Branch data Line data Source code
1 : : /* Analysis Utilities for Loop Vectorization.
2 : : Copyright (C) 2006-2024 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 : 10471628 : vect_get_range_info (tree var, wide_int *min_value, wide_int *max_value)
70 : : {
71 : 10471628 : int_range_max vr;
72 : 10471628 : tree vr_min, vr_max;
73 : 20943256 : get_range_query (cfun)->range_of_expr (vr, var);
74 : 10471628 : if (vr.undefined_p ())
75 : 27 : vr.set_varying (TREE_TYPE (var));
76 : 10471628 : value_range_kind vr_type = get_legacy_range (vr, vr_min, vr_max);
77 : 10471628 : *min_value = wi::to_wide (vr_min);
78 : 10471628 : *max_value = wi::to_wide (vr_max);
79 : 10471628 : wide_int nonzero = get_nonzero_bits (var);
80 : 10471628 : signop sgn = TYPE_SIGN (TREE_TYPE (var));
81 : 10471628 : if (intersect_range_with_nonzero_bits (vr_type, min_value, max_value,
82 : : nonzero, sgn) == VR_RANGE)
83 : : {
84 : 4784649 : if (dump_enabled_p ())
85 : : {
86 : 69237 : dump_generic_expr_loc (MSG_NOTE, vect_location, TDF_SLIM, var);
87 : 69237 : dump_printf (MSG_NOTE, " has range [");
88 : 69237 : dump_hex (MSG_NOTE, *min_value);
89 : 69237 : dump_printf (MSG_NOTE, ", ");
90 : 69237 : dump_hex (MSG_NOTE, *max_value);
91 : 69237 : dump_printf (MSG_NOTE, "]\n");
92 : : }
93 : 4784649 : return true;
94 : : }
95 : : else
96 : : {
97 : 5686979 : if (dump_enabled_p ())
98 : : {
99 : 84237 : dump_generic_expr_loc (MSG_NOTE, vect_location, TDF_SLIM, var);
100 : 84237 : dump_printf (MSG_NOTE, " has no range info\n");
101 : : }
102 : 5686979 : return false;
103 : : }
104 : 10471628 : }
105 : :
106 : : /* Report that we've found an instance of pattern PATTERN in
107 : : statement STMT. */
108 : :
109 : : static void
110 : 932780 : vect_pattern_detected (const char *name, gimple *stmt)
111 : : {
112 : 932780 : if (dump_enabled_p ())
113 : 25449 : dump_printf_loc (MSG_NOTE, vect_location, "%s: detected: %G", name, stmt);
114 : 932780 : }
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 : 1713251 : vect_init_pattern_stmt (vec_info *vinfo, gimple *pattern_stmt,
122 : : stmt_vec_info orig_stmt_info, tree vectype)
123 : : {
124 : 1713251 : stmt_vec_info pattern_stmt_info = vinfo->lookup_stmt (pattern_stmt);
125 : 1713251 : if (pattern_stmt_info == NULL)
126 : 1022827 : pattern_stmt_info = vinfo->add_stmt (pattern_stmt);
127 : 1713251 : gimple_set_bb (pattern_stmt, gimple_bb (orig_stmt_info->stmt));
128 : :
129 : 1713251 : pattern_stmt_info->pattern_stmt_p = true;
130 : 1713251 : STMT_VINFO_RELATED_STMT (pattern_stmt_info) = orig_stmt_info;
131 : 1713251 : STMT_VINFO_DEF_TYPE (pattern_stmt_info)
132 : 1713251 : = STMT_VINFO_DEF_TYPE (orig_stmt_info);
133 : 1713251 : STMT_VINFO_TYPE (pattern_stmt_info) = STMT_VINFO_TYPE (orig_stmt_info);
134 : 1713251 : if (!STMT_VINFO_VECTYPE (pattern_stmt_info))
135 : : {
136 : 1802817 : 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 : 1031255 : STMT_VINFO_VECTYPE (pattern_stmt_info) = vectype;
141 : 1031255 : pattern_stmt_info->mask_precision = orig_stmt_info->mask_precision;
142 : : }
143 : 1713251 : 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 : 733096 : vect_set_pattern_stmt (vec_info *vinfo, gimple *pattern_stmt,
152 : : stmt_vec_info orig_stmt_info, tree vectype)
153 : : {
154 : 733096 : STMT_VINFO_IN_PATTERN_P (orig_stmt_info) = true;
155 : 733096 : STMT_VINFO_RELATED_STMT (orig_stmt_info)
156 : 0 : = vect_init_pattern_stmt (vinfo, pattern_stmt, orig_stmt_info, vectype);
157 : 703476 : }
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 : 938523 : 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 : 1513318 : gcc_assert (!scalar_type_for_mask
172 : : == (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)));
173 : 938523 : if (vectype)
174 : : {
175 : 681996 : stmt_vec_info new_stmt_info = vinfo->add_stmt (new_stmt);
176 : 681996 : STMT_VINFO_VECTYPE (new_stmt_info) = vectype;
177 : 681996 : if (scalar_type_for_mask)
178 : 363728 : new_stmt_info->mask_precision
179 : 727456 : = GET_MODE_BITSIZE (SCALAR_TYPE_MODE (scalar_type_for_mask));
180 : : }
181 : 938523 : gimple_seq_add_stmt_without_update (&STMT_VINFO_PATTERN_DEF_SEQ (stmt_info),
182 : : new_stmt);
183 : 938523 : }
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 : 1338 : append_inv_pattern_def_seq (vec_info *vinfo, gimple *new_stmt)
192 : : {
193 : 1338 : 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 : 4765 : vect_get_external_def_edge (vec_info *vinfo, tree var)
204 : : {
205 : 4765 : edge e = NULL;
206 : 4765 : if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
207 : : {
208 : 495 : e = loop_preheader_edge (loop_vinfo->loop);
209 : 495 : if (!SSA_NAME_IS_DEFAULT_DEF (var))
210 : : {
211 : 410 : basic_block bb = gimple_bb (SSA_NAME_DEF_STMT (var));
212 : 410 : if (bb == NULL
213 : 410 : || !dominated_by_p (CDI_DOMINATORS, e->dest, bb))
214 : : e = NULL;
215 : : }
216 : : }
217 : 4765 : 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 : 2059 : 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 : 2059 : tree vecitype = get_vectype_for_scalar_type (vinfo, itype);
242 : 2059 : if (!vecitype)
243 : : return false;
244 : :
245 : 2059 : tree vecotype = get_vectype_for_scalar_type (vinfo, otype);
246 : 2059 : if (!vecotype)
247 : : return false;
248 : :
249 : 1949 : optab optab = optab_for_tree_code (code, vecitype, subtype);
250 : 1949 : if (!optab)
251 : : return false;
252 : :
253 : 1949 : insn_code icode = optab_handler (optab, TYPE_MODE (vecitype));
254 : 1949 : if (icode == CODE_FOR_nothing
255 : 1949 : || insn_data[icode].operand[0].mode != TYPE_MODE (vecotype))
256 : 1723 : 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 : 917 : 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 : 917 : tree vecitype = get_vectype_for_scalar_type (vinfo, itype);
280 : 917 : tree vecotype = get_vectype_for_scalar_type (vinfo, otype);
281 : 917 : if (!vecitype || !vecotype)
282 : : return false;
283 : :
284 : 817 : if (!directly_supported_p (code, vecotype, vecitype, subtype))
285 : : return false;
286 : :
287 : 425 : *vecotype_out = vecotype;
288 : 425 : if (vecitype_out)
289 : 425 : *vecitype_out = vecitype;
290 : : return true;
291 : : }
292 : :
293 : : /* Round bit precision PRECISION up to a full element. */
294 : :
295 : : static unsigned int
296 : 2493428 : vect_element_precision (unsigned int precision)
297 : : {
298 : 0 : precision = 1 << ceil_log2 (precision);
299 : 3672899 : 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 : 1429379 : vect_get_internal_def (vec_info *vinfo, tree op)
307 : : {
308 : 1429379 : stmt_vec_info def_stmt_info = vinfo->lookup_def (op);
309 : 1429379 : if (def_stmt_info
310 : 1369594 : && STMT_VINFO_DEF_TYPE (def_stmt_info) == vect_internal_def)
311 : 1360489 : 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 : 260880313 : inline vect_unpromoted_value::vect_unpromoted_value ()
339 : 260880313 : : op (NULL_TREE),
340 : 260880313 : type (NULL_TREE),
341 : 260880313 : dt (vect_uninitialized_def),
342 : 2349163 : 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 : 9352659 : vect_unpromoted_value::set_op (tree op_in, vect_def_type dt_in,
351 : : stmt_vec_info caster_in)
352 : : {
353 : 9352659 : op = op_in;
354 : 9352659 : type = TREE_TYPE (op);
355 : 9352659 : dt = dt_in;
356 : 9352659 : caster = caster_in;
357 : 9352659 : }
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 : 6979785 : vect_look_through_possible_promotion (vec_info *vinfo, tree op,
393 : : vect_unpromoted_value *unprom,
394 : : bool *single_use_p = NULL)
395 : : {
396 : 6979785 : tree op_type = TREE_TYPE (op);
397 : 6979785 : if (!INTEGRAL_TYPE_P (op_type))
398 : : return NULL_TREE;
399 : :
400 : 6955506 : tree res = NULL_TREE;
401 : 6955506 : unsigned int orig_precision = TYPE_PRECISION (op_type);
402 : 6955506 : unsigned int min_precision = orig_precision;
403 : 6955506 : stmt_vec_info caster = NULL;
404 : 8368478 : while (TREE_CODE (op) == SSA_NAME && INTEGRAL_TYPE_P (op_type))
405 : : {
406 : : /* See whether OP is simple enough to vectorize. */
407 : 8191775 : stmt_vec_info def_stmt_info;
408 : 8191775 : gimple *def_stmt;
409 : 8191775 : vect_def_type dt;
410 : 8191775 : 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 : 8186991 : 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 : 8065057 : if (!res
427 : 1205118 : || TYPE_PRECISION (unprom->type) == orig_precision
428 : 38112 : || TYPE_SIGN (unprom->type) == TYPE_SIGN (op_type)
429 : 8099737 : || (TYPE_UNSIGNED (op_type)
430 : 22682 : && TYPE_PRECISION (op_type) < TYPE_PRECISION (unprom->type)))
431 : : {
432 : 8031040 : unprom->set_op (op, dt, caster);
433 : 8031040 : 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 : 34017 : else if (TYPE_PRECISION (op_type)
438 : 34017 : != 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 : 8186950 : if (!def_stmt)
448 : : break;
449 : 8014943 : caster = def_stmt_info;
450 : :
451 : : /* Ignore pattern statements, since we don't link uses for them. */
452 : 8014943 : if (caster
453 : 8014943 : && single_use_p
454 : 1479579 : && !STMT_VINFO_RELATED_STMT (caster)
455 : 9363653 : && !has_single_use (res))
456 : 898439 : *single_use_p = false;
457 : :
458 : 14793746 : gassign *assign = dyn_cast <gassign *> (def_stmt);
459 : 5180137 : if (!assign || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (def_stmt)))
460 : : break;
461 : :
462 : : /* Continue with the input to the cast. */
463 : 1412972 : op = gimple_assign_rhs1 (def_stmt);
464 : 1412972 : 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 : 268479 : 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 : 268479 : unsigned int precision;
486 : 268479 : if (shift_p)
487 : : {
488 : 13004 : if (!wi::leu_p (wi::to_widest (op), TYPE_PRECISION (type) / 2))
489 : : return false;
490 : 10453 : precision = TREE_INT_CST_LOW (op);
491 : : }
492 : : else
493 : : {
494 : 255475 : precision = wi::min_precision (wi::to_widest (op),
495 : 255475 : TYPE_SIGN (*common_type));
496 : 255475 : 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 : 253268 : precision = vect_element_precision (precision);
503 : 253268 : if (TYPE_PRECISION (*common_type) < precision)
504 : 6146 : *common_type = build_nonstandard_integer_type
505 : 6146 : (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 : 38898 : vect_joust_widened_type (tree type, tree new_type, tree *common_type)
514 : : {
515 : 38898 : 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 : 6911 : if ((TYPE_PRECISION (new_type) < TYPE_PRECISION (*common_type))
520 : 6911 : && (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 : 6383 : if (TYPE_PRECISION (*common_type) < TYPE_PRECISION (new_type)
525 : 6383 : && (TYPE_UNSIGNED (*common_type) || !TYPE_UNSIGNED (new_type)))
526 : : {
527 : 347 : *common_type = new_type;
528 : 347 : 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 : 6036 : unsigned int precision = MAX (TYPE_PRECISION (*common_type),
535 : : TYPE_PRECISION (new_type));
536 : 6036 : precision *= 2;
537 : :
538 : 6036 : if (precision * 2 > TYPE_PRECISION (type))
539 : : return false;
540 : :
541 : 16 : *common_type = build_nonstandard_integer_type (precision, false);
542 : 16 : 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 : 110225839 : 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 : 110225839 : gimple* stmt = stmt_info->stmt;
572 : 110225839 : if (!(is_gimple_assign (stmt) || is_gimple_call (stmt)))
573 : : return 0;
574 : :
575 : 89044733 : code_helper rhs_code;
576 : 89044733 : if (is_gimple_assign (stmt))
577 : 75893260 : rhs_code = gimple_assign_rhs_code (stmt);
578 : 13151473 : else if (is_gimple_call (stmt))
579 : 13151473 : rhs_code = gimple_call_combined_fn (stmt);
580 : : else
581 : : return 0;
582 : :
583 : 89044733 : if (rhs_code != code
584 : 89044733 : && rhs_code != widened_code)
585 : : return 0;
586 : :
587 : 5469382 : tree lhs = gimple_get_lhs (stmt);
588 : 5469382 : tree type = TREE_TYPE (lhs);
589 : 5469382 : if (!INTEGRAL_TYPE_P (type))
590 : : return 0;
591 : :
592 : : /* Assume that both operands will be leaf operands. */
593 : 4926732 : max_nops -= 2;
594 : :
595 : : /* Check the operands. */
596 : 4926732 : unsigned int next_op = 0;
597 : 5596850 : for (unsigned int i = 0; i < 2; ++i)
598 : : {
599 : 5310338 : vect_unpromoted_value *this_unprom = &unprom[next_op];
600 : 5310338 : unsigned int nops = 1;
601 : 5310338 : tree op = gimple_arg (stmt, i);
602 : 5310338 : 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 : 268479 : this_unprom->set_op (op, vect_constant_def);
607 : 268479 : 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 : 5041859 : if (shift_p && i == 1)
614 : : return 0;
615 : :
616 : 5034400 : 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 : 1967 : if (TREE_CODE (op) != SSA_NAME
622 : 1967 : || !INTEGRAL_TYPE_P (TREE_TYPE (op)))
623 : 0 : return 0;
624 : :
625 : 1967 : stmt_vec_info def_stmt_info;
626 : 1967 : gimple *def_stmt;
627 : 1967 : vect_def_type dt;
628 : 1967 : if (!vect_is_simple_use (op, vinfo, &dt, &def_stmt_info,
629 : : &def_stmt))
630 : : return 0;
631 : 1967 : this_unprom->set_op (op, dt, NULL);
632 : : }
633 : 5032433 : else if (!vect_look_through_possible_promotion (vinfo, op,
634 : : this_unprom))
635 : : return 0;
636 : :
637 : 4940546 : 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 : 4513830 : if (rhs_code != code
643 : 4513830 : || max_nops == 0
644 : 4514256 : || 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 : 426 : max_nops += 1;
650 : :
651 : : /* Recursively process the definition of the operand. */
652 : 426 : stmt_vec_info def_stmt_info
653 : 426 : = vect_get_internal_def (vinfo, this_unprom->op);
654 : :
655 : 426 : 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 : 426 : if (nops == 0)
660 : : return 0;
661 : :
662 : 287 : max_nops -= nops;
663 : : }
664 : : else
665 : : {
666 : : /* Make sure that the operand is narrower than the result. */
667 : 426716 : if (TYPE_PRECISION (this_unprom->type) * 2
668 : 426716 : > TYPE_PRECISION (type))
669 : : return 0;
670 : :
671 : : /* Update COMMON_TYPE for the new operand. */
672 : 422373 : if (i == 0)
673 : 383475 : *common_type = this_unprom->type;
674 : 38898 : else if (!vect_joust_widened_type (type, this_unprom->type,
675 : : common_type))
676 : : {
677 : 6020 : 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 : 670118 : 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 : 1469446 : 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 : 30699 : vect_split_statement (vec_info *vinfo, stmt_vec_info stmt2_info, tree new_rhs,
711 : : gimple *stmt1, tree vectype)
712 : : {
713 : 30699 : 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 : 1079 : stmt_vec_info orig_stmt2_info = STMT_VINFO_RELATED_STMT (stmt2_info);
718 : 1079 : vect_init_pattern_stmt (vinfo, stmt1, orig_stmt2_info, vectype);
719 : :
720 : 1079 : 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 : 1079 : gimple_assign_set_rhs1 (stmt2_info->stmt, new_rhs);
728 : :
729 : 1079 : 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 : 1079 : gimple_seq *def_seq = &STMT_VINFO_PATTERN_DEF_SEQ (orig_stmt2_info);
737 : 1079 : 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 : 1079 : 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 : 0 : gimple_stmt_iterator gsi = gsi_for_stmt (stmt2_info->stmt, def_seq);
746 : 0 : gsi_insert_before_without_update (&gsi, stmt1, GSI_SAME_STMT);
747 : : }
748 : 1079 : 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 : 29620 : gcc_assert (!STMT_VINFO_RELATED_STMT (stmt2_info));
755 : 29620 : tree lhs_type = TREE_TYPE (gimple_get_lhs (stmt2_info->stmt));
756 : 29620 : tree lhs_vectype = get_vectype_for_scalar_type (vinfo, lhs_type);
757 : 29620 : if (!lhs_vectype)
758 : : return false;
759 : :
760 : 29620 : if (dump_enabled_p ())
761 : 2158 : 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 : 29620 : gimple_seq *def_seq = &STMT_VINFO_PATTERN_DEF_SEQ (stmt2_info);
766 : 29620 : vect_init_pattern_stmt (vinfo, stmt1, stmt2_info, vectype);
767 : 29620 : gimple_seq_add_stmt_without_update (def_seq, stmt1);
768 : :
769 : : /* Build the second of the two pattern statements. */
770 : 29620 : tree new_lhs = vect_recog_temp_ssa_var (lhs_type, NULL);
771 : 29620 : gassign *new_stmt2 = gimple_build_assign (new_lhs, NOP_EXPR, new_rhs);
772 : 29620 : vect_set_pattern_stmt (vinfo, new_stmt2, stmt2_info, lhs_vectype);
773 : :
774 : 29620 : if (dump_enabled_p ())
775 : : {
776 : 2158 : dump_printf_loc (MSG_NOTE, vect_location,
777 : : "into pattern statements: %G", stmt1);
778 : 2158 : dump_printf_loc (MSG_NOTE, vect_location, "and: %G",
779 : : (gimple *) new_stmt2);
780 : : }
781 : :
782 : 29620 : 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 : 18718877 : 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 : 18718877 : 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 : 18718877 : enum tree_code code = gimple_assign_rhs_code (abs_stmt);
810 : 18718877 : if (code != ABS_EXPR && code != ABSU_EXPR)
811 : : return false;
812 : :
813 : 21922 : tree abs_oprnd = gimple_assign_rhs1 (abs_stmt);
814 : 21922 : tree abs_type = TREE_TYPE (abs_oprnd);
815 : 21922 : if (!abs_oprnd)
816 : : return false;
817 : 16069 : if (!ANY_INTEGRAL_TYPE_P (abs_type)
818 : 6066 : || TYPE_OVERFLOW_WRAPS (abs_type)
819 : 27863 : || 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 : 5941 : vect_unpromoted_value unprom_diff;
828 : 5941 : abs_oprnd = vect_look_through_possible_promotion (vinfo, abs_oprnd,
829 : : &unprom_diff);
830 : 5941 : if (!abs_oprnd)
831 : : return false;
832 : 5728 : if (TYPE_PRECISION (unprom_diff.type) != TYPE_PRECISION (abs_type)
833 : 5728 : && 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 : 5728 : stmt_vec_info diff_stmt_vinfo = vect_get_internal_def (vinfo, abs_oprnd);
838 : 5728 : if (!diff_stmt_vinfo)
839 : : return false;
840 : :
841 : 5000 : gassign *diff = dyn_cast <gassign *> (STMT_VINFO_STMT (diff_stmt_vinfo));
842 : 5000 : if (diff_stmt && diff
843 : 4077 : && gimple_assign_rhs_code (diff) == MINUS_EXPR
844 : 7044 : && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (abs_oprnd)))
845 : 1052 : *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 : 5000 : if (vect_widened_op_tree (vinfo, diff_stmt_vinfo,
850 : : 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 : 407963 : 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 : 407963 : 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 : 407963 : if (types_compatible_p (type, TREE_TYPE (unprom->op)))
881 : 142440 : return unprom->op;
882 : :
883 : : /* Allow the caller to create constant vect_unpromoted_values. */
884 : 265523 : if (TREE_CODE (unprom->op) == INTEGER_CST)
885 : 163770 : return wide_int_to_tree (type, wi::to_widest (unprom->op));
886 : :
887 : 101753 : tree input = unprom->op;
888 : 101753 : if (unprom->caster)
889 : : {
890 : 52062 : tree lhs = gimple_get_lhs (unprom->caster->stmt);
891 : 52062 : 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 : 52062 : 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 : 50329 : else if (TYPE_PRECISION (lhs_type) > TYPE_PRECISION (type)
901 : 50329 : && 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 : 30699 : tree midtype = build_nonstandard_integer_type
915 : 30699 : (TYPE_PRECISION (type), TYPE_UNSIGNED (unprom->type));
916 : 30699 : tree vec_midtype = get_vectype_for_scalar_type (vinfo, midtype);
917 : 30699 : if (vec_midtype)
918 : : {
919 : 30699 : input = vect_recog_temp_ssa_var (midtype, NULL);
920 : 30699 : gassign *new_stmt = gimple_build_assign (input, NOP_EXPR,
921 : : unprom->op);
922 : 30699 : 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 : 52062 : if (types_compatible_p (type, TREE_TYPE (input)))
931 : : return input;
932 : : }
933 : :
934 : : /* We need a new conversion statement. */
935 : 78663 : tree new_op = vect_recog_temp_ssa_var (type, NULL);
936 : 78663 : 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 : 78663 : if (input == unprom->op && unprom->dt == vect_external_def)
941 : 4752 : if (edge e = vect_get_external_def_edge (vinfo, input))
942 : : {
943 : 482 : basic_block new_bb = gsi_insert_on_edge_immediate (e, new_stmt);
944 : 482 : gcc_assert (!new_bb);
945 : : return new_op;
946 : : }
947 : :
948 : : /* As a (common) last resort, add the statement to the pattern itself. */
949 : 78181 : append_pattern_def_seq (vinfo, stmt_info, new_stmt, vectype);
950 : 78181 : 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 : 206767 : 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 : 614471 : for (unsigned int i = 0; i < n; ++i)
964 : : {
965 : : unsigned int j;
966 : 608400 : for (j = 0; j < i; ++j)
967 : 200937 : if (unprom[j].op == unprom[i].op)
968 : : break;
969 : :
970 : 407704 : if (j < i)
971 : 241 : result[i] = result[j];
972 : : else
973 : 407463 : result[i] = vect_convert_input (vinfo, stmt_info,
974 : 407463 : type, &unprom[i], vectype, subtype);
975 : : }
976 : 206767 : }
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 : 230510 : vect_convert_output (vec_info *vinfo, stmt_vec_info stmt_info, tree type,
987 : : gimple *pattern_stmt, tree vecitype)
988 : : {
989 : 230510 : tree lhs = gimple_get_lhs (pattern_stmt);
990 : 230510 : if (!types_compatible_p (type, TREE_TYPE (lhs)))
991 : : {
992 : 206247 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt, vecitype);
993 : 206247 : tree cast_var = vect_recog_temp_ssa_var (type, NULL);
994 : 206247 : pattern_stmt = gimple_build_assign (cast_var, NOP_EXPR, lhs);
995 : : }
996 : 230510 : 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 : 82332340 : 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 : 82332340 : loop_vec_info loop_info = dyn_cast <loop_vec_info> (vinfo);
1013 : 9437101 : if (!loop_info)
1014 : : return false;
1015 : :
1016 : 9437101 : gassign *assign = dyn_cast <gassign *> (stmt_info->stmt);
1017 : 10282204 : 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 : 1994614 : class loop *loop = LOOP_VINFO_LOOP (loop_info);
1023 : 84229372 : if (loop && nested_in_vect_loop_p (loop, stmt_info))
1024 : : return false;
1025 : :
1026 : 1947133 : if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)
1027 : : {
1028 : 99117 : if (needs_fold_left_reduction_p (TREE_TYPE (gimple_assign_lhs (assign)),
1029 : : code))
1030 : : return false;
1031 : : }
1032 : 1848016 : else if (REDUC_GROUP_FIRST_ELEMENT (stmt_info) == NULL)
1033 : : return false;
1034 : :
1035 : 97582 : *op0_out = gimple_assign_rhs1 (assign);
1036 : 97582 : *op1_out = gimple_assign_rhs2 (assign);
1037 : 97582 : if (commutative_tree_code (code) && STMT_VINFO_REDUC_IDX (stmt_info) == 0)
1038 : 29276 : 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 : 27504562 : vect_recog_cond_expr_convert_pattern (vec_info *vinfo,
1095 : : stmt_vec_info stmt_vinfo, tree *type_out)
1096 : : {
1097 : 46303990 : gassign *last_stmt = dyn_cast <gassign *> (stmt_vinfo->stmt);
1098 : 18799486 : tree lhs, match[4], temp, type, new_lhs, op2;
1099 : 18799486 : gimple *cond_stmt;
1100 : 18799486 : gimple *pattern_stmt;
1101 : :
1102 : 18799486 : if (!last_stmt)
1103 : : return NULL;
1104 : :
1105 : 18799486 : lhs = gimple_assign_lhs (last_stmt);
1106 : :
1107 : : /* Find E = C cmp D ? (TYPE3) A ? (TYPE3) B;
1108 : : TYPE_PRECISION (A) == TYPE_PRECISION (C). */
1109 : 18799486 : if (!gimple_cond_expr_convert_p (lhs, &match[0], NULL))
1110 : : return NULL;
1111 : :
1112 : 58 : vect_pattern_detected ("vect_recog_cond_expr_convert_pattern", last_stmt);
1113 : :
1114 : 58 : op2 = match[2];
1115 : 58 : type = TREE_TYPE (match[1]);
1116 : 58 : if (TYPE_SIGN (type) != TYPE_SIGN (TREE_TYPE (match[2])))
1117 : : {
1118 : 24 : op2 = vect_recog_temp_ssa_var (type, NULL);
1119 : 24 : gimple* nop_stmt = gimple_build_assign (op2, NOP_EXPR, match[2]);
1120 : 24 : append_pattern_def_seq (vinfo, stmt_vinfo, nop_stmt,
1121 : : get_vectype_for_scalar_type (vinfo, type));
1122 : : }
1123 : :
1124 : 58 : temp = vect_recog_temp_ssa_var (type, NULL);
1125 : 58 : cond_stmt = gimple_build_assign (temp, build3 (COND_EXPR, type, match[3],
1126 : : match[1], op2));
1127 : 58 : append_pattern_def_seq (vinfo, stmt_vinfo, cond_stmt,
1128 : : get_vectype_for_scalar_type (vinfo, type));
1129 : 58 : new_lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL);
1130 : 58 : pattern_stmt = gimple_build_assign (new_lhs, NOP_EXPR, temp);
1131 : 58 : *type_out = STMT_VINFO_VECTYPE (stmt_vinfo);
1132 : :
1133 : 58 : if (dump_enabled_p ())
1134 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
1135 : : "created pattern stmt: %G", pattern_stmt);
1136 : : return pattern_stmt;
1137 : : }
1138 : :
1139 : : /* Function vect_recog_dot_prod_pattern
1140 : :
1141 : : Try to find the following pattern:
1142 : :
1143 : : type1a x_t
1144 : : type1b y_t;
1145 : : TYPE1 prod;
1146 : : TYPE2 sum = init;
1147 : : loop:
1148 : : sum_0 = phi <init, sum_1>
1149 : : S1 x_t = ...
1150 : : S2 y_t = ...
1151 : : S3 x_T = (TYPE1) x_t;
1152 : : S4 y_T = (TYPE1) y_t;
1153 : : S5 prod = x_T * y_T;
1154 : : [S6 prod = (TYPE2) prod; #optional]
1155 : : S7 sum_1 = prod + sum_0;
1156 : :
1157 : : where 'TYPE1' is exactly double the size of type 'type1a' and 'type1b',
1158 : : the sign of 'TYPE1' must be one of 'type1a' or 'type1b' but the sign of
1159 : : 'type1a' and 'type1b' can differ.
1160 : :
1161 : : Input:
1162 : :
1163 : : * STMT_VINFO: The stmt from which the pattern search begins. In the
1164 : : example, when this function is called with S7, the pattern {S3,S4,S5,S6,S7}
1165 : : will be detected.
1166 : :
1167 : : Output:
1168 : :
1169 : : * TYPE_OUT: The type of the output of this pattern.
1170 : :
1171 : : * Return value: A new stmt that will be used to replace the sequence of
1172 : : stmts that constitute the pattern. In this case it will be:
1173 : : WIDEN_DOT_PRODUCT <x_t, y_t, sum_0>
1174 : :
1175 : : Note: The dot-prod idiom is a widening reduction pattern that is
1176 : : vectorized without preserving all the intermediate results. It
1177 : : produces only N/2 (widened) results (by summing up pairs of
1178 : : intermediate results) rather than all N results. Therefore, we
1179 : : cannot allow this pattern when we want to get all the results and in
1180 : : the correct order (as is the case when this computation is in an
1181 : : inner-loop nested in an outer-loop that us being vectorized). */
1182 : :
1183 : : static gimple *
1184 : 27444468 : vect_recog_dot_prod_pattern (vec_info *vinfo,
1185 : : stmt_vec_info stmt_vinfo, tree *type_out)
1186 : : {
1187 : 27444468 : tree oprnd0, oprnd1;
1188 : 27444468 : gimple *last_stmt = stmt_vinfo->stmt;
1189 : 27444468 : tree type, half_type;
1190 : 27444468 : gimple *pattern_stmt;
1191 : 27444468 : tree var;
1192 : :
1193 : : /* Look for the following pattern
1194 : : DX = (TYPE1) X;
1195 : : DY = (TYPE1) Y;
1196 : : DPROD = DX * DY;
1197 : : DDPROD = (TYPE2) DPROD;
1198 : : sum_1 = DDPROD + sum_0;
1199 : : In which
1200 : : - DX is double the size of X
1201 : : - DY is double the size of Y
1202 : : - DX, DY, DPROD all have the same type but the sign
1203 : : between X, Y and DPROD can differ.
1204 : : - sum is the same size of DPROD or bigger
1205 : : - sum has been recognized as a reduction variable.
1206 : :
1207 : : This is equivalent to:
1208 : : DPROD = X w* Y; #widen mult
1209 : : sum_1 = DPROD w+ sum_0; #widen summation
1210 : : or
1211 : : DPROD = X w* Y; #widen mult
1212 : : sum_1 = DPROD + sum_0; #summation
1213 : : */
1214 : :
1215 : : /* Starting from LAST_STMT, follow the defs of its uses in search
1216 : : of the above pattern. */
1217 : :
1218 : 27444468 : if (!vect_reassociating_reduction_p (vinfo, stmt_vinfo, PLUS_EXPR,
1219 : : &oprnd0, &oprnd1))
1220 : : return NULL;
1221 : :
1222 : 32886 : type = TREE_TYPE (gimple_get_lhs (last_stmt));
1223 : :
1224 : 32886 : vect_unpromoted_value unprom_mult;
1225 : 32886 : oprnd0 = vect_look_through_possible_promotion (vinfo, oprnd0, &unprom_mult);
1226 : :
1227 : : /* So far so good. Since last_stmt was detected as a (summation) reduction,
1228 : : we know that oprnd1 is the reduction variable (defined by a loop-header
1229 : : phi), and oprnd0 is an ssa-name defined by a stmt in the loop body.
1230 : : Left to check that oprnd0 is defined by a (widen_)mult_expr */
1231 : 32886 : if (!oprnd0)
1232 : : return NULL;
1233 : :
1234 : 24678 : stmt_vec_info mult_vinfo = vect_get_internal_def (vinfo, oprnd0);
1235 : 24678 : if (!mult_vinfo)
1236 : : return NULL;
1237 : :
1238 : : /* FORNOW. Can continue analyzing the def-use chain when this stmt in a phi
1239 : : inside the loop (in case we are analyzing an outer-loop). */
1240 : 71907 : vect_unpromoted_value unprom0[2];
1241 : 23969 : enum optab_subtype subtype = optab_vector;
1242 : 23969 : if (!vect_widened_op_tree (vinfo, mult_vinfo, MULT_EXPR, WIDEN_MULT_EXPR,
1243 : : false, 2, unprom0, &half_type, &subtype))
1244 : : return NULL;
1245 : :
1246 : : /* If there are two widening operations, make sure they agree on the sign
1247 : : of the extension. The result of an optab_vector_mixed_sign operation
1248 : : is signed; otherwise, the result has the same sign as the operands. */
1249 : 951 : if (TYPE_PRECISION (unprom_mult.type) != TYPE_PRECISION (type)
1250 : 1567 : && (subtype == optab_vector_mixed_sign
1251 : 616 : ? TYPE_UNSIGNED (unprom_mult.type)
1252 : 428 : : TYPE_SIGN (unprom_mult.type) != TYPE_SIGN (half_type)))
1253 : : return NULL;
1254 : :
1255 : 843 : vect_pattern_detected ("vect_recog_dot_prod_pattern", last_stmt);
1256 : :
1257 : : /* If the inputs have mixed signs, canonicalize on using the signed
1258 : : input type for analysis. This also helps when emulating mixed-sign
1259 : : operations using signed operations. */
1260 : 843 : if (subtype == optab_vector_mixed_sign)
1261 : 150 : half_type = signed_type_for (half_type);
1262 : :
1263 : 843 : tree half_vectype;
1264 : 843 : if (!vect_supportable_conv_optab_p (vinfo, type, DOT_PROD_EXPR, half_type,
1265 : : type_out, &half_vectype, subtype))
1266 : : {
1267 : : /* We can emulate a mixed-sign dot-product using a sequence of
1268 : : signed dot-products; see vect_emulate_mixed_dot_prod for details. */
1269 : 424 : if (subtype != optab_vector_mixed_sign
1270 : 424 : || !vect_supportable_conv_optab_p (vinfo, signed_type_for (type),
1271 : : DOT_PROD_EXPR, half_type,
1272 : : type_out, &half_vectype,
1273 : : optab_vector))
1274 : 418 : return NULL;
1275 : :
1276 : 6 : *type_out = signed_or_unsigned_type_for (TYPE_UNSIGNED (type),
1277 : : *type_out);
1278 : : }
1279 : :
1280 : : /* Get the inputs in the appropriate types. */
1281 : 425 : tree mult_oprnd[2];
1282 : 425 : vect_convert_inputs (vinfo, stmt_vinfo, 2, mult_oprnd, half_type,
1283 : : unprom0, half_vectype, subtype);
1284 : :
1285 : 425 : var = vect_recog_temp_ssa_var (type, NULL);
1286 : 425 : pattern_stmt = gimple_build_assign (var, DOT_PROD_EXPR,
1287 : : mult_oprnd[0], mult_oprnd[1], oprnd1);
1288 : :
1289 : 425 : return pattern_stmt;
1290 : : }
1291 : :
1292 : :
1293 : : /* Function vect_recog_sad_pattern
1294 : :
1295 : : Try to find the following Sum of Absolute Difference (SAD) pattern:
1296 : :
1297 : : type x_t, y_t;
1298 : : signed TYPE1 diff, abs_diff;
1299 : : TYPE2 sum = init;
1300 : : loop:
1301 : : sum_0 = phi <init, sum_1>
1302 : : S1 x_t = ...
1303 : : S2 y_t = ...
1304 : : S3 x_T = (TYPE1) x_t;
1305 : : S4 y_T = (TYPE1) y_t;
1306 : : S5 diff = x_T - y_T;
1307 : : S6 abs_diff = ABS_EXPR <diff>;
1308 : : [S7 abs_diff = (TYPE2) abs_diff; #optional]
1309 : : S8 sum_1 = abs_diff + sum_0;
1310 : :
1311 : : where 'TYPE1' is at least double the size of type 'type', and 'TYPE2' is the
1312 : : same size of 'TYPE1' or bigger. This is a special case of a reduction
1313 : : computation.
1314 : :
1315 : : Input:
1316 : :
1317 : : * STMT_VINFO: The stmt from which the pattern search begins. In the
1318 : : example, when this function is called with S8, the pattern
1319 : : {S3,S4,S5,S6,S7,S8} will be detected.
1320 : :
1321 : : Output:
1322 : :
1323 : : * TYPE_OUT: The type of the output of this pattern.
1324 : :
1325 : : * Return value: A new stmt that will be used to replace the sequence of
1326 : : stmts that constitute the pattern. In this case it will be:
1327 : : SAD_EXPR <x_t, y_t, sum_0>
1328 : : */
1329 : :
1330 : : static gimple *
1331 : 27444049 : vect_recog_sad_pattern (vec_info *vinfo,
1332 : : stmt_vec_info stmt_vinfo, tree *type_out)
1333 : : {
1334 : 27444049 : gimple *last_stmt = stmt_vinfo->stmt;
1335 : 27444049 : tree half_type;
1336 : :
1337 : : /* Look for the following pattern
1338 : : DX = (TYPE1) X;
1339 : : DY = (TYPE1) Y;
1340 : : DDIFF = DX - DY;
1341 : : DAD = ABS_EXPR <DDIFF>;
1342 : : DDPROD = (TYPE2) DPROD;
1343 : : sum_1 = DAD + sum_0;
1344 : : In which
1345 : : - DX is at least double the size of X
1346 : : - DY is at least double the size of Y
1347 : : - DX, DY, DDIFF, DAD all have the same type
1348 : : - sum is the same size of DAD or bigger
1349 : : - sum has been recognized as a reduction variable.
1350 : :
1351 : : This is equivalent to:
1352 : : DDIFF = X w- Y; #widen sub
1353 : : DAD = ABS_EXPR <DDIFF>;
1354 : : sum_1 = DAD w+ sum_0; #widen summation
1355 : : or
1356 : : DDIFF = X w- Y; #widen sub
1357 : : DAD = ABS_EXPR <DDIFF>;
1358 : : sum_1 = DAD + sum_0; #summation
1359 : : */
1360 : :
1361 : : /* Starting from LAST_STMT, follow the defs of its uses in search
1362 : : of the above pattern. */
1363 : :
1364 : 27444049 : tree plus_oprnd0, plus_oprnd1;
1365 : 27444049 : if (!vect_reassociating_reduction_p (vinfo, stmt_vinfo, PLUS_EXPR,
1366 : : &plus_oprnd0, &plus_oprnd1))
1367 : : return NULL;
1368 : :
1369 : 32461 : tree sum_type = TREE_TYPE (gimple_get_lhs (last_stmt));
1370 : :
1371 : : /* Any non-truncating sequence of conversions is OK here, since
1372 : : with a successful match, the result of the ABS(U) is known to fit
1373 : : within the nonnegative range of the result type. (It cannot be the
1374 : : negative of the minimum signed value due to the range of the widening
1375 : : MINUS_EXPR.) */
1376 : 32461 : vect_unpromoted_value unprom_abs;
1377 : 32461 : plus_oprnd0 = vect_look_through_possible_promotion (vinfo, plus_oprnd0,
1378 : : &unprom_abs);
1379 : :
1380 : : /* So far so good. Since last_stmt was detected as a (summation) reduction,
1381 : : we know that plus_oprnd1 is the reduction variable (defined by a loop-header
1382 : : phi), and plus_oprnd0 is an ssa-name defined by a stmt in the loop body.
