp = vect_recog_temp_ssa_var (type, NULL);
5561 : 45848 : pattern_stmt = gimple_build_assign (tmp, COND_EXPR, var, cst1, cst0);
5562 : :
5563 : 45848 : if (!useless_type_conversion_p (type, TREE_TYPE (lhs)))
5564 : : {
5565 : 14602 : tree new_vectype = get_vectype_for_scalar_type (vinfo, type);
5566 : 14602 : append_pattern_def_seq (vinfo, stmt_vinfo,
5567 : : pattern_stmt, new_vectype);
5568 : :
5569 : 14602 : lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL);
5570 : 14602 : pattern_stmt = gimple_build_assign (lhs, CONVERT_EXPR, tmp);
5571 : : }
5572 : :
5573 : 45848 : *type_out = vectype;
5574 : 45848 : vect_pattern_detected ("vect_recog_bool_pattern", last_stmt);
5575 : :
5576 : 45848 : return pattern_stmt;
5577 : : }
5578 : 632717 : else if (rhs_code == COND_EXPR
5579 : 173729 : && TREE_CODE (var) == SSA_NAME)
5580 : : {
5581 : 173729 : vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (lhs));
5582 : 173729 : if (vectype == NULL_TREE)
5583 : : return NULL;
5584 : :
5585 : : /* Build a scalar type for the boolean result that when
5586 : : vectorized matches the vector type of the result in
5587 : : size and number of elements. */
5588 : 160336 : unsigned prec
5589 : 160336 : = vector_element_size (tree_to_poly_uint64 (TYPE_SIZE (vectype)),
5590 : : TYPE_VECTOR_SUBPARTS (vectype));
5591 : :
5592 : 160336 : tree type
5593 : 320672 : = build_nonstandard_integer_type (prec,
5594 : 160336 : TYPE_UNSIGNED (TREE_TYPE (var)));
5595 : 160336 : if (get_vectype_for_scalar_type (vinfo, type) == NULL_TREE)
5596 : : return NULL;
5597 : :
5598 : 160336 : enum vect_def_type dt;
5599 : 160336 : if (integer_type_for_mask (var, vinfo))
5600 : : return NULL;
5601 : 27244 : else if (TREE_CODE (TREE_TYPE (var)) == BOOLEAN_TYPE
5602 : 27244 : && vect_is_simple_use (var, vinfo, &dt)
5603 : 27244 : && (dt == vect_external_def
5604 : 27237 : || dt == vect_constant_def))
5605 : : {
5606 : : /* If the condition is already a boolean then manually convert it to a
5607 : : mask of the given integer type but don't set a vectype. */
5608 : 1229 : tree lhs_ivar = vect_recog_temp_ssa_var (type, NULL);
5609 : 1229 : pattern_stmt = gimple_build_assign (lhs_ivar, COND_EXPR, var,
5610 : : build_all_ones_cst (type),
5611 : : build_zero_cst (type));
5612 : 1229 : append_inv_pattern_def_seq (vinfo, pattern_stmt);
5613 : 1229 : var = lhs_ivar;
5614 : : }
5615 : :
5616 : 27244 : tree lhs_var = vect_recog_temp_ssa_var (boolean_type_node, NULL);
5617 : 27244 : pattern_stmt = gimple_build_assign (lhs_var, NE_EXPR, var,
5618 : 27244 : build_zero_cst (TREE_TYPE (var)));
5619 : :
5620 : 27244 : tree new_vectype = get_mask_type_for_scalar_type (vinfo, TREE_TYPE (var));
5621 : 27244 : if (!new_vectype)
5622 : : return NULL;
5623 : :
5624 : 27244 : new_vectype = truth_type_for (new_vectype);
5625 : 27244 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, new_vectype,
5626 : 27244 : TREE_TYPE (var));
5627 : :
5628 : 27244 : lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL);
5629 : 27244 : pattern_stmt
5630 : 27244 : = gimple_build_assign (lhs, COND_EXPR, lhs_var,
5631 : : gimple_assign_rhs2 (last_stmt),
5632 : : gimple_assign_rhs3 (last_stmt));
5633 : 27244 : *type_out = vectype;
5634 : 27244 : vect_pattern_detected ("vect_recog_bool_pattern", last_stmt);
5635 : :
5636 : 27244 : return pattern_stmt;
5637 : : }
5638 : 458988 : else if (rhs_code == SSA_NAME
5639 : 29235 : && STMT_VINFO_DATA_REF (stmt_vinfo))
5640 : : {
5641 : 7914 : stmt_vec_info pattern_stmt_info;
5642 : 7914 : vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (lhs));
5643 : 7914 : if (!vectype || !VECTOR_MODE_P (TYPE_MODE (vectype)))
5644 : 0 : return NULL;
5645 : :
5646 : 7914 : tree type = integer_type_for_mask (var, vinfo);
5647 : 7914 : tree cst0, cst1, new_vectype;
5648 : :
5649 : 7914 : if (!type)
5650 : : return NULL;
5651 : :
5652 : 4448 : if (TYPE_MODE (type) == TYPE_MODE (TREE_TYPE (vectype)))
5653 : 541 : type = TREE_TYPE (vectype);
5654 : :
5655 : 4448 : cst0 = build_int_cst (type, 0);
5656 : 4448 : cst1 = build_int_cst (type, 1);
5657 : 4448 : new_vectype = get_vectype_for_scalar_type (vinfo, type);
5658 : :
5659 : 4448 : rhs = vect_recog_temp_ssa_var (type, NULL);
5660 : 4448 : pattern_stmt = gimple_build_assign (rhs, COND_EXPR, var, cst1, cst0);
5661 : 4448 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, new_vectype);
5662 : :
5663 : 4448 : lhs = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (vectype), lhs);
5664 : 4448 : if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
5665 : : {
5666 : 3907 : tree rhs2 = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL);
5667 : 3907 : gimple *cast_stmt = gimple_build_assign (rhs2, NOP_EXPR, rhs);
5668 : 3907 : append_pattern_def_seq (vinfo, stmt_vinfo, cast_stmt);
5669 : 3907 : rhs = rhs2;
5670 : : }
5671 : 4448 : pattern_stmt = gimple_build_assign (lhs, SSA_NAME, rhs);
5672 : 4448 : pattern_stmt_info = vinfo->add_stmt (pattern_stmt);
5673 : 4448 : vinfo->move_dr (pattern_stmt_info, stmt_vinfo);
5674 : 4448 : *type_out = vectype;
5675 : 4448 : vect_pattern_detected ("vect_recog_bool_pattern", last_stmt);
5676 : :
5677 : 4448 : return pattern_stmt;
5678 : : }
5679 : : else
5680 : : return NULL;
5681 : 729447 : }
5682 : :
5683 : : /* A helper for vect_recog_mask_conversion_pattern. Build
5684 : : conversion of MASK to a type suitable for masking VECTYPE.
5685 : : Built statement gets required vectype and is appended to
5686 : : a pattern sequence of STMT_VINFO.
5687 : :
5688 : : Return converted mask. */
5689 : :
5690 : : static tree
5691 : 53432 : build_mask_conversion (vec_info *vinfo,
5692 : : tree mask, tree vectype, stmt_vec_info stmt_vinfo)
5693 : : {
5694 : 53432 : gimple *stmt;
5695 : 53432 : tree masktype, tmp;
5696 : :
5697 : 53432 : masktype = truth_type_for (vectype);
5698 : 53432 : tmp = vect_recog_temp_ssa_var (TREE_TYPE (masktype), NULL);
5699 : 53432 : stmt = gimple_build_assign (tmp, CONVERT_EXPR, mask);
5700 : 53432 : append_pattern_def_seq (vinfo, stmt_vinfo,
5701 : 53432 : stmt, masktype, TREE_TYPE (vectype));
5702 : :
5703 : 53432 : return tmp;
5704 : : }
5705 : :
5706 : :
5707 : : /* Function vect_recog_mask_conversion_pattern
5708 : :
5709 : : Try to find statements which require boolean type
5710 : : converison. Additional conversion statements are
5711 : : added to handle such cases. For example:
5712 : :
5713 : : bool m_1, m_2, m_3;
5714 : : int i_4, i_5;
5715 : : double d_6, d_7;
5716 : : char c_1, c_2, c_3;
5717 : :
5718 : : S1 m_1 = i_4 > i_5;
5719 : : S2 m_2 = d_6 < d_7;
5720 : : S3 m_3 = m_1 & m_2;
5721 : : S4 c_1 = m_3 ? c_2 : c_3;
5722 : :
5723 : : Will be transformed into:
5724 : :
5725 : : S1 m_1 = i_4 > i_5;
5726 : : S2 m_2 = d_6 < d_7;
5727 : : S3'' m_2' = (_Bool[bitsize=32])m_2
5728 : : S3' m_3' = m_1 & m_2';
5729 : : S4'' m_3'' = (_Bool[bitsize=8])m_3'
5730 : : S4' c_1' = m_3'' ? c_2 : c_3; */
5731 : :
5732 : : static gimple *
5733 : 30018492 : vect_recog_mask_conversion_pattern (vec_info *vinfo,
5734 : : stmt_vec_info stmt_vinfo, tree *type_out)
5735 : : {
5736 : 30018492 : gimple *last_stmt = stmt_vinfo->stmt;
5737 : 30018492 : enum tree_code rhs_code;
5738 : 30018492 : tree lhs = NULL_TREE, rhs1, rhs2, tmp, rhs1_type, rhs2_type;
5739 : 30018492 : tree vectype1, vectype2;
5740 : 30018492 : stmt_vec_info pattern_stmt_info;
5741 : :
5742 : : /* Check for MASK_LOAD and MASK_STORE as well as COND_OP calls requiring mask
5743 : : conversion. */
5744 : 30018492 : if (is_gimple_call (last_stmt)
5745 : 30018492 : && gimple_call_internal_p (last_stmt))
5746 : : {
5747 : 90438 : gcall *pattern_stmt;
5748 : :
5749 : 90438 : internal_fn ifn = gimple_call_internal_fn (last_stmt);
5750 : 90438 : int mask_argno = internal_fn_mask_index (ifn);
5751 : 90438 : if (mask_argno < 0)
5752 : : return NULL;
5753 : :
5754 : 8721 : bool store_p = internal_store_fn_p (ifn);
5755 : 8721 : bool load_p = internal_store_fn_p (ifn);
5756 : 8721 : if (store_p)
5757 : : {
5758 : 1577 : int rhs_index = internal_fn_stored_value_index (ifn);
5759 : 1577 : tree rhs = gimple_call_arg (last_stmt, rhs_index);
5760 : 1577 : vectype1 = get_vectype_for_scalar_type (vinfo, TREE_TYPE (rhs));
5761 : : }
5762 : : else
5763 : : {
5764 : 7144 : lhs = gimple_call_lhs (last_stmt);
5765 : 7144 : if (!lhs)
5766 : : return NULL;
5767 : 7144 : vectype1 = get_vectype_for_scalar_type (vinfo, TREE_TYPE (lhs));
5768 : : }
5769 : :
