]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* SLP - Basic Block Vectorization |
5624e564 | 2 | Copyright (C) 2007-2015 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" | |
78c60e3d | 25 | #include "dumpfile.h" |
ebfd146a | 26 | #include "tm.h" |
40e23961 MC |
27 | #include "hash-set.h" |
28 | #include "machmode.h" | |
29 | #include "vec.h" | |
30 | #include "double-int.h" | |
31 | #include "input.h" | |
32 | #include "alias.h" | |
33 | #include "symtab.h" | |
34 | #include "wide-int.h" | |
35 | #include "inchash.h" | |
ebfd146a | 36 | #include "tree.h" |
40e23961 | 37 | #include "fold-const.h" |
d8a2d370 | 38 | #include "stor-layout.h" |
ebfd146a | 39 | #include "target.h" |
60393bbc | 40 | #include "predict.h" |
60393bbc | 41 | #include "hard-reg-set.h" |
60393bbc | 42 | #include "function.h" |
ebfd146a | 43 | #include "basic-block.h" |
cf835838 | 44 | #include "gimple-pretty-print.h" |
2fb9a547 AM |
45 | #include "tree-ssa-alias.h" |
46 | #include "internal-fn.h" | |
47 | #include "gimple-expr.h" | |
48 | #include "is-a.h" | |
442b4905 | 49 | #include "gimple.h" |
5be5c238 | 50 | #include "gimple-iterator.h" |
442b4905 AM |
51 | #include "gimple-ssa.h" |
52 | #include "tree-phinodes.h" | |
53 | #include "ssa-iterators.h" | |
d8a2d370 | 54 | #include "stringpool.h" |
442b4905 | 55 | #include "tree-ssanames.h" |
7ee2468b | 56 | #include "tree-pass.h" |
ebfd146a | 57 | #include "cfgloop.h" |
36566b39 PK |
58 | #include "hashtab.h" |
59 | #include "rtl.h" | |
60 | #include "flags.h" | |
61 | #include "statistics.h" | |
62 | #include "real.h" | |
63 | #include "fixed-value.h" | |
64 | #include "insn-config.h" | |
65 | #include "expmed.h" | |
66 | #include "dojump.h" | |
67 | #include "explow.h" | |
68 | #include "calls.h" | |
69 | #include "emit-rtl.h" | |
70 | #include "varasm.h" | |
71 | #include "stmt.h" | |
ebfd146a | 72 | #include "expr.h" |
7ee2468b | 73 | #include "recog.h" /* FIXME: for insn_data */ |
b0710fe1 | 74 | #include "insn-codes.h" |
ebfd146a IR |
75 | #include "optabs.h" |
76 | #include "tree-vectorizer.h" | |
2635892a | 77 | #include "langhooks.h" |
642fce57 | 78 | #include "gimple-walk.h" |
ebfd146a | 79 | |
a70d6342 IR |
80 | /* Extract the location of the basic block in the source code. |
81 | Return the basic block location if succeed and NULL if not. */ | |
82 | ||
b05e0233 | 83 | source_location |
a70d6342 IR |
84 | find_bb_location (basic_block bb) |
85 | { | |
86 | gimple stmt = NULL; | |
87 | gimple_stmt_iterator si; | |
88 | ||
89 | if (!bb) | |
b05e0233 | 90 | return UNKNOWN_LOCATION; |
a70d6342 IR |
91 | |
92 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
93 | { | |
94 | stmt = gsi_stmt (si); | |
b05e0233 | 95 | if (gimple_location (stmt) != UNKNOWN_LOCATION) |
a70d6342 IR |
96 | return gimple_location (stmt); |
97 | } | |
98 | ||
b05e0233 | 99 | return UNKNOWN_LOCATION; |
a70d6342 IR |
100 | } |
101 | ||
102 | ||
ebfd146a IR |
103 | /* Recursively free the memory allocated for the SLP tree rooted at NODE. */ |
104 | ||
105 | static void | |
106 | vect_free_slp_tree (slp_tree node) | |
107 | { | |
d092494c | 108 | int i; |
d755c7ef | 109 | slp_tree child; |
d092494c | 110 | |
ebfd146a IR |
111 | if (!node) |
112 | return; | |
113 | ||
9771b263 | 114 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) |
d755c7ef | 115 | vect_free_slp_tree (child); |
b8698a0f | 116 | |
9771b263 DN |
117 | SLP_TREE_CHILDREN (node).release (); |
118 | SLP_TREE_SCALAR_STMTS (node).release (); | |
119 | SLP_TREE_VEC_STMTS (node).release (); | |
01d8bf07 | 120 | SLP_TREE_LOAD_PERMUTATION (node).release (); |
ebfd146a IR |
121 | |
122 | free (node); | |
123 | } | |
124 | ||
125 | ||
126 | /* Free the memory allocated for the SLP instance. */ | |
127 | ||
128 | void | |
129 | vect_free_slp_instance (slp_instance instance) | |
130 | { | |
131 | vect_free_slp_tree (SLP_INSTANCE_TREE (instance)); | |
9771b263 DN |
132 | SLP_INSTANCE_LOADS (instance).release (); |
133 | SLP_INSTANCE_BODY_COST_VEC (instance).release (); | |
c7e62a26 | 134 | free (instance); |
ebfd146a IR |
135 | } |
136 | ||
137 | ||
d092494c IR |
138 | /* Create an SLP node for SCALAR_STMTS. */ |
139 | ||
140 | static slp_tree | |
9771b263 | 141 | vect_create_new_slp_node (vec<gimple> scalar_stmts) |
d092494c | 142 | { |
d3cfd39e | 143 | slp_tree node; |
9771b263 | 144 | gimple stmt = scalar_stmts[0]; |
d092494c IR |
145 | unsigned int nops; |
146 | ||
147 | if (is_gimple_call (stmt)) | |
148 | nops = gimple_call_num_args (stmt); | |
149 | else if (is_gimple_assign (stmt)) | |
f7e531cf IR |
150 | { |
151 | nops = gimple_num_ops (stmt) - 1; | |
152 | if (gimple_assign_rhs_code (stmt) == COND_EXPR) | |
153 | nops++; | |
154 | } | |
d092494c IR |
155 | else |
156 | return NULL; | |
157 | ||
d3cfd39e | 158 | node = XNEW (struct _slp_tree); |
d092494c | 159 | SLP_TREE_SCALAR_STMTS (node) = scalar_stmts; |
9771b263 DN |
160 | SLP_TREE_VEC_STMTS (node).create (0); |
161 | SLP_TREE_CHILDREN (node).create (nops); | |
01d8bf07 | 162 | SLP_TREE_LOAD_PERMUTATION (node) = vNULL; |
d092494c IR |
163 | |
164 | return node; | |
165 | } | |
166 | ||
167 | ||
168 | /* Allocate operands info for NOPS operands, and GROUP_SIZE def-stmts for each | |
169 | operand. */ | |
9771b263 | 170 | static vec<slp_oprnd_info> |
d092494c IR |
171 | vect_create_oprnd_info (int nops, int group_size) |
172 | { | |
173 | int i; | |
174 | slp_oprnd_info oprnd_info; | |
9771b263 | 175 | vec<slp_oprnd_info> oprnds_info; |
d092494c | 176 | |
9771b263 | 177 | oprnds_info.create (nops); |
d092494c IR |
178 | for (i = 0; i < nops; i++) |
179 | { | |
180 | oprnd_info = XNEW (struct _slp_oprnd_info); | |
9771b263 | 181 | oprnd_info->def_stmts.create (group_size); |
d092494c | 182 | oprnd_info->first_dt = vect_uninitialized_def; |
793d9a16 | 183 | oprnd_info->first_op_type = NULL_TREE; |
d092494c | 184 | oprnd_info->first_pattern = false; |
9771b263 | 185 | oprnds_info.quick_push (oprnd_info); |
d092494c IR |
186 | } |
187 | ||
188 | return oprnds_info; | |
189 | } | |
190 | ||
191 | ||
d3cfd39e JJ |
192 | /* Free operands info. */ |
193 | ||
d092494c | 194 | static void |
9771b263 | 195 | vect_free_oprnd_info (vec<slp_oprnd_info> &oprnds_info) |
d092494c IR |
196 | { |
197 | int i; | |
198 | slp_oprnd_info oprnd_info; | |
199 | ||
9771b263 | 200 | FOR_EACH_VEC_ELT (oprnds_info, i, oprnd_info) |
d3cfd39e | 201 | { |
9771b263 | 202 | oprnd_info->def_stmts.release (); |
d3cfd39e JJ |
203 | XDELETE (oprnd_info); |
204 | } | |
d092494c | 205 | |
9771b263 | 206 | oprnds_info.release (); |
d092494c IR |
207 | } |
208 | ||
209 | ||
d755c7ef RB |
210 | /* Find the place of the data-ref in STMT in the interleaving chain that starts |
211 | from FIRST_STMT. Return -1 if the data-ref is not a part of the chain. */ | |
212 | ||
213 | static int | |
214 | vect_get_place_in_interleaving_chain (gimple stmt, gimple first_stmt) | |
215 | { | |
216 | gimple next_stmt = first_stmt; | |
217 | int result = 0; | |
218 | ||
219 | if (first_stmt != GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) | |
220 | return -1; | |
221 | ||
222 | do | |
223 | { | |
224 | if (next_stmt == stmt) | |
225 | return result; | |
226 | result++; | |
227 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); | |
228 | } | |
229 | while (next_stmt); | |
230 | ||
231 | return -1; | |
232 | } | |
233 | ||
234 | ||
d092494c IR |
235 | /* Get the defs for the rhs of STMT (collect them in OPRNDS_INFO), check that |
236 | they are of a valid type and that they match the defs of the first stmt of | |
b0b4483e RB |
237 | the SLP group (stored in OPRNDS_INFO). If there was a fatal error |
238 | return -1, if the error could be corrected by swapping operands of the | |
239 | operation return 1, if everything is ok return 0. */ | |
ebfd146a | 240 | |
b0b4483e | 241 | static int |
a70d6342 | 242 | vect_get_and_check_slp_defs (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo, |
23847df4 RB |
243 | gimple stmt, bool first, |
244 | vec<slp_oprnd_info> *oprnds_info) | |
ebfd146a IR |
245 | { |
246 | tree oprnd; | |
247 | unsigned int i, number_of_oprnds; | |
abf9bfbc | 248 | tree def; |
ebfd146a | 249 | gimple def_stmt; |
d092494c | 250 | enum vect_def_type dt = vect_uninitialized_def; |
a70d6342 | 251 | struct loop *loop = NULL; |
d092494c | 252 | bool pattern = false; |
abf9bfbc | 253 | slp_oprnd_info oprnd_info; |
b0b4483e RB |
254 | int first_op_idx = 1; |
255 | bool commutative = false; | |
256 | bool first_op_cond = false; | |
b8698a0f | 257 | |
a70d6342 IR |
258 | if (loop_vinfo) |
259 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a | 260 | |
d092494c | 261 | if (is_gimple_call (stmt)) |
190c2236 JJ |
262 | { |
263 | number_of_oprnds = gimple_call_num_args (stmt); | |
b0b4483e | 264 | first_op_idx = 3; |
190c2236 | 265 | } |
f7e531cf IR |
266 | else if (is_gimple_assign (stmt)) |
267 | { | |
b0b4483e | 268 | enum tree_code code = gimple_assign_rhs_code (stmt); |
f7e531cf IR |
269 | number_of_oprnds = gimple_num_ops (stmt) - 1; |
270 | if (gimple_assign_rhs_code (stmt) == COND_EXPR) | |
b0b4483e RB |
271 | { |
272 | first_op_cond = true; | |
273 | commutative = true; | |
274 | number_of_oprnds++; | |
275 | } | |
276 | else | |
277 | commutative = commutative_tree_code (code); | |
f7e531cf | 278 | } |
d092494c | 279 | else |
b0b4483e | 280 | return -1; |
ebfd146a | 281 | |
b0b4483e | 282 | bool swapped = false; |
ebfd146a IR |
283 | for (i = 0; i < number_of_oprnds; i++) |
284 | { | |
b0b4483e RB |
285 | again: |
286 | if (first_op_cond) | |
f7e531cf | 287 | { |
b0b4483e RB |
288 | if (i == 0 || i == 1) |
289 | oprnd = TREE_OPERAND (gimple_op (stmt, first_op_idx), | |
290 | swapped ? !i : i); | |
291 | else | |
292 | oprnd = gimple_op (stmt, first_op_idx + i - 1); | |
f7e531cf IR |
293 | } |
294 | else | |
b0b4483e | 295 | oprnd = gimple_op (stmt, first_op_idx + (swapped ? !i : i)); |
f7e531cf | 296 | |
9771b263 | 297 | oprnd_info = (*oprnds_info)[i]; |
ebfd146a | 298 | |
24ee1384 IR |
299 | if (!vect_is_simple_use (oprnd, NULL, loop_vinfo, bb_vinfo, &def_stmt, |
300 | &def, &dt) | |
d092494c | 301 | || (!def_stmt && dt != vect_constant_def)) |
ebfd146a | 302 | { |
73fbfcad | 303 | if (dump_enabled_p ()) |
ebfd146a | 304 | { |
78c60e3d SS |
305 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
306 | "Build SLP failed: can't find def for "); | |
307 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, oprnd); | |
e645e942 | 308 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a IR |
309 | } |
310 | ||
b0b4483e | 311 | return -1; |
ebfd146a IR |
312 | } |
313 | ||
a70d6342 | 314 | /* Check if DEF_STMT is a part of a pattern in LOOP and get the def stmt |
ff802fa1 | 315 | from the pattern. Check that all the stmts of the node are in the |
ebfd146a | 316 | pattern. */ |
f5709183 IR |
317 | if (def_stmt && gimple_bb (def_stmt) |
318 | && ((loop && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
319 | || (!loop && gimple_bb (def_stmt) == BB_VINFO_BB (bb_vinfo) | |
320 | && gimple_code (def_stmt) != GIMPLE_PHI)) | |
ebfd146a | 321 | && vinfo_for_stmt (def_stmt) |
83197f37 | 322 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (def_stmt)) |
f5709183 IR |
323 | && !STMT_VINFO_RELEVANT (vinfo_for_stmt (def_stmt)) |
