]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* SLP - Basic Block Vectorization |
818ab71a | 2 | Copyright (C) 2007-2016 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
c7131fb2 | 25 | #include "backend.h" |
957060b5 AM |
26 | #include "target.h" |
27 | #include "rtl.h" | |
ebfd146a | 28 | #include "tree.h" |
c7131fb2 | 29 | #include "gimple.h" |
957060b5 | 30 | #include "tree-pass.h" |
c7131fb2 | 31 | #include "ssa.h" |
957060b5 AM |
32 | #include "optabs-tree.h" |
33 | #include "insn-config.h" | |
34 | #include "recog.h" /* FIXME: for insn_data */ | |
957060b5 | 35 | #include "params.h" |
40e23961 | 36 | #include "fold-const.h" |
d8a2d370 | 37 | #include "stor-layout.h" |
5be5c238 | 38 | #include "gimple-iterator.h" |
ebfd146a | 39 | #include "cfgloop.h" |
ebfd146a | 40 | #include "tree-vectorizer.h" |
2635892a | 41 | #include "langhooks.h" |
642fce57 | 42 | #include "gimple-walk.h" |
428db0ba | 43 | #include "dbgcnt.h" |
a70d6342 IR |
44 | |
45 | ||
ebfd146a IR |
46 | /* Recursively free the memory allocated for the SLP tree rooted at NODE. */ |
47 | ||
48 | static void | |
49 | vect_free_slp_tree (slp_tree node) | |
50 | { | |
d092494c | 51 | int i; |
d755c7ef | 52 | slp_tree child; |
d092494c | 53 | |
9771b263 | 54 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) |
d755c7ef | 55 | vect_free_slp_tree (child); |
b8698a0f | 56 | |
78810bd3 RB |
57 | gimple *stmt; |
58 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt) | |
59 | /* After transform some stmts are removed and thus their vinfo is gone. */ | |
60 | if (vinfo_for_stmt (stmt)) | |
61 | { | |
62 | gcc_assert (STMT_VINFO_NUM_SLP_USES (vinfo_for_stmt (stmt)) > 0); | |
63 | STMT_VINFO_NUM_SLP_USES (vinfo_for_stmt (stmt))--; | |
64 | } | |
65 | ||
9771b263 DN |
66 | SLP_TREE_CHILDREN (node).release (); |
67 | SLP_TREE_SCALAR_STMTS (node).release (); | |
68 | SLP_TREE_VEC_STMTS (node).release (); | |
01d8bf07 | 69 | SLP_TREE_LOAD_PERMUTATION (node).release (); |
ebfd146a IR |
70 | |
71 | free (node); | |
72 | } | |
73 | ||
74 | ||
75 | /* Free the memory allocated for the SLP instance. */ | |
76 | ||
77 | void | |
78 | vect_free_slp_instance (slp_instance instance) | |
79 | { | |
80 | vect_free_slp_tree (SLP_INSTANCE_TREE (instance)); | |
9771b263 | 81 | SLP_INSTANCE_LOADS (instance).release (); |
c7e62a26 | 82 | free (instance); |
ebfd146a IR |
83 | } |
84 | ||
85 | ||
d092494c IR |
86 | /* Create an SLP node for SCALAR_STMTS. */ |
87 | ||
88 | static slp_tree | |
355fe088 | 89 | vect_create_new_slp_node (vec<gimple *> scalar_stmts) |
d092494c | 90 | { |
d3cfd39e | 91 | slp_tree node; |
355fe088 | 92 | gimple *stmt = scalar_stmts[0]; |
d092494c IR |
93 | unsigned int nops; |
94 | ||
95 | if (is_gimple_call (stmt)) | |
96 | nops = gimple_call_num_args (stmt); | |
97 | else if (is_gimple_assign (stmt)) | |
f7e531cf IR |
98 | { |
99 | nops = gimple_num_ops (stmt) - 1; | |
100 | if (gimple_assign_rhs_code (stmt) == COND_EXPR) | |
101 | nops++; | |
102 | } | |
d092494c IR |
103 | else |
104 | return NULL; | |
105 | ||
d3cfd39e | 106 | node = XNEW (struct _slp_tree); |
d092494c | 107 | SLP_TREE_SCALAR_STMTS (node) = scalar_stmts; |
9771b263 DN |
108 | SLP_TREE_VEC_STMTS (node).create (0); |
109 | SLP_TREE_CHILDREN (node).create (nops); | |
01d8bf07 | 110 | SLP_TREE_LOAD_PERMUTATION (node) = vNULL; |
6876e5bc | 111 | SLP_TREE_TWO_OPERATORS (node) = false; |
603cca93 | 112 | SLP_TREE_DEF_TYPE (node) = vect_internal_def; |
d092494c | 113 | |
78810bd3 RB |
114 | unsigned i; |
115 | FOR_EACH_VEC_ELT (scalar_stmts, i, stmt) | |
116 | STMT_VINFO_NUM_SLP_USES (vinfo_for_stmt (stmt))++; | |
117 | ||
d092494c IR |
118 | return node; |
119 | } | |
120 | ||
121 | ||
ddf56386 RB |
122 | /* This structure is used in creation of an SLP tree. Each instance |
123 | corresponds to the same operand in a group of scalar stmts in an SLP | |
124 | node. */ | |
125 | typedef struct _slp_oprnd_info | |
126 | { | |
127 | /* Def-stmts for the operands. */ | |
128 | vec<gimple *> def_stmts; | |
129 | /* Information about the first statement, its vector def-type, type, the | |
130 | operand itself in case it's constant, and an indication if it's a pattern | |
131 | stmt. */ | |
132 | enum vect_def_type first_dt; | |
133 | tree first_op_type; | |
134 | bool first_pattern; | |
135 | bool second_pattern; | |
136 | } *slp_oprnd_info; | |
137 | ||
138 | ||
d092494c IR |
139 | /* Allocate operands info for NOPS operands, and GROUP_SIZE def-stmts for each |
140 | operand. */ | |
9771b263 | 141 | static vec<slp_oprnd_info> |
d092494c IR |
142 | vect_create_oprnd_info (int nops, int group_size) |
143 | { | |
144 | int i; | |
145 | slp_oprnd_info oprnd_info; | |
9771b263 | 146 | vec<slp_oprnd_info> oprnds_info; |
d092494c | 147 | |
9771b263 | 148 | oprnds_info.create (nops); |
d092494c IR |
149 | for (i = 0; i < nops; i++) |
150 | { | |
151 | oprnd_info = XNEW (struct _slp_oprnd_info); | |
9771b263 | 152 | oprnd_info->def_stmts.create (group_size); |
d092494c | 153 | oprnd_info->first_dt = vect_uninitialized_def; |
793d9a16 | 154 | oprnd_info->first_op_type = NULL_TREE; |
d092494c | 155 | oprnd_info->first_pattern = false; |
effb52da | 156 | oprnd_info->second_pattern = false; |
9771b263 | 157 | oprnds_info.quick_push (oprnd_info); |
d092494c IR |
158 | } |
159 | ||
160 | return oprnds_info; | |
161 | } | |
162 | ||
163 | ||
d3cfd39e JJ |
164 | /* Free operands info. */ |
165 | ||
d092494c | 166 | static void |
9771b263 | 167 | vect_free_oprnd_info (vec<slp_oprnd_info> &oprnds_info) |
d092494c IR |
168 | { |
169 | int i; | |
170 | slp_oprnd_info oprnd_info; | |
171 | ||
9771b263 | 172 | FOR_EACH_VEC_ELT (oprnds_info, i, oprnd_info) |
d3cfd39e | 173 | { |
9771b263 | 174 | oprnd_info->def_stmts.release (); |
d3cfd39e JJ |
175 | XDELETE (oprnd_info); |
176 | } | |
d092494c | 177 | |
9771b263 | 178 | oprnds_info.release (); |
d092494c IR |
179 | } |
180 | ||
181 | ||
d755c7ef RB |
182 | /* Find the place of the data-ref in STMT in the interleaving chain that starts |
183 | from FIRST_STMT. Return -1 if the data-ref is not a part of the chain. */ | |
184 | ||
185 | static int | |
355fe088 | 186 | vect_get_place_in_interleaving_chain (gimple *stmt, gimple *first_stmt) |
d755c7ef | 187 | { |
355fe088 | 188 | gimple *next_stmt = first_stmt; |
d755c7ef RB |
189 | int result = 0; |
190 | ||
191 | if (first_stmt != GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) | |
192 | return -1; | |
193 | ||
194 | do | |
195 | { | |
196 | if (next_stmt == stmt) | |
197 | return result; | |
d755c7ef | 198 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
c8047699 RB |
199 | if (next_stmt) |
200 | result += GROUP_GAP (vinfo_for_stmt (next_stmt)); | |
d755c7ef RB |
201 | } |
202 | while (next_stmt); | |
203 | ||
204 | return -1; | |
205 | } | |
206 | ||
207 | ||
d092494c IR |
208 | /* Get the defs for the rhs of STMT (collect them in OPRNDS_INFO), check that |
209 | they are of a valid type and that they match the defs of the first stmt of | |
b0b4483e RB |
210 | the SLP group (stored in OPRNDS_INFO). If there was a fatal error |
211 | return -1, if the error could be corrected by swapping operands of the | |
212 | operation return 1, if everything is ok return 0. */ | |
ebfd146a | 213 | |
b0b4483e | 214 | static int |
310213d4 | 215 | vect_get_and_check_slp_defs (vec_info *vinfo, |
355fe088 | 216 | gimple *stmt, unsigned stmt_num, |
23847df4 | 217 | vec<slp_oprnd_info> *oprnds_info) |
ebfd146a IR |
218 | { |
219 | tree oprnd; | |
220 | unsigned int i, number_of_oprnds; | |
355fe088 | 221 | gimple *def_stmt; |
d092494c | 222 | enum vect_def_type dt = vect_uninitialized_def; |
d092494c | 223 | bool pattern = false; |
abf9bfbc | 224 | slp_oprnd_info oprnd_info; |
b0b4483e RB |
225 | int first_op_idx = 1; |
226 | bool commutative = false; | |
227 | bool first_op_cond = false; | |
effb52da RB |
228 | bool first = stmt_num == 0; |
229 | bool second = stmt_num == 1; | |
b8698a0f | 230 | |
d092494c | 231 | if (is_gimple_call (stmt)) |
190c2236 JJ |
232 | { |
233 | number_of_oprnds = gimple_call_num_args (stmt); | |
b0b4483e | 234 | first_op_idx = 3; |
190c2236 | 235 | } |
f7e531cf IR |
236 | else if (is_gimple_assign (stmt)) |
237 | { | |
b0b4483e | 238 | enum tree_code code = gimple_assign_rhs_code (stmt); |
f7e531cf | 239 | number_of_oprnds = gimple_num_ops (stmt) - 1; |
a414c77f IE |
240 | if (gimple_assign_rhs_code (stmt) == COND_EXPR |
241 | && COMPARISON_CLASS_P (gimple_assign_rhs1 (stmt))) | |
b0b4483e RB |
242 | { |
243 | first_op_cond = true; | |
244 | commutative = true; | |
245 | number_of_oprnds++; | |
246 | } | |
247 | else | |
248 | commutative = commutative_tree_code (code); | |
f7e531cf | 249 | } |
d092494c | 250 | else |
b0b4483e | 251 | return -1; |
ebfd146a | 252 | |
b0b4483e | 253 | bool swapped = false; |
ebfd146a IR |
254 | for (i = 0; i < number_of_oprnds; i++) |
255 | { | |
b0b4483e RB |
256 | again: |
257 | if (first_op_cond) | |
f7e531cf | 258 | { |
b0b4483e RB |
259 | if (i == 0 || i == 1) |
260 | oprnd = TREE_OPERAND (gimple_op (stmt, first_op_idx), | |
261 | swapped ? !i : i); | |
262 | else | |
263 | oprnd = gimple_op (stmt, first_op_idx + i - 1); | |
f7e531cf IR |
264 | } |
265 | else | |
b0b4483e | 266 | oprnd = gimple_op (stmt, first_op_idx + (swapped ? !i : i)); |
f7e531cf | 267 | |
9771b263 | 268 | oprnd_info = (*oprnds_info)[i]; |
ebfd146a | 269 | |
81c40241 | 270 | if (!vect_is_simple_use (oprnd, vinfo, &def_stmt, &dt)) |
ebfd146a | 271 | { |
73fbfcad | 272 | if (dump_enabled_p ()) |
ebfd146a | 273 | { |
78c60e3d | 274 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3fc356dc | 275 | "Build SLP failed: can't analyze def for "); |
78c60e3d | 276 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, oprnd); |
e645e942 | 277 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a IR |
278 | } |
279 | ||
b0b4483e | 280 | return -1; |
ebfd146a IR |
281 | } |
282 | ||
a70d6342 | 283 | /* Check if DEF_STMT is a part of a pattern in LOOP and get the def stmt |
ff802fa1 | 284 | from the pattern. Check that all the stmts of the node are in the |
ebfd146a | 285 | pattern. */ |
f5709183 | 286 | if (def_stmt && gimple_bb (def_stmt) |
61d371eb | 287 | && vect_stmt_in_region_p (vinfo, def_stmt) |
ebfd146a | 288 | && vinfo_for_stmt (def_stmt) |
83197f37 | 289 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (def_stmt)) |
f5709183 IR |
290 | && !STMT_VINFO_RELEVANT (vinfo_for_stmt (def_stmt)) |
291 | && !STMT_VINFO_LIVE_P (vinfo_for_stmt (def_stmt))) | |
ebfd146a | 292 | { |
d092494c | 293 | pattern = true; |
effb52da RB |
294 | if (!first && !oprnd_info->first_pattern |
295 | /* Allow different pattern state for the defs of the | |
296 | first stmt in reduction chains. */ | |
297 | && (oprnd_info->first_dt != vect_reduction_def | |
298 | || (!second && !oprnd_info->second_pattern))) | |
d092494c | 299 | { |
b0b4483e RB |
300 | if (i == 0 |
301 | && !swapped | |
302 | && commutative) | |
303 | { | |
304 | swapped = true; | |
305 | goto again; | |
306 | } | |
307 | ||
73fbfcad | 308 | if (dump_enabled_p ()) |
d092494c | 309 | { |
78c60e3d SS |
310 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
311 | "Build SLP failed: some of the stmts" | |
312 | " are in a pattern, and others are not "); | |
313 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, oprnd); | |
e645e942 | 314 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
d092494c | 315 | } |
ebfd146a | 316 | |
b0b4483e | 317 | return 1; |
ebfd146a IR |
318 | } |
319 | ||
320 | def_stmt = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (def_stmt)); | |
d092494c | 321 | dt = STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt)); |
ebfd146a | 322 | |
f7e531cf | 323 | if (dt == vect_unknown_def_type) |
ebfd146a | 324 | { |
73fbfcad | 325 | if (dump_enabled_p ()) |
78c60e3d | 326 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 327 | "Unsupported pattern.\n"); |
b0b4483e | 328 | return -1; |
ebfd146a IR |
329 | } |
330 | ||
331 | switch (gimple_code (def_stmt)) | |
332 | { | |
81c40241 RB |
333 | case GIMPLE_PHI: |
334 | case GIMPLE_ASSIGN: | |
335 | break; | |
336 | ||
337 | default: | |
338 | if (dump_enabled_p ()) | |
339 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
340 | "unsupported defining stmt:\n"); | |
341 | return -1; | |
ebfd146a IR |
342 | } |
343 | } | |
344 | ||
effb52da RB |
345 | if (second) |
346 | oprnd_info->second_pattern = pattern; | |
347 | ||
d092494c | 348 | if (first) |
ebfd146a | 349 | { |
d092494c IR |
350 | oprnd_info->first_dt = dt; |
351 | oprnd_info->first_pattern = pattern; | |
793d9a16 | 352 | oprnd_info->first_op_type = TREE_TYPE (oprnd); |
ebfd146a | 353 | } |
ebfd146a IR |
354 | else |
355 | { | |
d092494c IR |
356 | /* Not first stmt of the group, check that the def-stmt/s match |
357 | the def-stmt/s of the first stmt. Allow different definition | |
358 | types for reduction chains: the first stmt must be a | |
359 | vect_reduction_def (a phi node), and the rest | |
360 | vect_internal_def. */ | |
361 | if (((oprnd_info->first_dt != dt | |
362 | && !(oprnd_info->first_dt == vect_reduction_def | |
793d9a16 RB |
363 | && dt == vect_internal_def) |
364 | && !((oprnd_info->first_dt == vect_external_def | |
365 | || oprnd_info->first_dt == vect_constant_def) | |
366 | && (dt == vect_external_def | |
367 | || dt == vect_constant_def))) | |
368 | || !types_compatible_p (oprnd_info->first_op_type, | |
369 | TREE_TYPE (oprnd)))) | |
ebfd146a | 370 | { |
b0b4483e RB |
371 | /* Try swapping operands if we got a mismatch. */ |
372 | if (i == 0 | |
373 | && !swapped | |
374 | && commutative) | |
375 | { | |
376 | swapped = true; | |
377 | goto again; | |
378 | } | |
379 | ||
abf9bfbc RB |
380 | if (dump_enabled_p ()) |
381 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 382 | "Build SLP failed: different types\n"); |
d092494c | 383 | |
b0b4483e | 384 | return 1; |
ebfd146a IR |
385 | } |
386 | } | |
387 | ||
388 | /* Check the types of the definitions. */ | |
d092494c | 389 | switch (dt) |
ebfd146a IR |
390 | { |
391 | case vect_constant_def: | |
8644a673 | 392 | case vect_external_def: |
d092494c | 393 | case vect_reduction_def: |
ebfd146a | 394 | break; |
b8698a0f | 395 | |
8644a673 | 396 | case vect_internal_def: |
abf9bfbc | 397 | oprnd_info->def_stmts.quick_push (def_stmt); |
ebfd146a IR |
398 | break; |
399 | ||
400 | default: | |
401 | /* FORNOW: Not supported. */ | |
73fbfcad | 402 | if (dump_enabled_p ()) |
ebfd146a | 403 | { |
78c60e3d SS |
404 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
405 | "Build SLP failed: illegal type of def "); | |
81c40241 | 406 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, oprnd); |
e645e942 | 407 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a IR |
408 | } |
409 | ||
b0b4483e | 410 | return -1; |
ebfd146a IR |
411 | } |
412 | } | |
413 | ||
b0b4483e RB |
414 | /* Swap operands. */ |
415 | if (swapped) | |
416 | { | |
78810bd3 RB |
417 | /* If there are already uses of this stmt in a SLP instance then |
418 | we've committed to the operand order and can't swap it. */ | |
419 | if (STMT_VINFO_NUM_SLP_USES (vinfo_for_stmt (stmt)) != 0) | |
420 | { | |
421 | if (dump_enabled_p ()) | |
422 | { | |
423 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
424 | "Build SLP failed: cannot swap operands of " | |
425 | "shared stmt "); | |
426 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
427 | } | |
428 | return -1; | |
429 | } | |
430 | ||
b0b4483e RB |
431 | if (first_op_cond) |
432 | { | |
433 | tree cond = gimple_assign_rhs1 (stmt); | |
434 | swap_ssa_operands (stmt, &TREE_OPERAND (cond, 0), | |
435 | &TREE_OPERAND (cond, 1)); | |
436 | TREE_SET_CODE (cond, swap_tree_comparison (TREE_CODE (cond))); | |
437 | } | |
438 | else | |
439 | swap_ssa_operands (stmt, gimple_assign_rhs1_ptr (stmt), | |
440 | gimple_assign_rhs2_ptr (stmt)); | |
78810bd3 RB |
441 | if (dump_enabled_p ()) |
442 | { | |
443 | dump_printf_loc (MSG_NOTE, vect_location, | |
444 | "swapped operands to match def types in "); | |
445 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
446 | } | |
b0b4483e RB |
447 | } |
448 | ||
449 | return 0; | |
ebfd146a IR |
450 | } |
451 | ||
452 | ||
6983e6b5 RB |
453 | /* Verify if the scalar stmts STMTS are isomorphic, require data |
454 | permutation or are of unsupported types of operation. Return | |
455 | true if they are, otherwise return false and indicate in *MATCHES | |
456 | which stmts are not isomorphic to the first one. If MATCHES[0] | |
457 | is false then this indicates the comparison could not be | |
458 | carried out or the stmts will never be vectorized by SLP. */ | |
ebfd146a IR |
459 | |
460 | static bool | |
310213d4 | 461 | vect_build_slp_tree_1 (vec_info *vinfo, |
355fe088 | 462 | vec<gimple *> stmts, unsigned int group_size, |
6983e6b5 | 463 | unsigned nops, unsigned int *max_nunits, |
97a1a642 | 464 | bool *matches, bool *two_operators) |
ebfd146a | 465 | { |
ebfd146a | 466 | unsigned int i; |
355fe088 | 467 | gimple *first_stmt = stmts[0], *stmt = stmts[0]; |
6876e5bc RB |
468 | enum tree_code first_stmt_code = ERROR_MARK; |
469 | enum tree_code alt_stmt_code = ERROR_MARK; | |
470 | enum tree_code rhs_code = ERROR_MARK; | |
f7e531cf | 471 | enum tree_code first_cond_code = ERROR_MARK; |
ebfd146a | 472 | tree lhs; |
6983e6b5 | 473 | bool need_same_oprnds = false; |
e00cdb8a | 474 | tree vectype = NULL_TREE, scalar_type, first_op1 = NULL_TREE; |
ebfd146a IR |
475 | optab optab; |
476 | int icode; | |
ef4bddc2 RS |
477 | machine_mode optab_op2_mode; |
478 | machine_mode vec_mode; | |
ebfd146a | 479 | HOST_WIDE_INT dummy; |
355fe088 | 480 | gimple *first_load = NULL, *prev_first_load = NULL; |
d092494c | 481 | |
ebfd146a | 482 | /* For every stmt in NODE find its def stmt/s. */ |
9771b263 | 483 | FOR_EACH_VEC_ELT (stmts, i, stmt) |
ebfd146a | 484 | { |
6983e6b5 RB |
485 | matches[i] = false; |
486 | ||
73fbfcad | 487 | if (dump_enabled_p ()) |
ebfd146a | 488 | { |
78c60e3d SS |
489 | dump_printf_loc (MSG_NOTE, vect_location, "Build SLP for "); |
490 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
491 | } |
492 | ||
4b5caab7 IR |
493 | /* Fail to vectorize statements marked as unvectorizable. */ |
494 | if (!STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (stmt))) | |
495 | { | |
73fbfcad | 496 | if (dump_enabled_p ()) |
4b5caab7 | 497 | { |
78c60e3d SS |
498 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
499 | "Build SLP failed: unvectorizable statement "); | |
500 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 501 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4b5caab7 | 502 | } |
6983e6b5 RB |
503 | /* Fatal mismatch. */ |
504 | matches[0] = false; | |
4b5caab7 IR |
505 | return false; |
506 | } | |
507 | ||
ebfd146a IR |
508 | lhs = gimple_get_lhs (stmt); |
509 | if (lhs == NULL_TREE) | |
510 | { | |
73fbfcad | 511 | if (dump_enabled_p ()) |
ebfd146a | 512 | { |
78c60e3d SS |
513 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
514 | "Build SLP failed: not GIMPLE_ASSIGN nor " | |
515 | "GIMPLE_CALL "); | |
516 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 517 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a | 518 | } |
6983e6b5 RB |
519 | /* Fatal mismatch. */ |
520 | matches[0] = false; | |
ebfd146a IR |
521 | return false; |
522 | } | |
523 | ||
b8698a0f | 524 | scalar_type = vect_get_smallest_scalar_type (stmt, &dummy, &dummy); |
ebfd146a IR |
525 | vectype = get_vectype_for_scalar_type (scalar_type); |
526 | if (!vectype) | |
527 | { | |