5770 : 8721 : if (!vectype1)
5771 : : return NULL;
5772 : :
5773 : 8577 : tree mask_arg = gimple_call_arg (last_stmt, mask_argno);
5774 : 8577 : tree mask_arg_type = integer_type_for_mask (mask_arg, vinfo);
5775 : 8577 : if (mask_arg_type)
5776 : : {
5777 : 7623 : vectype2 = get_mask_type_for_scalar_type (vinfo, mask_arg_type);
5778 : :
5779 : 7623 : if (!vectype2
5780 : 7623 : || known_eq (TYPE_VECTOR_SUBPARTS (vectype1),
5781 : : TYPE_VECTOR_SUBPARTS (vectype2)))
5782 : 4489 : return NULL;
5783 : : }
5784 : 954 : else if (store_p || load_p)
5785 : : return NULL;
5786 : :
5787 : 3889 : tmp = build_mask_conversion (vinfo, mask_arg, vectype1, stmt_vinfo);
5788 : :
5789 : 3889 : auto_vec<tree, 8> args;
5790 : 3889 : unsigned int nargs = gimple_call_num_args (last_stmt);
5791 : 3889 : args.safe_grow (nargs, true);
5792 : 19445 : for (unsigned int i = 0; i < nargs; ++i)
5793 : 15556 : args[i] = ((int) i == mask_argno
5794 : 15556 : ? tmp
5795 : 11667 : : gimple_call_arg (last_stmt, i));
5796 : 3889 : pattern_stmt = gimple_build_call_internal_vec (ifn, args);
5797 : :
5798 : 3889 : if (!store_p)
5799 : : {
5800 : 3635 : lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL);
5801 : 3635 : gimple_call_set_lhs (pattern_stmt, lhs);
5802 : : }
5803 : :
5804 : 3635 : if (load_p || store_p)
5805 : 254 : gimple_call_set_nothrow (pattern_stmt, true);
5806 : :
5807 : 3889 : pattern_stmt_info = vinfo->add_stmt (pattern_stmt);
5808 : 3889 : if (STMT_VINFO_DATA_REF (stmt_vinfo))
5809 : 1533 : vinfo->move_dr (pattern_stmt_info, stmt_vinfo);
5810 : :
5811 : 3889 : *type_out = vectype1;
5812 : 3889 : vect_pattern_detected ("vect_recog_mask_conversion_pattern", last_stmt);
5813 : :
5814 : 3889 : return pattern_stmt;
5815 : 3889 : }
5816 : :
5817 : 29928054 : if (!is_gimple_assign (last_stmt))
5818 : : return NULL;
5819 : :
5820 : 20852884 : gimple *pattern_stmt;
5821 : 20852884 : lhs = gimple_assign_lhs (last_stmt);
5822 : 20852884 : rhs1 = gimple_assign_rhs1 (last_stmt);
5823 : 20852884 : rhs_code = gimple_assign_rhs_code (last_stmt);
5824 : :
5825 : : /* Check for cond expression requiring mask conversion. */
5826 : 20852884 : if (rhs_code == COND_EXPR)
5827 : : {
5828 : 165634 : vectype1 = get_vectype_for_scalar_type (vinfo, TREE_TYPE (lhs));
5829 : :
5830 : 165634 : gcc_assert (! COMPARISON_CLASS_P (rhs1));
5831 : 165634 : if (TREE_CODE (rhs1) == SSA_NAME)
5832 : : {
5833 : 165634 : rhs1_type = integer_type_for_mask (rhs1, vinfo);
5834 : 165634 : if (!rhs1_type)
5835 : : return NULL;
5836 : : }
5837 : : else
5838 : : return NULL;
5839 : :
5840 : 153960 : vectype2 = get_mask_type_for_scalar_type (vinfo, rhs1_type);
5841 : :
5842 : 153960 : if (!vectype1 || !vectype2)
5843 : : return NULL;
5844 : :
5845 : : /* Continue if a conversion is needed. Also continue if we have
5846 : : a comparison whose vector type would normally be different from
5847 : : VECTYPE2 when considered in isolation. In that case we'll
5848 : : replace the comparison with an SSA name (so that we can record
5849 : : its vector type) and behave as though the comparison was an SSA
5850 : : name from the outset. */
5851 : 152221 : if (known_eq (TYPE_VECTOR_SUBPARTS (vectype1),
5852 : : TYPE_VECTOR_SUBPARTS (vectype2)))
5853 : : return NULL;
5854 : :
5855 : 30456 : if (maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1),
5856 : 60912 : TYPE_VECTOR_SUBPARTS (vectype2)))
5857 : 30456 : tmp = build_mask_conversion (vinfo, rhs1, vectype1, stmt_vinfo);
5858 : : else
5859 : : tmp = rhs1;
5860 : :
5861 : 30456 : lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL);
5862 : 30456 : pattern_stmt = gimple_build_assign (lhs, COND_EXPR, tmp,
5863 : : gimple_assign_rhs2 (last_stmt),
5864 : : gimple_assign_rhs3 (last_stmt));
5865 : :
5866 : 30456 : *type_out = vectype1;
5867 : 30456 : vect_pattern_detected ("vect_recog_mask_conversion_pattern", last_stmt);
5868 : :
5869 : 30456 : return pattern_stmt;
5870 : : }
5871 : :
5872 : : /* Now check for binary boolean operations requiring conversion for
5873 : : one of operands. */
5874 : 20687250 : if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (lhs)))
5875 : : return NULL;
5876 : :
5877 : 1657643 : if (rhs_code != BIT_IOR_EXPR
5878 : : && rhs_code != BIT_XOR_EXPR
5879 : 1657643 : && rhs_code != BIT_AND_EXPR
5880 : 1303874 : && TREE_CODE_CLASS (rhs_code) != tcc_comparison)
5881 : : return NULL;
5882 : :
5883 : 1527548 : rhs2 = gimple_assign_rhs2 (last_stmt);
5884 : :
5885 : 1527548 : rhs1_type = integer_type_for_mask (rhs1, vinfo);
5886 : 1527548 : rhs2_type = integer_type_for_mask (rhs2, vinfo);
5887 : :
5888 : 1527548 : if (!rhs1_type || !rhs2_type
5889 : 1527548 : || TYPE_PRECISION (rhs1_type) == TYPE_PRECISION (rhs2_type))
5890 : : return NULL;
5891 : :
5892 : 19087 : if (TYPE_PRECISION (rhs1_type) < TYPE_PRECISION (rhs2_type))
5893 : : {
5894 : 11970 : vectype1 = get_mask_type_for_scalar_type (vinfo, rhs1_type);
5895 : 11970 : if (!vectype1)
5896 : : return NULL;
5897 : 11970 : rhs2 = build_mask_conversion (vinfo, rhs2, vectype1, stmt_vinfo);
5898 : : }
5899 : : else
5900 : : {
5901 : 7117 : vectype1 = get_mask_type_for_scalar_type (vinfo, rhs2_type);
5902 : 7117 : if (!vectype1)
5903 : : return NULL;
5904 : 7117 : rhs1 = build_mask_conversion (vinfo, rhs1, vectype1, stmt_vinfo);
5905 : : }
5906 : :
5907 : 19087 : lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL);
5908 : 19087 : pattern_stmt = gimple_build_assign (lhs, rhs_code, rhs1, rhs2);
5909 : :
5910 : 19087 : *type_out = vectype1;
5911 : 19087 : vect_pattern_detected ("vect_recog_mask_conversion_pattern", last_stmt);
5912 : :
5913 : 19087 : return pattern_stmt;
5914 : : }
5915 : :
5916 : : /* STMT_INFO is a load or store. If the load or store is conditional, return
5917 : : the boolean condition under which it occurs, otherwise return null. */
5918 : :
5919 : : static tree
5920 : 32844 : vect_get_load_store_mask (stmt_vec_info stmt_info)
5921 : : {
5922 : 32844 : if (gassign *def_assign = dyn_cast <gassign *> (stmt_info->stmt))
5923 : : {
5924 : 31402 : gcc_assert (gimple_assign_single_p (def_assign));
5925 : : return NULL_TREE;
5926 : : }
5927 : :
5928 : 1442 : if (gcall *def_call = dyn_cast <gcall *> (stmt_info->stmt))
5929 : : {
5930 : 1442 : internal_fn ifn = gimple_call_internal_fn (def_call);
5931 : 1442 : int mask_index = internal_fn_mask_index (ifn);
5932 : 1442 : return gimple_call_arg (def_call, mask_index);
5933 : : }
5934 : :
5935 : 0 : gcc_unreachable ();
5936 : : }
5937 : :
5938 : : /* Return MASK if MASK is suitable for masking an operation on vectors
5939 : : of type VECTYPE, otherwise convert it into such a form and return
5940 : : the result. Associate any conversion statements with STMT_INFO's
5941 : : pattern. */
5942 : :
5943 : : static tree
5944 : 0 : vect_convert_mask_for_vectype (tree mask, tree vectype,
5945 : : stmt_vec_info stmt_info, vec_info *vinfo)
5946 : : {
5947 : 0 : tree mask_type = integer_type_for_mask (mask, vinfo);
5948 : 0 : if (mask_type)
5949 : : {
5950 : 0 : tree mask_vectype = get_mask_type_for_scalar_type (vinfo, mask_type);
5951 : 0 : if (mask_vectype
5952 : 0 : && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype),
5953 : 0 : TYPE_VECTOR_SUBPARTS (mask_vectype)))
5954 : 0 : mask = build_mask_conversion (vinfo, mask, vectype, stmt_info);
5955 : : }
5956 : 0 : return mask;
5957 : : }
5958 : :
5959 : : /* Return the equivalent of:
5960 : :
5961 : : fold_convert (TYPE, VALUE)
5962 : :
5963 : : with the expectation that the operation will be vectorized.
5964 : : If new statements are needed, add them as pattern statements
5965 : : to STMT_INFO. */
5966 : :
5967 : : static tree
5968 : 0 : vect_add_conversion_to_pattern (vec_info *vinfo,
5969 : : tree type, tree value, stmt_vec_info stmt_info)
5970 : : {
5971 : 0 : if (useless_type_conversion_p (type, TREE_TYPE (value)))
5972 : : return value;
5973 : :
5974 : 0 : tree new_value = vect_recog_temp_ssa_var (type, NULL);
5975 : 0 : gassign *conversion = gimple_build_assign (new_value, CONVERT_EXPR, value);
5976 : 0 : append_pattern_def_seq (vinfo, stmt_info, conversion,
5977 : : get_vectype_for_scalar_type (vinfo, type));
5978 : 0 : return new_value;
5979 : : }
5980 : :
5981 : : /* Try to convert STMT_INFO into a call to a gather load or scatter store
5982 : : internal function. Return the final statement on success and set
5983 : : *TYPE_OUT to the vector type being loaded or stored.