324 | && !STMT_VINFO_LIVE_P (vinfo_for_stmt (def_stmt))) | |
ebfd146a | 325 | { |
d092494c IR |
326 | pattern = true; |
327 | if (!first && !oprnd_info->first_pattern) | |
328 | { | |
b0b4483e RB |
329 | if (i == 0 |
330 | && !swapped | |
331 | && commutative) | |
332 | { | |
333 | swapped = true; | |
334 | goto again; | |
335 | } | |
336 | ||
73fbfcad | 337 | if (dump_enabled_p ()) |
d092494c | 338 | { |
78c60e3d SS |
339 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
340 | "Build SLP failed: some of the stmts" | |
341 | " are in a pattern, and others are not "); | |
342 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, oprnd); | |
e645e942 | 343 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
d092494c | 344 | } |
ebfd146a | 345 | |
b0b4483e | 346 | return 1; |
ebfd146a IR |
347 | } |
348 | ||
349 | def_stmt = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (def_stmt)); | |
d092494c | 350 | dt = STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt)); |
ebfd146a | 351 | |
f7e531cf | 352 | if (dt == vect_unknown_def_type) |
ebfd146a | 353 | { |
73fbfcad | 354 | if (dump_enabled_p ()) |
78c60e3d | 355 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 356 | "Unsupported pattern.\n"); |
b0b4483e | 357 | return -1; |
ebfd146a IR |
358 | } |
359 | ||
360 | switch (gimple_code (def_stmt)) | |
361 | { | |
362 | case GIMPLE_PHI: | |
d092494c | 363 | def = gimple_phi_result (def_stmt); |
ebfd146a IR |
364 | break; |
365 | ||
366 | case GIMPLE_ASSIGN: | |
d092494c | 367 | def = gimple_assign_lhs (def_stmt); |
ebfd146a IR |
368 | break; |
369 | ||
370 | default: | |
73fbfcad | 371 | if (dump_enabled_p ()) |
78c60e3d | 372 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 373 | "unsupported defining stmt:\n"); |
b0b4483e | 374 | return -1; |
ebfd146a IR |
375 | } |
376 | } | |
377 | ||
d092494c | 378 | if (first) |
ebfd146a | 379 | { |
d092494c IR |
380 | oprnd_info->first_dt = dt; |
381 | oprnd_info->first_pattern = pattern; | |
793d9a16 | 382 | oprnd_info->first_op_type = TREE_TYPE (oprnd); |
ebfd146a | 383 | } |
ebfd146a IR |
384 | else |
385 | { | |
d092494c IR |
386 | /* Not first stmt of the group, check that the def-stmt/s match |
387 | the def-stmt/s of the first stmt. Allow different definition | |
388 | types for reduction chains: the first stmt must be a | |
389 | vect_reduction_def (a phi node), and the rest | |
390 | vect_internal_def. */ | |
391 | if (((oprnd_info->first_dt != dt | |
392 | && !(oprnd_info->first_dt == vect_reduction_def | |
793d9a16 RB |
393 | && dt == vect_internal_def) |
394 | && !((oprnd_info->first_dt == vect_external_def | |
395 | || oprnd_info->first_dt == vect_constant_def) | |
396 | && (dt == vect_external_def | |
397 | || dt == vect_constant_def))) | |
398 | || !types_compatible_p (oprnd_info->first_op_type, | |
399 | TREE_TYPE (oprnd)))) | |
ebfd146a | 400 | { |
b0b4483e RB |
401 | /* Try swapping operands if we got a mismatch. */ |
402 | if (i == 0 | |
403 | && !swapped | |
404 | && commutative) | |
405 | { | |
406 | swapped = true; | |
407 | goto again; | |
408 | } | |
409 | ||
abf9bfbc RB |
410 | if (dump_enabled_p ()) |
411 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 412 | "Build SLP failed: different types\n"); |
d092494c | 413 | |
b0b4483e | 414 | return 1; |
ebfd146a IR |
415 | } |
416 | } | |
417 | ||
418 | /* Check the types of the definitions. */ | |
d092494c | 419 | switch (dt) |
ebfd146a IR |
420 | { |
421 | case vect_constant_def: | |
8644a673 | 422 | case vect_external_def: |
d092494c | 423 | case vect_reduction_def: |
ebfd146a | 424 | break; |
b8698a0f | 425 | |
8644a673 | 426 | case vect_internal_def: |
abf9bfbc | 427 | oprnd_info->def_stmts.quick_push (def_stmt); |
ebfd146a IR |
428 | break; |
429 | ||
430 | default: | |
431 | /* FORNOW: Not supported. */ | |
73fbfcad | 432 | if (dump_enabled_p ()) |
ebfd146a | 433 | { |
78c60e3d SS |
434 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
435 | "Build SLP failed: illegal type of def "); | |
436 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, def); | |
e645e942 | 437 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a IR |
438 | } |
439 | ||
b0b4483e | 440 | return -1; |
ebfd146a IR |
441 | } |
442 | } | |
443 | ||
b0b4483e RB |
444 | /* Swap operands. */ |
445 | if (swapped) | |
446 | { | |
447 | if (first_op_cond) | |
448 | { | |
449 | tree cond = gimple_assign_rhs1 (stmt); | |
450 | swap_ssa_operands (stmt, &TREE_OPERAND (cond, 0), | |
451 | &TREE_OPERAND (cond, 1)); | |
452 | TREE_SET_CODE (cond, swap_tree_comparison (TREE_CODE (cond))); | |
453 | } | |
454 | else | |
455 | swap_ssa_operands (stmt, gimple_assign_rhs1_ptr (stmt), | |
456 | gimple_assign_rhs2_ptr (stmt)); | |
457 | } | |
458 | ||
459 | return 0; | |
ebfd146a IR |
460 | } |
461 | ||
462 | ||
6983e6b5 RB |
463 | /* Verify if the scalar stmts STMTS are isomorphic, require data |
464 | permutation or are of unsupported types of operation. Return | |
465 | true if they are, otherwise return false and indicate in *MATCHES | |
466 | which stmts are not isomorphic to the first one. If MATCHES[0] | |
467 | is false then this indicates the comparison could not be | |
468 | carried out or the stmts will never be vectorized by SLP. */ | |
ebfd146a IR |
469 | |
470 | static bool | |
6983e6b5 RB |
471 | vect_build_slp_tree_1 (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo, |
472 | vec<gimple> stmts, unsigned int group_size, | |
473 | unsigned nops, unsigned int *max_nunits, | |
474 | unsigned int vectorization_factor, bool *matches) | |
ebfd146a | 475 | { |
ebfd146a | 476 | unsigned int i; |
9771b263 | 477 | gimple stmt = stmts[0]; |
2200fc49 | 478 | enum tree_code first_stmt_code = ERROR_MARK, rhs_code = ERROR_MARK; |
f7e531cf | 479 | enum tree_code first_cond_code = ERROR_MARK; |
ebfd146a | 480 | tree lhs; |
6983e6b5 | 481 | bool need_same_oprnds = false; |
ebfd146a | 482 | tree vectype, scalar_type, first_op1 = NULL_TREE; |
ebfd146a IR |
483 | optab optab; |
484 | int icode; | |
ef4bddc2 RS |
485 | machine_mode optab_op2_mode; |
486 | machine_mode vec_mode; | |
ebfd146a | 487 | struct data_reference *first_dr; |
ebfd146a | 488 | HOST_WIDE_INT dummy; |
c3e7ee41 | 489 | gimple first_load = NULL, prev_first_load = NULL, old_first_load = NULL; |
f7e531cf | 490 | tree cond; |
d092494c | 491 | |
ebfd146a | 492 | /* For every stmt in NODE find its def stmt/s. */ |
9771b263 | 493 | FOR_EACH_VEC_ELT (stmts, i, stmt) |
ebfd146a | 494 | { |
6983e6b5 RB |
495 | matches[i] = false; |
496 | ||
73fbfcad | 497 | if (dump_enabled_p ()) |
ebfd146a | 498 | { |
78c60e3d SS |
499 | dump_printf_loc (MSG_NOTE, vect_location, "Build SLP for "); |
500 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 501 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
502 | } |
503 | ||
4b5caab7 IR |
504 | /* Fail to vectorize statements marked as unvectorizable. */ |
505 | if (!STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (stmt))) | |
506 | { | |
73fbfcad | 507 | if (dump_enabled_p ()) |
4b5caab7 | 508 | { |
78c60e3d SS |
509 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
510 | "Build SLP failed: unvectorizable statement "); | |
511 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 512 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4b5caab7 | 513 | } |
6983e6b5 RB |
514 | /* Fatal mismatch. */ |
515 | matches[0] = false; | |
4b5caab7 IR |
516 | return false; |
517 | } | |
518 | ||
ebfd146a IR |
519 | lhs = gimple_get_lhs (stmt); |
520 | if (lhs == NULL_TREE) | |
521 | { | |
73fbfcad | 522 | if (dump_enabled_p ()) |
ebfd146a | 523 | { |
78c60e3d SS |
524 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
525 | "Build SLP failed: not GIMPLE_ASSIGN nor " | |
526 | "GIMPLE_CALL "); | |
527 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 528 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a | 529 | } |
6983e6b5 RB |
530 | /* Fatal mismatch. */ |
531 | matches[0] = false; | |
ebfd146a IR |
532 | return false; |
533 | } | |
534 | ||
f7e531cf IR |
535 | if (is_gimple_assign (stmt) |
536 | && gimple_assign_rhs_code (stmt) == COND_EXPR | |
537 | && (cond = gimple_assign_rhs1 (stmt)) | |
538 | && !COMPARISON_CLASS_P (cond)) | |
539 | { | |
73fbfcad | 540 | if (dump_enabled_p ()) |
f7e531cf | 541 | { |
78c60e3d SS |
542 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
543 | "Build SLP failed: condition is not " | |
544 | "comparison "); | |
545 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 546 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
f7e531cf | 547 | } |
6983e6b5 RB |
548 | /* Fatal mismatch. */ |
549 | matches[0] = false; | |
f7e531cf IR |
550 | return false; |
551 | } | |
552 | ||
b8698a0f | 553 | scalar_type = vect_get_smallest_scalar_type (stmt, &dummy, &dummy); |
ebfd146a IR |
554 | vectype = get_vectype_for_scalar_type (scalar_type); |
555 | if (!vectype) | |
556 | { | |
73fbfcad | 557 | if (dump_enabled_p ()) |
ebfd146a | 558 | { |
78c60e3d SS |
559 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
560 | "Build SLP failed: unsupported data-type "); | |
561 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
562 | scalar_type); | |
e645e942 | 563 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a | 564 | } |
6983e6b5 RB |
565 | /* Fatal mismatch. */ |
566 | matches[0] = false; | |
ebfd146a IR |
567 | return false; |
568 | } | |
b8698a0f | 569 | |
4ef69dfc IR |
570 | /* In case of multiple types we need to detect the smallest type. */ |
571 | if (*max_nunits < TYPE_VECTOR_SUBPARTS (vectype)) | |
a70d6342 | 572 | { |
4ef69dfc IR |
573 | *max_nunits = TYPE_VECTOR_SUBPARTS (vectype); |
574 | if (bb_vinfo) | |
575 | vectorization_factor = *max_nunits; | |
a70d6342 | 576 | } |
b8698a0f | 577 | |
538dd0b7 | 578 | if (gcall *call_stmt = dyn_cast <gcall *> (stmt)) |
190c2236 JJ |
579 | { |
580 | rhs_code = CALL_EXPR; | |
538dd0b7 DM |
581 | if (gimple_call_internal_p (call_stmt) |
582 | || gimple_call_tail_p (call_stmt) | |
583 | || gimple_call_noreturn_p (call_stmt) | |
584 | || !gimple_call_nothrow_p (call_stmt) | |
585 | || gimple_call_chain (call_stmt)) | |
190c2236 | 586 | { |
73fbfcad | 587 | if (dump_enabled_p ()) |
190c2236 | 588 | { |
78c60e3d SS |
589 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
590 | "Build SLP failed: unsupported call type "); | |
538dd0b7 DM |
591 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, |
592 | call_stmt, 0); | |
e645e942 | 593 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
190c2236 | 594 | } |
6983e6b5 RB |
595 | /* Fatal mismatch. */ |
596 | matches[0] = false; | |
190c2236 JJ |
597 | return false; |
598 | } | |
599 | } | |
ebfd146a IR |
600 | else |
601 | rhs_code = gimple_assign_rhs_code (stmt); | |
602 | ||
603 | /* Check the operation. */ | |
604 | if (i == 0) | |
605 | { | |
606 | first_stmt_code = rhs_code; | |
607 | ||
b8698a0f | 608 | /* Shift arguments should be equal in all the packed stmts for a |
ebfd146a IR |
609 | vector shift with scalar shift operand. */ |
610 | if (rhs_code == LSHIFT_EXPR || rhs_code == RSHIFT_EXPR | |
611 | || rhs_code == LROTATE_EXPR | |
612 | || rhs_code == RROTATE_EXPR) | |
613 | { | |
614 | vec_mode = TYPE_MODE (vectype); | |
615 | ||
616 | /* First see if we have a vector/vector shift. */ | |
617 | optab = optab_for_tree_code (rhs_code, vectype, | |