73fbfcad | 528 | if (dump_enabled_p ()) |
ebfd146a | 529 | { |
78c60e3d SS |
530 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
531 | "Build SLP failed: unsupported data-type "); | |
532 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
533 | scalar_type); | |
e645e942 | 534 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a | 535 | } |
6983e6b5 RB |
536 | /* Fatal mismatch. */ |
537 | matches[0] = false; | |
ebfd146a IR |
538 | return false; |
539 | } | |
b8698a0f | 540 | |
dfc55d30 RB |
541 | /* If populating the vector type requires unrolling then fail |
542 | before adjusting *max_nunits for basic-block vectorization. */ | |
310213d4 | 543 | if (is_a <bb_vec_info> (vinfo) |
dfc55d30 RB |
544 | && TYPE_VECTOR_SUBPARTS (vectype) > group_size) |
545 | { | |
546 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
547 | "Build SLP failed: unrolling required " | |
548 | "in basic block SLP\n"); | |
549 | /* Fatal mismatch. */ | |
550 | matches[0] = false; | |
551 | return false; | |
552 | } | |
553 | ||
4ef69dfc IR |
554 | /* In case of multiple types we need to detect the smallest type. */ |
555 | if (*max_nunits < TYPE_VECTOR_SUBPARTS (vectype)) | |
97a1a642 | 556 | *max_nunits = TYPE_VECTOR_SUBPARTS (vectype); |
b8698a0f | 557 | |
538dd0b7 | 558 | if (gcall *call_stmt = dyn_cast <gcall *> (stmt)) |
190c2236 JJ |
559 | { |
560 | rhs_code = CALL_EXPR; | |
538dd0b7 DM |
561 | if (gimple_call_internal_p (call_stmt) |
562 | || gimple_call_tail_p (call_stmt) | |
563 | || gimple_call_noreturn_p (call_stmt) | |
564 | || !gimple_call_nothrow_p (call_stmt) | |
565 | || gimple_call_chain (call_stmt)) | |
190c2236 | 566 | { |
73fbfcad | 567 | if (dump_enabled_p ()) |
190c2236 | 568 | { |
78c60e3d SS |
569 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
570 | "Build SLP failed: unsupported call type "); | |
538dd0b7 DM |
571 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, |
572 | call_stmt, 0); | |
e645e942 | 573 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
190c2236 | 574 | } |
6983e6b5 RB |
575 | /* Fatal mismatch. */ |
576 | matches[0] = false; | |
190c2236 JJ |
577 | return false; |
578 | } | |
579 | } | |
ebfd146a IR |
580 | else |
581 | rhs_code = gimple_assign_rhs_code (stmt); | |
582 | ||
583 | /* Check the operation. */ | |
584 | if (i == 0) | |
585 | { | |
586 | first_stmt_code = rhs_code; | |
587 | ||
b8698a0f | 588 | /* Shift arguments should be equal in all the packed stmts for a |
ebfd146a IR |
589 | vector shift with scalar shift operand. */ |
590 | if (rhs_code == LSHIFT_EXPR || rhs_code == RSHIFT_EXPR | |
591 | || rhs_code == LROTATE_EXPR | |
592 | || rhs_code == RROTATE_EXPR) | |
593 | { | |
594 | vec_mode = TYPE_MODE (vectype); | |
595 | ||
596 | /* First see if we have a vector/vector shift. */ | |
597 | optab = optab_for_tree_code (rhs_code, vectype, | |
598 | optab_vector); | |
599 | ||
600 | if (!optab | |
947131ba | 601 | || optab_handler (optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
602 | { |
603 | /* No vector/vector shift, try for a vector/scalar shift. */ | |
604 | optab = optab_for_tree_code (rhs_code, vectype, | |
605 | optab_scalar); | |
606 | ||
607 | if (!optab) | |
608 | { | |
73fbfcad | 609 | if (dump_enabled_p ()) |
78c60e3d | 610 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 611 | "Build SLP failed: no optab.\n"); |
6983e6b5 RB |
612 | /* Fatal mismatch. */ |
613 | matches[0] = false; | |
ebfd146a IR |
614 | return false; |
615 | } | |
947131ba | 616 | icode = (int) optab_handler (optab, vec_mode); |
ebfd146a IR |
617 | if (icode == CODE_FOR_nothing) |
618 | { | |
73fbfcad | 619 | if (dump_enabled_p ()) |
78c60e3d SS |
620 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
621 | "Build SLP failed: " | |
e645e942 | 622 | "op not supported by target.\n"); |
6983e6b5 RB |
623 | /* Fatal mismatch. */ |
624 | matches[0] = false; | |
ebfd146a IR |
625 | return false; |
626 | } | |
627 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
628 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
629 | { | |
630 | need_same_oprnds = true; | |
631 | first_op1 = gimple_assign_rhs2 (stmt); | |
632 | } | |
633 | } | |
634 | } | |
36ba4aae IR |
635 | else if (rhs_code == WIDEN_LSHIFT_EXPR) |
636 | { | |
637 | need_same_oprnds = true; | |
638 | first_op1 = gimple_assign_rhs2 (stmt); | |
639 | } | |
ebfd146a IR |
640 | } |
641 | else | |
642 | { | |
6876e5bc RB |
643 | if (first_stmt_code != rhs_code |
644 | && alt_stmt_code == ERROR_MARK) | |
645 | alt_stmt_code = rhs_code; | |
ebfd146a IR |
646 | if (first_stmt_code != rhs_code |
647 | && (first_stmt_code != IMAGPART_EXPR | |
648 | || rhs_code != REALPART_EXPR) | |
649 | && (first_stmt_code != REALPART_EXPR | |
69f11a13 | 650 | || rhs_code != IMAGPART_EXPR) |
6876e5bc RB |
651 | /* Handle mismatches in plus/minus by computing both |
652 | and merging the results. */ | |
653 | && !((first_stmt_code == PLUS_EXPR | |
654 | || first_stmt_code == MINUS_EXPR) | |
655 | && (alt_stmt_code == PLUS_EXPR | |
656 | || alt_stmt_code == MINUS_EXPR) | |
657 | && rhs_code == alt_stmt_code) | |
0d0293ac | 658 | && !(STMT_VINFO_GROUPED_ACCESS (vinfo_for_stmt (stmt)) |
69f11a13 | 659 | && (first_stmt_code == ARRAY_REF |
38000232 | 660 | || first_stmt_code == BIT_FIELD_REF |
69f11a13 IR |
661 | || first_stmt_code == INDIRECT_REF |
662 | || first_stmt_code == COMPONENT_REF | |
663 | || first_stmt_code == MEM_REF))) | |
ebfd146a | 664 | { |
73fbfcad | 665 | if (dump_enabled_p ()) |
ebfd146a | 666 | { |
78c60e3d SS |
667 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
668 | "Build SLP failed: different operation " | |
669 | "in stmt "); | |
670 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
6876e5bc RB |
671 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
672 | "original stmt "); | |
673 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
674 | first_stmt, 0); | |
ebfd146a | 675 | } |
6983e6b5 RB |
676 | /* Mismatch. */ |
677 | continue; | |
ebfd146a | 678 | } |
b8698a0f L |
679 | |
680 | if (need_same_oprnds | |
ebfd146a IR |
681 | && !operand_equal_p (first_op1, gimple_assign_rhs2 (stmt), 0)) |
682 | { | |
73fbfcad | 683 | if (dump_enabled_p ()) |
ebfd146a | 684 | { |
78c60e3d SS |
685 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
686 | "Build SLP failed: different shift " | |
687 | "arguments in "); | |
688 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 689 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a | 690 | } |
6983e6b5 RB |
691 | /* Mismatch. */ |
692 | continue; | |
ebfd146a | 693 | } |
190c2236 JJ |
694 | |
695 | if (rhs_code == CALL_EXPR) | |
696 | { | |
355fe088 | 697 | gimple *first_stmt = stmts[0]; |
190c2236 JJ |
698 | if (gimple_call_num_args (stmt) != nops |
699 | || !operand_equal_p (gimple_call_fn (first_stmt), | |
700 | gimple_call_fn (stmt), 0) | |
701 | || gimple_call_fntype (first_stmt) | |
702 | != gimple_call_fntype (stmt)) | |
703 | { | |
73fbfcad | 704 | if (dump_enabled_p ()) |
190c2236 | 705 | { |
78c60e3d SS |
706 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
707 | "Build SLP failed: different calls in "); | |
708 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
709 | stmt, 0); | |
e645e942 | 710 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
190c2236 | 711 | } |
6983e6b5 RB |
712 | /* Mismatch. */ |
713 | continue; | |
190c2236 JJ |
714 | } |
715 | } | |
ebfd146a IR |
716 | } |
717 | ||
0d0293ac MM |
718 | /* Grouped store or load. */ |
719 | if (STMT_VINFO_GROUPED_ACCESS (vinfo_for_stmt (stmt))) | |
ebfd146a IR |
720 | { |
721 | if (REFERENCE_CLASS_P (lhs)) | |
722 | { | |
723 | /* Store. */ | |
6983e6b5 | 724 | ; |
ebfd146a | 725 | } |
b5aeb3bb IR |
726 | else |
727 | { | |
728 | /* Load. */ | |
e14c1050 | 729 | first_load = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)); |
b5aeb3bb IR |
730 | if (prev_first_load) |
731 | { | |
732 | /* Check that there are no loads from different interleaving | |
6983e6b5 RB |
733 | chains in the same node. */ |
734 | if (prev_first_load != first_load) | |
78c60e3d | 735 | { |
73fbfcad | 736 | if (dump_enabled_p ()) |
b5aeb3bb | 737 | { |
78c60e3d SS |
738 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, |
739 | vect_location, | |
740 | "Build SLP failed: different " | |
741 | "interleaving chains in one node "); | |
742 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
743 | stmt, 0); | |
e645e942 | 744 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
b5aeb3bb | 745 | } |
6983e6b5 RB |
746 | /* Mismatch. */ |
747 | continue; | |
b5aeb3bb IR |
748 | } |
749 | } | |
750 | else | |
751 | prev_first_load = first_load; | |
ebfd146a | 752 | } |
0d0293ac | 753 | } /* Grouped access. */ |
ebfd146a IR |
754 | else |
755 | { | |
756 | if (TREE_CODE_CLASS (rhs_code) == tcc_reference) | |
757 | { | |
0d0293ac | 758 | /* Not grouped load. */ |
73fbfcad | 759 | if (dump_enabled_p ()) |
ebfd146a | 760 | { |
78c60e3d SS |
761 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
762 | "Build SLP failed: not grouped load "); | |
763 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 764 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a IR |
765 | } |
766 | ||
0d0293ac | 767 | /* FORNOW: Not grouped loads are not supported. */ |
6983e6b5 RB |
768 | /* Fatal mismatch. */ |
769 | matches[0] = false; | |
ebfd146a IR |
770 | return false; |
771 | } | |
772 | ||
773 | /* Not memory operation. */ | |
774 | if (TREE_CODE_CLASS (rhs_code) != tcc_binary | |
f7e531cf | 775 | && TREE_CODE_CLASS (rhs_code) != tcc_unary |
effb52da | 776 | && TREE_CODE_CLASS (rhs_code) != tcc_expression |
42fd8198 | 777 | && TREE_CODE_CLASS (rhs_code) != tcc_comparison |
190c2236 | 778 | && rhs_code != CALL_EXPR) |
ebfd146a | 779 | { |
73fbfcad | 780 | if (dump_enabled_p ()) |
ebfd146a | 781 | { |
78c60e3d SS |
782 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
783 | "Build SLP failed: operation"); | |
784 | dump_printf (MSG_MISSED_OPTIMIZATION, " unsupported "); | |
785 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 786 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a | 787 | } |
6983e6b5 RB |
788 | /* Fatal mismatch. */ |
789 | matches[0] = false; | |
ebfd146a IR |
790 | return false; |
791 | } | |
792 | ||
f7e531cf IR |
793 | if (rhs_code == COND_EXPR) |
794 | { | |
795 | tree cond_expr = gimple_assign_rhs1 (stmt); | |
796 | ||
797 | if (i == 0) | |
798 | first_cond_code = TREE_CODE (cond_expr); | |
799 | else if (first_cond_code != TREE_CODE (cond_expr)) | |
800 | { | |
73fbfcad | 801 | if (dump_enabled_p ()) |
f7e531cf | 802 | { |
78c60e3d SS |
803 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
804 | "Build SLP failed: different" | |
805 | " operation"); | |
806 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
807 | stmt, 0); | |
e645e942 | 808 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
f7e531cf | 809 | } |
6983e6b5 RB |
810 | /* Mismatch. */ |
811 | continue; | |
f7e531cf IR |
812 | } |
813 | } | |
ebfd146a | 814 | } |
6983e6b5 RB |
815 | |
816 | matches[i] = true; | |
817 | } | |
818 | ||
819 | for (i = 0; i < group_size; ++i) | |
820 | if (!matches[i]) | |
821 | return false; | |
822 | ||
6876e5bc RB |
823 | /* If we allowed a two-operation SLP node verify the target can cope |
824 | with the permute we are going to use. */ | |
825 | if (alt_stmt_code != ERROR_MARK | |
826 | && TREE_CODE_CLASS (alt_stmt_code) != tcc_reference) | |
827 | { | |
828 | unsigned char *sel | |
829 | = XALLOCAVEC (unsigned char, TYPE_VECTOR_SUBPARTS (vectype)); | |
830 | for (i = 0; i < TYPE_VECTOR_SUBPARTS (vectype); ++i) | |
831 | { | |
832 | sel[i] = i; | |
833 | if (gimple_assign_rhs_code (stmts[i % group_size]) == alt_stmt_code) | |
834 | sel[i] += TYPE_VECTOR_SUBPARTS (vectype); | |
835 | } | |
836 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) | |
837 | { | |
838 | for (i = 0; i < group_size; ++i) | |
839 | if (gimple_assign_rhs_code (stmts[i]) == alt_stmt_code) | |
840 | { | |
841 | matches[i] = false; | |
842 | if (dump_enabled_p ()) | |
843 | { | |
844 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
845 | "Build SLP failed: different operation " | |
846 | "in stmt "); | |
847 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
848 | stmts[i], 0); | |
849 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
850 | "original stmt "); | |
851 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
852 | first_stmt, 0); | |
853 | } | |
854 | } | |
855 | return false; | |
856 | } | |
857 | *two_operators = true; | |
858 | } | |
859 | ||
6983e6b5 RB |
860 | return true; |
861 | } | |
862 | ||
863 | /* Recursively build an SLP tree starting from NODE. | |
864 | Fail (and return a value not equal to zero) if def-stmts are not | |
865 | isomorphic, require data permutation or are of unsupported types of | |
866 | operation. Otherwise, return 0. | |
867 | The value returned is the depth in the SLP tree where a mismatch | |
868 | was found. */ | |
869 | ||
e403d17e | 870 | static slp_tree |
310213d4 | 871 | vect_build_slp_tree (vec_info *vinfo, |
e403d17e | 872 | vec<gimple *> stmts, unsigned int group_size, |
6983e6b5 RB |
873 | unsigned int *max_nunits, |
874 | vec<slp_tree> *loads, | |
1428105c RB |
875 | bool *matches, unsigned *npermutes, unsigned *tree_size, |
876 | unsigned max_tree_size) | |
6983e6b5 | 877 | { |
e403d17e | 878 | unsigned nops, i, this_tree_size = 0, this_max_nunits = *max_nunits; |
355fe088 | 879 | gimple *stmt; |
e403d17e | 880 | slp_tree node; |
6983e6b5 | 881 | |
6983e6b5 RB |
882 | matches[0] = false; |
883 | ||
e403d17e | 884 | stmt = stmts[0]; |
6983e6b5 RB |
885 | if (is_gimple_call (stmt)) |
886 | nops = gimple_call_num_args (stmt); | |
887 | else if (is_gimple_assign (stmt)) | |
888 | { | |
889 | nops = gimple_num_ops (stmt) - 1; | |
890 | if (gimple_assign_rhs_code (stmt) == COND_EXPR) | |
891 | nops++; | |
ebfd146a | 892 | } |
6983e6b5 | 893 | else |
e403d17e | 894 | return NULL; |
6983e6b5 | 895 | |
6876e5bc | 896 | bool two_operators = false; |
310213d4 | 897 | if (!vect_build_slp_tree_1 (vinfo, |
e403d17e RB |
898 | stmts, group_size, nops, |
899 | &this_max_nunits, matches, &two_operators)) | |
900 | return NULL; | |
ebfd146a | 901 | |
6983e6b5 RB |
902 | /* If the SLP node is a load, terminate the recursion. */ |
903 | if (STMT_VINFO_GROUPED_ACCESS (vinfo_for_stmt (stmt)) | |
904 | && DR_IS_READ (STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt)))) | |
ebfd146a | 905 | { |
e403d17e RB |
906 | *max_nunits = this_max_nunits; |
907 | node = vect_create_new_slp_node (stmts); | |
908 | loads->safe_push (node); | |
909 | return node; | |
ebfd146a IR |
910 | } |
911 | ||
6983e6b5 RB |
912 | /* Get at the operands, verifying they are compatible. */ |
913 | vec<slp_oprnd_info> oprnds_info = vect_create_oprnd_info (nops, group_size); | |
914 | slp_oprnd_info oprnd_info; | |
e403d17e | 915 | FOR_EACH_VEC_ELT (stmts, i, stmt) |
6983e6b5 | 916 | { |
310213d4 | 917 | switch (vect_get_and_check_slp_defs (vinfo, stmt, i, &oprnds_info)) |
6983e6b5 | 918 | { |
b0b4483e RB |
919 | case 0: |
920 | break; | |
921 | case -1: | |
922 | matches[0] = false; | |
6983e6b5 | 923 | vect_free_oprnd_info (oprnds_info); |
e403d17e | 924 | return NULL; |
b0b4483e RB |
925 | case 1: |
926 | matches[i] = false; | |
927 | break; | |
6983e6b5 RB |
928 | } |
929 | } | |
b0b4483e RB |
930 | for (i = 0; i < group_size; ++i) |
931 | if (!matches[i]) | |
932 | { | |
933 | vect_free_oprnd_info (oprnds_info); | |
e403d17e | 934 | return NULL; |
b0b4483e | 935 | } |
6983e6b5 | 936 | |
e403d17e RB |
937 | auto_vec<slp_tree, 4> children; |
938 | auto_vec<slp_tree> this_loads; | |
939 | ||
940 | stmt = stmts[0]; | |
6983e6b5 | 941 | |
b8698a0f | 942 | /* Create SLP_TREE nodes for the definition node/s. */ |
9771b263 | 943 | FOR_EACH_VEC_ELT (oprnds_info, i, oprnd_info) |
ebfd146a | 944 | { |
d092494c | 945 | slp_tree child; |
e403d17e RB |
946 | unsigned old_nloads = this_loads.length (); |
947 | unsigned old_tree_size = this_tree_size; | |
948 | unsigned int j; | |
b8698a0f | 949 | |
d092494c IR |
950 | if (oprnd_info->first_dt != vect_internal_def) |
951 | continue; | |
ebfd146a | 952 | |
1428105c RB |
953 | if (++this_tree_size > max_tree_size) |
954 | { | |
e403d17e RB |
955 | FOR_EACH_VEC_ELT (children, j, child) |
956 | vect_free_slp_tree (child); | |
1428105c | 957 | vect_free_oprnd_info (oprnds_info); |
e403d17e | 958 | return NULL; |
1428105c RB |
959 | } |
960 | ||
e403d17e RB |
961 | if ((child = vect_build_slp_tree (vinfo, oprnd_info->def_stmts, |
962 | group_size, &this_max_nunits, | |
963 | &this_loads, matches, npermutes, | |
964 | &this_tree_size, | |
965 | max_tree_size)) != NULL) | |
6983e6b5 | 966 | { |
3fc356dc RB |
967 | /* If we have all children of child built up from scalars then just |
968 | throw that away and build it up this node from scalars. */ | |
995b6fe0 RB |
969 | if (!SLP_TREE_CHILDREN (child).is_empty () |
970 | /* ??? Rejecting patterns this way doesn't work. We'd have to | |
971 | do extra work to cancel the pattern so the uses see the | |
972 | scalar version. */ | |
973 | && !is_pattern_stmt_p | |
974 | (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (child)[0]))) | |
3fc356dc | 975 | { |
3fc356dc RB |
976 | slp_tree grandchild; |
977 | ||
978 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (child), j, grandchild) | |
603cca93 | 979 | if (SLP_TREE_DEF_TYPE (grandchild) == vect_internal_def) |
3fc356dc RB |
980 | break; |
981 | if (!grandchild) | |
982 | { | |
983 | /* Roll back. */ | |
e403d17e RB |
984 | this_loads.truncate (old_nloads); |
985 | this_tree_size = old_tree_size; | |
3fc356dc | 986 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (child), j, grandchild) |
603cca93 | 987 | vect_free_slp_tree (grandchild); |
3fc356dc RB |
988 | SLP_TREE_CHILDREN (child).truncate (0); |
989 | ||
990 | dump_printf_loc (MSG_NOTE, vect_location, | |
991 | "Building parent vector operands from " | |
992 | "scalars instead\n"); | |
993 | oprnd_info->def_stmts = vNULL; | |
603cca93 | 994 | SLP_TREE_DEF_TYPE (child) = vect_external_def; |
e403d17e | 995 | children.safe_push (child); |
3fc356dc RB |
996 | continue; |
997 | } | |
998 | } | |
999 | ||
6983e6b5 | 1000 | oprnd_info->def_stmts = vNULL; |
e403d17e | 1001 | children.safe_push (child); |
6983e6b5 RB |
1002 | continue; |
1003 | } | |
1004 | ||
90dd6e3d RB |
1005 | /* If the SLP build failed fatally and we analyze a basic-block |
1006 | simply treat nodes we fail to build as externally defined | |
1007 | (and thus build vectors from the scalar defs). | |
1008 | The cost model will reject outright expensive cases. | |
1009 | ??? This doesn't treat cases where permutation ultimatively | |
1010 | fails (or we don't try permutation below). Ideally we'd | |
1011 | even compute a permutation that will end up with the maximum | |
1012 | SLP tree size... */ | |
310213d4 | 1013 | if (is_a <bb_vec_info> (vinfo) |
90dd6e3d RB |
1014 | && !matches[0] |
1015 | /* ??? Rejecting patterns this way doesn't work. We'd have to | |
1016 | do extra work to cancel the pattern so the uses see the | |
1017 | scalar version. */ | |
1018 | && !is_pattern_stmt_p (vinfo_for_stmt (stmt))) | |
1019 | { | |
1020 | dump_printf_loc (MSG_NOTE, vect_location, | |
1021 | "Building vector operands from scalars\n"); | |
e403d17e | 1022 | child = vect_create_new_slp_node (oprnd_info->def_stmts); |
603cca93 | 1023 | SLP_TREE_DEF_TYPE (child) = vect_external_def; |
e403d17e RB |
1024 | children.safe_push (child); |