5984 : :
5985 : : This function only handles gathers and scatters that were recognized
5986 : : as such from the outset (indicated by STMT_VINFO_GATHER_SCATTER_P). */
5987 : :
5988 : : static gimple *
5989 : 30018492 : vect_recog_gather_scatter_pattern (vec_info *vinfo,
5990 : : stmt_vec_info stmt_info, tree *type_out)
5991 : : {
5992 : : /* Currently we only support this for loop vectorization. */
5993 : 33354319 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
5994 : 3335827 : if (!loop_vinfo)
5995 : : return NULL;
5996 : :
5997 : : /* Make sure that we're looking at a gather load or scatter store. */
5998 : 3335827 : data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
5999 : 3335827 : if (!dr || !STMT_VINFO_GATHER_SCATTER_P (stmt_info))
6000 : : return NULL;
6001 : :
6002 : : /* Get the boolean that controls whether the load or store happens.
6003 : : This is null if the operation is unconditional. */
6004 : 32844 : tree mask = vect_get_load_store_mask (stmt_info);
6005 : :
6006 : : /* Make sure that the target supports an appropriate internal
6007 : : function for the gather/scatter operation. */
6008 : 32844 : gather_scatter_info gs_info;
6009 : 32844 : if (!vect_check_gather_scatter (stmt_info, loop_vinfo, &gs_info)
6010 : 32844 : || gs_info.ifn == IFN_LAST)
6011 : : return NULL;
6012 : :
6013 : : /* Convert the mask to the right form. */
6014 : 0 : tree gs_vectype = get_vectype_for_scalar_type (loop_vinfo,
6015 : : gs_info.element_type);
6016 : 0 : if (mask)
6017 : 0 : mask = vect_convert_mask_for_vectype (mask, gs_vectype, stmt_info,
6018 : : loop_vinfo);
6019 : 0 : else if (gs_info.ifn == IFN_MASK_SCATTER_STORE
6020 : 0 : || gs_info.ifn == IFN_MASK_GATHER_LOAD
6021 : 0 : || gs_info.ifn == IFN_MASK_LEN_SCATTER_STORE
6022 : 0 : || gs_info.ifn == IFN_MASK_LEN_GATHER_LOAD)
6023 : 0 : mask = build_int_cst (TREE_TYPE (truth_type_for (gs_vectype)), -1);
6024 : :
6025 : : /* Get the invariant base and non-invariant offset, converting the
6026 : : latter to the same width as the vector elements. */
6027 : 0 : tree base = gs_info.base;
6028 : 0 : tree offset_type = TREE_TYPE (gs_info.offset_vectype);
6029 : 0 : tree offset = vect_add_conversion_to_pattern (vinfo, offset_type,
6030 : : gs_info.offset, stmt_info);
6031 : :
6032 : : /* Build the new pattern statement. */
6033 : 0 : tree scale = size_int (gs_info.scale);
6034 : 0 : gcall *pattern_stmt;
6035 : :
6036 : 0 : if (DR_IS_READ (dr))
6037 : : {
6038 : 0 : tree zero = build_zero_cst (gs_info.element_type);
6039 : 0 : if (mask != NULL)
6040 : : {
6041 : 0 : int elsval = MASK_LOAD_ELSE_ZERO;
6042 : :
6043 : 0 : tree vec_els
6044 : 0 : = vect_get_mask_load_else (elsval, TREE_TYPE (gs_vectype));
6045 : 0 : pattern_stmt = gimple_build_call_internal (gs_info.ifn, 6, base,
6046 : : offset, scale, zero, mask,
6047 : : vec_els);
6048 : : }
6049 : : else
6050 : 0 : pattern_stmt = gimple_build_call_internal (gs_info.ifn, 4, base,
6051 : : offset, scale, zero);
6052 : 0 : tree lhs = gimple_get_lhs (stmt_info->stmt);
6053 : 0 : tree load_lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL);
6054 : 0 : gimple_call_set_lhs (pattern_stmt, load_lhs);
6055 : : }
6056 : : else
6057 : : {
6058 : 0 : tree rhs = vect_get_store_rhs (stmt_info);
6059 : 0 : if (mask != NULL)
6060 : 0 : pattern_stmt = gimple_build_call_internal (gs_info.ifn, 5,
6061 : : base, offset, scale, rhs,
6062 : : mask);
6063 : : else
6064 : 0 : pattern_stmt = gimple_build_call_internal (gs_info.ifn, 4,
6065 : : base, offset, scale, rhs);
6066 : : }
6067 : 0 : gimple_call_set_nothrow (pattern_stmt, true);
6068 : :
6069 : : /* Copy across relevant vectorization info and associate DR with the
6070 : : new pattern statement instead of the original statement. */
6071 : 0 : stmt_vec_info pattern_stmt_info = loop_vinfo->add_stmt (pattern_stmt);
6072 : 0 : loop_vinfo->move_dr (pattern_stmt_info, stmt_info);
6073 : :
6074 : 0 : tree vectype = STMT_VINFO_VECTYPE (stmt_info);
6075 : 0 : *type_out = vectype;
6076 : 0 : vect_pattern_detected ("gather/scatter pattern", stmt_info->stmt);
6077 : :
6078 : 0 : return pattern_stmt;
6079 : : }
6080 : :
6081 : : /* Helper method of vect_recog_cond_store_pattern, checks to see if COND_ARG
6082 : : is points to a load statement that reads the same data as that of
6083 : : STORE_VINFO. */
6084 : :
6085 : : static bool
6086 : 27861 : vect_cond_store_pattern_same_ref (vec_info *vinfo,
6087 : : stmt_vec_info store_vinfo, tree cond_arg)
6088 : : {
6089 : 27861 : stmt_vec_info load_stmt_vinfo = vinfo->lookup_def (cond_arg);
6090 : 27861 : if (!load_stmt_vinfo
6091 : 15631 : || !STMT_VINFO_DATA_REF (load_stmt_vinfo)
6092 : 10479 : || DR_IS_WRITE (STMT_VINFO_DATA_REF (load_stmt_vinfo))
6093 : 38340 : || !same_data_refs (STMT_VINFO_DATA_REF (store_vinfo),
6094 : : STMT_VINFO_DATA_REF (load_stmt_vinfo)))
6095 : 20071 : return false;
6096 : :
6097 : : return true;
6098 : : }
6099 : :
6100 : : /* Function vect_recog_cond_store_pattern
6101 : :
6102 : : Try to find the following pattern:
6103 : :
6104 : : x = *_3;
6105 : : c = a CMP b;
6106 : : y = c ? t_20 : x;
6107 : : *_3 = y;
6108 : :
6109 : : where the store of _3 happens on a conditional select on a value loaded
6110 : : from the same location. In such case we can elide the initial load if
6111 : : MASK_STORE is supported and instead only conditionally write out the result.
6112 : :
6113 : : The pattern produces for the above:
6114 : :
6115 : : c = a CMP b;
6116 : : .MASK_STORE (_3, c, t_20)
6117 : :
6118 : : Input:
6119 : :
6120 : : * STMT_VINFO: The stmt from which the pattern search begins. In the
6121 : : example, when this function is called with _3 then the search begins.
6122 : :
6123 : : Output:
6124 : :
6125 : : * TYPE_OUT: The type of the output of this pattern.
6126 : :
6127 : : * Return value: A new stmt that will be used to replace the sequence. */
6128 : :
6129 : : static gimple *
6130 : 30018492 : vect_recog_cond_store_pattern (vec_info *vinfo,
6131 : : stmt_vec_info stmt_vinfo, tree *type_out)
6132 : : {
6133 : 30018492 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
6134 : 3335827 : if (!loop_vinfo)
6135 : : return NULL;
6136 : :
6137 : 3335827 : gimple *store_stmt = STMT_VINFO_STMT (stmt_vinfo);
6138 : :
6139 : : /* Needs to be a gimple store where we have DR info for. */
6140 : 3335827 : if (!STMT_VINFO_DATA_REF (stmt_vinfo)
6141 : 780162 : || DR_IS_READ (STMT_VINFO_DATA_REF (stmt_vinfo))
6142 : 3614379 : || !gimple_store_p (store_stmt))
6143 : 3058852 : return NULL;
6144 : :
6145 : 276975 : tree st_rhs = gimple_assign_rhs1 (store_stmt);
6146 : :
6147 : 276975 : if (TREE_CODE (st_rhs) != SSA_NAME)
6148 : : return NULL;
6149 : :
6150 : 227754 : auto cond_vinfo = vinfo->lookup_def (st_rhs);
6151 : :
6152 : : /* If the condition isn't part of the loop then bool recog wouldn't have seen
6153 : : it and so this transformation may not be valid. */
6154 : 227754 : if (!cond_vinfo)
6155 : : return NULL;
6156 : :
6157 : 213555 : cond_vinfo = vect_stmt_to_vectorize (cond_vinfo);
6158 : 30223177 : gassign *cond_stmt = dyn_cast<gassign *> (STMT_VINFO_STMT (cond_vinfo));
6159 : 267735 : if (!cond_stmt || gimple_assign_rhs_code (cond_stmt) != COND_EXPR)
6160 : : return NULL;
6161 : :
6162 : : /* Check if the else value matches the original loaded one. */
6163 : 14571 : bool invert = false;
6164 : 14571 : tree cmp_ls = gimple_arg (cond_stmt, 0);
6165 : 14571 : if (TREE_CODE (cmp_ls) != SSA_NAME)
6166 : : return NULL;
6167 : :
6168 : 14571 : tree cond_arg1 = gimple_arg (cond_stmt, 1);
6169 : 14571 : tree cond_arg2 = gimple_arg (cond_stmt, 2);
6170 : :
6171 : 14571 : if (!vect_cond_store_pattern_same_ref (vinfo, stmt_vinfo, cond_arg2)
6172 : 14571 : && !(invert = vect_cond_store_pattern_same_ref (vinfo, stmt_vinfo,
6173 : : cond_arg1)))
6174 : : return NULL;
6175 : :
6176 : 7790 : vect_pattern_detected ("vect_recog_cond_store_pattern", store_stmt);
6177 : :
6178 : 7790 : tree scalar_type = TREE_TYPE (st_rhs);
6179 : 7790 : if (VECTOR_TYPE_P (scalar_type))
6180 : : return NULL;
6181 : :
6182 : 7790 : tree vectype = get_vectype_for_scalar_type (vinfo, scalar_type);
6183 : 7790 : if (vectype == NULL_TREE)
6184 : : return NULL;
6185 : :
6186 : 7790 : machine_mode mask_mode;
6187 : 7790 : machine_mode vecmode = TYPE_MODE (vectype);
6188 : 1644 : if (!VECTOR_MODE_P (vecmode)
6189 : 7790 : || targetm.vectorize.conditional_operation_is_expensive (IFN_MASK_STORE)
6190 : 0 : || !targetm.vectorize.get_mask_mode (vecmode).exists (&mask_mode)
6191 : 7790 : || !can_vec_mask_load_store_p (vecmode, mask_mode, false))
6192 : 7790 : return NULL;