1383 : : Then check that plus_oprnd0 is defined by an abs_expr. */
1384 : :
1385 : 32461 : if (!plus_oprnd0)
1386 : : return NULL;
1387 : :
1388 : 24253 : stmt_vec_info abs_stmt_vinfo = vect_get_internal_def (vinfo, plus_oprnd0);
1389 : 24253 : if (!abs_stmt_vinfo)
1390 : : return NULL;
1391 : :
1392 : : /* FORNOW. Can continue analyzing the def-use chain when this stmt in a phi
1393 : : inside the loop (in case we are analyzing an outer-loop). */
1394 : 23544 : gassign *abs_stmt = dyn_cast <gassign *> (abs_stmt_vinfo->stmt);
1395 : 70632 : vect_unpromoted_value unprom[2];
1396 : :
1397 : 23544 : if (!abs_stmt)
1398 : : {
1399 : 27444102 : gcall *abd_stmt = dyn_cast <gcall *> (abs_stmt_vinfo->stmt);
1400 : 279 : if (!abd_stmt
1401 : 279 : || !gimple_call_internal_p (abd_stmt)
1402 : 0 : || gimple_call_num_args (abd_stmt) != 2)
1403 : : return NULL;
1404 : :
1405 : 0 : tree abd_oprnd0 = gimple_call_arg (abd_stmt, 0);
1406 : 0 : tree abd_oprnd1 = gimple_call_arg (abd_stmt, 1);
1407 : :
1408 : 0 : if (gimple_call_internal_fn (abd_stmt) == IFN_ABD)
1409 : : {
1410 : 0 : if (!vect_look_through_possible_promotion (vinfo, abd_oprnd0,
1411 : : &unprom[0])
1412 : 0 : || !vect_look_through_possible_promotion (vinfo, abd_oprnd1,
1413 : : &unprom[1]))
1414 : 0 : return NULL;
1415 : : }
1416 : 0 : else if (gimple_call_internal_fn (abd_stmt) == IFN_VEC_WIDEN_ABD)
1417 : : {
1418 : 0 : unprom[0].op = abd_oprnd0;
1419 : 0 : unprom[0].type = TREE_TYPE (abd_oprnd0);
1420 : 0 : unprom[1].op = abd_oprnd1;
1421 : 0 : unprom[1].type = TREE_TYPE (abd_oprnd1);
1422 : : }
1423 : : else
1424 : : return NULL;
1425 : :
1426 : 0 : half_type = unprom[0].type;
1427 : : }
1428 : 23214 : else if (!vect_recog_absolute_difference (vinfo, abs_stmt, &half_type,
1429 : : unprom, NULL))
1430 : : return NULL;
1431 : :
1432 : 367 : vect_pattern_detected ("vect_recog_sad_pattern", last_stmt);
1433 : :
1434 : 367 : tree half_vectype;
1435 : 367 : if (!vect_supportable_direct_optab_p (vinfo, sum_type, SAD_EXPR, half_type,
1436 : : type_out, &half_vectype))
1437 : : return NULL;
1438 : :
1439 : : /* Get the inputs to the SAD_EXPR in the appropriate types. */
1440 : 226 : tree sad_oprnd[2];
1441 : 226 : vect_convert_inputs (vinfo, stmt_vinfo, 2, sad_oprnd, half_type,
1442 : : unprom, half_vectype);
1443 : :
1444 : 226 : tree var = vect_recog_temp_ssa_var (sum_type, NULL);
1445 : 226 : gimple *pattern_stmt = gimple_build_assign (var, SAD_EXPR, sad_oprnd[0],
1446 : : sad_oprnd[1], plus_oprnd1);
1447 : :
1448 : 226 : return pattern_stmt;
1449 : : }
1450 : :
1451 : : /* Function vect_recog_abd_pattern
1452 : :
1453 : : Try to find the following ABsolute Difference (ABD) or
1454 : : widening ABD (WIDEN_ABD) pattern:
1455 : :
1456 : : TYPE1 x;
1457 : : TYPE2 y;
1458 : : TYPE3 x_cast = (TYPE3) x; // widening or no-op
1459 : : TYPE3 y_cast = (TYPE3) y; // widening or no-op
1460 : : TYPE3 diff = x_cast - y_cast;
1461 : : TYPE4 diff_cast = (TYPE4) diff; // widening or no-op
1462 : : TYPE5 abs = ABS(U)_EXPR <diff_cast>;
1463 : :
1464 : : WIDEN_ABD exists to optimize the case where TYPE4 is at least
1465 : : twice as wide as TYPE3.
1466 : :
1467 : : Input:
1468 : :
1469 : : * STMT_VINFO: The stmt from which the pattern search begins
1470 : :
1471 : : Output:
1472 : :
1473 : : * TYPE_OUT: The type of the output of this pattern
1474 : :
1475 : : * Return value: A new stmt that will be used to replace the sequence of
1476 : : stmts that constitute the pattern, principally:
1477 : : out = IFN_ABD (x, y)
1478 : : out = IFN_WIDEN_ABD (x, y)
1479 : : */
1480 : :
1481 : : static gimple *
1482 : 27400617 : vect_recog_abd_pattern (vec_info *vinfo,
1483 : : stmt_vec_info stmt_vinfo, tree *type_out)
1484 : : {
1485 : 46096280 : gassign *last_stmt = dyn_cast <gassign *> (STMT_VINFO_STMT (stmt_vinfo));
1486 : 18695663 : if (!last_stmt)
1487 : : return NULL;
1488 : :
1489 : 18695663 : tree out_type = TREE_TYPE (gimple_assign_lhs (last_stmt));
1490 : :
1491 : 56086989 : vect_unpromoted_value unprom[2];
1492 : 18695663 : gassign *diff_stmt = NULL;
1493 : 18695663 : tree abd_in_type;
1494 : 18695663 : if (!vect_recog_absolute_difference (vinfo, last_stmt, &abd_in_type,
1495 : : unprom, &diff_stmt))
1496 : : {
1497 : : /* We cannot try further without having a non-widening MINUS. */
1498 : 18694777 : if (!diff_stmt)
1499 : : return NULL;
1500 : :
1501 : 1052 : unprom[0].op = gimple_assign_rhs1 (diff_stmt);
1502 : 1052 : unprom[1].op = gimple_assign_rhs2 (diff_stmt);
1503 : 1052 : abd_in_type = signed_type_for (out_type);
1504 : : }
1505 : :
1506 : 1938 : tree abd_out_type = abd_in_type;
1507 : :
1508 : 1938 : tree vectype_in = get_vectype_for_scalar_type (vinfo, abd_in_type);
1509 : 1938 : if (!vectype_in)
1510 : : return NULL;
1511 : :
1512 : 1672 : internal_fn ifn = IFN_ABD;
1513 : 1672 : tree vectype_out = vectype_in;
1514 : :
1515 : 1672 : if (TYPE_PRECISION (out_type) >= TYPE_PRECISION (abd_in_type) * 2
1516 : 1672 : && stmt_vinfo->min_output_precision >= TYPE_PRECISION (abd_in_type) * 2)
1517 : : {
1518 : 779 : tree mid_type
1519 : 779 : = build_nonstandard_integer_type (TYPE_PRECISION (abd_in_type) * 2,
1520 : 779 : TYPE_UNSIGNED (abd_in_type));
1521 : 779 : tree mid_vectype = get_vectype_for_scalar_type (vinfo, mid_type);
1522 : :
1523 : 779 : code_helper dummy_code;
1524 : 779 : int dummy_int;
1525 : 779 : auto_vec<tree> dummy_vec;
1526 : 779 : if (mid_vectype
1527 : 779 : && supportable_widening_operation (vinfo, IFN_VEC_WIDEN_ABD,
1528 : : stmt_vinfo, mid_vectype,
1529 : : vectype_in,
1530 : : &dummy_code, &dummy_code,
1531 : : &dummy_int, &dummy_vec))
1532 : : {
1533 : 0 : ifn = IFN_VEC_WIDEN_ABD;
1534 : 0 : abd_out_type = mid_type;
1535 : 0 : vectype_out = mid_vectype;
1536 : : }
1537 : 779 : }
1538 : :
1539 : 779 : if (ifn == IFN_ABD
1540 : 1672 : && !direct_internal_fn_supported_p (ifn, vectype_in,
1541 : : OPTIMIZE_FOR_SPEED))
1542 : : return NULL;
1543 : :
1544 : 0 : vect_pattern_detected ("vect_recog_abd_pattern", last_stmt);
1545 : :
1546 : 0 : tree abd_oprnds[2];
1547 : 0 : vect_convert_inputs (vinfo, stmt_vinfo, 2, abd_oprnds,
1548 : : abd_in_type, unprom, vectype_in);
1549 : :
1550 : 0 : *type_out = get_vectype_for_scalar_type (vinfo, out_type);
1551 : :
1552 : 0 : tree abd_result = vect_recog_temp_ssa_var (abd_out_type, NULL);
1553 : 0 : gcall *abd_stmt = gimple_build_call_internal (ifn, 2,
1554 : : abd_oprnds[0], abd_oprnds[1]);
1555 : 0 : gimple_call_set_lhs (abd_stmt, abd_result);
1556 : 0 : gimple_set_location (abd_stmt, gimple_location (last_stmt));
1557 : :
1558 : 0 : gimple *stmt = abd_stmt;
1559 : 0 : if (TYPE_PRECISION (abd_in_type) == TYPE_PRECISION (abd_out_type)
1560 : 0 : && TYPE_PRECISION (abd_out_type) < TYPE_PRECISION (out_type)
1561 : 0 : && !TYPE_UNSIGNED (abd_out_type))
1562 : : {
1563 : 0 : tree unsign = unsigned_type_for (abd_out_type);
1564 : 0 : stmt = vect_convert_output (vinfo, stmt_vinfo, unsign, stmt, vectype_out);
1565 : 0 : vectype_out = get_vectype_for_scalar_type (vinfo, unsign);
1566 : : }
1567 : :
1568 : 0 : return vect_convert_output (vinfo, stmt_vinfo, out_type, stmt, vectype_out);
1569 : : }
1570 : :
1571 : : /* Recognize an operation that performs ORIG_CODE on widened inputs,
1572 : : so that it can be treated as though it had the form:
1573 : :
1574 : : A_TYPE a;
1575 : : B_TYPE b;
1576 : : HALF_TYPE a_cast = (HALF_TYPE) a; // possible no-op
1577 : : HALF_TYPE b_cast = (HALF_TYPE) b; // possible no-op
1578 : : | RES_TYPE a_extend = (RES_TYPE) a_cast; // promotion from HALF_TYPE
1579 : : | RES_TYPE b_extend = (RES_TYPE) b_cast; // promotion from HALF_TYPE
1580 : : | RES_TYPE res = a_extend ORIG_CODE b_extend;
1581 : :
1582 : : Try to replace the pattern with:
1583 : :
1584 : : A_TYPE a;
1585 : : B_TYPE b;
1586 : : HALF_TYPE a_cast = (HALF_TYPE) a; // possible no-op
1587 : : HALF_TYPE b_cast = (HALF_TYPE) b; // possible no-op
1588 : : | EXT_TYPE ext = a_cast WIDE_CODE b_cast;
1589 : : | RES_TYPE res = (EXT_TYPE) ext; // possible no-op
1590 : :
1591 : : where EXT_TYPE is wider than HALF_TYPE but has the same signedness.
1592 : :
1593 : : SHIFT_P is true if ORIG_CODE and WIDE_CODE are shifts. NAME is the
1594 : : name of the pattern being matched, for dump purposes. */
1595 : :
1596 : : static gimple *
1597 : 110173964 : vect_recog_widen_op_pattern (vec_info *vinfo,
1598 : : stmt_vec_info last_stmt_info, tree *type_out,
1599 : : tree_code orig_code, code_helper wide_code,
1600 : : bool shift_p, const char *name)
1601 : : {
1602 : 110173964 : gimple *last_stmt = last_stmt_info->stmt;
1603 : :
1604 : 330521892 : vect_unpromoted_value unprom[2];
1605 : 110173964 : tree half_type;
1606 : 110173964 : if (!vect_widened_op_tree (vinfo, last_stmt_info, orig_code, orig_code,
1607 : : shift_p, 2, unprom, &half_type))
1608 : :
1609 : : return NULL;
1610 : :
1611 : : /* Pattern detected. */
1612 : 280630 : vect_pattern_detected (name, last_stmt);
1613 : :
1614 : 280630 : tree type = TREE_TYPE (gimple_get_lhs (last_stmt));
1615 : 280630 : tree itype = type;
1616 : 280630 : if (TYPE_PRECISION (type) != TYPE_PRECISION (half_type) * 2
1617 : 280630 : || TYPE_UNSIGNED (type) != TYPE_UNSIGNED (half_type))
1618 : 197805 : itype = build_nonstandard_integer_type (TYPE_PRECISION (half_type) * 2,
1619 : 197805 : TYPE_UNSIGNED (half_type));
1620 : :
1621 : : /* Check target support */
1622 : 280630 : tree vectype = get_vectype_for_scalar_type (vinfo, half_type);
1623 : 280630 : tree vecitype = get_vectype_for_scalar_type (vinfo, itype);
1624 : 280630 : tree ctype = itype;
1625 : 280630 : tree vecctype = vecitype;
1626 : 280630 : if (orig_code == MINUS_EXPR
1627 : 7779 : && TYPE_UNSIGNED (itype)
1628 : 284169 : && TYPE_PRECISION (type) > TYPE_PRECISION (itype))
1629 : : {
1630 : : /* Subtraction is special, even if half_type is unsigned and no matter
1631 : : whether type is signed or unsigned, if type is wider than itype,
1632 : : we need to sign-extend from the widening operation result to the
1633 : : result type.
1634 : : Consider half_type unsigned char, operand 1 0xfe, operand 2 0xff,
1635 : : itype unsigned short and type either int or unsigned int.
1636 : : Widened (unsigned short) 0xfe - (unsigned short) 0xff is
1637 : : (unsigned short) 0xffff, but for type int we want the result -1
1638 : : and for type unsigned int 0xffffffff rather than 0xffff. */
1639 : 578 : ctype = build_nonstandard_integer_type (TYPE_PRECISION (itype), 0);
1640 : 578 : vecctype = get_vectype_for_scalar_type (vinfo, ctype);
1641 : : }
1642 : :
1643 : 280630 : code_helper dummy_code;
1644 : 280630 : int dummy_int;
1645 : 280630 : auto_vec<tree> dummy_vec;
1646 : 280630 : if (!vectype
1647 : 280630 : || !vecitype
1648 : 221221 : || !vecctype
1649 : 501851 : || !supportable_widening_operation (vinfo, wide_code, last_stmt_info,
1650 : : vecitype, vectype,
1651 : : &dummy_code, &dummy_code,
1652 : : &dummy_int, &dummy_vec))
1653 : 191297 : return NULL;
1654 : :
1655 : 89333 : *type_out = get_vectype_for_scalar_type (vinfo, type);
1656 : 89333 : if (!*type_out)
1657 : : return NULL;
1658 : :
1659 : 89333 : tree oprnd[2];
1660 : 89333 : vect_convert_inputs (vinfo, last_stmt_info,
1661 : : 2, oprnd, half_type, unprom, vectype);
1662 : :
1663 : 89333 : tree var = vect_recog_temp_ssa_var (itype, NULL);
1664 : 89333 : gimple *pattern_stmt = vect_gimple_build (var, wide_code, oprnd[0], oprnd[1]);
1665 : :
1666 : 89333 : if (vecctype != vecitype)
1667 : 0 : pattern_stmt = vect_convert_output (vinfo, last_stmt_info, ctype,
1668 : : pattern_stmt, vecitype);
1669 : :
1670 : 89333 : return vect_convert_output (vinfo, last_stmt_info,
1671 : 89333 : type, pattern_stmt, vecctype);
1672 : 280630 : }
1673 : :
1674 : : /* Try to detect multiplication on widened inputs, converting MULT_EXPR
1675 : : to WIDEN_MULT_EXPR. See vect_recog_widen_op_pattern for details. */
1676 : :
1677 : : static gimple *
1678 : 27465760 : vect_recog_widen_mult_pattern (vec_info *vinfo, stmt_vec_info last_stmt_info,
1679 : : tree *type_out)
1680 : : {
1681 : 27465760 : return vect_recog_widen_op_pattern (vinfo, last_stmt_info, type_out,
1682 : : MULT_EXPR, WIDEN_MULT_EXPR, false,
1683 : 27465760 : "vect_recog_widen_mult_pattern");
1684 : : }
1685 : :
1686 : : /* Try to detect addition on widened inputs, converting PLUS_EXPR
1687 : : to IFN_VEC_WIDEN_PLUS. See vect_recog_widen_op_pattern for details. */
1688 : :
1689 : : static gimple *
1690 : 27632087 : vect_recog_widen_plus_pattern (vec_info *vinfo, stmt_vec_info last_stmt_info,
1691 : : tree *type_out)
1692 : : {
1693 : 27632087 : return vect_recog_widen_op_pattern (vinfo, last_stmt_info, type_out,
1694 : : PLUS_EXPR, IFN_VEC_WIDEN_PLUS,
1695 : 27632087 : false, "vect_recog_widen_plus_pattern");
1696 : : }
1697 : :
1698 : : /* Try to detect subtraction on widened inputs, converting MINUS_EXPR
1699 : : to IFN_VEC_WIDEN_MINUS. See vect_recog_widen_op_pattern for details. */
1700 : : static gimple *
1701 : 27632087 : vect_recog_widen_minus_pattern (vec_info *vinfo, stmt_vec_info last_stmt_info,
1702 : : tree *type_out)
1703 : : {
1704 : 27632087 : return vect_recog_widen_op_pattern (vinfo, last_stmt_info, type_out,
1705 : : MINUS_EXPR, IFN_VEC_WIDEN_MINUS,
1706 : 27632087 : false, "vect_recog_widen_minus_pattern");
1707 : : }
1708 : :
1709 : : /* Try to detect abd on widened inputs, converting IFN_ABD
1710 : : to IFN_VEC_WIDEN_ABD. */
1711 : : static gimple *
1712 : 27632087 : vect_recog_widen_abd_pattern (vec_info *vinfo, stmt_vec_info stmt_vinfo,
1713 : : tree *type_out)
1714 : : {
1715 : 27632087 : gassign *last_stmt = dyn_cast <gassign *> (STMT_VINFO_STMT (stmt_vinfo));
1716 : 25601599 : if (!last_stmt || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (last_stmt)))
1717 : : return NULL;
1718 : :
1719 : 2592850 : tree last_rhs = gimple_assign_rhs1 (last_stmt);
1720 : :
1721 : 2592850 : tree in_type = TREE_TYPE (last_rhs);
1722 : 2592850 : tree out_type = TREE_TYPE (gimple_assign_lhs (last_stmt));
1723 : 2592850 : if (!INTEGRAL_TYPE_P (in_type)
1724 : 2322212 : || !INTEGRAL_TYPE_P (out_type)
1725 : 2232119 : || TYPE_PRECISION (in_type) * 2 != TYPE_PRECISION (out_type)
1726 : 3118823 : || !TYPE_UNSIGNED (in_type))
1727 : : return NULL;
1728 : :
1729 : 179921 : vect_unpromoted_value unprom;
1730 : 179921 : tree op = vect_look_through_possible_promotion (vinfo, last_rhs, &unprom);
1731 : 179921 : if (!op || TYPE_PRECISION (TREE_TYPE (op)) != TYPE_PRECISION (in_type))
1732 : : return NULL;
1733 : :
1734 : 179348 : stmt_vec_info abd_pattern_vinfo = vect_get_internal_def (vinfo, op);
1735 : 179348 : if (!abd_pattern_vinfo)
1736 : : return NULL;
1737 : :
1738 : 27641218 : gcall *abd_stmt = dyn_cast <gcall *> (STMT_VINFO_STMT (abd_pattern_vinfo));
1739 : 9131 : if (!abd_stmt
1740 : 9131 : || !gimple_call_internal_p (abd_stmt)
1741 : 250 : || gimple_call_internal_fn (abd_stmt) != IFN_ABD)
1742 : : return NULL;
1743 : :
1744 : 0 : tree vectype_in = get_vectype_for_scalar_type (vinfo, in_type);
1745 : 0 : tree vectype_out = get_vectype_for_scalar_type (vinfo, out_type);
1746 : :
1747 : 0 : code_helper dummy_code;
1748 : 0 : int dummy_int;
1749 : 0 : auto_vec<tree> dummy_vec;
1750 : 0 : if (!supportable_widening_operation (vinfo, IFN_VEC_WIDEN_ABD, stmt_vinfo,
1751 : : vectype_out, vectype_in,
1752 : : &dummy_code, &dummy_code,
1753 : : &dummy_int, &dummy_vec))
1754 : : return NULL;
1755 : :
1756 : 0 : vect_pattern_detected ("vect_recog_widen_abd_pattern", last_stmt);
1757 : :
1758 : 0 : *type_out = vectype_out;
1759 : :
1760 : 0 : tree abd_oprnd0 = gimple_call_arg (abd_stmt, 0);
1761 : 0 : tree abd_oprnd1 = gimple_call_arg (abd_stmt, 1);
1762 : 0 : tree widen_abd_result = vect_recog_temp_ssa_var (out_type, NULL);
1763 : 0 : gcall *widen_abd_stmt = gimple_build_call_internal (IFN_VEC_WIDEN_ABD, 2,
1764 : : abd_oprnd0, abd_oprnd1);
1765 : 0 : gimple_call_set_lhs (widen_abd_stmt, widen_abd_result);
1766 : 0 : gimple_set_location (widen_abd_stmt, gimple_location (last_stmt));
1767 : 0 : return widen_abd_stmt;
1768 : 0 : }
1769 : :
1770 : : /* Function vect_recog_ctz_ffs_pattern
1771 : :
1772 : : Try to find the following pattern:
1773 : :
1774 : : TYPE1 A;
1775 : : TYPE1 B;
1776 : :
1777 : : B = __builtin_ctz{,l,ll} (A);
1778 : :
1779 : : or
1780 : :
1781 : : B = __builtin_ffs{,l,ll} (A);
1782 : :
1783 : : Input:
1784 : :
1785 : : * STMT_VINFO: The stmt from which the pattern search begins.
1786 : : here it starts with B = __builtin_* (A);
1787 : :
1788 : : Output:
1789 : :
1790 : : * TYPE_OUT: The vector type of the output of this pattern.
1791 : :
1792 : : * Return value: A new stmt that will be used to replace the sequence of
1793 : : stmts that constitute the pattern, using clz or popcount builtins. */
1794 : :
1795 : : static gimple *
1796 : 27443934 : vect_recog_ctz_ffs_pattern (vec_info *vinfo, stmt_vec_info stmt_vinfo,
1797 : : tree *type_out)
1798 : : {
1799 : 27443934 : gimple *call_stmt = stmt_vinfo->stmt;
1800 : 27443934 : gimple *pattern_stmt;
1801 : 27443934 : tree rhs_oprnd, rhs_type, lhs_oprnd, lhs_type, vec_type, vec_rhs_type;
1802 : 27443934 : tree new_var;
1803 : 27443934 : internal_fn ifn = IFN_LAST, ifnnew = IFN_LAST;
1804 : 27443934 : bool defined_at_zero = true, defined_at_zero_new = false;
1805 : 27443934 : int val = 0, val_new = 0, val_cmp = 0;
1806 : 27443934 : int prec;
1807 : 27443934 : int sub = 0, add = 0;
1808 : 27443934 : location_t loc;
1809 : :
1810 : 27443934 : if (!is_gimple_call (call_stmt))
1811 : : return NULL;
1812 : :
1813 : 3289908 : if (gimple_call_num_args (call_stmt) != 1
1814 : 3289908 : && gimple_call_num_args (call_stmt) != 2)
1815 : : return NULL;
1816 : :
1817 : 1787287 : rhs_oprnd = gimple_call_arg (call_stmt, 0);
1818 : 1787287 : rhs_type = TREE_TYPE (rhs_oprnd);
1819 : 1787287 : lhs_oprnd = gimple_call_lhs (call_stmt);
1820 : 1787287 : if (!lhs_oprnd)
1821 : : return NULL;
1822 : 895562 : lhs_type = TREE_TYPE (lhs_oprnd);
1823 : 895562 : if (!INTEGRAL_TYPE_P (lhs_type)
1824 : 305157 : || !INTEGRAL_TYPE_P (rhs_type)
1825 : 45322 : || !type_has_mode_precision_p (rhs_type)
1826 : 939448 : || TREE_CODE (rhs_oprnd) != SSA_NAME)
1827 : 863046 : return NULL;
1828 : :
1829 : 32516 : switch (gimple_call_combined_fn (call_stmt))
1830 : : {
1831 : 1136 : CASE_CFN_CTZ:
1832 : 1136 : ifn = IFN_CTZ;
1833 : 1136 : if (!gimple_call_internal_p (call_stmt)
1834 : 1136 : || gimple_call_num_args (call_stmt) != 2)
1835 : : defined_at_zero = false;
1836 : : else
1837 : 48 : val = tree_to_shwi (gimple_call_arg (call_stmt, 1));
1838 : : break;
1839 : : CASE_CFN_FFS:
1840 : : ifn = IFN_FFS;
1841 : : break;
1842 : : default:
1843 : : return NULL;
1844 : : }
1845 : :
1846 : 1296 : prec = TYPE_PRECISION (rhs_type);
1847 : 1296 : loc = gimple_location (call_stmt);
1848 : :
1849 : 1296 : vec_type = get_vectype_for_scalar_type (vinfo, lhs_type);
1850 : 1296 : if (!vec_type)
1851 : : return NULL;
1852 : :
1853 : 1290 : vec_rhs_type = get_vectype_for_scalar_type (vinfo, rhs_type);
1854 : 1290 : if (!vec_rhs_type)
1855 : : return NULL;
1856 : :
1857 : : /* Do it only if the backend doesn't have ctz<vector_mode>2 or
1858 : : ffs<vector_mode>2 pattern but does have clz<vector_mode>2 or
1859 : : popcount<vector_mode>2. */
1860 : 1054 : if (!vec_type
1861 : 1054 : || direct_internal_fn_supported_p (ifn, vec_rhs_type,
1862 : : OPTIMIZE_FOR_SPEED))
1863 : : return NULL;
1864 : :
1865 : 1054 : if (ifn == IFN_FFS
1866 : 1054 : && direct_internal_fn_supported_p (IFN_CTZ, vec_rhs_type,
1867 : : OPTIMIZE_FOR_SPEED))
1868 : : {
1869 : 0 : ifnnew = IFN_CTZ;
1870 : 0 : defined_at_zero_new
1871 : 0 : = CTZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (rhs_type),
1872 : : val_new) == 2;
1873 : : }
1874 : 1054 : else if (direct_internal_fn_supported_p (IFN_CLZ, vec_rhs_type,
1875 : : OPTIMIZE_FOR_SPEED))
1876 : : {
1877 : 88 : ifnnew = IFN_CLZ;
1878 : 88 : defined_at_zero_new
1879 : 88 : = CLZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (rhs_type),
1880 : : val_new) == 2;
1881 : : }
1882 : 88 : if ((ifnnew == IFN_LAST
1883 : 88 : || (defined_at_zero && !defined_at_zero_new))
1884 : 966 : && direct_internal_fn_supported_p (IFN_POPCOUNT, vec_rhs_type,
1885 : : OPTIMIZE_FOR_SPEED))
1886 : : {
1887 : : ifnnew = IFN_POPCOUNT;
1888 : : defined_at_zero_new = true;
1889 : : val_new = prec;
1890 : : }
1891 : 1018 : if (ifnnew == IFN_LAST)
1892 : : return NULL;
1893 : :
1894 : 124 : vect_pattern_detected ("vec_recog_ctz_ffs_pattern", call_stmt);
1895 : :
1896 : 124 : val_cmp = val_new;
1897 : 124 : if ((ifnnew == IFN_CLZ
1898 : 124 : && defined_at_zero
1899 : 60 : && defined_at_zero_new
1900 : 60 : && val == prec
1901 : 31 : && val_new == prec)
1902 : 93 : || (ifnnew == IFN_POPCOUNT && ifn == IFN_CTZ))
1903 : : {
1904 : : /* .CTZ (X) = PREC - .CLZ ((X - 1) & ~X)
1905 : : .CTZ (X) = .POPCOUNT ((X - 1) & ~X). */
1906 : : if (ifnnew == IFN_CLZ)
1907 : : sub = prec;
1908 : 56 : val_cmp = prec;
1909 : :
1910 : 56 : if (!TYPE_UNSIGNED (rhs_type))
1911 : : {
1912 : 12 : rhs_type = unsigned_type_for (rhs_type);
1913 : 12 : vec_rhs_type = get_vectype_for_scalar_type (vinfo, rhs_type);
1914 : 12 : new_var = vect_recog_temp_ssa_var (rhs_type, NULL);
1915 : 12 : pattern_stmt = gimple_build_assign (new_var, NOP_EXPR, rhs_oprnd);
1916 : 12 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt,
1917 : : vec_rhs_type);
1918 : 12 : rhs_oprnd = new_var;
1919 : : }
1920 : :
1921 : 56 : tree m1 = vect_recog_temp_ssa_var (rhs_type, NULL);
1922 : 56 : pattern_stmt = gimple_build_assign (m1, PLUS_EXPR, rhs_oprnd,
1923 : 56 : build_int_cst (rhs_type, -1));
1924 : 56 : gimple_set_location (pattern_stmt, loc);
1925 : 56 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_rhs_type);
1926 : :
1927 : 56 : new_var = vect_recog_temp_ssa_var (rhs_type, NULL);
1928 : 56 : pattern_stmt = gimple_build_assign (new_var, BIT_NOT_EXPR, rhs_oprnd);
1929 : 56 : gimple_set_location (pattern_stmt, loc);
1930 : 56 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_rhs_type);
1931 : 56 : rhs_oprnd = new_var;
1932 : :
1933 : 56 : new_var = vect_recog_temp_ssa_var (rhs_type, NULL);
1934 : 56 : pattern_stmt = gimple_build_assign (new_var, BIT_AND_EXPR,
1935 : : m1, rhs_oprnd);
1936 : 56 : gimple_set_location (pattern_stmt, loc);
1937 : 56 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_rhs_type);
1938 : 56 : rhs_oprnd = new_var;
1939 : 56 : }
1940 : 68 : else if (ifnnew == IFN_CLZ)
1941 : : {
1942 : : /* .CTZ (X) = (PREC - 1) - .CLZ (X & -X)
1943 : : .FFS (X) = PREC - .CLZ (X & -X). */
1944 : 57 : sub = prec - (ifn == IFN_CTZ);
1945 : 57 : val_cmp = sub - val_new;
1946 : :
1947 : 57 : tree neg = vect_recog_temp_ssa_var (rhs_type, NULL);
1948 : 57 : pattern_stmt = gimple_build_assign (neg, NEGATE_EXPR, rhs_oprnd);
1949 : 57 : gimple_set_location (pattern_stmt, loc);
1950 : 57 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_rhs_type);
1951 : :
1952 : 57 : new_var = vect_recog_temp_ssa_var (rhs_type, NULL);
1953 : 57 : pattern_stmt = gimple_build_assign (new_var, BIT_AND_EXPR,
1954 : : rhs_oprnd, neg);
1955 : 57 : gimple_set_location (pattern_stmt, loc);
1956 : 57 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_rhs_type);
1957 : 57 : rhs_oprnd = new_var;
1958 : : }
1959 : 11 : else if (ifnnew == IFN_POPCOUNT)
1960 : : {
1961 : : /* .CTZ (X) = PREC - .POPCOUNT (X | -X)
1962 : : .FFS (X) = (PREC + 1) - .POPCOUNT (X | -X). */
1963 : 11 : sub = prec + (ifn == IFN_FFS);
1964 : 11 : val_cmp = sub;
1965 : :
1966 : 11 : tree neg = vect_recog_temp_ssa_var (rhs_type, NULL);
1967 : 11 : pattern_stmt = gimple_build_assign (neg, NEGATE_EXPR, rhs_oprnd);
1968 : 11 : gimple_set_location (pattern_stmt, loc);
1969 : 11 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_rhs_type);
1970 : :
1971 : 11 : new_var = vect_recog_temp_ssa_var (rhs_type, NULL);
1972 : 11 : pattern_stmt = gimple_build_assign (new_var, BIT_IOR_EXPR,
1973 : : rhs_oprnd, neg);
1974 : 11 : gimple_set_location (pattern_stmt, loc);
1975 : 11 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_rhs_type);
1976 : 11 : rhs_oprnd = new_var;
1977 : : }
1978 : 0 : else if (ifnnew == IFN_CTZ)
1979 : : {
1980 : : /* .FFS (X) = .CTZ (X) + 1. */
1981 : 0 : add = 1;
1982 : 0 : val_cmp++;
1983 : : }
1984 : :
1985 : : /* Create B = .IFNNEW (A). */
1986 : 124 : new_var = vect_recog_temp_ssa_var (lhs_type, NULL);
1987 : 124 : if ((ifnnew == IFN_CLZ || ifnnew == IFN_CTZ) && defined_at_zero_new)
1988 : 88 : pattern_stmt
1989 : 88 : = gimple_build_call_internal (ifnnew, 2, rhs_oprnd,
1990 : 88 : build_int_cst (integer_type_node,
1991 : : val_new));
1992 : : else
1993 : 36 : pattern_stmt = gimple_build_call_internal (ifnnew, 1, rhs_oprnd);
1994 : 124 : gimple_call_set_lhs (pattern_stmt, new_var);
1995 : 124 : gimple_set_location (pattern_stmt, loc);
1996 : 124 : *type_out = vec_type;
1997 : :
1998 : 124 : if (sub)
1999 : : {
2000 : 99 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_type);
2001 : 99 : tree ret_var = vect_recog_temp_ssa_var (lhs_type, NULL);
2002 : 99 : pattern_stmt = gimple_build_assign (ret_var, MINUS_EXPR,
2003 : 99 : build_int_cst (lhs_type, sub),
2004 : : new_var);
2005 : 99 : gimple_set_location (pattern_stmt, loc);
2006 : 99 : new_var = ret_var;
2007 : : }
2008 : 25 : else if (add)
2009 : : {
2010 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_type);
2011 : 0 : tree ret_var = vect_recog_temp_ssa_var (lhs_type, NULL);
2012 : 0 : pattern_stmt = gimple_build_assign (ret_var, PLUS_EXPR, new_var,
2013 : 0 : build_int_cst (lhs_type, add));
2014 : 0 : gimple_set_location (pattern_stmt, loc);
2015 : 0 : new_var = ret_var;
2016 : : }
2017 : :
2018 : 124 : if (defined_at_zero
2019 : 88 : && (!defined_at_zero_new || val != val_cmp))
2020 : : {
2021 : 11 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_type);
2022 : 11 : tree ret_var = vect_recog_temp_ssa_var (lhs_type, NULL);
2023 : 11 : rhs_oprnd = gimple_call_arg (call_stmt, 0);
2024 : 11 : rhs_type = TREE_TYPE (rhs_oprnd);
2025 : 11 : tree cmp = vect_recog_temp_ssa_var (boolean_type_node, NULL);
2026 : 11 : pattern_stmt = gimple_build_assign (cmp, NE_EXPR, rhs_oprnd,
2027 : : build_zero_cst (rhs_type));
2028 : 11 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt,
2029 : : truth_type_for (vec_type), rhs_type);
2030 : 11 : pattern_stmt = gimple_build_assign (ret_var, COND_EXPR, cmp,
2031 : : new_var,
2032 : 11 : build_int_cst (lhs_type, val));
2033 : : }
2034 : :
2035 : 124 : if (dump_enabled_p ())
2036 : 36 : dump_printf_loc (MSG_NOTE, vect_location,
2037 : : "created pattern stmt: %G", pattern_stmt);
2038 : :
2039 : : return pattern_stmt;
2040 : : }
2041 : :
2042 : : /* Function vect_recog_popcount_clz_ctz_ffs_pattern
2043 : :
2044 : : Try to find the following pattern:
2045 : :
2046 : : UTYPE1 A;
2047 : : TYPE1 B;
2048 : : UTYPE2 temp_in;
2049 : : TYPE3 temp_out;
2050 : : temp_in = (UTYPE2)A;
2051 : :
2052 : : temp_out = __builtin_popcount{,l,ll} (temp_in);
2053 : : B = (TYPE1) temp_out;
2054 : :
2055 : : TYPE2 may or may not be equal to TYPE3.
2056 : : i.e. TYPE2 is equal to TYPE3 for __builtin_popcount
2057 : : i.e. TYPE2 is not equal to TYPE3 for __builtin_popcountll
2058 : :
2059 : : Input:
2060 : :
2061 : : * STMT_VINFO: The stmt from which the pattern search begins.
2062 : : here it starts with B = (TYPE1) temp_out;
2063 : :
2064 : : Output:
2065 : :
2066 : : * TYPE_OUT: The vector type of the output of this pattern.
2067 : :
2068 : : * Return value: A new stmt that will be used to replace the sequence of
2069 : : stmts that constitute the pattern. In this case it will be:
2070 : : B = .POPCOUNT (A);
2071 : :
2072 : : Similarly for clz, ctz and ffs.
2073 : : */
2074 : :
2075 : : static gimple *
2076 : 27443814 : vect_recog_popcount_clz_ctz_ffs_pattern (vec_info *vinfo,
2077 : : stmt_vec_info stmt_vinfo,
2078 : : tree *type_out)
2079 : : {
2080 : 27443814 : gassign *last_stmt = dyn_cast <gassign *> (stmt_vinfo->stmt);
2081 : 18738680 : gimple *call_stmt, *pattern_stmt;
2082 : 18738680 : tree rhs_oprnd, rhs_origin, lhs_oprnd, lhs_type, vec_type, new_var;
2083 : 46182335 : internal_fn ifn = IFN_LAST;
2084 : 27443655 : int addend = 0;
2085 : :
2086 : : /* Find B = (TYPE1) temp_out. */
2087 : 18738680 : if (!last_stmt)
2088 : : return NULL;
2089 : 18738680 : tree_code code = gimple_assign_rhs_code (last_stmt);
2090 : 18738680 : if (!CONVERT_EXPR_CODE_P (code))
2091 : : return NULL;
2092 : :
2093 : 2558680 : lhs_oprnd = gimple_assign_lhs (last_stmt);
2094 : 2558680 : lhs_type = TREE_TYPE (lhs_oprnd);
2095 : 2558680 : if (!INTEGRAL_TYPE_P (lhs_type))
2096 : : return NULL;
2097 : :
2098 : 2408339 : rhs_oprnd = gimple_assign_rhs1 (last_stmt);
2099 : 2408339 : if (TREE_CODE (rhs_oprnd) != SSA_NAME
2100 : 2408339 : || !has_single_use (rhs_oprnd))
2101 : : return NULL;
2102 : 1259806 : call_stmt = SSA_NAME_DEF_STMT (rhs_oprnd);
2103 : :
2104 : : /* Find temp_out = __builtin_popcount{,l,ll} (temp_in); */
2105 : 1259806 : if (!is_gimple_call (call_stmt))
2106 : : return NULL;
2107 : 95998 : switch (gimple_call_combined_fn (call_stmt))
2108 : : {
2109 : : int val;
2110 : : CASE_CFN_POPCOUNT:
2111 : : ifn = IFN_POPCOUNT;
2112 : : break;
2113 : 2258 : CASE_CFN_CLZ:
2114 : 2258 : ifn = IFN_CLZ;
2115 : : /* Punt if call result is unsigned and defined value at zero
2116 : : is negative, as the negative value doesn't extend correctly. */
2117 : 2258 : if (TYPE_UNSIGNED (TREE_TYPE (rhs_oprnd))
2118 : 0 : && gimple_call_internal_p (call_stmt)
2119 : 2258 : && CLZ_DEFINED_VALUE_AT_ZERO
2120 : : (SCALAR_INT_TYPE_MODE (TREE_TYPE (rhs_oprnd)), val) == 2
2121 : 2258 : && val < 0)
2122 : : return NULL;
2123 : : break;
2124 : 578 : CASE_CFN_CTZ:
2125 : 578 : ifn = IFN_CTZ;
2126 : : /* Punt if call result is unsigned and defined value at zero
2127 : : is negative, as the negative value doesn't extend correctly. */
2128 : 578 : if (TYPE_UNSIGNED (TREE_TYPE (rhs_oprnd))
2129 : 0 : && gimple_call_internal_p (call_stmt)
2130 : 578 : && CTZ_DEFINED_VALUE_AT_ZERO
2131 : : (SCALAR_INT_TYPE_MODE (TREE_TYPE (rhs_oprnd)), val) == 2
2132 : 578 : && val < 0)
2133 : : return NULL;
2134 : : break;
2135 : 18 : CASE_CFN_FFS:
2136 : 18 : ifn = IFN_FFS;
2137 : 18 : break;
2138 : : default:
2139 : : return NULL;
2140 : : }
2141 : :
2142 : 3110 : if (gimple_call_num_args (call_stmt) != 1
2143 : 3110 : && gimple_call_num_args (call_stmt) != 2)
2144 : : return NULL;
2145 : :
2146 : 3110 : rhs_oprnd = gimple_call_arg (call_stmt, 0);
2147 : 3110 : vect_unpromoted_value unprom_diff;
2148 : 3110 : rhs_origin
2149 : 3110 : = vect_look_through_possible_promotion (vinfo, rhs_oprnd, &unprom_diff);
2150 : :
2151 : 3110 : if (!rhs_origin)
2152 : : return NULL;
2153 : :
2154 : : /* Input and output of .POPCOUNT should be same-precision integer. */
2155 : 3110 : if (TYPE_PRECISION (unprom_diff.type) != TYPE_PRECISION (lhs_type))
2156 : : return NULL;
2157 : :
2158 : : /* Also A should be unsigned or same precision as temp_in, otherwise
2159 : : different builtins/internal functions have different behaviors. */
2160 : 1298 : if (TYPE_PRECISION (unprom_diff.type)
2161 : 1298 : != TYPE_PRECISION (TREE_TYPE (rhs_oprnd)))
2162 : 170 : switch (ifn)
2163 : : {
2164 : 79 : case IFN_POPCOUNT:
2165 : : /* For popcount require zero extension, which doesn't add any
2166 : : further bits to the count. */
2167 : 79 : if (!TYPE_UNSIGNED (unprom_diff.type))
2168 : : return NULL;
2169 : : break;
2170 : 73 : case IFN_CLZ:
2171 : : /* clzll (x) == clz (x) + 32 for unsigned x != 0, so ok
2172 : : if it is undefined at zero or if it matches also for the
2173 : : defined value there. */
2174 : 73 : if (!TYPE_UNSIGNED (unprom_diff.type))
2175 : : return NULL;
2176 : 73 : if (!type_has_mode_precision_p (lhs_type)
2177 : 73 : || !type_has_mode_precision_p (TREE_TYPE (rhs_oprnd)))
2178 : 0 : return NULL;
2179 : 73 : addend = (TYPE_PRECISION (TREE_TYPE (rhs_oprnd))
2180 : 73 : - TYPE_PRECISION (lhs_type));
2181 : 73 : if (gimple_call_internal_p (call_stmt)
2182 : 73 : && gimple_call_num_args (call_stmt) == 2)
2183 : : {
2184 : 0 : int val1, val2;
2185 : 0 : val1 = tree_to_shwi (gimple_call_arg (call_stmt, 1));
2186 : 0 : int d2
2187 : 0 : = CLZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (lhs_type),
2188 : : val2);
2189 : 0 : if (d2 != 2 || val1 != val2 + addend)
2190 : : return NULL;
2191 : : }
2192 : : break;
2193 : 13 : case IFN_CTZ:
2194 : : /* ctzll (x) == ctz (x) for unsigned or signed x != 0, so ok
2195 : : if it is undefined at zero or if it matches also for the
2196 : : defined value there. */
2197 : 13 : if (gimple_call_internal_p (call_stmt)
2198 : 13 : && gimple_call_num_args (call_stmt) == 2)
2199 : : {
2200 : 0 : int val1, val2;
2201 : 0 : val1 = tree_to_shwi (gimple_call_arg (call_stmt, 1));
2202 : 0 : int d2
2203 : 0 : = CTZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (lhs_type),
2204 : : val2);
2205 : 0 : if (d2 != 2 || val1 != val2)
2206 : : return NULL;
2207 : : }
2208 : : break;
2209 : : case IFN_FFS:
2210 : : /* ffsll (x) == ffs (x) for unsigned or signed x. */
2211 : : break;
2212 : 0 : default:
2213 : 0 : gcc_unreachable ();
2214 : : }
2215 : :
2216 : 1298 : vec_type = get_vectype_for_scalar_type (vinfo, lhs_type);
2217 : : /* Do it only if the backend has popcount<vector_mode>2 etc. pattern. */
2218 : 1298 : if (!vec_type)
2219 : : return NULL;
2220 : :
2221 : 1160 : bool supported
2222 : 1160 : = direct_internal_fn_supported_p (ifn, vec_type, OPTIMIZE_FOR_SPEED);
2223 : 1160 : if (!supported)
2224 : 1057 : switch (ifn)
2225 : : {
2226 : : case IFN_POPCOUNT:
2227 : : case IFN_CLZ:
2228 : : return NULL;
2229 : 18 : case IFN_FFS:
2230 : : /* vect_recog_ctz_ffs_pattern can implement ffs using ctz. */
2231 : 18 : if (direct_internal_fn_supported_p (IFN_CTZ, vec_type,
2232 : : OPTIMIZE_FOR_SPEED))
2233 : : break;
2234 : : /* FALLTHRU */
2235 : 365 : case IFN_CTZ:
2236 : : /* vect_recog_ctz_ffs_pattern can implement ffs or ctz using
2237 : : clz or popcount. */
2238 : 365 : if (direct_internal_fn_supported_p (IFN_CLZ, vec_type,
2239 : : OPTIMIZE_FOR_SPEED))
2240 : : break;
2241 : 331 : if (direct_internal_fn_supported_p (IFN_POPCOUNT, vec_type,
2242 : : OPTIMIZE_FOR_SPEED))
2243 : : break;
2244 : : return NULL;
2245 : 0 : default:
2246 : 0 : gcc_unreachable ();
2247 : : }
2248 : :
2249 : 159 : vect_pattern_detected ("vec_recog_popcount_clz_ctz_ffs_pattern",
2250 : : call_stmt);
2251 : :
2252 : : /* Create B = .POPCOUNT (A). */
2253 : 159 : new_var = vect_recog_temp_ssa_var (lhs_type, NULL);
2254 : 159 : tree arg2 = NULL_TREE;
2255 : 159 : int val;
2256 : 159 : if (ifn == IFN_CLZ
2257 : 191 : && CLZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (lhs_type),
2258 : : val) == 2)
2259 : 30 : arg2 = build_int_cst (integer_type_node, val);
2260 : 129 : else if (ifn == IFN_CTZ
2261 : 167 : && CTZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (lhs_type),
2262 : : val) == 2)
2263 : 38 : arg2 = build_int_cst (integer_type_node, val);
2264 : 159 : if (arg2)
2265 : 68 : pattern_stmt = gimple_build_call_internal (ifn, 2, unprom_diff.op, arg2);
2266 : : else
2267 : 91 : pattern_stmt = gimple_build_call_internal (ifn, 1, unprom_diff.op);
2268 : 159 : gimple_call_set_lhs (pattern_stmt, new_var);
2269 : 159 : gimple_set_location (pattern_stmt, gimple_location (last_stmt));
2270 : 159 : *type_out = vec_type;
2271 : :
2272 : 159 : if (dump_enabled_p ())
2273 : 24 : dump_printf_loc (MSG_NOTE, vect_location,
2274 : : "created pattern stmt: %G", pattern_stmt);
2275 : :
2276 : 159 : if (addend)
2277 : : {
2278 : 6 : gcc_assert (supported);
2279 : 6 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, vec_type);
2280 : 6 : tree ret_var = vect_recog_temp_ssa_var (lhs_type, NULL);
2281 : 6 : pattern_stmt = gimple_build_assign (ret_var, PLUS_EXPR, new_var,
2282 : 6 : build_int_cst (lhs_type, addend));
2283 : : }
2284 : 153 : else if (!supported)
2285 : : {
2286 : 56 : stmt_vec_info new_stmt_info = vinfo->add_stmt (pattern_stmt);
2287 : 56 : STMT_VINFO_VECTYPE (new_stmt_info) = vec_type;
2288 : 56 : pattern_stmt
2289 : 56 : = vect_recog_ctz_ffs_pattern (vinfo, new_stmt_info, type_out);
2290 : 56 : if (pattern_stmt == NULL)
2291 : : return NULL;
2292 : 56 : if (gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (new_stmt_info))
2293 : : {
2294 : 56 : gimple_seq *pseq = &STMT_VINFO_PATTERN_DEF_SEQ (stmt_vinfo);
2295 : 56 : gimple_seq_add_seq_without_update (pseq, seq);
2296 : : }
2297 : : }
2298 : : return pattern_stmt;
2299 : : }
2300 : :
2301 : : /* Function vect_recog_pow_pattern
2302 : :
2303 : : Try to find the following pattern:
2304 : :
2305 : : x = POW (y, N);
2306 : :
2307 : : with POW being one of pow, powf, powi, powif and N being
2308 : : either 2 or 0.5.