618 | optab_vector); | |
619 | ||
620 | if (!optab | |
947131ba | 621 | || optab_handler (optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
622 | { |
623 | /* No vector/vector shift, try for a vector/scalar shift. */ | |
624 | optab = optab_for_tree_code (rhs_code, vectype, | |
625 | optab_scalar); | |
626 | ||
627 | if (!optab) | |
628 | { | |
73fbfcad | 629 | if (dump_enabled_p ()) |
78c60e3d | 630 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 631 | "Build SLP failed: no optab.\n"); |
6983e6b5 RB |
632 | /* Fatal mismatch. */ |
633 | matches[0] = false; | |
ebfd146a IR |
634 | return false; |
635 | } | |
947131ba | 636 | icode = (int) optab_handler (optab, vec_mode); |
ebfd146a IR |
637 | if (icode == CODE_FOR_nothing) |
638 | { | |
73fbfcad | 639 | if (dump_enabled_p ()) |
78c60e3d SS |
640 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
641 | "Build SLP failed: " | |
e645e942 | 642 | "op not supported by target.\n"); |
6983e6b5 RB |
643 | /* Fatal mismatch. */ |
644 | matches[0] = false; | |
ebfd146a IR |
645 | return false; |
646 | } | |
647 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
648 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
649 | { | |
650 | need_same_oprnds = true; | |
651 | first_op1 = gimple_assign_rhs2 (stmt); | |
652 | } | |
653 | } | |
654 | } | |
36ba4aae IR |
655 | else if (rhs_code == WIDEN_LSHIFT_EXPR) |
656 | { | |
657 | need_same_oprnds = true; | |
658 | first_op1 = gimple_assign_rhs2 (stmt); | |
659 | } | |
ebfd146a IR |
660 | } |
661 | else | |
662 | { | |
663 | if (first_stmt_code != rhs_code | |
664 | && (first_stmt_code != IMAGPART_EXPR | |
665 | || rhs_code != REALPART_EXPR) | |
666 | && (first_stmt_code != REALPART_EXPR | |
69f11a13 | 667 | || rhs_code != IMAGPART_EXPR) |
0d0293ac | 668 | && !(STMT_VINFO_GROUPED_ACCESS (vinfo_for_stmt (stmt)) |
69f11a13 | 669 | && (first_stmt_code == ARRAY_REF |
38000232 | 670 | || first_stmt_code == BIT_FIELD_REF |
69f11a13 IR |
671 | || first_stmt_code == INDIRECT_REF |
672 | || first_stmt_code == COMPONENT_REF | |
673 | || first_stmt_code == MEM_REF))) | |
ebfd146a | 674 | { |
73fbfcad | 675 | if (dump_enabled_p ()) |
ebfd146a | 676 | { |
78c60e3d SS |
677 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
678 | "Build SLP failed: different operation " | |
679 | "in stmt "); | |
680 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 681 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a | 682 | } |
6983e6b5 RB |
683 | /* Mismatch. */ |
684 | continue; | |
ebfd146a | 685 | } |
b8698a0f L |
686 | |
687 | if (need_same_oprnds | |
ebfd146a IR |
688 | && !operand_equal_p (first_op1, gimple_assign_rhs2 (stmt), 0)) |
689 | { | |
73fbfcad | 690 | if (dump_enabled_p ()) |
ebfd146a | 691 | { |
78c60e3d SS |
692 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
693 | "Build SLP failed: different shift " | |
694 | "arguments in "); | |
695 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 696 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a | 697 | } |
6983e6b5 RB |
698 | /* Mismatch. */ |
699 | continue; | |
ebfd146a | 700 | } |
190c2236 JJ |
701 | |
702 | if (rhs_code == CALL_EXPR) | |
703 | { | |
9771b263 | 704 | gimple first_stmt = stmts[0]; |
190c2236 JJ |
705 | if (gimple_call_num_args (stmt) != nops |
706 | || !operand_equal_p (gimple_call_fn (first_stmt), | |
707 | gimple_call_fn (stmt), 0) | |
708 | || gimple_call_fntype (first_stmt) | |
709 | != gimple_call_fntype (stmt)) | |
710 | { | |
73fbfcad | 711 | if (dump_enabled_p ()) |
190c2236 | 712 | { |
78c60e3d SS |
713 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
714 | "Build SLP failed: different calls in "); | |
715 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
716 | stmt, 0); | |
e645e942 | 717 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
190c2236 | 718 | } |
6983e6b5 RB |
719 | /* Mismatch. */ |
720 | continue; | |
190c2236 JJ |
721 | } |
722 | } | |
ebfd146a IR |
723 | } |
724 | ||
0d0293ac MM |
725 | /* Grouped store or load. */ |
726 | if (STMT_VINFO_GROUPED_ACCESS (vinfo_for_stmt (stmt))) | |
ebfd146a IR |
727 | { |
728 | if (REFERENCE_CLASS_P (lhs)) | |
729 | { | |
730 | /* Store. */ | |
6983e6b5 | 731 | ; |
ebfd146a | 732 | } |
b5aeb3bb IR |
733 | else |
734 | { | |
735 | /* Load. */ | |
314f64eb RB |
736 | unsigned unrolling_factor |
737 | = least_common_multiple | |
738 | (*max_nunits, group_size) / group_size; | |
a64b9c26 RB |
739 | /* FORNOW: Check that there is no gap between the loads |
740 | and no gap between the groups when we need to load | |
741 | multiple groups at once. | |
742 | ??? We should enhance this to only disallow gaps | |
743 | inside vectors. */ | |
314f64eb | 744 | if ((unrolling_factor > 1 |
7ef95f9c RB |
745 | && ((GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) == stmt |
746 | && GROUP_GAP (vinfo_for_stmt (stmt)) != 0) | |
747 | /* If the group is split up then GROUP_GAP | |
748 | isn't correct here, nor is GROUP_FIRST_ELEMENT. */ | |
749 | || GROUP_SIZE (vinfo_for_stmt (stmt)) > group_size)) | |
a64b9c26 RB |
750 | || (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) != stmt |
751 | && GROUP_GAP (vinfo_for_stmt (stmt)) != 1)) | |
b5aeb3bb | 752 | { |
73fbfcad | 753 | if (dump_enabled_p ()) |
b5aeb3bb | 754 | { |
78c60e3d SS |
755 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
756 | "Build SLP failed: grouped " | |
757 | "loads have gaps "); | |
758 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
759 | stmt, 0); | |
e645e942 | 760 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
b5aeb3bb | 761 | } |
6983e6b5 RB |
762 | /* Fatal mismatch. */ |
763 | matches[0] = false; | |
b5aeb3bb IR |
764 | return false; |
765 | } | |
2f0fa28e | 766 | |
b5aeb3bb IR |
767 | /* Check that the size of interleaved loads group is not |
768 | greater than the SLP group size. */ | |
314f64eb RB |
769 | unsigned ncopies |
770 | = vectorization_factor / TYPE_VECTOR_SUBPARTS (vectype); | |
6aa904c4 | 771 | if (loop_vinfo |
a64b9c26 RB |
772 | && GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) == stmt |
773 | && ((GROUP_SIZE (vinfo_for_stmt (stmt)) | |
774 | - GROUP_GAP (vinfo_for_stmt (stmt))) | |
775 | > ncopies * group_size)) | |
b5aeb3bb | 776 | { |
73fbfcad | 777 | if (dump_enabled_p ()) |
b5aeb3bb | 778 | { |
78c60e3d SS |
779 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
780 | "Build SLP failed: the number " | |
781 | "of interleaved loads is greater than " | |
782 | "the SLP group size "); | |
783 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
784 | stmt, 0); | |
e645e942 | 785 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
b5aeb3bb | 786 | } |
6983e6b5 RB |
787 | /* Fatal mismatch. */ |
788 | matches[0] = false; | |
b5aeb3bb IR |
789 | return false; |
790 | } | |
791 | ||
c3e7ee41 | 792 | old_first_load = first_load; |
e14c1050 | 793 | first_load = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)); |
b5aeb3bb IR |
794 | if (prev_first_load) |
795 | { | |
796 | /* Check that there are no loads from different interleaving | |
6983e6b5 RB |
797 | chains in the same node. */ |
798 | if (prev_first_load != first_load) | |
78c60e3d | 799 | { |
73fbfcad | 800 | if (dump_enabled_p ()) |
b5aeb3bb | 801 | { |
78c60e3d SS |
802 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, |
803 | vect_location, | |
804 | "Build SLP failed: different " | |
805 | "interleaving chains in one node "); | |
806 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
807 | stmt, 0); | |
e645e942 | 808 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
b5aeb3bb | 809 | } |
6983e6b5 RB |
810 | /* Mismatch. */ |
811 | continue; | |
b5aeb3bb IR |
812 | } |
813 | } | |
814 | else | |
815 | prev_first_load = first_load; | |
b8698a0f | 816 | |
c3e7ee41 BS |
817 | /* In some cases a group of loads is just the same load |
818 | repeated N times. Only analyze its cost once. */ | |
819 | if (first_load == stmt && old_first_load != first_load) | |
ebfd146a IR |
820 | { |
821 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt)); | |
720f5239 | 822 | if (vect_supportable_dr_alignment (first_dr, false) |
ebfd146a IR |
823 | == dr_unaligned_unsupported) |
824 | { | |
73fbfcad | 825 | if (dump_enabled_p ()) |
ebfd146a | 826 | { |
78c60e3d SS |
827 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, |
828 | vect_location, | |
829 | "Build SLP failed: unsupported " | |
830 | "unaligned load "); | |
831 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
832 | stmt, 0); | |
e645e942 | 833 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a | 834 | } |
6983e6b5 RB |
835 | /* Fatal mismatch. */ |
836 | matches[0] = false; | |
ebfd146a IR |
837 | return false; |
838 | } | |
ebfd146a | 839 | } |
ebfd146a | 840 | } |
0d0293ac | 841 | } /* Grouped access. */ |
ebfd146a IR |
842 | else |
843 | { | |
844 | if (TREE_CODE_CLASS (rhs_code) == tcc_reference) | |
845 | { | |
0d0293ac | 846 | /* Not grouped load. */ |
73fbfcad | 847 | if (dump_enabled_p ()) |
ebfd146a | 848 | { |
78c60e3d SS |
849 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
850 | "Build SLP failed: not grouped load "); | |
851 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 852 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a IR |
853 | } |
854 | ||
0d0293ac | 855 | /* FORNOW: Not grouped loads are not supported. */ |
6983e6b5 RB |
856 | /* Fatal mismatch. */ |
857 | matches[0] = false; | |
ebfd146a IR |
858 | return false; |
859 | } | |
860 | ||
861 | /* Not memory operation. */ | |
862 | if (TREE_CODE_CLASS (rhs_code) != tcc_binary | |
f7e531cf | 863 | && TREE_CODE_CLASS (rhs_code) != tcc_unary |
190c2236 JJ |
864 | && rhs_code != COND_EXPR |
865 | && rhs_code != CALL_EXPR) | |
ebfd146a | 866 | { |
73fbfcad | 867 | if (dump_enabled_p ()) |
ebfd146a | 868 | { |
78c60e3d SS |
869 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
870 | "Build SLP failed: operation"); | |
871 | dump_printf (MSG_MISSED_OPTIMIZATION, " unsupported "); | |
872 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 873 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a | 874 | } |
6983e6b5 RB |
875 | /* Fatal mismatch. */ |
876 | matches[0] = false; | |
ebfd146a IR |
877 | return false; |
878 | } | |
879 | ||
f7e531cf IR |
880 | if (rhs_code == COND_EXPR) |
881 | { | |
882 | tree cond_expr = gimple_assign_rhs1 (stmt); | |
883 | ||
884 | if (i == 0) | |
885 | first_cond_code = TREE_CODE (cond_expr); | |
886 | else if (first_cond_code != TREE_CODE (cond_expr)) | |
887 | { | |
73fbfcad | 888 | if (dump_enabled_p ()) |
f7e531cf | 889 | { |
78c60e3d SS |
890 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
891 | "Build SLP failed: different" | |
892 | " operation"); | |
893 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
894 | stmt, 0); | |
e645e942 | 895 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
f7e531cf | 896 | } |
6983e6b5 RB |
897 | /* Mismatch. */ |
898 | continue; | |
f7e531cf IR |
899 | } |
900 | } | |
ebfd146a | 901 | } |
6983e6b5 RB |
902 | |
903 | matches[i] = true; | |
904 | } | |
905 | ||
906 | for (i = 0; i < group_size; ++i) | |
907 | if (!matches[i]) | |
908 | return false; | |
909 | ||
910 | return true; | |
911 | } | |
912 | ||
913 | /* Recursively build an SLP tree starting from NODE. | |