6193 : :
6194 : 0 : tree base = DR_REF (STMT_VINFO_DATA_REF (stmt_vinfo));
6195 : 0 : if (may_be_nonaddressable_p (base))
6196 : : return NULL;
6197 : :
6198 : : /* We need to use the false parameter of the conditional select. */
6199 : 0 : tree cond_store_arg = invert ? cond_arg2 : cond_arg1;
6200 : 0 : tree cond_load_arg = invert ? cond_arg1 : cond_arg2;
6201 : 0 : gimple *load_stmt = SSA_NAME_DEF_STMT (cond_load_arg);
6202 : :
6203 : : /* This is a rough estimation to check that there aren't any aliasing stores
6204 : : in between the load and store. It's a bit strict, but for now it's good
6205 : : enough. */
6206 : 0 : if (gimple_vuse (load_stmt) != gimple_vuse (store_stmt))
6207 : : return NULL;
6208 : :
6209 : : /* If we have to invert the condition, i.e. use the true argument rather than
6210 : : the false argument, we have to negate the mask. */
6211 : 0 : if (invert)
6212 : : {
6213 : 0 : tree var = vect_recog_temp_ssa_var (boolean_type_node, NULL);
6214 : :
6215 : : /* Invert the mask using ^ 1. */
6216 : 0 : tree itype = TREE_TYPE (cmp_ls);
6217 : 0 : gassign *conv = gimple_build_assign (var, BIT_XOR_EXPR, cmp_ls,
6218 : : build_int_cst (itype, 1));
6219 : :
6220 : 0 : tree mask_vec_type = get_mask_type_for_scalar_type (vinfo, itype);
6221 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, conv, mask_vec_type, itype);
6222 : 0 : cmp_ls= var;
6223 : : }
6224 : :
6225 : 0 : if (TREE_CODE (base) != MEM_REF)
6226 : 0 : base = build_fold_addr_expr (base);
6227 : :
6228 : 0 : tree ptr = build_int_cst (reference_alias_ptr_type (base),
6229 : 0 : get_object_alignment (base));
6230 : :
6231 : : /* Convert the mask to the right form. */
6232 : 0 : tree mask = vect_convert_mask_for_vectype (cmp_ls, vectype, stmt_vinfo,
6233 : : vinfo);
6234 : :
6235 : 0 : gcall *call
6236 : 0 : = gimple_build_call_internal (IFN_MASK_STORE, 4, base, ptr, mask,
6237 : : cond_store_arg);
6238 : 0 : gimple_set_location (call, gimple_location (store_stmt));
6239 : :
6240 : : /* Copy across relevant vectorization info and associate DR with the
6241 : : new pattern statement instead of the original statement. */
6242 : 0 : stmt_vec_info pattern_stmt_info = loop_vinfo->add_stmt (call);
6243 : 0 : loop_vinfo->move_dr (pattern_stmt_info, stmt_vinfo);
6244 : :
6245 : 0 : *type_out = vectype;
6246 : 0 : return call;
6247 : : }
6248 : :
6249 : : /* Return true if TYPE is a non-boolean integer type. These are the types
6250 : : that we want to consider for narrowing. */
6251 : :
6252 : : static bool
6253 : 60307150 : vect_narrowable_type_p (tree type)
6254 : : {
6255 : 60307150 : return INTEGRAL_TYPE_P (type) && !VECT_SCALAR_BOOLEAN_TYPE_P (type);
6256 : : }
6257 : :
6258 : : /* Return true if the operation given by CODE can be truncated to N bits
6259 : : when only N bits of the output are needed. This is only true if bit N+1
6260 : : of the inputs has no effect on the low N bits of the result. */
6261 : :
6262 : : static bool
6263 : 14724555 : vect_truncatable_operation_p (tree_code code)
6264 : : {
6265 : 14724555 : switch (code)
6266 : : {
6267 : : case NEGATE_EXPR:
6268 : : case PLUS_EXPR:
6269 : : case MINUS_EXPR:
6270 : : case MULT_EXPR:
6271 : : case BIT_NOT_EXPR:
6272 : : case BIT_AND_EXPR:
6273 : : case BIT_IOR_EXPR:
6274 : : case BIT_XOR_EXPR:
6275 : : case COND_EXPR:
6276 : : return true;
6277 : :
6278 : 5765854 : default:
6279 : 5765854 : return false;
6280 : : }
6281 : : }
6282 : :
6283 : : /* Record that STMT_INFO could be changed from operating on TYPE to
6284 : : operating on a type with the precision and sign given by PRECISION
6285 : : and SIGN respectively. PRECISION is an arbitrary bit precision;
6286 : : it might not be a whole number of bytes. */
6287 : :
6288 : : static void
6289 : 2129599 : vect_set_operation_type (stmt_vec_info stmt_info, tree type,
6290 : : unsigned int precision, signop sign)
6291 : : {
6292 : : /* Round the precision up to a whole number of bytes. */
6293 : 2129599 : precision = vect_element_precision (precision);
6294 : 2129599 : if (precision < TYPE_PRECISION (type)
6295 : 2129599 : && (!stmt_info->operation_precision
6296 : 40369 : || stmt_info->operation_precision > precision))
6297 : : {
6298 : 1296789 : stmt_info->operation_precision = precision;
6299 : 1296789 : stmt_info->operation_sign = sign;
6300 : : }
6301 : 2129599 : }
6302 : :
6303 : : /* Record that STMT_INFO only requires MIN_INPUT_PRECISION from its
6304 : : non-boolean inputs, all of which have type TYPE. MIN_INPUT_PRECISION
6305 : : is an arbitrary bit precision; it might not be a whole number of bytes. */
6306 : :
6307 : : static void
6308 : 10662543 : vect_set_min_input_precision (stmt_vec_info stmt_info, tree type,
6309 : : unsigned int min_input_precision)
6310 : : {
6311 : : /* This operation in isolation only requires the inputs to have
6312 : : MIN_INPUT_PRECISION of precision, However, that doesn't mean
6313 : : that MIN_INPUT_PRECISION is a natural precision for the chain
6314 : : as a whole. E.g. consider something like:
6315 : :
6316 : : unsigned short *x, *y;
6317 : : *y = ((*x & 0xf0) >> 4) | (*y << 4);
6318 : :
6319 : : The right shift can be done on unsigned chars, and only requires the
6320 : : result of "*x & 0xf0" to be done on unsigned chars. But taking that
6321 : : approach would mean turning a natural chain of single-vector unsigned
6322 : : short operations into one that truncates "*x" and then extends
6323 : : "(*x & 0xf0) >> 4", with two vectors for each unsigned short
6324 : : operation and one vector for each unsigned char operation.
6325 : : This would be a significant pessimization.
6326 : :
6327 : : Instead only propagate the maximum of this precision and the precision
6328 : : required by the users of the result. This means that we don't pessimize
6329 : : the case above but continue to optimize things like:
6330 : :
6331 : : unsigned char *y;
6332 : : unsigned short *x;
6333 : : *y = ((*x & 0xf0) >> 4) | (*y << 4);
6334 : :
6335 : : Here we would truncate two vectors of *x to a single vector of
6336 : : unsigned chars and use single-vector unsigned char operations for
6337 : : everything else, rather than doing two unsigned short copies of
6338 : : "(*x & 0xf0) >> 4" and then truncating the result. */
6339 : 10662543 : min_input_precision = MAX (min_input_precision,
6340 : : stmt_info->min_output_precision);
6341 : :
6342 : 10662543 : if (min_input_precision < TYPE_PRECISION (type)
6343 : 10662543 : && (!stmt_info->min_input_precision
6344 : 66914 : || stmt_info->min_input_precision > min_input_precision))
6345 : 584264 : stmt_info->min_input_precision = min_input_precision;
6346 : 10662543 : }
6347 : :
6348 : : /* Subroutine of vect_determine_min_output_precision. Return true if
6349 : : we can calculate a reduced number of output bits for STMT_INFO,
6350 : : whose result is LHS. */
6351 : :
6352 : : static bool
6353 : 14122183 : vect_determine_min_output_precision_1 (vec_info *vinfo,
6354 : : stmt_vec_info stmt_info, tree lhs)
6355 : : {
6356 : : /* Take the maximum precision required by users of the result. */
6357 : 14122183 : unsigned int precision = 0;
6358 : 14122183 : imm_use_iterator iter;
6359 : 14122183 : use_operand_p use;
6360 : 15051407 : FOR_EACH_IMM_USE_FAST (use, iter, lhs)
6361 : : {
6362 : 14753915 : gimple *use_stmt = USE_STMT (use);
6363 : 14753915 : if (is_gimple_debug (use_stmt))
6364 : 621161 : continue;
6365 : 14132754 : stmt_vec_info use_stmt_info = vinfo->lookup_stmt (use_stmt);
6366 : 14132754 : if (!use_stmt_info || !use_stmt_info->min_input_precision)
6367 : : return false;
6368 : : /* The input precision recorded for COND_EXPRs applies only to the
6369 : : "then" and "else" values. */
6370 : 308452 : gassign *assign = dyn_cast <gassign *> (stmt_info->stmt);
6371 : 264577 : if (assign
6372 : 264577 : && gimple_assign_rhs_code (assign) == COND_EXPR
6373 : 389 : && use->use != gimple_assign_rhs2_ptr (assign)
6374 : 389 : && use->use != gimple_assign_rhs3_ptr (assign))
6375 : : return false;
6376 : 931708 : precision = MAX (precision, use_stmt_info->min_input_precision);
6377 : : }
6378 : :
6379 : 297492 : if (dump_enabled_p ())
6380 : 6670 : dump_printf_loc (MSG_NOTE, vect_location,
6381 : : "only the low %d bits of %T are significant\n",
6382 : : precision, lhs);
6383 : 297492 : stmt_info->min_output_precision = precision;
6384 : 297492 : return true;
6385 : : }
6386 : :
6387 : : /* Calculate min_output_precision for STMT_INFO. */
6388 : :
6389 : : static void
6390 : 36939893 : vect_determine_min_output_precision (vec_info *vinfo, stmt_vec_info stmt_info)