2309 : :
2310 : : Input:
2311 : :
2312 : : * STMT_VINFO: The stmt from which the pattern search begins.
2313 : :
2314 : : Output:
2315 : :
2316 : : * TYPE_OUT: The type of the output of this pattern.
2317 : :
2318 : : * Return value: A new stmt that will be used to replace the sequence of
2319 : : stmts that constitute the pattern. In this case it will be:
2320 : : x = x * x
2321 : : or
2322 : : x = sqrt (x)
2323 : : */
2324 : :
2325 : : static gimple *
2326 : 27443823 : vect_recog_pow_pattern (vec_info *vinfo,
2327 : : stmt_vec_info stmt_vinfo, tree *type_out)
2328 : : {
2329 : 27443823 : gimple *last_stmt = stmt_vinfo->stmt;
2330 : 27443823 : tree base, exp;
2331 : 27443823 : gimple *stmt;
2332 : 27443823 : tree var;
2333 : :
2334 : 27443823 : if (!is_gimple_call (last_stmt) || gimple_call_lhs (last_stmt) == NULL)
2335 : : return NULL;
2336 : :
2337 : 1415203 : switch (gimple_call_combined_fn (last_stmt))
2338 : : {
2339 : 274 : CASE_CFN_POW:
2340 : 274 : CASE_CFN_POWI:
2341 : 274 : break;
2342 : :
2343 : : default:
2344 : : return NULL;
2345 : : }
2346 : :
2347 : 274 : base = gimple_call_arg (last_stmt, 0);
2348 : 274 : exp = gimple_call_arg (last_stmt, 1);
2349 : 274 : if (TREE_CODE (exp) != REAL_CST
2350 : 248 : && TREE_CODE (exp) != INTEGER_CST)
2351 : : {
2352 : 248 : if (flag_unsafe_math_optimizations
2353 : 26 : && TREE_CODE (base) == REAL_CST
2354 : 250 : && gimple_call_builtin_p (last_stmt, BUILT_IN_NORMAL))
2355 : : {
2356 : 2 : combined_fn log_cfn;
2357 : 2 : built_in_function exp_bfn;
2358 : 2 : switch (DECL_FUNCTION_CODE (gimple_call_fndecl (last_stmt)))
2359 : : {
2360 : : case BUILT_IN_POW:
2361 : : log_cfn = CFN_BUILT_IN_LOG;
2362 : : exp_bfn = BUILT_IN_EXP;
2363 : : break;
2364 : 0 : case BUILT_IN_POWF:
2365 : 0 : log_cfn = CFN_BUILT_IN_LOGF;
2366 : 0 : exp_bfn = BUILT_IN_EXPF;
2367 : 0 : break;
2368 : 0 : case BUILT_IN_POWL:
2369 : 0 : log_cfn = CFN_BUILT_IN_LOGL;
2370 : 0 : exp_bfn = BUILT_IN_EXPL;
2371 : 0 : break;
2372 : : default:
2373 : : return NULL;
2374 : : }
2375 : 2 : tree logc = fold_const_call (log_cfn, TREE_TYPE (base), base);
2376 : 2 : tree exp_decl = builtin_decl_implicit (exp_bfn);
2377 : : /* Optimize pow (C, x) as exp (log (C) * x). Normally match.pd
2378 : : does that, but if C is a power of 2, we want to use
2379 : : exp2 (log2 (C) * x) in the non-vectorized version, but for
2380 : : vectorization we don't have vectorized exp2. */
2381 : 2 : if (logc
2382 : 2 : && TREE_CODE (logc) == REAL_CST
2383 : 2 : && exp_decl
2384 : 4 : && lookup_attribute ("omp declare simd",
2385 : 2 : DECL_ATTRIBUTES (exp_decl)))
2386 : : {
2387 : 2 : cgraph_node *node = cgraph_node::get_create (exp_decl);
2388 : 2 : if (node->simd_clones == NULL)
2389 : : {
2390 : 2 : if (targetm.simd_clone.compute_vecsize_and_simdlen == NULL
2391 : 2 : || node->definition)
2392 : : return NULL;
2393 : 2 : expand_simd_clones (node);
2394 : 2 : if (node->simd_clones == NULL)
2395 : : return NULL;
2396 : : }
2397 : 2 : *type_out = get_vectype_for_scalar_type (vinfo, TREE_TYPE (base));
2398 : 2 : if (!*type_out)
2399 : : return NULL;
2400 : 2 : tree def = vect_recog_temp_ssa_var (TREE_TYPE (base), NULL);
2401 : 2 : gimple *g = gimple_build_assign (def, MULT_EXPR, exp, logc);
2402 : 2 : append_pattern_def_seq (vinfo, stmt_vinfo, g);
2403 : 2 : tree res = vect_recog_temp_ssa_var (TREE_TYPE (base), NULL);
2404 : 2 : g = gimple_build_call (exp_decl, 1, def);
2405 : 2 : gimple_call_set_lhs (g, res);
2406 : 2 : return g;
2407 : : }
2408 : : }
2409 : :
2410 : 246 : return NULL;
2411 : : }
2412 : :
2413 : : /* We now have a pow or powi builtin function call with a constant
2414 : : exponent. */
2415 : :
2416 : : /* Catch squaring. */
2417 : 26 : if ((tree_fits_shwi_p (exp)
2418 : 0 : && tree_to_shwi (exp) == 2)
2419 : 26 : || (TREE_CODE (exp) == REAL_CST
2420 : 26 : && real_equal (&TREE_REAL_CST (exp), &dconst2)))
2421 : : {
2422 : 0 : if (!vect_supportable_direct_optab_p (vinfo, TREE_TYPE (base), MULT_EXPR,
2423 : 0 : TREE_TYPE (base), type_out))
2424 : : return NULL;
2425 : :
2426 : 0 : var = vect_recog_temp_ssa_var (TREE_TYPE (base), NULL);
2427 : 0 : stmt = gimple_build_assign (var, MULT_EXPR, base, base);
2428 : 0 : return stmt;
2429 : : }
2430 : :
2431 : : /* Catch square root. */
2432 : 26 : if (TREE_CODE (exp) == REAL_CST
2433 : 26 : && real_equal (&TREE_REAL_CST (exp), &dconsthalf))
2434 : : {
2435 : 11 : *type_out = get_vectype_for_scalar_type (vinfo, TREE_TYPE (base));
2436 : 11 : if (*type_out
2437 : 11 : && direct_internal_fn_supported_p (IFN_SQRT, *type_out,
2438 : : OPTIMIZE_FOR_SPEED))
2439 : : {
2440 : 9 : gcall *stmt = gimple_build_call_internal (IFN_SQRT, 1, base);
2441 : 9 : var = vect_recog_temp_ssa_var (TREE_TYPE (base), stmt);
2442 : 9 : gimple_call_set_lhs (stmt, var);
2443 : 9 : gimple_call_set_nothrow (stmt, true);
2444 : 9 : return stmt;
2445 : : }
2446 : : }
2447 : :
2448 : : return NULL;
2449 : : }
2450 : :
2451 : :
2452 : : /* Function vect_recog_widen_sum_pattern
2453 : :
2454 : : Try to find the following pattern:
2455 : :
2456 : : type x_t;
2457 : : TYPE x_T, sum = init;
2458 : : loop:
2459 : : sum_0 = phi <init, sum_1>
2460 : : S1 x_t = *p;
2461 : : S2 x_T = (TYPE) x_t;
2462 : : S3 sum_1 = x_T + sum_0;
2463 : :
2464 : : where type 'TYPE' is at least double the size of type 'type', i.e - we're
2465 : : summing elements of type 'type' into an accumulator of type 'TYPE'. This is
2466 : : a special case of a reduction computation.
2467 : :
2468 : : Input:
2469 : :
2470 : : * STMT_VINFO: The stmt from which the pattern search begins. In the example,
2471 : : when this function is called with S3, the pattern {S2,S3} will be detected.
2472 : :
2473 : : Output:
2474 : :
2475 : : * TYPE_OUT: The type of the output of this pattern.
2476 : :
2477 : : * Return value: A new stmt that will be used to replace the sequence of
2478 : : stmts that constitute the pattern. In this case it will be:
2479 : : WIDEN_SUM <x_t, sum_0>
2480 : :
2481 : : Note: The widening-sum idiom is a widening reduction pattern that is
2482 : : vectorized without preserving all the intermediate results. It
2483 : : produces only N/2 (widened) results (by summing up pairs of
2484 : : intermediate results) rather than all N results. Therefore, we
2485 : : cannot allow this pattern when we want to get all the results and in
2486 : : the correct order (as is the case when this computation is in an
2487 : : inner-loop nested in an outer-loop that us being vectorized). */
2488 : :
2489 : : static gimple *
2490 : 27443823 : vect_recog_widen_sum_pattern (vec_info *vinfo,
2491 : : stmt_vec_info stmt_vinfo, tree *type_out)
2492 : : {
2493 : 27443823 : gimple *last_stmt = stmt_vinfo->stmt;
2494 : 27443823 : tree oprnd0, oprnd1;
2495 : 27443823 : tree type;
2496 : 27443823 : gimple *pattern_stmt;
2497 : 27443823 : tree var;
2498 : :
2499 : : /* Look for the following pattern
2500 : : DX = (TYPE) X;
2501 : : sum_1 = DX + sum_0;
2502 : : In which DX is at least double the size of X, and sum_1 has been
2503 : : recognized as a reduction variable.
2504 : : */
2505 : :
2506 : : /* Starting from LAST_STMT, follow the defs of its uses in search
2507 : : of the above pattern. */
2508 : :
2509 : 27443823 : if (!vect_reassociating_reduction_p (vinfo, stmt_vinfo, PLUS_EXPR,
2510 : : &oprnd0, &oprnd1)
2511 : 32235 : || TREE_CODE (oprnd0) != SSA_NAME
2512 : 27475881 : || !vinfo->lookup_def (oprnd0))
2513 : 27411823 : return NULL;
2514 : :
2515 : 32000 : type = TREE_TYPE (gimple_get_lhs (last_stmt));
2516 : :
2517 : : /* So far so good. Since last_stmt was detected as a (summation) reduction,
2518 : : we know that oprnd1 is the reduction variable (defined by a loop-header
2519 : : phi), and oprnd0 is an ssa-name defined by a stmt in the loop body.
2520 : : Left to check that oprnd0 is defined by a cast from type 'type' to type
2521 : : 'TYPE'. */
2522 : :
2523 : 32000 : vect_unpromoted_value unprom0;
2524 : 32000 : if (!vect_look_through_possible_promotion (vinfo, oprnd0, &unprom0)
2525 : 32000 : || TYPE_PRECISION (unprom0.type) * 2 > TYPE_PRECISION (type))
2526 : : return NULL;
2527 : :
2528 : 1692 : vect_pattern_detected ("vect_recog_widen_sum_pattern", last_stmt);
2529 : :
2530 : 1692 : if (!vect_supportable_direct_optab_p (vinfo, type, WIDEN_SUM_EXPR,
2531 : : unprom0.type, type_out))
2532 : : return NULL;
2533 : :
2534 : 0 : var = vect_recog_temp_ssa_var (type, NULL);
2535 : 0 : pattern_stmt = gimple_build_assign (var, WIDEN_SUM_EXPR, unprom0.op, oprnd1);
2536 : :
2537 : 0 : return pattern_stmt;
2538 : : }
2539 : :
2540 : : /* Function vect_recog_bitfield_ref_pattern
2541 : :
2542 : : Try to find the following pattern:
2543 : :
2544 : : bf_value = BIT_FIELD_REF (container, bitsize, bitpos);
2545 : : result = (type_out) bf_value;
2546 : :
2547 : : or
2548 : :
2549 : : if (BIT_FIELD_REF (container, bitsize, bitpos) `cmp` <constant>)
2550 : :
2551 : : where type_out is a non-bitfield type, that is to say, it's precision matches
2552 : : 2^(TYPE_SIZE(type_out) - (TYPE_UNSIGNED (type_out) ? 1 : 2)).
2553 : :
2554 : : Input:
2555 : :
2556 : : * STMT_VINFO: The stmt from which the pattern search begins.
2557 : : here it starts with:
2558 : : result = (type_out) bf_value;
2559 : :
2560 : : or
2561 : :
2562 : : if (BIT_FIELD_REF (container, bitsize, bitpos) `cmp` <constant>)
2563 : :
2564 : : Output:
2565 : :
2566 : : * TYPE_OUT: The vector type of the output of this pattern.
2567 : :
2568 : : * Return value: A new stmt that will be used to replace the sequence of
2569 : : stmts that constitute the pattern. If the precision of type_out is bigger
2570 : : than the precision type of _1 we perform the widening before the shifting,
2571 : : since the new precision will be large enough to shift the value and moving
2572 : : widening operations up the statement chain enables the generation of
2573 : : widening loads. If we are widening and the operation after the pattern is
2574 : : an addition then we mask first and shift later, to enable the generation of
2575 : : shifting adds. In the case of narrowing we will always mask first, shift
2576 : : last and then perform a narrowing operation. This will enable the
2577 : : generation of narrowing shifts.
2578 : :
2579 : : Widening with mask first, shift later:
2580 : : container = (type_out) container;
2581 : : masked = container & (((1 << bitsize) - 1) << bitpos);
2582 : : result = masked >> bitpos;
2583 : :
2584 : : Widening with shift first, mask last:
2585 : : container = (type_out) container;
2586 : : shifted = container >> bitpos;
2587 : : result = shifted & ((1 << bitsize) - 1);
2588 : :
2589 : : Narrowing:
2590 : : masked = container & (((1 << bitsize) - 1) << bitpos);
2591 : : result = masked >> bitpos;
2592 : : result = (type_out) result;
2593 : :
2594 : : If the bitfield is signed and it's wider than type_out, we need to
2595 : : keep the result sign-extended:
2596 : : container = (type) container;
2597 : : masked = container << (prec - bitsize - bitpos);
2598 : : result = (type_out) (masked >> (prec - bitsize));
2599 : :
2600 : : Here type is the signed variant of the wider of type_out and the type
2601 : : of container.
2602 : :
2603 : : The shifting is always optional depending on whether bitpos != 0.
2604 : :
2605 : : When the original bitfield was inside a gcond then an new gcond is also
2606 : : generated with the newly `result` as the operand to the comparison.
2607 : :
2608 : : */
2609 : :
2610 : : static gimple *
2611 : 27397249 : vect_recog_bitfield_ref_pattern (vec_info *vinfo, stmt_vec_info stmt_info,
2612 : : tree *type_out)
2613 : : {
2614 : 27397249 : gimple *bf_stmt = NULL;
2615 : 27397249 : tree lhs = NULL_TREE;
2616 : 27397249 : tree ret_type = NULL_TREE;
2617 : 27397249 : gimple *stmt = STMT_VINFO_STMT (stmt_info);
2618 : 27397249 : if (gcond *cond_stmt = dyn_cast <gcond *> (stmt))
2619 : : {
2620 : 4513539 : tree op = gimple_cond_lhs (cond_stmt);
2621 : 4513539 : if (TREE_CODE (op) != SSA_NAME)
2622 : : return NULL;
2623 : 4443878 : bf_stmt = dyn_cast <gassign *> (SSA_NAME_DEF_STMT (op));
2624 : 4443878 : if (TREE_CODE (gimple_cond_rhs (cond_stmt)) != INTEGER_CST)
2625 : : return NULL;
2626 : : }
2627 : 22883710 : else if (is_gimple_assign (stmt)
2628 : 18691700 : && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
2629 : 25377877 : && TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME)
2630 : : {
2631 : 2452377 : gimple *second_stmt = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt));
2632 : 2452377 : bf_stmt = dyn_cast <gassign *> (second_stmt);
2633 : 2452377 : lhs = gimple_assign_lhs (stmt);
2634 : 2452377 : ret_type = TREE_TYPE (lhs);
2635 : : }
2636 : :
2637 : 5447591 : if (!bf_stmt
2638 : 5447591 : || gimple_assign_rhs_code (bf_stmt) != BIT_FIELD_REF)
2639 : : return NULL;
2640 : :
2641 : 13862 : tree bf_ref = gimple_assign_rhs1 (bf_stmt);
2642 : 13862 : tree container = TREE_OPERAND (bf_ref, 0);
2643 : 13862 : ret_type = ret_type ? ret_type : TREE_TYPE (container);
2644 : :
2645 : 13862 : if (!bit_field_offset (bf_ref).is_constant ()
2646 : 13862 : || !bit_field_size (bf_ref).is_constant ()
2647 : 13862 : || !tree_fits_uhwi_p (TYPE_SIZE (TREE_TYPE (container))))
2648 : 0 : return NULL;
2649 : :
2650 : 27346 : if (!INTEGRAL_TYPE_P (TREE_TYPE (bf_ref))
2651 : 13860 : || !INTEGRAL_TYPE_P (TREE_TYPE (container))
2652 : 15837 : || TYPE_MODE (TREE_TYPE (container)) == E_BLKmode)
2653 : 11887 : return NULL;
2654 : :
2655 : 1975 : gimple *use_stmt, *pattern_stmt;
2656 : 1975 : use_operand_p use_p;
2657 : 1975 : bool shift_first = true;
2658 : 1975 : tree container_type = TREE_TYPE (container);
2659 : 1975 : tree vectype = get_vectype_for_scalar_type (vinfo, container_type);
2660 : :
2661 : : /* Calculate shift_n before the adjustments for widening loads, otherwise
2662 : : the container may change and we have to consider offset change for
2663 : : widening loads on big endianness. The shift_n calculated here can be
2664 : : independent of widening. */
2665 : 1975 : unsigned HOST_WIDE_INT shift_n = bit_field_offset (bf_ref).to_constant ();
2666 : 1975 : unsigned HOST_WIDE_INT mask_width = bit_field_size (bf_ref).to_constant ();
2667 : 1975 : unsigned HOST_WIDE_INT prec = tree_to_uhwi (TYPE_SIZE (container_type));
2668 : 1975 : if (BYTES_BIG_ENDIAN)
2669 : : shift_n = prec - shift_n - mask_width;
2670 : :
2671 : 1975 : bool ref_sext = (!TYPE_UNSIGNED (TREE_TYPE (bf_ref)) &&
2672 : 1353 : TYPE_PRECISION (ret_type) > mask_width);
2673 : 1975 : bool load_widen = (TYPE_PRECISION (TREE_TYPE (container)) <
2674 : 1975 : TYPE_PRECISION (ret_type));
2675 : :
2676 : : /* We move the conversion earlier if the loaded type is smaller than the
2677 : : return type to enable the use of widening loads. And if we need a
2678 : : sign extension, we need to convert the loaded value early to a signed
2679 : : type as well. */
2680 : 1975 : if (ref_sext || load_widen)
2681 : : {
2682 : 919 : tree type = load_widen ? ret_type : container_type;
2683 : 919 : if (ref_sext)
2684 : 891 : type = gimple_signed_type (type);
2685 : 919 : pattern_stmt = gimple_build_assign (vect_recog_temp_ssa_var (type),
2686 : : NOP_EXPR, container);
2687 : 919 : container = gimple_get_lhs (pattern_stmt);
2688 : 919 : container_type = TREE_TYPE (container);
2689 : 919 : prec = tree_to_uhwi (TYPE_SIZE (container_type));
2690 : 919 : vectype = get_vectype_for_scalar_type (vinfo, container_type);
2691 : 919 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt, vectype);
2692 : : }
2693 : 1056 : else if (!useless_type_conversion_p (TREE_TYPE (container), ret_type))
2694 : : /* If we are doing the conversion last then also delay the shift as we may
2695 : : be able to combine the shift and conversion in certain cases. */
2696 : : shift_first = false;
2697 : :
2698 : : /* If the only use of the result of this BIT_FIELD_REF + CONVERT is a
2699 : : PLUS_EXPR then do the shift last as some targets can combine the shift and
2700 : : add into a single instruction. */
2701 : 1272 : if (lhs && !is_pattern_stmt_p (stmt_info)
2702 : 3247 : && single_imm_use (lhs, &use_p, &use_stmt))
2703 : : {
2704 : 917 : if (gimple_code (use_stmt) == GIMPLE_ASSIGN
2705 : 917 : && gimple_assign_rhs_code (use_stmt) == PLUS_EXPR)
2706 : : shift_first = false;
2707 : : }
2708 : :
2709 : : /* If we don't have to shift we only generate the mask, so just fix the
2710 : : code-path to shift_first. */
2711 : 1975 : if (shift_n == 0)
2712 : 687 : shift_first = true;
2713 : :
2714 : 1975 : tree result;
2715 : 1975 : if (shift_first && !ref_sext)
2716 : : {
2717 : 426 : tree shifted = container;
2718 : 426 : if (shift_n)
2719 : : {
2720 : 53 : pattern_stmt
2721 : 53 : = gimple_build_assign (vect_recog_temp_ssa_var (container_type),
2722 : : RSHIFT_EXPR, container,
2723 : 53 : build_int_cst (sizetype, shift_n));
2724 : 53 : shifted = gimple_assign_lhs (pattern_stmt);
2725 : 53 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt, vectype);
2726 : : }
2727 : :
2728 : 426 : tree mask = wide_int_to_tree (container_type,
2729 : 426 : wi::mask (mask_width, false, prec));
2730 : :
2731 : 426 : pattern_stmt
2732 : 426 : = gimple_build_assign (vect_recog_temp_ssa_var (container_type),
2733 : : BIT_AND_EXPR, shifted, mask);
2734 : 426 : result = gimple_assign_lhs (pattern_stmt);
2735 : : }
2736 : : else
2737 : : {
2738 : 1549 : tree temp = vect_recog_temp_ssa_var (container_type);
2739 : 1549 : if (!ref_sext)
2740 : : {
2741 : 658 : tree mask = wide_int_to_tree (container_type,
2742 : 658 : wi::shifted_mask (shift_n,
2743 : : mask_width,
2744 : : false, prec));
2745 : 658 : pattern_stmt = gimple_build_assign (temp, BIT_AND_EXPR,
2746 : : container, mask);
2747 : : }
2748 : : else
2749 : : {
2750 : 891 : HOST_WIDE_INT shl = prec - shift_n - mask_width;
2751 : 891 : shift_n += shl;
2752 : 891 : pattern_stmt = gimple_build_assign (temp, LSHIFT_EXPR,
2753 : : container,
2754 : 891 : build_int_cst (sizetype,
2755 : : shl));
2756 : : }
2757 : :
2758 : 1549 : tree masked = gimple_assign_lhs (pattern_stmt);
2759 : 1549 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt, vectype);
2760 : 1549 : pattern_stmt
2761 : 1549 : = gimple_build_assign (vect_recog_temp_ssa_var (container_type),
2762 : : RSHIFT_EXPR, masked,
2763 : 1549 : build_int_cst (sizetype, shift_n));
2764 : 1549 : result = gimple_assign_lhs (pattern_stmt);
2765 : : }
2766 : :
2767 : 1975 : if (!useless_type_conversion_p (TREE_TYPE (result), ret_type))
2768 : : {
2769 : 1293 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt, vectype);
2770 : 1293 : pattern_stmt
2771 : 1293 : = gimple_build_assign (vect_recog_temp_ssa_var (ret_type),
2772 : : NOP_EXPR, result);
2773 : : }
2774 : :
2775 : 1975 : if (!lhs)
2776 : : {
2777 : 703 : if (!vectype)
2778 : : return NULL;
2779 : :
2780 : 595 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt, vectype);
2781 : 595 : vectype = truth_type_for (vectype);
2782 : :
2783 : : /* FIXME: This part extracts the boolean value out of the bitfield in the
2784 : : same way as vect_recog_gcond_pattern does. However because
2785 : : patterns cannot match the same root twice, when we handle and
2786 : : lower the bitfield in the gcond, vect_recog_gcond_pattern can't
2787 : : apply anymore. We should really fix it so that we don't need to
2788 : : duplicate transformations like these. */
2789 : 595 : tree new_lhs = vect_recog_temp_ssa_var (boolean_type_node, NULL);
2790 : 595 : gcond *cond_stmt = dyn_cast <gcond *> (stmt_info->stmt);
2791 : 595 : tree cond_cst = gimple_cond_rhs (cond_stmt);
2792 : 595 : gimple *new_stmt
2793 : 595 : = gimple_build_assign (new_lhs, gimple_cond_code (cond_stmt),
2794 : : gimple_get_lhs (pattern_stmt),
2795 : : fold_convert (container_type, cond_cst));
2796 : 595 : append_pattern_def_seq (vinfo, stmt_info, new_stmt, vectype, container_type);
2797 : 595 : pattern_stmt
2798 : 595 : = gimple_build_cond (NE_EXPR, new_lhs,
2799 : 595 : build_zero_cst (TREE_TYPE (new_lhs)),
2800 : : NULL_TREE, NULL_TREE);
2801 : : }
2802 : :
2803 : 1867 : *type_out = STMT_VINFO_VECTYPE (stmt_info);
2804 : 1867 : vect_pattern_detected ("bitfield_ref pattern", stmt_info->stmt);
2805 : :
2806 : 1867 : return pattern_stmt;
2807 : : }
2808 : :
2809 : : /* Function vect_recog_bit_insert_pattern
2810 : :
2811 : : Try to find the following pattern:
2812 : :
2813 : : written = BIT_INSERT_EXPR (container, value, bitpos);
2814 : :
2815 : : Input:
2816 : :
2817 : : * STMT_VINFO: The stmt we want to replace.
2818 : :
2819 : : Output:
2820 : :
2821 : : * TYPE_OUT: The vector type of the output of this pattern.
2822 : :
2823 : : * Return value: A new stmt that will be used to replace the sequence of
2824 : : stmts that constitute the pattern. In this case it will be:
2825 : : value = (container_type) value; // Make sure
2826 : : shifted = value << bitpos; // Shift value into place
2827 : : masked = shifted & (mask << bitpos); // Mask off the non-relevant bits in
2828 : : // the 'to-write value'.
2829 : : cleared = container & ~(mask << bitpos); // Clearing the bits we want to
2830 : : // write to from the value we want
2831 : : // to write to.
2832 : : written = cleared | masked; // Write bits.
2833 : :
2834 : :
2835 : : where mask = ((1 << TYPE_PRECISION (value)) - 1), a mask to keep the number of
2836 : : bits corresponding to the real size of the bitfield value we are writing to.
2837 : : The shifting is always optional depending on whether bitpos != 0.
2838 : :
2839 : : */
2840 : :
2841 : : static gimple *
2842 : 27400177 : vect_recog_bit_insert_pattern (vec_info *vinfo, stmt_vec_info stmt_info,
2843 : : tree *type_out)
2844 : : {
2845 : 27400177 : gassign *bf_stmt = dyn_cast <gassign *> (stmt_info->stmt);
2846 : 25130657 : if (!bf_stmt || gimple_assign_rhs_code (bf_stmt) != BIT_INSERT_EXPR)
2847 : : return NULL;
2848 : :
2849 : 604 : tree container = gimple_assign_rhs1 (bf_stmt);
2850 : 604 : tree value = gimple_assign_rhs2 (bf_stmt);
2851 : 604 : tree shift = gimple_assign_rhs3 (bf_stmt);
2852 : :
2853 : 604 : tree bf_type = TREE_TYPE (value);
2854 : 604 : tree container_type = TREE_TYPE (container);
2855 : :
2856 : 604 : if (!INTEGRAL_TYPE_P (container_type)
2857 : 604 : || !tree_fits_uhwi_p (TYPE_SIZE (container_type)))
2858 : : return NULL;
2859 : :
2860 : 500 : gimple *pattern_stmt;
2861 : :
2862 : 500 : vect_unpromoted_value unprom;
2863 : 500 : unprom.set_op (value, vect_internal_def);
2864 : 500 : value = vect_convert_input (vinfo, stmt_info, container_type, &unprom,
2865 : : get_vectype_for_scalar_type (vinfo,
2866 : : container_type));
2867 : :
2868 : 500 : unsigned HOST_WIDE_INT mask_width = TYPE_PRECISION (bf_type);
2869 : 500 : unsigned HOST_WIDE_INT prec = tree_to_uhwi (TYPE_SIZE (container_type));
2870 : 500 : unsigned HOST_WIDE_INT shift_n = tree_to_uhwi (shift);
2871 : 500 : if (BYTES_BIG_ENDIAN)
2872 : : {
2873 : : shift_n = prec - shift_n - mask_width;
2874 : : shift = build_int_cst (TREE_TYPE (shift), shift_n);
2875 : : }
2876 : :
2877 : 500 : if (!useless_type_conversion_p (TREE_TYPE (value), container_type))
2878 : : {
2879 : 0 : pattern_stmt =
2880 : 0 : gimple_build_assign (vect_recog_temp_ssa_var (container_type),
2881 : : NOP_EXPR, value);
2882 : 0 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt);
2883 : 0 : value = gimple_get_lhs (pattern_stmt);
2884 : : }
2885 : :
2886 : : /* Shift VALUE into place. */
2887 : 500 : tree shifted = value;
2888 : 500 : if (shift_n)
2889 : : {
2890 : 261 : gimple_seq stmts = NULL;
2891 : 261 : shifted
2892 : 261 : = gimple_build (&stmts, LSHIFT_EXPR, container_type, value, shift);
2893 : 261 : if (!gimple_seq_empty_p (stmts))
2894 : 112 : append_pattern_def_seq (vinfo, stmt_info,
2895 : : gimple_seq_first_stmt (stmts));
2896 : : }
2897 : :
2898 : 500 : tree mask_t
2899 : 500 : = wide_int_to_tree (container_type,
2900 : 500 : wi::shifted_mask (shift_n, mask_width, false, prec));
2901 : :
2902 : : /* Clear bits we don't want to write back from SHIFTED. */
2903 : 500 : gimple_seq stmts = NULL;
2904 : 500 : tree masked = gimple_build (&stmts, BIT_AND_EXPR, container_type, shifted,
2905 : : mask_t);
2906 : 500 : if (!gimple_seq_empty_p (stmts))
2907 : : {
2908 : 110 : pattern_stmt = gimple_seq_first_stmt (stmts);
2909 : 110 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt);
2910 : : }
2911 : :
2912 : : /* Mask off the bits in the container that we are to write to. */
2913 : 500 : mask_t = wide_int_to_tree (container_type,
2914 : 500 : wi::shifted_mask (shift_n, mask_width, true, prec));
2915 : 500 : tree cleared = vect_recog_temp_ssa_var (container_type);
2916 : 500 : pattern_stmt = gimple_build_assign (cleared, BIT_AND_EXPR, container, mask_t);
2917 : 500 : append_pattern_def_seq (vinfo, stmt_info, pattern_stmt);
2918 : :
2919 : : /* Write MASKED into CLEARED. */
2920 : 500 : pattern_stmt
2921 : 500 : = gimple_build_assign (vect_recog_temp_ssa_var (container_type),
2922 : : BIT_IOR_EXPR, cleared, masked);
2923 : :
2924 : 500 : *type_out = STMT_VINFO_VECTYPE (stmt_info);
2925 : 500 : vect_pattern_detected ("bit_insert pattern", stmt_info->stmt);
2926 : :
2927 : 500 : return pattern_stmt;
2928 : : }
2929 : :
2930 : :
2931 : : /* Recognize cases in which an operation is performed in one type WTYPE
2932 : : but could be done more efficiently in a narrower type NTYPE. For example,
2933 : : if we have:
2934 : :
2935 : : ATYPE a; // narrower than NTYPE
2936 : : BTYPE b; // narrower than NTYPE
2937 : : WTYPE aw = (WTYPE) a;
2938 : : WTYPE bw = (WTYPE) b;
2939 : : WTYPE res = aw + bw; // only uses of aw and bw
2940 : :
2941 : : then it would be more efficient to do:
2942 : :
2943 : : NTYPE an = (NTYPE) a;
2944 : : NTYPE bn = (NTYPE) b;
2945 : : NTYPE resn = an + bn;
2946 : : WTYPE res = (WTYPE) resn;
2947 : :
2948 : : Other situations include things like:
2949 : :
2950 : : ATYPE a; // NTYPE or narrower
2951 : : WTYPE aw = (WTYPE) a;
2952 : : WTYPE res = aw + b;
2953 : :
2954 : : when only "(NTYPE) res" is significant. In that case it's more efficient
2955 : : to truncate "b" and do the operation on NTYPE instead:
2956 : :
2957 : : NTYPE an = (NTYPE) a;
2958 : : NTYPE bn = (NTYPE) b; // truncation
2959 : : NTYPE resn = an + bn;
2960 : : WTYPE res = (WTYPE) resn;
2961 : :
2962 : : All users of "res" should then use "resn" instead, making the final
2963 : : statement dead (not marked as relevant). The final statement is still
2964 : : needed to maintain the type correctness of the IR.
2965 : :
2966 : : vect_determine_precisions has already determined the minimum
2967 : : precison of the operation and the minimum precision required
2968 : : by users of the result. */
2969 : :
2970 : : static gimple *
2971 : 27400617 : vect_recog_over_widening_pattern (vec_info *vinfo,
2972 : : stmt_vec_info last_stmt_info, tree *type_out)
2973 : : {
2974 : 27400617 : gassign *last_stmt = dyn_cast <gassign *> (last_stmt_info->stmt);
2975 : 18695663 : if (!last_stmt)
2976 : : return NULL;
2977 : :
2978 : : /* See whether we have found that this operation can be done on a
2979 : : narrower type without changing its semantics. */
2980 : 18695663 : unsigned int new_precision = last_stmt_info->operation_precision;
2981 : 18695663 : if (!new_precision)
2982 : : return NULL;
2983 : :
2984 : 1183312 : tree lhs = gimple_assign_lhs (last_stmt);
2985 : 1183312 : tree type = TREE_TYPE (lhs);
2986 : 1183312 : tree_code code = gimple_assign_rhs_code (last_stmt);
2987 : :
2988 : : /* Punt for reductions where we don't handle the type conversions. */
2989 : 1183312 : if (STMT_VINFO_DEF_TYPE (last_stmt_info) == vect_reduction_def)
2990 : : return NULL;
2991 : :
2992 : : /* Keep the first operand of a COND_EXPR as-is: only the other two
2993 : : operands are interesting. */
2994 : 1179900 : unsigned int first_op = (code == COND_EXPR ? 2 : 1);
2995 : :
2996 : : /* Check the operands. */
2997 : 1179900 : unsigned int nops = gimple_num_ops (last_stmt) - first_op;
2998 : 1179900 : auto_vec <vect_unpromoted_value, 3> unprom (nops);
2999 : 1179900 : unprom.quick_grow_cleared (nops);
3000 : 1179900 : unsigned int min_precision = 0;
3001 : 1179900 : bool single_use_p = false;
3002 : 3528361 : for (unsigned int i = 0; i < nops; ++i)
3003 : : {
3004 : 2348890 : tree op = gimple_op (last_stmt, first_op + i);
3005 : 2348890 : if (TREE_CODE (op) == INTEGER_CST)
3006 : 1050673 : unprom[i].set_op (op, vect_constant_def);
3007 : 1298217 : else if (TREE_CODE (op) == SSA_NAME)
3008 : : {
3009 : 1298217 : bool op_single_use_p = true;
3010 : 1298217 : if (!vect_look_through_possible_promotion (vinfo, op, &unprom[i],
3011 : : &op_single_use_p))
3012 : 429 : return NULL;
3013 : : /* If:
3014 : :
3015 : : (1) N bits of the result are needed;
3016 : : (2) all inputs are widened from M<N bits; and
3017 : : (3) one operand OP is a single-use SSA name
3018 : :
3019 : : we can shift the M->N widening from OP to the output
3020 : : without changing the number or type of extensions involved.
3021 : : This then reduces the number of copies of STMT_INFO.
3022 : :
3023 : : If instead of (3) more than one operand is a single-use SSA name,
3024 : : shifting the extension to the output is even more of a win.
3025 : :
3026 : : If instead:
3027 : :
3028 : : (1) N bits of the result are needed;
3029 : : (2) one operand OP2 is widened from M2<N bits;
3030 : : (3) another operand OP1 is widened from M1<M2 bits; and
3031 : : (4) both OP1 and OP2 are single-use
3032 : :
3033 : : the choice is between:
3034 : :
3035 : : (a) truncating OP2 to M1, doing the operation on M1,
3036 : : and then widening the result to N
3037 : :
3038 : : (b) widening OP1 to M2, doing the operation on M2, and then
3039 : : widening the result to N
3040 : :
3041 : : Both shift the M2->N widening of the inputs to the output.
3042 : : (a) additionally shifts the M1->M2 widening to the output;
3043 : : it requires fewer copies of STMT_INFO but requires an extra
3044 : : M2->M1 truncation.
3045 : :
3046 : : Which is better will depend on the complexity and cost of
3047 : : STMT_INFO, which is hard to predict at this stage. However,
3048 : : a clear tie-breaker in favor of (b) is the fact that the
3049 : : truncation in (a) increases the length of the operation chain.
3050 : :
3051 : : If instead of (4) only one of OP1 or OP2 is single-use,
3052 : : (b) is still a win over doing the operation in N bits:
3053 : : it still shifts the M2->N widening on the single-use operand
3054 : : to the output and reduces the number of STMT_INFO copies.