1025 | oprnd_info->def_stmts = vNULL; | |
90dd6e3d RB |
1026 | continue; |
1027 | } | |
1028 | ||
6983e6b5 RB |
1029 | /* If the SLP build for operand zero failed and operand zero |
1030 | and one can be commutated try that for the scalar stmts | |
1031 | that failed the match. */ | |
1032 | if (i == 0 | |
1033 | /* A first scalar stmt mismatch signals a fatal mismatch. */ | |
1034 | && matches[0] | |
1035 | /* ??? For COND_EXPRs we can swap the comparison operands | |
1036 | as well as the arms under some constraints. */ | |
1037 | && nops == 2 | |
1038 | && oprnds_info[1]->first_dt == vect_internal_def | |
1039 | && is_gimple_assign (stmt) | |
1040 | && commutative_tree_code (gimple_assign_rhs_code (stmt)) | |
e403d17e | 1041 | && ! two_operators |
6983e6b5 RB |
1042 | /* Do so only if the number of not successful permutes was nor more |
1043 | than a cut-ff as re-trying the recursive match on | |
1044 | possibly each level of the tree would expose exponential | |
1045 | behavior. */ | |
1046 | && *npermutes < 4) | |
1047 | { | |
78810bd3 RB |
1048 | /* Verify if we can safely swap or if we committed to a specific |
1049 | operand order already. */ | |
1050 | for (j = 0; j < group_size; ++j) | |
1051 | if (!matches[j] | |
1052 | && STMT_VINFO_NUM_SLP_USES (vinfo_for_stmt (stmts[j])) != 0) | |
1053 | { | |
1054 | if (dump_enabled_p ()) | |
1055 | { | |
1056 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1057 | "Build SLP failed: cannot swap operands " | |
1058 | "of shared stmt "); | |
1059 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
1060 | stmts[j], 0); | |
1061 | } | |
1062 | goto fail; | |
1063 | } | |
1064 | ||
6983e6b5 | 1065 | /* Swap mismatched definition stmts. */ |
b0b4483e RB |
1066 | dump_printf_loc (MSG_NOTE, vect_location, |
1067 | "Re-trying with swapped operands of stmts "); | |
e72baed7 | 1068 | for (j = 0; j < group_size; ++j) |
6983e6b5 RB |
1069 | if (!matches[j]) |
1070 | { | |
6b4db501 MM |
1071 | std::swap (oprnds_info[0]->def_stmts[j], |
1072 | oprnds_info[1]->def_stmts[j]); | |
b0b4483e | 1073 | dump_printf (MSG_NOTE, "%d ", j); |
6983e6b5 | 1074 | } |
b0b4483e | 1075 | dump_printf (MSG_NOTE, "\n"); |
74574669 RB |
1076 | /* And try again with scratch 'matches' ... */ |
1077 | bool *tem = XALLOCAVEC (bool, group_size); | |
e403d17e RB |
1078 | if ((child = vect_build_slp_tree (vinfo, oprnd_info->def_stmts, |
1079 | group_size, &this_max_nunits, | |
1080 | &this_loads, tem, npermutes, | |
1081 | &this_tree_size, | |
1082 | max_tree_size)) != NULL) | |
6983e6b5 | 1083 | { |
60f2b864 RB |
1084 | /* ... so if successful we can apply the operand swapping |
1085 | to the GIMPLE IL. This is necessary because for example | |
1086 | vect_get_slp_defs uses operand indexes and thus expects | |
1087 | canonical operand order. This is also necessary even | |
1088 | if we end up building the operand from scalars as | |
1089 | we'll continue to process swapped operand two. */ | |
1090 | for (j = 0; j < group_size; ++j) | |
f47cda24 | 1091 | { |
e403d17e | 1092 | gimple *stmt = stmts[j]; |
f47cda24 RB |
1093 | gimple_set_plf (stmt, GF_PLF_1, false); |
1094 | } | |
1095 | for (j = 0; j < group_size; ++j) | |
1096 | { | |
e403d17e | 1097 | gimple *stmt = stmts[j]; |
f47cda24 RB |
1098 | if (!matches[j]) |
1099 | { | |
1100 | /* Avoid swapping operands twice. */ | |
1101 | if (gimple_plf (stmt, GF_PLF_1)) | |
1102 | continue; | |
1103 | swap_ssa_operands (stmt, gimple_assign_rhs1_ptr (stmt), | |
1104 | gimple_assign_rhs2_ptr (stmt)); | |
1105 | gimple_set_plf (stmt, GF_PLF_1, true); | |
1106 | } | |
1107 | } | |
1108 | /* Verify we swap all duplicates or none. */ | |
1109 | if (flag_checking) | |
1110 | for (j = 0; j < group_size; ++j) | |
60f2b864 | 1111 | { |
e403d17e | 1112 | gimple *stmt = stmts[j]; |
f47cda24 | 1113 | gcc_assert (gimple_plf (stmt, GF_PLF_1) == ! matches[j]); |
60f2b864 RB |
1114 | } |
1115 | ||
85c69b0b RB |
1116 | /* If we have all children of child built up from scalars then |
1117 | just throw that away and build it up this node from scalars. */ | |
995b6fe0 RB |
1118 | if (!SLP_TREE_CHILDREN (child).is_empty () |
1119 | /* ??? Rejecting patterns this way doesn't work. We'd have | |
1120 | to do extra work to cancel the pattern so the uses see the | |
1121 | scalar version. */ | |
1122 | && !is_pattern_stmt_p | |
1123 | (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (child)[0]))) | |
85c69b0b RB |
1124 | { |
1125 | unsigned int j; | |
1126 | slp_tree grandchild; | |
1127 | ||
1128 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (child), j, grandchild) | |
603cca93 | 1129 | if (SLP_TREE_DEF_TYPE (grandchild) == vect_internal_def) |
85c69b0b RB |
1130 | break; |
1131 | if (!grandchild) | |
1132 | { | |
1133 | /* Roll back. */ | |
e403d17e RB |
1134 | this_loads.truncate (old_nloads); |
1135 | this_tree_size = old_tree_size; | |
85c69b0b RB |
1136 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (child), j, grandchild) |
1137 | vect_free_slp_tree (grandchild); | |
1138 | SLP_TREE_CHILDREN (child).truncate (0); | |
1139 | ||
1140 | dump_printf_loc (MSG_NOTE, vect_location, | |
1141 | "Building parent vector operands from " | |
1142 | "scalars instead\n"); | |
1143 | oprnd_info->def_stmts = vNULL; | |
603cca93 | 1144 | SLP_TREE_DEF_TYPE (child) = vect_external_def; |
e403d17e | 1145 | children.safe_push (child); |
85c69b0b RB |
1146 | continue; |
1147 | } | |
1148 | } | |
1149 | ||
6983e6b5 | 1150 | oprnd_info->def_stmts = vNULL; |
e403d17e | 1151 | children.safe_push (child); |
6983e6b5 RB |
1152 | continue; |
1153 | } | |
1154 | ||
1155 | ++*npermutes; | |
1156 | } | |
1157 | ||
78810bd3 | 1158 | fail: |
e403d17e RB |
1159 | gcc_assert (child == NULL); |
1160 | FOR_EACH_VEC_ELT (children, j, child) | |
1161 | vect_free_slp_tree (child); | |
6983e6b5 | 1162 | vect_free_oprnd_info (oprnds_info); |
e403d17e | 1163 | return NULL; |
ebfd146a IR |
1164 | } |
1165 | ||
e403d17e RB |
1166 | vect_free_oprnd_info (oprnds_info); |
1167 | ||
1428105c RB |
1168 | if (tree_size) |
1169 | *tree_size += this_tree_size; | |
e403d17e RB |
1170 | *max_nunits = this_max_nunits; |
1171 | loads->safe_splice (this_loads); | |
1428105c | 1172 | |
e403d17e RB |
1173 | node = vect_create_new_slp_node (stmts); |
1174 | SLP_TREE_TWO_OPERATORS (node) = two_operators; | |
1175 | SLP_TREE_CHILDREN (node).splice (children); | |
1176 | return node; | |
ebfd146a IR |
1177 | } |
1178 | ||
78c60e3d | 1179 | /* Dump a slp tree NODE using flags specified in DUMP_KIND. */ |
ebfd146a IR |
1180 | |
1181 | static void | |
c2a12ca0 | 1182 | vect_print_slp_tree (int dump_kind, location_t loc, slp_tree node) |
ebfd146a IR |
1183 | { |
1184 | int i; | |
355fe088 | 1185 | gimple *stmt; |
d755c7ef | 1186 | slp_tree child; |
ebfd146a | 1187 | |
603cca93 RB |
1188 | dump_printf_loc (dump_kind, loc, "node%s\n", |
1189 | SLP_TREE_DEF_TYPE (node) != vect_internal_def | |
1190 | ? " (external)" : ""); | |
9771b263 | 1191 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt) |
ebfd146a | 1192 | { |
c2a12ca0 | 1193 | dump_printf_loc (dump_kind, loc, "\tstmt %d ", i); |
78c60e3d | 1194 | dump_gimple_stmt (dump_kind, TDF_SLIM, stmt, 0); |
ebfd146a | 1195 | } |
9771b263 | 1196 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) |
c2a12ca0 | 1197 | vect_print_slp_tree (dump_kind, loc, child); |
ebfd146a IR |
1198 | } |
1199 | ||
1200 | ||
b8698a0f L |
1201 | /* Mark the tree rooted at NODE with MARK (PURE_SLP or HYBRID). |
1202 | If MARK is HYBRID, it refers to a specific stmt in NODE (the stmt at index | |
ff802fa1 | 1203 | J). Otherwise, MARK is PURE_SLP and J is -1, which indicates that all the |
ebfd146a IR |
1204 | stmts in NODE are to be marked. */ |
1205 | ||
1206 | static void | |
1207 | vect_mark_slp_stmts (slp_tree node, enum slp_vect_type mark, int j) | |
1208 | { | |
1209 | int i; | |
355fe088 | 1210 | gimple *stmt; |
d755c7ef | 1211 | slp_tree child; |
ebfd146a | 1212 | |
603cca93 | 1213 | if (SLP_TREE_DEF_TYPE (node) != vect_internal_def) |
ebfd146a IR |
1214 | return; |
1215 | ||
9771b263 | 1216 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt) |
ebfd146a IR |
1217 | if (j < 0 || i == j) |
1218 | STMT_SLP_TYPE (vinfo_for_stmt (stmt)) = mark; | |
1219 | ||
9771b263 | 1220 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) |
d755c7ef | 1221 | vect_mark_slp_stmts (child, mark, j); |
ebfd146a IR |
1222 | } |
1223 | ||
1224 | ||
a70d6342 IR |
1225 | /* Mark the statements of the tree rooted at NODE as relevant (vect_used). */ |
1226 | ||
1227 | static void | |
1228 | vect_mark_slp_stmts_relevant (slp_tree node) | |
1229 | { | |
1230 | int i; | |
355fe088 | 1231 | gimple *stmt; |
a70d6342 | 1232 | stmt_vec_info stmt_info; |
d755c7ef | 1233 | slp_tree child; |
a70d6342 | 1234 | |
603cca93 | 1235 | if (SLP_TREE_DEF_TYPE (node) != vect_internal_def) |
a70d6342 IR |
1236 | return; |
1237 | ||
9771b263 | 1238 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt) |
a70d6342 IR |
1239 | { |
1240 | stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 1241 | gcc_assert (!STMT_VINFO_RELEVANT (stmt_info) |
a70d6342 IR |
1242 | || STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_scope); |
1243 | STMT_VINFO_RELEVANT (stmt_info) = vect_used_in_scope; | |
1244 | } | |
1245 | ||
9771b263 | 1246 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) |
d755c7ef | 1247 | vect_mark_slp_stmts_relevant (child); |
a70d6342 IR |
1248 | } |
1249 | ||
1250 | ||
b5aeb3bb IR |
1251 | /* Rearrange the statements of NODE according to PERMUTATION. */ |
1252 | ||
1253 | static void | |
1254 | vect_slp_rearrange_stmts (slp_tree node, unsigned int group_size, | |
01d8bf07 | 1255 | vec<unsigned> permutation) |
b5aeb3bb | 1256 | { |
355fe088 TS |
1257 | gimple *stmt; |
1258 | vec<gimple *> tmp_stmts; | |
d755c7ef RB |
1259 | unsigned int i; |
1260 | slp_tree child; | |
b5aeb3bb | 1261 | |
9771b263 | 1262 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) |
d755c7ef | 1263 | vect_slp_rearrange_stmts (child, group_size, permutation); |
b5aeb3bb | 1264 | |
9771b263 DN |
1265 | gcc_assert (group_size == SLP_TREE_SCALAR_STMTS (node).length ()); |
1266 | tmp_stmts.create (group_size); | |
d755c7ef | 1267 | tmp_stmts.quick_grow_cleared (group_size); |
b5aeb3bb | 1268 | |
9771b263 | 1269 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt) |
d755c7ef | 1270 | tmp_stmts[permutation[i]] = stmt; |
b5aeb3bb | 1271 | |
9771b263 | 1272 | SLP_TREE_SCALAR_STMTS (node).release (); |
b5aeb3bb IR |
1273 | SLP_TREE_SCALAR_STMTS (node) = tmp_stmts; |
1274 | } | |
1275 | ||
1276 | ||
b266b968 RB |
1277 | /* Attempt to reorder stmts in a reduction chain so that we don't |
1278 | require any load permutation. Return true if that was possible, | |
1279 | otherwise return false. */ | |
1280 | ||
1281 | static bool | |
1282 | vect_attempt_slp_rearrange_stmts (slp_instance slp_instn) | |
1283 | { | |
1284 | unsigned int group_size = SLP_INSTANCE_GROUP_SIZE (slp_instn); | |
1285 | unsigned int i, j; | |
1286 | sbitmap load_index; | |
1287 | unsigned int lidx; | |
1288 | slp_tree node, load; | |
1289 | ||
1290 | /* Compare all the permutation sequences to the first one. We know | |
1291 | that at least one load is permuted. */ | |
1292 | node = SLP_INSTANCE_LOADS (slp_instn)[0]; | |
1293 | if (!node->load_permutation.exists ()) | |
1294 | return false; | |
1295 | for (i = 1; SLP_INSTANCE_LOADS (slp_instn).iterate (i, &load); ++i) | |
1296 | { | |
1297 | if (!load->load_permutation.exists ()) | |
1298 | return false; | |
1299 | FOR_EACH_VEC_ELT (load->load_permutation, j, lidx) | |
1300 | if (lidx != node->load_permutation[j]) | |
1301 | return false; | |
1302 | } | |
1303 | ||
1304 | /* Check that the loads in the first sequence are different and there | |
1305 | are no gaps between them. */ | |
1306 | load_index = sbitmap_alloc (group_size); | |
1307 | bitmap_clear (load_index); | |
1308 | FOR_EACH_VEC_ELT (node->load_permutation, i, lidx) | |
1309 | { | |
41eefe13 | 1310 | if (lidx >= group_size) |
6e078af8 ML |
1311 | { |
1312 | sbitmap_free (load_index); | |
1313 | return false; | |
1314 | } | |
b266b968 RB |
1315 | if (bitmap_bit_p (load_index, lidx)) |
1316 | { | |
1317 | sbitmap_free (load_index); | |
1318 | return false; | |
1319 | } | |
1320 | bitmap_set_bit (load_index, lidx); | |
1321 | } | |
1322 | for (i = 0; i < group_size; i++) | |
1323 | if (!bitmap_bit_p (load_index, i)) | |
1324 | { | |
1325 | sbitmap_free (load_index); | |
1326 | return false; | |
1327 | } | |
1328 | sbitmap_free (load_index); | |
1329 | ||
1330 | /* This permutation is valid for reduction. Since the order of the | |
1331 | statements in the nodes is not important unless they are memory | |
1332 | accesses, we can rearrange the statements in all the nodes | |
1333 | according to the order of the loads. */ | |
1334 | vect_slp_rearrange_stmts (SLP_INSTANCE_TREE (slp_instn), group_size, | |
1335 | node->load_permutation); | |
1336 | ||
1337 | /* We are done, no actual permutations need to be generated. */ | |
c4e360f4 | 1338 | unsigned int unrolling_factor = SLP_INSTANCE_UNROLLING_FACTOR (slp_instn); |
b266b968 | 1339 | FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node) |
c4e360f4 RB |
1340 | { |
1341 | gimple *first_stmt = SLP_TREE_SCALAR_STMTS (node)[0]; | |
1342 | first_stmt = GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt)); | |
1343 | /* But we have to keep those permutations that are required because | |
1344 | of handling of gaps. */ | |
1345 | if (unrolling_factor == 1 | |
1346 | || (group_size == GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
1347 | && GROUP_GAP (vinfo_for_stmt (first_stmt)) == 0)) | |
1348 | SLP_TREE_LOAD_PERMUTATION (node).release (); | |
cbd400b4 RB |
1349 | else |
1350 | for (j = 0; j < SLP_TREE_LOAD_PERMUTATION (node).length (); ++j) | |
1351 | SLP_TREE_LOAD_PERMUTATION (node)[j] = j; | |
c4e360f4 RB |
1352 | } |
1353 | ||
b266b968 RB |
1354 | return true; |
1355 | } | |
1356 | ||
01d8bf07 RB |
1357 | /* Check if the required load permutations in the SLP instance |
1358 | SLP_INSTN are supported. */ | |
ebfd146a IR |
1359 | |
1360 | static bool | |
01d8bf07 | 1361 | vect_supported_load_permutation_p (slp_instance slp_instn) |
ebfd146a | 1362 | { |
01d8bf07 RB |
1363 | unsigned int group_size = SLP_INSTANCE_GROUP_SIZE (slp_instn); |
1364 | unsigned int i, j, k, next; | |
6983e6b5 | 1365 | slp_tree node; |
a5b50aa1 | 1366 | gimple *stmt, *load, *next_load; |
ebfd146a | 1367 | |
73fbfcad | 1368 | if (dump_enabled_p ()) |
ebfd146a | 1369 | { |
78c60e3d | 1370 | dump_printf_loc (MSG_NOTE, vect_location, "Load permutation "); |
01d8bf07 RB |
1371 | FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node) |
1372 | if (node->load_permutation.exists ()) | |
1373 | FOR_EACH_VEC_ELT (node->load_permutation, j, next) | |
1374 | dump_printf (MSG_NOTE, "%d ", next); | |
1375 | else | |
bddc974e TJ |
1376 | for (k = 0; k < group_size; ++k) |
1377 | dump_printf (MSG_NOTE, "%d ", k); | |
e645e942 | 1378 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1379 | } |
1380 | ||
b5aeb3bb IR |
1381 | /* In case of reduction every load permutation is allowed, since the order |
1382 | of the reduction statements is not important (as opposed to the case of | |
0d0293ac | 1383 | grouped stores). The only condition we need to check is that all the |
b5aeb3bb IR |
1384 | load nodes are of the same size and have the same permutation (and then |
1385 | rearrange all the nodes of the SLP instance according to this | |
1386 | permutation). */ | |
1387 | ||
1388 | /* Check that all the load nodes are of the same size. */ | |
01d8bf07 | 1389 | /* ??? Can't we assert this? */ |
9771b263 | 1390 | FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node) |
6983e6b5 RB |
1391 | if (SLP_TREE_SCALAR_STMTS (node).length () != (unsigned) group_size) |
1392 | return false; | |
2200fc49 | 1393 | |
b5aeb3bb | 1394 | node = SLP_INSTANCE_TREE (slp_instn); |
9771b263 | 1395 | stmt = SLP_TREE_SCALAR_STMTS (node)[0]; |
b5aeb3bb | 1396 | |
b010117a | 1397 | /* Reduction (there are no data-refs in the root). |
b266b968 | 1398 | In reduction chain the order of the loads is not important. */ |
b010117a IR |
1399 | if (!STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt)) |
1400 | && !GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) | |
c4e360f4 | 1401 | vect_attempt_slp_rearrange_stmts (slp_instn); |
b5aeb3bb | 1402 | |
6aa904c4 IR |
1403 | /* In basic block vectorization we allow any subchain of an interleaving |
1404 | chain. | |
1405 | FORNOW: not supported in loop SLP because of realignment compications. */ | |
01d8bf07 | 1406 | if (STMT_VINFO_BB_VINFO (vinfo_for_stmt (stmt))) |
6aa904c4 | 1407 | { |
240a94da RB |
1408 | /* Check whether the loads in an instance form a subchain and thus |
1409 | no permutation is necessary. */ | |
9771b263 | 1410 | FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node) |
6aa904c4 | 1411 | { |
9626d143 RB |
1412 | if (!SLP_TREE_LOAD_PERMUTATION (node).exists ()) |
1413 | continue; | |
240a94da | 1414 | bool subchain_p = true; |
6aa904c4 | 1415 | next_load = NULL; |
9771b263 | 1416 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), j, load) |
6aa904c4 | 1417 | { |
5b5826c4 RB |
1418 | if (j != 0 |
1419 | && (next_load != load | |
1420 | || GROUP_GAP (vinfo_for_stmt (load)) != 1)) | |
240a94da RB |
1421 | { |
1422 | subchain_p = false; | |
1423 | break; | |
1424 | } | |
6aa904c4 IR |
1425 | next_load = GROUP_NEXT_ELEMENT (vinfo_for_stmt (load)); |
1426 | } | |
240a94da RB |
1427 | if (subchain_p) |
1428 | SLP_TREE_LOAD_PERMUTATION (node).release (); | |
1429 | else | |
1430 | { | |
1431 | /* Verify the permutation can be generated. */ | |
1432 | vec<tree> tem; | |
1433 | if (!vect_transform_slp_perm_load (node, tem, NULL, | |
1434 | 1, slp_instn, true)) | |
1435 | { | |
1436 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, | |
1437 | vect_location, | |
1438 | "unsupported load permutation\n"); | |
1439 | return false; | |
1440 | } | |
1441 | } | |
6aa904c4 | 1442 | } |
01d8bf07 | 1443 | return true; |
6aa904c4 IR |
1444 | } |
1445 | ||
9b999e8c | 1446 | /* For loop vectorization verify we can generate the permutation. */ |
01d8bf07 RB |
1447 | FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node) |
1448 | if (node->load_permutation.exists () | |
1449 | && !vect_transform_slp_perm_load | |
1450 | (node, vNULL, NULL, | |
1451 | SLP_INSTANCE_UNROLLING_FACTOR (slp_instn), slp_instn, true)) | |
1452 | return false; | |
9b999e8c | 1453 | |
01d8bf07 | 1454 | return true; |
ebfd146a IR |
1455 | } |
1456 | ||
1457 | ||
e4a707c4 | 1458 | /* Find the last store in SLP INSTANCE. */ |
ff802fa1 | 1459 | |
64900538 | 1460 | gimple * |
2e8ab70c | 1461 | vect_find_last_scalar_stmt_in_slp (slp_tree node) |
e4a707c4 | 1462 | { |
355fe088 | 1463 | gimple *last = NULL, *stmt; |
e4a707c4 | 1464 | |
2e8ab70c RB |
1465 | for (int i = 0; SLP_TREE_SCALAR_STMTS (node).iterate (i, &stmt); i++) |
1466 | { | |
1467 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
1468 | if (is_pattern_stmt_p (stmt_vinfo)) | |
1469 | last = get_later_stmt (STMT_VINFO_RELATED_STMT (stmt_vinfo), last); | |
1470 | else | |
1471 | last = get_later_stmt (stmt, last); | |
1472 | } | |
e4a707c4 | 1473 | |
2e8ab70c | 1474 | return last; |
e4a707c4 IR |
1475 | } |
1476 | ||
23847df4 RB |
1477 | /* Compute the cost for the SLP node NODE in the SLP instance INSTANCE. */ |
1478 | ||
1479 | static void | |
1a4b99c1 | 1480 | vect_analyze_slp_cost_1 (slp_instance instance, slp_tree node, |