6391 : : {
6392 : : /* We're only interested in statements with a narrowable result. */
6393 : 36939893 : tree lhs = gimple_get_lhs (stmt_info->stmt);
6394 : 36939893 : if (!lhs
6395 : 28990991 : || TREE_CODE (lhs) != SSA_NAME
6396 : 61157096 : || !vect_narrowable_type_p (TREE_TYPE (lhs)))
6397 : : return;
6398 : :
6399 : 14122183 : if (!vect_determine_min_output_precision_1 (vinfo, stmt_info, lhs))
6400 : 13824691 : stmt_info->min_output_precision = TYPE_PRECISION (TREE_TYPE (lhs));
6401 : : }
6402 : :
6403 : : /* Use range information to decide whether STMT (described by STMT_INFO)
6404 : : could be done in a narrower type. This is effectively a forward
6405 : : propagation, since it uses context-independent information that applies
6406 : : to all users of an SSA name. */
6407 : :
6408 : : static void
6409 : 20286152 : vect_determine_precisions_from_range (stmt_vec_info stmt_info, gassign *stmt)
6410 : : {
6411 : 20286152 : tree lhs = gimple_assign_lhs (stmt);
6412 : 20286152 : if (!lhs || TREE_CODE (lhs) != SSA_NAME)
6413 : 18280373 : return;
6414 : :
6415 : 15803795 : tree type = TREE_TYPE (lhs);
6416 : 15803795 : if (!vect_narrowable_type_p (type))
6417 : : return;
6418 : :
6419 : : /* First see whether we have any useful range information for the result. */
6420 : 10683974 : unsigned int precision = TYPE_PRECISION (type);
6421 : 10683974 : signop sign = TYPE_SIGN (type);
6422 : 10683974 : wide_int min_value, max_value;
6423 : 10683974 : if (!vect_get_range_info (lhs, &min_value, &max_value))
6424 : : return;
6425 : :
6426 : 5018105 : tree_code code = gimple_assign_rhs_code (stmt);
6427 : 5018105 : unsigned int nops = gimple_num_ops (stmt);
6428 : :
6429 : 5018105 : if (!vect_truncatable_operation_p (code))
6430 : : {
6431 : : /* Handle operations that can be computed in type T if all inputs
6432 : : and outputs can be represented in type T. Also handle left and
6433 : : right shifts, where (in addition) the maximum shift amount must
6434 : : be less than the number of bits in T. */
6435 : 1932835 : bool is_shift;
6436 : 1932835 : switch (code)
6437 : : {
6438 : : case LSHIFT_EXPR:
6439 : : case RSHIFT_EXPR:
6440 : : is_shift = true;
6441 : : break;
6442 : :
6443 : 245854 : case ABS_EXPR:
6444 : 245854 : case MIN_EXPR:
6445 : 245854 : case MAX_EXPR:
6446 : 245854 : case TRUNC_DIV_EXPR:
6447 : 245854 : case CEIL_DIV_EXPR:
6448 : 245854 : case FLOOR_DIV_EXPR:
6449 : 245854 : case ROUND_DIV_EXPR:
6450 : 245854 : case EXACT_DIV_EXPR:
6451 : : /* Modulus is excluded because it is typically calculated by doing
6452 : : a division, for which minimum signed / -1 isn't representable in
6453 : : the original signed type. We could take the division range into
6454 : : account instead, if handling modulus ever becomes important. */
6455 : 245854 : is_shift = false;
6456 : 245854 : break;
6457 : :
6458 : : default:
6459 : : return;
6460 : : }
6461 : 1261814 : for (unsigned int i = 1; i < nops; ++i)
6462 : : {
6463 : 968289 : tree op = gimple_op (stmt, i);
6464 : 968289 : wide_int op_min_value, op_max_value;
6465 : 968289 : if (TREE_CODE (op) == INTEGER_CST)
6466 : : {
6467 : 297753 : unsigned int op_precision = TYPE_PRECISION (TREE_TYPE (op));
6468 : 297753 : op_min_value = op_max_value = wi::to_wide (op, op_precision);
6469 : : }
6470 : 670536 : else if (TREE_CODE (op) == SSA_NAME)
6471 : : {
6472 : 670536 : if (!vect_get_range_info (op, &op_min_value, &op_max_value))
6473 : : return;
6474 : : }
6475 : : else
6476 : : return;
6477 : :
6478 : 644757 : if (is_shift && i == 2)
6479 : : {
6480 : : /* There needs to be one more bit than the maximum shift amount.
6481 : :
6482 : : If the maximum shift amount is already 1 less than PRECISION
6483 : : then we can't narrow the shift further. Dealing with that
6484 : : case first ensures that we can safely use an unsigned range
6485 : : below.
6486 : :
6487 : : op_min_value isn't relevant, since shifts by negative amounts
6488 : : are UB. */
6489 : 195962 : if (wi::geu_p (op_max_value, precision - 1))
6490 : : return;
6491 : 179114 : unsigned int min_bits = op_max_value.to_uhwi () + 1;
6492 : :
6493 : : /* As explained below, we can convert a signed shift into an
6494 : : unsigned shift if the sign bit is always clear. At this
6495 : : point we've already processed the ranges of the output and
6496 : : the first input. */
6497 : 179114 : auto op_sign = sign;
6498 : 179114 : if (sign == SIGNED && !wi::neg_p (min_value))
6499 : : op_sign = UNSIGNED;
6500 : 358228 : op_min_value = wide_int::from (wi::min_value (min_bits, op_sign),
6501 : 179114 : precision, op_sign);
6502 : 358228 : op_max_value = wide_int::from (wi::max_value (min_bits, op_sign),
6503 : 179114 : precision, op_sign);
6504 : : }
6505 : 627909 : min_value = wi::min (min_value, op_min_value, sign);
6506 : 627909 : max_value = wi::max (max_value, op_max_value, sign);
6507 : 968289 : }
6508 : : }
6509 : :
6510 : : /* Try to switch signed types for unsigned types if we can.
6511 : : This is better for two reasons. First, unsigned ops tend
6512 : : to be cheaper than signed ops. Second, it means that we can
6513 : : handle things like:
6514 : :
6515 : : signed char c;
6516 : : int res = (int) c & 0xff00; // range [0x0000, 0xff00]
6517 : :
6518 : : as:
6519 : :
6520 : : signed char c;
6521 : : unsigned short res_1 = (unsigned short) c & 0xff00;
6522 : : int res = (int) res_1;
6523 : :
6524 : : where the intermediate result res_1 has unsigned rather than
6525 : : signed type. */
6526 : 3378795 : if (sign == SIGNED && !wi::neg_p (min_value))
6527 : : sign = UNSIGNED;
6528 : :
6529 : : /* See what precision is required for MIN_VALUE and MAX_VALUE. */
6530 : 3378795 : unsigned int precision1 = wi::min_precision (min_value, sign);
6531 : 3378795 : unsigned int precision2 = wi::min_precision (max_value, sign);
6532 : 3378795 : unsigned int value_precision = MAX (precision1, precision2);
6533 : 3378795 : if (value_precision >= precision)
6534 : : return;
6535 : :
6536 : 2005779 : if (dump_enabled_p ())
6537 : 78997 : dump_printf_loc (MSG_NOTE, vect_location, "can narrow to %s:%d"
6538 : : " without loss of precision: %G",
6539 : : sign == SIGNED ? "signed" : "unsigned",
6540 : : value_precision, (gimple *) stmt);
6541 : :
6542 : 2005779 : vect_set_operation_type (stmt_info, type, value_precision, sign);
6543 : 2005779 : vect_set_min_input_precision (stmt_info, type, value_precision);
6544 : 10683974 : }
6545 : :
6546 : : /* Use information about the users of STMT's result to decide whether
6547 : : STMT (described by STMT_INFO) could be done in a narrower type.
6548 : : This is effectively a backward propagation. */
6549 : :
6550 : : static void
6551 : 20286152 : vect_determine_precisions_from_users (stmt_vec_info stmt_info, gassign *stmt)
6552 : : {
6553 : 20286152 : tree_code code = gimple_assign_rhs_code (stmt);
6554 : 20286152 : unsigned int opno = (code == COND_EXPR ? 2 : 1);
6555 : 20286152 : tree type = TREE_TYPE (gimple_op (stmt, opno));
6556 : 20286152 : if (!vect_narrowable_type_p (type))
6557 : 11629388 : return;
6558 : :
6559 : 12523539 : unsigned int precision = TYPE_PRECISION (type);
6560 : 12523539 : unsigned int operation_precision, min_input_precision;
6561 : 12523539 : switch (code)
6562 : : {
6563 : 2318557 : CASE_CONVERT:
6564 : : /* Only the bits that contribute to the output matter. Don't change
6565 : : the precision of the operation itself. */
6566 : 2318557 : operation_precision = precision;
6567 : 2318557 : min_input_precision = stmt_info->min_output_precision;
6568 : 2318557 : break;
6569 : :
6570 : 498532 : case LSHIFT_EXPR:
6571 : 498532 : case RSHIFT_EXPR:
6572 : 498532 : {
6573 : 498532 : tree shift = gimple_assign_rhs2 (stmt);
6574 : 498532 : unsigned int min_const_shift, max_const_shift;
6575 : 498532 : wide_int min_shift, max_shift;
6576 : 498532 : if (TREE_CODE (shift) == SSA_NAME
6577 : 96908 : && vect_get_range_info (shift, &min_shift, &max_shift)
6578 : 73157 : && wi::ge_p (min_shift, 0, TYPE_SIGN (TREE_TYPE (shift)))
6579 : 568974 : && wi::lt_p (max_shift, TYPE_PRECISION (type),
6580 : 70442 : TYPE_SIGN (TREE_TYPE (shift))))
6581 : : {
6582 : 63260 : min_const_shift = min_shift.to_uhwi ();
6583 : 63260 : max_const_shift = max_shift.to_uhwi ();
6584 : : }
6585 : 435272 : else if (TREE_CODE (shift) == INTEGER_CST
6586 : 836896 : && wi::ltu_p (wi::to_widest (shift), precision))
6587 : 401516 : min_const_shift = max_const_shift = TREE_INT_CST_LOW (shift);
6588 : : else
6589 : 33756 : return;
6590 : 464776 : if (code == LSHIFT_EXPR)
6591 : : {
6592 : : /* Avoid creating an undefined shift.