3055 : :
3056 : : If neither operand is single-use then operating on fewer than
3057 : : N bits might lead to more extensions overall. Whether it does
3058 : : or not depends on global information about the vectorization
3059 : : region, and whether that's a good trade-off would again
3060 : : depend on the complexity and cost of the statements involved,
3061 : : as well as things like register pressure that are not normally
3062 : : modelled at this stage. We therefore ignore these cases
3063 : : and just optimize the clear single-use wins above.
3064 : :
3065 : : Thus we take the maximum precision of the unpromoted operands
3066 : : and record whether any operand is single-use. */
3067 : 1297788 : if (unprom[i].dt == vect_internal_def)
3068 : : {
3069 : 932211 : min_precision = MAX (min_precision,
3070 : : TYPE_PRECISION (unprom[i].type));
3071 : 932211 : single_use_p |= op_single_use_p;
3072 : : }
3073 : : }
3074 : : else
3075 : : return NULL;
3076 : : }
3077 : :
3078 : : /* Although the operation could be done in operation_precision, we have
3079 : : to balance that against introducing extra truncations or extensions.
3080 : : Calculate the minimum precision that can be handled efficiently.
3081 : :
3082 : : The loop above determined that the operation could be handled
3083 : : efficiently in MIN_PRECISION if SINGLE_USE_P; this would shift an
3084 : : extension from the inputs to the output without introducing more
3085 : : instructions, and would reduce the number of instructions required
3086 : : for STMT_INFO itself.
3087 : :
3088 : : vect_determine_precisions has also determined that the result only
3089 : : needs min_output_precision bits. Truncating by a factor of N times
3090 : : requires a tree of N - 1 instructions, so if TYPE is N times wider
3091 : : than min_output_precision, doing the operation in TYPE and truncating
3092 : : the result requires N + (N - 1) = 2N - 1 instructions per output vector.
3093 : : In contrast:
3094 : :
3095 : : - truncating the input to a unary operation and doing the operation
3096 : : in the new type requires at most N - 1 + 1 = N instructions per
3097 : : output vector
3098 : :
3099 : : - doing the same for a binary operation requires at most
3100 : : (N - 1) * 2 + 1 = 2N - 1 instructions per output vector
3101 : :
3102 : : Both unary and binary operations require fewer instructions than
3103 : : this if the operands were extended from a suitable truncated form.
3104 : : Thus there is usually nothing to lose by doing operations in
3105 : : min_output_precision bits, but there can be something to gain. */
3106 : 1179471 : if (!single_use_p)
3107 : 869929 : min_precision = last_stmt_info->min_output_precision;
3108 : : else
3109 : 309542 : min_precision = MIN (min_precision, last_stmt_info->min_output_precision);
3110 : :
3111 : : /* Apply the minimum efficient precision we just calculated. */
3112 : 1179471 : if (new_precision < min_precision)
3113 : : new_precision = min_precision;
3114 : 1179471 : new_precision = vect_element_precision (new_precision);
3115 : 1179471 : if (new_precision >= TYPE_PRECISION (type))
3116 : : return NULL;
3117 : :
3118 : 130585 : vect_pattern_detected ("vect_recog_over_widening_pattern", last_stmt);
3119 : :
3120 : 130585 : *type_out = get_vectype_for_scalar_type (vinfo, type);
3121 : 130585 : if (!*type_out)
3122 : : return NULL;
3123 : :
3124 : : /* We've found a viable pattern. Get the new type of the operation. */
3125 : 116326 : bool unsigned_p = (last_stmt_info->operation_sign == UNSIGNED);
3126 : 116326 : tree new_type = build_nonstandard_integer_type (new_precision, unsigned_p);
3127 : :
3128 : : /* If we're truncating an operation, we need to make sure that we
3129 : : don't introduce new undefined overflow. The codes tested here are
3130 : : a subset of those accepted by vect_truncatable_operation_p. */
3131 : 116326 : tree op_type = new_type;
3132 : 116326 : if (TYPE_OVERFLOW_UNDEFINED (new_type)
3133 : 148330 : && (code == PLUS_EXPR || code == MINUS_EXPR || code == MULT_EXPR))
3134 : 24394 : op_type = build_nonstandard_integer_type (new_precision, true);
3135 : :
3136 : : /* We specifically don't check here whether the target supports the
3137 : : new operation, since it might be something that a later pattern
3138 : : wants to rewrite anyway. If targets have a minimum element size
3139 : : for some optabs, we should pattern-match smaller ops to larger ops
3140 : : where beneficial. */
3141 : 116326 : tree new_vectype = get_vectype_for_scalar_type (vinfo, new_type);
3142 : 116326 : tree op_vectype = get_vectype_for_scalar_type (vinfo, op_type);
3143 : 116326 : if (!new_vectype || !op_vectype)
3144 : : return NULL;
3145 : :
3146 : 116326 : if (dump_enabled_p ())
3147 : 4035 : dump_printf_loc (MSG_NOTE, vect_location, "demoting %T to %T\n",
3148 : : type, new_type);
3149 : :
3150 : : /* Calculate the rhs operands for an operation on OP_TYPE. */
3151 : 116326 : tree ops[3] = {};
3152 : 116535 : for (unsigned int i = 1; i < first_op; ++i)
3153 : 209 : ops[i - 1] = gimple_op (last_stmt, i);
3154 : 116326 : vect_convert_inputs (vinfo, last_stmt_info, nops, &ops[first_op - 1],
3155 : 116326 : op_type, &unprom[0], op_vectype);
3156 : :
3157 : : /* Use the operation to produce a result of type OP_TYPE. */
3158 : 116326 : tree new_var = vect_recog_temp_ssa_var (op_type, NULL);
3159 : 116326 : gimple *pattern_stmt = gimple_build_assign (new_var, code,
3160 : : ops[0], ops[1], ops[2]);
3161 : 116326 : gimple_set_location (pattern_stmt, gimple_location (last_stmt));
3162 : :
3163 : 116326 : if (dump_enabled_p ())
3164 : 4035 : dump_printf_loc (MSG_NOTE, vect_location,
3165 : : "created pattern stmt: %G", pattern_stmt);
3166 : :
3167 : : /* Convert back to the original signedness, if OP_TYPE is different
3168 : : from NEW_TYPE. */
3169 : 116326 : if (op_type != new_type)
3170 : 24394 : pattern_stmt = vect_convert_output (vinfo, last_stmt_info, new_type,
3171 : : pattern_stmt, op_vectype);
3172 : :
3173 : : /* Promote the result to the original type. */
3174 : 116326 : pattern_stmt = vect_convert_output (vinfo, last_stmt_info, type,
3175 : : pattern_stmt, new_vectype);
3176 : :
3177 : 116326 : return pattern_stmt;
3178 : 1179900 : }
3179 : :
3180 : : /* Recognize the following patterns:
3181 : :
3182 : : ATYPE a; // narrower than TYPE
3183 : : BTYPE b; // narrower than TYPE
3184 : :
3185 : : 1) Multiply high with scaling
3186 : : TYPE res = ((TYPE) a * (TYPE) b) >> c;
3187 : : Here, c is bitsize (TYPE) / 2 - 1.
3188 : :
3189 : : 2) ... or also with rounding
3190 : : TYPE res = (((TYPE) a * (TYPE) b) >> d + 1) >> 1;
3191 : : Here, d is bitsize (TYPE) / 2 - 2.
3192 : :
3193 : : 3) Normal multiply high
3194 : : TYPE res = ((TYPE) a * (TYPE) b) >> e;
3195 : : Here, e is bitsize (TYPE) / 2.
3196 : :
3197 : : where only the bottom half of res is used. */
3198 : :
3199 : : static gimple *
3200 : 27504606 : vect_recog_mulhs_pattern (vec_info *vinfo,
3201 : : stmt_vec_info last_stmt_info, tree *type_out)
3202 : : {
3203 : : /* Check for a right shift. */
3204 : 27504606 : gassign *last_stmt = dyn_cast <gassign *> (last_stmt_info->stmt);
3205 : 18799530 : if (!last_stmt
3206 : 18799530 : || gimple_assign_rhs_code (last_stmt) != RSHIFT_EXPR)
3207 : : return NULL;
3208 : :
3209 : : /* Check that the shift result is wider than the users of the
3210 : : result need (i.e. that narrowing would be a natural choice). */
3211 : 302396 : tree lhs_type = TREE_TYPE (gimple_assign_lhs (last_stmt));
3212 : 302396 : unsigned int target_precision
3213 : 302396 : = vect_element_precision (last_stmt_info->min_output_precision);
3214 : 302396 : if (!INTEGRAL_TYPE_P (lhs_type)
3215 : 302396 : || target_precision >= TYPE_PRECISION (lhs_type))
3216 : : return NULL;
3217 : :
3218 : : /* Look through any change in sign on the outer shift input. */
3219 : 40094 : vect_unpromoted_value unprom_rshift_input;
3220 : 40094 : tree rshift_input = vect_look_through_possible_promotion
3221 : 40094 : (vinfo, gimple_assign_rhs1 (last_stmt), &unprom_rshift_input);
3222 : 40094 : if (!rshift_input
3223 : 40094 : || TYPE_PRECISION (TREE_TYPE (rshift_input))
3224 : 39473 : != TYPE_PRECISION (lhs_type))
3225 : : return NULL;
3226 : :
3227 : : /* Get the definition of the shift input. */
3228 : 37547 : stmt_vec_info rshift_input_stmt_info
3229 : 37547 : = vect_get_internal_def (vinfo, rshift_input);
3230 : 37547 : if (!rshift_input_stmt_info)
3231 : : return NULL;
3232 : 33499 : gassign *rshift_input_stmt
3233 : 27533917 : = dyn_cast <gassign *> (rshift_input_stmt_info->stmt);
3234 : 29380 : if (!rshift_input_stmt)
3235 : : return NULL;
3236 : :
3237 : 29380 : stmt_vec_info mulh_stmt_info;
3238 : 29380 : tree scale_term;
3239 : 29380 : bool rounding_p = false;
3240 : :
3241 : : /* Check for the presence of the rounding term. */
3242 : 35502 : if (gimple_assign_rhs_code (rshift_input_stmt) == PLUS_EXPR)
3243 : : {
3244 : : /* Check that the outer shift was by 1. */
3245 : 16554 : if (!integer_onep (gimple_assign_rhs2 (last_stmt)))
3246 : 8229 : return NULL;
3247 : :
3248 : : /* Check that the second operand of the PLUS_EXPR is 1. */
3249 : 1326 : if (!integer_onep (gimple_assign_rhs2 (rshift_input_stmt)))
3250 : : return NULL;
3251 : :
3252 : : /* Look through any change in sign on the addition input. */
3253 : 91 : vect_unpromoted_value unprom_plus_input;
3254 : 91 : tree plus_input = vect_look_through_possible_promotion
3255 : 91 : (vinfo, gimple_assign_rhs1 (rshift_input_stmt), &unprom_plus_input);
3256 : 91 : if (!plus_input
3257 : 91 : || TYPE_PRECISION (TREE_TYPE (plus_input))
3258 : 91 : != TYPE_PRECISION (TREE_TYPE (rshift_input)))
3259 : : return NULL;
3260 : :
3261 : : /* Get the definition of the multiply-high-scale part. */
3262 : 91 : stmt_vec_info plus_input_stmt_info
3263 : 91 : = vect_get_internal_def (vinfo, plus_input);
3264 : 91 : if (!plus_input_stmt_info)
3265 : : return NULL;
3266 : 91 : gassign *plus_input_stmt
3267 : 8320 : = dyn_cast <gassign *> (plus_input_stmt_info->stmt);
3268 : 91 : if (!plus_input_stmt
3269 : 91 : || gimple_assign_rhs_code (plus_input_stmt) != RSHIFT_EXPR)
3270 : : return NULL;
3271 : :
3272 : : /* Look through any change in sign on the scaling input. */
3273 : 48 : vect_unpromoted_value unprom_scale_input;
3274 : 48 : tree scale_input = vect_look_through_possible_promotion
3275 : 48 : (vinfo, gimple_assign_rhs1 (plus_input_stmt), &unprom_scale_input);
3276 : 48 : if (!scale_input
3277 : 48 : || TYPE_PRECISION (TREE_TYPE (scale_input))
3278 : 48 : != TYPE_PRECISION (TREE_TYPE (plus_input)))
3279 : : return NULL;
3280 : :
3281 : : /* Get the definition of the multiply-high part. */
3282 : 48 : mulh_stmt_info = vect_get_internal_def (vinfo, scale_input);
3283 : 48 : if (!mulh_stmt_info)
3284 : : return NULL;
3285 : :
3286 : : /* Get the scaling term. */
3287 : 48 : scale_term = gimple_assign_rhs2 (plus_input_stmt);
3288 : 48 : rounding_p = true;
3289 : : }
3290 : : else
3291 : : {
3292 : 21103 : mulh_stmt_info = rshift_input_stmt_info;
3293 : 21103 : scale_term = gimple_assign_rhs2 (last_stmt);
3294 : : }
3295 : :
3296 : : /* Check that the scaling factor is constant. */
3297 : 21151 : if (TREE_CODE (scale_term) != INTEGER_CST)
3298 : : return NULL;
3299 : :
3300 : : /* Check whether the scaling input term can be seen as two widened
3301 : : inputs multiplied together. */
3302 : 60615 : vect_unpromoted_value unprom_mult[2];
3303 : 20205 : tree new_type;
3304 : 20205 : unsigned int nops
3305 : 20205 : = vect_widened_op_tree (vinfo, mulh_stmt_info, MULT_EXPR, WIDEN_MULT_EXPR,
3306 : : false, 2, unprom_mult, &new_type);
3307 : 20205 : if (nops != 2)
3308 : : return NULL;
3309 : :
3310 : : /* Adjust output precision. */
3311 : 2479 : if (TYPE_PRECISION (new_type) < target_precision)
3312 : 0 : new_type = build_nonstandard_integer_type
3313 : 0 : (target_precision, TYPE_UNSIGNED (new_type));
3314 : :
3315 : 2479 : unsigned mult_precision = TYPE_PRECISION (new_type);
3316 : 2479 : internal_fn ifn;
3317 : : /* Check that the scaling factor is expected. Instead of
3318 : : target_precision, we should use the one that we actually
3319 : : use for internal function. */
3320 : 2479 : if (rounding_p)
3321 : : {
3322 : : /* Check pattern 2). */
3323 : 96 : if (wi::to_widest (scale_term) + mult_precision + 2
3324 : 144 : != TYPE_PRECISION (lhs_type))
3325 : : return NULL;
3326 : :
3327 : : ifn = IFN_MULHRS;
3328 : : }
3329 : : else
3330 : : {
3331 : : /* Check for pattern 1). */
3332 : 4862 : if (wi::to_widest (scale_term) + mult_precision + 1
3333 : 7293 : == TYPE_PRECISION (lhs_type))
3334 : : ifn = IFN_MULHS;
3335 : : /* Check for pattern 3). */
3336 : 2387 : else if (wi::to_widest (scale_term) + mult_precision
3337 : 4774 : == TYPE_PRECISION (lhs_type))
3338 : : ifn = IFN_MULH;
3339 : : else
3340 : : return NULL;
3341 : : }
3342 : :
3343 : 2407 : vect_pattern_detected ("vect_recog_mulhs_pattern", last_stmt);
3344 : :
3345 : : /* Check for target support. */
3346 : 2407 : tree new_vectype = get_vectype_for_scalar_type (vinfo, new_type);
3347 : 2407 : if (!new_vectype
3348 : 4800 : || !direct_internal_fn_supported_p
3349 : 2393 : (ifn, new_vectype, OPTIMIZE_FOR_SPEED))
3350 : 2338 : return NULL;
3351 : :
3352 : : /* The IR requires a valid vector type for the cast result, even though
3353 : : it's likely to be discarded. */
3354 : 69 : *type_out = get_vectype_for_scalar_type (vinfo, lhs_type);
3355 : 69 : if (!*type_out)
3356 : : return NULL;
3357 : :
3358 : : /* Generate the IFN_MULHRS call. */
3359 : 69 : tree new_var = vect_recog_temp_ssa_var (new_type, NULL);
3360 : 69 : tree new_ops[2];
3361 : 69 : vect_convert_inputs (vinfo, last_stmt_info, 2, new_ops, new_type,
3362 : : unprom_mult, new_vectype);
3363 : 69 : gcall *mulhrs_stmt
3364 : 69 : = gimple_build_call_internal (ifn, 2, new_ops[0], new_ops[1]);
3365 : 69 : gimple_call_set_lhs (mulhrs_stmt, new_var);
3366 : 69 : gimple_set_location (mulhrs_stmt, gimple_location (last_stmt));
3367 : :
3368 : 69 : if (dump_enabled_p ())
3369 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
3370 : : "created pattern stmt: %G", (gimple *) mulhrs_stmt);
3371 : :
3372 : 69 : return vect_convert_output (vinfo, last_stmt_info, lhs_type,
3373 : 69 : mulhrs_stmt, new_vectype);
3374 : : }
3375 : :
3376 : : /* Recognize the patterns:
3377 : :
3378 : : ATYPE a; // narrower than TYPE
3379 : : BTYPE b; // narrower than TYPE
3380 : : (1) TYPE avg = ((TYPE) a + (TYPE) b) >> 1;
3381 : : or (2) TYPE avg = ((TYPE) a + (TYPE) b + 1) >> 1;
3382 : :
3383 : : where only the bottom half of avg is used. Try to transform them into:
3384 : :
3385 : : (1) NTYPE avg' = .AVG_FLOOR ((NTYPE) a, (NTYPE) b);
3386 : : or (2) NTYPE avg' = .AVG_CEIL ((NTYPE) a, (NTYPE) b);
3387 : :
3388 : : followed by:
3389 : :
3390 : : TYPE avg = (TYPE) avg';
3391 : :
3392 : : where NTYPE is no wider than half of TYPE. Since only the bottom half
3393 : : of avg is used, all or part of the cast of avg' should become redundant.
3394 : :
3395 : : If there is no target support available, generate code to distribute rshift
3396 : : over plus and add a carry. */
3397 : :
3398 : : static gimple *
3399 : 27502875 : vect_recog_average_pattern (vec_info *vinfo,
3400 : : stmt_vec_info last_stmt_info, tree *type_out)
3401 : : {
3402 : : /* Check for a shift right by one bit. */
3403 : 27502875 : gassign *last_stmt = dyn_cast <gassign *> (last_stmt_info->stmt);
3404 : 18797921 : if (!last_stmt
3405 : 18797921 : || gimple_assign_rhs_code (last_stmt) != RSHIFT_EXPR
3406 : 302252 : || !integer_onep (gimple_assign_rhs2 (last_stmt)))
3407 : 27458032 : return NULL;
3408 : :
3409 : : /* Check that the shift result is wider than the users of the
3410 : : result need (i.e. that narrowing would be a natural choice). */
3411 : 44843 : tree lhs = gimple_assign_lhs (last_stmt);
3412 : 44843 : tree type = TREE_TYPE (lhs);
3413 : 44843 : unsigned int target_precision
3414 : 44843 : = vect_element_precision (last_stmt_info->min_output_precision);
3415 : 44843 : if (!INTEGRAL_TYPE_P (type) || target_precision >= TYPE_PRECISION (type))
3416 : : return NULL;
3417 : :
3418 : : /* Look through any change in sign on the shift input. */
3419 : 2303 : tree rshift_rhs = gimple_assign_rhs1 (last_stmt);
3420 : 2303 : vect_unpromoted_value unprom_plus;
3421 : 2303 : rshift_rhs = vect_look_through_possible_promotion (vinfo, rshift_rhs,
3422 : : &unprom_plus);
3423 : 2303 : if (!rshift_rhs
3424 : 2303 : || TYPE_PRECISION (TREE_TYPE (rshift_rhs)) != TYPE_PRECISION (type))
3425 : : return NULL;
3426 : :
3427 : : /* Get the definition of the shift input. */
3428 : 2293 : stmt_vec_info plus_stmt_info = vect_get_internal_def (vinfo, rshift_rhs);
3429 : 2293 : if (!plus_stmt_info)
3430 : : return NULL;
3431 : :
3432 : : /* Check whether the shift input can be seen as a tree of additions on
3433 : : 2 or 3 widened inputs.
3434 : :
3435 : : Note that the pattern should be a win even if the result of one or
3436 : : more additions is reused elsewhere: if the pattern matches, we'd be
3437 : : replacing 2N RSHIFT_EXPRs and N VEC_PACK_*s with N IFN_AVG_*s. */
3438 : 9100 : internal_fn ifn = IFN_AVG_FLOOR;
3439 : 9100 : vect_unpromoted_value unprom[3];
3440 : 2275 : tree new_type;
3441 : 2275 : unsigned int nops = vect_widened_op_tree (vinfo, plus_stmt_info, PLUS_EXPR,
3442 : : IFN_VEC_WIDEN_PLUS, false, 3,
3443 : : unprom, &new_type);
3444 : 2275 : if (nops == 0)
3445 : : return NULL;
3446 : 912 : if (nops == 3)
3447 : : {
3448 : : /* Check that one operand is 1. */
3449 : : unsigned int i;
3450 : 921 : for (i = 0; i < 3; ++i)
3451 : 861 : if (integer_onep (unprom[i].op))
3452 : : break;
3453 : 287 : if (i == 3)
3454 : : return NULL;
3455 : : /* Throw away the 1 operand and keep the other two. */
3456 : 227 : if (i < 2)
3457 : 0 : unprom[i] = unprom[2];
3458 : : ifn = IFN_AVG_CEIL;
3459 : : }
3460 : :
3461 : 852 : vect_pattern_detected ("vect_recog_average_pattern", last_stmt);
3462 : :
3463 : : /* We know that:
3464 : :
3465 : : (a) the operation can be viewed as:
3466 : :
3467 : : TYPE widened0 = (TYPE) UNPROM[0];
3468 : : TYPE widened1 = (TYPE) UNPROM[1];
3469 : : TYPE tmp1 = widened0 + widened1 {+ 1};
3470 : : TYPE tmp2 = tmp1 >> 1; // LAST_STMT_INFO
3471 : :
3472 : : (b) the first two statements are equivalent to:
3473 : :
3474 : : TYPE widened0 = (TYPE) (NEW_TYPE) UNPROM[0];
3475 : : TYPE widened1 = (TYPE) (NEW_TYPE) UNPROM[1];
3476 : :
3477 : : (c) vect_recog_over_widening_pattern has already tried to narrow TYPE
3478 : : where sensible;
3479 : :
3480 : : (d) all the operations can be performed correctly at twice the width of
3481 : : NEW_TYPE, due to the nature of the average operation; and
3482 : :
3483 : : (e) users of the result of the right shift need only TARGET_PRECISION
3484 : : bits, where TARGET_PRECISION is no more than half of TYPE's
3485 : : precision.
3486 : :
3487 : : Under these circumstances, the only situation in which NEW_TYPE
3488 : : could be narrower than TARGET_PRECISION is if widened0, widened1
3489 : : and an addition result are all used more than once. Thus we can
3490 : : treat any widening of UNPROM[0] and UNPROM[1] to TARGET_PRECISION
3491 : : as "free", whereas widening the result of the average instruction
3492 : : from NEW_TYPE to TARGET_PRECISION would be a new operation. It's
3493 : : therefore better not to go narrower than TARGET_PRECISION. */
3494 : 852 : if (TYPE_PRECISION (new_type) < target_precision)
3495 : 8 : new_type = build_nonstandard_integer_type (target_precision,
3496 : 8 : TYPE_UNSIGNED (new_type));
3497 : :
3498 : : /* Check for target support. */
3499 : 852 : tree new_vectype = get_vectype_for_scalar_type (vinfo, new_type);
3500 : 852 : if (!new_vectype)
3501 : : return NULL;
3502 : :
3503 : 852 : bool fallback_p = false;
3504 : :
3505 : 852 : if (direct_internal_fn_supported_p (ifn, new_vectype, OPTIMIZE_FOR_SPEED))
3506 : : ;
3507 : 730 : else if (TYPE_UNSIGNED (new_type)
3508 : 267 : && optab_for_tree_code (RSHIFT_EXPR, new_vectype, optab_scalar)
3509 : 267 : && optab_for_tree_code (PLUS_EXPR, new_vectype, optab_default)
3510 : 267 : && optab_for_tree_code (BIT_IOR_EXPR, new_vectype, optab_default)
3511 : 997 : && optab_for_tree_code (BIT_AND_EXPR, new_vectype, optab_default))
3512 : : fallback_p = true;
3513 : : else
3514 : 463 : return NULL;
3515 : :
3516 : : /* The IR requires a valid vector type for the cast result, even though
3517 : : it's likely to be discarded. */
3518 : 389 : *type_out = get_vectype_for_scalar_type (vinfo, type);
3519 : 389 : if (!*type_out)
3520 : : return NULL;
3521 : :
3522 : 388 : tree new_var = vect_recog_temp_ssa_var (new_type, NULL);
3523 : 388 : tree new_ops[2];
3524 : 388 : vect_convert_inputs (vinfo, last_stmt_info, 2, new_ops, new_type,
3525 : : unprom, new_vectype);
3526 : :
3527 : 388 : if (fallback_p)
3528 : : {
3529 : : /* As a fallback, generate code for following sequence:
3530 : :
3531 : : shifted_op0 = new_ops[0] >> 1;
3532 : : shifted_op1 = new_ops[1] >> 1;
3533 : : sum_of_shifted = shifted_op0 + shifted_op1;
3534 : : unmasked_carry = new_ops[0] and/or new_ops[1];
3535 : : carry = unmasked_carry & 1;
3536 : : new_var = sum_of_shifted + carry;
3537 : : */
3538 : :
3539 : 266 : tree one_cst = build_one_cst (new_type);
3540 : 266 : gassign *g;
3541 : :
3542 : 266 : tree shifted_op0 = vect_recog_temp_ssa_var (new_type, NULL);
3543 : 266 : g = gimple_build_assign (shifted_op0, RSHIFT_EXPR, new_ops[0], one_cst);
3544 : 266 : append_pattern_def_seq (vinfo, last_stmt_info, g, new_vectype);
3545 : :
3546 : 266 : tree shifted_op1 = vect_recog_temp_ssa_var (new_type, NULL);
3547 : 266 : g = gimple_build_assign (shifted_op1, RSHIFT_EXPR, new_ops[1], one_cst);
3548 : 266 : append_pattern_def_seq (vinfo, last_stmt_info, g, new_vectype);
3549 : :
3550 : 266 : tree sum_of_shifted = vect_recog_temp_ssa_var (new_type, NULL);
3551 : 266 : g = gimple_build_assign (sum_of_shifted, PLUS_EXPR,
3552 : : shifted_op0, shifted_op1);
3553 : 266 : append_pattern_def_seq (vinfo, last_stmt_info, g, new_vectype);
3554 : :
3555 : 266 : tree unmasked_carry = vect_recog_temp_ssa_var (new_type, NULL);
3556 : 266 : tree_code c = (ifn == IFN_AVG_CEIL) ? BIT_IOR_EXPR : BIT_AND_EXPR;
3557 : 266 : g = gimple_build_assign (unmasked_carry, c, new_ops[0], new_ops[1]);
3558 : 266 : append_pattern_def_seq (vinfo, last_stmt_info, g, new_vectype);
3559 : :
3560 : 266 : tree carry = vect_recog_temp_ssa_var (new_type, NULL);
3561 : 266 : g = gimple_build_assign (carry, BIT_AND_EXPR, unmasked_carry, one_cst);
3562 : 266 : append_pattern_def_seq (vinfo, last_stmt_info, g, new_vectype);
3563 : :
3564 : 266 : g = gimple_build_assign (new_var, PLUS_EXPR, sum_of_shifted, carry);
3565 : 266 : return vect_convert_output (vinfo, last_stmt_info, type, g, new_vectype);
3566 : : }
3567 : :
3568 : : /* Generate the IFN_AVG* call. */
3569 : 122 : gcall *average_stmt = gimple_build_call_internal (ifn, 2, new_ops[0],
3570 : : new_ops[1]);
3571 : 122 : gimple_call_set_lhs (average_stmt, new_var);
3572 : 122 : gimple_set_location (average_stmt, gimple_location (last_stmt));
3573 : :
3574 : 122 : if (dump_enabled_p ())
3575 : 42 : dump_printf_loc (MSG_NOTE, vect_location,
3576 : : "created pattern stmt: %G", (gimple *) average_stmt);
3577 : :
3578 : 122 : return vect_convert_output (vinfo, last_stmt_info,
3579 : 122 : type, average_stmt, new_vectype);
3580 : : }
3581 : :
3582 : : /* Recognize cases in which the input to a cast is wider than its
3583 : : output, and the input is fed by a widening operation. Fold this
3584 : : by removing the unnecessary intermediate widening. E.g.:
3585 : :
3586 : : unsigned char a;
3587 : : unsigned int b = (unsigned int) a;
3588 : : unsigned short c = (unsigned short) b;
3589 : :
3590 : : -->
3591 : :
3592 : : unsigned short c = (unsigned short) a;
3593 : :
3594 : : Although this is rare in input IR, it is an expected side-effect
3595 : : of the over-widening pattern above.
3596 : :
3597 : : This is beneficial also for integer-to-float conversions, if the
3598 : : widened integer has more bits than the float, and if the unwidened
3599 : : input doesn't. */
3600 : :
3601 : : static gimple *
3602 : 27504606 : vect_recog_cast_forwprop_pattern (vec_info *vinfo,
3603 : : stmt_vec_info last_stmt_info, tree *type_out)
3604 : : {
3605 : : /* Check for a cast, including an integer-to-float conversion. */
3606 : 46265221 : gassign *last_stmt = dyn_cast <gassign *> (last_stmt_info->stmt);
3607 : 18799461 : if (!last_stmt)
3608 : : return NULL;
3609 : 18799461 : tree_code code = gimple_assign_rhs_code (last_stmt);
3610 : 18799461 : if (!CONVERT_EXPR_CODE_P (code) && code != FLOAT_EXPR)
3611 : : return NULL;
3612 : :
3613 : : /* Make sure that the rhs is a scalar with a natural bitsize. */
3614 : 2652566 : tree lhs = gimple_assign_lhs (last_stmt);
3615 : 2652566 : if (!lhs)
3616 : : return NULL;
3617 : 2652566 : tree lhs_type = TREE_TYPE (lhs);
3618 : 2652566 : scalar_mode lhs_mode;
3619 : 2635401 : if (VECT_SCALAR_BOOLEAN_TYPE_P (lhs_type)
3620 : 5286274 : || !is_a <scalar_mode> (TYPE_MODE (lhs_type), &lhs_mode))
3621 : 22070 : return NULL;
3622 : :
3623 : : /* Check for a narrowing operation (from a vector point of view). */
3624 : 2630496 : tree rhs = gimple_assign_rhs1 (last_stmt);
3625 : 2630496 : tree rhs_type = TREE_TYPE (rhs);
3626 : 2630496 : if (!INTEGRAL_TYPE_P (rhs_type)
3627 : 2361968 : || VECT_SCALAR_BOOLEAN_TYPE_P (rhs_type)
3628 : 7183504 : || TYPE_PRECISION (rhs_type) <= GET_MODE_BITSIZE (lhs_mode))
3629 : : return NULL;
3630 : :
3631 : : /* Try to find an unpromoted input. */
3632 : 320280 : vect_unpromoted_value unprom;
3633 : 320280 : if (!vect_look_through_possible_promotion (vinfo, rhs, &unprom)
3634 : 320280 : || TYPE_PRECISION (unprom.type) >= TYPE_PRECISION (rhs_type))
3635 : : return NULL;
3636 : :
3637 : : /* If the bits above RHS_TYPE matter, make sure that they're the
3638 : : same when extending from UNPROM as they are when extending from RHS. */
3639 : 38968 : if (!INTEGRAL_TYPE_P (lhs_type)
3640 : 38968 : && TYPE_SIGN (rhs_type) != TYPE_SIGN (unprom.type))
3641 : : return NULL;
3642 : :
3643 : : /* We can get the same result by casting UNPROM directly, to avoid
3644 : : the unnecessary widening and narrowing. */
3645 : 38848 : vect_pattern_detected ("vect_recog_cast_forwprop_pattern", last_stmt);
3646 : :
3647 : 38848 : *type_out = get_vectype_for_scalar_type (vinfo, lhs_type);
3648 : 38848 : if (!*type_out)
3649 : : return NULL;
3650 : :
3651 : 38846 : tree new_var = vect_recog_temp_ssa_var (lhs_type, NULL);
3652 : 38846 : gimple *pattern_stmt = gimple_build_assign (new_var, code, unprom.op);
3653 : 38846 : gimple_set_location (pattern_stmt, gimple_location (last_stmt));
3654 : :
3655 : 38846 : return pattern_stmt;
3656 : : }
3657 : :
3658 : : /* Try to detect a shift left of a widened input, converting LSHIFT_EXPR
3659 : : to WIDEN_LSHIFT_EXPR. See vect_recog_widen_op_pattern for details. */
3660 : :
3661 : : static gimple *
3662 : 27444030 : vect_recog_widen_shift_pattern (vec_info *vinfo,
3663 : : stmt_vec_info last_stmt_info, tree *type_out)
3664 : : {
3665 : 27444030 : return vect_recog_widen_op_pattern (vinfo, last_stmt_info, type_out,
3666 : : LSHIFT_EXPR, WIDEN_LSHIFT_EXPR, true,
3667 : 27444030 : "vect_recog_widen_shift_pattern");
3668 : : }
3669 : :
3670 : : /* Detect a rotate pattern wouldn't be otherwise vectorized:
3671 : :
3672 : : type a_t, b_t, c_t;
3673 : :
3674 : : S0 a_t = b_t r<< c_t;
3675 : :
3676 : : Input/Output:
3677 : :
3678 : : * STMT_VINFO: The stmt from which the pattern search begins,
3679 : : i.e. the shift/rotate stmt. The original stmt (S0) is replaced
3680 : : with a sequence:
3681 : :
3682 : : S1 d_t = -c_t;
3683 : : S2 e_t = d_t & (B - 1);
3684 : : S3 f_t = b_t << c_t;
3685 : : S4 g_t = b_t >> e_t;
3686 : : S0 a_t = f_t | g_t;
3687 : :
3688 : : where B is element bitsize of type.
3689 : :
3690 : : Output:
3691 : :
3692 : : * TYPE_OUT: The type of the output of this pattern.
3693 : :
3694 : : * Return value: A new stmt that will be used to replace the rotate
3695 : : S0 stmt. */
3696 : :
3697 : : static gimple *
3698 : 27444030 : vect_recog_rotate_pattern (vec_info *vinfo,
3699 : : stmt_vec_info stmt_vinfo, tree *type_out)
3700 : : {
3701 : 27444030 : gimple *last_stmt = stmt_vinfo->stmt;
3702 : 27444030 : tree oprnd0, oprnd1, lhs, var, var1, var2, vectype, type, stype, def, def2;
3703 : 27444030 : gimple *pattern_stmt, *def_stmt;
3704 : 27444030 : enum tree_code rhs_code;
3705 : 27444030 : enum vect_def_type dt;
3706 : 27444030 : optab optab1, optab2;
3707 : 27444030 : edge ext_def = NULL;
3708 : 27444030 : bool bswap16_p = false;
3709 : :
3710 : 27444030 : if (is_gimple_assign (last_stmt))
3711 : : {
3712 : 18738859 : rhs_code = gimple_assign_rhs_code (last_stmt);
3713 : 18738859 : switch (rhs_code)
3714 : : {
3715 : 4303 : case LROTATE_EXPR:
3716 : 4303 : case RROTATE_EXPR:
3717 : 4303 : break;
3718 : : default:
3719 : : return NULL;
3720 : : }
3721 : :
3722 : 4303 : lhs = gimple_assign_lhs (last_stmt);
3723 : 4303 : oprnd0 = gimple_assign_rhs1 (last_stmt);
3724 : 4303 : type = TREE_TYPE (oprnd0);
3725 : 4303 : oprnd1 = gimple_assign_rhs2 (last_stmt);
3726 : : }
3727 : 8705171 : else if (gimple_call_builtin_p (last_stmt, BUILT_IN_BSWAP16))
3728 : : {
3729 : : /* __builtin_bswap16 (x) is another form of x r>> 8.