23847df4 | 1481 | stmt_vector_for_cost *prologue_cost_vec, |
1a4b99c1 | 1482 | stmt_vector_for_cost *body_cost_vec, |
23847df4 RB |
1483 | unsigned ncopies_for_cost) |
1484 | { | |
603cca93 | 1485 | unsigned i, j; |
23847df4 | 1486 | slp_tree child; |
8155f4d8 | 1487 | gimple *stmt; |
23847df4 RB |
1488 | stmt_vec_info stmt_info; |
1489 | tree lhs; | |
23847df4 RB |
1490 | |
1491 | /* Recurse down the SLP tree. */ | |
1492 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) | |
603cca93 | 1493 | if (SLP_TREE_DEF_TYPE (child) == vect_internal_def) |
1a4b99c1 RB |
1494 | vect_analyze_slp_cost_1 (instance, child, prologue_cost_vec, |
1495 | body_cost_vec, ncopies_for_cost); | |
23847df4 RB |
1496 | |
1497 | /* Look at the first scalar stmt to determine the cost. */ | |
1498 | stmt = SLP_TREE_SCALAR_STMTS (node)[0]; | |
1499 | stmt_info = vinfo_for_stmt (stmt); | |
1500 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1501 | { | |
1502 | if (DR_IS_WRITE (STMT_VINFO_DATA_REF (stmt_info))) | |
1503 | vect_model_store_cost (stmt_info, ncopies_for_cost, false, | |
1504 | vect_uninitialized_def, | |
1505 | node, prologue_cost_vec, body_cost_vec); | |
1506 | else | |
1507 | { | |
23847df4 | 1508 | gcc_checking_assert (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info))); |
52eab378 RB |
1509 | if (SLP_TREE_LOAD_PERMUTATION (node).exists ()) |
1510 | { | |
8155f4d8 RB |
1511 | /* If the load is permuted then the alignment is determined by |
1512 | the first group element not by the first scalar stmt DR. */ | |
52eab378 RB |
1513 | stmt = GROUP_FIRST_ELEMENT (stmt_info); |
1514 | stmt_info = vinfo_for_stmt (stmt); | |
8155f4d8 RB |
1515 | /* Record the cost for the permutation. */ |
1516 | record_stmt_cost (body_cost_vec, ncopies_for_cost, vec_perm, | |
1517 | stmt_info, 0, vect_body); | |
1518 | /* And adjust the number of loads performed. */ | |
1519 | unsigned nunits | |
1520 | = TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info)); | |
1521 | ncopies_for_cost | |
1522 | = (GROUP_SIZE (stmt_info) - GROUP_GAP (stmt_info) | |
1523 | + nunits - 1) / nunits; | |
1524 | ncopies_for_cost *= SLP_INSTANCE_UNROLLING_FACTOR (instance); | |
52eab378 | 1525 | } |
8155f4d8 | 1526 | /* Record the cost for the vector loads. */ |
23847df4 RB |
1527 | vect_model_load_cost (stmt_info, ncopies_for_cost, false, |
1528 | node, prologue_cost_vec, body_cost_vec); | |
89483f99 | 1529 | return; |
23847df4 RB |
1530 | } |
1531 | } | |
89483f99 | 1532 | else |
6876e5bc RB |
1533 | { |
1534 | record_stmt_cost (body_cost_vec, ncopies_for_cost, vector_stmt, | |
1535 | stmt_info, 0, vect_body); | |
89483f99 RB |
1536 | if (SLP_TREE_TWO_OPERATORS (node)) |
1537 | { | |
1538 | record_stmt_cost (body_cost_vec, ncopies_for_cost, vector_stmt, | |
1539 | stmt_info, 0, vect_body); | |
1540 | record_stmt_cost (body_cost_vec, ncopies_for_cost, vec_perm, | |
1541 | stmt_info, 0, vect_body); | |
1542 | } | |
6876e5bc | 1543 | } |
23847df4 | 1544 | |
603cca93 RB |
1545 | /* Push SLP node def-type to stmts. */ |
1546 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) | |
1547 | if (SLP_TREE_DEF_TYPE (child) != vect_internal_def) | |
1548 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (child), j, stmt) | |
1549 | STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmt)) = SLP_TREE_DEF_TYPE (child); | |
1550 | ||
23847df4 RB |
1551 | /* Scan operands and account for prologue cost of constants/externals. |
1552 | ??? This over-estimates cost for multiple uses and should be | |
1553 | re-engineered. */ | |
603cca93 | 1554 | stmt = SLP_TREE_SCALAR_STMTS (node)[0]; |
23847df4 RB |
1555 | lhs = gimple_get_lhs (stmt); |
1556 | for (i = 0; i < gimple_num_ops (stmt); ++i) | |
1557 | { | |
81c40241 | 1558 | tree op = gimple_op (stmt, i); |
355fe088 | 1559 | gimple *def_stmt; |
23847df4 RB |
1560 | enum vect_def_type dt; |
1561 | if (!op || op == lhs) | |
1562 | continue; | |
81c40241 | 1563 | if (vect_is_simple_use (op, stmt_info->vinfo, &def_stmt, &dt)) |
2e8ab70c RB |
1564 | { |
1565 | /* Without looking at the actual initializer a vector of | |
1566 | constants can be implemented as load from the constant pool. | |
1567 | ??? We need to pass down stmt_info for a vector type | |
1568 | even if it points to the wrong stmt. */ | |
1569 | if (dt == vect_constant_def) | |
1570 | record_stmt_cost (prologue_cost_vec, 1, vector_load, | |
1571 | stmt_info, 0, vect_prologue); | |
1572 | else if (dt == vect_external_def) | |
1573 | record_stmt_cost (prologue_cost_vec, 1, vec_construct, | |
1574 | stmt_info, 0, vect_prologue); | |
1575 | } | |
23847df4 | 1576 | } |
603cca93 RB |
1577 | |
1578 | /* Restore stmt def-types. */ | |
1579 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) | |
1580 | if (SLP_TREE_DEF_TYPE (child) != vect_internal_def) | |
1581 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (child), j, stmt) | |
1582 | STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmt)) = vect_internal_def; | |
23847df4 RB |
1583 | } |
1584 | ||
1585 | /* Compute the cost for the SLP instance INSTANCE. */ | |
1586 | ||
1587 | static void | |
1a4b99c1 | 1588 | vect_analyze_slp_cost (slp_instance instance, void *data) |
23847df4 RB |
1589 | { |
1590 | stmt_vector_for_cost body_cost_vec, prologue_cost_vec; | |
1591 | unsigned ncopies_for_cost; | |
1592 | stmt_info_for_cost *si; | |
1593 | unsigned i; | |
1594 | ||
b939ea86 RB |
1595 | if (dump_enabled_p ()) |
1596 | dump_printf_loc (MSG_NOTE, vect_location, | |
1597 | "=== vect_analyze_slp_cost ===\n"); | |
1598 | ||
23847df4 RB |
1599 | /* Calculate the number of vector stmts to create based on the unrolling |
1600 | factor (number of vectors is 1 if NUNITS >= GROUP_SIZE, and is | |
1601 | GROUP_SIZE / NUNITS otherwise. */ | |
1602 | unsigned group_size = SLP_INSTANCE_GROUP_SIZE (instance); | |
1a4b99c1 RB |
1603 | slp_tree node = SLP_INSTANCE_TREE (instance); |
1604 | stmt_vec_info stmt_info = vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (node)[0]); | |
1605 | /* Adjust the group_size by the vectorization factor which is always one | |
1606 | for basic-block vectorization. */ | |
1607 | if (STMT_VINFO_LOOP_VINFO (stmt_info)) | |
1608 | group_size *= LOOP_VINFO_VECT_FACTOR (STMT_VINFO_LOOP_VINFO (stmt_info)); | |
1609 | unsigned nunits = TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info)); | |
1610 | /* For reductions look at a reduction operand in case the reduction | |
1611 | operation is widening like DOT_PROD or SAD. */ | |
1612 | if (!STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1613 | { | |
355fe088 | 1614 | gimple *stmt = SLP_TREE_SCALAR_STMTS (node)[0]; |
1a4b99c1 RB |
1615 | switch (gimple_assign_rhs_code (stmt)) |
1616 | { | |
1617 | case DOT_PROD_EXPR: | |
1618 | case SAD_EXPR: | |
1619 | nunits = TYPE_VECTOR_SUBPARTS (get_vectype_for_scalar_type | |
1620 | (TREE_TYPE (gimple_assign_rhs1 (stmt)))); | |
1621 | break; | |
1622 | default:; | |
1623 | } | |
1624 | } | |
23847df4 RB |
1625 | ncopies_for_cost = least_common_multiple (nunits, group_size) / nunits; |
1626 | ||
1627 | prologue_cost_vec.create (10); | |
1628 | body_cost_vec.create (10); | |
1a4b99c1 RB |
1629 | vect_analyze_slp_cost_1 (instance, SLP_INSTANCE_TREE (instance), |
1630 | &prologue_cost_vec, &body_cost_vec, | |
1631 | ncopies_for_cost); | |
23847df4 RB |
1632 | |
1633 | /* Record the prologue costs, which were delayed until we were | |
1a4b99c1 | 1634 | sure that SLP was successful. */ |
23847df4 RB |
1635 | FOR_EACH_VEC_ELT (prologue_cost_vec, i, si) |
1636 | { | |
1637 | struct _stmt_vec_info *stmt_info | |
1638 | = si->stmt ? vinfo_for_stmt (si->stmt) : NULL; | |
1639 | (void) add_stmt_cost (data, si->count, si->kind, stmt_info, | |
1640 | si->misalign, vect_prologue); | |
1641 | } | |
1642 | ||
1a4b99c1 RB |
1643 | /* Record the instance's instructions in the target cost model. */ |
1644 | FOR_EACH_VEC_ELT (body_cost_vec, i, si) | |
1645 | { | |
1646 | struct _stmt_vec_info *stmt_info | |
1647 | = si->stmt ? vinfo_for_stmt (si->stmt) : NULL; | |
1648 | (void) add_stmt_cost (data, si->count, si->kind, stmt_info, | |
1649 | si->misalign, vect_body); | |
1650 | } | |
1651 | ||
23847df4 | 1652 | prologue_cost_vec.release (); |
1a4b99c1 | 1653 | body_cost_vec.release (); |
23847df4 | 1654 | } |
e4a707c4 | 1655 | |
1ba91a49 AL |
1656 | /* Splits a group of stores, currently beginning at FIRST_STMT, into two groups: |
1657 | one (still beginning at FIRST_STMT) of size GROUP1_SIZE (also containing | |
1658 | the first GROUP1_SIZE stmts, since stores are consecutive), the second | |
1659 | containing the remainder. | |
1660 | Return the first stmt in the second group. */ | |
1661 | ||
1662 | static gimple * | |
1663 | vect_split_slp_store_group (gimple *first_stmt, unsigned group1_size) | |
1664 | { | |
1665 | stmt_vec_info first_vinfo = vinfo_for_stmt (first_stmt); | |
1666 | gcc_assert (GROUP_FIRST_ELEMENT (first_vinfo) == first_stmt); | |
1667 | gcc_assert (group1_size > 0); | |
1668 | int group2_size = GROUP_SIZE (first_vinfo) - group1_size; | |
1669 | gcc_assert (group2_size > 0); | |
1670 | GROUP_SIZE (first_vinfo) = group1_size; | |
1671 | ||
1672 | gimple *stmt = first_stmt; | |
1673 | for (unsigned i = group1_size; i > 1; i--) | |
1674 | { | |
1675 | stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (stmt)); | |
1676 | gcc_assert (GROUP_GAP (vinfo_for_stmt (stmt)) == 1); | |
1677 | } | |
1678 | /* STMT is now the last element of the first group. */ | |
1679 | gimple *group2 = GROUP_NEXT_ELEMENT (vinfo_for_stmt (stmt)); | |
1680 | GROUP_NEXT_ELEMENT (vinfo_for_stmt (stmt)) = 0; | |
1681 | ||
1682 | GROUP_SIZE (vinfo_for_stmt (group2)) = group2_size; | |
1683 | for (stmt = group2; stmt; stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (stmt))) | |
1684 | { | |
1685 | GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) = group2; | |
1686 | gcc_assert (GROUP_GAP (vinfo_for_stmt (stmt)) == 1); | |
1687 | } | |
1688 | ||
1689 | /* For the second group, the GROUP_GAP is that before the original group, | |
1690 | plus skipping over the first vector. */ | |
1691 | GROUP_GAP (vinfo_for_stmt (group2)) = | |
1692 | GROUP_GAP (first_vinfo) + group1_size; | |
1693 | ||
1694 | /* GROUP_GAP of the first group now has to skip over the second group too. */ | |
1695 | GROUP_GAP (first_vinfo) += group2_size; | |
1696 | ||
1697 | if (dump_enabled_p ()) | |
1698 | dump_printf_loc (MSG_NOTE, vect_location, "Split group into %d and %d\n", | |
1699 | group1_size, group2_size); | |
1700 | ||
1701 | return group2; | |
1702 | } | |
1703 | ||
0d0293ac | 1704 | /* Analyze an SLP instance starting from a group of grouped stores. Call |
b8698a0f | 1705 | vect_build_slp_tree to build a tree of packed stmts if possible. |
ebfd146a IR |
1706 | Return FALSE if it's impossible to SLP any stmt in the loop. */ |
1707 | ||
1708 | static bool | |
310213d4 | 1709 | vect_analyze_slp_instance (vec_info *vinfo, |
355fe088 | 1710 | gimple *stmt, unsigned max_tree_size) |
ebfd146a IR |
1711 | { |
1712 | slp_instance new_instance; | |
d092494c | 1713 | slp_tree node; |
e14c1050 | 1714 | unsigned int group_size = GROUP_SIZE (vinfo_for_stmt (stmt)); |
ebfd146a | 1715 | unsigned int unrolling_factor = 1, nunits; |
b5aeb3bb | 1716 | tree vectype, scalar_type = NULL_TREE; |
355fe088 | 1717 | gimple *next; |
1ba91a49 | 1718 | unsigned int i; |
ebfd146a | 1719 | unsigned int max_nunits = 0; |
9771b263 | 1720 | vec<slp_tree> loads; |
b5aeb3bb | 1721 | struct data_reference *dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt)); |
355fe088 | 1722 | vec<gimple *> scalar_stmts; |
b5aeb3bb | 1723 | |
b010117a | 1724 | if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) |
b5aeb3bb | 1725 | { |
b010117a IR |
1726 | if (dr) |
1727 | { | |
1728 | scalar_type = TREE_TYPE (DR_REF (dr)); | |
1729 | vectype = get_vectype_for_scalar_type (scalar_type); | |
1730 | } | |
1731 | else | |
1732 | { | |
310213d4 | 1733 | gcc_assert (is_a <loop_vec_info> (vinfo)); |
b010117a IR |
1734 | vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt)); |
1735 | } | |
1736 | ||
e14c1050 | 1737 | group_size = GROUP_SIZE (vinfo_for_stmt (stmt)); |
b5aeb3bb IR |
1738 | } |
1739 | else | |
1740 | { | |
310213d4 | 1741 | gcc_assert (is_a <loop_vec_info> (vinfo)); |
b5aeb3bb | 1742 | vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt)); |
310213d4 | 1743 | group_size = as_a <loop_vec_info> (vinfo)->reductions.length (); |
b5aeb3bb | 1744 | } |
b8698a0f | 1745 | |
ebfd146a IR |
1746 | if (!vectype) |
1747 | { | |
73fbfcad | 1748 | if (dump_enabled_p ()) |
ebfd146a | 1749 | { |
78c60e3d SS |
1750 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1751 | "Build SLP failed: unsupported data-type "); | |
1752 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, scalar_type); | |
e645e942 | 1753 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a | 1754 | } |
b5aeb3bb | 1755 | |
ebfd146a IR |
1756 | return false; |
1757 | } | |
ebfd146a | 1758 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
a70d6342 | 1759 | |
a70d6342 IR |
1760 | /* Calculate the unrolling factor. */ |
1761 | unrolling_factor = least_common_multiple (nunits, group_size) / group_size; | |
310213d4 | 1762 | if (unrolling_factor != 1 && is_a <bb_vec_info> (vinfo)) |
a70d6342 | 1763 | { |
73fbfcad | 1764 | if (dump_enabled_p ()) |
e645e942 | 1765 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
78c60e3d | 1766 | "Build SLP failed: unrolling required in basic" |
e645e942 | 1767 | " block SLP\n"); |
b8698a0f | 1768 | |
a70d6342 IR |
1769 | return false; |
1770 | } | |
1771 | ||
0d0293ac | 1772 | /* Create a node (a root of the SLP tree) for the packed grouped stores. */ |
9771b263 | 1773 | scalar_stmts.create (group_size); |
ebfd146a | 1774 | next = stmt; |
b010117a | 1775 | if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) |
ebfd146a | 1776 | { |
b5aeb3bb IR |
1777 | /* Collect the stores and store them in SLP_TREE_SCALAR_STMTS. */ |
1778 | while (next) | |
1779 | { | |
f7e531cf IR |
1780 | if (STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (next)) |
1781 | && STMT_VINFO_RELATED_STMT (vinfo_for_stmt (next))) | |
9771b263 DN |
1782 | scalar_stmts.safe_push ( |
1783 | STMT_VINFO_RELATED_STMT (vinfo_for_stmt (next))); | |
f7e531cf | 1784 | else |
9771b263 | 1785 | scalar_stmts.safe_push (next); |
e14c1050 | 1786 | next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next)); |
b5aeb3bb | 1787 | } |
14a61437 RB |
1788 | /* Mark the first element of the reduction chain as reduction to properly |
1789 | transform the node. In the reduction analysis phase only the last | |
1790 | element of the chain is marked as reduction. */ | |
1791 | if (!STMT_VINFO_GROUPED_ACCESS (vinfo_for_stmt (stmt))) | |
1792 | STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmt)) = vect_reduction_def; | |
b5aeb3bb IR |
1793 | } |
1794 | else | |
1795 | { | |
1796 | /* Collect reduction statements. */ | |
310213d4 | 1797 | vec<gimple *> reductions = as_a <loop_vec_info> (vinfo)->reductions; |
9771b263 DN |
1798 | for (i = 0; reductions.iterate (i, &next); i++) |
1799 | scalar_stmts.safe_push (next); | |
ebfd146a IR |
1800 | } |
1801 | ||
9771b263 | 1802 | loads.create (group_size); |
ebfd146a IR |
1803 | |
1804 | /* Build the tree for the SLP instance. */ | |
89d390e5 RB |
1805 | bool *matches = XALLOCAVEC (bool, group_size); |
1806 | unsigned npermutes = 0; | |
e403d17e RB |
1807 | if ((node = vect_build_slp_tree (vinfo, scalar_stmts, group_size, |
1808 | &max_nunits, &loads, matches, &npermutes, | |
1809 | NULL, max_tree_size)) != NULL) | |
ebfd146a | 1810 | { |
4ef69dfc | 1811 | /* Calculate the unrolling factor based on the smallest type. */ |
ebfd146a IR |
1812 | if (max_nunits > nunits) |
1813 | unrolling_factor = least_common_multiple (max_nunits, group_size) | |
1814 | / group_size; | |
b8698a0f | 1815 | |
310213d4 | 1816 | if (unrolling_factor != 1 && is_a <bb_vec_info> (vinfo)) |
4ef69dfc | 1817 | { |
73fbfcad | 1818 | if (dump_enabled_p ()) |
e645e942 | 1819 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
78c60e3d | 1820 | "Build SLP failed: unrolling required in basic" |
e645e942 | 1821 | " block SLP\n"); |
c7e62a26 | 1822 | vect_free_slp_tree (node); |
9771b263 | 1823 | loads.release (); |
4ef69dfc IR |
1824 | return false; |
1825 | } | |
1826 | ||
1827 | /* Create a new SLP instance. */ | |
1828 | new_instance = XNEW (struct _slp_instance); | |
1829 | SLP_INSTANCE_TREE (new_instance) = node; | |
1830 | SLP_INSTANCE_GROUP_SIZE (new_instance) = group_size; | |
ebfd146a | 1831 | SLP_INSTANCE_UNROLLING_FACTOR (new_instance) = unrolling_factor; |
ebfd146a | 1832 | SLP_INSTANCE_LOADS (new_instance) = loads; |
abf9bfbc RB |
1833 | |
1834 | /* Compute the load permutation. */ | |
1835 | slp_tree load_node; | |
1836 | bool loads_permuted = false; | |
abf9bfbc RB |
1837 | FOR_EACH_VEC_ELT (loads, i, load_node) |
1838 | { | |
01d8bf07 | 1839 | vec<unsigned> load_permutation; |
abf9bfbc | 1840 | int j; |
355fe088 | 1841 | gimple *load, *first_stmt; |
01d8bf07 RB |
1842 | bool this_load_permuted = false; |
1843 | load_permutation.create (group_size); | |
6983e6b5 RB |
1844 | first_stmt = GROUP_FIRST_ELEMENT |
1845 | (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (load_node)[0])); | |
abf9bfbc RB |
1846 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (load_node), j, load) |
1847 | { | |
6983e6b5 RB |
1848 | int load_place |
1849 | = vect_get_place_in_interleaving_chain (load, first_stmt); | |
1850 | gcc_assert (load_place != -1); | |
1851 | if (load_place != j) | |
01d8bf07 | 1852 | this_load_permuted = true; |
abf9bfbc RB |
1853 | load_permutation.safe_push (load_place); |
1854 | } | |
fe2bef71 RB |
1855 | if (!this_load_permuted |
1856 | /* The load requires permutation when unrolling exposes | |
1857 | a gap either because the group is larger than the SLP | |
1858 | group-size or because there is a gap between the groups. */ | |
1859 | && (unrolling_factor == 1 | |
1860 | || (group_size == GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
1861 | && GROUP_GAP (vinfo_for_stmt (first_stmt)) == 0))) | |
01d8bf07 RB |
1862 | { |
1863 | load_permutation.release (); | |
1864 | continue; | |
1865 | } | |
1866 | SLP_TREE_LOAD_PERMUTATION (load_node) = load_permutation; | |
1867 | loads_permuted = true; | |
abf9bfbc | 1868 | } |
6aa904c4 IR |
1869 | |
1870 | if (loads_permuted) | |
ebfd146a | 1871 | { |
01d8bf07 | 1872 | if (!vect_supported_load_permutation_p (new_instance)) |
ebfd146a | 1873 | { |
73fbfcad | 1874 | if (dump_enabled_p ()) |
ebfd146a | 1875 | { |
e645e942 | 1876 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
78c60e3d SS |
1877 | "Build SLP failed: unsupported load " |
1878 | "permutation "); | |
1879 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 1880 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a | 1881 | } |
ebfd146a IR |
1882 | vect_free_slp_instance (new_instance); |
1883 | return false; | |
1884 | } | |
ebfd146a | 1885 | } |
ebfd146a | 1886 | |
bb0f5ca7 AL |
1887 | /* If the loads and stores can be handled with load/store-lane |
1888 | instructions do not generate this SLP instance. */ | |
1889 | if (is_a <loop_vec_info> (vinfo) | |
1890 | && loads_permuted | |
1891 | && dr && vect_store_lanes_supported (vectype, group_size)) | |
1892 | { | |
1893 | slp_tree load_node; | |
1894 | FOR_EACH_VEC_ELT (loads, i, load_node) | |
1895 | { | |
1896 | gimple *first_stmt = GROUP_FIRST_ELEMENT | |