6593 : :
6594 : : ??? We could instead use min_output_precision as-is and
6595 : : optimize out-of-range shifts to zero. However, only
6596 : : degenerate testcases shift away all their useful input data,
6597 : : and it isn't natural to drop input operations in the middle
6598 : : of vectorization. This sort of thing should really be
6599 : : handled before vectorization. */
6600 : 143445 : operation_precision = MAX (stmt_info->min_output_precision,
6601 : : max_const_shift + 1);
6602 : : /* We need CONST_SHIFT fewer bits of the input. */
6603 : 143445 : min_input_precision = (MAX (operation_precision, max_const_shift)
6604 : : - min_const_shift);
6605 : : }
6606 : : else
6607 : : {
6608 : : /* We need CONST_SHIFT extra bits to do the operation. */
6609 : 321331 : operation_precision = (stmt_info->min_output_precision
6610 : : + max_const_shift);
6611 : 321331 : min_input_precision = operation_precision;
6612 : : }
6613 : 464776 : break;
6614 : 498532 : }
6615 : :
6616 : 9706450 : default:
6617 : 9706450 : if (vect_truncatable_operation_p (code))
6618 : : {
6619 : : /* Input bit N has no effect on output bits N-1 and lower. */
6620 : 5873431 : operation_precision = stmt_info->min_output_precision;
6621 : 5873431 : min_input_precision = operation_precision;
6622 : 5873431 : break;
6623 : : }
6624 : : return;
6625 : : }
6626 : :
6627 : 8656764 : if (operation_precision < precision)
6628 : : {
6629 : 123820 : if (dump_enabled_p ())
6630 : 3385 : dump_printf_loc (MSG_NOTE, vect_location, "can narrow to %s:%d"
6631 : : " without affecting users: %G",
6632 : 3385 : TYPE_UNSIGNED (type) ? "unsigned" : "signed",
6633 : : operation_precision, (gimple *) stmt);
6634 : 247640 : vect_set_operation_type (stmt_info, type, operation_precision,
6635 : 123820 : TYPE_SIGN (type));
6636 : : }
6637 : 8656764 : vect_set_min_input_precision (stmt_info, type, min_input_precision);
6638 : : }
6639 : :
6640 : : /* Return true if the statement described by STMT_INFO sets a boolean
6641 : : SSA_NAME and if we know how to vectorize this kind of statement using
6642 : : vector mask types. */
6643 : :
6644 : : static bool
6645 : 36939893 : possible_vector_mask_operation_p (stmt_vec_info stmt_info)
6646 : : {
6647 : 36939893 : tree lhs = gimple_get_lhs (stmt_info->stmt);
6648 : 36939893 : tree_code code = ERROR_MARK;
6649 : 36939893 : gassign *assign = NULL;
6650 : 36939893 : gcond *cond = NULL;
6651 : :
6652 : 36939893 : if ((assign = dyn_cast <gassign *> (stmt_info->stmt)))
6653 : 20286152 : code = gimple_assign_rhs_code (assign);
6654 : 16653741 : else if ((cond = dyn_cast <gcond *> (stmt_info->stmt)))
6655 : : {
6656 : 4954678 : lhs = gimple_cond_lhs (cond);
6657 : 4954678 : code = gimple_cond_code (cond);
6658 : : }
6659 : :
6660 : 36939893 : if (!lhs
6661 : 33945669 : || TREE_CODE (lhs) != SSA_NAME
6662 : 66111544 : || !VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (lhs)))
6663 : : return false;
6664 : :
6665 : 2106290 : if (code != ERROR_MARK)
6666 : : {
6667 : 1797109 : switch (code)
6668 : : {
6669 : : CASE_CONVERT:
6670 : : case SSA_NAME:
6671 : : case BIT_NOT_EXPR:
6672 : : case BIT_IOR_EXPR:
6673 : : case BIT_XOR_EXPR:
6674 : : case BIT_AND_EXPR:
6675 : : return true;
6676 : :
6677 : 1403758 : default:
6678 : 1403758 : return TREE_CODE_CLASS (code) == tcc_comparison;
6679 : : }
6680 : : }
6681 : 309181 : else if (is_a <gphi *> (stmt_info->stmt))
6682 : 199370 : return true;
6683 : : return false;
6684 : : }
6685 : :
6686 : : /* If STMT_INFO sets a boolean SSA_NAME, see whether we should use
6687 : : a vector mask type instead of a normal vector type. Record the
6688 : : result in STMT_INFO->mask_precision. */
6689 : :
6690 : : static void
6691 : 36939893 : vect_determine_mask_precision (vec_info *vinfo, stmt_vec_info stmt_info)
6692 : : {
6693 : 36939893 : if (!possible_vector_mask_operation_p (stmt_info))
6694 : : return;
6695 : :
6696 : : /* If at least one boolean input uses a vector mask type,
6697 : : pick the mask type with the narrowest elements.
6698 : :
6699 : : ??? This is the traditional behavior. It should always produce
6700 : : the smallest number of operations, but isn't necessarily the
6701 : : optimal choice. For example, if we have:
6702 : :
6703 : : a = b & c
6704 : :
6705 : : where:
6706 : :
6707 : : - the user of a wants it to have a mask type for 16-bit elements (M16)
6708 : : - b also uses M16
6709 : : - c uses a mask type for 8-bit elements (M8)
6710 : :
6711 : : then picking M8 gives:
6712 : :
6713 : : - 1 M16->M8 pack for b
6714 : : - 1 M8 AND for a
6715 : : - 2 M8->M16 unpacks for the user of a
6716 : :
6717 : : whereas picking M16 would have given:
6718 : :
6719 : : - 2 M8->M16 unpacks for c
6720 : : - 2 M16 ANDs for a
6721 : :
6722 : : The number of operations are equal, but M16 would have given
6723 : : a shorter dependency chain and allowed more ILP. */
6724 : 1956069 : unsigned int precision = ~0U;
6725 : 1956069 : gimple *stmt = STMT_VINFO_STMT (stmt_info);
6726 : :
6727 : : /* If the statement compares two values that shouldn't use vector masks,
6728 : : try comparing the values as normal scalars instead. */
6729 : 1956069 : tree_code code = ERROR_MARK;
6730 : 1956069 : tree op0_type;
6731 : 1956069 : unsigned int nops = -1;
6732 : 1956069 : unsigned int ops_start = 0;
6733 : :
6734 : 1956069 : if (gassign *assign = dyn_cast <gassign *> (stmt))
6735 : : {
6736 : 1203434 : code = gimple_assign_rhs_code (assign);
6737 : 1203434 : op0_type = TREE_TYPE (gimple_assign_rhs1 (assign));
6738 : 1203434 : nops = gimple_num_ops (assign);
6739 : 1203434 : ops_start = 1;
6740 : : }
6741 : 752635 : else if (gcond *cond = dyn_cast <gcond *> (stmt))
6742 : : {
6743 : 553265 : code = gimple_cond_code (cond);
6744 : 553265 : op0_type = TREE_TYPE (gimple_cond_lhs (cond));
6745 : 553265 : nops = 2;
6746 : 553265 : ops_start = 0;
6747 : : }
6748 : :
6749 : 1756699 : if (code != ERROR_MARK)
6750 : : {
6751 : 5230515 : for (unsigned int i = ops_start; i < nops; ++i)
6752 : : {
6753 : 3473816 : tree rhs = gimple_op (stmt, i);
6754 : 3473816 : if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (rhs)))
6755 : 1630973 : continue;
6756 : :
6757 : 1842843 : stmt_vec_info def_stmt_info = vinfo->lookup_def (rhs);
6758 : 1842843 : if (!def_stmt_info)
6759 : : /* Don't let external or constant operands influence the choice.
6760 : : We can convert them to whichever vector type we pick. */
6761 : 566734 : continue;
6762 : :
6763 : 1276109 : if (def_stmt_info->mask_precision)
6764 : : {
6765 : 1064042 : if (precision > def_stmt_info->mask_precision)
6766 : 3473816 : precision = def_stmt_info->mask_precision;
6767 : : }
6768 : : }
6769 : :
6770 : 1756699 : if (precision == ~0U
6771 : 1413131 : && TREE_CODE_CLASS (code) == tcc_comparison)
6772 : : {
6773 : 1214106 : scalar_mode mode;
6774 : 1214106 : tree vectype, mask_type;
6775 : 1214106 : if (is_a <scalar_mode> (TYPE_MODE (op0_type), &mode)
6776 : 1214106 : && (vectype = get_vectype_for_scalar_type (vinfo, op0_type))
6777 : 1050229 : && (mask_type = get_mask_type_for_scalar_type (vinfo, op0_type))
6778 : 1050229 : && expand_vec_cmp_expr_p (vectype, mask_type, code))
6779 : 1614564 : precision = GET_MODE_BITSIZE (mode);
6780 : : }
6781 : : }
6782 : : else
6783 : : {
6784 : 199370 : gphi *phi = as_a <gphi *> (stmt_info->stmt);
6785 : 715167 : for (unsigned i = 0; i < gimple_phi_num_args (phi); ++i)
6786 : : {
6787 : 515797 : tree rhs = gimple_phi_arg_def (phi, i);
6788 : :
6789 : 515797 : stmt_vec_info def_stmt_info = vinfo->lookup_def (rhs);
6790 : 515797 : if (!def_stmt_info)
6791 : : /* Don't let external or constant operands influence the choice.