3730 : : The vectorizer has bswap support, but only if the argument isn't
3731 : : promoted. */
3732 : 138 : lhs = gimple_call_lhs (last_stmt);
3733 : 138 : oprnd0 = gimple_call_arg (last_stmt, 0);
3734 : 138 : type = TREE_TYPE (oprnd0);
3735 : 138 : if (!lhs
3736 : 138 : || TYPE_PRECISION (TREE_TYPE (lhs)) != 16
3737 : 138 : || TYPE_PRECISION (type) <= 16
3738 : 126 : || TREE_CODE (oprnd0) != SSA_NAME
3739 : 264 : || BITS_PER_UNIT != 8)
3740 : 96 : return NULL;
3741 : :
3742 : 126 : stmt_vec_info def_stmt_info;
3743 : 126 : if (!vect_is_simple_use (oprnd0, vinfo, &dt, &def_stmt_info, &def_stmt))
3744 : : return NULL;
3745 : :
3746 : 126 : if (dt != vect_internal_def)
3747 : : return NULL;
3748 : :
3749 : 122 : if (gimple_assign_cast_p (def_stmt))
3750 : : {
3751 : 84 : def = gimple_assign_rhs1 (def_stmt);
3752 : 168 : if (INTEGRAL_TYPE_P (TREE_TYPE (def))
3753 : 168 : && TYPE_PRECISION (TREE_TYPE (def)) == 16)
3754 : : oprnd0 = def;
3755 : : }
3756 : :
3757 : 122 : type = TREE_TYPE (lhs);
3758 : 122 : vectype = get_vectype_for_scalar_type (vinfo, type);
3759 : 122 : if (vectype == NULL_TREE)
3760 : : return NULL;
3761 : :
3762 : 122 : if (tree char_vectype = get_same_sized_vectype (char_type_node, vectype))
3763 : : {
3764 : : /* The encoding uses one stepped pattern for each byte in the
3765 : : 16-bit word. */
3766 : 122 : vec_perm_builder elts (TYPE_VECTOR_SUBPARTS (char_vectype), 2, 3);
3767 : 488 : for (unsigned i = 0; i < 3; ++i)
3768 : 1098 : for (unsigned j = 0; j < 2; ++j)
3769 : 732 : elts.quick_push ((i + 1) * 2 - j - 1);
3770 : :
3771 : 122 : vec_perm_indices indices (elts, 1,
3772 : 122 : TYPE_VECTOR_SUBPARTS (char_vectype));
3773 : 122 : machine_mode vmode = TYPE_MODE (char_vectype);
3774 : 122 : if (can_vec_perm_const_p (vmode, vmode, indices))
3775 : : {
3776 : : /* vectorizable_bswap can handle the __builtin_bswap16 if we
3777 : : undo the argument promotion. */
3778 : 80 : if (!useless_type_conversion_p (type, TREE_TYPE (oprnd0)))
3779 : : {
3780 : 29 : def = vect_recog_temp_ssa_var (type, NULL);
3781 : 29 : def_stmt = gimple_build_assign (def, NOP_EXPR, oprnd0);
3782 : 29 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
3783 : 29 : oprnd0 = def;
3784 : : }
3785 : :
3786 : : /* Pattern detected. */
3787 : 80 : vect_pattern_detected ("vect_recog_rotate_pattern", last_stmt);
3788 : :
3789 : 80 : *type_out = vectype;
3790 : :
3791 : : /* Pattern supported. Create a stmt to be used to replace the
3792 : : pattern, with the unpromoted argument. */
3793 : 80 : var = vect_recog_temp_ssa_var (type, NULL);
3794 : 80 : pattern_stmt = gimple_build_call (gimple_call_fndecl (last_stmt),
3795 : : 1, oprnd0);
3796 : 80 : gimple_call_set_lhs (pattern_stmt, var);
3797 : 80 : gimple_call_set_fntype (as_a <gcall *> (pattern_stmt),
3798 : : gimple_call_fntype (last_stmt));
3799 : 80 : return pattern_stmt;
3800 : : }
3801 : 122 : }
3802 : :
3803 : 42 : oprnd1 = build_int_cst (integer_type_node, 8);
3804 : 42 : rhs_code = LROTATE_EXPR;
3805 : 42 : bswap16_p = true;
3806 : : }
3807 : : else
3808 : : return NULL;
3809 : :
3810 : 4345 : if (TREE_CODE (oprnd0) != SSA_NAME
3811 : 4225 : || !INTEGRAL_TYPE_P (type)
3812 : 8561 : || TYPE_PRECISION (TREE_TYPE (lhs)) != TYPE_PRECISION (type))
3813 : : return NULL;
3814 : :
3815 : 4216 : stmt_vec_info def_stmt_info;
3816 : 4216 : if (!vect_is_simple_use (oprnd1, vinfo, &dt, &def_stmt_info, &def_stmt))
3817 : : return NULL;
3818 : :
3819 : 4216 : if (dt != vect_internal_def
3820 : 4007 : && dt != vect_constant_def
3821 : 21 : && dt != vect_external_def)
3822 : : return NULL;
3823 : :
3824 : 4210 : vectype = get_vectype_for_scalar_type (vinfo, type);
3825 : 4210 : if (vectype == NULL_TREE)
3826 : : return NULL;
3827 : :
3828 : : /* If vector/vector or vector/scalar rotate is supported by the target,
3829 : : don't do anything here. */
3830 : 3991 : optab1 = optab_for_tree_code (rhs_code, vectype, optab_vector);
3831 : 3991 : if (optab1
3832 : 3991 : && can_implement_p (optab1, TYPE_MODE (vectype)))
3833 : : {
3834 : 34 : use_rotate:
3835 : 34 : if (bswap16_p)
3836 : : {
3837 : 0 : if (!useless_type_conversion_p (type, TREE_TYPE (oprnd0)))
3838 : : {
3839 : 0 : def = vect_recog_temp_ssa_var (type, NULL);
3840 : 0 : def_stmt = gimple_build_assign (def, NOP_EXPR, oprnd0);
3841 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
3842 : 0 : oprnd0 = def;
3843 : : }
3844 : :
3845 : : /* Pattern detected. */
3846 : 0 : vect_pattern_detected ("vect_recog_rotate_pattern", last_stmt);
3847 : :
3848 : 0 : *type_out = vectype;
3849 : :
3850 : : /* Pattern supported. Create a stmt to be used to replace the
3851 : : pattern. */
3852 : 0 : var = vect_recog_temp_ssa_var (type, NULL);
3853 : 0 : pattern_stmt = gimple_build_assign (var, LROTATE_EXPR, oprnd0,
3854 : : oprnd1);
3855 : 0 : return pattern_stmt;
3856 : : }
3857 : : return NULL;
3858 : : }
3859 : :
3860 : 3957 : if (is_a <bb_vec_info> (vinfo) || dt != vect_internal_def)
3861 : : {
3862 : 3896 : optab2 = optab_for_tree_code (rhs_code, vectype, optab_scalar);
3863 : 3896 : if (optab2
3864 : 3896 : && can_implement_p (optab2, TYPE_MODE (vectype)))
3865 : 0 : goto use_rotate;
3866 : : }
3867 : :
3868 : 3957 : tree utype = unsigned_type_for (type);
3869 : 3957 : tree uvectype = get_vectype_for_scalar_type (vinfo, utype);
3870 : 3957 : if (!uvectype)
3871 : : return NULL;
3872 : :
3873 : : /* If vector/vector or vector/scalar shifts aren't supported by the target,
3874 : : don't do anything here either. */
3875 : 3957 : optab1 = optab_for_tree_code (LSHIFT_EXPR, uvectype, optab_vector);
3876 : 3957 : optab2 = optab_for_tree_code (RSHIFT_EXPR, uvectype, optab_vector);
3877 : 3957 : if (!optab1
3878 : 3957 : || !can_implement_p (optab1, TYPE_MODE (uvectype))
3879 : 331 : || !optab2
3880 : 4288 : || !can_implement_p (optab2, TYPE_MODE (uvectype)))
3881 : : {
3882 : 3626 : if (! is_a <bb_vec_info> (vinfo) && dt == vect_internal_def)
3883 : : return NULL;
3884 : 3583 : optab1 = optab_for_tree_code (LSHIFT_EXPR, uvectype, optab_scalar);
3885 : 3583 : optab2 = optab_for_tree_code (RSHIFT_EXPR, uvectype, optab_scalar);
3886 : 3583 : if (!optab1
3887 : 3583 : || !can_implement_p (optab1, TYPE_MODE (uvectype))
3888 : 2952 : || !optab2
3889 : 6535 : || !can_implement_p (optab2, TYPE_MODE (uvectype)))
3890 : 631 : return NULL;
3891 : : }
3892 : :
3893 : 3283 : *type_out = vectype;
3894 : :
3895 : 3283 : if (!useless_type_conversion_p (utype, TREE_TYPE (oprnd0)))
3896 : : {
3897 : 63 : def = vect_recog_temp_ssa_var (utype, NULL);
3898 : 63 : def_stmt = gimple_build_assign (def, NOP_EXPR, oprnd0);
3899 : 63 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, uvectype);
3900 : 63 : oprnd0 = def;
3901 : : }
3902 : :
3903 : 3283 : if (dt == vect_external_def && TREE_CODE (oprnd1) == SSA_NAME)
3904 : 13 : ext_def = vect_get_external_def_edge (vinfo, oprnd1);
3905 : :
3906 : 3283 : def = NULL_TREE;
3907 : 3283 : scalar_int_mode mode = SCALAR_INT_TYPE_MODE (utype);
3908 : 3283 : if (dt != vect_internal_def || TYPE_MODE (TREE_TYPE (oprnd1)) == mode)
3909 : : def = oprnd1;
3910 : 28 : else if (def_stmt && gimple_assign_cast_p (def_stmt))
3911 : : {
3912 : 0 : tree rhs1 = gimple_assign_rhs1 (def_stmt);
3913 : 0 : if (TYPE_MODE (TREE_TYPE (rhs1)) == mode
3914 : 0 : && TYPE_PRECISION (TREE_TYPE (rhs1))
3915 : 0 : == TYPE_PRECISION (type))
3916 : : def = rhs1;
3917 : : }
3918 : :
3919 : 3255 : if (def == NULL_TREE)
3920 : : {
3921 : 28 : def = vect_recog_temp_ssa_var (utype, NULL);
3922 : 28 : def_stmt = gimple_build_assign (def, NOP_EXPR, oprnd1);
3923 : 28 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, uvectype);
3924 : : }
3925 : 3283 : stype = TREE_TYPE (def);
3926 : :
3927 : 3283 : if (TREE_CODE (def) == INTEGER_CST)
3928 : : {
3929 : 3172 : if (!tree_fits_uhwi_p (def)
3930 : 3172 : || tree_to_uhwi (def) >= GET_MODE_PRECISION (mode)
3931 : 6344 : || integer_zerop (def))
3932 : 0 : return NULL;
3933 : 3172 : def2 = build_int_cst (stype,
3934 : 3172 : GET_MODE_PRECISION (mode) - tree_to_uhwi (def));
3935 : : }
3936 : : else
3937 : : {
3938 : 111 : tree vecstype = get_vectype_for_scalar_type (vinfo, stype);
3939 : :
3940 : 111 : if (vecstype == NULL_TREE)
3941 : : return NULL;
3942 : 111 : def2 = vect_recog_temp_ssa_var (stype, NULL);
3943 : 111 : def_stmt = gimple_build_assign (def2, NEGATE_EXPR, def);
3944 : 111 : if (ext_def)
3945 : : {
3946 : 13 : basic_block new_bb
3947 : 13 : = gsi_insert_on_edge_immediate (ext_def, def_stmt);
3948 : 13 : gcc_assert (!new_bb);
3949 : : }
3950 : : else
3951 : 98 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, vecstype);
3952 : :
3953 : 111 : def2 = vect_recog_temp_ssa_var (stype, NULL);
3954 : 111 : tree mask = build_int_cst (stype, GET_MODE_PRECISION (mode) - 1);
3955 : 111 : def_stmt = gimple_build_assign (def2, BIT_AND_EXPR,
3956 : : gimple_assign_lhs (def_stmt), mask);
3957 : 111 : if (ext_def)
3958 : : {
3959 : 13 : basic_block new_bb
3960 : 13 : = gsi_insert_on_edge_immediate (ext_def, def_stmt);
3961 : 13 : gcc_assert (!new_bb);
3962 : : }
3963 : : else
3964 : 98 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, vecstype);
3965 : : }
3966 : :
3967 : 3283 : var1 = vect_recog_temp_ssa_var (utype, NULL);
3968 : 6438 : def_stmt = gimple_build_assign (var1, rhs_code == LROTATE_EXPR
3969 : : ? LSHIFT_EXPR : RSHIFT_EXPR,
3970 : : oprnd0, def);
3971 : 3283 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, uvectype);
3972 : :
3973 : 3283 : var2 = vect_recog_temp_ssa_var (utype, NULL);
3974 : 6438 : def_stmt = gimple_build_assign (var2, rhs_code == LROTATE_EXPR
3975 : : ? RSHIFT_EXPR : LSHIFT_EXPR,
3976 : : oprnd0, def2);
3977 : 3283 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, uvectype);
3978 : :
3979 : : /* Pattern detected. */
3980 : 3283 : vect_pattern_detected ("vect_recog_rotate_pattern", last_stmt);
3981 : :
3982 : : /* Pattern supported. Create a stmt to be used to replace the pattern. */
3983 : 3283 : var = vect_recog_temp_ssa_var (utype, NULL);
3984 : 3283 : pattern_stmt = gimple_build_assign (var, BIT_IOR_EXPR, var1, var2);
3985 : :
3986 : 3283 : if (!useless_type_conversion_p (type, utype))
3987 : : {
3988 : 49 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, uvectype);
3989 : 49 : tree result = vect_recog_temp_ssa_var (type, NULL);
3990 : 49 : pattern_stmt = gimple_build_assign (result, NOP_EXPR, var);
3991 : : }
3992 : : return pattern_stmt;
3993 : : }
3994 : :
3995 : : /* Detect a vector by vector shift pattern that wouldn't be otherwise
3996 : : vectorized:
3997 : :
3998 : : type a_t;
3999 : : TYPE b_T, res_T;
4000 : :
4001 : : S1 a_t = ;
4002 : : S2 b_T = ;
4003 : : S3 res_T = b_T op a_t;
4004 : :
4005 : : where type 'TYPE' is a type with different size than 'type',
4006 : : and op is <<, >> or rotate.
4007 : :
4008 : : Also detect cases:
4009 : :
4010 : : type a_t;
4011 : : TYPE b_T, c_T, res_T;
4012 : :
4013 : : S0 c_T = ;
4014 : : S1 a_t = (type) c_T;
4015 : : S2 b_T = ;
4016 : : S3 res_T = b_T op a_t;
4017 : :
4018 : : Input/Output:
4019 : :
4020 : : * STMT_VINFO: The stmt from which the pattern search begins,
4021 : : i.e. the shift/rotate stmt. The original stmt (S3) is replaced
4022 : : with a shift/rotate which has same type on both operands, in the
4023 : : second case just b_T op c_T, in the first case with added cast
4024 : : from a_t to c_T in STMT_VINFO_PATTERN_DEF_SEQ.
4025 : :
4026 : : Output:
4027 : :
4028 : : * TYPE_OUT: The type of the output of this pattern.
4029 : :
4030 : : * Return value: A new stmt that will be used to replace the shift/rotate
4031 : : S3 stmt. */
4032 : :
4033 : : static gimple *
4034 : 27447598 : vect_recog_vector_vector_shift_pattern (vec_info *vinfo,
4035 : : stmt_vec_info stmt_vinfo,
4036 : : tree *type_out)
4037 : : {
4038 : 27447598 : gimple *last_stmt = stmt_vinfo->stmt;
4039 : 27447598 : tree oprnd0, oprnd1, lhs, var;
4040 : 27447598 : gimple *pattern_stmt;
4041 : 27447598 : enum tree_code rhs_code;
4042 : :
4043 : 27447598 : if (!is_gimple_assign (last_stmt))
4044 : : return NULL;
4045 : :
4046 : 18742549 : rhs_code = gimple_assign_rhs_code (last_stmt);
4047 : 18742549 : switch (rhs_code)
4048 : : {
4049 : 470715 : case LSHIFT_EXPR:
4050 : 470715 : case RSHIFT_EXPR:
4051 : 470715 : case LROTATE_EXPR:
4052 : 470715 : case RROTATE_EXPR:
4053 : 470715 : break;
4054 : : default:
4055 : : return NULL;
4056 : : }
4057 : :
4058 : 470715 : lhs = gimple_assign_lhs (last_stmt);
4059 : 470715 : oprnd0 = gimple_assign_rhs1 (last_stmt);
4060 : 470715 : oprnd1 = gimple_assign_rhs2 (last_stmt);
4061 : 470715 : if (TREE_CODE (oprnd0) != SSA_NAME
4062 : 422201 : || TREE_CODE (oprnd1) != SSA_NAME
4063 : 47135 : || TYPE_MODE (TREE_TYPE (oprnd0)) == TYPE_MODE (TREE_TYPE (oprnd1))
4064 : 18556 : || !INTEGRAL_TYPE_P (TREE_TYPE (oprnd0))
4065 : 18248 : || !type_has_mode_precision_p (TREE_TYPE (oprnd1))
4066 : 488963 : || TYPE_PRECISION (TREE_TYPE (lhs))
4067 : 18248 : != TYPE_PRECISION (TREE_TYPE (oprnd0)))
4068 : 452467 : return NULL;
4069 : :
4070 : 18248 : stmt_vec_info def_vinfo = vect_get_internal_def (vinfo, oprnd1);
4071 : 18248 : if (!def_vinfo)
4072 : : return NULL;
4073 : :
4074 : 16281 : *type_out = get_vectype_for_scalar_type (vinfo, TREE_TYPE (oprnd0));
4075 : 16281 : if (*type_out == NULL_TREE)
4076 : : return NULL;
4077 : :
4078 : 11043 : tree def = NULL_TREE;
4079 : 11043 : gassign *def_stmt = dyn_cast <gassign *> (def_vinfo->stmt);
4080 : 9277 : if (def_stmt && gimple_assign_cast_p (def_stmt))
4081 : : {
4082 : 1976 : tree rhs1 = gimple_assign_rhs1 (def_stmt);
4083 : 1976 : if (TYPE_MODE (TREE_TYPE (rhs1)) == TYPE_MODE (TREE_TYPE (oprnd0))
4084 : 1976 : && TYPE_PRECISION (TREE_TYPE (rhs1))
4085 : 598 : == TYPE_PRECISION (TREE_TYPE (oprnd0)))
4086 : : {
4087 : 598 : if (TYPE_PRECISION (TREE_TYPE (oprnd1))
4088 : 598 : >= TYPE_PRECISION (TREE_TYPE (rhs1)))
4089 : : def = rhs1;
4090 : : else
4091 : : {
4092 : 593 : tree mask
4093 : 593 : = build_low_bits_mask (TREE_TYPE (rhs1),
4094 : 593 : TYPE_PRECISION (TREE_TYPE (oprnd1)));
4095 : 593 : def = vect_recog_temp_ssa_var (TREE_TYPE (rhs1), NULL);
4096 : 593 : def_stmt = gimple_build_assign (def, BIT_AND_EXPR, rhs1, mask);
4097 : 593 : tree vecstype = get_vectype_for_scalar_type (vinfo,
4098 : 593 : TREE_TYPE (rhs1));
4099 : 593 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, vecstype);
4100 : : }
4101 : : }
4102 : : }
4103 : :
4104 : 598 : if (def == NULL_TREE)
4105 : : {
4106 : 10445 : def = vect_recog_temp_ssa_var (TREE_TYPE (oprnd0), NULL);
4107 : 10445 : def_stmt = gimple_build_assign (def, NOP_EXPR, oprnd1);
4108 : 10445 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
4109 : : }
4110 : :
4111 : : /* Pattern detected. */
4112 : 11043 : vect_pattern_detected ("vect_recog_vector_vector_shift_pattern", last_stmt);
4113 : :
4114 : : /* Pattern supported. Create a stmt to be used to replace the pattern. */
4115 : 11043 : var = vect_recog_temp_ssa_var (TREE_TYPE (oprnd0), NULL);
4116 : 11043 : pattern_stmt = gimple_build_assign (var, rhs_code, oprnd0, def);
4117 : :
4118 : 11043 : return pattern_stmt;
4119 : : }
4120 : :
4121 : : /* Return true iff the target has a vector optab implementing the operation
4122 : : CODE on type VECTYPE. */
4123 : :
4124 : : static bool
4125 : 532886 : target_has_vecop_for_code (tree_code code, tree vectype)
4126 : : {
4127 : 532886 : optab voptab = optab_for_tree_code (code, vectype, optab_vector);
4128 : 532886 : return voptab
4129 : 532886 : && can_implement_p (voptab, TYPE_MODE (vectype));
4130 : : }
4131 : :
4132 : : /* Verify that the target has optabs of VECTYPE to perform all the steps
4133 : : needed by the multiplication-by-immediate synthesis algorithm described by
4134 : : ALG and VAR. If SYNTH_SHIFT_P is true ensure that vector addition is
4135 : : present. Return true iff the target supports all the steps. */
4136 : :
4137 : : static bool
4138 : 232507 : target_supports_mult_synth_alg (struct algorithm *alg, mult_variant var,
4139 : : tree vectype, bool synth_shift_p)
4140 : : {
4141 : 232507 : if (alg->op[0] != alg_zero && alg->op[0] != alg_m)
4142 : : return false;
4143 : :
4144 : 232507 : bool supports_vminus = target_has_vecop_for_code (MINUS_EXPR, vectype);
4145 : 232507 : bool supports_vplus = target_has_vecop_for_code (PLUS_EXPR, vectype);
4146 : :
4147 : 232507 : if (var == negate_variant
4148 : 232507 : && !target_has_vecop_for_code (NEGATE_EXPR, vectype))
4149 : : return false;
4150 : :
4151 : : /* If we must synthesize shifts with additions make sure that vector
4152 : : addition is available. */
4153 : 232089 : if ((var == add_variant || synth_shift_p) && !supports_vplus)
4154 : : return false;
4155 : :
4156 : 129964 : for (int i = 1; i < alg->ops; i++)
4157 : : {
4158 : 101863 : switch (alg->op[i])
4159 : : {
4160 : : case alg_shift:
4161 : : break;
4162 : 26796 : case alg_add_t_m2:
4163 : 26796 : case alg_add_t2_m:
4164 : 26796 : case alg_add_factor:
4165 : 26796 : if (!supports_vplus)
4166 : : return false;
4167 : : break;
4168 : 17105 : case alg_sub_t_m2:
4169 : 17105 : case alg_sub_t2_m:
4170 : 17105 : case alg_sub_factor:
4171 : 17105 : if (!supports_vminus)
4172 : : return false;
4173 : : break;
4174 : : case alg_unknown:
4175 : : case alg_m:
4176 : : case alg_zero:
4177 : : case alg_impossible:
4178 : : return false;
4179 : 0 : default:
4180 : 0 : gcc_unreachable ();
4181 : : }
4182 : : }
4183 : :
4184 : : return true;
4185 : : }
4186 : :
4187 : : /* Synthesize a left shift of OP by AMNT bits using a series of additions and
4188 : : putting the final result in DEST. Append all statements but the last into
4189 : : VINFO. Return the last statement. */
4190 : :
4191 : : static gimple *
4192 : 0 : synth_lshift_by_additions (vec_info *vinfo,
4193 : : tree dest, tree op, HOST_WIDE_INT amnt,
4194 : : stmt_vec_info stmt_info)
4195 : : {
4196 : 0 : HOST_WIDE_INT i;
4197 : 0 : tree itype = TREE_TYPE (op);
4198 : 0 : tree prev_res = op;
4199 : 0 : gcc_assert (amnt >= 0);
4200 : 0 : for (i = 0; i < amnt; i++)
4201 : : {
4202 : 0 : tree tmp_var = (i < amnt - 1) ? vect_recog_temp_ssa_var (itype, NULL)
4203 : 0 : : dest;
4204 : 0 : gimple *stmt
4205 : 0 : = gimple_build_assign (tmp_var, PLUS_EXPR, prev_res, prev_res);
4206 : 0 : prev_res = tmp_var;
4207 : 0 : if (i < amnt - 1)
4208 : 0 : append_pattern_def_seq (vinfo, stmt_info, stmt);
4209 : : else
4210 : 0 : return stmt;
4211 : : }
4212 : 0 : gcc_unreachable ();
4213 : : return NULL;
4214 : : }
4215 : :
4216 : : /* Helper for vect_synth_mult_by_constant. Apply a binary operation
4217 : : CODE to operands OP1 and OP2, creating a new temporary SSA var in
4218 : : the process if necessary. Append the resulting assignment statements
4219 : : to the sequence in STMT_VINFO. Return the SSA variable that holds the
4220 : : result of the binary operation. If SYNTH_SHIFT_P is true synthesize
4221 : : left shifts using additions. */
4222 : :
4223 : : static tree
4224 : 43808 : apply_binop_and_append_stmt (vec_info *vinfo,
4225 : : tree_code code, tree op1, tree op2,
4226 : : stmt_vec_info stmt_vinfo, bool synth_shift_p)
4227 : : {
4228 : 43808 : if (integer_zerop (op2)
4229 : 43808 : && (code == LSHIFT_EXPR
4230 : 38346 : || code == PLUS_EXPR))
4231 : : {
4232 : 38346 : gcc_assert (TREE_CODE (op1) == SSA_NAME);
4233 : : return op1;
4234 : : }
4235 : :
4236 : 5462 : gimple *stmt;
4237 : 5462 : tree itype = TREE_TYPE (op1);
4238 : 5462 : tree tmp_var = vect_recog_temp_ssa_var (itype, NULL);
4239 : :
4240 : 5462 : if (code == LSHIFT_EXPR
4241 : 5462 : && synth_shift_p)
4242 : : {
4243 : 0 : stmt = synth_lshift_by_additions (vinfo, tmp_var, op1,
4244 : 0 : TREE_INT_CST_LOW (op2), stmt_vinfo);
4245 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, stmt);
4246 : 0 : return tmp_var;
4247 : : }
4248 : :
4249 : 5462 : stmt = gimple_build_assign (tmp_var, code, op1, op2);
4250 : 5462 : append_pattern_def_seq (vinfo, stmt_vinfo, stmt);
4251 : 5462 : return tmp_var;
4252 : : }
4253 : :
4254 : : /* Synthesize a multiplication of OP by an INTEGER_CST VAL using shifts
4255 : : and simple arithmetic operations to be vectorized. Record the statements
4256 : : produced in STMT_VINFO and return the last statement in the sequence or
4257 : : NULL if it's not possible to synthesize such a multiplication.
4258 : : This function mirrors the behavior of expand_mult_const in expmed.cc but
4259 : : works on tree-ssa form. */
4260 : :
4261 : : static gimple *
4262 : 235063 : vect_synth_mult_by_constant (vec_info *vinfo, tree op, tree val,
4263 : : stmt_vec_info stmt_vinfo)
4264 : : {
4265 : 235063 : tree itype = TREE_TYPE (op);
4266 : 235063 : machine_mode mode = TYPE_MODE (itype);
4267 : 235063 : struct algorithm alg;
4268 : 235063 : mult_variant variant;
4269 : 235063 : if (!tree_fits_shwi_p (val))
4270 : : return NULL;
4271 : :
4272 : : /* Multiplication synthesis by shifts, adds and subs can introduce
4273 : : signed overflow where the original operation didn't. Perform the
4274 : : operations on an unsigned type and cast back to avoid this.
4275 : : In the future we may want to relax this for synthesis algorithms
4276 : : that we can prove do not cause unexpected overflow. */
4277 : 232507 : bool cast_to_unsigned_p = !TYPE_OVERFLOW_WRAPS (itype);
4278 : :
4279 : 49592 : tree multtype = cast_to_unsigned_p ? unsigned_type_for (itype) : itype;
4280 : 232507 : tree vectype = get_vectype_for_scalar_type (vinfo, multtype);
4281 : 232507 : if (!vectype)
4282 : : return NULL;
4283 : :
4284 : : /* Targets that don't support vector shifts but support vector additions
4285 : : can synthesize shifts that way. */
4286 : 232507 : bool synth_shift_p = !vect_supportable_shift (vinfo, LSHIFT_EXPR, multtype);
4287 : :
4288 : 232507 : HOST_WIDE_INT hwval = tree_to_shwi (val);
4289 : : /* Use MAX_COST here as we don't want to limit the sequence on rtx costs.
4290 : : The vectorizer's benefit analysis will decide whether it's beneficial
4291 : : to do this. */
4292 : 465014 : bool possible = choose_mult_variant (VECTOR_MODE_P (TYPE_MODE (vectype))
4293 : 232507 : ? TYPE_MODE (vectype) : mode,
4294 : : hwval, &alg, &variant, MAX_COST);
4295 : 232507 : if (!possible)
4296 : : return NULL;
4297 : :
4298 : 232507 : if (!target_supports_mult_synth_alg (&alg, variant, vectype, synth_shift_p))
4299 : : return NULL;
4300 : :
4301 : 28101 : tree accumulator;
4302 : :
4303 : : /* Clear out the sequence of statements so we can populate it below. */
4304 : 28101 : gimple *stmt = NULL;
4305 : :
4306 : 28101 : if (cast_to_unsigned_p)
4307 : : {
4308 : 11383 : tree tmp_op = vect_recog_temp_ssa_var (multtype, NULL);
4309 : 11383 : stmt = gimple_build_assign (tmp_op, CONVERT_EXPR, op);
4310 : 11383 : append_pattern_def_seq (vinfo, stmt_vinfo, stmt);
4311 : 11383 : op = tmp_op;
4312 : : }
4313 : :
4314 : 28101 : if (alg.op[0] == alg_zero)
4315 : 173 : accumulator = build_int_cst (multtype, 0);
4316 : : else
4317 : : accumulator = op;
4318 : :
4319 : 28101 : bool needs_fixup = (variant == negate_variant)
4320 : 28101 : || (variant == add_variant);
4321 : :
4322 : 129811 : for (int i = 1; i < alg.ops; i++)
4323 : : {
4324 : 101710 : tree shft_log = build_int_cst (multtype, alg.log[i]);
4325 : 101710 : tree accum_tmp = vect_recog_temp_ssa_var (multtype, NULL);
4326 : 101710 : tree tmp_var = NULL_TREE;
4327 : :
4328 : 101710 : switch (alg.op[i])
4329 : : {
4330 : 57902 : case alg_shift:
4331 : 57902 : if (synth_shift_p)
4332 : 0 : stmt
4333 : 0 : = synth_lshift_by_additions (vinfo, accum_tmp, accumulator,
4334 : 0 : alg.log[i], stmt_vinfo);
4335 : : else
4336 : 57902 : stmt = gimple_build_assign (accum_tmp, LSHIFT_EXPR, accumulator,
4337 : : shft_log);
4338 : : break;
4339 : 22104 : case alg_add_t_m2:
4340 : 22104 : tmp_var
4341 : 22104 : = apply_binop_and_append_stmt (vinfo, LSHIFT_EXPR, op, shft_log,
4342 : : stmt_vinfo, synth_shift_p);
4343 : 22104 : stmt = gimple_build_assign (accum_tmp, PLUS_EXPR, accumulator,
4344 : : tmp_var);
4345 : 22104 : break;
4346 : 16420 : case alg_sub_t_m2:
4347 : 16420 : tmp_var = apply_binop_and_append_stmt (vinfo, LSHIFT_EXPR, op,
4348 : : shft_log, stmt_vinfo,
4349 : : synth_shift_p);
4350 : : /* In some algorithms the first step involves zeroing the
4351 : : accumulator. If subtracting from such an accumulator
4352 : : just emit the negation directly. */
4353 : 16420 : if (integer_zerop (accumulator))
4354 : 173 : stmt = gimple_build_assign (accum_tmp, NEGATE_EXPR, tmp_var);
4355 : : else
4356 : 16247 : stmt = gimple_build_assign (accum_tmp, MINUS_EXPR, accumulator,
4357 : : tmp_var);
4358 : : break;
4359 : 0 : case alg_add_t2_m:
4360 : 0 : tmp_var
4361 : 0 : = apply_binop_and_append_stmt (vinfo, LSHIFT_EXPR, accumulator,
4362 : : shft_log, stmt_vinfo, synth_shift_p);
4363 : 0 : stmt = gimple_build_assign (accum_tmp, PLUS_EXPR, tmp_var, op);
4364 : 0 : break;
4365 : 0 : case alg_sub_t2_m:
4366 : 0 : tmp_var
4367 : 0 : = apply_binop_and_append_stmt (vinfo, LSHIFT_EXPR, accumulator,
4368 : : shft_log, stmt_vinfo, synth_shift_p);
4369 : 0 : stmt = gimple_build_assign (accum_tmp, MINUS_EXPR, tmp_var, op);
4370 : 0 : break;
4371 : 4634 : case alg_add_factor:
4372 : 4634 : tmp_var
4373 : 4634 : = apply_binop_and_append_stmt (vinfo, LSHIFT_EXPR, accumulator,
4374 : : shft_log, stmt_vinfo, synth_shift_p);
4375 : 4634 : stmt = gimple_build_assign (accum_tmp, PLUS_EXPR, accumulator,
4376 : : tmp_var);
4377 : 4634 : break;
4378 : 650 : case alg_sub_factor:
4379 : 650 : tmp_var
4380 : 650 : = apply_binop_and_append_stmt (vinfo, LSHIFT_EXPR, accumulator,
4381 : : shft_log, stmt_vinfo, synth_shift_p);
4382 : 650 : stmt = gimple_build_assign (accum_tmp, MINUS_EXPR, tmp_var,
4383 : : accumulator);
4384 : 650 : break;
4385 : 0 : default:
4386 : 0 : gcc_unreachable ();
4387 : : }
4388 : : /* We don't want to append the last stmt in the sequence to stmt_vinfo
4389 : : but rather return it directly. */
4390 : :
4391 : 101710 : if ((i < alg.ops - 1) || needs_fixup || cast_to_unsigned_p)
4392 : 85193 : append_pattern_def_seq (vinfo, stmt_vinfo, stmt);
4393 : 101710 : accumulator = accum_tmp;
4394 : : }
4395 : 28101 : if (variant == negate_variant)
4396 : : {
4397 : 326 : tree accum_tmp = vect_recog_temp_ssa_var (multtype, NULL);
4398 : 326 : stmt = gimple_build_assign (accum_tmp, NEGATE_EXPR, accumulator);
4399 : 326 : accumulator = accum_tmp;
4400 : 326 : if (cast_to_unsigned_p)
4401 : 135 : append_pattern_def_seq (vinfo, stmt_vinfo, stmt);
4402 : : }
4403 : 27775 : else if (variant == add_variant)
4404 : : {
4405 : 83 : tree accum_tmp = vect_recog_temp_ssa_var (multtype, NULL);
4406 : 83 : stmt = gimple_build_assign (accum_tmp, PLUS_EXPR, accumulator, op);
4407 : 83 : accumulator = accum_tmp;
4408 : 83 : if (cast_to_unsigned_p)
4409 : 73 : append_pattern_def_seq (vinfo, stmt_vinfo, stmt);
4410 : : }
4411 : : /* Move back to a signed if needed. */
4412 : 27900 : if (cast_to_unsigned_p)
4413 : : {
4414 : 11383 : tree accum_tmp = vect_recog_temp_ssa_var (itype, NULL);
4415 : 11383 : stmt = gimple_build_assign (accum_tmp, CONVERT_EXPR, accumulator);
4416 : : }
4417 : :
4418 : : return stmt;
4419 : : }
4420 : :
4421 : : /* Detect multiplication by constant and convert it into a sequence of
4422 : : shifts and additions, subtractions, negations. We reuse the
4423 : : choose_mult_variant algorithms from expmed.cc
4424 : :
4425 : : Input/Output:
4426 : :
4427 : : STMT_VINFO: The stmt from which the pattern search begins,
4428 : : i.e. the mult stmt.
4429 : :
4430 : : Output:
4431 : :
4432 : : * TYPE_OUT: The type of the output of this pattern.
4433 : :
4434 : : * Return value: A new stmt that will be used to replace
4435 : : the multiplication. */
4436 : :
4437 : : static gimple *
4438 : 27578497 : vect_recog_mult_pattern (vec_info *vinfo,
4439 : : stmt_vec_info stmt_vinfo, tree *type_out)
4440 : : {
4441 : 27578497 : gimple *last_stmt = stmt_vinfo->stmt;
4442 : 27578497 : tree oprnd0, oprnd1, vectype, itype;
4443 : 27578497 : gimple *pattern_stmt;
4444 : :
4445 : 27578497 : if (!is_gimple_assign (last_stmt))
4446 : : return NULL;
4447 : :
4448 : 18873448 : if (gimple_assign_rhs_code (last_stmt) != MULT_EXPR)
4449 : : return NULL;
4450 : :
4451 : 1230628 : oprnd0 = gimple_assign_rhs1 (last_stmt);
4452 : 1230628 : oprnd1 = gimple_assign_rhs2 (last_stmt);
4453 : 1230628 : itype = TREE_TYPE (oprnd0);
4454 : :
4455 : 1230628 : if (TREE_CODE (oprnd0) != SSA_NAME
4456 : 1230565 : || TREE_CODE (oprnd1) != INTEGER_CST
4457 : 725096 : || !INTEGRAL_TYPE_P (itype)
4458 : 1955724 : || !type_has_mode_precision_p (itype))
4459 : 505584 : return NULL;
4460 : :
4461 : 725044 : vectype = get_vectype_for_scalar_type (vinfo, itype);
4462 : 725044 : if (vectype == NULL_TREE)
4463 : : return NULL;
4464 : :
4465 : : /* If the target can handle vectorized multiplication natively,
4466 : : don't attempt to optimize this. */
4467 : 591547 : optab mul_optab = optab_for_tree_code (MULT_EXPR, vectype, optab_default);
4468 : 591547 : if (mul_optab != unknown_optab
4469 : 591547 : && can_implement_p (mul_optab, TYPE_MODE (vectype)))
4470 : : return NULL;
4471 : :
4472 : 235063 : pattern_stmt = vect_synth_mult_by_constant (vinfo,
4473 : : oprnd0, oprnd1, stmt_vinfo);
4474 : 235063 : if (!pattern_stmt)
4475 : : return NULL;
4476 : :
4477 : : /* Pattern detected. */
4478 : 28101 : vect_pattern_detected ("vect_recog_mult_pattern", last_stmt);
4479 : :
4480 : 28101 : *type_out = vectype;
4481 : :
4482 : 28101 : return pattern_stmt;
4483 : : }
4484 : :
4485 : : extern bool gimple_unsigned_integer_sat_add (tree, tree*, tree (*)(tree));
4486 : : extern bool gimple_unsigned_integer_sat_sub (tree, tree*, tree (*)(tree));
4487 : : extern bool gimple_unsigned_integer_sat_trunc (tree, tree*, tree (*)(tree));
4488 : :
4489 : : extern bool gimple_signed_integer_sat_add (tree, tree*, tree (*)(tree));
4490 : : extern bool gimple_signed_integer_sat_sub (tree, tree*, tree (*)(tree));
4491 : : extern bool gimple_signed_integer_sat_trunc (tree, tree*, tree (*)(tree));
4492 : :
4493 : : static gimple *
4494 : 239 : vect_recog_build_binary_gimple_stmt (vec_info *vinfo, stmt_vec_info stmt_info,
4495 : : internal_fn fn, tree *type_out,
4496 : : tree lhs, tree op_0, tree op_1)
4497 : : {
4498 : 239 : tree itype = TREE_TYPE (op_0);
4499 : 239 : tree otype = TREE_TYPE (lhs);
4500 : 239 : tree v_itype = get_vectype_for_scalar_type (vinfo, itype);
4501 : 239 : tree v_otype = get_vectype_for_scalar_type (vinfo, otype);
4502 : :
4503 : 239 : if (v_itype != NULL_TREE && v_otype != NULL_TREE
4504 : 239 : && direct_internal_fn_supported_p (fn, v_itype, OPTIMIZE_FOR_BOTH))
4505 : : {
4506 : 61 : gcall *call = gimple_build_call_internal (fn, 2, op_0, op_1);
4507 : 61 : tree in_ssa = vect_recog_temp_ssa_var (itype, NULL);
4508 : :
4509 : 61 : gimple_call_set_lhs (call, in_ssa);
4510 : 61 : gimple_call_set_nothrow (call, /* nothrow_p */ false);
4511 : 61 : gimple_set_location (call, gimple_location (STMT_VINFO_STMT (stmt_info)));
4512 : :
4513 : 61 : *type_out = v_otype;
4514 : :
4515 : 61 : if (types_compatible_p (itype, otype))
4516 : : return call;
4517 : : else
4518 : : {
4519 : 0 : append_pattern_def_seq (vinfo, stmt_info, call, v_itype);
4520 : 0 : tree out_ssa = vect_recog_temp_ssa_var (otype, NULL);
4521 : :
4522 : 0 : return gimple_build_assign (out_ssa, NOP_EXPR, in_ssa);
4523 : : }
4524 : : }
4525 : :
4526 : : return NULL;
4527 : : }
4528 : :
4529 : : /*
4530 : : * Try to detect saturation add pattern (SAT_ADD), aka below gimple:
4531 : : * _7 = _4 + _6;
4532 : : * _8 = _4 > _7;
4533 : : * _9 = (long unsigned int) _8;
4534 : : * _10 = -_9;
4535 : : * _12 = _7 | _10;
4536 : : *
4537 : : * And then simplied to
4538 : : * _12 = .SAT_ADD (_4, _6);
4539 : : */
4540 : :
4541 : : static gimple *
4542 : 27656364 : vect_recog_sat_add_pattern (vec_info *vinfo, stmt_vec_info stmt_vinfo,
4543 : : tree *type_out)
4544 : : {
4545 : 27656364 : gimple *last_stmt = STMT_VINFO_STMT (stmt_vinfo);
4546 : :
4547 : 27656364 : if (!is_gimple_assign (last_stmt))
4548 : : return NULL;
4549 : :
4550 : 18951315 : tree ops[2];
4551 : 18951315 : tree lhs = gimple_assign_lhs (last_stmt);
4552 : :
4553 : 18951315 : if (gimple_unsigned_integer_sat_add (lhs, ops, NULL)
4554 : 18951315 : || gimple_signed_integer_sat_add (lhs, ops, NULL))
4555 : : {
4556 : 50 : if (TREE_CODE (ops[1]) == INTEGER_CST)
4557 : 12 : ops[1] = fold_convert (TREE_TYPE (ops[0]), ops[1]);
4558 : :
4559 : 50 : gimple *stmt = vect_recog_build_binary_gimple_stmt (vinfo, stmt_vinfo,
4560 : : IFN_SAT_ADD, type_out,
4561 : : lhs, ops[0], ops[1]);
4562 : 50 : if (stmt)
4563 : : {
4564 : 32 : vect_pattern_detected ("vect_recog_sat_add_pattern", last_stmt);
4565 : 32 : return stmt;
4566 : : }
4567 : : }
4568 : :
4569 : : return NULL;
4570 : : }
4571 : :
4572 : : /*
4573 : : * Try to transform the truncation for .SAT_SUB pattern, mostly occurs in
4574 : : * the benchmark zip. Aka:
4575 : : *
4576 : : * unsigned int _1;
4577 : : * unsigned int _2;
4578 : : * unsigned short int _4;
4579 : : * _9 = (unsigned short int).SAT_SUB (_1, _2);
4580 : : *
4581 : : * if _1 is known to be in the range of unsigned short int. For example
4582 : : * there is a def _1 = (unsigned short int)_4. Then we can transform the
4583 : : * truncation to:
4584 : : *
4585 : : * _3 = (unsigned short int) MIN (65535, _2); // aka _3 = .SAT_TRUNC (_2);
4586 : : * _9 = .SAT_SUB (_4, _3);
4587 : : *
4588 : : * Then, we can better vectorized code and avoid the unnecessary narrowing
4589 : : * stmt during vectorization with below stmt(s).
4590 : : *
4591 : : * _3 = .SAT_TRUNC(_2); // SI => HI
4592 : : * _9 = .SAT_SUB (_4, _3);
4593 : : */
4594 : : static void
4595 : 189 : vect_recog_sat_sub_pattern_transform (vec_info *vinfo,
4596 : : stmt_vec_info stmt_vinfo,
4597 : : tree lhs, tree *ops)
4598 : : {
4599 : 189 : tree otype = TREE_TYPE (lhs);
4600 : 189 : tree itype = TREE_TYPE (ops[0]);
4601 : 189 : unsigned itype_prec = TYPE_PRECISION (itype);
4602 : 189 : unsigned otype_prec = TYPE_PRECISION (otype);
4603 : :
4604 : 189 : if (types_compatible_p (otype, itype) || otype_prec >= itype_prec)
4605 : 189 : return;
4606 : :
4607 : 0 : tree v_otype = get_vectype_for_scalar_type (vinfo, otype);
4608 : 0 : tree v_itype = get_vectype_for_scalar_type (vinfo, itype);
4609 : 0 : tree_pair v_pair = tree_pair (v_otype, v_itype);
4610 : :
4611 : 0 : if (v_otype == NULL_TREE || v_itype == NULL_TREE
4612 : 0 : || !direct_internal_fn_supported_p (IFN_SAT_TRUNC, v_pair,
4613 : : OPTIMIZE_FOR_BOTH))
4614 : 0 : return;
4615 : :
4616 : : /* 1. Find the _4 and update ops[0] as above example. */
4617 : 0 : vect_unpromoted_value unprom;
4618 : 0 : tree tmp = vect_look_through_possible_promotion (vinfo, ops[0], &unprom);
4619 : :
4620 : 0 : if (tmp == NULL_TREE || TYPE_PRECISION (unprom.type) != otype_prec)
4621 : : return;
4622 : :
4623 : 0 : ops[0] = tmp;
4624 : :
4625 : : /* 2. Generate _3 = .SAT_TRUNC (_2) and update ops[1] as above example. */
4626 : 0 : tree trunc_lhs_ssa = vect_recog_temp_ssa_var (otype, NULL);
4627 : 0 : gcall *call = gimple_build_call_internal (IFN_SAT_TRUNC, 1, ops[1]);
4628 : :
4629 : 0 : gimple_call_set_lhs (call, trunc_lhs_ssa);
4630 : 0 : gimple_call_set_nothrow (call, /* nothrow_p */ false);
4631 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, call, v_otype);
4632 : :
4633 : 0 : ops[1] = trunc_lhs_ssa;
4634 : : }
4635 : :
4636 : : /*
4637 : : * Try to detect saturation sub pattern (SAT_ADD), aka below gimple:
4638 : : * Unsigned:
4639 : : * _7 = _1 >= _2;
4640 : : * _8 = _1 - _2;
4641 : : * _10 = (long unsigned int) _7;
4642 : : * _9 = _8 * _10;
4643 : : *
4644 : : * And then simplied to
4645 : : * _9 = .SAT_SUB (_1, _2);
4646 : : *
4647 : : * Signed:
4648 : : * x.0_4 = (unsigned char) x_16;
4649 : : * y.1_5 = (unsigned char) y_18;
4650 : : * _6 = x.0_4 - y.1_5;
4651 : : * minus_19 = (int8_t) _6;
4652 : : * _7 = x_16 ^ y_18;
4653 : : * _8 = x_16 ^ minus_19;
4654 : : * _44 = _7 < 0;
4655 : : * _23 = x_16 < 0;
4656 : : * _24 = (signed char) _23;
4657 : : * _58 = (unsigned char) _24;
4658 : : * _59 = -_58;
4659 : : * _25 = (signed char) _59;
4660 : : * _26 = _25 ^ 127;
4661 : : * _42 = _8 < 0;
4662 : : * _41 = _42 & _44;
4663 : : * iftmp.2_11 = _41 ? _26 : minus_19;
4664 : : *
4665 : : * And then simplied to
4666 : : * iftmp.2_11 = .SAT_SUB (x_16, y_18);
4667 : : */
4668 : :
4669 : : static gimple *
4670 : 27656332 : vect_recog_sat_sub_pattern (vec_info *vinfo, stmt_vec_info stmt_vinfo,
4671 : : tree *type_out)
4672 : : {
4673 : 27656332 : gimple *last_stmt = STMT_VINFO_STMT (stmt_vinfo);
4674 : :
4675 : 27656332 : if (!is_gimple_assign (last_stmt))
4676 : : return NULL;
4677 : :
4678 : 18951283 : tree ops[2];
4679 : 18951283 : tree lhs = gimple_assign_lhs (last_stmt);
4680 : :
4681 : 18951283 : if (gimple_unsigned_integer_sat_sub (lhs, ops, NULL)
4682 : 18951283 : || gimple_signed_integer_sat_sub (lhs, ops, NULL))
4683 : : {
4684 : 189 : vect_recog_sat_sub_pattern_transform (vinfo, stmt_vinfo, lhs, ops);
4685 : 189 : gimple *stmt = vect_recog_build_binary_gimple_stmt (vinfo, stmt_vinfo,
4686 : : IFN_SAT_SUB, type_out,
4687 : : lhs, ops[0], ops[1]);
4688 : 189 : if (stmt)
4689 : : {
4690 : 29 : vect_pattern_detected ("vect_recog_sat_sub_pattern", last_stmt);
4691 : 29 : return stmt;
4692 : : }
4693 : : }
4694 : :
4695 : : return NULL;
4696 : : }
4697 : :
4698 : : /*
4699 : : * Try to detect saturation truncation pattern (SAT_TRUNC), aka below gimple:
4700 : : * overflow_5 = x_4(D) > 4294967295;
4701 : : * _1 = (unsigned int) x_4(D);
4702 : : * _2 = (unsigned int) overflow_5;
4703 : : * _3 = -_2;
4704 : : * _6 = _1 | _3;
4705 : : *
4706 : : * And then simplied to
4707 : : * _6 = .SAT_TRUNC (x_4(D));
4708 : : */
4709 : :
4710 : : static gimple *
4711 : 27656303 : vect_recog_sat_trunc_pattern (vec_info *vinfo, stmt_vec_info stmt_vinfo,
4712 : : tree *type_out)
4713 : : {
4714 : 27656303 : gimple *last_stmt = STMT_VINFO_STMT (stmt_vinfo);
4715 : :
4716 : 27656303 : if (!is_gimple_assign (last_stmt))
4717 : : return NULL;
4718 : :
4719 : 18951254 : tree ops[1];
4720 : 18951254 : tree lhs = gimple_assign_lhs (last_stmt);
4721 : 18951254 : tree otype = TREE_TYPE (lhs);
4722 : :
4723 : 18951254 : if ((gimple_unsigned_integer_sat_trunc (lhs, ops, NULL)
4724 : 18951039 : || gimple_signed_integer_sat_trunc (lhs, ops, NULL))
4725 : 18951254 : && type_has_mode_precision_p (otype))
4726 : : {
4727 : 203 : tree itype = TREE_TYPE (ops[0]);
4728 : 203 : tree v_itype = get_vectype_for_scalar_type (vinfo, itype);
4729 : 203 : tree v_otype = get_vectype_for_scalar_type (vinfo, otype);
4730 : 203 : internal_fn fn = IFN_SAT_TRUNC;
4731 : :
4732 : 197 : if (v_itype != NULL_TREE && v_otype != NULL_TREE
4733 : 400 : && direct_internal_fn_supported_p (fn, tree_pair (v_otype, v_itype),
4734 : : OPTIMIZE_FOR_BOTH))
4735 : : {
4736 : 0 : gcall *call = gimple_build_call_internal (fn, 1, ops[0]);
4737 : 0 : tree out_ssa = vect_recog_temp_ssa_var (otype, NULL);
4738 : :
4739 : 0 : gimple_call_set_lhs (call, out_ssa);
4740 : 0 : gimple_call_set_nothrow (call, /* nothrow_p */ false);
4741 : 0 : gimple_set_location (call, gimple_location (last_stmt));
4742 : :
4743 : 0 : *type_out = v_otype;
4744 : :
4745 : 0 : return call;
4746 : : }
4747 : : }
4748 : :
4749 : : return NULL;
4750 : : }
4751 : :
4752 : : /* Detect a signed division by a constant that wouldn't be
4753 : : otherwise vectorized:
4754 : :
4755 : : type a_t, b_t;
4756 : :
4757 : : S1 a_t = b_t / N;
4758 : :
4759 : : where type 'type' is an integral type and N is a constant.