1897 | (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (load_node)[0])); | |
1898 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (first_stmt); | |
1899 | /* Use SLP for strided accesses (or if we can't load-lanes). */ | |
1900 | if (STMT_VINFO_STRIDED_P (stmt_vinfo) | |
1901 | || ! vect_load_lanes_supported | |
1902 | (STMT_VINFO_VECTYPE (stmt_vinfo), | |
1903 | GROUP_SIZE (stmt_vinfo))) | |
1904 | break; | |
1905 | } | |
1906 | if (i == loads.length ()) | |
1907 | { | |
1908 | if (dump_enabled_p ()) | |
1909 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1910 | "Built SLP cancelled: can use " | |
1911 | "load/store-lanes\n"); | |
1912 | vect_free_slp_instance (new_instance); | |
1913 | return false; | |
1914 | } | |
1915 | } | |
1916 | ||
310213d4 | 1917 | vinfo->slp_instances.safe_push (new_instance); |
b8698a0f | 1918 | |
73fbfcad | 1919 | if (dump_enabled_p ()) |
c2a12ca0 RB |
1920 | { |
1921 | dump_printf_loc (MSG_NOTE, vect_location, | |
1922 | "Final SLP tree for instance:\n"); | |
1923 | vect_print_slp_tree (MSG_NOTE, vect_location, node); | |
1924 | } | |
ebfd146a IR |
1925 | |
1926 | return true; | |
1927 | } | |
1928 | ||
1929 | /* Failed to SLP. */ | |
1930 | /* Free the allocated memory. */ | |
e403d17e | 1931 | scalar_stmts.release (); |
9771b263 | 1932 | loads.release (); |
b8698a0f | 1933 | |
1ba91a49 | 1934 | /* For basic block SLP, try to break the group up into multiples of the |
97a1a642 | 1935 | vector size. */ |
1ba91a49 AL |
1936 | if (is_a <bb_vec_info> (vinfo) |
1937 | && GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) | |
1938 | && STMT_VINFO_GROUPED_ACCESS (vinfo_for_stmt (stmt))) | |
1939 | { | |
1940 | /* We consider breaking the group only on VF boundaries from the existing | |
1941 | start. */ | |
1942 | for (i = 0; i < group_size; i++) | |
1943 | if (!matches[i]) break; | |
1944 | ||
97a1a642 | 1945 | if (i >= nunits && i < group_size) |
1ba91a49 AL |
1946 | { |
1947 | /* Split into two groups at the first vector boundary before i. */ | |
97a1a642 RB |
1948 | gcc_assert ((nunits & (nunits - 1)) == 0); |
1949 | unsigned group1_size = i & ~(nunits - 1); | |
1ba91a49 AL |
1950 | |
1951 | gimple *rest = vect_split_slp_store_group (stmt, group1_size); | |
1952 | bool res = vect_analyze_slp_instance (vinfo, stmt, max_tree_size); | |
1953 | /* If the first non-match was in the middle of a vector, | |
1954 | skip the rest of that vector. */ | |
1955 | if (group1_size < i) | |
1956 | { | |
97a1a642 | 1957 | i = group1_size + nunits; |
1ba91a49 | 1958 | if (i < group_size) |
97a1a642 | 1959 | rest = vect_split_slp_store_group (rest, nunits); |
1ba91a49 AL |
1960 | } |
1961 | if (i < group_size) | |
1962 | res |= vect_analyze_slp_instance (vinfo, rest, max_tree_size); | |
1963 | return res; | |
1964 | } | |
1965 | /* Even though the first vector did not all match, we might be able to SLP | |
1966 | (some) of the remainder. FORNOW ignore this possibility. */ | |
1967 | } | |
1968 | ||
a70d6342 | 1969 | return false; |
ebfd146a IR |
1970 | } |
1971 | ||
1972 | ||
ff802fa1 | 1973 | /* Check if there are stmts in the loop can be vectorized using SLP. Build SLP |
ebfd146a IR |
1974 | trees of packed scalar stmts if SLP is possible. */ |
1975 | ||
1976 | bool | |
310213d4 | 1977 | vect_analyze_slp (vec_info *vinfo, unsigned max_tree_size) |
ebfd146a IR |
1978 | { |
1979 | unsigned int i; | |
355fe088 | 1980 | gimple *first_element; |
a70d6342 | 1981 | bool ok = false; |
ebfd146a | 1982 | |
73fbfcad | 1983 | if (dump_enabled_p ()) |
e645e942 | 1984 | dump_printf_loc (MSG_NOTE, vect_location, "=== vect_analyze_slp ===\n"); |
ebfd146a | 1985 | |
0d0293ac | 1986 | /* Find SLP sequences starting from groups of grouped stores. */ |
310213d4 RB |
1987 | FOR_EACH_VEC_ELT (vinfo->grouped_stores, i, first_element) |
1988 | if (vect_analyze_slp_instance (vinfo, first_element, max_tree_size)) | |
a70d6342 | 1989 | ok = true; |
ebfd146a | 1990 | |
310213d4 | 1991 | if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo)) |
b010117a | 1992 | { |
310213d4 RB |
1993 | if (loop_vinfo->reduction_chains.length () > 0) |
1994 | { | |
1995 | /* Find SLP sequences starting from reduction chains. */ | |
1996 | FOR_EACH_VEC_ELT (loop_vinfo->reduction_chains, i, first_element) | |
1997 | if (vect_analyze_slp_instance (vinfo, first_element, | |
1998 | max_tree_size)) | |
1999 | ok = true; | |
2000 | else | |
2001 | return false; | |
b010117a | 2002 | |
310213d4 RB |
2003 | /* Don't try to vectorize SLP reductions if reduction chain was |
2004 | detected. */ | |
2005 | return ok; | |
2006 | } | |
b010117a | 2007 | |
310213d4 RB |
2008 | /* Find SLP sequences starting from groups of reductions. */ |
2009 | if (loop_vinfo->reductions.length () > 1 | |
2010 | && vect_analyze_slp_instance (vinfo, loop_vinfo->reductions[0], | |
2011 | max_tree_size)) | |
2012 | ok = true; | |
2013 | } | |
b5aeb3bb | 2014 | |
ebfd146a IR |
2015 | return true; |
2016 | } | |
2017 | ||
2018 | ||
2019 | /* For each possible SLP instance decide whether to SLP it and calculate overall | |
437f4a00 IR |
2020 | unrolling factor needed to SLP the loop. Return TRUE if decided to SLP at |
2021 | least one instance. */ | |
ebfd146a | 2022 | |
437f4a00 | 2023 | bool |
ebfd146a IR |
2024 | vect_make_slp_decision (loop_vec_info loop_vinfo) |
2025 | { | |
2026 | unsigned int i, unrolling_factor = 1; | |
9771b263 | 2027 | vec<slp_instance> slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo); |
ebfd146a IR |
2028 | slp_instance instance; |
2029 | int decided_to_slp = 0; | |
2030 | ||
73fbfcad | 2031 | if (dump_enabled_p ()) |
e645e942 TJ |
2032 | dump_printf_loc (MSG_NOTE, vect_location, "=== vect_make_slp_decision ===" |
2033 | "\n"); | |
ebfd146a | 2034 | |
9771b263 | 2035 | FOR_EACH_VEC_ELT (slp_instances, i, instance) |
ebfd146a IR |
2036 | { |
2037 | /* FORNOW: SLP if you can. */ | |
2038 | if (unrolling_factor < SLP_INSTANCE_UNROLLING_FACTOR (instance)) | |
2039 | unrolling_factor = SLP_INSTANCE_UNROLLING_FACTOR (instance); | |
2040 | ||
ff802fa1 | 2041 | /* Mark all the stmts that belong to INSTANCE as PURE_SLP stmts. Later we |
b8698a0f | 2042 | call vect_detect_hybrid_slp () to find stmts that need hybrid SLP and |
ff802fa1 | 2043 | loop-based vectorization. Such stmts will be marked as HYBRID. */ |
ebfd146a IR |
2044 | vect_mark_slp_stmts (SLP_INSTANCE_TREE (instance), pure_slp, -1); |
2045 | decided_to_slp++; | |
2046 | } | |
2047 | ||
2048 | LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo) = unrolling_factor; | |
2049 | ||
73fbfcad | 2050 | if (decided_to_slp && dump_enabled_p ()) |
ccb3ad87 | 2051 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 2052 | "Decided to SLP %d instances. Unrolling factor %d\n", |
78c60e3d | 2053 | decided_to_slp, unrolling_factor); |
437f4a00 IR |
2054 | |
2055 | return (decided_to_slp > 0); | |
ebfd146a IR |
2056 | } |
2057 | ||
2058 | ||
2059 | /* Find stmts that must be both vectorized and SLPed (since they feed stmts that | |
ff802fa1 | 2060 | can't be SLPed) in the tree rooted at NODE. Mark such stmts as HYBRID. */ |
ebfd146a IR |
2061 | |
2062 | static void | |
642fce57 | 2063 | vect_detect_hybrid_slp_stmts (slp_tree node, unsigned i, slp_vect_type stype) |
ebfd146a | 2064 | { |
355fe088 | 2065 | gimple *stmt = SLP_TREE_SCALAR_STMTS (node)[i]; |
ebfd146a | 2066 | imm_use_iterator imm_iter; |
355fe088 | 2067 | gimple *use_stmt; |
642fce57 | 2068 | stmt_vec_info use_vinfo, stmt_vinfo = vinfo_for_stmt (stmt); |
d755c7ef | 2069 | slp_tree child; |
f2c74cc4 | 2070 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
642fce57 RB |
2071 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
2072 | int j; | |
2073 | ||
2074 | /* Propagate hybrid down the SLP tree. */ | |
2075 | if (stype == hybrid) | |
2076 | ; | |
2077 | else if (HYBRID_SLP_STMT (stmt_vinfo)) | |
2078 | stype = hybrid; | |
2079 | else | |
2080 | { | |
2081 | /* Check if a pure SLP stmt has uses in non-SLP stmts. */ | |
2082 | gcc_checking_assert (PURE_SLP_STMT (stmt_vinfo)); | |
2935d994 RB |
2083 | /* If we get a pattern stmt here we have to use the LHS of the |
2084 | original stmt for immediate uses. */ | |
2085 | if (! STMT_VINFO_IN_PATTERN_P (stmt_vinfo) | |
2086 | && STMT_VINFO_RELATED_STMT (stmt_vinfo)) | |
29764870 | 2087 | stmt = STMT_VINFO_RELATED_STMT (stmt_vinfo); |
642fce57 RB |
2088 | if (TREE_CODE (gimple_op (stmt, 0)) == SSA_NAME) |
2089 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, gimple_op (stmt, 0)) | |
29764870 RB |
2090 | { |
2091 | if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) | |
2092 | continue; | |
2093 | use_vinfo = vinfo_for_stmt (use_stmt); | |
2094 | if (STMT_VINFO_IN_PATTERN_P (use_vinfo) | |
2095 | && STMT_VINFO_RELATED_STMT (use_vinfo)) | |
2096 | use_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (use_vinfo)); | |
2097 | if (!STMT_SLP_TYPE (use_vinfo) | |
2098 | && (STMT_VINFO_RELEVANT (use_vinfo) | |
2099 | || VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (use_vinfo))) | |
2100 | && !(gimple_code (use_stmt) == GIMPLE_PHI | |
2101 | && STMT_VINFO_DEF_TYPE (use_vinfo) == vect_reduction_def)) | |
502f0263 RB |
2102 | { |
2103 | if (dump_enabled_p ()) | |
2104 | { | |
2105 | dump_printf_loc (MSG_NOTE, vect_location, "use of SLP " | |
2106 | "def in non-SLP stmt: "); | |
2107 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, use_stmt, 0); | |
2108 | } | |
2109 | stype = hybrid; | |
2110 | } | |
29764870 | 2111 | } |
642fce57 | 2112 | } |
ebfd146a | 2113 | |
502f0263 RB |
2114 | if (stype == hybrid |
2115 | && !HYBRID_SLP_STMT (stmt_vinfo)) | |
b1af7da6 RB |
2116 | { |
2117 | if (dump_enabled_p ()) | |
2118 | { | |
2119 | dump_printf_loc (MSG_NOTE, vect_location, "marking hybrid: "); | |
2120 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
2121 | } | |
2122 | STMT_SLP_TYPE (stmt_vinfo) = hybrid; | |
2123 | } | |
ebfd146a | 2124 | |
642fce57 | 2125 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), j, child) |
603cca93 | 2126 | if (SLP_TREE_DEF_TYPE (child) != vect_external_def) |
90dd6e3d | 2127 | vect_detect_hybrid_slp_stmts (child, i, stype); |
642fce57 | 2128 | } |
f2c74cc4 | 2129 | |
642fce57 | 2130 | /* Helpers for vect_detect_hybrid_slp walking pattern stmt uses. */ |
ebfd146a | 2131 | |
642fce57 RB |
2132 | static tree |
2133 | vect_detect_hybrid_slp_1 (tree *tp, int *, void *data) | |
2134 | { | |
2135 | walk_stmt_info *wi = (walk_stmt_info *)data; | |
2136 | struct loop *loopp = (struct loop *)wi->info; | |
2137 | ||
2138 | if (wi->is_lhs) | |
2139 | return NULL_TREE; | |
2140 | ||
2141 | if (TREE_CODE (*tp) == SSA_NAME | |
2142 | && !SSA_NAME_IS_DEFAULT_DEF (*tp)) | |
2143 | { | |
355fe088 | 2144 | gimple *def_stmt = SSA_NAME_DEF_STMT (*tp); |
642fce57 RB |
2145 | if (flow_bb_inside_loop_p (loopp, gimple_bb (def_stmt)) |
2146 | && PURE_SLP_STMT (vinfo_for_stmt (def_stmt))) | |
b1af7da6 RB |
2147 | { |
2148 | if (dump_enabled_p ()) | |
2149 | { | |
2150 | dump_printf_loc (MSG_NOTE, vect_location, "marking hybrid: "); | |
2151 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
2152 | } | |
2153 | STMT_SLP_TYPE (vinfo_for_stmt (def_stmt)) = hybrid; | |
2154 | } | |
642fce57 RB |
2155 | } |
2156 | ||
2157 | return NULL_TREE; | |
ebfd146a IR |
2158 | } |
2159 | ||
642fce57 RB |
2160 | static tree |
2161 | vect_detect_hybrid_slp_2 (gimple_stmt_iterator *gsi, bool *handled, | |
2162 | walk_stmt_info *) | |
2163 | { | |
2164 | /* If the stmt is in a SLP instance then this isn't a reason | |
2165 | to mark use definitions in other SLP instances as hybrid. */ | |
2166 | if (STMT_SLP_TYPE (vinfo_for_stmt (gsi_stmt (*gsi))) != loop_vect) | |
2167 | *handled = true; | |
2168 | return NULL_TREE; | |
2169 | } | |
ebfd146a IR |
2170 | |
2171 | /* Find stmts that must be both vectorized and SLPed. */ | |
2172 | ||
2173 | void | |
2174 | vect_detect_hybrid_slp (loop_vec_info loop_vinfo) | |
2175 | { | |
2176 | unsigned int i; | |
9771b263 | 2177 | vec<slp_instance> slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo); |
ebfd146a IR |
2178 | slp_instance instance; |
2179 | ||
73fbfcad | 2180 | if (dump_enabled_p ()) |
e645e942 TJ |
2181 | dump_printf_loc (MSG_NOTE, vect_location, "=== vect_detect_hybrid_slp ===" |
2182 | "\n"); | |
ebfd146a | 2183 | |
642fce57 RB |
2184 | /* First walk all pattern stmt in the loop and mark defs of uses as |
2185 | hybrid because immediate uses in them are not recorded. */ | |
2186 | for (i = 0; i < LOOP_VINFO_LOOP (loop_vinfo)->num_nodes; ++i) | |
2187 | { | |
2188 | basic_block bb = LOOP_VINFO_BBS (loop_vinfo)[i]; | |
2189 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi); | |
2190 | gsi_next (&gsi)) | |
2191 | { | |
355fe088 | 2192 | gimple *stmt = gsi_stmt (gsi); |
642fce57 RB |
2193 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
2194 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
2195 | { | |
2196 | walk_stmt_info wi; | |
2197 | memset (&wi, 0, sizeof (wi)); | |
2198 | wi.info = LOOP_VINFO_LOOP (loop_vinfo); | |
2199 | gimple_stmt_iterator gsi2 | |
2200 | = gsi_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
2201 | walk_gimple_stmt (&gsi2, vect_detect_hybrid_slp_2, | |
2202 | vect_detect_hybrid_slp_1, &wi); | |
2203 | walk_gimple_seq (STMT_VINFO_PATTERN_DEF_SEQ (stmt_info), | |
2204 | vect_detect_hybrid_slp_2, | |
2205 | vect_detect_hybrid_slp_1, &wi); | |
2206 | } | |
2207 | } | |
2208 | } | |
2209 | ||
2210 | /* Then walk the SLP instance trees marking stmts with uses in | |
2211 | non-SLP stmts as hybrid, also propagating hybrid down the | |
2212 | SLP tree, collecting the above info on-the-fly. */ | |
9771b263 | 2213 | FOR_EACH_VEC_ELT (slp_instances, i, instance) |
642fce57 RB |
2214 | { |
2215 | for (unsigned i = 0; i < SLP_INSTANCE_GROUP_SIZE (instance); ++i) | |
2216 | vect_detect_hybrid_slp_stmts (SLP_INSTANCE_TREE (instance), | |
2217 | i, pure_slp); | |
2218 | } | |
ebfd146a IR |
2219 | } |
2220 | ||
a70d6342 IR |
2221 | |
2222 | /* Create and initialize a new bb_vec_info struct for BB, as well as | |
2223 | stmt_vec_info structs for all the stmts in it. */ | |
b8698a0f | 2224 | |
a70d6342 | 2225 | static bb_vec_info |
61d371eb RB |
2226 | new_bb_vec_info (gimple_stmt_iterator region_begin, |
2227 | gimple_stmt_iterator region_end) | |
a70d6342 | 2228 | { |
61d371eb | 2229 | basic_block bb = gsi_bb (region_begin); |
a70d6342 IR |
2230 | bb_vec_info res = NULL; |
2231 | gimple_stmt_iterator gsi; | |
2232 | ||
2233 | res = (bb_vec_info) xcalloc (1, sizeof (struct _bb_vec_info)); | |
310213d4 | 2234 | res->kind = vec_info::bb; |
a70d6342 | 2235 | BB_VINFO_BB (res) = bb; |
61d371eb RB |
2236 | res->region_begin = region_begin; |
2237 | res->region_end = region_end; | |
a70d6342 | 2238 | |
61d371eb RB |
2239 | for (gsi = region_begin; gsi_stmt (gsi) != gsi_stmt (region_end); |
2240 | gsi_next (&gsi)) | |
a70d6342 | 2241 | { |
355fe088 | 2242 | gimple *stmt = gsi_stmt (gsi); |
a70d6342 | 2243 | gimple_set_uid (stmt, 0); |
310213d4 | 2244 | set_vinfo_for_stmt (stmt, new_stmt_vec_info (stmt, res)); |
a70d6342 IR |
2245 | } |
2246 | ||
9771b263 DN |
2247 | BB_VINFO_GROUPED_STORES (res).create (10); |
2248 | BB_VINFO_SLP_INSTANCES (res).create (2); | |
c3e7ee41 | 2249 | BB_VINFO_TARGET_COST_DATA (res) = init_cost (NULL); |
a70d6342 IR |
2250 | |
2251 | bb->aux = res; | |
2252 | return res; | |
2253 | } | |
2254 | ||
2255 | ||
2256 | /* Free BB_VINFO struct, as well as all the stmt_vec_info structs of all the | |
2257 | stmts in the basic block. */ | |
2258 | ||
2259 | static void | |
2260 | destroy_bb_vec_info (bb_vec_info bb_vinfo) | |
2261 | { | |
c7e62a26 | 2262 | slp_instance instance; |
c7e62a26 | 2263 | unsigned i; |
a70d6342 IR |
2264 | |
2265 | if (!bb_vinfo) | |
2266 | return; | |
2267 | ||
78810bd3 RB |
2268 | vect_destroy_datarefs (bb_vinfo); |
2269 | free_dependence_relations (BB_VINFO_DDRS (bb_vinfo)); | |
2270 | BB_VINFO_GROUPED_STORES (bb_vinfo).release (); | |
2271 | FOR_EACH_VEC_ELT (BB_VINFO_SLP_INSTANCES (bb_vinfo), i, instance) | |
2272 | vect_free_slp_instance (instance); | |
2273 | BB_VINFO_SLP_INSTANCES (bb_vinfo).release (); | |
2274 | destroy_cost_data (BB_VINFO_TARGET_COST_DATA (bb_vinfo)); | |
a70d6342 | 2275 | |
78810bd3 | 2276 | for (gimple_stmt_iterator si = bb_vinfo->region_begin; |
61d371eb | 2277 | gsi_stmt (si) != gsi_stmt (bb_vinfo->region_end); gsi_next (&si)) |
a70d6342 | 2278 | { |
355fe088 | 2279 | gimple *stmt = gsi_stmt (si); |
a70d6342 IR |
2280 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
2281 | ||
2282 | if (stmt_info) | |
2283 | /* Free stmt_vec_info. */ | |
2284 | free_stmt_vec_info (stmt); | |
61d371eb RB |
2285 | |
2286 | /* Reset region marker. */ | |
2287 | gimple_set_uid (stmt, -1); | |
a70d6342 IR |
2288 | } |
2289 | ||
78810bd3 | 2290 | BB_VINFO_BB (bb_vinfo)->aux = NULL; |
a70d6342 | 2291 | free (bb_vinfo); |
a70d6342 IR |
2292 | } |
2293 | ||
2294 | ||
2295 | /* Analyze statements contained in SLP tree node after recursively analyzing | |
2296 | the subtree. Return TRUE if the operations are supported. */ | |
2297 | ||
2298 | static bool | |
a12e42fc | 2299 | vect_slp_analyze_node_operations (slp_tree node) |
a70d6342 IR |
2300 | { |
2301 | bool dummy; | |
603cca93 | 2302 | int i, j; |
355fe088 | 2303 | gimple *stmt; |
d755c7ef | 2304 | slp_tree child; |
a70d6342 | 2305 | |
603cca93 | 2306 | if (SLP_TREE_DEF_TYPE (node) != vect_internal_def) |
a70d6342 IR |
2307 | return true; |
2308 | ||
9771b263 | 2309 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) |
a12e42fc | 2310 | if (!vect_slp_analyze_node_operations (child)) |
d092494c | 2311 | return false; |
a70d6342 | 2312 | |
603cca93 | 2313 | bool res = true; |
9771b263 | 2314 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt) |
a70d6342 IR |
2315 | { |
2316 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2317 | gcc_assert (stmt_info); | |
a12e42fc | 2318 | gcc_assert (STMT_SLP_TYPE (stmt_info) != loop_vect); |
a70d6342 | 2319 | |
6379dfb5 RB |
2320 | /* Push SLP node def-type to stmt operands. */ |
2321 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), j, child) | |
2322 | if (SLP_TREE_DEF_TYPE (child) != vect_internal_def) | |
2323 | STMT_VINFO_DEF_TYPE (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (child)[i])) | |
2324 | = SLP_TREE_DEF_TYPE (child); | |
2325 | res = vect_analyze_stmt (stmt, &dummy, node); | |
2326 | /* Restore def-types. */ | |
2327 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), j, child) | |
2328 | if (SLP_TREE_DEF_TYPE (child) != vect_internal_def) | |
2329 | STMT_VINFO_DEF_TYPE (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (child)[i])) | |
2330 | = vect_internal_def; | |
2331 | if (! res) | |
2332 | break; | |
a70d6342 IR |
2333 | } |
2334 | ||
603cca93 | 2335 | return res; |
a70d6342 IR |
2336 | } |
2337 | ||
2338 | ||
ff802fa1 | 2339 | /* Analyze statements in SLP instances of the basic block. Return TRUE if the |
a70d6342 IR |
2340 | operations are supported. */ |
2341 | ||
a12e42fc | 2342 | bool |
1a4b99c1 | 2343 | vect_slp_analyze_operations (vec<slp_instance> slp_instances, void *data) |
a70d6342 | 2344 | { |
a70d6342 IR |
2345 | slp_instance instance; |
2346 | int i; | |
2347 | ||
a12e42fc RB |
2348 | if (dump_enabled_p ()) |
2349 | dump_printf_loc (MSG_NOTE, vect_location, | |
2350 | "=== vect_slp_analyze_operations ===\n"); | |
2351 | ||
9771b263 | 2352 | for (i = 0; slp_instances.iterate (i, &instance); ) |
a70d6342 | 2353 | { |
a12e42fc | 2354 | if (!vect_slp_analyze_node_operations (SLP_INSTANCE_TREE (instance))) |
a70d6342 | 2355 | { |
a12e42fc RB |