6792 : : We can convert them to whichever vector type we pick. */
6793 : 304918 : continue;
6794 : :
6795 : 210879 : if (def_stmt_info->mask_precision)
6796 : : {
6797 : 185770 : if (precision > def_stmt_info->mask_precision)
6798 : 515797 : precision = def_stmt_info->mask_precision;
6799 : : }
6800 : : }
6801 : : }
6802 : :
6803 : 1956069 : if (dump_enabled_p ())
6804 : : {
6805 : 7235 : if (precision == ~0U)
6806 : 1676 : dump_printf_loc (MSG_NOTE, vect_location,
6807 : : "using normal nonmask vectors for %G",
6808 : : stmt_info->stmt);
6809 : : else
6810 : 5559 : dump_printf_loc (MSG_NOTE, vect_location,
6811 : : "using boolean precision %d for %G",
6812 : : precision, stmt_info->stmt);
6813 : : }
6814 : :
6815 : 1956069 : stmt_info->mask_precision = precision;
6816 : : }
6817 : :
6818 : : /* Handle vect_determine_precisions for STMT_INFO, given that we
6819 : : have already done so for the users of its result. */
6820 : :
6821 : : void
6822 : 36939893 : vect_determine_stmt_precisions (vec_info *vinfo, stmt_vec_info stmt_info)
6823 : : {
6824 : 36939893 : vect_determine_min_output_precision (vinfo, stmt_info);
6825 : 36939893 : if (gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt))
6826 : : {
6827 : 20286152 : vect_determine_precisions_from_range (stmt_info, stmt);
6828 : 20286152 : vect_determine_precisions_from_users (stmt_info, stmt);
6829 : : }
6830 : 36939893 : }
6831 : :
6832 : : /* Walk backwards through the vectorizable region to determine the
6833 : : values of these fields:
6834 : :
6835 : : - min_output_precision
6836 : : - min_input_precision
6837 : : - operation_precision
6838 : : - operation_sign. */
6839 : :
6840 : : void
6841 : 955241 : vect_determine_precisions (vec_info *vinfo)
6842 : : {
6843 : 955241 : basic_block *bbs = vinfo->bbs;
6844 : 955241 : unsigned int nbbs = vinfo->nbbs;
6845 : :
6846 : 955241 : DUMP_VECT_SCOPE ("vect_determine_precisions");
6847 : :
6848 : 12588507 : for (unsigned int i = 0; i < nbbs; i++)
6849 : : {
6850 : 11633266 : basic_block bb = bbs[i];
6851 : 19044710 : for (auto gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
6852 : : {
6853 : 7411444 : stmt_vec_info stmt_info = vinfo->lookup_stmt (gsi.phi ());
6854 : 7411444 : if (stmt_info && STMT_VINFO_VECTORIZABLE (stmt_info))
6855 : 7224089 : vect_determine_mask_precision (vinfo, stmt_info);
6856 : : }
6857 : 112249449 : for (auto gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
6858 : : {
6859 : 88982917 : stmt_vec_info stmt_info = vinfo->lookup_stmt (gsi_stmt (gsi));
6860 : 88982917 : if (stmt_info && STMT_VINFO_VECTORIZABLE (stmt_info))
6861 : 29715804 : vect_determine_mask_precision (vinfo, stmt_info);
6862 : : }
6863 : : }
6864 : 12588507 : for (unsigned int i = 0; i < nbbs; i++)
6865 : : {
6866 : 11633266 : basic_block bb = bbs[nbbs - i - 1];
6867 : 201232366 : for (auto gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi))
6868 : : {
6869 : 88982917 : stmt_vec_info stmt_info = vinfo->lookup_stmt (gsi_stmt (gsi));
6870 : 88982917 : if (stmt_info && STMT_VINFO_VECTORIZABLE (stmt_info))
6871 : 29715804 : vect_determine_stmt_precisions (vinfo, stmt_info);
6872 : : }
6873 : 19044710 : for (auto gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
6874 : : {
6875 : 7411444 : stmt_vec_info stmt_info = vinfo->lookup_stmt (gsi.phi ());
6876 : 7411444 : if (stmt_info && STMT_VINFO_VECTORIZABLE (stmt_info))
6877 : 7224089 : vect_determine_stmt_precisions (vinfo, stmt_info);
6878 : : }
6879 : : }
6880 : 955241 : }
6881 : :
6882 : : typedef gimple *(*vect_recog_func_ptr) (vec_info *, stmt_vec_info, tree *);
6883 : :
6884 : : struct vect_recog_func
6885 : : {
6886 : : vect_recog_func_ptr fn;
6887 : : const char *name;
6888 : : };
6889 : :
6890 : : /* Note that ordering matters - the first pattern matching on a stmt is
6891 : : taken which means usually the more complex one needs to preceed the
6892 : : less comples onex (widen_sum only after dot_prod or sad for example). */
6893 : : static vect_recog_func vect_vect_recog_func_ptrs[] = {
6894 : : { vect_recog_bitfield_ref_pattern, "bitfield_ref" },
6895 : : { vect_recog_bit_insert_pattern, "bit_insert" },
6896 : : { vect_recog_abd_pattern, "abd" },
6897 : : { vect_recog_over_widening_pattern, "over_widening" },
6898 : : /* Must come after over_widening, which narrows the shift as much as
6899 : : possible beforehand. */
6900 : : { vect_recog_average_pattern, "average" },
6901 : : { vect_recog_cond_expr_convert_pattern, "cond_expr_convert" },
6902 : : { vect_recog_mulhs_pattern, "mult_high" },
6903 : : { vect_recog_cast_forwprop_pattern, "cast_forwprop" },
6904 : : { vect_recog_widen_mult_pattern, "widen_mult" },
6905 : : { vect_recog_dot_prod_pattern, "dot_prod" },
6906 : : { vect_recog_sad_pattern, "sad" },
6907 : : { vect_recog_widen_sum_pattern, "widen_sum" },
6908 : : { vect_recog_pow_pattern, "pow" },
6909 : : { vect_recog_popcount_clz_ctz_ffs_pattern, "popcount_clz_ctz_ffs" },
6910 : : { vect_recog_ctz_ffs_pattern, "ctz_ffs" },
6911 : : { vect_recog_widen_shift_pattern, "widen_shift" },
6912 : : { vect_recog_rotate_pattern, "rotate" },
6913 : : { vect_recog_vector_vector_shift_pattern, "vector_vector_shift" },
6914 : : { vect_recog_divmod_pattern, "divmod" },
6915 : : { vect_recog_mod_var_pattern, "modvar" },
6916 : : { vect_recog_mult_pattern, "mult" },
6917 : : { vect_recog_sat_add_pattern, "sat_add" },
6918 : : { vect_recog_sat_sub_pattern, "sat_sub" },
6919 : : { vect_recog_sat_trunc_pattern, "sat_trunc" },
6920 : : { vect_recog_gcond_pattern, "gcond" },
6921 : : { vect_recog_bool_pattern, "bool" },
6922 : : /* This must come before mask conversion, and includes the parts
6923 : : of mask conversion that are needed for gather and scatter
6924 : : internal functions. */
6925 : : { vect_recog_gather_scatter_pattern, "gather_scatter" },
6926 : : { vect_recog_cond_store_pattern, "cond_store" },
6927 : : { vect_recog_mask_conversion_pattern, "mask_conversion" },
6928 : : { vect_recog_widen_plus_pattern, "widen_plus" },
6929 : : { vect_recog_widen_minus_pattern, "widen_minus" },
6930 : : { vect_recog_widen_abd_pattern, "widen_abd" },
6931 : : /* These must come after the double widening ones. */
6932 : : };
6933 : :
6934 : : /* Mark statements that are involved in a pattern. */
6935 : :
6936 : : void
6937 : 826665 : vect_mark_pattern_stmts (vec_info *vinfo,
6938 : : stmt_vec_info orig_stmt_info, gimple *pattern_stmt,
6939 : : tree pattern_vectype)
6940 : : {
6941 : 826665 : stmt_vec_info orig_stmt_info_saved = orig_stmt_info;
6942 : 826665 : gimple *def_seq = STMT_VINFO_PATTERN_DEF_SEQ (orig_stmt_info);
6943 : :
6944 : 826665 : gimple *orig_pattern_stmt = NULL;
6945 : 826665 : if (is_pattern_stmt_p (orig_stmt_info))
6946 : : {
6947 : : /* We're replacing a statement in an existing pattern definition
6948 : : sequence. */
6949 : 12616 : orig_pattern_stmt = orig_stmt_info->stmt;
6950 : 12616 : if (dump_enabled_p ())
6951 : 679 : dump_printf_loc (MSG_NOTE, vect_location,
6952 : : "replacing earlier pattern %G", orig_pattern_stmt);
6953 : :
6954 : : /* To keep the book-keeping simple, just swap the lhs of the
6955 : : old and new statements, so that the old one has a valid but
6956 : : unused lhs. */
6957 : 12616 : tree old_lhs = gimple_get_lhs (orig_pattern_stmt);
6958 : 12616 : gimple_set_lhs (orig_pattern_stmt, gimple_get_lhs (pattern_stmt));
6959 : 12616 : gimple_set_lhs (pattern_stmt, old_lhs);
6960 : :
6961 : 12616 : if (dump_enabled_p ())
6962 : 679 : dump_printf_loc (MSG_NOTE, vect_location, "with %G", pattern_stmt);
6963 : :
6964 : : /* Switch to the statement that ORIG replaces. */
6965 : 12616 : orig_stmt_info = STMT_VINFO_RELATED_STMT (orig_stmt_info);
6966 : :
6967 : : /* We shouldn't be replacing the main pattern statement. */
6968 : 12616 : gcc_assert (STMT_VINFO_RELATED_STMT (orig_stmt_info)->stmt
6969 : : != orig_pattern_stmt);
6970 : : }
6971 : :
6972 : 826665 : if (def_seq)
6973 : : for (gimple_stmt_iterator si = gsi_start (def_seq);
6974 : 1825602 : !gsi_end_p (si); gsi_next (&si))
6975 : : {
6976 : 1116821 : if (dump_enabled_p ())
6977 : 23592 : dump_printf_loc (MSG_NOTE, vect_location,
6978 : : "extra pattern stmt: %G", gsi_stmt (si));
6979 : 1116821 : stmt_vec_info pattern_stmt_info
6980 : 1116821 : = vect_init_pattern_stmt (vinfo, gsi_stmt (si),
6981 : : orig_stmt_info, pattern_vectype);
6982 : : /* Stmts in the def sequence are not vectorizable cycle or
6983 : : induction defs, instead they should all be vect_internal_def
6984 : : feeding the main pattern stmt which retains this def type. */
6985 : 1116821 : STMT_VINFO_DEF_TYPE (pattern_stmt_info) = vect_internal_def;
6986 : : }
6987 : :
6988 : 826665 : if (orig_pattern_stmt)
6989 : : {
6990 : 12616 : vect_init_pattern_stmt (vinfo, pattern_stmt,
6991 : : orig_stmt_info, pattern_vectype);
6992 : :
6993 : : /* Insert all the new pattern statements before the original one. */
6994 : 12616 : gimple_seq *orig_def_seq = &STMT_VINFO_PATTERN_DEF_SEQ (orig_stmt_info);
6995 : 12616 : gimple_stmt_iterator gsi = gsi_for_stmt (orig_pattern_stmt,
6996 : : orig_def_seq);
6997 : 12616 : gsi_insert_seq_before_without_update (&gsi, def_seq, GSI_SAME_STMT);
6998 : 12616 : gsi_insert_before_without_update (&gsi, pattern_stmt, GSI_SAME_STMT);
6999 : :
7000 : : /* Remove the pattern statement that this new pattern replaces. */
7001 : 12616 : gsi_remove (&gsi, false);
7002 : : }
7003 : : else
7004 : 814049 : vect_set_pattern_stmt (vinfo,
7005 : : pattern_stmt, orig_stmt_info, pattern_vectype);
7006 : :
7007 : : /* For any conditionals mark them as vect_condition_def. */
7008 : 826665 : if (is_a <gcond *> (pattern_stmt))
7009 : 264291 : STMT_VINFO_DEF_TYPE (STMT_VINFO_RELATED_STMT (orig_stmt_info)) = vect_condition_def;
7010 : :
7011 : : /* Transfer reduction path info to the pattern. */
7012 : 826665 : if (STMT_VINFO_REDUC_IDX (orig_stmt_info_saved) != -1)
7013 : : {
7014 : 9072 : gimple_match_op op;
7015 : 9072 : if (!gimple_extract_op (orig_stmt_info_saved->stmt, &op))
7016 : 0 : gcc_unreachable ();
7017 : 9072 : tree lookfor = op.ops[STMT_VINFO_REDUC_IDX (orig_stmt_info)];
7018 : : /* Search the pattern def sequence and the main pattern stmt. Note
7019 : : we may have inserted all into a containing pattern def sequence
7020 : : so the following is a bit awkward. */
7021 : 9072 : gimple_stmt_iterator si;
7022 : 9072 : gimple *s;
7023 : 9072 : if (def_seq)
7024 : : {
7025 : 8418 : si = gsi_start (def_seq);
7026 : 8418 : s = gsi_stmt (si);
7027 : 8418 : gsi_next (&si);
7028 : : }
7029 : : else
7030 : : {
7031 : : si = gsi_none ();
7032 : : s = pattern_stmt;
7033 : : }
7034 : 17605 : do
7035 : : {
7036 : 17605 : bool found = false;
7037 : 17605 : if (gimple_extract_op (s, &op))
7038 : 42240 : for (unsigned i = 0; i < op.num_ops; ++i)
7039 : 33868 : if (op.ops[i] == lookfor)
7040 : : {
7041 : 9233 : STMT_VINFO_REDUC_IDX (vinfo->lookup_stmt (s)) = i;
7042 : 9233 : lookfor = gimple_get_lhs (s);
7043 : 9233 : found = true;
7044 : 9233 : break;
7045 : : }
7046 : 17605 : if (s == pattern_stmt)
7047 : : {
7048 : 9072 : if (!found && dump_enabled_p ())
7049 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
7050 : : "failed to update reduction index.\n");
7051 : 9072 : break;
7052 : : }
7053 : 8533 : if (gsi_end_p (si))
7054 : : s = pattern_stmt;
7055 : : else
7056 : : {
7057 : 115 : s = gsi_stmt (si);
7058 : 115 : if (s == pattern_stmt)
7059 : : /* Found the end inside a bigger pattern def seq. */
7060 : : si = gsi_none ();
7061 : : else
7062 : 115 : gsi_next (&si);
7063 : : }
7064 : : } while (1);
7065 : : }
7066 : 826665 : }
7067 : :
7068 : : /* Function vect_pattern_recog_1
7069 : :
7070 : : Input:
7071 : : PATTERN_RECOG_FUNC: A pointer to a function that detects a certain
7072 : : computation pattern.