914 | Fail (and return a value not equal to zero) if def-stmts are not | |
915 | isomorphic, require data permutation or are of unsupported types of | |
916 | operation. Otherwise, return 0. | |
917 | The value returned is the depth in the SLP tree where a mismatch | |
918 | was found. */ | |
919 | ||
920 | static bool | |
921 | vect_build_slp_tree (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo, | |
922 | slp_tree *node, unsigned int group_size, | |
923 | unsigned int *max_nunits, | |
924 | vec<slp_tree> *loads, | |
925 | unsigned int vectorization_factor, | |
1428105c RB |
926 | bool *matches, unsigned *npermutes, unsigned *tree_size, |
927 | unsigned max_tree_size) | |
6983e6b5 | 928 | { |
1428105c | 929 | unsigned nops, i, this_npermutes = 0, this_tree_size = 0; |
6983e6b5 RB |
930 | gimple stmt; |
931 | ||
932 | if (!matches) | |
933 | matches = XALLOCAVEC (bool, group_size); | |
934 | if (!npermutes) | |
935 | npermutes = &this_npermutes; | |
936 | ||
937 | matches[0] = false; | |
938 | ||
939 | stmt = SLP_TREE_SCALAR_STMTS (*node)[0]; | |
940 | if (is_gimple_call (stmt)) | |
941 | nops = gimple_call_num_args (stmt); | |
942 | else if (is_gimple_assign (stmt)) | |
943 | { | |
944 | nops = gimple_num_ops (stmt) - 1; | |
945 | if (gimple_assign_rhs_code (stmt) == COND_EXPR) | |
946 | nops++; | |
ebfd146a | 947 | } |
6983e6b5 RB |
948 | else |
949 | return false; | |
950 | ||
951 | if (!vect_build_slp_tree_1 (loop_vinfo, bb_vinfo, | |
952 | SLP_TREE_SCALAR_STMTS (*node), group_size, nops, | |
953 | max_nunits, vectorization_factor, matches)) | |
954 | return false; | |
ebfd146a | 955 | |
6983e6b5 RB |
956 | /* If the SLP node is a load, terminate the recursion. */ |
957 | if (STMT_VINFO_GROUPED_ACCESS (vinfo_for_stmt (stmt)) | |
958 | && DR_IS_READ (STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt)))) | |
ebfd146a | 959 | { |
9771b263 | 960 | loads->safe_push (*node); |
ebfd146a IR |
961 | return true; |
962 | } | |
963 | ||
6983e6b5 RB |
964 | /* Get at the operands, verifying they are compatible. */ |
965 | vec<slp_oprnd_info> oprnds_info = vect_create_oprnd_info (nops, group_size); | |
966 | slp_oprnd_info oprnd_info; | |
967 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (*node), i, stmt) | |
968 | { | |
b0b4483e RB |
969 | switch (vect_get_and_check_slp_defs (loop_vinfo, bb_vinfo, |
970 | stmt, (i == 0), &oprnds_info)) | |
6983e6b5 | 971 | { |
b0b4483e RB |
972 | case 0: |
973 | break; | |
974 | case -1: | |
975 | matches[0] = false; | |
6983e6b5 RB |
976 | vect_free_oprnd_info (oprnds_info); |
977 | return false; | |
b0b4483e RB |
978 | case 1: |
979 | matches[i] = false; | |
980 | break; | |
6983e6b5 RB |
981 | } |
982 | } | |
b0b4483e RB |
983 | for (i = 0; i < group_size; ++i) |
984 | if (!matches[i]) | |
985 | { | |
986 | vect_free_oprnd_info (oprnds_info); | |
987 | return false; | |
988 | } | |
6983e6b5 RB |
989 | |
990 | stmt = SLP_TREE_SCALAR_STMTS (*node)[0]; | |
991 | ||
b8698a0f | 992 | /* Create SLP_TREE nodes for the definition node/s. */ |
9771b263 | 993 | FOR_EACH_VEC_ELT (oprnds_info, i, oprnd_info) |
ebfd146a | 994 | { |
d092494c | 995 | slp_tree child; |
6983e6b5 RB |
996 | unsigned old_nloads = loads->length (); |
997 | unsigned old_max_nunits = *max_nunits; | |
b8698a0f | 998 | |
d092494c IR |
999 | if (oprnd_info->first_dt != vect_internal_def) |
1000 | continue; | |
ebfd146a | 1001 | |
1428105c RB |
1002 | if (++this_tree_size > max_tree_size) |
1003 | { | |
1004 | vect_free_oprnd_info (oprnds_info); | |
1005 | return false; | |
1006 | } | |
1007 | ||
d092494c | 1008 | child = vect_create_new_slp_node (oprnd_info->def_stmts); |
6983e6b5 RB |
1009 | if (!child) |
1010 | { | |
9771b263 | 1011 | vect_free_oprnd_info (oprnds_info); |
6983e6b5 | 1012 | return false; |
d092494c | 1013 | } |
b8698a0f | 1014 | |
6983e6b5 RB |
1015 | bool *matches = XALLOCAVEC (bool, group_size); |
1016 | if (vect_build_slp_tree (loop_vinfo, bb_vinfo, &child, | |
1017 | group_size, max_nunits, loads, | |
1428105c RB |
1018 | vectorization_factor, matches, |
1019 | npermutes, &this_tree_size, max_tree_size)) | |
6983e6b5 RB |
1020 | { |
1021 | oprnd_info->def_stmts = vNULL; | |
1022 | SLP_TREE_CHILDREN (*node).quick_push (child); | |
1023 | continue; | |
1024 | } | |
1025 | ||
1026 | /* If the SLP build for operand zero failed and operand zero | |
1027 | and one can be commutated try that for the scalar stmts | |
1028 | that failed the match. */ | |
1029 | if (i == 0 | |
1030 | /* A first scalar stmt mismatch signals a fatal mismatch. */ | |
1031 | && matches[0] | |
1032 | /* ??? For COND_EXPRs we can swap the comparison operands | |
1033 | as well as the arms under some constraints. */ | |
1034 | && nops == 2 | |
1035 | && oprnds_info[1]->first_dt == vect_internal_def | |
1036 | && is_gimple_assign (stmt) | |
1037 | && commutative_tree_code (gimple_assign_rhs_code (stmt)) | |
1038 | /* Do so only if the number of not successful permutes was nor more | |
1039 | than a cut-ff as re-trying the recursive match on | |
1040 | possibly each level of the tree would expose exponential | |
1041 | behavior. */ | |
1042 | && *npermutes < 4) | |
1043 | { | |
1044 | /* Roll back. */ | |
1045 | *max_nunits = old_max_nunits; | |
1046 | loads->truncate (old_nloads); | |
1047 | /* Swap mismatched definition stmts. */ | |
b0b4483e RB |
1048 | dump_printf_loc (MSG_NOTE, vect_location, |
1049 | "Re-trying with swapped operands of stmts "); | |
6983e6b5 RB |
1050 | for (unsigned j = 0; j < group_size; ++j) |
1051 | if (!matches[j]) | |
1052 | { | |
1053 | gimple tem = oprnds_info[0]->def_stmts[j]; | |
1054 | oprnds_info[0]->def_stmts[j] = oprnds_info[1]->def_stmts[j]; | |
1055 | oprnds_info[1]->def_stmts[j] = tem; | |
b0b4483e | 1056 | dump_printf (MSG_NOTE, "%d ", j); |
6983e6b5 | 1057 | } |
b0b4483e | 1058 | dump_printf (MSG_NOTE, "\n"); |
6983e6b5 RB |
1059 | /* And try again ... */ |
1060 | if (vect_build_slp_tree (loop_vinfo, bb_vinfo, &child, | |
1061 | group_size, max_nunits, loads, | |
1062 | vectorization_factor, | |
1428105c RB |
1063 | matches, npermutes, &this_tree_size, |
1064 | max_tree_size)) | |
6983e6b5 RB |
1065 | { |
1066 | oprnd_info->def_stmts = vNULL; | |
1067 | SLP_TREE_CHILDREN (*node).quick_push (child); | |
1068 | continue; | |
1069 | } | |
1070 | ||
1071 | ++*npermutes; | |
1072 | } | |
1073 | ||
1074 | oprnd_info->def_stmts = vNULL; | |
1075 | vect_free_slp_tree (child); | |
1076 | vect_free_oprnd_info (oprnds_info); | |
1077 | return false; | |
ebfd146a IR |
1078 | } |
1079 | ||
1428105c RB |
1080 | if (tree_size) |
1081 | *tree_size += this_tree_size; | |
1082 | ||
9771b263 | 1083 | vect_free_oprnd_info (oprnds_info); |
ebfd146a IR |
1084 | return true; |
1085 | } | |
1086 | ||
78c60e3d | 1087 | /* Dump a slp tree NODE using flags specified in DUMP_KIND. */ |
ebfd146a IR |
1088 | |
1089 | static void | |
78c60e3d | 1090 | vect_print_slp_tree (int dump_kind, slp_tree node) |
ebfd146a IR |
1091 | { |
1092 | int i; | |
1093 | gimple stmt; | |
d755c7ef | 1094 | slp_tree child; |
ebfd146a IR |
1095 | |
1096 | if (!node) | |
1097 | return; | |
1098 | ||
78c60e3d | 1099 | dump_printf (dump_kind, "node "); |
9771b263 | 1100 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt) |
ebfd146a | 1101 | { |
78c60e3d SS |
1102 | dump_printf (dump_kind, "\n\tstmt %d ", i); |
1103 | dump_gimple_stmt (dump_kind, TDF_SLIM, stmt, 0); | |
ebfd146a | 1104 | } |
78c60e3d | 1105 | dump_printf (dump_kind, "\n"); |
ebfd146a | 1106 | |
9771b263 | 1107 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) |
d755c7ef | 1108 | vect_print_slp_tree (dump_kind, child); |
ebfd146a IR |
1109 | } |
1110 | ||
1111 | ||
b8698a0f L |
1112 | /* Mark the tree rooted at NODE with MARK (PURE_SLP or HYBRID). |
1113 | If MARK is HYBRID, it refers to a specific stmt in NODE (the stmt at index | |
ff802fa1 | 1114 | J). Otherwise, MARK is PURE_SLP and J is -1, which indicates that all the |
ebfd146a IR |
1115 | stmts in NODE are to be marked. */ |
1116 | ||
1117 | static void | |
1118 | vect_mark_slp_stmts (slp_tree node, enum slp_vect_type mark, int j) | |
1119 | { | |
1120 | int i; | |
1121 | gimple stmt; | |
d755c7ef | 1122 | slp_tree child; |
ebfd146a IR |
1123 | |
1124 | if (!node) | |
1125 | return; | |
1126 | ||
9771b263 | 1127 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt) |
ebfd146a IR |
1128 | if (j < 0 || i == j) |
1129 | STMT_SLP_TYPE (vinfo_for_stmt (stmt)) = mark; | |
1130 | ||
9771b263 | 1131 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) |
d755c7ef | 1132 | vect_mark_slp_stmts (child, mark, j); |
ebfd146a IR |
1133 | } |
1134 | ||
1135 | ||
a70d6342 IR |
1136 | /* Mark the statements of the tree rooted at NODE as relevant (vect_used). */ |
1137 | ||
1138 | static void | |
1139 | vect_mark_slp_stmts_relevant (slp_tree node) | |
1140 | { | |
1141 | int i; | |
1142 | gimple stmt; | |
1143 | stmt_vec_info stmt_info; | |
d755c7ef | 1144 | slp_tree child; |
a70d6342 IR |
1145 | |
1146 | if (!node) | |
1147 | return; | |
1148 | ||
9771b263 | 1149 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt) |
a70d6342 IR |
1150 | { |
1151 | stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 1152 | gcc_assert (!STMT_VINFO_RELEVANT (stmt_info) |
a70d6342 IR |
1153 | || STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_scope); |
1154 | STMT_VINFO_RELEVANT (stmt_info) = vect_used_in_scope; | |
1155 | } | |
1156 | ||
9771b263 | 1157 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) |
d755c7ef | 1158 | vect_mark_slp_stmts_relevant (child); |
a70d6342 IR |
1159 | } |
1160 | ||
1161 | ||
b5aeb3bb IR |
1162 | /* Rearrange the statements of NODE according to PERMUTATION. */ |
1163 | ||
1164 | static void | |
1165 | vect_slp_rearrange_stmts (slp_tree node, unsigned int group_size, | |
01d8bf07 | 1166 | vec<unsigned> permutation) |
b5aeb3bb IR |
1167 | { |
1168 | gimple stmt; | |
9771b263 | 1169 | vec<gimple> tmp_stmts; |
d755c7ef RB |
1170 | unsigned int i; |
1171 | slp_tree child; | |
b5aeb3bb | 1172 | |
9771b263 | 1173 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) |
d755c7ef | 1174 | vect_slp_rearrange_stmts (child, group_size, permutation); |
b5aeb3bb | 1175 | |
9771b263 DN |
1176 | gcc_assert (group_size == SLP_TREE_SCALAR_STMTS (node).length ()); |
1177 | tmp_stmts.create (group_size); | |
d755c7ef | 1178 | tmp_stmts.quick_grow_cleared (group_size); |
b5aeb3bb | 1179 | |
9771b263 | 1180 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt) |
d755c7ef | 1181 | tmp_stmts[permutation[i]] = stmt; |
b5aeb3bb | 1182 | |
9771b263 | 1183 | SLP_TREE_SCALAR_STMTS (node).release (); |
b5aeb3bb IR |
1184 | SLP_TREE_SCALAR_STMTS (node) = tmp_stmts; |
1185 | } | |
1186 | ||
1187 | ||
01d8bf07 RB |
1188 | /* Check if the required load permutations in the SLP instance |
1189 | SLP_INSTN are supported. */ | |
ebfd146a IR |
1190 | |
1191 | static bool | |
01d8bf07 | 1192 | vect_supported_load_permutation_p (slp_instance slp_instn) |
ebfd146a | 1193 | { |
01d8bf07 RB |
1194 | unsigned int group_size = SLP_INSTANCE_GROUP_SIZE (slp_instn); |
1195 | unsigned int i, j, k, next; | |
7417f6c0 | 1196 | sbitmap load_index; |
6983e6b5 RB |
1197 | slp_tree node; |
1198 | gimple stmt, load, next_load, first_load; | |
6aa904c4 | 1199 | struct data_reference *dr; |