2356 | dump_printf_loc (MSG_NOTE, vect_location, |
2357 | "removing SLP instance operations starting from: "); | |
2358 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, | |
2359 | SLP_TREE_SCALAR_STMTS | |
2360 | (SLP_INSTANCE_TREE (instance))[0], 0); | |
2361 | vect_free_slp_instance (instance); | |
9771b263 | 2362 | slp_instances.ordered_remove (i); |
a70d6342 IR |
2363 | } |
2364 | else | |
1a4b99c1 RB |
2365 | { |
2366 | /* Compute the costs of the SLP instance. */ | |
2367 | vect_analyze_slp_cost (instance, data); | |
2368 | i++; | |
2369 | } | |
b8698a0f L |
2370 | } |
2371 | ||
9771b263 | 2372 | if (!slp_instances.length ()) |
a70d6342 IR |
2373 | return false; |
2374 | ||
2375 | return true; | |
2376 | } | |
2377 | ||
6eddf228 RB |
2378 | |
2379 | /* Compute the scalar cost of the SLP node NODE and its children | |
2380 | and return it. Do not account defs that are marked in LIFE and | |
2381 | update LIFE according to uses of NODE. */ | |
2382 | ||
2383 | static unsigned | |
292cba13 | 2384 | vect_bb_slp_scalar_cost (basic_block bb, |
ff4c81cc | 2385 | slp_tree node, vec<bool, va_heap> *life) |
6eddf228 RB |
2386 | { |
2387 | unsigned scalar_cost = 0; | |
2388 | unsigned i; | |
355fe088 | 2389 | gimple *stmt; |
6eddf228 RB |
2390 | slp_tree child; |
2391 | ||
2392 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt) | |
2393 | { | |
2394 | unsigned stmt_cost; | |
2395 | ssa_op_iter op_iter; | |
2396 | def_operand_p def_p; | |
2397 | stmt_vec_info stmt_info; | |
2398 | ||
ff4c81cc | 2399 | if ((*life)[i]) |
6eddf228 RB |
2400 | continue; |
2401 | ||
2402 | /* If there is a non-vectorized use of the defs then the scalar | |
2403 | stmt is kept live in which case we do not account it or any | |
2404 | required defs in the SLP children in the scalar cost. This | |
2405 | way we make the vectorization more costly when compared to | |
2406 | the scalar cost. */ | |
2407 | FOR_EACH_SSA_DEF_OPERAND (def_p, stmt, op_iter, SSA_OP_DEF) | |
2408 | { | |
2409 | imm_use_iterator use_iter; | |
355fe088 | 2410 | gimple *use_stmt; |
6eddf228 | 2411 | FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, DEF_FROM_PTR (def_p)) |
f30a0ba5 | 2412 | if (!is_gimple_debug (use_stmt) |
61d371eb RB |
2413 | && (! vect_stmt_in_region_p (vinfo_for_stmt (stmt)->vinfo, |
2414 | use_stmt) | |
603cca93 | 2415 | || ! PURE_SLP_STMT (vinfo_for_stmt (use_stmt)))) |
6eddf228 | 2416 | { |
ff4c81cc | 2417 | (*life)[i] = true; |
6eddf228 RB |
2418 | BREAK_FROM_IMM_USE_STMT (use_iter); |
2419 | } | |
2420 | } | |
ff4c81cc | 2421 | if ((*life)[i]) |
6eddf228 RB |
2422 | continue; |
2423 | ||
b555a2e4 RB |
2424 | /* Count scalar stmts only once. */ |
2425 | if (gimple_visited_p (stmt)) | |
2426 | continue; | |
2427 | gimple_set_visited (stmt, true); | |
2428 | ||
6eddf228 RB |
2429 | stmt_info = vinfo_for_stmt (stmt); |
2430 | if (STMT_VINFO_DATA_REF (stmt_info)) | |
2431 | { | |
2432 | if (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info))) | |
2433 | stmt_cost = vect_get_stmt_cost (scalar_load); | |
2434 | else | |
2435 | stmt_cost = vect_get_stmt_cost (scalar_store); | |
2436 | } | |
2437 | else | |
2438 | stmt_cost = vect_get_stmt_cost (scalar_stmt); | |
2439 | ||
2440 | scalar_cost += stmt_cost; | |
2441 | } | |
2442 | ||
2443 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) | |
603cca93 | 2444 | if (SLP_TREE_DEF_TYPE (child) == vect_internal_def) |
90dd6e3d | 2445 | scalar_cost += vect_bb_slp_scalar_cost (bb, child, life); |
6eddf228 RB |
2446 | |
2447 | return scalar_cost; | |
2448 | } | |
2449 | ||
69f11a13 IR |
2450 | /* Check if vectorization of the basic block is profitable. */ |
2451 | ||
2452 | static bool | |
2453 | vect_bb_vectorization_profitable_p (bb_vec_info bb_vinfo) | |
2454 | { | |
9771b263 | 2455 | vec<slp_instance> slp_instances = BB_VINFO_SLP_INSTANCES (bb_vinfo); |
69f11a13 | 2456 | slp_instance instance; |
1a4b99c1 | 2457 | int i; |
c3e7ee41 | 2458 | unsigned int vec_inside_cost = 0, vec_outside_cost = 0, scalar_cost = 0; |
92345349 | 2459 | unsigned int vec_prologue_cost = 0, vec_epilogue_cost = 0; |
69f11a13 IR |
2460 | |
2461 | /* Calculate scalar cost. */ | |
6eddf228 | 2462 | FOR_EACH_VEC_ELT (slp_instances, i, instance) |
69f11a13 | 2463 | { |
00f96dc9 | 2464 | auto_vec<bool, 20> life; |
ff4c81cc | 2465 | life.safe_grow_cleared (SLP_INSTANCE_GROUP_SIZE (instance)); |
292cba13 RB |
2466 | scalar_cost += vect_bb_slp_scalar_cost (BB_VINFO_BB (bb_vinfo), |
2467 | SLP_INSTANCE_TREE (instance), | |
ff4c81cc | 2468 | &life); |
69f11a13 IR |
2469 | } |
2470 | ||
b555a2e4 RB |
2471 | /* Unset visited flag. */ |
2472 | for (gimple_stmt_iterator gsi = bb_vinfo->region_begin; | |
2473 | gsi_stmt (gsi) != gsi_stmt (bb_vinfo->region_end); gsi_next (&gsi)) | |
2474 | gimple_set_visited (gsi_stmt (gsi), false); | |
2475 | ||
c3e7ee41 | 2476 | /* Complete the target-specific cost calculation. */ |
92345349 BS |
2477 | finish_cost (BB_VINFO_TARGET_COST_DATA (bb_vinfo), &vec_prologue_cost, |
2478 | &vec_inside_cost, &vec_epilogue_cost); | |
2479 | ||
2480 | vec_outside_cost = vec_prologue_cost + vec_epilogue_cost; | |
c3e7ee41 | 2481 | |
73fbfcad | 2482 | if (dump_enabled_p ()) |
69f11a13 | 2483 | { |
78c60e3d SS |
2484 | dump_printf_loc (MSG_NOTE, vect_location, "Cost model analysis: \n"); |
2485 | dump_printf (MSG_NOTE, " Vector inside of basic block cost: %d\n", | |
2486 | vec_inside_cost); | |
2487 | dump_printf (MSG_NOTE, " Vector prologue cost: %d\n", vec_prologue_cost); | |
2488 | dump_printf (MSG_NOTE, " Vector epilogue cost: %d\n", vec_epilogue_cost); | |
e645e942 | 2489 | dump_printf (MSG_NOTE, " Scalar cost of basic block: %d\n", scalar_cost); |
69f11a13 IR |
2490 | } |
2491 | ||
a6524bba RB |
2492 | /* Vectorization is profitable if its cost is more than the cost of scalar |
2493 | version. Note that we err on the vector side for equal cost because | |
2494 | the cost estimate is otherwise quite pessimistic (constant uses are | |
2495 | free on the scalar side but cost a load on the vector side for | |
2496 | example). */ | |
2497 | if (vec_outside_cost + vec_inside_cost > scalar_cost) | |
69f11a13 IR |
2498 | return false; |
2499 | ||
2500 | return true; | |
2501 | } | |
2502 | ||
a5b50aa1 RB |
2503 | /* Check if the basic block can be vectorized. Returns a bb_vec_info |
2504 | if so and sets fatal to true if failure is independent of | |
2505 | current_vector_size. */ | |
a70d6342 | 2506 | |
8e19f5a1 | 2507 | static bb_vec_info |
61d371eb RB |
2508 | vect_slp_analyze_bb_1 (gimple_stmt_iterator region_begin, |
2509 | gimple_stmt_iterator region_end, | |
a5b50aa1 RB |
2510 | vec<data_reference_p> datarefs, int n_stmts, |
2511 | bool &fatal) | |
a70d6342 IR |
2512 | { |
2513 | bb_vec_info bb_vinfo; | |
a70d6342 | 2514 | slp_instance instance; |
8e19f5a1 | 2515 | int i; |
777e1f09 | 2516 | int min_vf = 2; |
e4a707c4 | 2517 | |
a5b50aa1 RB |
2518 | /* The first group of checks is independent of the vector size. */ |
2519 | fatal = true; | |
2520 | ||
61d371eb RB |
2521 | if (n_stmts > PARAM_VALUE (PARAM_SLP_MAX_INSNS_IN_BB)) |
2522 | { | |
2523 | if (dump_enabled_p ()) | |
2524 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2525 | "not vectorized: too many instructions in " | |
2526 | "basic block.\n"); | |
2527 | free_data_refs (datarefs); | |
2528 | return NULL; | |
2529 | } | |
2530 | ||
2531 | bb_vinfo = new_bb_vec_info (region_begin, region_end); | |
a70d6342 IR |
2532 | if (!bb_vinfo) |
2533 | return NULL; | |
2534 | ||
61d371eb | 2535 | BB_VINFO_DATAREFS (bb_vinfo) = datarefs; |
428db0ba RB |
2536 | |
2537 | /* Analyze the data references. */ | |
2538 | ||
2539 | if (!vect_analyze_data_refs (bb_vinfo, &min_vf)) | |
a70d6342 | 2540 | { |
73fbfcad | 2541 | if (dump_enabled_p ()) |
78c60e3d SS |
2542 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2543 | "not vectorized: unhandled data-ref in basic " | |
2544 | "block.\n"); | |
b8698a0f | 2545 | |
a70d6342 IR |
2546 | destroy_bb_vec_info (bb_vinfo); |
2547 | return NULL; | |
2548 | } | |
2549 | ||
fcac74a1 | 2550 | if (BB_VINFO_DATAREFS (bb_vinfo).length () < 2) |
a70d6342 | 2551 | { |
73fbfcad | 2552 | if (dump_enabled_p ()) |
78c60e3d SS |
2553 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2554 | "not vectorized: not enough data-refs in " | |
2555 | "basic block.\n"); | |
a70d6342 IR |
2556 | |
2557 | destroy_bb_vec_info (bb_vinfo); | |
2558 | return NULL; | |
2559 | } | |
2560 | ||
310213d4 | 2561 | if (!vect_analyze_data_ref_accesses (bb_vinfo)) |
5abe1e05 RB |
2562 | { |
2563 | if (dump_enabled_p ()) | |
2564 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2565 | "not vectorized: unhandled data access in " | |
2566 | "basic block.\n"); | |
2567 | ||
2568 | destroy_bb_vec_info (bb_vinfo); | |
2569 | return NULL; | |
2570 | } | |
2571 | ||
a5b50aa1 RB |
2572 | /* If there are no grouped stores in the region there is no need |
2573 | to continue with pattern recog as vect_analyze_slp will fail | |
2574 | anyway. */ | |
2575 | if (bb_vinfo->grouped_stores.is_empty ()) | |
a70d6342 | 2576 | { |
73fbfcad | 2577 | if (dump_enabled_p ()) |
a5b50aa1 RB |
2578 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2579 | "not vectorized: no grouped stores in " | |
2580 | "basic block.\n"); | |
b8698a0f | 2581 | |
a70d6342 IR |
2582 | destroy_bb_vec_info (bb_vinfo); |
2583 | return NULL; | |
2584 | } | |
b8698a0f | 2585 | |
a5b50aa1 RB |
2586 | /* While the rest of the analysis below depends on it in some way. */ |
2587 | fatal = false; | |
2588 | ||
2589 | vect_pattern_recog (bb_vinfo); | |
2590 | ||
a70d6342 IR |
2591 | /* Check the SLP opportunities in the basic block, analyze and build SLP |
2592 | trees. */ | |
310213d4 | 2593 | if (!vect_analyze_slp (bb_vinfo, n_stmts)) |
a70d6342 | 2594 | { |
73fbfcad | 2595 | if (dump_enabled_p ()) |
effb52da RB |
2596 | { |
2597 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2598 | "Failed to SLP the basic block.\n"); | |
2599 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2600 | "not vectorized: failed to find SLP opportunities " | |
2601 | "in basic block.\n"); | |
2602 | } | |
a70d6342 IR |
2603 | |
2604 | destroy_bb_vec_info (bb_vinfo); | |
2605 | return NULL; | |
2606 | } | |
b8698a0f | 2607 | |
c2a12ca0 RB |
2608 | /* Analyze and verify the alignment of data references and the |
2609 | dependence in the SLP instances. */ | |
a5b50aa1 RB |
2610 | for (i = 0; BB_VINFO_SLP_INSTANCES (bb_vinfo).iterate (i, &instance); ) |
2611 | { | |
c2a12ca0 RB |
2612 | if (! vect_slp_analyze_and_verify_instance_alignment (instance) |
2613 | || ! vect_slp_analyze_instance_dependence (instance)) | |
a5b50aa1 RB |
2614 | { |
2615 | dump_printf_loc (MSG_NOTE, vect_location, | |
2616 | "removing SLP instance operations starting from: "); | |
2617 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, | |
2618 | SLP_TREE_SCALAR_STMTS | |
2619 | (SLP_INSTANCE_TREE (instance))[0], 0); | |
2620 | vect_free_slp_instance (instance); | |
2621 | BB_VINFO_SLP_INSTANCES (bb_vinfo).ordered_remove (i); | |
2622 | continue; | |
2623 | } | |
c2a12ca0 RB |
2624 | |
2625 | /* Mark all the statements that we want to vectorize as pure SLP and | |
2626 | relevant. */ | |
2627 | vect_mark_slp_stmts (SLP_INSTANCE_TREE (instance), pure_slp, -1); | |
2628 | vect_mark_slp_stmts_relevant (SLP_INSTANCE_TREE (instance)); | |
2629 | ||
a5b50aa1 RB |
2630 | i++; |
2631 | } | |
a5b50aa1 RB |
2632 | if (! BB_VINFO_SLP_INSTANCES (bb_vinfo).length ()) |
2633 | { | |
2634 | destroy_bb_vec_info (bb_vinfo); | |
2635 | return NULL; | |
2636 | } | |
2637 | ||
1a4b99c1 RB |
2638 | if (!vect_slp_analyze_operations (BB_VINFO_SLP_INSTANCES (bb_vinfo), |
2639 | BB_VINFO_TARGET_COST_DATA (bb_vinfo))) | |
a70d6342 | 2640 | { |
73fbfcad | 2641 | if (dump_enabled_p ()) |
e645e942 | 2642 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
78c60e3d | 2643 | "not vectorized: bad operation in basic block.\n"); |
a70d6342 IR |
2644 | |
2645 | destroy_bb_vec_info (bb_vinfo); | |
2646 | return NULL; | |
2647 | } | |
2648 | ||
69f11a13 | 2649 | /* Cost model: check if the vectorization is worthwhile. */ |
8b5e1202 | 2650 | if (!unlimited_cost_model (NULL) |
69f11a13 IR |
2651 | && !vect_bb_vectorization_profitable_p (bb_vinfo)) |
2652 | { | |
73fbfcad | 2653 | if (dump_enabled_p ()) |
78c60e3d SS |
2654 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2655 | "not vectorized: vectorization is not " | |
2656 | "profitable.\n"); | |
69f11a13 IR |
2657 | |
2658 | destroy_bb_vec_info (bb_vinfo); | |
2659 | return NULL; | |
2660 | } | |
2661 | ||
73fbfcad | 2662 | if (dump_enabled_p ()) |
78c60e3d SS |
2663 | dump_printf_loc (MSG_NOTE, vect_location, |
2664 | "Basic block will be vectorized using SLP\n"); | |
a70d6342 IR |
2665 | |
2666 | return bb_vinfo; | |
2667 | } | |
2668 | ||
2669 | ||
428db0ba RB |
2670 | /* Main entry for the BB vectorizer. Analyze and transform BB, returns |
2671 | true if anything in the basic-block was vectorized. */ | |
2672 | ||
2673 | bool | |
2674 | vect_slp_bb (basic_block bb) | |
8e19f5a1 IR |
2675 | { |
2676 | bb_vec_info bb_vinfo; | |
8e19f5a1 IR |
2677 | gimple_stmt_iterator gsi; |
2678 | unsigned int vector_sizes; | |
61d371eb | 2679 | bool any_vectorized = false; |
8e19f5a1 | 2680 | |
73fbfcad | 2681 | if (dump_enabled_p ()) |
78c60e3d | 2682 | dump_printf_loc (MSG_NOTE, vect_location, "===vect_slp_analyze_bb===\n"); |
8e19f5a1 | 2683 | |
8e19f5a1 IR |
2684 | /* Autodetect first vector size we try. */ |
2685 | current_vector_size = 0; | |
2686 | vector_sizes = targetm.vectorize.autovectorize_vector_sizes (); | |
2687 | ||
61d371eb RB |
2688 | gsi = gsi_start_bb (bb); |
2689 | ||
8e19f5a1 IR |
2690 | while (1) |
2691 | { | |
61d371eb RB |
2692 | if (gsi_end_p (gsi)) |
2693 | break; | |
2694 | ||
2695 | gimple_stmt_iterator region_begin = gsi; | |
2696 | vec<data_reference_p> datarefs = vNULL; | |
2697 | int insns = 0; | |
2698 | ||
2699 | for (; !gsi_end_p (gsi); gsi_next (&gsi)) | |
428db0ba | 2700 | { |
61d371eb RB |
2701 | gimple *stmt = gsi_stmt (gsi); |
2702 | if (is_gimple_debug (stmt)) | |
2703 | continue; | |
2704 | insns++; | |
2705 | ||
2706 | if (gimple_location (stmt) != UNKNOWN_LOCATION) | |
2707 | vect_location = gimple_location (stmt); | |
2708 | ||
2709 | if (!find_data_references_in_stmt (NULL, stmt, &datarefs)) | |
2710 | break; | |
2711 | } | |
2712 | ||
2713 | /* Skip leading unhandled stmts. */ | |
2714 | if (gsi_stmt (region_begin) == gsi_stmt (gsi)) | |
2715 | { | |
2716 | gsi_next (&gsi); | |
2717 | continue; | |
2718 | } | |
428db0ba | 2719 | |
61d371eb RB |
2720 | gimple_stmt_iterator region_end = gsi; |
2721 | ||
2722 | bool vectorized = false; | |
a5b50aa1 | 2723 | bool fatal = false; |
61d371eb | 2724 | bb_vinfo = vect_slp_analyze_bb_1 (region_begin, region_end, |
a5b50aa1 | 2725 | datarefs, insns, fatal); |
61d371eb RB |
2726 | if (bb_vinfo |
2727 | && dbg_cnt (vect_slp)) | |
2728 | { | |
428db0ba | 2729 | if (dump_enabled_p ()) |
61d371eb | 2730 | dump_printf_loc (MSG_NOTE, vect_location, "SLPing BB part\n"); |
428db0ba RB |
2731 | |
2732 | vect_schedule_slp (bb_vinfo); | |
2733 | ||
2734 | if (dump_enabled_p ()) | |
2735 | dump_printf_loc (MSG_NOTE, vect_location, | |
61d371eb | 2736 | "basic block part vectorized\n"); |
428db0ba RB |
2737 | |
2738 | destroy_bb_vec_info (bb_vinfo); | |
2739 | ||
61d371eb | 2740 | vectorized = true; |
428db0ba | 2741 | } |
61d371eb RB |
2742 | else |
2743 | destroy_bb_vec_info (bb_vinfo); | |
8e19f5a1 | 2744 | |
61d371eb | 2745 | any_vectorized |= vectorized; |
8e19f5a1 IR |
2746 | |
2747 | vector_sizes &= ~current_vector_size; | |
61d371eb RB |
2748 | if (vectorized |
2749 | || vector_sizes == 0 | |
a5b50aa1 RB |
2750 | || current_vector_size == 0 |
2751 | /* If vect_slp_analyze_bb_1 signaled that analysis for all | |
2752 | vector sizes will fail do not bother iterating. */ | |
2753 | || fatal) | |
61d371eb RB |
2754 | { |
2755 | if (gsi_end_p (region_end)) | |
2756 | break; | |
8e19f5a1 | 2757 | |
61d371eb RB |
2758 | /* Skip the unhandled stmt. */ |
2759 | gsi_next (&gsi); | |
2760 | ||
2761 | /* And reset vector sizes. */ | |
2762 | current_vector_size = 0; | |
2763 | vector_sizes = targetm.vectorize.autovectorize_vector_sizes (); | |
2764 | } | |
2765 | else | |
2766 | { | |
2767 | /* Try the next biggest vector size. */ | |
2768 | current_vector_size = 1 << floor_log2 (vector_sizes); | |
2769 | if (dump_enabled_p ()) | |
2770 | dump_printf_loc (MSG_NOTE, vect_location, | |
2771 | "***** Re-trying analysis with " | |
2772 | "vector size %d\n", current_vector_size); | |
2773 | ||
2774 | /* Start over. */ | |
2775 | gsi = region_begin; | |
2776 | } | |
8e19f5a1 | 2777 | } |
61d371eb RB |
2778 | |
2779 | return any_vectorized; | |
8e19f5a1 IR |
2780 | } |
2781 | ||
2782 | ||
e4af0bc4 IE |
2783 | /* Return 1 if vector type of boolean constant which is OPNUM |
2784 | operand in statement STMT is a boolean vector. */ | |
2785 | ||
2786 | static bool | |
2787 | vect_mask_constant_operand_p (gimple *stmt, int opnum) | |
2788 | { | |
2789 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
2790 | enum tree_code code = gimple_expr_code (stmt); | |
2791 | tree op, vectype; | |
2792 | gimple *def_stmt; | |
2793 | enum vect_def_type dt; | |
2794 | ||
2795 | /* For comparison and COND_EXPR type is chosen depending | |
2796 | on the other comparison operand. */ | |
2797 | if (TREE_CODE_CLASS (code) == tcc_comparison) | |
2798 | { | |
2799 | if (opnum) | |
2800 | op = gimple_assign_rhs1 (stmt); | |
2801 | else | |
2802 | op = gimple_assign_rhs2 (stmt); | |
2803 | ||
2804 | if (!vect_is_simple_use (op, stmt_vinfo->vinfo, &def_stmt, | |
2805 | &dt, &vectype)) | |
2806 | gcc_unreachable (); | |
2807 | ||
2808 | return !vectype || VECTOR_BOOLEAN_TYPE_P (vectype); | |
2809 | } | |
2810 | ||
2811 | if (code == COND_EXPR) | |
2812 | { | |
2813 | tree cond = gimple_assign_rhs1 (stmt); | |
2814 | ||
2815 | if (TREE_CODE (cond) == SSA_NAME) | |
2816 | return false; | |
2817 | ||
2818 | if (opnum) | |
2819 | op = TREE_OPERAND (cond, 1); | |
2820 | else | |
2821 | op = TREE_OPERAND (cond, 0); | |
2822 | ||
2823 | if (!vect_is_simple_use (op, stmt_vinfo->vinfo, &def_stmt, | |
2824 | &dt, &vectype)) | |
2825 | gcc_unreachable (); | |
2826 | ||
2827 | return !vectype || VECTOR_BOOLEAN_TYPE_P (vectype); | |
2828 | } | |
2829 | ||
2830 | return VECTOR_BOOLEAN_TYPE_P (STMT_VINFO_VECTYPE (stmt_vinfo)); | |
2831 | } | |
2832 | ||
2833 | ||
b8698a0f L |
2834 | /* For constant and loop invariant defs of SLP_NODE this function returns |
2835 | (vector) defs (VEC_OPRNDS) that will be used in the vectorized stmts. | |
d59dc888 IR |
2836 | OP_NUM determines if we gather defs for operand 0 or operand 1 of the RHS of |
2837 | scalar stmts. NUMBER_OF_VECTORS is the number of vector defs to create. | |
b5aeb3bb IR |
2838 | REDUC_INDEX is the index of the reduction operand in the statements, unless |
2839 | it is -1. */ | |
ebfd146a IR |
2840 | |
2841 | static void | |
9dc3f7de | 2842 | vect_get_constant_vectors (tree op, slp_tree slp_node, |
9771b263 | 2843 | vec<tree> *vec_oprnds, |
b5aeb3bb IR |
2844 | unsigned int op_num, unsigned int number_of_vectors, |
2845 | int reduc_index) | |
ebfd146a | 2846 | { |
355fe088 TS |
2847 | vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
2848 | gimple *stmt = stmts[0]; | |
ebfd146a | 2849 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
d2a12ae7 | 2850 | unsigned nunits; |
ebfd146a | 2851 | tree vec_cst; |
d2a12ae7 RG |
2852 | tree *elts; |