4760 : :
4761 : : Similarly handle modulo by a constant:
4762 : :
4763 : : S4 a_t = b_t % N;
4764 : :
4765 : : Input/Output:
4766 : :
4767 : : * STMT_VINFO: The stmt from which the pattern search begins,
4768 : : i.e. the division stmt. S1 is replaced by if N is a power
4769 : : of two constant and type is signed:
4770 : : S3 y_t = b_t < 0 ? N - 1 : 0;
4771 : : S2 x_t = b_t + y_t;
4772 : : S1' a_t = x_t >> log2 (N);
4773 : :
4774 : : S4 is replaced if N is a power of two constant and
4775 : : type is signed by (where *_T temporaries have unsigned type):
4776 : : S9 y_T = b_t < 0 ? -1U : 0U;
4777 : : S8 z_T = y_T >> (sizeof (type_t) * CHAR_BIT - log2 (N));
4778 : : S7 z_t = (type) z_T;
4779 : : S6 w_t = b_t + z_t;
4780 : : S5 x_t = w_t & (N - 1);
4781 : : S4' a_t = x_t - z_t;
4782 : :
4783 : : Output:
4784 : :
4785 : : * TYPE_OUT: The type of the output of this pattern.
4786 : :
4787 : : * Return value: A new stmt that will be used to replace the division
4788 : : S1 or modulo S4 stmt. */
4789 : :
4790 : : static gimple *
4791 : 27447593 : vect_recog_divmod_pattern (vec_info *vinfo,
4792 : : stmt_vec_info stmt_vinfo, tree *type_out)
4793 : : {
4794 : 27447593 : gimple *last_stmt = stmt_vinfo->stmt;
4795 : 27447593 : tree oprnd0, oprnd1, vectype, itype, cond;
4796 : 27447593 : gimple *pattern_stmt, *def_stmt;
4797 : 27447593 : enum tree_code rhs_code;
4798 : 27447593 : optab optab;
4799 : 27447593 : tree q, cst;
4800 : 27447593 : int prec;
4801 : :
4802 : 27447593 : if (!is_gimple_assign (last_stmt))
4803 : : return NULL;
4804 : :
4805 : 18742544 : rhs_code = gimple_assign_rhs_code (last_stmt);
4806 : 18742544 : switch (rhs_code)
4807 : : {
4808 : 245773 : case TRUNC_DIV_EXPR:
4809 : 245773 : case EXACT_DIV_EXPR:
4810 : 245773 : case TRUNC_MOD_EXPR:
4811 : 245773 : break;
4812 : : default:
4813 : : return NULL;
4814 : : }
4815 : :
4816 : 245773 : oprnd0 = gimple_assign_rhs1 (last_stmt);
4817 : 245773 : oprnd1 = gimple_assign_rhs2 (last_stmt);
4818 : 245773 : itype = TREE_TYPE (oprnd0);
4819 : 245773 : if (TREE_CODE (oprnd0) != SSA_NAME
4820 : 229144 : || TREE_CODE (oprnd1) != INTEGER_CST
4821 : 136526 : || TREE_CODE (itype) != INTEGER_TYPE
4822 : 382299 : || !type_has_mode_precision_p (itype))
4823 : 109247 : return NULL;
4824 : :
4825 : 136526 : scalar_int_mode itype_mode = SCALAR_INT_TYPE_MODE (itype);
4826 : 136526 : vectype = get_vectype_for_scalar_type (vinfo, itype);
4827 : 136526 : if (vectype == NULL_TREE)
4828 : : return NULL;
4829 : :
4830 : 111437 : if (optimize_bb_for_size_p (gimple_bb (last_stmt)))
4831 : : {
4832 : : /* If the target can handle vectorized division or modulo natively,
4833 : : don't attempt to optimize this, since native division is likely
4834 : : to give smaller code. */
4835 : 1134 : optab = optab_for_tree_code (rhs_code, vectype, optab_default);
4836 : 1134 : if (optab != unknown_optab
4837 : 1134 : && can_implement_p (optab, TYPE_MODE (vectype)))
4838 : : return NULL;
4839 : : }
4840 : :
4841 : 111437 : prec = TYPE_PRECISION (itype);
4842 : 111437 : if (integer_pow2p (oprnd1))
4843 : : {
4844 : 53417 : if (TYPE_UNSIGNED (itype) || tree_int_cst_sgn (oprnd1) != 1)
4845 : 3640 : return NULL;
4846 : :
4847 : : /* Pattern detected. */
4848 : 49777 : vect_pattern_detected ("vect_recog_divmod_pattern", last_stmt);
4849 : :
4850 : 49777 : *type_out = vectype;
4851 : :
4852 : : /* Check if the target supports this internal function. */
4853 : 49777 : internal_fn ifn = IFN_DIV_POW2;
4854 : 49777 : if (direct_internal_fn_supported_p (ifn, vectype, OPTIMIZE_FOR_SPEED))
4855 : : {
4856 : 0 : tree shift = build_int_cst (itype, tree_log2 (oprnd1));
4857 : :
4858 : 0 : tree var_div = vect_recog_temp_ssa_var (itype, NULL);
4859 : 0 : gimple *div_stmt = gimple_build_call_internal (ifn, 2, oprnd0, shift);
4860 : 0 : gimple_call_set_lhs (div_stmt, var_div);
4861 : :
4862 : 0 : if (rhs_code == TRUNC_MOD_EXPR)
4863 : : {
4864 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, div_stmt);
4865 : 0 : def_stmt
4866 : 0 : = gimple_build_assign (vect_recog_temp_ssa_var (itype, NULL),
4867 : : LSHIFT_EXPR, var_div, shift);
4868 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
4869 : 0 : pattern_stmt
4870 : 0 : = gimple_build_assign (vect_recog_temp_ssa_var (itype, NULL),
4871 : : MINUS_EXPR, oprnd0,
4872 : : gimple_assign_lhs (def_stmt));
4873 : : }
4874 : : else
4875 : : pattern_stmt = div_stmt;
4876 : 0 : gimple_set_location (pattern_stmt, gimple_location (last_stmt));
4877 : :
4878 : 0 : return pattern_stmt;
4879 : : }
4880 : :
4881 : 49777 : cond = vect_recog_temp_ssa_var (boolean_type_node, NULL);
4882 : 49777 : def_stmt = gimple_build_assign (cond, LT_EXPR, oprnd0,
4883 : 49777 : build_int_cst (itype, 0));
4884 : 49777 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt,
4885 : : truth_type_for (vectype), itype);
4886 : 49777 : if (rhs_code == TRUNC_DIV_EXPR
4887 : 49777 : || rhs_code == EXACT_DIV_EXPR)
4888 : : {
4889 : 47183 : tree var = vect_recog_temp_ssa_var (itype, NULL);
4890 : 47183 : tree shift;
4891 : 47183 : def_stmt
4892 : 47183 : = gimple_build_assign (var, COND_EXPR, cond,
4893 : 47183 : fold_build2 (MINUS_EXPR, itype, oprnd1,
4894 : : build_int_cst (itype, 1)),
4895 : 47183 : build_int_cst (itype, 0));
4896 : 47183 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
4897 : 47183 : var = vect_recog_temp_ssa_var (itype, NULL);
4898 : 47183 : def_stmt
4899 : 47183 : = gimple_build_assign (var, PLUS_EXPR, oprnd0,
4900 : : gimple_assign_lhs (def_stmt));
4901 : 47183 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
4902 : :
4903 : 47183 : shift = build_int_cst (itype, tree_log2 (oprnd1));
4904 : 47183 : pattern_stmt
4905 : 47183 : = gimple_build_assign (vect_recog_temp_ssa_var (itype, NULL),
4906 : : RSHIFT_EXPR, var, shift);
4907 : : }
4908 : : else
4909 : : {
4910 : 2594 : tree signmask;
4911 : 2594 : if (compare_tree_int (oprnd1, 2) == 0)
4912 : : {
4913 : 1091 : signmask = vect_recog_temp_ssa_var (itype, NULL);
4914 : 1091 : def_stmt = gimple_build_assign (signmask, COND_EXPR, cond,
4915 : 1091 : build_int_cst (itype, 1),
4916 : 1091 : build_int_cst (itype, 0));
4917 : 1091 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
4918 : : }
4919 : : else
4920 : : {
4921 : 1503 : tree utype
4922 : 1503 : = build_nonstandard_integer_type (prec, 1);
4923 : 1503 : tree vecutype = get_vectype_for_scalar_type (vinfo, utype);
4924 : 1503 : tree shift
4925 : 1503 : = build_int_cst (utype, GET_MODE_BITSIZE (itype_mode)
4926 : 1503 : - tree_log2 (oprnd1));
4927 : 1503 : tree var = vect_recog_temp_ssa_var (utype, NULL);
4928 : :
4929 : 1503 : def_stmt = gimple_build_assign (var, COND_EXPR, cond,
4930 : 1503 : build_int_cst (utype, -1),
4931 : 1503 : build_int_cst (utype, 0));
4932 : 1503 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, vecutype);
4933 : 1503 : var = vect_recog_temp_ssa_var (utype, NULL);
4934 : 1503 : def_stmt = gimple_build_assign (var, RSHIFT_EXPR,
4935 : : gimple_assign_lhs (def_stmt),
4936 : : shift);
4937 : 1503 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, vecutype);
4938 : 1503 : signmask = vect_recog_temp_ssa_var (itype, NULL);
4939 : 1503 : def_stmt
4940 : 1503 : = gimple_build_assign (signmask, NOP_EXPR, var);
4941 : 1503 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
4942 : : }
4943 : 2594 : def_stmt
4944 : 2594 : = gimple_build_assign (vect_recog_temp_ssa_var (itype, NULL),
4945 : : PLUS_EXPR, oprnd0, signmask);
4946 : 2594 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
4947 : 2594 : def_stmt
4948 : 2594 : = gimple_build_assign (vect_recog_temp_ssa_var (itype, NULL),
4949 : : BIT_AND_EXPR, gimple_assign_lhs (def_stmt),
4950 : 2594 : fold_build2 (MINUS_EXPR, itype, oprnd1,
4951 : : build_int_cst (itype, 1)));
4952 : 2594 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
4953 : :
4954 : 2594 : pattern_stmt
4955 : 2594 : = gimple_build_assign (vect_recog_temp_ssa_var (itype, NULL),
4956 : : MINUS_EXPR, gimple_assign_lhs (def_stmt),
4957 : : signmask);
4958 : : }
4959 : :
4960 : 49777 : return pattern_stmt;
4961 : : }
4962 : :
4963 : 58020 : if ((cst = uniform_integer_cst_p (oprnd1))
4964 : 58020 : && TYPE_UNSIGNED (itype)
4965 : : && rhs_code == TRUNC_DIV_EXPR
4966 : 34088 : && vectype
4967 : 80510 : && targetm.vectorize.preferred_div_as_shifts_over_mult (vectype))
4968 : : {
4969 : : /* We can use the relationship:
4970 : :
4971 : : x // N == ((x+N+2) // (N+1) + x) // (N+1) for 0 <= x < N(N+3)
4972 : :
4973 : : to optimize cases where N+1 is a power of 2, and where // (N+1)
4974 : : is therefore a shift right. When operating in modes that are
4975 : : multiples of a byte in size, there are two cases:
4976 : :
4977 : : (1) N(N+3) is not representable, in which case the question
4978 : : becomes whether the replacement expression overflows.
4979 : : It is enough to test that x+N+2 does not overflow,
4980 : : i.e. that x < MAX-(N+1).
4981 : :
4982 : : (2) N(N+3) is representable, in which case it is the (only)
4983 : : bound that we need to check.
4984 : :
4985 : : ??? For now we just handle the case where // (N+1) is a shift
4986 : : right by half the precision, since some architectures can
4987 : : optimize the associated addition and shift combinations
4988 : : into single instructions. */
4989 : :
4990 : 15106 : auto wcst = wi::to_wide (cst);
4991 : 15106 : int pow = wi::exact_log2 (wcst + 1);
4992 : 15106 : if (pow == prec / 2)
4993 : : {
4994 : 472 : gimple *stmt = SSA_NAME_DEF_STMT (oprnd0);
4995 : :
4996 : 472 : gimple_ranger ranger;
4997 : 472 : int_range_max r;
4998 : :
4999 : : /* Check that no overflow will occur. If we don't have range
5000 : : information we can't perform the optimization. */
5001 : :
5002 : 472 : if (ranger.range_of_expr (r, oprnd0, stmt) && !r.undefined_p ())
5003 : : {
5004 : 470 : wide_int max = r.upper_bound ();
5005 : 470 : wide_int one = wi::shwi (1, prec);
5006 : 470 : wide_int adder = wi::add (one, wi::lshift (one, pow));
5007 : 470 : wi::overflow_type ovf;
5008 : 470 : wi::add (max, adder, UNSIGNED, &ovf);
5009 : 470 : if (ovf == wi::OVF_NONE)
5010 : : {
5011 : 305 : *type_out = vectype;
5012 : 305 : tree tadder = wide_int_to_tree (itype, adder);
5013 : 305 : tree rshift = wide_int_to_tree (itype, pow);
5014 : :
5015 : 305 : tree new_lhs1 = vect_recog_temp_ssa_var (itype, NULL);
5016 : 305 : gassign *patt1
5017 : 305 : = gimple_build_assign (new_lhs1, PLUS_EXPR, oprnd0, tadder);
5018 : 305 : append_pattern_def_seq (vinfo, stmt_vinfo, patt1, vectype);
5019 : :
5020 : 305 : tree new_lhs2 = vect_recog_temp_ssa_var (itype, NULL);
5021 : 305 : patt1 = gimple_build_assign (new_lhs2, RSHIFT_EXPR, new_lhs1,
5022 : : rshift);
5023 : 305 : append_pattern_def_seq (vinfo, stmt_vinfo, patt1, vectype);
5024 : :
5025 : 305 : tree new_lhs3 = vect_recog_temp_ssa_var (itype, NULL);
5026 : 305 : patt1 = gimple_build_assign (new_lhs3, PLUS_EXPR, new_lhs2,
5027 : : oprnd0);
5028 : 305 : append_pattern_def_seq (vinfo, stmt_vinfo, patt1, vectype);
5029 : :
5030 : 305 : tree new_lhs4 = vect_recog_temp_ssa_var (itype, NULL);
5031 : 305 : pattern_stmt = gimple_build_assign (new_lhs4, RSHIFT_EXPR,
5032 : : new_lhs3, rshift);
5033 : :
5034 : 305 : return pattern_stmt;
5035 : : }
5036 : 470 : }
5037 : 472 : }
5038 : : }
5039 : :
5040 : 57715 : if (prec > HOST_BITS_PER_WIDE_INT
5041 : 57715 : || integer_zerop (oprnd1))
5042 : 500 : return NULL;
5043 : :
5044 : 57215 : if (!can_mult_highpart_p (TYPE_MODE (vectype), TYPE_UNSIGNED (itype)))
5045 : : return NULL;
5046 : :
5047 : 15257 : if (TYPE_UNSIGNED (itype))
5048 : : {
5049 : 10190 : unsigned HOST_WIDE_INT mh, ml;
5050 : 10190 : int pre_shift, post_shift;
5051 : 10190 : unsigned HOST_WIDE_INT d = (TREE_INT_CST_LOW (oprnd1)
5052 : 10190 : & GET_MODE_MASK (itype_mode));
5053 : 10190 : tree t1, t2, t3, t4;
5054 : :
5055 : 10190 : if (d >= (HOST_WIDE_INT_1U << (prec - 1)))
5056 : : /* FIXME: Can transform this into oprnd0 >= oprnd1 ? 1 : 0. */
5057 : 22 : return NULL;
5058 : :
5059 : : /* Find a suitable multiplier and right shift count instead of
5060 : : directly dividing by D. */
5061 : 10168 : mh = choose_multiplier (d, prec, prec, &ml, &post_shift);
5062 : :
5063 : : /* If the suggested multiplier is more than PREC bits, we can do better
5064 : : for even divisors, using an initial right shift. */
5065 : 10168 : if (mh != 0 && (d & 1) == 0)
5066 : : {
5067 : 374 : pre_shift = ctz_or_zero (d);
5068 : 374 : mh = choose_multiplier (d >> pre_shift, prec, prec - pre_shift,
5069 : : &ml, &post_shift);
5070 : 374 : gcc_assert (!mh);
5071 : : }
5072 : : else
5073 : : pre_shift = 0;
5074 : :
5075 : 686 : if (mh != 0)
5076 : : {
5077 : 686 : if (post_shift - 1 >= prec)
5078 : : return NULL;
5079 : :
5080 : : /* t1 = oprnd0 h* ml;
5081 : : t2 = oprnd0 - t1;
5082 : : t3 = t2 >> 1;
5083 : : t4 = t1 + t3;
5084 : : q = t4 >> (post_shift - 1); */
5085 : 686 : t1 = vect_recog_temp_ssa_var (itype, NULL);
5086 : 686 : def_stmt = gimple_build_assign (t1, MULT_HIGHPART_EXPR, oprnd0,
5087 : 686 : build_int_cst (itype, ml));
5088 : 686 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5089 : :
5090 : 686 : t2 = vect_recog_temp_ssa_var (itype, NULL);
5091 : 686 : def_stmt
5092 : 686 : = gimple_build_assign (t2, MINUS_EXPR, oprnd0, t1);
5093 : 686 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5094 : :
5095 : 686 : t3 = vect_recog_temp_ssa_var (itype, NULL);
5096 : 686 : def_stmt
5097 : 686 : = gimple_build_assign (t3, RSHIFT_EXPR, t2, integer_one_node);
5098 : 686 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5099 : :
5100 : 686 : t4 = vect_recog_temp_ssa_var (itype, NULL);
5101 : 686 : def_stmt
5102 : 686 : = gimple_build_assign (t4, PLUS_EXPR, t1, t3);
5103 : :
5104 : 686 : if (post_shift != 1)
5105 : : {
5106 : 686 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5107 : :
5108 : 686 : q = vect_recog_temp_ssa_var (itype, NULL);
5109 : 686 : pattern_stmt
5110 : 686 : = gimple_build_assign (q, RSHIFT_EXPR, t4,
5111 : 686 : build_int_cst (itype, post_shift - 1));
5112 : : }
5113 : : else
5114 : : {
5115 : : q = t4;
5116 : : pattern_stmt = def_stmt;
5117 : : }
5118 : : }
5119 : : else
5120 : : {
5121 : 9482 : if (pre_shift >= prec || post_shift >= prec)
5122 : : return NULL;
5123 : :
5124 : : /* t1 = oprnd0 >> pre_shift;
5125 : : t2 = t1 h* ml;
5126 : : q = t2 >> post_shift; */
5127 : 9482 : if (pre_shift)
5128 : : {
5129 : 374 : t1 = vect_recog_temp_ssa_var (itype, NULL);
5130 : 374 : def_stmt
5131 : 374 : = gimple_build_assign (t1, RSHIFT_EXPR, oprnd0,
5132 : 374 : build_int_cst (NULL, pre_shift));
5133 : 374 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5134 : : }
5135 : : else
5136 : : t1 = oprnd0;
5137 : :
5138 : 9482 : t2 = vect_recog_temp_ssa_var (itype, NULL);
5139 : 9482 : def_stmt = gimple_build_assign (t2, MULT_HIGHPART_EXPR, t1,
5140 : 9482 : build_int_cst (itype, ml));
5141 : :
5142 : 9482 : if (post_shift)
5143 : : {
5144 : 9476 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5145 : :
5146 : 9476 : q = vect_recog_temp_ssa_var (itype, NULL);
5147 : 9476 : def_stmt
5148 : 9476 : = gimple_build_assign (q, RSHIFT_EXPR, t2,
5149 : 9476 : build_int_cst (itype, post_shift));
5150 : : }
5151 : : else
5152 : : q = t2;
5153 : :
5154 : : pattern_stmt = def_stmt;
5155 : : }
5156 : : }
5157 : : else
5158 : : {
5159 : 5067 : unsigned HOST_WIDE_INT ml;
5160 : 5067 : int post_shift;
5161 : 5067 : HOST_WIDE_INT d = TREE_INT_CST_LOW (oprnd1);
5162 : 5067 : unsigned HOST_WIDE_INT abs_d;
5163 : 5067 : bool add = false;
5164 : 5067 : tree t1, t2, t3, t4;
5165 : :
5166 : : /* Give up for -1. */
5167 : 5067 : if (d == -1)
5168 : 0 : return NULL;
5169 : :
5170 : : /* Since d might be INT_MIN, we have to cast to
5171 : : unsigned HOST_WIDE_INT before negating to avoid
5172 : : undefined signed overflow. */
5173 : 5067 : abs_d = (d >= 0
5174 : 5067 : ? (unsigned HOST_WIDE_INT) d
5175 : : : - (unsigned HOST_WIDE_INT) d);
5176 : :
5177 : : /* n rem d = n rem -d */
5178 : 5067 : if (rhs_code == TRUNC_MOD_EXPR && d < 0)
5179 : : {
5180 : 0 : d = abs_d;
5181 : 0 : oprnd1 = build_int_cst (itype, abs_d);
5182 : : }
5183 : 5067 : if (HOST_BITS_PER_WIDE_INT >= prec
5184 : 5067 : && abs_d == HOST_WIDE_INT_1U << (prec - 1))
5185 : : /* This case is not handled correctly below. */
5186 : : return NULL;
5187 : :
5188 : 5067 : choose_multiplier (abs_d, prec, prec - 1, &ml, &post_shift);
5189 : 5067 : if (ml >= HOST_WIDE_INT_1U << (prec - 1))
5190 : : {
5191 : 1448 : add = true;
5192 : 1448 : ml |= HOST_WIDE_INT_M1U << (prec - 1);
5193 : : }
5194 : 5067 : if (post_shift >= prec)
5195 : : return NULL;
5196 : :
5197 : : /* t1 = oprnd0 h* ml; */
5198 : 5067 : t1 = vect_recog_temp_ssa_var (itype, NULL);
5199 : 5067 : def_stmt = gimple_build_assign (t1, MULT_HIGHPART_EXPR, oprnd0,
5200 : 5067 : build_int_cst (itype, ml));
5201 : :
5202 : 5067 : if (add)
5203 : : {
5204 : : /* t2 = t1 + oprnd0; */
5205 : 1448 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5206 : 1448 : t2 = vect_recog_temp_ssa_var (itype, NULL);
5207 : 1448 : def_stmt = gimple_build_assign (t2, PLUS_EXPR, t1, oprnd0);
5208 : : }
5209 : : else
5210 : : t2 = t1;
5211 : :
5212 : 5067 : if (post_shift)
5213 : : {
5214 : : /* t3 = t2 >> post_shift; */
5215 : 4437 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5216 : 4437 : t3 = vect_recog_temp_ssa_var (itype, NULL);
5217 : 4437 : def_stmt = gimple_build_assign (t3, RSHIFT_EXPR, t2,
5218 : 4437 : build_int_cst (itype, post_shift));
5219 : : }
5220 : : else
5221 : : t3 = t2;
5222 : :
5223 : 5067 : int msb = 1;
5224 : 5067 : int_range_max r;
5225 : 10134 : get_range_query (cfun)->range_of_expr (r, oprnd0);
5226 : 5067 : if (!r.varying_p () && !r.undefined_p ())
5227 : : {
5228 : 2994 : if (!wi::neg_p (r.lower_bound (), TYPE_SIGN (itype)))
5229 : : msb = 0;
5230 : 688 : else if (wi::neg_p (r.upper_bound (), TYPE_SIGN (itype)))
5231 : : msb = -1;
5232 : : }
5233 : :
5234 : 2306 : if (msb == 0 && d >= 0)
5235 : : {
5236 : : /* q = t3; */
5237 : : q = t3;
5238 : : pattern_stmt = def_stmt;
5239 : : }
5240 : : else
5241 : : {
5242 : : /* t4 = oprnd0 >> (prec - 1);
5243 : : or if we know from VRP that oprnd0 >= 0
5244 : : t4 = 0;
5245 : : or if we know from VRP that oprnd0 < 0
5246 : : t4 = -1; */
5247 : 2827 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5248 : 2827 : t4 = vect_recog_temp_ssa_var (itype, NULL);
5249 : 2827 : if (msb != 1)
5250 : 74 : def_stmt = gimple_build_assign (t4, INTEGER_CST,
5251 : 74 : build_int_cst (itype, msb));
5252 : : else
5253 : 2753 : def_stmt = gimple_build_assign (t4, RSHIFT_EXPR, oprnd0,
5254 : 2753 : build_int_cst (itype, prec - 1));
5255 : 2827 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5256 : :
5257 : : /* q = t3 - t4; or q = t4 - t3; */
5258 : 2827 : q = vect_recog_temp_ssa_var (itype, NULL);
5259 : 5493 : pattern_stmt = gimple_build_assign (q, MINUS_EXPR, d < 0 ? t4 : t3,
5260 : : d < 0 ? t3 : t4);
5261 : : }
5262 : 5067 : }
5263 : :
5264 : 15235 : if (rhs_code == TRUNC_MOD_EXPR)
5265 : : {
5266 : 6121 : tree r, t1;
5267 : :
5268 : : /* We divided. Now finish by:
5269 : : t1 = q * oprnd1;
5270 : : r = oprnd0 - t1; */
5271 : 6121 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt);
5272 : :
5273 : 6121 : t1 = vect_recog_temp_ssa_var (itype, NULL);
5274 : 6121 : def_stmt = gimple_build_assign (t1, MULT_EXPR, q, oprnd1);
5275 : 6121 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
5276 : :
5277 : 6121 : r = vect_recog_temp_ssa_var (itype, NULL);
5278 : 6121 : pattern_stmt = gimple_build_assign (r, MINUS_EXPR, oprnd0, t1);
5279 : : }
5280 : :
5281 : : /* Pattern detected. */
5282 : 15235 : vect_pattern_detected ("vect_recog_divmod_pattern", last_stmt);
5283 : :
5284 : 15235 : *type_out = vectype;
5285 : 15235 : return pattern_stmt;
5286 : : }
5287 : :
5288 : : /* Detects pattern with a modulo operation (S1) where both arguments
5289 : : are variables of integral type.
5290 : : The statement is replaced by division, multiplication, and subtraction.
5291 : : The last statement (S4) is returned.
5292 : :
5293 : : Example:
5294 : : S1 c_t = a_t % b_t;
5295 : :
5296 : : is replaced by
5297 : : S2 x_t = a_t / b_t;
5298 : : S3 y_t = x_t * b_t;
5299 : : S4 z_t = a_t - y_t; */
5300 : :
5301 : : static gimple *
5302 : 27578497 : vect_recog_mod_var_pattern (vec_info *vinfo,
5303 : : stmt_vec_info stmt_vinfo, tree *type_out)
5304 : : {
5305 : 27578497 : gimple *last_stmt = STMT_VINFO_STMT (stmt_vinfo);
5306 : 27578497 : tree oprnd0, oprnd1, vectype, itype;
5307 : 27578497 : gimple *pattern_stmt, *def_stmt;
5308 : 27578497 : enum tree_code rhs_code;
5309 : :
5310 : 27578497 : if (!is_gimple_assign (last_stmt))
5311 : : return NULL;
5312 : :
5313 : 18873448 : rhs_code = gimple_assign_rhs_code (last_stmt);
5314 : 18873448 : if (rhs_code != TRUNC_MOD_EXPR)
5315 : : return NULL;
5316 : :
5317 : 68085 : oprnd0 = gimple_assign_rhs1 (last_stmt);
5318 : 68085 : oprnd1 = gimple_assign_rhs2 (last_stmt);
5319 : 68085 : itype = TREE_TYPE (oprnd0);
5320 : 68085 : if (TREE_CODE (oprnd0) != SSA_NAME
5321 : 60284 : || TREE_CODE (oprnd1) != SSA_NAME
5322 : 42628 : || TREE_CODE (itype) != INTEGER_TYPE)
5323 : : return NULL;
5324 : :
5325 : 42524 : vectype = get_vectype_for_scalar_type (vinfo, itype);
5326 : :
5327 : 42524 : if (!vectype
5328 : 33564 : || target_has_vecop_for_code (TRUNC_MOD_EXPR, vectype)
5329 : 33564 : || !target_has_vecop_for_code (TRUNC_DIV_EXPR, vectype)
5330 : 0 : || !target_has_vecop_for_code (MULT_EXPR, vectype)
5331 : 42524 : || !target_has_vecop_for_code (MINUS_EXPR, vectype))
5332 : 42524 : return NULL;
5333 : :
5334 : 0 : tree q, tmp, r;
5335 : 0 : q = vect_recog_temp_ssa_var (itype, NULL);
5336 : 0 : def_stmt = gimple_build_assign (q, TRUNC_DIV_EXPR, oprnd0, oprnd1);
5337 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, vectype);
5338 : :
5339 : 0 : tmp = vect_recog_temp_ssa_var (itype, NULL);
5340 : 0 : def_stmt = gimple_build_assign (tmp, MULT_EXPR, q, oprnd1);
5341 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, vectype);
5342 : :
5343 : 0 : r = vect_recog_temp_ssa_var (itype, NULL);
5344 : 0 : pattern_stmt = gimple_build_assign (r, MINUS_EXPR, oprnd0, tmp);
5345 : :
5346 : : /* Pattern detected. */
5347 : 0 : *type_out = vectype;
5348 : 0 : vect_pattern_detected ("vect_recog_mod_var_pattern", last_stmt);
5349 : :
5350 : 0 : return pattern_stmt;
5351 : : }
5352 : :
5353 : :
5354 : : /* Return the proper type for converting bool VAR into
5355 : : an integer value or NULL_TREE if no such type exists.
5356 : : The type is chosen so that the converted value has the
5357 : : same number of elements as VAR's vector type. */
5358 : :
5359 : : static tree
5360 : 3224109 : integer_type_for_mask (tree var, vec_info *vinfo)
5361 : : {
5362 : 3224109 : if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (var)))
5363 : : return NULL_TREE;
5364 : :
5365 : 1136719 : stmt_vec_info def_stmt_info = vect_get_internal_def (vinfo, var);
5366 : 1136719 : if (!def_stmt_info || !vect_use_mask_type_p (def_stmt_info))
5367 : : return NULL_TREE;
5368 : :
5369 : 606722 : return build_nonstandard_integer_type (def_stmt_info->mask_precision, 1);
5370 : : }
5371 : :
5372 : : /* Function vect_recog_gcond_pattern
5373 : :
5374 : : Try to find pattern like following:
5375 : :
5376 : : if (a op b)
5377 : :
5378 : : where operator 'op' is not != and convert it to an adjusted boolean pattern
5379 : :
5380 : : mask = a op b
5381 : : if (mask != 0)
5382 : :
5383 : : and set the mask type on MASK.
5384 : :
5385 : : Input:
5386 : :
5387 : : * STMT_VINFO: The stmt at the end from which the pattern
5388 : : search begins, i.e. cast of a bool to
5389 : : an integer type.
5390 : :
5391 : : Output:
5392 : :
5393 : : * TYPE_OUT: The type of the output of this pattern.
5394 : :
5395 : : * Return value: A new stmt that will be used to replace the pattern. */
5396 : :
5397 : : static gimple *
5398 : 27656303 : vect_recog_gcond_pattern (vec_info *vinfo,
5399 : : stmt_vec_info stmt_vinfo, tree *type_out)
5400 : : {
5401 : : /* Currently we only support this for loop vectorization and when multiple
5402 : : exits. */
5403 : 27656303 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
5404 : 3160762 : if (!loop_vinfo || !LOOP_VINFO_EARLY_BREAKS (loop_vinfo))
5405 : : return NULL;
5406 : :
5407 : 1005576 : gimple *last_stmt = STMT_VINFO_STMT (stmt_vinfo);
5408 : 1005576 : gcond* cond = NULL;
5409 : 27671641 : if (!(cond = dyn_cast <gcond *> (last_stmt)))
5410 : : return NULL;
5411 : :
5412 : 255713 : auto lhs = gimple_cond_lhs (cond);
5413 : 255713 : auto rhs = gimple_cond_rhs (cond);
5414 : 255713 : auto code = gimple_cond_code (cond);
5415 : :
5416 : 255713 : tree scalar_type = TREE_TYPE (lhs);
5417 : 255713 : if (VECTOR_TYPE_P (scalar_type))
5418 : : return NULL;
5419 : :
5420 : : /* If the input is a boolean then try to figure out the precision that the
5421 : : vector type should use. We cannot use the scalar precision as this would
5422 : : later mismatch. This is similar to what recog_bool does. */
5423 : 255713 : if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type))
5424 : : {
5425 : 10483 : if (tree stype = integer_type_for_mask (lhs, vinfo))
5426 : 255713 : scalar_type = stype;
5427 : : }
5428 : :
5429 : 255713 : tree vectype = get_mask_type_for_scalar_type (vinfo, scalar_type);
5430 : 255713 : if (vectype == NULL_TREE)
5431 : : return NULL;
5432 : :
5433 : 240375 : tree new_lhs = vect_recog_temp_ssa_var (boolean_type_node, NULL);
5434 : 240375 : gimple *new_stmt = gimple_build_assign (new_lhs, code, lhs, rhs);
5435 : 240375 : append_pattern_def_seq (vinfo, stmt_vinfo, new_stmt, vectype, scalar_type);
5436 : :
5437 : 240375 : gimple *pattern_stmt
5438 : 240375 : = gimple_build_cond (NE_EXPR, new_lhs,
5439 : 240375 : build_int_cst (TREE_TYPE (new_lhs), 0),
5440 : : NULL_TREE, NULL_TREE);
5441 : 240375 : *type_out = vectype;
5442 : 240375 : vect_pattern_detected ("vect_recog_gcond_pattern", last_stmt);
5443 : 240375 : return pattern_stmt;
5444 : : }
5445 : :
5446 : : /* Function vect_recog_bool_pattern
5447 : :
5448 : : Try to find pattern like following:
5449 : :
5450 : : bool a_b, b_b, c_b, d_b, e_b;
5451 : : TYPE f_T;
5452 : : loop:
5453 : : S1 a_b = x1 CMP1 y1;
5454 : : S2 b_b = x2 CMP2 y2;
5455 : : S3 c_b = a_b & b_b;
5456 : : S4 d_b = x3 CMP3 y3;
5457 : : S5 e_b = c_b | d_b;
5458 : : S6 f_T = (TYPE) e_b;
5459 : :
5460 : : where type 'TYPE' is an integral type. Or a similar pattern
5461 : : ending in
5462 : :
5463 : : S6 f_Y = e_b ? r_Y : s_Y;
5464 : :
5465 : : as results from if-conversion of a complex condition.
5466 : :
5467 : : Input:
5468 : :
5469 : : * STMT_VINFO: The stmt at the end from which the pattern
5470 : : search begins, i.e. cast of a bool to
5471 : : an integer type.
5472 : :
5473 : : Output:
5474 : :
5475 : : * TYPE_OUT: The type of the output of this pattern.
5476 : :
5477 : : * Return value: A new stmt that will be used to replace the pattern.