ebfd146a | 1200 | |
73fbfcad | 1201 | if (dump_enabled_p ()) |
ebfd146a | 1202 | { |
78c60e3d | 1203 | dump_printf_loc (MSG_NOTE, vect_location, "Load permutation "); |
01d8bf07 RB |
1204 | FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node) |
1205 | if (node->load_permutation.exists ()) | |
1206 | FOR_EACH_VEC_ELT (node->load_permutation, j, next) | |
1207 | dump_printf (MSG_NOTE, "%d ", next); | |
1208 | else | |
bddc974e TJ |
1209 | for (k = 0; k < group_size; ++k) |
1210 | dump_printf (MSG_NOTE, "%d ", k); | |
e645e942 | 1211 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1212 | } |
1213 | ||
b5aeb3bb IR |
1214 | /* In case of reduction every load permutation is allowed, since the order |
1215 | of the reduction statements is not important (as opposed to the case of | |
0d0293ac | 1216 | grouped stores). The only condition we need to check is that all the |
b5aeb3bb IR |
1217 | load nodes are of the same size and have the same permutation (and then |
1218 | rearrange all the nodes of the SLP instance according to this | |
1219 | permutation). */ | |
1220 | ||
1221 | /* Check that all the load nodes are of the same size. */ | |
01d8bf07 | 1222 | /* ??? Can't we assert this? */ |
9771b263 | 1223 | FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node) |
6983e6b5 RB |
1224 | if (SLP_TREE_SCALAR_STMTS (node).length () != (unsigned) group_size) |
1225 | return false; | |
2200fc49 | 1226 | |
b5aeb3bb | 1227 | node = SLP_INSTANCE_TREE (slp_instn); |
9771b263 | 1228 | stmt = SLP_TREE_SCALAR_STMTS (node)[0]; |
b5aeb3bb | 1229 | |
b010117a IR |
1230 | /* Reduction (there are no data-refs in the root). |
1231 | In reduction chain the order of the loads is important. */ | |
1232 | if (!STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt)) | |
1233 | && !GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) | |
b5aeb3bb | 1234 | { |
01d8bf07 RB |
1235 | slp_tree load; |
1236 | unsigned int lidx; | |
b5aeb3bb | 1237 | |
01d8bf07 RB |
1238 | /* Compare all the permutation sequences to the first one. We know |
1239 | that at least one load is permuted. */ | |
1240 | node = SLP_INSTANCE_LOADS (slp_instn)[0]; | |
1241 | if (!node->load_permutation.exists ()) | |
1242 | return false; | |
1243 | for (i = 1; SLP_INSTANCE_LOADS (slp_instn).iterate (i, &load); ++i) | |
1244 | { | |
1245 | if (!load->load_permutation.exists ()) | |
1246 | return false; | |
1247 | FOR_EACH_VEC_ELT (load->load_permutation, j, lidx) | |
1248 | if (lidx != node->load_permutation[j]) | |
1249 | return false; | |
1250 | } | |
c9c1e775 | 1251 | |
01d8bf07 RB |
1252 | /* Check that the loads in the first sequence are different and there |
1253 | are no gaps between them. */ | |
1254 | load_index = sbitmap_alloc (group_size); | |
1255 | bitmap_clear (load_index); | |
1256 | FOR_EACH_VEC_ELT (node->load_permutation, i, lidx) | |
1257 | { | |
1258 | if (bitmap_bit_p (load_index, lidx)) | |
1259 | { | |
1260 | sbitmap_free (load_index); | |
1261 | return false; | |
1262 | } | |
1263 | bitmap_set_bit (load_index, lidx); | |
1264 | } | |
1265 | for (i = 0; i < group_size; i++) | |
1266 | if (!bitmap_bit_p (load_index, i)) | |
1267 | { | |
1268 | sbitmap_free (load_index); | |
1269 | return false; | |
1270 | } | |
1271 | sbitmap_free (load_index); | |
1272 | ||
1273 | /* This permutation is valid for reduction. Since the order of the | |
1274 | statements in the nodes is not important unless they are memory | |
1275 | accesses, we can rearrange the statements in all the nodes | |
1276 | according to the order of the loads. */ | |
1277 | vect_slp_rearrange_stmts (SLP_INSTANCE_TREE (slp_instn), group_size, | |
1278 | node->load_permutation); | |
1279 | ||
1280 | /* We are done, no actual permutations need to be generated. */ | |
1281 | FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node) | |
1282 | SLP_TREE_LOAD_PERMUTATION (node).release (); | |
1283 | return true; | |
b5aeb3bb IR |
1284 | } |
1285 | ||
6aa904c4 IR |
1286 | /* In basic block vectorization we allow any subchain of an interleaving |
1287 | chain. | |
1288 | FORNOW: not supported in loop SLP because of realignment compications. */ | |
01d8bf07 | 1289 | if (STMT_VINFO_BB_VINFO (vinfo_for_stmt (stmt))) |
6aa904c4 | 1290 | { |
01d8bf07 RB |
1291 | /* Check that for every node in the instance the loads |
1292 | form a subchain. */ | |
9771b263 | 1293 | FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node) |
6aa904c4 IR |
1294 | { |
1295 | next_load = NULL; | |
9771b263 | 1296 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), j, load) |
6aa904c4 | 1297 | { |
6aa904c4 | 1298 | if (j != 0 && next_load != load) |
01d8bf07 | 1299 | return false; |
6aa904c4 IR |
1300 | next_load = GROUP_NEXT_ELEMENT (vinfo_for_stmt (load)); |
1301 | } | |
6aa904c4 IR |
1302 | } |
1303 | ||
1304 | /* Check that the alignment of the first load in every subchain, i.e., | |
01d8bf07 RB |
1305 | the first statement in every load node, is supported. |
1306 | ??? This belongs in alignment checking. */ | |
1307 | FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node) | |
1308 | { | |
1309 | first_load = SLP_TREE_SCALAR_STMTS (node)[0]; | |
1310 | if (first_load != GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_load))) | |
1311 | { | |
1312 | dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_load)); | |
1313 | if (vect_supportable_dr_alignment (dr, false) | |
1314 | == dr_unaligned_unsupported) | |
1315 | { | |
1316 | if (dump_enabled_p ()) | |
1317 | { | |
1318 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, | |
1319 | vect_location, | |
1320 | "unsupported unaligned load "); | |
1321 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
1322 | first_load, 0); | |
e645e942 | 1323 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
01d8bf07 RB |
1324 | } |
1325 | return false; | |
1326 | } | |
1327 | } | |
1328 | } | |
6aa904c4 | 1329 | |
01d8bf07 RB |
1330 | /* We are done, no actual permutations need to be generated. */ |
1331 | FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node) | |
1332 | SLP_TREE_LOAD_PERMUTATION (node).release (); | |
1333 | return true; | |
6aa904c4 IR |
1334 | } |
1335 | ||
b8698a0f L |
1336 | /* FORNOW: the only supported permutation is 0..01..1.. of length equal to |
1337 | GROUP_SIZE and where each sequence of same drs is of GROUP_SIZE length as | |
b5aeb3bb | 1338 | well (unless it's reduction). */ |
01d8bf07 | 1339 | if (SLP_INSTANCE_LOADS (slp_instn).length () != group_size) |
ebfd146a | 1340 | return false; |
01d8bf07 RB |
1341 | FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node) |
1342 | if (!node->load_permutation.exists ()) | |
1343 | return false; | |
ebfd146a | 1344 | |
7417f6c0 | 1345 | load_index = sbitmap_alloc (group_size); |
f61e445a | 1346 | bitmap_clear (load_index); |
01d8bf07 RB |
1347 | FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node) |
1348 | { | |
1349 | unsigned int lidx = node->load_permutation[0]; | |
1350 | if (bitmap_bit_p (load_index, lidx)) | |
1351 | { | |
1352 | sbitmap_free (load_index); | |
1353 | return false; | |
1354 | } | |
1355 | bitmap_set_bit (load_index, lidx); | |
1356 | FOR_EACH_VEC_ELT (node->load_permutation, j, k) | |
1357 | if (k != lidx) | |
1358 | { | |
1359 | sbitmap_free (load_index); | |
1360 | return false; | |
1361 | } | |
ebfd146a | 1362 | } |
01d8bf07 RB |
1363 | for (i = 0; i < group_size; i++) |
1364 | if (!bitmap_bit_p (load_index, i)) | |
b8d381a3 JJ |
1365 | { |
1366 | sbitmap_free (load_index); | |
1367 | return false; | |
1368 | } | |
7417f6c0 | 1369 | sbitmap_free (load_index); |
ebfd146a | 1370 | |
01d8bf07 RB |
1371 | FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node) |
1372 | if (node->load_permutation.exists () | |
1373 | && !vect_transform_slp_perm_load | |
1374 | (node, vNULL, NULL, | |
1375 | SLP_INSTANCE_UNROLLING_FACTOR (slp_instn), slp_instn, true)) | |
1376 | return false; | |
1377 | return true; | |
ebfd146a IR |
1378 | } |
1379 | ||
1380 | ||
b8698a0f | 1381 | /* Find the first load in the loop that belongs to INSTANCE. |
ebfd146a | 1382 | When loads are in several SLP nodes, there can be a case in which the first |
b8698a0f | 1383 | load does not appear in the first SLP node to be transformed, causing |
ff802fa1 | 1384 | incorrect order of statements. Since we generate all the loads together, |
ebfd146a IR |
1385 | they must be inserted before the first load of the SLP instance and not |
1386 | before the first load of the first node of the instance. */ | |
ff802fa1 | 1387 | |
b8698a0f L |
1388 | static gimple |
1389 | vect_find_first_load_in_slp_instance (slp_instance instance) | |
ebfd146a IR |
1390 | { |
1391 | int i, j; | |
1392 | slp_tree load_node; | |
1393 | gimple first_load = NULL, load; | |
1394 | ||
9771b263 DN |
1395 | FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (instance), i, load_node) |
1396 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (load_node), j, load) | |
ebfd146a | 1397 | first_load = get_earlier_stmt (load, first_load); |
b8698a0f | 1398 | |
ebfd146a IR |
1399 | return first_load; |
1400 | } | |
1401 | ||
1402 | ||
e4a707c4 | 1403 | /* Find the last store in SLP INSTANCE. */ |
ff802fa1 | 1404 | |
e4a707c4 IR |
1405 | static gimple |
1406 | vect_find_last_store_in_slp_instance (slp_instance instance) | |
1407 | { | |
1408 | int i; | |
1409 | slp_tree node; | |
1410 | gimple last_store = NULL, store; | |
1411 | ||
1412 | node = SLP_INSTANCE_TREE (instance); | |
9771b263 | 1413 | for (i = 0; SLP_TREE_SCALAR_STMTS (node).iterate (i, &store); i++) |
e4a707c4 IR |
1414 | last_store = get_later_stmt (store, last_store); |
1415 | ||
1416 | return last_store; | |
1417 | } | |
1418 | ||
23847df4 RB |
1419 | /* Compute the cost for the SLP node NODE in the SLP instance INSTANCE. */ |
1420 | ||
1421 | static void | |
1422 | vect_analyze_slp_cost_1 (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo, | |
1423 | slp_instance instance, slp_tree node, | |
1424 | stmt_vector_for_cost *prologue_cost_vec, | |
1425 | unsigned ncopies_for_cost) | |
1426 | { | |
1427 | stmt_vector_for_cost *body_cost_vec = &SLP_INSTANCE_BODY_COST_VEC (instance); | |
1428 | ||
1429 | unsigned i; | |
1430 | slp_tree child; | |
1431 | gimple stmt, s; | |
1432 | stmt_vec_info stmt_info; | |
1433 | tree lhs; | |
1434 | unsigned group_size = SLP_INSTANCE_GROUP_SIZE (instance); | |
1435 | ||
1436 | /* Recurse down the SLP tree. */ | |
1437 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) | |
1438 | vect_analyze_slp_cost_1 (loop_vinfo, bb_vinfo, | |
1439 | instance, child, prologue_cost_vec, | |
1440 | ncopies_for_cost); | |
1441 | ||
1442 | /* Look at the first scalar stmt to determine the cost. */ | |
1443 | stmt = SLP_TREE_SCALAR_STMTS (node)[0]; | |
1444 | stmt_info = vinfo_for_stmt (stmt); | |
1445 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1446 | { | |
1447 | if (DR_IS_WRITE (STMT_VINFO_DATA_REF (stmt_info))) | |
1448 | vect_model_store_cost (stmt_info, ncopies_for_cost, false, | |
1449 | vect_uninitialized_def, | |
1450 | node, prologue_cost_vec, body_cost_vec); | |
1451 | else | |
1452 | { | |
1453 | int i; | |
1454 | gcc_checking_assert (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info))); | |