7073 : : STMT_INFO: A stmt from which the pattern search should start.
7074 : :
7075 : : If PATTERN_RECOG_FUNC successfully detected the pattern, it creates
7076 : : a sequence of statements that has the same functionality and can be
7077 : : used to replace STMT_INFO. It returns the last statement in the sequence
7078 : : and adds any earlier statements to STMT_INFO's STMT_VINFO_PATTERN_DEF_SEQ.
7079 : : PATTERN_RECOG_FUNC also sets *TYPE_OUT to the vector type of the final
7080 : : statement, having first checked that the target supports the new operation
7081 : : in that type.
7082 : :
7083 : : This function also does some bookkeeping, as explained in the documentation
7084 : : for vect_recog_pattern. */
7085 : :
7086 : : static void
7087 : 967639477 : vect_pattern_recog_1 (vec_info *vinfo,
7088 : : const vect_recog_func &recog_func, stmt_vec_info stmt_info)
7089 : : {
7090 : 967639477 : gimple *pattern_stmt;
7091 : 967639477 : tree pattern_vectype;
7092 : :
7093 : : /* If this statement has already been replaced with pattern statements,
7094 : : leave the original statement alone, since the first match wins.
7095 : : Instead try to match against the definition statements that feed
7096 : : the main pattern statement. */
7097 : 967639477 : if (STMT_VINFO_IN_PATTERN_P (stmt_info))
7098 : : {
7099 : 11535676 : gimple_stmt_iterator gsi;
7100 : 11535676 : for (gsi = gsi_start (STMT_VINFO_PATTERN_DEF_SEQ (stmt_info));
7101 : 28269425 : !gsi_end_p (gsi); gsi_next (&gsi))
7102 : 16733749 : vect_pattern_recog_1 (vinfo, recog_func,
7103 : : vinfo->lookup_stmt (gsi_stmt (gsi)));
7104 : : return;
7105 : : }
7106 : :
7107 : 956103801 : gcc_assert (!STMT_VINFO_PATTERN_DEF_SEQ (stmt_info));
7108 : 956103801 : pattern_stmt = recog_func.fn (vinfo, stmt_info, &pattern_vectype);
7109 : 956103801 : if (!pattern_stmt)
7110 : : {
7111 : : /* Clear any half-formed pattern definition sequence. */
7112 : 955277136 : STMT_VINFO_PATTERN_DEF_SEQ (stmt_info) = NULL;
7113 : 955277136 : return;
7114 : : }
7115 : :
7116 : : /* Found a vectorizable pattern. */
7117 : 826665 : if (dump_enabled_p ())
7118 : 18798 : dump_printf_loc (MSG_NOTE, vect_location,
7119 : : "%s pattern recognized: %G",
7120 : 18798 : recog_func.name, pattern_stmt);
7121 : :
7122 : : /* Mark the stmts that are involved in the pattern. */
7123 : 826665 : vect_mark_pattern_stmts (vinfo, stmt_info, pattern_stmt, pattern_vectype);
7124 : : }
7125 : :
7126 : :
7127 : : /* Function vect_pattern_recog
7128 : :
7129 : : Input:
7130 : : LOOP_VINFO - a struct_loop_info of a loop in which we want to look for
7131 : : computation idioms.
7132 : :
7133 : : Output - for each computation idiom that is detected we create a new stmt
7134 : : that provides the same functionality and that can be vectorized. We
7135 : : also record some information in the struct_stmt_info of the relevant
7136 : : stmts, as explained below:
7137 : :
7138 : : At the entry to this function we have the following stmts, with the
7139 : : following initial value in the STMT_VINFO fields:
7140 : :
7141 : : stmt in_pattern_p related_stmt vec_stmt
7142 : : S1: a_i = .... - - -
7143 : : S2: a_2 = ..use(a_i).. - - -
7144 : : S3: a_1 = ..use(a_2).. - - -
7145 : : S4: a_0 = ..use(a_1).. - - -
7146 : : S5: ... = ..use(a_0).. - - -
7147 : :
7148 : : Say the sequence {S1,S2,S3,S4} was detected as a pattern that can be
7149 : : represented by a single stmt. We then:
7150 : : - create a new stmt S6 equivalent to the pattern (the stmt is not
7151 : : inserted into the code)
7152 : : - fill in the STMT_VINFO fields as follows:
7153 : :
7154 : : in_pattern_p related_stmt vec_stmt
7155 : : S1: a_i = .... - - -
7156 : : S2: a_2 = ..use(a_i).. - - -
7157 : : S3: a_1 = ..use(a_2).. - - -
7158 : : S4: a_0 = ..use(a_1).. true S6 -
7159 : : '---> S6: a_new = .... - S4 -
7160 : : S5: ... = ..use(a_0).. - - -
7161 : :
7162 : : (the last stmt in the pattern (S4) and the new pattern stmt (S6) point
7163 : : to each other through the RELATED_STMT field).
7164 : :
7165 : : S6 will be marked as relevant in vect_mark_stmts_to_be_vectorized instead
7166 : : of S4 because it will replace all its uses. Stmts {S1,S2,S3} will
7167 : : remain irrelevant unless used by stmts other than S4.
7168 : :
7169 : : If vectorization succeeds, vect_transform_stmt will skip over {S1,S2,S3}
7170 : : (because they are marked as irrelevant). It will vectorize S6, and record
7171 : : a pointer to the new vector stmt VS6 from S6 (as usual).
7172 : : S4 will be skipped, and S5 will be vectorized as usual:
7173 : :
7174 : : in_pattern_p related_stmt vec_stmt
7175 : : S1: a_i = .... - - -
7176 : : S2: a_2 = ..use(a_i).. - - -
7177 : : S3: a_1 = ..use(a_2).. - - -
7178 : : > VS6: va_new = .... - - -
7179 : : S4: a_0 = ..use(a_1).. true S6 VS6
7180 : : '---> S6: a_new = .... - S4 VS6
7181 : : > VS5: ... = ..vuse(va_new).. - - -
7182 : : S5: ... = ..use(a_0).. - - -
7183 : :
7184 : : DCE could then get rid of {S1,S2,S3,S4,S5} (if their defs are not used
7185 : : elsewhere), and we'll end up with:
7186 : :
7187 : : VS6: va_new = ....
7188 : : VS5: ... = ..vuse(va_new)..
7189 : :
7190 : : In case of more than one pattern statements, e.g., widen-mult with
7191 : : intermediate type:
7192 : :
7193 : : S1 a_t = ;
7194 : : S2 a_T = (TYPE) a_t;
7195 : : '--> S3: a_it = (interm_type) a_t;
7196 : : S4 prod_T = a_T * CONST;
7197 : : '--> S5: prod_T' = a_it w* CONST;
7198 : :
7199 : : there may be other users of a_T outside the pattern. In that case S2 will
7200 : : be marked as relevant (as well as S3), and both S2 and S3 will be analyzed
7201 : : and vectorized. The vector stmt VS2 will be recorded in S2, and VS3 will
7202 : : be recorded in S3. */
7203 : :
7204 : : void
7205 : 955241 : vect_pattern_recog (vec_info *vinfo)
7206 : : {
7207 : 955241 : basic_block *bbs = vinfo->bbs;
7208 : 955241 : unsigned int nbbs = vinfo->nbbs;
7209 : :
7210 : 955241 : vect_determine_precisions (vinfo);
7211 : :
7212 : 955241 : DUMP_VECT_SCOPE ("vect_pattern_recog");
7213 : :
7214 : : /* Scan through the stmts in the region, applying the pattern recognition
7215 : : functions starting at each stmt visited. */
7216 : 12588507 : for (unsigned i = 0; i < nbbs; i++)
7217 : : {
7218 : 11633266 : basic_block bb = bbs[i];
7219 : :
7220 : 112249449 : for (auto si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
7221 : : {
7222 : 88982917 : stmt_vec_info stmt_info = vinfo->lookup_stmt (gsi_stmt (si));
7223 : :
7224 : 88982917 : if (!stmt_info || !STMT_VINFO_VECTORIZABLE (stmt_info))
7225 : 59267113 : continue;
7226 : :
7227 : : /* Scan over all generic vect_recog_xxx_pattern functions. */
7228 : 980621532 : for (const auto &func_ptr : vect_vect_recog_func_ptrs)
7229 : 950905728 : vect_pattern_recog_1 (vinfo, func_ptr,
7230 : : stmt_info);
7231 : : }
7232 : : }
7233 : :
7234 : : /* After this no more add_stmt calls are allowed. */
7235 : 955241 : vinfo->stmt_vec_info_ro = true;
7236 : 955241 : }
7237 : :
7238 : : /* Build a GIMPLE_ASSIGN or GIMPLE_CALL with the tree_code,
7239 : : or internal_fn contained in ch, respectively. */
7240 : : gimple *
7241 : 135219 : vect_gimple_build (tree lhs, code_helper ch, tree op0, tree op1)
7242 : : {
7243 : 135219 : gcc_assert (op0 != NULL_TREE);
7244 : 135219 : if (ch.is_tree_code ())
7245 : 135219 : return gimple_build_assign (lhs, (tree_code) ch, op0, op1);
7246 : :
7247 : 0 : gcc_assert (ch.is_internal_fn ());
7248 : 0 : gimple* stmt = gimple_build_call_internal (as_internal_fn ((combined_fn) ch),
7249 : : op1 == NULL_TREE ? 1 : 2,
7250 : : op0, op1);
7251 : 0 : gimple_call_set_lhs (stmt, lhs);
7252 : 0 : return stmt;
7253 : : }
|