5478 : :
5479 : : Assuming size of TYPE is the same as size of all comparisons
5480 : : (otherwise some casts would be added where needed), the above
5481 : : sequence we create related pattern stmts:
5482 : : S1' a_T = x1 CMP1 y1 ? 1 : 0;
5483 : : S3' c_T = x2 CMP2 y2 ? a_T : 0;
5484 : : S4' d_T = x3 CMP3 y3 ? 1 : 0;
5485 : : S5' e_T = c_T | d_T;
5486 : : S6' f_T = e_T;
5487 : :
5488 : : Instead of the above S3' we could emit:
5489 : : S2' b_T = x2 CMP2 y2 ? 1 : 0;
5490 : : S3' c_T = a_T | b_T;
5491 : : but the above is more efficient. */
5492 : :
5493 : : static gimple *
5494 : 27656303 : vect_recog_bool_pattern (vec_info *vinfo,
5495 : : stmt_vec_info stmt_vinfo, tree *type_out)
5496 : : {
5497 : 27656303 : gimple *last_stmt = stmt_vinfo->stmt;
5498 : 27656303 : enum tree_code rhs_code;
5499 : 27656303 : tree var, lhs, rhs, vectype;
5500 : 27656303 : gimple *pattern_stmt;
5501 : :
5502 : 27656303 : if (!is_gimple_assign (last_stmt))
5503 : : return NULL;
5504 : :
5505 : 19191629 : var = gimple_assign_rhs1 (last_stmt);
5506 : 19191629 : lhs = gimple_assign_lhs (last_stmt);
5507 : 19191629 : rhs_code = gimple_assign_rhs_code (last_stmt);
5508 : :
5509 : 19191629 : if (rhs_code == VIEW_CONVERT_EXPR)
5510 : 167460 : var = TREE_OPERAND (var, 0);
5511 : :
5512 : 19191629 : if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (var)))
5513 : : return NULL;
5514 : :
5515 : 666513 : hash_set<gimple *> bool_stmts;
5516 : :
5517 : 666513 : if (CONVERT_EXPR_CODE_P (rhs_code)
5518 : 577839 : || rhs_code == VIEW_CONVERT_EXPR)
5519 : : {
5520 : 184152 : if (! INTEGRAL_TYPE_P (TREE_TYPE (lhs))
5521 : 184021 : || VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (lhs)))
5522 : : return NULL;
5523 : 88594 : vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (lhs));
5524 : :
5525 : 88594 : tree type = integer_type_for_mask (var, vinfo);
5526 : 88594 : tree cst0, cst1, tmp;
5527 : :
5528 : 88594 : if (!type)
5529 : : return NULL;
5530 : :
5531 : : /* We may directly use cond with narrowed type to avoid multiple cond
5532 : : exprs with following result packing and perform single cond with
5533 : : packed mask instead. In case of widening we better make cond first
5534 : : and then extract results. */
5535 : 40661 : if (TYPE_MODE (type) == TYPE_MODE (TREE_TYPE (lhs)))
5536 : 28279 : type = TREE_TYPE (lhs);
5537 : :
5538 : 40661 : cst0 = build_int_cst (type, 0);
5539 : 40661 : cst1 = build_int_cst (type, 1);
5540 : 40661 : tmp = vect_recog_temp_ssa_var (type, NULL);
5541 : 40661 : pattern_stmt = gimple_build_assign (tmp, COND_EXPR, var, cst1, cst0);
5542 : :
5543 : 40661 : if (!useless_type_conversion_p (type, TREE_TYPE (lhs)))
5544 : : {
5545 : 12382 : tree new_vectype = get_vectype_for_scalar_type (vinfo, type);
5546 : 12382 : append_pattern_def_seq (vinfo, stmt_vinfo,
5547 : : pattern_stmt, new_vectype);
5548 : :
5549 : 12382 : lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL);
5550 : 12382 : pattern_stmt = gimple_build_assign (lhs, CONVERT_EXPR, tmp);
5551 : : }
5552 : :
5553 : 40661 : *type_out = vectype;
5554 : 40661 : vect_pattern_detected ("vect_recog_bool_pattern", last_stmt);
5555 : :
5556 : 40661 : return pattern_stmt;
5557 : : }
5558 : 574050 : else if (rhs_code == COND_EXPR
5559 : 144861 : && TREE_CODE (var) == SSA_NAME)
5560 : : {
5561 : 144861 : vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (lhs));
5562 : 144861 : if (vectype == NULL_TREE)
5563 : : return NULL;
5564 : :
5565 : : /* Build a scalar type for the boolean result that when
5566 : : vectorized matches the vector type of the result in
5567 : : size and number of elements. */
5568 : 132954 : unsigned prec
5569 : 132954 : = vector_element_size (tree_to_poly_uint64 (TYPE_SIZE (vectype)),
5570 : : TYPE_VECTOR_SUBPARTS (vectype));
5571 : :
5572 : 132954 : tree type
5573 : 265908 : = build_nonstandard_integer_type (prec,
5574 : 132954 : TYPE_UNSIGNED (TREE_TYPE (var)));
5575 : 132954 : if (get_vectype_for_scalar_type (vinfo, type) == NULL_TREE)
5576 : : return NULL;
5577 : :
5578 : 132954 : enum vect_def_type dt;
5579 : 132954 : if (integer_type_for_mask (var, vinfo))
5580 : : return NULL;
5581 : 26189 : else if (TREE_CODE (TREE_TYPE (var)) == BOOLEAN_TYPE
5582 : 26189 : && vect_is_simple_use (var, vinfo, &dt)
5583 : 26189 : && (dt == vect_external_def
5584 : 26182 : || dt == vect_constant_def))
5585 : : {
5586 : : /* If the condition is already a boolean then manually convert it to a
5587 : : mask of the given integer type but don't set a vectype. */
5588 : 1338 : tree lhs_ivar = vect_recog_temp_ssa_var (type, NULL);
5589 : 1338 : pattern_stmt = gimple_build_assign (lhs_ivar, COND_EXPR, var,
5590 : : build_all_ones_cst (type),
5591 : : build_zero_cst (type));
5592 : 1338 : append_inv_pattern_def_seq (vinfo, pattern_stmt);
5593 : 1338 : var = lhs_ivar;
5594 : : }
5595 : :
5596 : 26189 : tree lhs_var = vect_recog_temp_ssa_var (boolean_type_node, NULL);
5597 : 26189 : pattern_stmt = gimple_build_assign (lhs_var, NE_EXPR, var,
5598 : 26189 : build_zero_cst (TREE_TYPE (var)));
5599 : :
5600 : 26189 : tree new_vectype = get_mask_type_for_scalar_type (vinfo, TREE_TYPE (var));
5601 : 26189 : if (!new_vectype)
5602 : : return NULL;
5603 : :
5604 : 26189 : new_vectype = truth_type_for (new_vectype);
5605 : 26189 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, new_vectype,
5606 : 26189 : TREE_TYPE (var));
5607 : :
5608 : 26189 : lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL);
5609 : 26189 : pattern_stmt
5610 : 26189 : = gimple_build_assign (lhs, COND_EXPR, lhs_var,
5611 : : gimple_assign_rhs2 (last_stmt),
5612 : : gimple_assign_rhs3 (last_stmt));
5613 : 26189 : *type_out = vectype;
5614 : 26189 : vect_pattern_detected ("vect_recog_bool_pattern", last_stmt);
5615 : :
5616 : 26189 : return pattern_stmt;
5617 : : }
5618 : 429189 : else if (rhs_code == SSA_NAME
5619 : 22541 : && STMT_VINFO_DATA_REF (stmt_vinfo))
5620 : : {
5621 : 7768 : stmt_vec_info pattern_stmt_info;
5622 : 7768 : vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (lhs));
5623 : 7768 : if (!vectype || !VECTOR_MODE_P (TYPE_MODE (vectype)))
5624 : 0 : return NULL;
5625 : :
5626 : 7768 : tree type = integer_type_for_mask (var, vinfo);
5627 : 7768 : tree cst0, cst1, new_vectype;
5628 : :
5629 : 7768 : if (!type)
5630 : : return NULL;
5631 : :
5632 : 4451 : if (TYPE_MODE (type) == TYPE_MODE (TREE_TYPE (vectype)))
5633 : 553 : type = TREE_TYPE (vectype);
5634 : :
5635 : 4451 : cst0 = build_int_cst (type, 0);
5636 : 4451 : cst1 = build_int_cst (type, 1);
5637 : 4451 : new_vectype = get_vectype_for_scalar_type (vinfo, type);
5638 : :
5639 : 4451 : rhs = vect_recog_temp_ssa_var (type, NULL);
5640 : 4451 : pattern_stmt = gimple_build_assign (rhs, COND_EXPR, var, cst1, cst0);
5641 : 4451 : append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, new_vectype);
5642 : :
5643 : 4451 : lhs = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (vectype), lhs);
5644 : 4451 : if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
5645 : : {
5646 : 3898 : tree rhs2 = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL);
5647 : 3898 : gimple *cast_stmt = gimple_build_assign (rhs2, NOP_EXPR, rhs);
5648 : 3898 : append_pattern_def_seq (vinfo, stmt_vinfo, cast_stmt);
5649 : 3898 : rhs = rhs2;
5650 : : }
5651 : 4451 : pattern_stmt = gimple_build_assign (lhs, SSA_NAME, rhs);
5652 : 4451 : pattern_stmt_info = vinfo->add_stmt (pattern_stmt);
5653 : 4451 : vinfo->move_dr (pattern_stmt_info, stmt_vinfo);
5654 : 4451 : *type_out = vectype;
5655 : 4451 : vect_pattern_detected ("vect_recog_bool_pattern", last_stmt);
5656 : :
5657 : 4451 : return pattern_stmt;
5658 : : }
5659 : : else
5660 : : return NULL;
5661 : 666513 : }
5662 : :
5663 : : /* A helper for vect_recog_mask_conversion_pattern. Build
5664 : : conversion of MASK to a type suitable for masking VECTYPE.
5665 : : Built statement gets required vectype and is appended to
5666 : : a pattern sequence of STMT_VINFO.
5667 : :
5668 : : Return converted mask. */
5669 : :
5670 : : static tree
5671 : 46781 : build_mask_conversion (vec_info *vinfo,
5672 : : tree mask, tree vectype, stmt_vec_info stmt_vinfo)
5673 : : {
5674 : 46781 : gimple *stmt;
5675 : 46781 : tree masktype, tmp;
5676 : :
5677 : 46781 : masktype = truth_type_for (vectype);
5678 : 46781 : tmp = vect_recog_temp_ssa_var (TREE_TYPE (masktype), NULL);
5679 : 46781 : stmt = gimple_build_assign (tmp, CONVERT_EXPR, mask);
5680 : 46781 : append_pattern_def_seq (vinfo, stmt_vinfo,
5681 : 46781 : stmt, masktype, TREE_TYPE (vectype));
5682 : :
5683 : 46781 : return tmp;
5684 : : }
5685 : :
5686 : :
5687 : : /* Function vect_recog_mask_conversion_pattern
5688 : :
5689 : : Try to find statements which require boolean type
5690 : : converison. Additional conversion statements are
5691 : : added to handle such cases. For example:
5692 : :
5693 : : bool m_1, m_2, m_3;
5694 : : int i_4, i_5;
5695 : : double d_6, d_7;
5696 : : char c_1, c_2, c_3;
5697 : :
5698 : : S1 m_1 = i_4 > i_5;
5699 : : S2 m_2 = d_6 < d_7;
5700 : : S3 m_3 = m_1 & m_2;
5701 : : S4 c_1 = m_3 ? c_2 : c_3;
5702 : :
5703 : : Will be transformed into:
5704 : :
5705 : : S1 m_1 = i_4 > i_5;
5706 : : S2 m_2 = d_6 < d_7;
5707 : : S3'' m_2' = (_Bool[bitsize=32])m_2
5708 : : S3' m_3' = m_1 & m_2';
5709 : : S4'' m_3'' = (_Bool[bitsize=8])m_3'
5710 : : S4' c_1' = m_3'' ? c_2 : c_3; */
5711 : :
5712 : : static gimple *
5713 : 27632001 : vect_recog_mask_conversion_pattern (vec_info *vinfo,
5714 : : stmt_vec_info stmt_vinfo, tree *type_out)
5715 : : {
5716 : 27632001 : gimple *last_stmt = stmt_vinfo->stmt;
5717 : 27632001 : enum tree_code rhs_code;
5718 : 27632001 : tree lhs = NULL_TREE, rhs1, rhs2, tmp, rhs1_type, rhs2_type;
5719 : 27632001 : tree vectype1, vectype2;
5720 : 27632001 : stmt_vec_info pattern_stmt_info;
5721 : :
5722 : : /* Check for MASK_LOAD and MASK_STORE as well as COND_OP calls requiring mask
5723 : : conversion. */
5724 : 27632001 : if (is_gimple_call (last_stmt)
5725 : 27632001 : && gimple_call_internal_p (last_stmt))
5726 : : {
5727 : 87936 : gcall *pattern_stmt;
5728 : :
5729 : 87936 : internal_fn ifn = gimple_call_internal_fn (last_stmt);
5730 : 87936 : int mask_argno = internal_fn_mask_index (ifn);
5731 : 87936 : if (mask_argno < 0)
5732 : : return NULL;
5733 : :
5734 : 8759 : bool store_p = internal_store_fn_p (ifn);
5735 : 8759 : bool load_p = internal_store_fn_p (ifn);
5736 : 8759 : if (store_p)
5737 : : {
5738 : 1712 : int rhs_index = internal_fn_stored_value_index (ifn);
5739 : 1712 : tree rhs = gimple_call_arg (last_stmt, rhs_index);
5740 : 1712 : vectype1 = get_vectype_for_scalar_type (vinfo, TREE_TYPE (rhs));
5741 : : }
5742 : : else
5743 : : {
5744 : 7047 : lhs = gimple_call_lhs (last_stmt);
5745 : 7047 : if (!lhs)
5746 : : return NULL;
5747 : 7047 : vectype1 = get_vectype_for_scalar_type (vinfo, TREE_TYPE (lhs));
5748 : : }
5749 : :
5750 : 8759 : if (!vectype1)
5751 : : return NULL;
5752 : :
5753 : 8630 : tree mask_arg = gimple_call_arg (last_stmt, mask_argno);
5754 : 8630 : tree mask_arg_type = integer_type_for_mask (mask_arg, vinfo);
5755 : 8630 : if (mask_arg_type)
5756 : : {
5757 : 7592 : vectype2 = get_mask_type_for_scalar_type (vinfo, mask_arg_type);
5758 : :
5759 : 7592 : if (!vectype2
5760 : 7592 : || known_eq (TYPE_VECTOR_SUBPARTS (vectype1),
5761 : : TYPE_VECTOR_SUBPARTS (vectype2)))
5762 : 4454 : return NULL;
5763 : : }
5764 : 1038 : else if (store_p || load_p)
5765 : : return NULL;
5766 : :
5767 : 3977 : tmp = build_mask_conversion (vinfo, mask_arg, vectype1, stmt_vinfo);
5768 : :
5769 : 3977 : auto_vec<tree, 8> args;
5770 : 3977 : unsigned int nargs = gimple_call_num_args (last_stmt);
5771 : 3977 : args.safe_grow (nargs, true);
5772 : 19885 : for (unsigned int i = 0; i < nargs; ++i)
5773 : 15908 : args[i] = ((int) i == mask_argno
5774 : 15908 : ? tmp
5775 : 11931 : : gimple_call_arg (last_stmt, i));
5776 : 3977 : pattern_stmt = gimple_build_call_internal_vec (ifn, args);
5777 : :
5778 : 3977 : if (!store_p)
5779 : : {
5780 : 3596 : lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL);
5781 : 3596 : gimple_call_set_lhs (pattern_stmt, lhs);
5782 : : }
5783 : :
5784 : 3596 : if (load_p || store_p)
5785 : 381 : gimple_call_set_nothrow (pattern_stmt, true);
5786 : :
5787 : 3977 : pattern_stmt_info = vinfo->add_stmt (pattern_stmt);
5788 : 3977 : if (STMT_VINFO_DATA_REF (stmt_vinfo))
5789 : 1729 : vinfo->move_dr (pattern_stmt_info, stmt_vinfo);
5790 : :
5791 : 3977 : *type_out = vectype1;
5792 : 3977 : vect_pattern_detected ("vect_recog_mask_conversion_pattern", last_stmt);
5793 : :
5794 : 3977 : return pattern_stmt;
5795 : 3977 : }
5796 : :
5797 : 27544065 : if (!is_gimple_assign (last_stmt))
5798 : : return NULL;
5799 : :
5800 : 19167327 : gimple *pattern_stmt;
5801 : 19167327 : lhs = gimple_assign_lhs (last_stmt);
5802 : 19167327 : rhs1 = gimple_assign_rhs1 (last_stmt);
5803 : 19167327 : rhs_code = gimple_assign_rhs_code (last_stmt);
5804 : :
5805 : : /* Check for cond expression requiring mask conversion. */
5806 : 19167327 : if (rhs_code == COND_EXPR)
5807 : : {
5808 : 135604 : vectype1 = get_vectype_for_scalar_type (vinfo, TREE_TYPE (lhs));
5809 : :
5810 : 135604 : gcc_assert (! COMPARISON_CLASS_P (rhs1));
5811 : 135604 : if (TREE_CODE (rhs1) == SSA_NAME)
5812 : : {
5813 : 135604 : rhs1_type = integer_type_for_mask (rhs1, vinfo);
5814 : 135604 : if (!rhs1_type)
5815 : : return NULL;
5816 : : }
5817 : : else
5818 : : return NULL;
5819 : :
5820 : 125272 : vectype2 = get_mask_type_for_scalar_type (vinfo, rhs1_type);
5821 : :
5822 : 125272 : if (!vectype1 || !vectype2)
5823 : : return NULL;
5824 : :
5825 : : /* Continue if a conversion is needed. Also continue if we have
5826 : : a comparison whose vector type would normally be different from
5827 : : VECTYPE2 when considered in isolation. In that case we'll
5828 : : replace the comparison with an SSA name (so that we can record
5829 : : its vector type) and behave as though the comparison was an SSA
5830 : : name from the outset. */
5831 : 123677 : if (known_eq (TYPE_VECTOR_SUBPARTS (vectype1),
5832 : : TYPE_VECTOR_SUBPARTS (vectype2)))
5833 : : return NULL;
5834 : :
5835 : 27275 : if (maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1),
5836 : 54550 : TYPE_VECTOR_SUBPARTS (vectype2)))
5837 : 27275 : tmp = build_mask_conversion (vinfo, rhs1, vectype1, stmt_vinfo);
5838 : : else
5839 : : tmp = rhs1;
5840 : :
5841 : 27275 : lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL);
5842 : 27275 : pattern_stmt = gimple_build_assign (lhs, COND_EXPR, tmp,
5843 : : gimple_assign_rhs2 (last_stmt),
5844 : : gimple_assign_rhs3 (last_stmt));
5845 : :
5846 : 27275 : *type_out = vectype1;
5847 : 27275 : vect_pattern_detected ("vect_recog_mask_conversion_pattern", last_stmt);
5848 : :
5849 : 27275 : return pattern_stmt;
5850 : : }
5851 : :
5852 : : /* Now check for binary boolean operations requiring conversion for
5853 : : one of operands. */
5854 : 19031723 : if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (lhs)))
5855 : : return NULL;
5856 : :
5857 : 1532061 : if (rhs_code != BIT_IOR_EXPR
5858 : : && rhs_code != BIT_XOR_EXPR
5859 : 1532061 : && rhs_code != BIT_AND_EXPR
5860 : 1193681 : && TREE_CODE_CLASS (rhs_code) != tcc_comparison)
5861 : : return NULL;
5862 : :
5863 : 1420038 : rhs2 = gimple_assign_rhs2 (last_stmt);
5864 : :
5865 : 1420038 : rhs1_type = integer_type_for_mask (rhs1, vinfo);
5866 : 1420038 : rhs2_type = integer_type_for_mask (rhs2, vinfo);
5867 : :
5868 : 1420038 : if (!rhs1_type || !rhs2_type
5869 : 1420038 : || TYPE_PRECISION (rhs1_type) == TYPE_PRECISION (rhs2_type))
5870 : : return NULL;
5871 : :
5872 : 15529 : if (TYPE_PRECISION (rhs1_type) < TYPE_PRECISION (rhs2_type))
5873 : : {
5874 : 10028 : vectype1 = get_mask_type_for_scalar_type (vinfo, rhs1_type);
5875 : 10028 : if (!vectype1)
5876 : : return NULL;
5877 : 10028 : rhs2 = build_mask_conversion (vinfo, rhs2, vectype1, stmt_vinfo);
5878 : : }
5879 : : else
5880 : : {
5881 : 5501 : vectype1 = get_mask_type_for_scalar_type (vinfo, rhs2_type);
5882 : 5501 : if (!vectype1)
5883 : : return NULL;
5884 : 5501 : rhs1 = build_mask_conversion (vinfo, rhs1, vectype1, stmt_vinfo);
5885 : : }
5886 : :
5887 : 15529 : lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL);
5888 : 15529 : pattern_stmt = gimple_build_assign (lhs, rhs_code, rhs1, rhs2);
5889 : :
5890 : 15529 : *type_out = vectype1;
5891 : 15529 : vect_pattern_detected ("vect_recog_mask_conversion_pattern", last_stmt);
5892 : :
5893 : 15529 : return pattern_stmt;
5894 : : }
5895 : :
5896 : : /* STMT_INFO is a load or store. If the load or store is conditional, return
5897 : : the boolean condition under which it occurs, otherwise return null. */
5898 : :
5899 : : static tree
5900 : 31800 : vect_get_load_store_mask (stmt_vec_info stmt_info)
5901 : : {
5902 : 31800 : if (gassign *def_assign = dyn_cast <gassign *> (stmt_info->stmt))
5903 : : {
5904 : 30369 : gcc_assert (gimple_assign_single_p (def_assign));
5905 : : return NULL_TREE;
5906 : : }
5907 : :
5908 : 1431 : if (gcall *def_call = dyn_cast <gcall *> (stmt_info->stmt))
5909 : : {
5910 : 1431 : internal_fn ifn = gimple_call_internal_fn (def_call);
5911 : 1431 : int mask_index = internal_fn_mask_index (ifn);
5912 : 1431 : return gimple_call_arg (def_call, mask_index);
5913 : : }
5914 : :
5915 : 0 : gcc_unreachable ();
5916 : : }
5917 : :
5918 : : /* Return MASK if MASK is suitable for masking an operation on vectors
5919 : : of type VECTYPE, otherwise convert it into such a form and return
5920 : : the result. Associate any conversion statements with STMT_INFO's
5921 : : pattern. */
5922 : :
5923 : : static tree
5924 : 0 : vect_convert_mask_for_vectype (tree mask, tree vectype,
5925 : : stmt_vec_info stmt_info, vec_info *vinfo)
5926 : : {
5927 : 0 : tree mask_type = integer_type_for_mask (mask, vinfo);
5928 : 0 : if (mask_type)
5929 : : {
5930 : 0 : tree mask_vectype = get_mask_type_for_scalar_type (vinfo, mask_type);
5931 : 0 : if (mask_vectype
5932 : 0 : && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype),
5933 : 0 : TYPE_VECTOR_SUBPARTS (mask_vectype)))
5934 : 0 : mask = build_mask_conversion (vinfo, mask, vectype, stmt_info);
5935 : : }
5936 : 0 : return mask;
5937 : : }
5938 : :
5939 : : /* Return the equivalent of:
5940 : :
5941 : : fold_convert (TYPE, VALUE)
5942 : :
5943 : : with the expectation that the operation will be vectorized.
5944 : : If new statements are needed, add them as pattern statements
5945 : : to STMT_INFO. */
5946 : :
5947 : : static tree
5948 : 0 : vect_add_conversion_to_pattern (vec_info *vinfo,
5949 : : tree type, tree value, stmt_vec_info stmt_info)
5950 : : {
5951 : 0 : if (useless_type_conversion_p (type, TREE_TYPE (value)))
5952 : : return value;
5953 : :
5954 : 0 : tree new_value = vect_recog_temp_ssa_var (type, NULL);
5955 : 0 : gassign *conversion = gimple_build_assign (new_value, CONVERT_EXPR, value);
5956 : 0 : append_pattern_def_seq (vinfo, stmt_info, conversion,
5957 : : get_vectype_for_scalar_type (vinfo, type));
5958 : 0 : return new_value;
5959 : : }
5960 : :
5961 : : /* Try to convert STMT_INFO into a call to a gather load or scatter store
5962 : : internal function. Return the final statement on success and set
5963 : : *TYPE_OUT to the vector type being loaded or stored.
5964 : :
5965 : : This function only handles gathers and scatters that were recognized
5966 : : as such from the outset (indicated by STMT_VINFO_GATHER_SCATTER_P). */
5967 : :
5968 : : static gimple *
5969 : 27632001 : vect_recog_gather_scatter_pattern (vec_info *vinfo,
5970 : : stmt_vec_info stmt_info, tree *type_out)
5971 : : {
5972 : : /* Currently we only support this for loop vectorization. */
5973 : 30792234 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
5974 : 3160233 : if (!loop_vinfo)
5975 : : return NULL;
5976 : :
5977 : : /* Make sure that we're looking at a gather load or scatter store. */
5978 : 3160233 : data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
5979 : 3160233 : if (!dr || !STMT_VINFO_GATHER_SCATTER_P (stmt_info))
5980 : : return NULL;
5981 : :
5982 : : /* Get the boolean that controls whether the load or store happens.
5983 : : This is null if the operation is unconditional. */
5984 : 31800 : tree mask = vect_get_load_store_mask (stmt_info);
5985 : :
5986 : : /* Make sure that the target supports an appropriate internal
5987 : : function for the gather/scatter operation. */
5988 : 31800 : gather_scatter_info gs_info;
5989 : 31800 : if (!vect_check_gather_scatter (stmt_info, loop_vinfo, &gs_info)
5990 : 31800 : || gs_info.ifn == IFN_LAST)
5991 : : return NULL;
5992 : :
5993 : : /* Convert the mask to the right form. */
5994 : 0 : tree gs_vectype = get_vectype_for_scalar_type (loop_vinfo,
5995 : : gs_info.element_type);
5996 : 0 : if (mask)
5997 : 0 : mask = vect_convert_mask_for_vectype (mask, gs_vectype, stmt_info,
5998 : : loop_vinfo);
5999 : 0 : else if (gs_info.ifn == IFN_MASK_SCATTER_STORE
6000 : 0 : || gs_info.ifn == IFN_MASK_GATHER_LOAD
6001 : 0 : || gs_info.ifn == IFN_MASK_LEN_SCATTER_STORE
6002 : 0 : || gs_info.ifn == IFN_MASK_LEN_GATHER_LOAD)
6003 : 0 : mask = build_int_cst (TREE_TYPE (truth_type_for (gs_vectype)), -1);
6004 : :
6005 : : /* Get the invariant base and non-invariant offset, converting the
6006 : : latter to the same width as the vector elements. */
6007 : 0 : tree base = gs_info.base;
6008 : 0 : tree offset_type = TREE_TYPE (gs_info.offset_vectype);
6009 : 0 : tree offset = vect_add_conversion_to_pattern (vinfo, offset_type,
6010 : : gs_info.offset, stmt_info);
6011 : :
6012 : : /* Build the new pattern statement. */
6013 : 0 : tree scale = size_int (gs_info.scale);
6014 : 0 : gcall *pattern_stmt;
6015 : :
6016 : 0 : if (DR_IS_READ (dr))
6017 : : {
6018 : 0 : tree zero = build_zero_cst (gs_info.element_type);
6019 : 0 : if (mask != NULL)
6020 : : {
6021 : 0 : int elsval = MASK_LOAD_ELSE_ZERO;
6022 : :
6023 : 0 : tree vec_els
6024 : 0 : = vect_get_mask_load_else (elsval, TREE_TYPE (gs_vectype));
6025 : 0 : pattern_stmt = gimple_build_call_internal (gs_info.ifn, 6, base,
6026 : : offset, scale, zero, mask,
6027 : : vec_els);
6028 : : }
6029 : : else
6030 : 0 : pattern_stmt = gimple_build_call_internal (gs_info.ifn, 4, base,
6031 : : offset, scale, zero);
6032 : 0 : tree load_lhs = vect_recog_temp_ssa_var (gs_info.element_type, NULL);
6033 : 0 : gimple_call_set_lhs (pattern_stmt, load_lhs);
6034 : : }
6035 : : else
6036 : : {
6037 : 0 : tree rhs = vect_get_store_rhs (stmt_info);
6038 : 0 : if (mask != NULL)
6039 : 0 : pattern_stmt = gimple_build_call_internal (gs_info.ifn, 5,
6040 : : base, offset, scale, rhs,
6041 : : mask);
6042 : : else
6043 : 0 : pattern_stmt = gimple_build_call_internal (gs_info.ifn, 4,
6044 : : base, offset, scale, rhs);
6045 : : }
6046 : 0 : gimple_call_set_nothrow (pattern_stmt, true);
6047 : :
6048 : : /* Copy across relevant vectorization info and associate DR with the
6049 : : new pattern statement instead of the original statement. */
6050 : 0 : stmt_vec_info pattern_stmt_info = loop_vinfo->add_stmt (pattern_stmt);
6051 : 0 : loop_vinfo->move_dr (pattern_stmt_info, stmt_info);
6052 : :
6053 : 0 : tree vectype = STMT_VINFO_VECTYPE (stmt_info);
6054 : 0 : *type_out = vectype;
6055 : 0 : vect_pattern_detected ("gather/scatter pattern", stmt_info->stmt);
6056 : :
6057 : 0 : return pattern_stmt;
6058 : : }
6059 : :
6060 : : /* Helper method of vect_recog_cond_store_pattern, checks to see if COND_ARG
6061 : : is points to a load statement that reads the same data as that of
6062 : : STORE_VINFO. */
6063 : :
6064 : : static bool
6065 : 27473 : vect_cond_store_pattern_same_ref (vec_info *vinfo,
6066 : : stmt_vec_info store_vinfo, tree cond_arg)
6067 : : {
6068 : 27473 : stmt_vec_info load_stmt_vinfo = vinfo->lookup_def (cond_arg);
6069 : 27473 : if (!load_stmt_vinfo
6070 : 15536 : || !STMT_VINFO_DATA_REF (load_stmt_vinfo)
6071 : 10426 : || DR_IS_WRITE (STMT_VINFO_DATA_REF (load_stmt_vinfo))
6072 : 37899 : || !same_data_refs (STMT_VINFO_DATA_REF (store_vinfo),
6073 : : STMT_VINFO_DATA_REF (load_stmt_vinfo)))
6074 : 19662 : return false;
6075 : :
6076 : : return true;
6077 : : }
6078 : :
6079 : : /* Function vect_recog_cond_store_pattern
6080 : :
6081 : : Try to find the following pattern:
6082 : :
6083 : : x = *_3;
6084 : : c = a CMP b;
6085 : : y = c ? t_20 : x;
6086 : : *_3 = y;
6087 : :
6088 : : where the store of _3 happens on a conditional select on a value loaded
6089 : : from the same location. In such case we can elide the initial load if
6090 : : MASK_STORE is supported and instead only conditionally write out the result.
6091 : :
6092 : : The pattern produces for the above:
6093 : :
6094 : : c = a CMP b;
6095 : : .MASK_STORE (_3, c, t_20)
6096 : :
6097 : : Input:
6098 : :
6099 : : * STMT_VINFO: The stmt from which the pattern search begins. In the
6100 : : example, when this function is called with _3 then the search begins.
6101 : :
6102 : : Output:
6103 : :
6104 : : * TYPE_OUT: The type of the output of this pattern.
6105 : :
6106 : : * Return value: A new stmt that will be used to replace the sequence. */
6107 : :
6108 : : static gimple *
6109 : 27632001 : vect_recog_cond_store_pattern (vec_info *vinfo,
6110 : : stmt_vec_info stmt_vinfo, tree *type_out)
6111 : : {
6112 : 27632001 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
6113 : 3160233 : if (!loop_vinfo)
6114 : : return NULL;
6115 : :
6116 : 3160233 : gimple *store_stmt = STMT_VINFO_STMT (stmt_vinfo);
6117 : :
6118 : : /* Needs to be a gimple store where we have DR info for. */
6119 : 3160233 : if (!STMT_VINFO_DATA_REF (stmt_vinfo)
6120 : 747674 : || DR_IS_READ (STMT_VINFO_DATA_REF (stmt_vinfo))
6121 : 3423960 : || !gimple_store_p (store_stmt))
6122 : 2898218 : return NULL;
6123 : :
6124 : 262015 : tree st_rhs = gimple_assign_rhs1 (store_stmt);
6125 : :
6126 : 262015 : if (TREE_CODE (st_rhs) != SSA_NAME)
6127 : : return NULL;
6128 : :
6129 : 213857 : auto cond_vinfo = vinfo->lookup_def (st_rhs);
6130 : :
6131 : : /* If the condition isn't part of the loop then bool recog wouldn't have seen
6132 : : it and so this transformation may not be valid. */
6133 : 213857 : if (!cond_vinfo)
6134 : : return NULL;
6135 : :
6136 : 198772 : cond_vinfo = vect_stmt_to_vectorize (cond_vinfo);
6137 : 27822084 : gassign *cond_stmt = dyn_cast<gassign *> (STMT_VINFO_STMT (cond_vinfo));
6138 : 246028 : if (!cond_stmt || gimple_assign_rhs_code (cond_stmt) != COND_EXPR)
6139 : : return NULL;
6140 : :
6141 : : /* Check if the else value matches the original loaded one. */
6142 : 14388 : bool invert = false;
6143 : 14388 : tree cmp_ls = gimple_arg (cond_stmt, 0);
6144 : 14388 : if (TREE_CODE (cmp_ls) != SSA_NAME)
6145 : : return NULL;
6146 : :
6147 : 14388 : tree cond_arg1 = gimple_arg (cond_stmt, 1);
6148 : 14388 : tree cond_arg2 = gimple_arg (cond_stmt, 2);
6149 : :
6150 : 14388 : if (!vect_cond_store_pattern_same_ref (vinfo, stmt_vinfo, cond_arg2)
6151 : 14388 : && !(invert = vect_cond_store_pattern_same_ref (vinfo, stmt_vinfo,
6152 : : cond_arg1)))
6153 : : return NULL;
6154 : :
6155 : 7811 : vect_pattern_detected ("vect_recog_cond_store_pattern", store_stmt);
6156 : :
6157 : 7811 : tree scalar_type = TREE_TYPE (st_rhs);
6158 : 7811 : if (VECTOR_TYPE_P (scalar_type))
6159 : : return NULL;
6160 : :
6161 : 7811 : tree vectype = get_vectype_for_scalar_type (vinfo, scalar_type);
6162 : 7811 : if (vectype == NULL_TREE)
6163 : : return NULL;
6164 : :
6165 : 7811 : machine_mode mask_mode;
6166 : 7811 : machine_mode vecmode = TYPE_MODE (vectype);
6167 : 1644 : if (!VECTOR_MODE_P (vecmode)
6168 : 7811 : || targetm.vectorize.conditional_operation_is_expensive (IFN_MASK_STORE)
6169 : 0 : || !targetm.vectorize.get_mask_mode (vecmode).exists (&mask_mode)
6170 : 7811 : || !can_vec_mask_load_store_p (vecmode, mask_mode, false))
6171 : 7811 : return NULL;
6172 : :
6173 : 0 : tree base = DR_REF (STMT_VINFO_DATA_REF (stmt_vinfo));
6174 : 0 : if (may_be_nonaddressable_p (base))
6175 : : return NULL;
6176 : :
6177 : : /* We need to use the false parameter of the conditional select. */
6178 : 0 : tree cond_store_arg = invert ? cond_arg2 : cond_arg1;
6179 : 0 : tree cond_load_arg = invert ? cond_arg1 : cond_arg2;
6180 : 0 : gimple *load_stmt = SSA_NAME_DEF_STMT (cond_load_arg);
6181 : :
6182 : : /* This is a rough estimation to check that there aren't any aliasing stores
6183 : : in between the load and store. It's a bit strict, but for now it's good
6184 : : enough. */
6185 : 0 : if (gimple_vuse (load_stmt) != gimple_vuse (store_stmt))
6186 : : return NULL;
6187 : :
6188 : : /* If we have to invert the condition, i.e. use the true argument rather than
6189 : : the false argument, we have to negate the mask. */
6190 : 0 : if (invert)
6191 : : {
6192 : 0 : tree var = vect_recog_temp_ssa_var (boolean_type_node, NULL);
6193 : :
6194 : : /* Invert the mask using ^ 1. */
6195 : 0 : tree itype = TREE_TYPE (cmp_ls);
6196 : 0 : gassign *conv = gimple_build_assign (var, BIT_XOR_EXPR, cmp_ls,
6197 : 0 : build_int_cst (itype, 1));
6198 : :
6199 : 0 : tree mask_vec_type = get_mask_type_for_scalar_type (vinfo, itype);
6200 : 0 : append_pattern_def_seq (vinfo, stmt_vinfo, conv, mask_vec_type, itype);
6201 : 0 : cmp_ls= var;
6202 : : }
6203 : :
6204 : 0 : if (TREE_CODE (base) != MEM_REF)
6205 : 0 : base = build_fold_addr_expr (base);
6206 : :
6207 : 0 : tree ptr = build_int_cst (reference_alias_ptr_type (base),
6208 : 0 : get_object_alignment (base));
6209 : :
6210 : : /* Convert the mask to the right form. */
6211 : 0 : tree mask = vect_convert_mask_for_vectype (cmp_ls, vectype, stmt_vinfo,
6212 : : vinfo);
6213 : :
6214 : 0 : gcall *call
6215 : 0 : = gimple_build_call_internal (IFN_MASK_STORE, 4, base, ptr, mask,
6216 : : cond_store_arg);
6217 : 0 : gimple_set_location (call, gimple_location (store_stmt));
6218 : :
6219 : : /* Copy across relevant vectorization info and associate DR with the
6220 : : new pattern statement instead of the original statement. */
6221 : 0 : stmt_vec_info pattern_stmt_info = loop_vinfo->add_stmt (call);
6222 : 0 : loop_vinfo->move_dr (pattern_stmt_info, stmt_vinfo);
6223 : :
6224 : 0 : *type_out = vectype;
6225 : 0 : return call;
6226 : : }
6227 : :
6228 : : /* Return true if TYPE is a non-boolean integer type. These are the types
6229 : : that we want to consider for narrowing. */
6230 : :
6231 : : static bool
6232 : 55813308 : vect_narrowable_type_p (tree type)
6233 : : {
6234 : 55813308 : return INTEGRAL_TYPE_P (type) && !VECT_SCALAR_BOOLEAN_TYPE_P (type);
6235 : : }
6236 : :
6237 : : /* Return true if the operation given by CODE can be truncated to N bits
6238 : : when only N bits of the output are needed. This is only true if bit N+1
6239 : : of the inputs has no effect on the low N bits of the result. */
6240 : :
6241 : : static bool
6242 : 13510952 : vect_truncatable_operation_p (tree_code code)
6243 : : {
6244 : 13510952 : switch (code)
6245 : : {
6246 : : case NEGATE_EXPR:
6247 : : case PLUS_EXPR:
6248 : : case MINUS_EXPR:
6249 : : case MULT_EXPR:
6250 : : case BIT_NOT_EXPR:
6251 : : case BIT_AND_EXPR:
6252 : : case BIT_IOR_EXPR:
6253 : : case BIT_XOR_EXPR:
6254 : : case COND_EXPR:
6255 : : return true;
6256 : :
6257 : 5252323 : default:
6258 : 5252323 : return false;
6259 : : }
6260 : : }
6261 : :
6262 : : /* Record that STMT_INFO could be changed from operating on TYPE to
6263 : : operating on a type with the precision and sign given by PRECISION
6264 : : and SIGN respectively. PRECISION is an arbitrary bit precision;
6265 : : it might not be a whole number of bytes. */
6266 : :
6267 : : static void
6268 : 1892921 : vect_set_operation_type (stmt_vec_info stmt_info, tree type,
6269 : : unsigned int precision, signop sign)
6270 : : {
6271 : : /* Round the precision up to a whole number of bytes. */
6272 : 1892921 : precision = vect_element_precision (precision);
6273 : 1892921 : if (precision < TYPE_PRECISION (type)
6274 : 1892921 : && (!stmt_info->operation_precision
6275 : 30371 : || stmt_info->operation_precision > precision))
6276 : : {
6277 : 1188378 : stmt_info->operation_precision = precision;
6278 : 1188378 : stmt_info->operation_sign = sign;
6279 : : }
6280 : 1892921 : }
6281 : :
6282 : : /* Record that STMT_INFO only requires MIN_INPUT_PRECISION from its
6283 : : non-boolean inputs, all of which have type TYPE. MIN_INPUT_PRECISION
6284 : : is an arbitrary bit precision; it might not be a whole number of bytes. */
6285 : :
6286 : : static void
6287 : 9753320 : vect_set_min_input_precision (stmt_vec_info stmt_info, tree type,
6288 : : unsigned int min_input_precision)
6289 : : {
6290 : : /* This operation in isolation only requires the inputs to have
6291 : : MIN_INPUT_PRECISION of precision, However, that doesn't mean
6292 : : that MIN_INPUT_PRECISION is a natural precision for the chain
6293 : : as a whole. E.g. consider something like:
6294 : :
6295 : : unsigned short *x, *y;
6296 : : *y = ((*x & 0xf0) >> 4) | (*y << 4);
6297 : :
6298 : : The right shift can be done on unsigned chars, and only requires the
6299 : : result of "*x & 0xf0" to be done on unsigned chars. But taking that
6300 : : approach would mean turning a natural chain of single-vector unsigned
6301 : : short operations into one that truncates "*x" and then extends
6302 : : "(*x & 0xf0) >> 4", with two vectors for each unsigned short
6303 : : operation and one vector for each unsigned char operation.
6304 : : This would be a significant pessimization.