1455 | vect_model_load_cost (stmt_info, ncopies_for_cost, false, | |
1456 | node, prologue_cost_vec, body_cost_vec); | |
1457 | /* If the load is permuted record the cost for the permutation. | |
1458 | ??? Loads from multiple chains are let through here only | |
1459 | for a single special case involving complex numbers where | |
1460 | in the end no permutation is necessary. */ | |
1461 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, s) | |
1462 | if ((STMT_VINFO_GROUP_FIRST_ELEMENT (vinfo_for_stmt (s)) | |
1463 | == STMT_VINFO_GROUP_FIRST_ELEMENT (stmt_info)) | |
1464 | && vect_get_place_in_interleaving_chain | |
1465 | (s, STMT_VINFO_GROUP_FIRST_ELEMENT (stmt_info)) != i) | |
1466 | { | |
1467 | record_stmt_cost (body_cost_vec, group_size, vec_perm, | |
1468 | stmt_info, 0, vect_body); | |
1469 | break; | |
1470 | } | |
1471 | } | |
1472 | } | |
1473 | else | |
1474 | record_stmt_cost (body_cost_vec, ncopies_for_cost, vector_stmt, | |
1475 | stmt_info, 0, vect_body); | |
1476 | ||
1477 | /* Scan operands and account for prologue cost of constants/externals. | |
1478 | ??? This over-estimates cost for multiple uses and should be | |
1479 | re-engineered. */ | |
1480 | lhs = gimple_get_lhs (stmt); | |
1481 | for (i = 0; i < gimple_num_ops (stmt); ++i) | |
1482 | { | |
1483 | tree def, op = gimple_op (stmt, i); | |
1484 | gimple def_stmt; | |
1485 | enum vect_def_type dt; | |
1486 | if (!op || op == lhs) | |
1487 | continue; | |
1488 | if (vect_is_simple_use (op, NULL, loop_vinfo, bb_vinfo, | |
1489 | &def_stmt, &def, &dt) | |
1490 | && (dt == vect_constant_def || dt == vect_external_def)) | |
1491 | record_stmt_cost (prologue_cost_vec, 1, vector_stmt, | |
1492 | stmt_info, 0, vect_prologue); | |
1493 | } | |
1494 | } | |
1495 | ||
1496 | /* Compute the cost for the SLP instance INSTANCE. */ | |
1497 | ||
1498 | static void | |
1499 | vect_analyze_slp_cost (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo, | |
1500 | slp_instance instance, unsigned nunits) | |
1501 | { | |
1502 | stmt_vector_for_cost body_cost_vec, prologue_cost_vec; | |
1503 | unsigned ncopies_for_cost; | |
1504 | stmt_info_for_cost *si; | |
1505 | unsigned i; | |
1506 | ||
1507 | /* Calculate the number of vector stmts to create based on the unrolling | |
1508 | factor (number of vectors is 1 if NUNITS >= GROUP_SIZE, and is | |
1509 | GROUP_SIZE / NUNITS otherwise. */ | |
1510 | unsigned group_size = SLP_INSTANCE_GROUP_SIZE (instance); | |
1511 | ncopies_for_cost = least_common_multiple (nunits, group_size) / nunits; | |
1512 | ||
1513 | prologue_cost_vec.create (10); | |
1514 | body_cost_vec.create (10); | |
1515 | SLP_INSTANCE_BODY_COST_VEC (instance) = body_cost_vec; | |
1516 | vect_analyze_slp_cost_1 (loop_vinfo, bb_vinfo, | |
1517 | instance, SLP_INSTANCE_TREE (instance), | |
1518 | &prologue_cost_vec, ncopies_for_cost); | |
1519 | ||
1520 | /* Record the prologue costs, which were delayed until we were | |
1521 | sure that SLP was successful. Unlike the body costs, we know | |
1522 | the final values now regardless of the loop vectorization factor. */ | |
1523 | void *data = (loop_vinfo ? LOOP_VINFO_TARGET_COST_DATA (loop_vinfo) | |
1524 | : BB_VINFO_TARGET_COST_DATA (bb_vinfo)); | |
1525 | FOR_EACH_VEC_ELT (prologue_cost_vec, i, si) | |
1526 | { | |
1527 | struct _stmt_vec_info *stmt_info | |
1528 | = si->stmt ? vinfo_for_stmt (si->stmt) : NULL; | |
1529 | (void) add_stmt_cost (data, si->count, si->kind, stmt_info, | |
1530 | si->misalign, vect_prologue); | |
1531 | } | |
1532 | ||
1533 | prologue_cost_vec.release (); | |
1534 | } | |
e4a707c4 | 1535 | |
0d0293ac | 1536 | /* Analyze an SLP instance starting from a group of grouped stores. Call |
b8698a0f | 1537 | vect_build_slp_tree to build a tree of packed stmts if possible. |
ebfd146a IR |
1538 | Return FALSE if it's impossible to SLP any stmt in the loop. */ |
1539 | ||
1540 | static bool | |
a70d6342 | 1541 | vect_analyze_slp_instance (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo, |
1428105c | 1542 | gimple stmt, unsigned max_tree_size) |
ebfd146a IR |
1543 | { |
1544 | slp_instance new_instance; | |
d092494c | 1545 | slp_tree node; |
e14c1050 | 1546 | unsigned int group_size = GROUP_SIZE (vinfo_for_stmt (stmt)); |
ebfd146a | 1547 | unsigned int unrolling_factor = 1, nunits; |
b5aeb3bb | 1548 | tree vectype, scalar_type = NULL_TREE; |
ebfd146a | 1549 | gimple next; |
0f900dfa | 1550 | unsigned int vectorization_factor = 0; |
23847df4 | 1551 | int i; |
ebfd146a | 1552 | unsigned int max_nunits = 0; |
9771b263 | 1553 | vec<slp_tree> loads; |
b5aeb3bb | 1554 | struct data_reference *dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt)); |
9771b263 | 1555 | vec<gimple> scalar_stmts; |
b5aeb3bb | 1556 | |
b010117a | 1557 | if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) |
b5aeb3bb | 1558 | { |
b010117a IR |
1559 | if (dr) |
1560 | { | |
1561 | scalar_type = TREE_TYPE (DR_REF (dr)); | |
1562 | vectype = get_vectype_for_scalar_type (scalar_type); | |
1563 | } | |
1564 | else | |
1565 | { | |
1566 | gcc_assert (loop_vinfo); | |
1567 | vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt)); | |
1568 | } | |
1569 | ||
e14c1050 | 1570 | group_size = GROUP_SIZE (vinfo_for_stmt (stmt)); |
b5aeb3bb IR |
1571 | } |
1572 | else | |
1573 | { | |
1574 | gcc_assert (loop_vinfo); | |
1575 | vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt)); | |
9771b263 | 1576 | group_size = LOOP_VINFO_REDUCTIONS (loop_vinfo).length (); |
b5aeb3bb | 1577 | } |
b8698a0f | 1578 | |
ebfd146a IR |
1579 | if (!vectype) |
1580 | { | |
73fbfcad | 1581 | if (dump_enabled_p ()) |
ebfd146a | 1582 | { |
78c60e3d SS |
1583 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1584 | "Build SLP failed: unsupported data-type "); | |
1585 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, scalar_type); | |
e645e942 | 1586 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a | 1587 | } |
b5aeb3bb | 1588 | |
ebfd146a IR |
1589 | return false; |
1590 | } | |
1591 | ||
1592 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
a70d6342 IR |
1593 | if (loop_vinfo) |
1594 | vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
1595 | else | |
a70d6342 IR |
1596 | vectorization_factor = nunits; |
1597 | ||
a70d6342 IR |
1598 | /* Calculate the unrolling factor. */ |
1599 | unrolling_factor = least_common_multiple (nunits, group_size) / group_size; | |
1600 | if (unrolling_factor != 1 && !loop_vinfo) | |
1601 | { | |
73fbfcad | 1602 | if (dump_enabled_p ()) |
e645e942 | 1603 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
78c60e3d | 1604 | "Build SLP failed: unrolling required in basic" |
e645e942 | 1605 | " block SLP\n"); |
b8698a0f | 1606 | |
a70d6342 IR |
1607 | return false; |
1608 | } | |
1609 | ||
0d0293ac | 1610 | /* Create a node (a root of the SLP tree) for the packed grouped stores. */ |
9771b263 | 1611 | scalar_stmts.create (group_size); |
ebfd146a | 1612 | next = stmt; |
b010117a | 1613 | if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) |
ebfd146a | 1614 | { |
b5aeb3bb IR |
1615 | /* Collect the stores and store them in SLP_TREE_SCALAR_STMTS. */ |
1616 | while (next) | |
1617 | { | |
f7e531cf IR |
1618 | if (STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (next)) |
1619 | && STMT_VINFO_RELATED_STMT (vinfo_for_stmt (next))) | |
9771b263 DN |
1620 | scalar_stmts.safe_push ( |
1621 | STMT_VINFO_RELATED_STMT (vinfo_for_stmt (next))); | |
f7e531cf | 1622 | else |
9771b263 | 1623 | scalar_stmts.safe_push (next); |
e14c1050 | 1624 | next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next)); |
b5aeb3bb IR |
1625 | } |
1626 | } | |
1627 | else | |
1628 | { | |
1629 | /* Collect reduction statements. */ | |
9771b263 DN |
1630 | vec<gimple> reductions = LOOP_VINFO_REDUCTIONS (loop_vinfo); |
1631 | for (i = 0; reductions.iterate (i, &next); i++) | |
1632 | scalar_stmts.safe_push (next); | |
ebfd146a IR |
1633 | } |
1634 | ||
d092494c | 1635 | node = vect_create_new_slp_node (scalar_stmts); |
ebfd146a | 1636 | |
9771b263 | 1637 | loads.create (group_size); |
ebfd146a IR |
1638 | |
1639 | /* Build the tree for the SLP instance. */ | |
b8698a0f | 1640 | if (vect_build_slp_tree (loop_vinfo, bb_vinfo, &node, group_size, |
abf9bfbc | 1641 | &max_nunits, &loads, |
1428105c RB |
1642 | vectorization_factor, NULL, NULL, NULL, |
1643 | max_tree_size)) | |
ebfd146a | 1644 | { |
4ef69dfc | 1645 | /* Calculate the unrolling factor based on the smallest type. */ |
ebfd146a IR |
1646 | if (max_nunits > nunits) |
1647 | unrolling_factor = least_common_multiple (max_nunits, group_size) | |
1648 | / group_size; | |
b8698a0f | 1649 | |
4ef69dfc IR |
1650 | if (unrolling_factor != 1 && !loop_vinfo) |
1651 | { | |
73fbfcad | 1652 | if (dump_enabled_p ()) |
e645e942 | 1653 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
78c60e3d | 1654 | "Build SLP failed: unrolling required in basic" |
e645e942 | 1655 | " block SLP\n"); |
c7e62a26 | 1656 | vect_free_slp_tree (node); |
9771b263 | 1657 | loads.release (); |
4ef69dfc IR |
1658 | return false; |
1659 | } | |
1660 | ||
1661 | /* Create a new SLP instance. */ | |
1662 | new_instance = XNEW (struct _slp_instance); | |
1663 | SLP_INSTANCE_TREE (new_instance) = node; | |
1664 | SLP_INSTANCE_GROUP_SIZE (new_instance) = group_size; | |
ebfd146a | 1665 | SLP_INSTANCE_UNROLLING_FACTOR (new_instance) = unrolling_factor; |
23847df4 | 1666 | SLP_INSTANCE_BODY_COST_VEC (new_instance) = vNULL; |
ebfd146a IR |
1667 | SLP_INSTANCE_LOADS (new_instance) = loads; |
1668 | SLP_INSTANCE_FIRST_LOAD_STMT (new_instance) = NULL; | |
abf9bfbc RB |
1669 | |
1670 | /* Compute the load permutation. */ | |
1671 | slp_tree load_node; | |
1672 | bool loads_permuted = false; | |
abf9bfbc RB |
1673 | FOR_EACH_VEC_ELT (loads, i, load_node) |
1674 | { | |
01d8bf07 | 1675 | vec<unsigned> load_permutation; |
abf9bfbc | 1676 | int j; |
6983e6b5 | 1677 | gimple load, first_stmt; |
01d8bf07 RB |
1678 | bool this_load_permuted = false; |
1679 | load_permutation.create (group_size); | |
6983e6b5 RB |
1680 | first_stmt = GROUP_FIRST_ELEMENT |
1681 | (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (load_node)[0])); | |
abf9bfbc RB |
1682 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (load_node), j, load) |
1683 | { | |
6983e6b5 RB |
1684 | int load_place |
1685 | = vect_get_place_in_interleaving_chain (load, first_stmt); | |
1686 | gcc_assert (load_place != -1); | |
1687 | if (load_place != j) | |
01d8bf07 | 1688 | this_load_permuted = true; |
abf9bfbc RB |
1689 | load_permutation.safe_push (load_place); |
1690 | } | |
01d8bf07 RB |
1691 | if (!this_load_permuted) |
1692 | { | |
1693 | load_permutation.release (); | |
1694 | continue; | |
1695 | } | |
1696 | SLP_TREE_LOAD_PERMUTATION (load_node) = load_permutation; | |
1697 | loads_permuted = true; | |
abf9bfbc | 1698 | } |
6aa904c4 IR |
1699 | |
1700 | if (loads_permuted) | |
ebfd146a | 1701 | { |
01d8bf07 | 1702 | if (!vect_supported_load_permutation_p (new_instance)) |
ebfd146a | 1703 | { |
73fbfcad | 1704 | if (dump_enabled_p ()) |
ebfd146a | 1705 | { |
e645e942 | 1706 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
78c60e3d SS |
1707 | "Build SLP failed: unsupported load " |
1708 | "permutation "); | |
1709 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 1710 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a | 1711 | } |