2853 | unsigned j, number_of_places_left_in_vector; | |
ebfd146a | 2854 | tree vector_type; |
9dc3f7de | 2855 | tree vop; |
9771b263 | 2856 | int group_size = stmts.length (); |
ebfd146a | 2857 | unsigned int vec_num, i; |
d2a12ae7 | 2858 | unsigned number_of_copies = 1; |
9771b263 DN |
2859 | vec<tree> voprnds; |
2860 | voprnds.create (number_of_vectors); | |
ebfd146a | 2861 | bool constant_p, is_store; |
b5aeb3bb | 2862 | tree neutral_op = NULL; |
bac430c9 | 2863 | enum tree_code code = gimple_expr_code (stmt); |
355fe088 | 2864 | gimple *def_stmt; |
0e93a64e | 2865 | struct loop *loop; |
13396b6e | 2866 | gimple_seq ctor_seq = NULL; |
b5aeb3bb | 2867 | |
42fd8198 IE |
2868 | /* Check if vector type is a boolean vector. */ |
2869 | if (TREE_CODE (TREE_TYPE (op)) == BOOLEAN_TYPE | |
e4af0bc4 | 2870 | && vect_mask_constant_operand_p (stmt, op_num)) |
42fd8198 IE |
2871 | vector_type |
2872 | = build_same_sized_truth_vector_type (STMT_VINFO_VECTYPE (stmt_vinfo)); | |
2873 | else | |
2874 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
afbe6325 RB |
2875 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); |
2876 | ||
29ed4920 IR |
2877 | if (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
2878 | && reduc_index != -1) | |
b5aeb3bb | 2879 | { |
afbe6325 RB |
2880 | op_num = reduc_index; |
2881 | op = gimple_op (stmt, op_num + 1); | |
b5aeb3bb | 2882 | /* For additional copies (see the explanation of NUMBER_OF_COPIES below) |
ff802fa1 | 2883 | we need either neutral operands or the original operands. See |
b5aeb3bb IR |
2884 | get_initial_def_for_reduction() for details. */ |
2885 | switch (code) | |
2886 | { | |
2887 | case WIDEN_SUM_EXPR: | |
2888 | case DOT_PROD_EXPR: | |
afbe6325 | 2889 | case SAD_EXPR: |
b5aeb3bb IR |
2890 | case PLUS_EXPR: |
2891 | case MINUS_EXPR: | |
2892 | case BIT_IOR_EXPR: | |
2893 | case BIT_XOR_EXPR: | |
2894 | if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (op))) | |
2895 | neutral_op = build_real (TREE_TYPE (op), dconst0); | |
2896 | else | |
2897 | neutral_op = build_int_cst (TREE_TYPE (op), 0); | |
2898 | ||
2899 | break; | |
2900 | ||
2901 | case MULT_EXPR: | |
b5aeb3bb IR |
2902 | if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (op))) |
2903 | neutral_op = build_real (TREE_TYPE (op), dconst1); | |
2904 | else | |
2905 | neutral_op = build_int_cst (TREE_TYPE (op), 1); | |
2906 | ||
2907 | break; | |
2908 | ||
c1e822d5 IR |
2909 | case BIT_AND_EXPR: |
2910 | neutral_op = build_int_cst (TREE_TYPE (op), -1); | |
2911 | break; | |
2912 | ||
f1485e5b RB |
2913 | /* For MIN/MAX we don't have an easy neutral operand but |
2914 | the initial values can be used fine here. Only for | |
2915 | a reduction chain we have to force a neutral element. */ | |
2916 | case MAX_EXPR: | |
2917 | case MIN_EXPR: | |
2918 | if (!GROUP_FIRST_ELEMENT (stmt_vinfo)) | |
2919 | neutral_op = NULL; | |
2920 | else | |
2921 | { | |
2922 | def_stmt = SSA_NAME_DEF_STMT (op); | |
2923 | loop = (gimple_bb (stmt))->loop_father; | |
2924 | neutral_op = PHI_ARG_DEF_FROM_EDGE (def_stmt, | |
2925 | loop_preheader_edge (loop)); | |
2926 | } | |
2927 | break; | |
0e93a64e | 2928 | |
b5aeb3bb | 2929 | default: |
afbe6325 | 2930 | gcc_assert (!GROUP_FIRST_ELEMENT (stmt_vinfo)); |
0e93a64e | 2931 | neutral_op = NULL; |
b5aeb3bb IR |
2932 | } |
2933 | } | |
ebfd146a IR |
2934 | |
2935 | if (STMT_VINFO_DATA_REF (stmt_vinfo)) | |
2936 | { | |
2937 | is_store = true; | |
2938 | op = gimple_assign_rhs1 (stmt); | |
2939 | } | |
2940 | else | |
9dc3f7de IR |
2941 | is_store = false; |
2942 | ||
2943 | gcc_assert (op); | |
ebfd146a IR |
2944 | |
2945 | if (CONSTANT_CLASS_P (op)) | |
d59dc888 | 2946 | constant_p = true; |
ebfd146a | 2947 | else |
d59dc888 IR |
2948 | constant_p = false; |
2949 | ||
ebfd146a | 2950 | /* NUMBER_OF_COPIES is the number of times we need to use the same values in |
b8698a0f | 2951 | created vectors. It is greater than 1 if unrolling is performed. |
ebfd146a IR |
2952 | |
2953 | For example, we have two scalar operands, s1 and s2 (e.g., group of | |
2954 | strided accesses of size two), while NUNITS is four (i.e., four scalars | |
f7e531cf IR |
2955 | of this type can be packed in a vector). The output vector will contain |
2956 | two copies of each scalar operand: {s1, s2, s1, s2}. (NUMBER_OF_COPIES | |
ebfd146a IR |
2957 | will be 2). |
2958 | ||
b8698a0f | 2959 | If GROUP_SIZE > NUNITS, the scalars will be split into several vectors |
ebfd146a IR |
2960 | containing the operands. |
2961 | ||
2962 | For example, NUNITS is four as before, and the group size is 8 | |
f7e531cf | 2963 | (s1, s2, ..., s8). We will create two vectors {s1, s2, s3, s4} and |
ebfd146a | 2964 | {s5, s6, s7, s8}. */ |
b8698a0f | 2965 | |
14a61437 | 2966 | number_of_copies = nunits * number_of_vectors / group_size; |
ebfd146a IR |
2967 | |
2968 | number_of_places_left_in_vector = nunits; | |
d2a12ae7 | 2969 | elts = XALLOCAVEC (tree, nunits); |
90dd6e3d | 2970 | bool place_after_defs = false; |
ebfd146a IR |
2971 | for (j = 0; j < number_of_copies; j++) |
2972 | { | |
9771b263 | 2973 | for (i = group_size - 1; stmts.iterate (i, &stmt); i--) |
ebfd146a IR |
2974 | { |
2975 | if (is_store) | |
2976 | op = gimple_assign_rhs1 (stmt); | |
bac430c9 | 2977 | else |
f7e531cf | 2978 | { |
bac430c9 | 2979 | switch (code) |
f7e531cf | 2980 | { |
bac430c9 | 2981 | case COND_EXPR: |
a989bcc3 IE |
2982 | { |
2983 | tree cond = gimple_assign_rhs1 (stmt); | |
2984 | if (TREE_CODE (cond) == SSA_NAME) | |
2985 | op = gimple_op (stmt, op_num + 1); | |
2986 | else if (op_num == 0 || op_num == 1) | |
bac430c9 | 2987 | op = TREE_OPERAND (cond, op_num); |
a989bcc3 IE |
2988 | else |
2989 | { | |
2990 | if (op_num == 2) | |
2991 | op = gimple_assign_rhs2 (stmt); | |
2992 | else | |
2993 | op = gimple_assign_rhs3 (stmt); | |
2994 | } | |
2995 | } | |
bac430c9 IR |
2996 | break; |
2997 | ||
2998 | case CALL_EXPR: | |
2999 | op = gimple_call_arg (stmt, op_num); | |
3000 | break; | |
3001 | ||
b84b294a JJ |
3002 | case LSHIFT_EXPR: |
3003 | case RSHIFT_EXPR: | |
3004 | case LROTATE_EXPR: | |
3005 | case RROTATE_EXPR: | |
3006 | op = gimple_op (stmt, op_num + 1); | |
3007 | /* Unlike the other binary operators, shifts/rotates have | |
3008 | the shift count being int, instead of the same type as | |
3009 | the lhs, so make sure the scalar is the right type if | |
3010 | we are dealing with vectors of | |
3011 | long long/long/short/char. */ | |
793d9a16 | 3012 | if (op_num == 1 && TREE_CODE (op) == INTEGER_CST) |
b84b294a JJ |
3013 | op = fold_convert (TREE_TYPE (vector_type), op); |
3014 | break; | |
3015 | ||
bac430c9 IR |
3016 | default: |
3017 | op = gimple_op (stmt, op_num + 1); | |
b84b294a | 3018 | break; |
f7e531cf IR |
3019 | } |
3020 | } | |
b8698a0f | 3021 | |
b5aeb3bb IR |
3022 | if (reduc_index != -1) |
3023 | { | |
0e93a64e IR |
3024 | loop = (gimple_bb (stmt))->loop_father; |
3025 | def_stmt = SSA_NAME_DEF_STMT (op); | |
b5aeb3bb IR |
3026 | |
3027 | gcc_assert (loop); | |
b010117a IR |
3028 | |
3029 | /* Get the def before the loop. In reduction chain we have only | |
3030 | one initial value. */ | |
3031 | if ((j != (number_of_copies - 1) | |
3032 | || (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) | |
3033 | && i != 0)) | |
3034 | && neutral_op) | |
b5aeb3bb | 3035 | op = neutral_op; |
b010117a IR |
3036 | else |
3037 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, | |
3038 | loop_preheader_edge (loop)); | |
b5aeb3bb IR |
3039 | } |
3040 | ||
ebfd146a | 3041 | /* Create 'vect_ = {op0,op1,...,opn}'. */ |
ebfd146a | 3042 | number_of_places_left_in_vector--; |
90dd6e3d | 3043 | tree orig_op = op; |
13396b6e | 3044 | if (!types_compatible_p (TREE_TYPE (vector_type), TREE_TYPE (op))) |
50eeef09 | 3045 | { |
793d9a16 | 3046 | if (CONSTANT_CLASS_P (op)) |
13396b6e | 3047 | { |
42fd8198 IE |
3048 | if (VECTOR_BOOLEAN_TYPE_P (vector_type)) |
3049 | { | |
3050 | /* Can't use VIEW_CONVERT_EXPR for booleans because | |
3051 | of possibly different sizes of scalar value and | |
3052 | vector element. */ | |
3053 | if (integer_zerop (op)) | |
3054 | op = build_int_cst (TREE_TYPE (vector_type), 0); | |
3055 | else if (integer_onep (op)) | |
3056 | op = build_int_cst (TREE_TYPE (vector_type), 1); | |
3057 | else | |
3058 | gcc_unreachable (); | |
3059 | } | |
3060 | else | |
3061 | op = fold_unary (VIEW_CONVERT_EXPR, | |
3062 | TREE_TYPE (vector_type), op); | |
13396b6e JJ |
3063 | gcc_assert (op && CONSTANT_CLASS_P (op)); |
3064 | } | |
3065 | else | |
3066 | { | |
b731b390 | 3067 | tree new_temp = make_ssa_name (TREE_TYPE (vector_type)); |
355fe088 | 3068 | gimple *init_stmt; |
262a363f JJ |
3069 | if (VECTOR_BOOLEAN_TYPE_P (vector_type)) |
3070 | { | |
7c285ab9 | 3071 | gcc_assert (INTEGRAL_TYPE_P (TREE_TYPE (op))); |
262a363f JJ |
3072 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, op); |
3073 | } | |
262a363f JJ |
3074 | else |
3075 | { | |
3076 | op = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (vector_type), | |
3077 | op); | |
3078 | init_stmt | |
3079 | = gimple_build_assign (new_temp, VIEW_CONVERT_EXPR, | |
3080 | op); | |
3081 | } | |
13396b6e JJ |
3082 | gimple_seq_add_stmt (&ctor_seq, init_stmt); |
3083 | op = new_temp; | |
3084 | } | |
50eeef09 | 3085 | } |
d2a12ae7 | 3086 | elts[number_of_places_left_in_vector] = op; |
793d9a16 RB |
3087 | if (!CONSTANT_CLASS_P (op)) |
3088 | constant_p = false; | |
90dd6e3d RB |
3089 | if (TREE_CODE (orig_op) == SSA_NAME |
3090 | && !SSA_NAME_IS_DEFAULT_DEF (orig_op) | |
3091 | && STMT_VINFO_BB_VINFO (stmt_vinfo) | |
3092 | && (STMT_VINFO_BB_VINFO (stmt_vinfo)->bb | |
3093 | == gimple_bb (SSA_NAME_DEF_STMT (orig_op)))) | |
3094 | place_after_defs = true; | |
ebfd146a IR |
3095 | |
3096 | if (number_of_places_left_in_vector == 0) | |
3097 | { | |
3098 | number_of_places_left_in_vector = nunits; | |
3099 | ||
3100 | if (constant_p) | |
d2a12ae7 | 3101 | vec_cst = build_vector (vector_type, elts); |
ebfd146a | 3102 | else |
d2a12ae7 | 3103 | { |
9771b263 | 3104 | vec<constructor_elt, va_gc> *v; |
d2a12ae7 | 3105 | unsigned k; |
9771b263 | 3106 | vec_alloc (v, nunits); |
d2a12ae7 RG |
3107 | for (k = 0; k < nunits; ++k) |
3108 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, elts[k]); | |
3109 | vec_cst = build_constructor (vector_type, v); | |
3110 | } | |
90dd6e3d RB |
3111 | tree init; |
3112 | gimple_stmt_iterator gsi; | |
3113 | if (place_after_defs) | |
3114 | { | |
3115 | gsi = gsi_for_stmt | |
3116 | (vect_find_last_scalar_stmt_in_slp (slp_node)); | |
3117 | init = vect_init_vector (stmt, vec_cst, vector_type, &gsi); | |
3118 | } | |
3119 | else | |
3120 | init = vect_init_vector (stmt, vec_cst, vector_type, NULL); | |
13396b6e JJ |
3121 | if (ctor_seq != NULL) |
3122 | { | |
90dd6e3d | 3123 | gsi = gsi_for_stmt (SSA_NAME_DEF_STMT (init)); |
13396b6e JJ |
3124 | gsi_insert_seq_before_without_update (&gsi, ctor_seq, |
3125 | GSI_SAME_STMT); | |
3126 | ctor_seq = NULL; | |
3127 | } | |
90dd6e3d RB |
3128 | voprnds.quick_push (init); |
3129 | place_after_defs = false; | |
ebfd146a IR |
3130 | } |
3131 | } | |
3132 | } | |
3133 | ||
b8698a0f | 3134 | /* Since the vectors are created in the reverse order, we should invert |
ebfd146a | 3135 | them. */ |
9771b263 | 3136 | vec_num = voprnds.length (); |
d2a12ae7 | 3137 | for (j = vec_num; j != 0; j--) |
ebfd146a | 3138 | { |
9771b263 DN |
3139 | vop = voprnds[j - 1]; |
3140 | vec_oprnds->quick_push (vop); | |
ebfd146a IR |
3141 | } |
3142 | ||
9771b263 | 3143 | voprnds.release (); |
ebfd146a IR |
3144 | |
3145 | /* In case that VF is greater than the unrolling factor needed for the SLP | |
b8698a0f L |
3146 | group of stmts, NUMBER_OF_VECTORS to be created is greater than |
3147 | NUMBER_OF_SCALARS/NUNITS or NUNITS/NUMBER_OF_SCALARS, and hence we have | |
ebfd146a | 3148 | to replicate the vectors. */ |
9771b263 | 3149 | while (number_of_vectors > vec_oprnds->length ()) |
ebfd146a | 3150 | { |
b5aeb3bb IR |
3151 | tree neutral_vec = NULL; |
3152 | ||
3153 | if (neutral_op) | |
3154 | { | |
3155 | if (!neutral_vec) | |
b9acc9f1 | 3156 | neutral_vec = build_vector_from_val (vector_type, neutral_op); |
b5aeb3bb | 3157 | |
9771b263 | 3158 | vec_oprnds->quick_push (neutral_vec); |
b5aeb3bb IR |
3159 | } |
3160 | else | |
3161 | { | |
9771b263 DN |
3162 | for (i = 0; vec_oprnds->iterate (i, &vop) && i < vec_num; i++) |
3163 | vec_oprnds->quick_push (vop); | |
b5aeb3bb | 3164 | } |
ebfd146a IR |
3165 | } |
3166 | } | |
3167 | ||
3168 | ||
3169 | /* Get vectorized definitions from SLP_NODE that contains corresponding | |
3170 | vectorized def-stmts. */ | |
3171 | ||
3172 | static void | |
9771b263 | 3173 | vect_get_slp_vect_defs (slp_tree slp_node, vec<tree> *vec_oprnds) |
ebfd146a IR |
3174 | { |
3175 | tree vec_oprnd; | |
355fe088 | 3176 | gimple *vec_def_stmt; |
ebfd146a IR |
3177 | unsigned int i; |
3178 | ||
9771b263 | 3179 | gcc_assert (SLP_TREE_VEC_STMTS (slp_node).exists ()); |
ebfd146a | 3180 | |
9771b263 | 3181 | FOR_EACH_VEC_ELT (SLP_TREE_VEC_STMTS (slp_node), i, vec_def_stmt) |
ebfd146a IR |
3182 | { |
3183 | gcc_assert (vec_def_stmt); | |
3184 | vec_oprnd = gimple_get_lhs (vec_def_stmt); | |
9771b263 | 3185 | vec_oprnds->quick_push (vec_oprnd); |
ebfd146a IR |
3186 | } |
3187 | } | |
3188 | ||
3189 | ||
b8698a0f L |
3190 | /* Get vectorized definitions for SLP_NODE. |
3191 | If the scalar definitions are loop invariants or constants, collect them and | |
ebfd146a IR |
3192 | call vect_get_constant_vectors() to create vector stmts. |
3193 | Otherwise, the def-stmts must be already vectorized and the vectorized stmts | |
d092494c IR |
3194 | must be stored in the corresponding child of SLP_NODE, and we call |
3195 | vect_get_slp_vect_defs () to retrieve them. */ | |
b8698a0f | 3196 | |
ebfd146a | 3197 | void |
9771b263 | 3198 | vect_get_slp_defs (vec<tree> ops, slp_tree slp_node, |
37b5ec8f | 3199 | vec<vec<tree> > *vec_oprnds, int reduc_index) |
ebfd146a | 3200 | { |
355fe088 | 3201 | gimple *first_stmt; |
d092494c IR |
3202 | int number_of_vects = 0, i; |
3203 | unsigned int child_index = 0; | |
b8698a0f | 3204 | HOST_WIDE_INT lhs_size_unit, rhs_size_unit; |
d092494c | 3205 | slp_tree child = NULL; |
37b5ec8f | 3206 | vec<tree> vec_defs; |
e44978dc | 3207 | tree oprnd; |
d092494c | 3208 | bool vectorized_defs; |
ebfd146a | 3209 | |
9771b263 DN |
3210 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
3211 | FOR_EACH_VEC_ELT (ops, i, oprnd) | |
ebfd146a | 3212 | { |
d092494c IR |
3213 | /* For each operand we check if it has vectorized definitions in a child |
3214 | node or we need to create them (for invariants and constants). We | |
3215 | check if the LHS of the first stmt of the next child matches OPRND. | |
3216 | If it does, we found the correct child. Otherwise, we call | |
3217 | vect_get_constant_vectors (), and not advance CHILD_INDEX in order | |
3218 | to check this child node for the next operand. */ | |
3219 | vectorized_defs = false; | |
9771b263 | 3220 | if (SLP_TREE_CHILDREN (slp_node).length () > child_index) |
ebfd146a | 3221 | { |
01d8bf07 | 3222 | child = SLP_TREE_CHILDREN (slp_node)[child_index]; |
d092494c | 3223 | |
e44978dc | 3224 | /* We have to check both pattern and original def, if available. */ |
603cca93 | 3225 | if (SLP_TREE_DEF_TYPE (child) == vect_internal_def) |
e44978dc | 3226 | { |
355fe088 TS |
3227 | gimple *first_def = SLP_TREE_SCALAR_STMTS (child)[0]; |
3228 | gimple *related | |
90dd6e3d RB |
3229 | = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (first_def)); |
3230 | ||
3231 | if (operand_equal_p (oprnd, gimple_get_lhs (first_def), 0) | |
3232 | || (related | |
3233 | && operand_equal_p (oprnd, gimple_get_lhs (related), 0))) | |
3234 | { | |
3235 | /* The number of vector defs is determined by the number of | |
3236 | vector statements in the node from which we get those | |
3237 | statements. */ | |
3238 | number_of_vects = SLP_TREE_NUMBER_OF_VEC_STMTS (child); | |
3239 | vectorized_defs = true; | |
3240 | child_index++; | |
3241 | } | |
e44978dc | 3242 | } |
90dd6e3d RB |
3243 | else |
3244 | child_index++; | |
d092494c | 3245 | } |
ebfd146a | 3246 | |
d092494c IR |
3247 | if (!vectorized_defs) |
3248 | { | |
3249 | if (i == 0) | |
3250 | { | |
3251 | number_of_vects = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
3252 | /* Number of vector stmts was calculated according to LHS in | |
3253 | vect_schedule_slp_instance (), fix it by replacing LHS with | |
3254 | RHS, if necessary. See vect_get_smallest_scalar_type () for | |
3255 | details. */ | |
3256 | vect_get_smallest_scalar_type (first_stmt, &lhs_size_unit, | |
3257 | &rhs_size_unit); | |
3258 | if (rhs_size_unit != lhs_size_unit) | |
3259 | { | |
3260 | number_of_vects *= rhs_size_unit; | |
3261 | number_of_vects /= lhs_size_unit; | |
3262 | } | |
3263 | } | |
3264 | } | |
b5aeb3bb | 3265 | |
d092494c | 3266 | /* Allocate memory for vectorized defs. */ |
37b5ec8f JJ |
3267 | vec_defs = vNULL; |
3268 | vec_defs.create (number_of_vects); | |
ebfd146a | 3269 | |
d092494c IR |
3270 | /* For reduction defs we call vect_get_constant_vectors (), since we are |
3271 | looking for initial loop invariant values. */ | |
3272 | if (vectorized_defs && reduc_index == -1) | |
3273 | /* The defs are already vectorized. */ | |
37b5ec8f | 3274 | vect_get_slp_vect_defs (child, &vec_defs); |
d092494c IR |
3275 | else |
3276 | /* Build vectors from scalar defs. */ | |
37b5ec8f | 3277 | vect_get_constant_vectors (oprnd, slp_node, &vec_defs, i, |
d092494c | 3278 | number_of_vects, reduc_index); |
ebfd146a | 3279 | |
37b5ec8f | 3280 | vec_oprnds->quick_push (vec_defs); |
ebfd146a | 3281 | |
d092494c IR |
3282 | /* For reductions, we only need initial values. */ |
3283 | if (reduc_index != -1) | |
3284 | return; | |
3285 | } | |
ebfd146a IR |
3286 | } |
3287 | ||
a70d6342 | 3288 | |
b8698a0f | 3289 | /* Create NCOPIES permutation statements using the mask MASK_BYTES (by |
ebfd146a IR |
3290 | building a vector of type MASK_TYPE from it) and two input vectors placed in |
3291 | DR_CHAIN at FIRST_VEC_INDX and SECOND_VEC_INDX for the first copy and | |
3292 | shifting by STRIDE elements of DR_CHAIN for every copy. | |
3293 | (STRIDE is the number of vectorized stmts for NODE divided by the number of | |
b8698a0f | 3294 | copies). |
ebfd146a IR |
3295 | VECT_STMTS_COUNTER specifies the index in the vectorized stmts of NODE, where |
3296 | the created stmts must be inserted. */ | |
3297 | ||
3298 | static inline void | |
355fe088 | 3299 | vect_create_mask_and_perm (gimple *stmt, |
faf63e39 | 3300 | tree mask, int first_vec_indx, int second_vec_indx, |
b8698a0f | 3301 | gimple_stmt_iterator *gsi, slp_tree node, |
9771b263 | 3302 | tree vectype, vec<tree> dr_chain, |
ebfd146a IR |
3303 | int ncopies, int vect_stmts_counter) |
3304 | { | |
faf63e39 | 3305 | tree perm_dest; |
355fe088 | 3306 | gimple *perm_stmt = NULL; |
7706cb01 | 3307 | int i, stride_in, stride_out; |
ebfd146a | 3308 | tree first_vec, second_vec, data_ref; |
ebfd146a | 3309 | |
7706cb01 RB |
3310 | stride_out = SLP_TREE_NUMBER_OF_VEC_STMTS (node) / ncopies; |
3311 | stride_in = dr_chain.length () / ncopies; | |
ebfd146a | 3312 | |
b8698a0f | 3313 | /* Initialize the vect stmts of NODE to properly insert the generated |
ebfd146a | 3314 | stmts later. */ |
9771b263 | 3315 | for (i = SLP_TREE_VEC_STMTS (node).length (); |
ebfd146a | 3316 | i < (int) SLP_TREE_NUMBER_OF_VEC_STMTS (node); i++) |
9771b263 | 3317 | SLP_TREE_VEC_STMTS (node).quick_push (NULL); |
ebfd146a IR |
3318 | |