6305 : :
6306 : : Instead only propagate the maximum of this precision and the precision
6307 : : required by the users of the result. This means that we don't pessimize
6308 : : the case above but continue to optimize things like:
6309 : :
6310 : : unsigned char *y;
6311 : : unsigned short *x;
6312 : : *y = ((*x & 0xf0) >> 4) | (*y << 4);
6313 : :
6314 : : Here we would truncate two vectors of *x to a single vector of
6315 : : unsigned chars and use single-vector unsigned char operations for
6316 : : everything else, rather than doing two unsigned short copies of
6317 : : "(*x & 0xf0) >> 4" and then truncating the result. */
6318 : 9753320 : min_input_precision = MAX (min_input_precision,
6319 : : stmt_info->min_output_precision);
6320 : :
6321 : 9753320 : if (min_input_precision < TYPE_PRECISION (type)
6322 : 9753320 : && (!stmt_info->min_input_precision
6323 : 50302 : || stmt_info->min_input_precision > min_input_precision))
6324 : 508553 : stmt_info->min_input_precision = min_input_precision;
6325 : 9753320 : }
6326 : :
6327 : : /* Subroutine of vect_determine_min_output_precision. Return true if
6328 : : we can calculate a reduced number of output bits for STMT_INFO,
6329 : : whose result is LHS. */
6330 : :
6331 : : static bool
6332 : 13084078 : vect_determine_min_output_precision_1 (vec_info *vinfo,
6333 : : stmt_vec_info stmt_info, tree lhs)
6334 : : {
6335 : : /* Take the maximum precision required by users of the result. */
6336 : 13084078 : unsigned int precision = 0;
6337 : 13084078 : imm_use_iterator iter;
6338 : 13084078 : use_operand_p use;
6339 : 13825161 : FOR_EACH_IMM_USE_FAST (use, iter, lhs)
6340 : : {
6341 : 13576259 : gimple *use_stmt = USE_STMT (use);
6342 : 13576259 : if (is_gimple_debug (use_stmt))
6343 : 483451 : continue;
6344 : 13092808 : stmt_vec_info use_stmt_info = vinfo->lookup_stmt (use_stmt);
6345 : 13092808 : if (!use_stmt_info || !use_stmt_info->min_input_precision)
6346 : : return false;
6347 : : /* The input precision recorded for COND_EXPRs applies only to the
6348 : : "then" and "else" values. */
6349 : 258091 : gassign *assign = dyn_cast <gassign *> (stmt_info->stmt);
6350 : 221334 : if (assign
6351 : 221334 : && gimple_assign_rhs_code (assign) == COND_EXPR
6352 : 459 : && use->use != gimple_assign_rhs2_ptr (assign)
6353 : 459 : && use->use != gimple_assign_rhs3_ptr (assign))
6354 : : return false;
6355 : 257632 : precision = MAX (precision, use_stmt_info->min_input_precision);
6356 : : }
6357 : :
6358 : 248902 : if (dump_enabled_p ())
6359 : 6619 : dump_printf_loc (MSG_NOTE, vect_location,
6360 : : "only the low %d bits of %T are significant\n",
6361 : : precision, lhs);
6362 : 248902 : stmt_info->min_output_precision = precision;
6363 : 248902 : return true;
6364 : : }
6365 : :
6366 : : /* Calculate min_output_precision for STMT_INFO. */
6367 : :
6368 : : static void
6369 : 33978850 : vect_determine_min_output_precision (vec_info *vinfo, stmt_vec_info stmt_info)
6370 : : {
6371 : : /* We're only interested in statements with a narrowable result. */
6372 : 33978850 : tree lhs = gimple_get_lhs (stmt_info->stmt);
6373 : 33978850 : if (!lhs
6374 : 26688313 : || TREE_CODE (lhs) != SSA_NAME
6375 : 56398890 : || !vect_narrowable_type_p (TREE_TYPE (lhs)))
6376 : : return;
6377 : :
6378 : 13084078 : if (!vect_determine_min_output_precision_1 (vinfo, stmt_info, lhs))
6379 : 12835176 : stmt_info->min_output_precision = TYPE_PRECISION (TREE_TYPE (lhs));
6380 : : }
6381 : :
6382 : : /* Use range information to decide whether STMT (described by STMT_INFO)
6383 : : could be done in a narrower type. This is effectively a forward
6384 : : propagation, since it uses context-independent information that applies
6385 : : to all users of an SSA name. */
6386 : :
6387 : : static void
6388 : 18691700 : vect_determine_precisions_from_range (stmt_vec_info stmt_info, gassign *stmt)
6389 : : {
6390 : 18691700 : tree lhs = gimple_assign_lhs (stmt);
6391 : 18691700 : if (!lhs || TREE_CODE (lhs) != SSA_NAME)
6392 : 16892334 : return;
6393 : :
6394 : 14701568 : tree type = TREE_TYPE (lhs);
6395 : 14701568 : if (!vect_narrowable_type_p (type))
6396 : : return;
6397 : :
6398 : : /* First see whether we have any useful range information for the result. */
6399 : 9882957 : unsigned int precision = TYPE_PRECISION (type);
6400 : 9882957 : signop sign = TYPE_SIGN (type);
6401 : 9882957 : wide_int min_value, max_value;
6402 : 9882957 : if (!vect_get_range_info (lhs, &min_value, &max_value))
6403 : : return;
6404 : :
6405 : 4476758 : tree_code code = gimple_assign_rhs_code (stmt);
6406 : 4476758 : unsigned int nops = gimple_num_ops (stmt);
6407 : :
6408 : 4476758 : if (!vect_truncatable_operation_p (code))
6409 : : {
6410 : : /* Handle operations that can be computed in type T if all inputs
6411 : : and outputs can be represented in type T. Also handle left and
6412 : : right shifts, where (in addition) the maximum shift amount must
6413 : : be less than the number of bits in T. */
6414 : 1679696 : bool is_shift;
6415 : 1679696 : switch (code)
6416 : : {
6417 : : case LSHIFT_EXPR:
6418 : : case RSHIFT_EXPR:
6419 : : is_shift = true;
6420 : : break;
6421 : :
6422 : 201983 : case ABS_EXPR:
6423 : 201983 : case MIN_EXPR:
6424 : 201983 : case MAX_EXPR:
6425 : 201983 : case TRUNC_DIV_EXPR:
6426 : 201983 : case CEIL_DIV_EXPR:
6427 : 201983 : case FLOOR_DIV_EXPR:
6428 : 201983 : case ROUND_DIV_EXPR:
6429 : 201983 : case EXACT_DIV_EXPR:
6430 : : /* Modulus is excluded because it is typically calculated by doing
6431 : : a division, for which minimum signed / -1 isn't representable in
6432 : : the original signed type. We could take the division range into
6433 : : account instead, if handling modulus ever becomes important. */
6434 : 201983 : is_shift = false;
6435 : 201983 : break;
6436 : :
6437 : : default:
6438 : : return;
6439 : : }
6440 : 1101964 : for (unsigned int i = 1; i < nops; ++i)
6441 : : {
6442 : 847861 : tree op = gimple_op (stmt, i);
6443 : 847861 : wide_int op_min_value, op_max_value;
6444 : 847861 : if (TREE_CODE (op) == INTEGER_CST)
6445 : : {
6446 : 259268 : unsigned int op_precision = TYPE_PRECISION (TREE_TYPE (op));
6447 : 259268 : op_min_value = op_max_value = wi::to_wide (op, op_precision);
6448 : : }
6449 : 588593 : else if (TREE_CODE (op) == SSA_NAME)
6450 : : {
6451 : 588593 : if (!vect_get_range_info (op, &op_min_value, &op_max_value))
6452 : : return;
6453 : : }
6454 : : else
6455 : : return;
6456 : :
6457 : 567155 : if (is_shift && i == 2)
6458 : : {
6459 : : /* There needs to be one more bit than the maximum shift amount.
6460 : :
6461 : : If the maximum shift amount is already 1 less than PRECISION
6462 : : then we can't narrow the shift further. Dealing with that
6463 : : case first ensures that we can safely use an unsigned range
6464 : : below.
6465 : :
6466 : : op_min_value isn't relevant, since shifts by negative amounts
6467 : : are UB. */
6468 : 168175 : if (wi::geu_p (op_max_value, precision - 1))
6469 : : return;
6470 : 152083 : unsigned int min_bits = op_max_value.to_uhwi () + 1;
6471 : :
6472 : : /* As explained below, we can convert a signed shift into an
6473 : : unsigned shift if the sign bit is always clear. At this
6474 : : point we've already processed the ranges of the output and
6475 : : the first input. */
6476 : 152083 : auto op_sign = sign;
6477 : 152083 : if (sign == SIGNED && !wi::neg_p (min_value))
6478 : : op_sign = UNSIGNED;
6479 : 304166 : op_min_value = wide_int::from (wi::min_value (min_bits, op_sign),
6480 : 152083 : precision, op_sign);
6481 : 304166 : op_max_value = wide_int::from (wi::max_value (min_bits, op_sign),
6482 : 152083 : precision, op_sign);
6483 : : }
6484 : 551063 : min_value = wi::min (min_value, op_min_value, sign);
6485 : 551063 : max_value = wi::max (max_value, op_max_value, sign);
6486 : 847861 : }
6487 : : }
6488 : :
6489 : : /* Try to switch signed types for unsigned types if we can.
6490 : : This is better for two reasons. First, unsigned ops tend
6491 : : to be cheaper than signed ops. Second, it means that we can
6492 : : handle things like:
6493 : :
6494 : : signed char c;
6495 : : int res = (int) c & 0xff00; // range [0x0000, 0xff00]
6496 : :
6497 : : as:
6498 : :
6499 : : signed char c;
6500 : : unsigned short res_1 = (unsigned short) c & 0xff00;
6501 : : int res = (int) res_1;
6502 : :
6503 : : where the intermediate result res_1 has unsigned rather than
6504 : : signed type. */
6505 : 3051165 : if (sign == SIGNED && !wi::neg_p (min_value))
6506 : : sign = UNSIGNED;
6507 : :
6508 : : /* See what precision is required for MIN_VALUE and MAX_VALUE. */
6509 : 3051165 : unsigned int precision1 = wi::min_precision (min_value, sign);
6510 : 3051165 : unsigned int precision2 = wi::min_precision (max_value, sign);
6511 : 3051165 : unsigned int value_precision = MAX (precision1, precision2);
6512 : 3051165 : if (value_precision >= precision)
6513 : : return;
6514 : :
6515 : 1799366 : if (dump_enabled_p ())
6516 : 77953 : dump_printf_loc (MSG_NOTE, vect_location, "can narrow to %s:%d"
6517 : : " without loss of precision: %G",
6518 : : sign == SIGNED ? "signed" : "unsigned",
6519 : : value_precision, (gimple *) stmt);
6520 : :
6521 : 1799366 : vect_set_operation_type (stmt_info, type, value_precision, sign);
6522 : 1799366 : vect_set_min_input_precision (stmt_info, type, value_precision);
6523 : 9882957 : }
6524 : :
6525 : : /* Use information about the users of STMT's result to decide whether
6526 : : STMT (described by STMT_INFO) could be done in a narrower type.
6527 : : This is effectively a backward propagation. */
6528 : :
6529 : : static void
6530 : 18691700 : vect_determine_precisions_from_users (stmt_vec_info stmt_info, gassign *stmt)
6531 : : {
6532 : 18691700 : tree_code code = gimple_assign_rhs_code (stmt);
6533 : 18691700 : unsigned int opno = (code == COND_EXPR ? 2 : 1);
6534 : 18691700 : tree type = TREE_TYPE (gimple_op (stmt, opno));
6535 : 18691700 : if (!vect_narrowable_type_p (type))
6536 : 10737746 : return;
6537 : :
6538 : 11620935 : unsigned int precision = TYPE_PRECISION (type);
6539 : 11620935 : unsigned int operation_precision, min_input_precision;
6540 : 11620935 : switch (code)
6541 : : {
6542 : 2134827 : CASE_CONVERT:
6543 : : /* Only the bits that contribute to the output matter. Don't change
6544 : : the precision of the operation itself. */
6545 : 2134827 : operation_precision = precision;
6546 : 2134827 : min_input_precision = stmt_info->min_output_precision;
6547 : 2134827 : break;
6548 : :
6549 : 451914 : case LSHIFT_EXPR:
6550 : 451914 : case RSHIFT_EXPR:
6551 : 451914 : {
6552 : 451914 : tree shift = gimple_assign_rhs2 (stmt);
6553 : 451914 : if (TREE_CODE (shift) != INTEGER_CST
6554 : 809510 : || !wi::ltu_p (wi::to_widest (shift), precision))
6555 : 94354 : return;
6556 : 357560 : unsigned int const_shift = TREE_INT_CST_LOW (shift);
6557 : 357560 : if (code == LSHIFT_EXPR)
6558 : : {
6559 : : /* Avoid creating an undefined shift.
6560 : :
6561 : : ??? We could instead use min_output_precision as-is and
6562 : : optimize out-of-range shifts to zero. However, only
6563 : : degenerate testcases shift away all their useful input data,
6564 : : and it isn't natural to drop input operations in the middle
6565 : : of vectorization. This sort of thing should really be
6566 : : handled before vectorization. */
6567 : 92537 : operation_precision = MAX (stmt_info->min_output_precision,
6568 : : const_shift + 1);
6569 : : /* We need CONST_SHIFT fewer bits of the input. */
6570 : 92537 : min_input_precision = (MAX (operation_precision, const_shift)
6571 : 92537 : - const_shift);
6572 : : }
6573 : : else
6574 : : {
6575 : : /* We need CONST_SHIFT extra bits to do the operation. */
6576 : 265023 : operation_precision = (stmt_info->min_output_precision
6577 : : + const_shift);
6578 : 265023 : min_input_precision = operation_precision;
6579 : : }
6580 : : break;
6581 : : }
6582 : :
6583 : 9034194 : default:
6584 : 9034194 : if (vect_truncatable_operation_p (code))
6585 : : {
6586 : : /* Input bit N has no effect on output bits N-1 and lower. */
6587 : 5461567 : operation_precision = stmt_info->min_output_precision;
6588 : 5461567 : min_input_precision = operation_precision;
6589 : 5461567 : break;
6590 : : }
6591 : : return;
6592 : : }
6593 : :
6594 : 7953954 : if (operation_precision < precision)
6595 : : {
6596 : 93555 : if (dump_enabled_p ())
6597 : 3327 : dump_printf_loc (MSG_NOTE, vect_location, "can narrow to %s:%d"
6598 : : " without affecting users: %G",
6599 : 3327 : TYPE_UNSIGNED (type) ? "unsigned" : "signed",
6600 : : operation_precision, (gimple *) stmt);
6601 : 187110 : vect_set_operation_type (stmt_info, type, operation_precision,
6602 : 93555 : TYPE_SIGN (type));
6603 : : }
6604 : 7953954 : vect_set_min_input_precision (stmt_info, type, min_input_precision);
6605 : : }
6606 : :
6607 : : /* Return true if the statement described by STMT_INFO sets a boolean
6608 : : SSA_NAME and if we know how to vectorize this kind of statement using
6609 : : vector mask types. */
6610 : :
6611 : : static bool
6612 : 33978850 : possible_vector_mask_operation_p (stmt_vec_info stmt_info)
6613 : : {
6614 : 33978850 : tree lhs = gimple_get_lhs (stmt_info->stmt);
6615 : 33978850 : tree_code code = ERROR_MARK;
6616 : 33978850 : gassign *assign = NULL;
6617 : 33978850 : gcond *cond = NULL;
6618 : :
6619 : 33978850 : if ((assign = dyn_cast <gassign *> (stmt_info->stmt)))
6620 : 25127134 : code = gimple_assign_rhs_code (assign);
6621 : 15287150 : else if ((cond = dyn_cast <gcond *> (stmt_info->stmt)))
6622 : : {
6623 : 4513539 : lhs = gimple_cond_lhs (cond);
6624 : 4513539 : code = gimple_cond_code (cond);
6625 : : }
6626 : :
6627 : 33978850 : if (!lhs
6628 : 31201852 : || TREE_CODE (lhs) != SSA_NAME
6629 : 60842768 : || !VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (lhs)))
6630 : : return false;
6631 : :
6632 : 1942559 : if (code != ERROR_MARK)
6633 : : {
6634 : 1681084 : switch (code)
6635 : : {
6636 : : CASE_CONVERT:
6637 : : case SSA_NAME:
6638 : : case BIT_NOT_EXPR:
6639 : : case BIT_IOR_EXPR:
6640 : : case BIT_XOR_EXPR:
6641 : : case BIT_AND_EXPR:
6642 : : return true;
6643 : :
6644 : 1306120 : default:
6645 : 1306120 : return TREE_CODE_CLASS (code) == tcc_comparison;
6646 : : }
6647 : : }
6648 : 261475 : else if (is_a <gphi *> (stmt_info->stmt))
6649 : 165193 : return true;
6650 : : return false;
6651 : : }
6652 : :
6653 : : /* If STMT_INFO sets a boolean SSA_NAME, see whether we should use
6654 : : a vector mask type instead of a normal vector type. Record the
6655 : : result in STMT_INFO->mask_precision. */
6656 : :
6657 : : static void
6658 : 33978850 : vect_determine_mask_precision (vec_info *vinfo, stmt_vec_info stmt_info)
6659 : : {
6660 : 33978850 : if (!possible_vector_mask_operation_p (stmt_info))
6661 : : return;
6662 : :
6663 : : /* If at least one boolean input uses a vector mask type,
6664 : : pick the mask type with the narrowest elements.
6665 : :
6666 : : ??? This is the traditional behavior. It should always produce
6667 : : the smallest number of operations, but isn't necessarily the
6668 : : optimal choice. For example, if we have:
6669 : :
6670 : : a = b & c
6671 : :
6672 : : where:
6673 : :
6674 : : - the user of a wants it to have a mask type for 16-bit elements (M16)
6675 : : - b also uses M16
6676 : : - c uses a mask type for 8-bit elements (M8)
6677 : :
6678 : : then picking M8 gives:
6679 : :
6680 : : - 1 M16->M8 pack for b
6681 : : - 1 M8 AND for a
6682 : : - 2 M8->M16 unpacks for the user of a
6683 : :
6684 : : whereas picking M16 would have given:
6685 : :
6686 : : - 2 M8->M16 unpacks for c
6687 : : - 2 M16 ANDs for a
6688 : :
6689 : : The number of operations are equal, but M16 would have given
6690 : : a shorter dependency chain and allowed more ILP. */
6691 : 1808401 : unsigned int precision = ~0U;
6692 : 1808401 : gimple *stmt = STMT_VINFO_STMT (stmt_info);
6693 : :
6694 : : /* If the statement compares two values that shouldn't use vector masks,
6695 : : try comparing the values as normal scalars instead. */
6696 : 1808401 : tree_code code = ERROR_MARK;
6697 : 1808401 : tree op0_type;
6698 : 1808401 : unsigned int nops = -1;
6699 : 1808401 : unsigned int ops_start = 0;
6700 : :
6701 : 1808401 : if (gassign *assign = dyn_cast <gassign *> (stmt))
6702 : : {
6703 : 1139675 : code = gimple_assign_rhs_code (assign);
6704 : 1139675 : op0_type = TREE_TYPE (gimple_assign_rhs1 (assign));
6705 : 1139675 : nops = gimple_num_ops (assign);
6706 : 1139675 : ops_start = 1;
6707 : : }
6708 : 668726 : else if (gcond *cond = dyn_cast <gcond *> (stmt))
6709 : : {
6710 : 503533 : code = gimple_cond_code (cond);
6711 : 503533 : op0_type = TREE_TYPE (gimple_cond_lhs (cond));
6712 : 503533 : nops = 2;
6713 : 503533 : ops_start = 0;
6714 : : }
6715 : :
6716 : 1643208 : if (code != ERROR_MARK)
6717 : : {
6718 : 4893040 : for (unsigned int i = ops_start; i < nops; ++i)
6719 : : {
6720 : 3249832 : tree rhs = gimple_op (stmt, i);
6721 : 3249832 : if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (rhs)))
6722 : 1539264 : continue;
6723 : :
6724 : 1710568 : stmt_vec_info def_stmt_info = vinfo->lookup_def (rhs);
6725 : 1710568 : if (!def_stmt_info)
6726 : : /* Don't let external or constant operands influence the choice.
6727 : : We can convert them to whichever vector type we pick. */
6728 : 514996 : continue;
6729 : :
6730 : 1195572 : if (def_stmt_info->mask_precision)
6731 : : {
6732 : 1011349 : if (precision > def_stmt_info->mask_precision)
6733 : 3249832 : precision = def_stmt_info->mask_precision;
6734 : : }
6735 : : }
6736 : :
6737 : 1643208 : if (precision == ~0U
6738 : 1317046 : && TREE_CODE_CLASS (code) == tcc_comparison)
6739 : : {
6740 : 1127443 : scalar_mode mode;
6741 : 1127443 : tree vectype, mask_type;
6742 : 1127443 : if (is_a <scalar_mode> (TYPE_MODE (op0_type), &mode)
6743 : 1127443 : && (vectype = get_vectype_for_scalar_type (vinfo, op0_type))
6744 : 971553 : && (mask_type = get_mask_type_for_scalar_type (vinfo, op0_type))
6745 : 971553 : && expand_vec_cmp_expr_p (vectype, mask_type, code))
6746 : 1489432 : precision = GET_MODE_BITSIZE (mode);
6747 : : }
6748 : : }
6749 : : else
6750 : : {
6751 : 165193 : gphi *phi = as_a <gphi *> (stmt_info->stmt);
6752 : 603140 : for (unsigned i = 0; i < gimple_phi_num_args (phi); ++i)
6753 : : {
6754 : 437947 : tree rhs = gimple_phi_arg_def (phi, i);
6755 : :
6756 : 437947 : stmt_vec_info def_stmt_info = vinfo->lookup_def (rhs);
6757 : 437947 : if (!def_stmt_info)
6758 : : /* Don't let external or constant operands influence the choice.
6759 : : We can convert them to whichever vector type we pick. */
6760 : 258100 : continue;
6761 : :
6762 : 179847 : if (def_stmt_info->mask_precision)
6763 : : {
6764 : 155874 : if (precision > def_stmt_info->mask_precision)
6765 : 437947 : precision = def_stmt_info->mask_precision;
6766 : : }
6767 : : }
6768 : : }
6769 : :
6770 : 1808401 : if (dump_enabled_p ())
6771 : : {
6772 : 6908 : if (precision == ~0U)
6773 : 1667 : dump_printf_loc (MSG_NOTE, vect_location,
6774 : : "using normal nonmask vectors for %G",
6775 : : stmt_info->stmt);
6776 : : else
6777 : 5241 : dump_printf_loc (MSG_NOTE, vect_location,
6778 : : "using boolean precision %d for %G",
6779 : : precision, stmt_info->stmt);
6780 : : }
6781 : :
6782 : 1808401 : stmt_info->mask_precision = precision;
6783 : : }
6784 : :
6785 : : /* Handle vect_determine_precisions for STMT_INFO, given that we
6786 : : have already done so for the users of its result. */
6787 : :
6788 : : void
6789 : 33978850 : vect_determine_stmt_precisions (vec_info *vinfo, stmt_vec_info stmt_info)
6790 : : {
6791 : 33978850 : vect_determine_min_output_precision (vinfo, stmt_info);
6792 : 33978850 : if (gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt))
6793 : : {
6794 : 18691700 : vect_determine_precisions_from_range (stmt_info, stmt);
6795 : 18691700 : vect_determine_precisions_from_users (stmt_info, stmt);
6796 : : }
6797 : 33978850 : }
6798 : :
6799 : : /* Walk backwards through the vectorizable region to determine the
6800 : : values of these fields:
6801 : :
6802 : : - min_output_precision
6803 : : - min_input_precision
6804 : : - operation_precision
6805 : : - operation_sign. */
6806 : :
6807 : : void
6808 : 892880 : vect_determine_precisions (vec_info *vinfo)
6809 : : {
6810 : 892880 : basic_block *bbs = vinfo->bbs;
6811 : 892880 : unsigned int nbbs = vinfo->nbbs;
6812 : :
6813 : 892880 : DUMP_VECT_SCOPE ("vect_determine_precisions");
6814 : :
6815 : 11432570 : for (unsigned int i = 0; i < nbbs; i++)
6816 : : {
6817 : 10539690 : basic_block bb = bbs[i];
6818 : 17298973 : for (auto gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
6819 : : {
6820 : 6759283 : stmt_vec_info stmt_info = vinfo->lookup_stmt (gsi.phi ());
6821 : 6759283 : if (stmt_info && STMT_VINFO_VECTORIZABLE (stmt_info))
6822 : 6581601 : vect_determine_mask_precision (vinfo, stmt_info);
6823 : : }
6824 : 98247556 : for (auto gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
6825 : : {
6826 : 77168176 : stmt_vec_info stmt_info = vinfo->lookup_stmt (gsi_stmt (gsi));
6827 : 77168176 : if (stmt_info && STMT_VINFO_VECTORIZABLE (stmt_info))
6828 : 27397249 : vect_determine_mask_precision (vinfo, stmt_info);
6829 : : }
6830 : : }
6831 : 11432570 : for (unsigned int i = 0; i < nbbs; i++)
6832 : : {
6833 : 10539690 : basic_block bb = bbs[nbbs - i - 1];
6834 : 175415732 : for (auto gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi))
6835 : : {
6836 : 77168176 : stmt_vec_info stmt_info = vinfo->lookup_stmt (gsi_stmt (gsi));
6837 : 77168176 : if (stmt_info && STMT_VINFO_VECTORIZABLE (stmt_info))
6838 : 27397249 : vect_determine_stmt_precisions (vinfo, stmt_info);
6839 : : }
6840 : 17298973 : for (auto gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
6841 : : {
6842 : 6759283 : stmt_vec_info stmt_info = vinfo->lookup_stmt (gsi.phi ());
6843 : 6759283 : if (stmt_info && STMT_VINFO_VECTORIZABLE (stmt_info))
6844 : 6581601 : vect_determine_stmt_precisions (vinfo, stmt_info);
6845 : : }
6846 : : }
6847 : 892880 : }
6848 : :
6849 : : typedef gimple *(*vect_recog_func_ptr) (vec_info *, stmt_vec_info, tree *);
6850 : :
6851 : : struct vect_recog_func
6852 : : {
6853 : : vect_recog_func_ptr fn;
6854 : : const char *name;
6855 : : };
6856 : :
6857 : : /* Note that ordering matters - the first pattern matching on a stmt is
6858 : : taken which means usually the more complex one needs to preceed the
6859 : : less comples onex (widen_sum only after dot_prod or sad for example). */
6860 : : static vect_recog_func vect_vect_recog_func_ptrs[] = {
6861 : : { vect_recog_bitfield_ref_pattern, "bitfield_ref" },
6862 : : { vect_recog_bit_insert_pattern, "bit_insert" },
6863 : : { vect_recog_abd_pattern, "abd" },
6864 : : { vect_recog_over_widening_pattern, "over_widening" },
6865 : : /* Must come after over_widening, which narrows the shift as much as
6866 : : possible beforehand. */
6867 : : { vect_recog_average_pattern, "average" },
6868 : : { vect_recog_cond_expr_convert_pattern, "cond_expr_convert" },
6869 : : { vect_recog_mulhs_pattern, "mult_high" },
6870 : : { vect_recog_cast_forwprop_pattern, "cast_forwprop" },
6871 : : { vect_recog_widen_mult_pattern, "widen_mult" },
6872 : : { vect_recog_dot_prod_pattern, "dot_prod" },
6873 : : { vect_recog_sad_pattern, "sad" },
6874 : : { vect_recog_widen_sum_pattern, "widen_sum" },
6875 : : { vect_recog_pow_pattern, "pow" },
6876 : : { vect_recog_popcount_clz_ctz_ffs_pattern, "popcount_clz_ctz_ffs" },
6877 : : { vect_recog_ctz_ffs_pattern, "ctz_ffs" },
6878 : : { vect_recog_widen_shift_pattern, "widen_shift" },
6879 : : { vect_recog_rotate_pattern, "rotate" },
6880 : : { vect_recog_vector_vector_shift_pattern, "vector_vector_shift" },
6881 : : { vect_recog_divmod_pattern, "divmod" },
6882 : : { vect_recog_mod_var_pattern, "modvar" },
6883 : : { vect_recog_mult_pattern, "mult" },
6884 : : { vect_recog_sat_add_pattern, "sat_add" },
6885 : : { vect_recog_sat_sub_pattern, "sat_sub" },
6886 : : { vect_recog_sat_trunc_pattern, "sat_trunc" },
6887 : : { vect_recog_gcond_pattern, "gcond" },
6888 : : { vect_recog_bool_pattern, "bool" },
6889 : : /* This must come before mask conversion, and includes the parts
6890 : : of mask conversion that are needed for gather and scatter
6891 : : internal functions. */
6892 : : { vect_recog_gather_scatter_pattern, "gather_scatter" },
6893 : : { vect_recog_cond_store_pattern, "cond_store" },
6894 : : { vect_recog_mask_conversion_pattern, "mask_conversion" },
6895 : : { vect_recog_widen_plus_pattern, "widen_plus" },
6896 : : { vect_recog_widen_minus_pattern, "widen_minus" },
6897 : : { vect_recog_widen_abd_pattern, "widen_abd" },
6898 : : /* These must come after the double widening ones. */
6899 : : };
6900 : :
6901 : : /* Mark statements that are involved in a pattern. */
6902 : :
6903 : : void
6904 : 714618 : vect_mark_pattern_stmts (vec_info *vinfo,
6905 : : stmt_vec_info orig_stmt_info, gimple *pattern_stmt,
6906 : : tree pattern_vectype)
6907 : : {
6908 : 714618 : stmt_vec_info orig_stmt_info_saved = orig_stmt_info;
6909 : 714618 : gimple *def_seq = STMT_VINFO_PATTERN_DEF_SEQ (orig_stmt_info);
6910 : :
6911 : 714618 : gimple *orig_pattern_stmt = NULL;
6912 : 714618 : if (is_pattern_stmt_p (orig_stmt_info))
6913 : : {
6914 : : /* We're replacing a statement in an existing pattern definition
6915 : : sequence. */
6916 : 11142 : orig_pattern_stmt = orig_stmt_info->stmt;
6917 : 11142 : if (dump_enabled_p ())
6918 : 603 : dump_printf_loc (MSG_NOTE, vect_location,
6919 : : "replacing earlier pattern %G", orig_pattern_stmt);
6920 : :
6921 : : /* To keep the book-keeping simple, just swap the lhs of the
6922 : : old and new statements, so that the old one has a valid but
6923 : : unused lhs. */
6924 : 11142 : tree old_lhs = gimple_get_lhs (orig_pattern_stmt);
6925 : 11142 : gimple_set_lhs (orig_pattern_stmt, gimple_get_lhs (pattern_stmt));
6926 : 11142 : gimple_set_lhs (pattern_stmt, old_lhs);
6927 : :
6928 : 11142 : if (dump_enabled_p ())
6929 : 603 : dump_printf_loc (MSG_NOTE, vect_location, "with %G", pattern_stmt);
6930 : :
6931 : : /* Switch to the statement that ORIG replaces. */
6932 : 11142 : orig_stmt_info = STMT_VINFO_RELATED_STMT (orig_stmt_info);
6933 : :
6934 : : /* We shouldn't be replacing the main pattern statement. */
6935 : 11142 : gcc_assert (STMT_VINFO_RELATED_STMT (orig_stmt_info)->stmt
6936 : : != orig_pattern_stmt);
6937 : : }
6938 : :
6939 : 714618 : if (def_seq)
6940 : : for (gimple_stmt_iterator si = gsi_start (def_seq);
6941 : 1552086 : !gsi_end_p (si); gsi_next (&si))
6942 : : {
6943 : 938314 : if (dump_enabled_p ())
6944 : 21988 : dump_printf_loc (MSG_NOTE, vect_location,
6945 : : "extra pattern stmt: %G", gsi_stmt (si));
6946 : 938314 : stmt_vec_info pattern_stmt_info
6947 : 938314 : = vect_init_pattern_stmt (vinfo, gsi_stmt (si),
6948 : : orig_stmt_info, pattern_vectype);
6949 : : /* Stmts in the def sequence are not vectorizable cycle or
6950 : : induction defs, instead they should all be vect_internal_def
6951 : : feeding the main pattern stmt which retains this def type. */
6952 : 938314 : STMT_VINFO_DEF_TYPE (pattern_stmt_info) = vect_internal_def;
6953 : : }
6954 : :
6955 : 714618 : if (orig_pattern_stmt)
6956 : : {
6957 : 11142 : vect_init_pattern_stmt (vinfo, pattern_stmt,
6958 : : orig_stmt_info, pattern_vectype);
6959 : :
6960 : : /* Insert all the new pattern statements before the original one. */
6961 : 11142 : gimple_seq *orig_def_seq = &STMT_VINFO_PATTERN_DEF_SEQ (orig_stmt_info);
6962 : 11142 : gimple_stmt_iterator gsi = gsi_for_stmt (orig_pattern_stmt,
6963 : : orig_def_seq);
6964 : 11142 : gsi_insert_seq_before_without_update (&gsi, def_seq, GSI_SAME_STMT);
6965 : 11142 : gsi_insert_before_without_update (&gsi, pattern_stmt, GSI_SAME_STMT);
6966 : :
6967 : : /* Remove the pattern statement that this new pattern replaces. */
6968 : 11142 : gsi_remove (&gsi, false);
6969 : : }
6970 : : else
6971 : 703476 : vect_set_pattern_stmt (vinfo,
6972 : : pattern_stmt, orig_stmt_info, pattern_vectype);
6973 : :
6974 : : /* For any conditionals mark them as vect_condition_def. */
6975 : 714618 : if (is_a <gcond *> (pattern_stmt))
6976 : 240970 : STMT_VINFO_DEF_TYPE (STMT_VINFO_RELATED_STMT (orig_stmt_info)) = vect_condition_def;
6977 : :
6978 : : /* Transfer reduction path info to the pattern. */
6979 : 714618 : if (STMT_VINFO_REDUC_IDX (orig_stmt_info_saved) != -1)
6980 : : {
6981 : 8259 : gimple_match_op op;
6982 : 8259 : if (!gimple_extract_op (orig_stmt_info_saved->stmt, &op))
6983 : 0 : gcc_unreachable ();
6984 : 8259 : tree lookfor = op.ops[STMT_VINFO_REDUC_IDX (orig_stmt_info)];
6985 : : /* Search the pattern def sequence and the main pattern stmt. Note
6986 : : we may have inserted all into a containing pattern def sequence
6987 : : so the following is a bit awkward. */
6988 : 8259 : gimple_stmt_iterator si;
6989 : 8259 : gimple *s;
6990 : 8259 : if (def_seq)
6991 : : {
6992 : 7599 : si = gsi_start (def_seq);
6993 : 7599 : s = gsi_stmt (si);
6994 : 7599 : gsi_next (&si);
6995 : : }
6996 : : else
6997 : : {
6998 : : si = gsi_none ();
6999 : : s = pattern_stmt;
7000 : : }
7001 : 15973 : do
7002 : : {
7003 : 15973 : bool found = false;
7004 : 15973 : if (gimple_extract_op (s, &op))
7005 : 38240 : for (unsigned i = 0; i < op.num_ops; ++i)
7006 : 30688 : if (op.ops[i] == lookfor)
7007 : : {
7008 : 8421 : STMT_VINFO_REDUC_IDX (vinfo->lookup_stmt (s)) = i;
7009 : 8421 : lookfor = gimple_get_lhs (s);
7010 : 8421 : found = true;
7011 : 8421 : break;
7012 : : }
7013 : 15973 : if (s == pattern_stmt)
7014 : : {
7015 : 8259 : if (!found && dump_enabled_p ())
7016 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
7017 : : "failed to update reduction index.\n");
7018 : 8259 : break;
7019 : : }
7020 : 7714 : if (gsi_end_p (si))
7021 : : s = pattern_stmt;
7022 : : else
7023 : : {
7024 : 115 : s = gsi_stmt (si);
7025 : 115 : if (s == pattern_stmt)
7026 : : /* Found the end inside a bigger pattern def seq. */
7027 : : si = gsi_none ();
7028 : : else
7029 : 115 : gsi_next (&si);
7030 : : }
7031 : : } while (1);
7032 : : }
7033 : 714618 : }
7034 : :
7035 : : /* Function vect_pattern_recog_1
7036 : :
7037 : : Input:
7038 : : PATTERN_RECOG_FUNC: A pointer to a function that detects a certain
7039 : : computation pattern.
7040 : : STMT_INFO: A stmt from which the pattern search should start.
7041 : :
7042 : : If PATTERN_RECOG_FUNC successfully detected the pattern, it creates
7043 : : a sequence of statements that has the same functionality and can be
7044 : : used to replace STMT_INFO. It returns the last statement in the sequence
7045 : : and adds any earlier statements to STMT_INFO's STMT_VINFO_PATTERN_DEF_SEQ.
7046 : : PATTERN_RECOG_FUNC also sets *TYPE_OUT to the vector type of the final
7047 : : statement, having first checked that the target supports the new operation
7048 : : in that type.
7049 : :
7050 : : This function also does some bookkeeping, as explained in the documentation
7051 : : for vect_recog_pattern. */
7052 : :
7053 : : static void
7054 : 890552645 : vect_pattern_recog_1 (vec_info *vinfo,
7055 : : const vect_recog_func &recog_func, stmt_vec_info stmt_info)
7056 : : {
7057 : 890552645 : gimple *pattern_stmt;
7058 : 890552645 : tree pattern_vectype;
7059 : :
7060 : : /* If this statement has already been replaced with pattern statements,
7061 : : leave the original statement alone, since the first match wins.
7062 : : Instead try to match against the definition statements that feed
7063 : : the main pattern statement. */
7064 : 890552645 : if (STMT_VINFO_IN_PATTERN_P (stmt_info))
7065 : : {
7066 : 9793607 : gimple_stmt_iterator gsi;
7067 : 9793607 : for (gsi = gsi_start (STMT_VINFO_PATTERN_DEF_SEQ (stmt_info));
7068 : 23634284 : !gsi_end_p (gsi); gsi_next (&gsi))
7069 : 13840677 : vect_pattern_recog_1 (vinfo, recog_func,
7070 : : vinfo->lookup_stmt (gsi_stmt (gsi)));
7071 : : return;
7072 : : }
7073 : :
7074 : 880759038 : gcc_assert (!STMT_VINFO_PATTERN_DEF_SEQ (stmt_info));
7075 : 880759038 : pattern_stmt = recog_func.fn (vinfo, stmt_info, &pattern_vectype);
7076 : 880759038 : if (!pattern_stmt)
7077 : : {
7078 : : /* Clear any half-formed pattern definition sequence. */
7079 : 880044420 : STMT_VINFO_PATTERN_DEF_SEQ (stmt_info) = NULL;
7080 : 880044420 : return;
7081 : : }
7082 : :
7083 : : /* Found a vectorizable pattern. */
7084 : 714618 : if (dump_enabled_p ())
7085 : 17659 : dump_printf_loc (MSG_NOTE, vect_location,
7086 : : "%s pattern recognized: %G",
7087 : 17659 : recog_func.name, pattern_stmt);
7088 : :
7089 : : /* Mark the stmts that are involved in the pattern. */
7090 : 714618 : vect_mark_pattern_stmts (vinfo, stmt_info, pattern_stmt, pattern_vectype);
7091 : : }
7092 : :
7093 : :
7094 : : /* Function vect_pattern_recog
7095 : :
7096 : : Input:
7097 : : LOOP_VINFO - a struct_loop_info of a loop in which we want to look for
7098 : : computation idioms.
7099 : :
7100 : : Output - for each computation idiom that is detected we create a new stmt
7101 : : that provides the same functionality and that can be vectorized. We
7102 : : also record some information in the struct_stmt_info of the relevant
7103 : : stmts, as explained below:
7104 : :
7105 : : At the entry to this function we have the following stmts, with the
7106 : : following initial value in the STMT_VINFO fields:
7107 : :
7108 : : stmt in_pattern_p related_stmt vec_stmt
7109 : : S1: a_i = .... - - -
7110 : : S2: a_2 = ..use(a_i).. - - -
7111 : : S3: a_1 = ..use(a_2).. - - -
7112 : : S4: a_0 = ..use(a_1).. - - -
7113 : : S5: ... = ..use(a_0).. - - -
7114 : :
7115 : : Say the sequence {S1,S2,S3,S4} was detected as a pattern that can be
7116 : : represented by a single stmt. We then:
7117 : : - create a new stmt S6 equivalent to the pattern (the stmt is not
7118 : : inserted into the code)
7119 : : - fill in the STMT_VINFO fields as follows:
7120 : :
7121 : : in_pattern_p related_stmt vec_stmt
7122 : : S1: a_i = .... - - -
7123 : : S2: a_2 = ..use(a_i).. - - -
7124 : : S3: a_1 = ..use(a_2).. - - -
7125 : : S4: a_0 = ..use(a_1).. true S6 -
7126 : : '---> S6: a_new = .... - S4 -
7127 : : S5: ... = ..use(a_0).. - - -
7128 : :
7129 : : (the last stmt in the pattern (S4) and the new pattern stmt (S6) point
7130 : : to each other through the RELATED_STMT field).
7131 : :
7132 : : S6 will be marked as relevant in vect_mark_stmts_to_be_vectorized instead
7133 : : of S4 because it will replace all its uses. Stmts {S1,S2,S3} will
7134 : : remain irrelevant unless used by stmts other than S4.
7135 : :
7136 : : If vectorization succeeds, vect_transform_stmt will skip over {S1,S2,S3}
7137 : : (because they are marked as irrelevant). It will vectorize S6, and record
7138 : : a pointer to the new vector stmt VS6 from S6 (as usual).
7139 : : S4 will be skipped, and S5 will be vectorized as usual:
7140 : :
7141 : : 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 : : > VS6: va_new = .... - - -
7146 : : S4: a_0 = ..use(a_1).. true S6 VS6
7147 : : '---> S6: a_new = .... - S4 VS6
7148 : : > VS5: ... = ..vuse(va_new).. - - -
7149 : : S5: ... = ..use(a_0).. - - -
7150 : :
7151 : : DCE could then get rid of {S1,S2,S3,S4,S5} (if their defs are not used
7152 : : elsewhere), and we'll end up with:
7153 : :
7154 : : VS6: va_new = ....
7155 : : VS5: ... = ..vuse(va_new)..
7156 : :
7157 : : In case of more than one pattern statements, e.g., widen-mult with
7158 : : intermediate type:
7159 : :
7160 : : S1 a_t = ;
7161 : : S2 a_T = (TYPE) a_t;
7162 : : '--> S3: a_it = (interm_type) a_t;
7163 : : S4 prod_T = a_T * CONST;
7164 : : '--> S5: prod_T' = a_it w* CONST;
7165 : :
7166 : : there may be other users of a_T outside the pattern. In that case S2 will
7167 : : be marked as relevant (as well as S3), and both S2 and S3 will be analyzed
7168 : : and vectorized. The vector stmt VS2 will be recorded in S2, and VS3 will
7169 : : be recorded in S3. */
7170 : :
7171 : : void
7172 : 892880 : vect_pattern_recog (vec_info *vinfo)
7173 : : {
7174 : 892880 : basic_block *bbs = vinfo->bbs;
7175 : 892880 : unsigned int nbbs = vinfo->nbbs;
7176 : :
7177 : 892880 : vect_determine_precisions (vinfo);
7178 : :
7179 : 892880 : DUMP_VECT_SCOPE ("vect_pattern_recog");
7180 : :
7181 : : /* Scan through the stmts in the region, applying the pattern recognition
7182 : : functions starting at each stmt visited. */
7183 : 11432570 : for (unsigned i = 0; i < nbbs; i++)
7184 : : {
7185 : 10539690 : basic_block bb = bbs[i];
7186 : :
7187 : 98247556 : for (auto si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
7188 : : {
7189 : 77168176 : stmt_vec_info stmt_info = vinfo->lookup_stmt (gsi_stmt (si));
7190 : :
7191 : 77168176 : if (!stmt_info || !STMT_VINFO_VECTORIZABLE (stmt_info))
7192 : 49770927 : continue;
7193 : :
7194 : : /* Scan over all generic vect_recog_xxx_pattern functions. */
7195 : 904109217 : for (const auto &func_ptr : vect_vect_recog_func_ptrs)
7196 : 876711968 : vect_pattern_recog_1 (vinfo, func_ptr,
7197 : : stmt_info);
7198 : : }
7199 : : }
7200 : :
7201 : : /* After this no more add_stmt calls are allowed. */
7202 : 892880 : vinfo->stmt_vec_info_ro = true;
7203 : 892880 : }
7204 : :
7205 : : /* Build a GIMPLE_ASSIGN or GIMPLE_CALL with the tree_code,
7206 : : or internal_fn contained in ch, respectively. */
7207 : : gimple *
7208 : 127822 : vect_gimple_build (tree lhs, code_helper ch, tree op0, tree op1)
7209 : : {
7210 : 127822 : gcc_assert (op0 != NULL_TREE);
7211 : 127822 : if (ch.is_tree_code ())
7212 : 127822 : return gimple_build_assign (lhs, (tree_code) ch, op0, op1);
7213 : :
7214 : 0 : gcc_assert (ch.is_internal_fn ());
7215 : 0 : gimple* stmt = gimple_build_call_internal (as_internal_fn ((combined_fn) ch),
7216 : : op1 == NULL_TREE ? 1 : 2,
7217 : : op0, op1);
7218 : 0 : gimple_call_set_lhs (stmt, lhs);
7219 : 0 : return stmt;
7220 : : }
|