ebfd146a IR |
1712 | vect_free_slp_instance (new_instance); |
1713 | return false; | |
1714 | } | |
1715 | ||
1716 | SLP_INSTANCE_FIRST_LOAD_STMT (new_instance) | |
01d8bf07 | 1717 | = vect_find_first_load_in_slp_instance (new_instance); |
ebfd146a | 1718 | } |
ebfd146a | 1719 | |
23847df4 RB |
1720 | /* Compute the costs of this SLP instance. */ |
1721 | vect_analyze_slp_cost (loop_vinfo, bb_vinfo, | |
1722 | new_instance, TYPE_VECTOR_SUBPARTS (vectype)); | |
92345349 | 1723 | |
a70d6342 | 1724 | if (loop_vinfo) |
9771b263 | 1725 | LOOP_VINFO_SLP_INSTANCES (loop_vinfo).safe_push (new_instance); |
a70d6342 | 1726 | else |
9771b263 | 1727 | BB_VINFO_SLP_INSTANCES (bb_vinfo).safe_push (new_instance); |
b8698a0f | 1728 | |
73fbfcad | 1729 | if (dump_enabled_p ()) |
78c60e3d | 1730 | vect_print_slp_tree (MSG_NOTE, node); |
ebfd146a IR |
1731 | |
1732 | return true; | |
1733 | } | |
1734 | ||
1735 | /* Failed to SLP. */ | |
1736 | /* Free the allocated memory. */ | |
1737 | vect_free_slp_tree (node); | |
9771b263 | 1738 | loads.release (); |
b8698a0f | 1739 | |
a70d6342 | 1740 | return false; |
ebfd146a IR |
1741 | } |
1742 | ||
1743 | ||
ff802fa1 | 1744 | /* Check if there are stmts in the loop can be vectorized using SLP. Build SLP |
ebfd146a IR |
1745 | trees of packed scalar stmts if SLP is possible. */ |
1746 | ||
1747 | bool | |
1428105c RB |
1748 | vect_analyze_slp (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo, |
1749 | unsigned max_tree_size) | |
ebfd146a IR |
1750 | { |
1751 | unsigned int i; | |
9771b263 | 1752 | vec<gimple> grouped_stores; |
6e1aa848 DN |
1753 | vec<gimple> reductions = vNULL; |
1754 | vec<gimple> reduc_chains = vNULL; | |
b010117a | 1755 | gimple first_element; |
a70d6342 | 1756 | bool ok = false; |
ebfd146a | 1757 | |
73fbfcad | 1758 | if (dump_enabled_p ()) |
e645e942 | 1759 | dump_printf_loc (MSG_NOTE, vect_location, "=== vect_analyze_slp ===\n"); |
ebfd146a | 1760 | |
a70d6342 | 1761 | if (loop_vinfo) |
b5aeb3bb | 1762 | { |
0d0293ac | 1763 | grouped_stores = LOOP_VINFO_GROUPED_STORES (loop_vinfo); |
b010117a | 1764 | reduc_chains = LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo); |
b5aeb3bb IR |
1765 | reductions = LOOP_VINFO_REDUCTIONS (loop_vinfo); |
1766 | } | |
a70d6342 | 1767 | else |
0d0293ac | 1768 | grouped_stores = BB_VINFO_GROUPED_STORES (bb_vinfo); |
b8698a0f | 1769 | |
0d0293ac | 1770 | /* Find SLP sequences starting from groups of grouped stores. */ |
9771b263 | 1771 | FOR_EACH_VEC_ELT (grouped_stores, i, first_element) |
1428105c RB |
1772 | if (vect_analyze_slp_instance (loop_vinfo, bb_vinfo, first_element, |
1773 | max_tree_size)) | |
a70d6342 | 1774 | ok = true; |
ebfd146a | 1775 | |
b8698a0f | 1776 | if (bb_vinfo && !ok) |
a70d6342 | 1777 | { |
73fbfcad | 1778 | if (dump_enabled_p ()) |
78c60e3d | 1779 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1780 | "Failed to SLP the basic block.\n"); |
a70d6342 IR |
1781 | |
1782 | return false; | |
1783 | } | |
ebfd146a | 1784 | |
b010117a | 1785 | if (loop_vinfo |
9771b263 | 1786 | && LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo).length () > 0) |
b010117a IR |
1787 | { |
1788 | /* Find SLP sequences starting from reduction chains. */ | |
9771b263 | 1789 | FOR_EACH_VEC_ELT (reduc_chains, i, first_element) |
1428105c RB |
1790 | if (vect_analyze_slp_instance (loop_vinfo, bb_vinfo, first_element, |
1791 | max_tree_size)) | |
b010117a IR |
1792 | ok = true; |
1793 | else | |
1794 | return false; | |
1795 | ||
1796 | /* Don't try to vectorize SLP reductions if reduction chain was | |
1797 | detected. */ | |
1798 | return ok; | |
1799 | } | |
1800 | ||
b5aeb3bb | 1801 | /* Find SLP sequences starting from groups of reductions. */ |
9771b263 | 1802 | if (loop_vinfo && LOOP_VINFO_REDUCTIONS (loop_vinfo).length () > 1 |
1428105c RB |
1803 | && vect_analyze_slp_instance (loop_vinfo, bb_vinfo, reductions[0], |
1804 | max_tree_size)) | |
b5aeb3bb IR |
1805 | ok = true; |
1806 | ||
ebfd146a IR |
1807 | return true; |
1808 | } | |
1809 | ||
1810 | ||
1811 | /* For each possible SLP instance decide whether to SLP it and calculate overall | |
437f4a00 IR |
1812 | unrolling factor needed to SLP the loop. Return TRUE if decided to SLP at |
1813 | least one instance. */ | |
ebfd146a | 1814 | |
437f4a00 | 1815 | bool |
ebfd146a IR |
1816 | vect_make_slp_decision (loop_vec_info loop_vinfo) |
1817 | { | |
1818 | unsigned int i, unrolling_factor = 1; | |
9771b263 | 1819 | vec<slp_instance> slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo); |
ebfd146a IR |
1820 | slp_instance instance; |
1821 | int decided_to_slp = 0; | |
1822 | ||
73fbfcad | 1823 | if (dump_enabled_p ()) |
e645e942 TJ |
1824 | dump_printf_loc (MSG_NOTE, vect_location, "=== vect_make_slp_decision ===" |
1825 | "\n"); | |
ebfd146a | 1826 | |
9771b263 | 1827 | FOR_EACH_VEC_ELT (slp_instances, i, instance) |
ebfd146a IR |
1828 | { |
1829 | /* FORNOW: SLP if you can. */ | |
1830 | if (unrolling_factor < SLP_INSTANCE_UNROLLING_FACTOR (instance)) | |
1831 | unrolling_factor = SLP_INSTANCE_UNROLLING_FACTOR (instance); | |
1832 | ||
ff802fa1 | 1833 | /* Mark all the stmts that belong to INSTANCE as PURE_SLP stmts. Later we |
b8698a0f | 1834 | call vect_detect_hybrid_slp () to find stmts that need hybrid SLP and |
ff802fa1 | 1835 | loop-based vectorization. Such stmts will be marked as HYBRID. */ |
ebfd146a IR |
1836 | vect_mark_slp_stmts (SLP_INSTANCE_TREE (instance), pure_slp, -1); |
1837 | decided_to_slp++; | |
1838 | } | |
1839 | ||
1840 | LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo) = unrolling_factor; | |
1841 | ||
73fbfcad | 1842 | if (decided_to_slp && dump_enabled_p ()) |
ccb3ad87 | 1843 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1844 | "Decided to SLP %d instances. Unrolling factor %d\n", |
78c60e3d | 1845 | decided_to_slp, unrolling_factor); |
437f4a00 IR |
1846 | |
1847 | return (decided_to_slp > 0); | |
ebfd146a IR |
1848 | } |
1849 | ||
1850 | ||
1851 | /* Find stmts that must be both vectorized and SLPed (since they feed stmts that | |
ff802fa1 | 1852 | can't be SLPed) in the tree rooted at NODE. Mark such stmts as HYBRID. */ |
ebfd146a IR |
1853 | |
1854 | static void | |
642fce57 | 1855 | vect_detect_hybrid_slp_stmts (slp_tree node, unsigned i, slp_vect_type stype) |
ebfd146a | 1856 | { |
642fce57 | 1857 | gimple stmt = SLP_TREE_SCALAR_STMTS (node)[i]; |
ebfd146a IR |
1858 | imm_use_iterator imm_iter; |
1859 | gimple use_stmt; | |
642fce57 | 1860 | stmt_vec_info use_vinfo, stmt_vinfo = vinfo_for_stmt (stmt); |
d755c7ef | 1861 | slp_tree child; |
f2c74cc4 | 1862 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
642fce57 RB |
1863 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
1864 | int j; | |
1865 | ||
1866 | /* Propagate hybrid down the SLP tree. */ | |
1867 | if (stype == hybrid) | |
1868 | ; | |
1869 | else if (HYBRID_SLP_STMT (stmt_vinfo)) | |
1870 | stype = hybrid; | |
1871 | else | |
1872 | { | |
1873 | /* Check if a pure SLP stmt has uses in non-SLP stmts. */ | |
1874 | gcc_checking_assert (PURE_SLP_STMT (stmt_vinfo)); | |
1875 | if (TREE_CODE (gimple_op (stmt, 0)) == SSA_NAME) | |
1876 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, gimple_op (stmt, 0)) | |
1877 | if (gimple_bb (use_stmt) | |
1878 | && flow_bb_inside_loop_p (loop, gimple_bb (use_stmt)) | |
1879 | && (use_vinfo = vinfo_for_stmt (use_stmt)) | |
1880 | && !STMT_SLP_TYPE (use_vinfo) | |
1881 | && (STMT_VINFO_RELEVANT (use_vinfo) | |
1882 | || VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (use_vinfo)) | |
1883 | || (STMT_VINFO_IN_PATTERN_P (use_vinfo) | |
1884 | && STMT_VINFO_RELATED_STMT (use_vinfo) | |
1885 | && !STMT_SLP_TYPE (vinfo_for_stmt | |
1886 | (STMT_VINFO_RELATED_STMT (use_vinfo))))) | |
1887 | && !(gimple_code (use_stmt) == GIMPLE_PHI | |
1888 | && STMT_VINFO_DEF_TYPE (use_vinfo) == vect_reduction_def)) | |
1889 | stype = hybrid; | |
1890 | } | |
ebfd146a | 1891 | |
642fce57 RB |
1892 | if (stype == hybrid) |
1893 | STMT_SLP_TYPE (stmt_vinfo) = hybrid; | |
ebfd146a | 1894 | |
642fce57 RB |
1895 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), j, child) |
1896 | vect_detect_hybrid_slp_stmts (child, i, stype); | |
1897 | } | |
f2c74cc4 | 1898 | |
642fce57 | 1899 | /* Helpers for vect_detect_hybrid_slp walking pattern stmt uses. */ |
ebfd146a | 1900 | |
642fce57 RB |
1901 | static tree |
1902 | vect_detect_hybrid_slp_1 (tree *tp, int *, void *data) | |
1903 | { | |
1904 | walk_stmt_info *wi = (walk_stmt_info *)data; | |
1905 | struct loop *loopp = (struct loop *)wi->info; | |
1906 | ||
1907 | if (wi->is_lhs) | |
1908 | return NULL_TREE; | |
1909 | ||
1910 | if (TREE_CODE (*tp) == SSA_NAME | |
1911 | && !SSA_NAME_IS_DEFAULT_DEF (*tp)) | |
1912 | { | |
1913 | gimple def_stmt = SSA_NAME_DEF_STMT (*tp); | |
1914 | if (flow_bb_inside_loop_p (loopp, gimple_bb (def_stmt)) | |
1915 | && PURE_SLP_STMT (vinfo_for_stmt (def_stmt))) | |
1916 | STMT_SLP_TYPE (vinfo_for_stmt (def_stmt)) = hybrid; | |
1917 | } | |
1918 | ||
1919 | return NULL_TREE; | |
ebfd146a IR |
1920 | } |
1921 | ||
642fce57 RB |
1922 | static tree |
1923 | vect_detect_hybrid_slp_2 (gimple_stmt_iterator *gsi, bool *handled, | |
1924 | walk_stmt_info *) | |
1925 | { | |
1926 | /* If the stmt is in a SLP instance then this isn't a reason | |
1927 | to mark use definitions in other SLP instances as hybrid. */ | |
1928 | if (STMT_SLP_TYPE (vinfo_for_stmt (gsi_stmt (*gsi))) != loop_vect) | |
1929 | *handled = true; | |
1930 | return NULL_TREE; | |
1931 | } | |
ebfd146a IR |
1932 | |
1933 | /* Find stmts that must be both vectorized and SLPed. */ | |
1934 | ||
1935 | void | |
1936 | vect_detect_hybrid_slp (loop_vec_info loop_vinfo) | |
1937 | { | |
1938 | unsigned int i; | |
9771b263 | 1939 | vec<slp_instance> slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo); |
ebfd146a IR |
1940 | slp_instance instance; |
1941 | ||
73fbfcad | 1942 | if (dump_enabled_p ()) |
e645e942 TJ |
1943 | dump_printf_loc (MSG_NOTE, vect_location, "=== vect_detect_hybrid_slp ===" |
1944 | "\n"); | |
ebfd146a | 1945 | |
642fce57 RB |
1946 | /* First walk all pattern stmt in the loop and mark defs of uses as |
1947 | hybrid because immediate uses in them are not recorded. */ | |
1948 | for (i = 0; i < LOOP_VINFO_LOOP (loop_vinfo)->num_nodes; ++i) | |
1949 | { | |
1950 | basic_block bb = LOOP_VINFO_BBS (loop_vinfo)[i]; | |
1951 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi); | |
1952 | gsi_next (&gsi)) | |
1953 | { | |
1954 | gimple stmt = gsi_stmt (gsi); | |
1955 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1956 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
1957 | { | |
1958 | walk_stmt_info wi; | |
1959 | memset (&wi, 0, sizeof (wi)); | |
1960 | wi.info = LOOP_VINFO_LOOP (loop_vinfo); | |
1961 | gimple_stmt_iterator gsi2 | |
1962 | = gsi_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
1963 | walk_gimple_stmt (&gsi2, vect_detect_hybrid_slp_2, | |
1964 | vect_detect_hybrid_slp_1, &wi); | |
1965 | walk_gimple_seq (STMT_VINFO_PATTERN_DEF_SEQ (stmt_info), | |
1966 | vect_detect_hybrid_slp_2, | |
1967 | vect_detect_hybrid_slp_1, &wi); | |
1968 | } | |
1969 | } | |
1970 | } | |
1971 | ||