3319 | perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype); | |
3320 | for (i = 0; i < ncopies; i++) | |
3321 | { | |
9771b263 DN |
3322 | first_vec = dr_chain[first_vec_indx]; |
3323 | second_vec = dr_chain[second_vec_indx]; | |
ebfd146a | 3324 | |
be377c80 RB |
3325 | /* Generate the permute statement if necessary. */ |
3326 | if (mask) | |
3327 | { | |
3328 | perm_stmt = gimple_build_assign (perm_dest, VEC_PERM_EXPR, | |
3329 | first_vec, second_vec, mask); | |
3330 | data_ref = make_ssa_name (perm_dest, perm_stmt); | |
3331 | gimple_set_lhs (perm_stmt, data_ref); | |
3332 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); | |
3333 | } | |
3334 | else | |
3335 | /* If mask was NULL_TREE generate the requested identity transform. */ | |
3336 | perm_stmt = SSA_NAME_DEF_STMT (first_vec); | |
ebfd146a | 3337 | |
b8698a0f | 3338 | /* Store the vector statement in NODE. */ |
7706cb01 RB |
3339 | SLP_TREE_VEC_STMTS (node)[stride_out * i + vect_stmts_counter] |
3340 | = perm_stmt; | |
ebfd146a | 3341 | |
7706cb01 RB |
3342 | first_vec_indx += stride_in; |
3343 | second_vec_indx += stride_in; | |
ebfd146a | 3344 | } |
ebfd146a IR |
3345 | } |
3346 | ||
3347 | ||
ebfd146a IR |
3348 | /* Generate vector permute statements from a list of loads in DR_CHAIN. |
3349 | If ANALYZE_ONLY is TRUE, only check that it is possible to create valid | |
01d8bf07 RB |
3350 | permute statements for the SLP node NODE of the SLP instance |
3351 | SLP_NODE_INSTANCE. */ | |
3352 | ||
ebfd146a | 3353 | bool |
01d8bf07 | 3354 | vect_transform_slp_perm_load (slp_tree node, vec<tree> dr_chain, |
ebfd146a IR |
3355 | gimple_stmt_iterator *gsi, int vf, |
3356 | slp_instance slp_node_instance, bool analyze_only) | |
3357 | { | |
355fe088 | 3358 | gimple *stmt = SLP_TREE_SCALAR_STMTS (node)[0]; |
ebfd146a IR |
3359 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
3360 | tree mask_element_type = NULL_TREE, mask_type; | |
2ce27200 | 3361 | int nunits, vec_index = 0; |
2635892a | 3362 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 3363 | int group_size = SLP_INSTANCE_GROUP_SIZE (slp_node_instance); |
2ce27200 | 3364 | int unroll_factor, mask_element, ncopies; |
22e4dee7 | 3365 | unsigned char *mask; |
ef4bddc2 | 3366 | machine_mode mode; |
ebfd146a | 3367 | |
91ff1504 RB |
3368 | if (!STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
3369 | return false; | |
3370 | ||
3371 | stmt_info = vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info)); | |
3372 | ||
22e4dee7 RH |
3373 | mode = TYPE_MODE (vectype); |
3374 | ||
2635892a RH |
3375 | /* The generic VEC_PERM_EXPR code always uses an integral type of the |
3376 | same size as the vector element being permuted. */ | |
96f9265a RG |
3377 | mask_element_type = lang_hooks.types.type_for_mode |
3378 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
ebfd146a | 3379 | mask_type = get_vectype_for_scalar_type (mask_element_type); |
ebfd146a | 3380 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 | 3381 | mask = XALLOCAVEC (unsigned char, nunits); |
ebfd146a IR |
3382 | unroll_factor = SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance); |
3383 | ||
b8698a0f | 3384 | /* Number of copies is determined by the final vectorization factor |
ebfd146a | 3385 | relatively to SLP_NODE_INSTANCE unrolling factor. */ |
b8698a0f | 3386 | ncopies = vf / SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance); |
ebfd146a | 3387 | |
b8698a0f L |
3388 | /* Generate permutation masks for every NODE. Number of masks for each NODE |
3389 | is equal to GROUP_SIZE. | |
3390 | E.g., we have a group of three nodes with three loads from the same | |
3391 | location in each node, and the vector size is 4. I.e., we have a | |
3392 | a0b0c0a1b1c1... sequence and we need to create the following vectors: | |
ebfd146a IR |
3393 | for a's: a0a0a0a1 a1a1a2a2 a2a3a3a3 |
3394 | for b's: b0b0b0b1 b1b1b2b2 b2b3b3b3 | |
3395 | ... | |
3396 | ||
2635892a | 3397 | The masks for a's should be: {0,0,0,3} {3,3,6,6} {6,9,9,9}. |
b8698a0f | 3398 | The last mask is illegal since we assume two operands for permute |
ff802fa1 IR |
3399 | operation, and the mask element values can't be outside that range. |
3400 | Hence, the last mask must be converted into {2,5,5,5}. | |
b8698a0f | 3401 | For the first two permutations we need the first and the second input |
ebfd146a | 3402 | vectors: {a0,b0,c0,a1} and {b1,c1,a2,b2}, and for the last permutation |
b8698a0f | 3403 | we need the second and the third vectors: {b1,c1,a2,b2} and |
ebfd146a IR |
3404 | {c2,a3,b3,c3}. */ |
3405 | ||
2ce27200 RB |
3406 | int vect_stmts_counter = 0; |
3407 | int index = 0; | |
3408 | int first_vec_index = -1; | |
3409 | int second_vec_index = -1; | |
be377c80 | 3410 | bool noop_p = true; |
ebfd146a | 3411 | |
2ce27200 RB |
3412 | for (int j = 0; j < unroll_factor; j++) |
3413 | { | |
3414 | for (int k = 0; k < group_size; k++) | |
3415 | { | |
3416 | int i = (SLP_TREE_LOAD_PERMUTATION (node)[k] | |
3417 | + j * STMT_VINFO_GROUP_SIZE (stmt_info)); | |
3418 | vec_index = i / nunits; | |
3419 | mask_element = i % nunits; | |
3420 | if (vec_index == first_vec_index | |
3421 | || first_vec_index == -1) | |
3422 | { | |
3423 | first_vec_index = vec_index; | |
3424 | } | |
3425 | else if (vec_index == second_vec_index | |
3426 | || second_vec_index == -1) | |
3427 | { | |
3428 | second_vec_index = vec_index; | |
3429 | mask_element += nunits; | |
3430 | } | |
3431 | else | |
3432 | { | |
3433 | if (dump_enabled_p ()) | |
3434 | { | |
3435 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3436 | "permutation requires at " | |
3437 | "least three vectors "); | |
3438 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
3439 | stmt, 0); | |
3440 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
3441 | } | |
3442 | return false; | |
3443 | } | |
ebfd146a | 3444 | |
2ce27200 RB |
3445 | gcc_assert (mask_element >= 0 |
3446 | && mask_element < 2 * nunits); | |
be377c80 RB |
3447 | if (mask_element != index) |
3448 | noop_p = false; | |
2ce27200 RB |
3449 | mask[index++] = mask_element; |
3450 | ||
3451 | if (index == nunits) | |
3452 | { | |
be377c80 RB |
3453 | if (! noop_p |
3454 | && ! can_vec_perm_p (mode, false, mask)) | |
2ce27200 RB |
3455 | { |
3456 | if (dump_enabled_p ()) | |
22e4dee7 | 3457 | { |
2ce27200 RB |
3458 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, |
3459 | vect_location, | |
3460 | "unsupported vect permute { "); | |
3461 | for (i = 0; i < nunits; ++i) | |
3462 | dump_printf (MSG_MISSED_OPTIMIZATION, "%d ", mask[i]); | |
3463 | dump_printf (MSG_MISSED_OPTIMIZATION, "}\n"); | |
22e4dee7 | 3464 | } |
2ce27200 RB |
3465 | return false; |
3466 | } | |
22e4dee7 | 3467 | |
2ce27200 RB |
3468 | if (!analyze_only) |
3469 | { | |
be377c80 RB |
3470 | tree mask_vec = NULL_TREE; |
3471 | ||
3472 | if (! noop_p) | |
3473 | { | |
3474 | tree *mask_elts = XALLOCAVEC (tree, nunits); | |
3475 | for (int l = 0; l < nunits; ++l) | |
3476 | mask_elts[l] = build_int_cst (mask_element_type, | |
3477 | mask[l]); | |
3478 | mask_vec = build_vector (mask_type, mask_elts); | |
3479 | } | |
2ce27200 RB |
3480 | |
3481 | if (second_vec_index == -1) | |
3482 | second_vec_index = first_vec_index; | |
3483 | vect_create_mask_and_perm (stmt, mask_vec, first_vec_index, | |
3484 | second_vec_index, | |
3485 | gsi, node, vectype, dr_chain, | |
3486 | ncopies, vect_stmts_counter++); | |
3487 | } | |
ebfd146a | 3488 | |
2ce27200 RB |
3489 | index = 0; |
3490 | first_vec_index = -1; | |
3491 | second_vec_index = -1; | |
be377c80 | 3492 | noop_p = true; |
2ce27200 RB |
3493 | } |
3494 | } | |
b8698a0f | 3495 | } |
ebfd146a | 3496 | |
ebfd146a IR |
3497 | return true; |
3498 | } | |
3499 | ||
3500 | ||
3501 | ||
3502 | /* Vectorize SLP instance tree in postorder. */ | |
3503 | ||
3504 | static bool | |
3505 | vect_schedule_slp_instance (slp_tree node, slp_instance instance, | |
a70d6342 | 3506 | unsigned int vectorization_factor) |
ebfd146a | 3507 | { |
355fe088 | 3508 | gimple *stmt; |
0d0293ac | 3509 | bool grouped_store, is_store; |
ebfd146a IR |
3510 | gimple_stmt_iterator si; |
3511 | stmt_vec_info stmt_info; | |
3512 | unsigned int vec_stmts_size, nunits, group_size; | |
3513 | tree vectype; | |
603cca93 | 3514 | int i, j; |
d755c7ef | 3515 | slp_tree child; |
ebfd146a | 3516 | |
603cca93 | 3517 | if (SLP_TREE_DEF_TYPE (node) != vect_internal_def) |
ebfd146a IR |
3518 | return false; |
3519 | ||
9771b263 | 3520 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) |
d755c7ef | 3521 | vect_schedule_slp_instance (child, instance, vectorization_factor); |
b8698a0f | 3522 | |
603cca93 RB |
3523 | /* Push SLP node def-type to stmts. */ |
3524 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) | |
3525 | if (SLP_TREE_DEF_TYPE (child) != vect_internal_def) | |
3526 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (child), j, stmt) | |
3527 | STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmt)) = SLP_TREE_DEF_TYPE (child); | |
3528 | ||
9771b263 | 3529 | stmt = SLP_TREE_SCALAR_STMTS (node)[0]; |
ebfd146a IR |
3530 | stmt_info = vinfo_for_stmt (stmt); |
3531 | ||
3532 | /* VECTYPE is the type of the destination. */ | |
b690cc0f | 3533 | vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a IR |
3534 | nunits = (unsigned int) TYPE_VECTOR_SUBPARTS (vectype); |
3535 | group_size = SLP_INSTANCE_GROUP_SIZE (instance); | |
3536 | ||
3537 | /* For each SLP instance calculate number of vector stmts to be created | |
ff802fa1 | 3538 | for the scalar stmts in each node of the SLP tree. Number of vector |
ebfd146a IR |
3539 | elements in one vector iteration is the number of scalar elements in |
3540 | one scalar iteration (GROUP_SIZE) multiplied by VF divided by vector | |
14a61437 RB |
3541 | size. |
3542 | Unless this is a SLP reduction in which case the number of vector | |
3543 | stmts is equal to the number of vector stmts of the children. */ | |
3544 | if (GROUP_FIRST_ELEMENT (stmt_info) | |
3545 | && !STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
3546 | vec_stmts_size = SLP_TREE_NUMBER_OF_VEC_STMTS (SLP_TREE_CHILDREN (node)[0]); | |
3547 | else | |
3548 | vec_stmts_size = (vectorization_factor * group_size) / nunits; | |
ebfd146a | 3549 | |
9771b263 | 3550 | if (!SLP_TREE_VEC_STMTS (node).exists ()) |
ebfd146a | 3551 | { |
9771b263 | 3552 | SLP_TREE_VEC_STMTS (node).create (vec_stmts_size); |
ebfd146a IR |
3553 | SLP_TREE_NUMBER_OF_VEC_STMTS (node) = vec_stmts_size; |
3554 | } | |
3555 | ||
73fbfcad | 3556 | if (dump_enabled_p ()) |
ebfd146a | 3557 | { |
78c60e3d SS |
3558 | dump_printf_loc (MSG_NOTE,vect_location, |
3559 | "------>vectorizing SLP node starting from: "); | |
3560 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 3561 | dump_printf (MSG_NOTE, "\n"); |
b8698a0f | 3562 | } |
ebfd146a | 3563 | |
2e8ab70c RB |
3564 | /* Vectorized stmts go before the last scalar stmt which is where |
3565 | all uses are ready. */ | |
3566 | si = gsi_for_stmt (vect_find_last_scalar_stmt_in_slp (node)); | |
e4a707c4 | 3567 | |
b010117a IR |
3568 | /* Mark the first element of the reduction chain as reduction to properly |
3569 | transform the node. In the analysis phase only the last element of the | |
3570 | chain is marked as reduction. */ | |
0d0293ac | 3571 | if (GROUP_FIRST_ELEMENT (stmt_info) && !STMT_VINFO_GROUPED_ACCESS (stmt_info) |
b010117a IR |
3572 | && GROUP_FIRST_ELEMENT (stmt_info) == stmt) |
3573 | { | |
3574 | STMT_VINFO_DEF_TYPE (stmt_info) = vect_reduction_def; | |
3575 | STMT_VINFO_TYPE (stmt_info) = reduc_vec_info_type; | |
3576 | } | |
3577 | ||
6876e5bc RB |
3578 | /* Handle two-operation SLP nodes by vectorizing the group with |
3579 | both operations and then performing a merge. */ | |
3580 | if (SLP_TREE_TWO_OPERATORS (node)) | |
3581 | { | |
3582 | enum tree_code code0 = gimple_assign_rhs_code (stmt); | |
567a3691 | 3583 | enum tree_code ocode = ERROR_MARK; |
355fe088 | 3584 | gimple *ostmt; |
6876e5bc | 3585 | unsigned char *mask = XALLOCAVEC (unsigned char, group_size); |
6876e5bc RB |
3586 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, ostmt) |
3587 | if (gimple_assign_rhs_code (ostmt) != code0) | |
3588 | { | |
3589 | mask[i] = 1; | |
6876e5bc RB |
3590 | ocode = gimple_assign_rhs_code (ostmt); |
3591 | } | |
3592 | else | |
3593 | mask[i] = 0; | |
567a3691 | 3594 | if (ocode != ERROR_MARK) |
6876e5bc | 3595 | { |
355fe088 TS |
3596 | vec<gimple *> v0; |
3597 | vec<gimple *> v1; | |
6876e5bc RB |
3598 | unsigned j; |
3599 | tree tmask = NULL_TREE; | |
3600 | vect_transform_stmt (stmt, &si, &grouped_store, node, instance); | |
3601 | v0 = SLP_TREE_VEC_STMTS (node).copy (); | |
3602 | SLP_TREE_VEC_STMTS (node).truncate (0); | |
3603 | gimple_assign_set_rhs_code (stmt, ocode); | |
3604 | vect_transform_stmt (stmt, &si, &grouped_store, node, instance); | |
3605 | gimple_assign_set_rhs_code (stmt, code0); | |
3606 | v1 = SLP_TREE_VEC_STMTS (node).copy (); | |
3607 | SLP_TREE_VEC_STMTS (node).truncate (0); | |
3608 | tree meltype = build_nonstandard_integer_type | |
3609 | (GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
3610 | tree mvectype = get_same_sized_vectype (meltype, vectype); | |
3611 | unsigned k = 0, l; | |
3612 | for (j = 0; j < v0.length (); ++j) | |
3613 | { | |
3614 | tree *melts = XALLOCAVEC (tree, TYPE_VECTOR_SUBPARTS (vectype)); | |
3615 | for (l = 0; l < TYPE_VECTOR_SUBPARTS (vectype); ++l) | |
3616 | { | |
1ece8d4c | 3617 | if (k >= group_size) |
6876e5bc RB |
3618 | k = 0; |
3619 | melts[l] = build_int_cst | |
3620 | (meltype, mask[k++] * TYPE_VECTOR_SUBPARTS (vectype) + l); | |
3621 | } | |
3622 | tmask = build_vector (mvectype, melts); | |
3623 | ||
3624 | /* ??? Not all targets support a VEC_PERM_EXPR with a | |
3625 | constant mask that would translate to a vec_merge RTX | |
3626 | (with their vec_perm_const_ok). We can either not | |
3627 | vectorize in that case or let veclower do its job. | |
3628 | Unfortunately that isn't too great and at least for | |
3629 | plus/minus we'd eventually like to match targets | |
3630 | vector addsub instructions. */ | |
355fe088 | 3631 | gimple *vstmt; |
6876e5bc RB |
3632 | vstmt = gimple_build_assign (make_ssa_name (vectype), |
3633 | VEC_PERM_EXPR, | |
3634 | gimple_assign_lhs (v0[j]), | |
3635 | gimple_assign_lhs (v1[j]), tmask); | |
3636 | vect_finish_stmt_generation (stmt, vstmt, &si); | |
3637 | SLP_TREE_VEC_STMTS (node).quick_push (vstmt); | |
3638 | } | |
3639 | v0.release (); | |
3640 | v1.release (); | |
3641 | return false; | |
3642 | } | |
3643 | } | |
0d0293ac | 3644 | is_store = vect_transform_stmt (stmt, &si, &grouped_store, node, instance); |
603cca93 RB |
3645 | |
3646 | /* Restore stmt def-types. */ | |
3647 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) | |
3648 | if (SLP_TREE_DEF_TYPE (child) != vect_internal_def) | |
3649 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (child), j, stmt) | |
3650 | STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmt)) = vect_internal_def; | |
3651 | ||
b5aeb3bb | 3652 | return is_store; |
ebfd146a IR |
3653 | } |
3654 | ||
dd34c087 JJ |
3655 | /* Replace scalar calls from SLP node NODE with setting of their lhs to zero. |
3656 | For loop vectorization this is done in vectorizable_call, but for SLP | |
3657 | it needs to be deferred until end of vect_schedule_slp, because multiple | |
3658 | SLP instances may refer to the same scalar stmt. */ | |
3659 | ||
3660 | static void | |
3661 | vect_remove_slp_scalar_calls (slp_tree node) | |
3662 | { | |
355fe088 | 3663 | gimple *stmt, *new_stmt; |
dd34c087 JJ |
3664 | gimple_stmt_iterator gsi; |
3665 | int i; | |
d755c7ef | 3666 | slp_tree child; |
dd34c087 JJ |
3667 | tree lhs; |
3668 | stmt_vec_info stmt_info; | |
3669 | ||
603cca93 | 3670 | if (SLP_TREE_DEF_TYPE (node) != vect_internal_def) |
dd34c087 JJ |
3671 | return; |
3672 | ||
9771b263 | 3673 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) |
d755c7ef | 3674 | vect_remove_slp_scalar_calls (child); |
dd34c087 | 3675 | |
9771b263 | 3676 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt) |
dd34c087 JJ |
3677 | { |
3678 | if (!is_gimple_call (stmt) || gimple_bb (stmt) == NULL) | |
3679 | continue; | |
3680 | stmt_info = vinfo_for_stmt (stmt); | |
3681 | if (stmt_info == NULL | |
3682 | || is_pattern_stmt_p (stmt_info) | |
3683 | || !PURE_SLP_STMT (stmt_info)) | |
3684 | continue; | |
3685 | lhs = gimple_call_lhs (stmt); | |
3686 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
3687 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
3688 | set_vinfo_for_stmt (stmt, NULL); | |
3689 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
3690 | gsi = gsi_for_stmt (stmt); | |
3691 | gsi_replace (&gsi, new_stmt, false); | |
3692 | SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt; | |
3693 | } | |
3694 | } | |
ebfd146a | 3695 | |
ff802fa1 IR |
3696 | /* Generate vector code for all SLP instances in the loop/basic block. */ |
3697 | ||
ebfd146a | 3698 | bool |
310213d4 | 3699 | vect_schedule_slp (vec_info *vinfo) |
ebfd146a | 3700 | { |
9771b263 | 3701 | vec<slp_instance> slp_instances; |
ebfd146a | 3702 | slp_instance instance; |
01d8bf07 | 3703 | unsigned int i, vf; |
ebfd146a IR |
3704 | bool is_store = false; |
3705 | ||
310213d4 RB |
3706 | slp_instances = vinfo->slp_instances; |
3707 | if (is_a <loop_vec_info> (vinfo)) | |
3708 | vf = as_a <loop_vec_info> (vinfo)->vectorization_factor; | |
a70d6342 | 3709 | else |
310213d4 | 3710 | vf = 1; |
a70d6342 | 3711 | |
9771b263 | 3712 | FOR_EACH_VEC_ELT (slp_instances, i, instance) |
ebfd146a IR |
3713 | { |
3714 | /* Schedule the tree of INSTANCE. */ | |
3715 | is_store = vect_schedule_slp_instance (SLP_INSTANCE_TREE (instance), | |
a70d6342 | 3716 | instance, vf); |
73fbfcad | 3717 | if (dump_enabled_p ()) |
78c60e3d | 3718 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3719 | "vectorizing stmts using SLP.\n"); |
ebfd146a IR |
3720 | } |
3721 | ||
9771b263 | 3722 | FOR_EACH_VEC_ELT (slp_instances, i, instance) |
b5aeb3bb IR |
3723 | { |
3724 | slp_tree root = SLP_INSTANCE_TREE (instance); | |
355fe088 | 3725 | gimple *store; |
b5aeb3bb IR |
3726 | unsigned int j; |
3727 | gimple_stmt_iterator gsi; | |
3728 | ||
c40eced0 RB |
3729 | /* Remove scalar call stmts. Do not do this for basic-block |
3730 | vectorization as not all uses may be vectorized. | |
3731 | ??? Why should this be necessary? DCE should be able to | |
3732 | remove the stmts itself. | |
3733 | ??? For BB vectorization we can as well remove scalar | |
3734 | stmts starting from the SLP tree root if they have no | |
3735 | uses. */ | |
310213d4 | 3736 | if (is_a <loop_vec_info> (vinfo)) |
c40eced0 | 3737 | vect_remove_slp_scalar_calls (root); |
dd34c087 | 3738 | |
9771b263 | 3739 | for (j = 0; SLP_TREE_SCALAR_STMTS (root).iterate (j, &store) |
b5aeb3bb IR |
3740 | && j < SLP_INSTANCE_GROUP_SIZE (instance); j++) |
3741 | { | |
3742 | if (!STMT_VINFO_DATA_REF (vinfo_for_stmt (store))) | |
3743 | break; | |
3744 | ||
a024e70e IR |
3745 | if (is_pattern_stmt_p (vinfo_for_stmt (store))) |
3746 | store = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (store)); | |
b5aeb3bb IR |
3747 | /* Free the attached stmt_vec_info and remove the stmt. */ |
3748 | gsi = gsi_for_stmt (store); | |
3d3f2249 | 3749 | unlink_stmt_vdef (store); |
b5aeb3bb | 3750 | gsi_remove (&gsi, true); |
3d3f2249 | 3751 | release_defs (store); |
b5aeb3bb IR |
3752 | free_stmt_vec_info (store); |
3753 | } | |
3754 | } | |
3755 | ||
ebfd146a IR |
3756 | return is_store; |
3757 | } |