1972 | /* Then walk the SLP instance trees marking stmts with uses in | |
1973 | non-SLP stmts as hybrid, also propagating hybrid down the | |
1974 | SLP tree, collecting the above info on-the-fly. */ | |
9771b263 | 1975 | FOR_EACH_VEC_ELT (slp_instances, i, instance) |
642fce57 RB |
1976 | { |
1977 | for (unsigned i = 0; i < SLP_INSTANCE_GROUP_SIZE (instance); ++i) | |
1978 | vect_detect_hybrid_slp_stmts (SLP_INSTANCE_TREE (instance), | |
1979 | i, pure_slp); | |
1980 | } | |
ebfd146a IR |
1981 | } |
1982 | ||
a70d6342 IR |
1983 | |
1984 | /* Create and initialize a new bb_vec_info struct for BB, as well as | |
1985 | stmt_vec_info structs for all the stmts in it. */ | |
b8698a0f | 1986 | |
a70d6342 IR |
1987 | static bb_vec_info |
1988 | new_bb_vec_info (basic_block bb) | |
1989 | { | |
1990 | bb_vec_info res = NULL; | |
1991 | gimple_stmt_iterator gsi; | |
1992 | ||
1993 | res = (bb_vec_info) xcalloc (1, sizeof (struct _bb_vec_info)); | |
1994 | BB_VINFO_BB (res) = bb; | |
1995 | ||
1996 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1997 | { | |
1998 | gimple stmt = gsi_stmt (gsi); | |
1999 | gimple_set_uid (stmt, 0); | |
2000 | set_vinfo_for_stmt (stmt, new_stmt_vec_info (stmt, NULL, res)); | |
2001 | } | |
2002 | ||
9771b263 DN |
2003 | BB_VINFO_GROUPED_STORES (res).create (10); |
2004 | BB_VINFO_SLP_INSTANCES (res).create (2); | |
c3e7ee41 | 2005 | BB_VINFO_TARGET_COST_DATA (res) = init_cost (NULL); |
a70d6342 IR |
2006 | |
2007 | bb->aux = res; | |
2008 | return res; | |
2009 | } | |
2010 | ||
2011 | ||
2012 | /* Free BB_VINFO struct, as well as all the stmt_vec_info structs of all the | |
2013 | stmts in the basic block. */ | |
2014 | ||
2015 | static void | |
2016 | destroy_bb_vec_info (bb_vec_info bb_vinfo) | |
2017 | { | |
9771b263 | 2018 | vec<slp_instance> slp_instances; |
c7e62a26 | 2019 | slp_instance instance; |
a70d6342 IR |
2020 | basic_block bb; |
2021 | gimple_stmt_iterator si; | |
c7e62a26 | 2022 | unsigned i; |
a70d6342 IR |
2023 | |
2024 | if (!bb_vinfo) | |
2025 | return; | |
2026 | ||
2027 | bb = BB_VINFO_BB (bb_vinfo); | |
2028 | ||
2029 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
2030 | { | |
2031 | gimple stmt = gsi_stmt (si); | |
2032 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2033 | ||
2034 | if (stmt_info) | |
2035 | /* Free stmt_vec_info. */ | |
2036 | free_stmt_vec_info (stmt); | |
2037 | } | |
2038 | ||
c716e67f | 2039 | vect_destroy_datarefs (NULL, bb_vinfo); |
01be8516 | 2040 | free_dependence_relations (BB_VINFO_DDRS (bb_vinfo)); |
9771b263 | 2041 | BB_VINFO_GROUPED_STORES (bb_vinfo).release (); |
c7e62a26 | 2042 | slp_instances = BB_VINFO_SLP_INSTANCES (bb_vinfo); |
9771b263 | 2043 | FOR_EACH_VEC_ELT (slp_instances, i, instance) |
c7e62a26 | 2044 | vect_free_slp_instance (instance); |
9771b263 | 2045 | BB_VINFO_SLP_INSTANCES (bb_vinfo).release (); |
c3e7ee41 | 2046 | destroy_cost_data (BB_VINFO_TARGET_COST_DATA (bb_vinfo)); |
a70d6342 IR |
2047 | free (bb_vinfo); |
2048 | bb->aux = NULL; | |
2049 | } | |
2050 | ||
2051 | ||
2052 | /* Analyze statements contained in SLP tree node after recursively analyzing | |
2053 | the subtree. Return TRUE if the operations are supported. */ | |
2054 | ||
2055 | static bool | |
2056 | vect_slp_analyze_node_operations (bb_vec_info bb_vinfo, slp_tree node) | |
2057 | { | |
2058 | bool dummy; | |
2059 | int i; | |
2060 | gimple stmt; | |
d755c7ef | 2061 | slp_tree child; |
a70d6342 IR |
2062 | |
2063 | if (!node) | |
2064 | return true; | |
2065 | ||
9771b263 | 2066 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) |
d755c7ef | 2067 | if (!vect_slp_analyze_node_operations (bb_vinfo, child)) |
d092494c | 2068 | return false; |
a70d6342 | 2069 | |
9771b263 | 2070 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt) |
a70d6342 IR |
2071 | { |
2072 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2073 | gcc_assert (stmt_info); | |
2074 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
2075 | ||
2076 | if (!vect_analyze_stmt (stmt, &dummy, node)) | |
2077 | return false; | |
2078 | } | |
2079 | ||
2080 | return true; | |
2081 | } | |
2082 | ||
2083 | ||
ff802fa1 | 2084 | /* Analyze statements in SLP instances of the basic block. Return TRUE if the |
a70d6342 IR |
2085 | operations are supported. */ |
2086 | ||
2087 | static bool | |
2088 | vect_slp_analyze_operations (bb_vec_info bb_vinfo) | |
2089 | { | |
9771b263 | 2090 | vec<slp_instance> slp_instances = BB_VINFO_SLP_INSTANCES (bb_vinfo); |
a70d6342 IR |
2091 | slp_instance instance; |
2092 | int i; | |
2093 | ||
9771b263 | 2094 | for (i = 0; slp_instances.iterate (i, &instance); ) |
a70d6342 | 2095 | { |
b8698a0f | 2096 | if (!vect_slp_analyze_node_operations (bb_vinfo, |
a70d6342 IR |
2097 | SLP_INSTANCE_TREE (instance))) |
2098 | { | |
2099 | vect_free_slp_instance (instance); | |
9771b263 | 2100 | slp_instances.ordered_remove (i); |
a70d6342 IR |
2101 | } |
2102 | else | |
2103 | i++; | |
b8698a0f L |
2104 | } |
2105 | ||
9771b263 | 2106 | if (!slp_instances.length ()) |
a70d6342 IR |
2107 | return false; |
2108 | ||
2109 | return true; | |
2110 | } | |
2111 | ||
6eddf228 RB |
2112 | |
2113 | /* Compute the scalar cost of the SLP node NODE and its children | |
2114 | and return it. Do not account defs that are marked in LIFE and | |
2115 | update LIFE according to uses of NODE. */ | |
2116 | ||
2117 | static unsigned | |
292cba13 | 2118 | vect_bb_slp_scalar_cost (basic_block bb, |
ff4c81cc | 2119 | slp_tree node, vec<bool, va_heap> *life) |
6eddf228 RB |
2120 | { |
2121 | unsigned scalar_cost = 0; | |
2122 | unsigned i; | |
2123 | gimple stmt; | |
2124 | slp_tree child; | |
2125 | ||
2126 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt) | |
2127 | { | |
2128 | unsigned stmt_cost; | |
2129 | ssa_op_iter op_iter; | |
2130 | def_operand_p def_p; | |
2131 | stmt_vec_info stmt_info; | |
2132 | ||
ff4c81cc | 2133 | if ((*life)[i]) |
6eddf228 RB |
2134 | continue; |
2135 | ||
2136 | /* If there is a non-vectorized use of the defs then the scalar | |
2137 | stmt is kept live in which case we do not account it or any | |
2138 | required defs in the SLP children in the scalar cost. This | |
2139 | way we make the vectorization more costly when compared to | |
2140 | the scalar cost. */ | |
2141 | FOR_EACH_SSA_DEF_OPERAND (def_p, stmt, op_iter, SSA_OP_DEF) | |
2142 | { | |
2143 | imm_use_iterator use_iter; | |
2144 | gimple use_stmt; | |
2145 | FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, DEF_FROM_PTR (def_p)) | |
f30a0ba5 RB |
2146 | if (!is_gimple_debug (use_stmt) |
2147 | && (gimple_code (use_stmt) == GIMPLE_PHI | |
2148 | || gimple_bb (use_stmt) != bb | |
2149 | || !STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (use_stmt)))) | |
6eddf228 | 2150 | { |
ff4c81cc | 2151 | (*life)[i] = true; |
6eddf228 RB |
2152 | BREAK_FROM_IMM_USE_STMT (use_iter); |
2153 | } | |
2154 | } | |
ff4c81cc | 2155 | if ((*life)[i]) |
6eddf228 RB |
2156 | continue; |
2157 | ||
2158 | stmt_info = vinfo_for_stmt (stmt); | |
2159 | if (STMT_VINFO_DATA_REF (stmt_info)) | |
2160 | { | |
2161 | if (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info))) | |
2162 | stmt_cost = vect_get_stmt_cost (scalar_load); | |
2163 | else | |
2164 | stmt_cost = vect_get_stmt_cost (scalar_store); | |
2165 | } | |
2166 | else | |
2167 | stmt_cost = vect_get_stmt_cost (scalar_stmt); | |
2168 | ||
2169 | scalar_cost += stmt_cost; | |
2170 | } | |
2171 | ||
2172 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) | |
292cba13 | 2173 | scalar_cost += vect_bb_slp_scalar_cost (bb, child, life); |
6eddf228 RB |
2174 | |
2175 | return scalar_cost; | |
2176 | } | |
2177 | ||
69f11a13 IR |
2178 | /* Check if vectorization of the basic block is profitable. */ |
2179 | ||
2180 | static bool | |
2181 | vect_bb_vectorization_profitable_p (bb_vec_info bb_vinfo) | |
2182 | { | |
9771b263 | 2183 | vec<slp_instance> slp_instances = BB_VINFO_SLP_INSTANCES (bb_vinfo); |
69f11a13 | 2184 | slp_instance instance; |
c3e7ee41 BS |
2185 | int i, j; |
2186 | unsigned int vec_inside_cost = 0, vec_outside_cost = 0, scalar_cost = 0; | |
92345349 | 2187 | unsigned int vec_prologue_cost = 0, vec_epilogue_cost = 0; |
92345349 | 2188 | void *target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); |
69f11a13 | 2189 | stmt_vec_info stmt_info = NULL; |
92345349 | 2190 | stmt_vector_for_cost body_cost_vec; |
c3e7ee41 | 2191 | stmt_info_for_cost *ci; |
69f11a13 IR |
2192 | |
2193 | /* Calculate vector costs. */ | |
9771b263 | 2194 | FOR_EACH_VEC_ELT (slp_instances, i, instance) |
69f11a13 | 2195 | { |
92345349 | 2196 | body_cost_vec = SLP_INSTANCE_BODY_COST_VEC (instance); |
c3e7ee41 | 2197 | |
9771b263 | 2198 | FOR_EACH_VEC_ELT (body_cost_vec, j, ci) |
92345349 BS |
2199 | { |
2200 | stmt_info = ci->stmt ? vinfo_for_stmt (ci->stmt) : NULL; | |
2201 | (void) add_stmt_cost (target_cost_data, ci->count, ci->kind, | |
2202 | stmt_info, ci->misalign, vect_body); | |
2203 | } | |
69f11a13 IR |
2204 | } |
2205 | ||
2206 | /* Calculate scalar cost. */ | |
6eddf228 | 2207 | FOR_EACH_VEC_ELT (slp_instances, i, instance) |
69f11a13 | 2208 | { |
00f96dc9 | 2209 | auto_vec<bool, 20> life; |
ff4c81cc | 2210 | life.safe_grow_cleared (SLP_INSTANCE_GROUP_SIZE (instance)); |
292cba13 RB |
2211 | scalar_cost += vect_bb_slp_scalar_cost (BB_VINFO_BB (bb_vinfo), |
2212 | SLP_INSTANCE_TREE (instance), | |
ff4c81cc | 2213 | &life); |
69f11a13 IR |
2214 | } |
2215 | ||
c3e7ee41 | 2216 | /* Complete the target-specific cost calculation. */ |
92345349 BS |
2217 | finish_cost (BB_VINFO_TARGET_COST_DATA (bb_vinfo), &vec_prologue_cost, |
2218 | &vec_inside_cost, &vec_epilogue_cost); | |
2219 | ||
2220 | vec_outside_cost = vec_prologue_cost + vec_epilogue_cost; | |
c3e7ee41 | 2221 | |
73fbfcad | 2222 | if (dump_enabled_p ()) |
69f11a13 | 2223 | { |
78c60e3d SS |
2224 | dump_printf_loc (MSG_NOTE, vect_location, "Cost model analysis: \n"); |
2225 | dump_printf (MSG_NOTE, " Vector inside of basic block cost: %d\n", | |
2226 | vec_inside_cost); | |
2227 | dump_printf (MSG_NOTE, " Vector prologue cost: %d\n", vec_prologue_cost); | |
2228 | dump_printf (MSG_NOTE, " Vector epilogue cost: %d\n", vec_epilogue_cost); | |
e645e942 | 2229 | dump_printf (MSG_NOTE, " Scalar cost of basic block: %d\n", scalar_cost); |
69f11a13 IR |
2230 | } |
2231 | ||
2232 | /* Vectorization is profitable if its cost is less than the cost of scalar | |
2233 | version. */ | |
2234 | if (vec_outside_cost + vec_inside_cost >= scalar_cost) | |
2235 | return false; | |
2236 | ||
2237 | return true; | |
2238 | } | |
2239 | ||
2240 | /* Check if the basic block can be vectorized. */ | |
a70d6342 | 2241 | |
8e19f5a1 IR |
2242 | static bb_vec_info |
2243 | vect_slp_analyze_bb_1 (basic_block bb) | |
a70d6342 IR |
2244 | { |
2245 | bb_vec_info bb_vinfo; | |
9771b263 | 2246 | vec<slp_instance> slp_instances; |
a70d6342 | 2247 | slp_instance instance; |
8e19f5a1 | 2248 | int i; |
777e1f09 | 2249 | int min_vf = 2; |
1428105c | 2250 | unsigned n_stmts = 0; |
e4a707c4 | 2251 | |
a70d6342 IR |
2252 | bb_vinfo = new_bb_vec_info (bb); |
2253 | if (!bb_vinfo) | |
2254 | return NULL; | |
2255 | ||
1428105c | 2256 | if (!vect_analyze_data_refs (NULL, bb_vinfo, &min_vf, &n_stmts)) |
a70d6342 | 2257 | { |
73fbfcad | 2258 | if (dump_enabled_p ()) |
78c60e3d SS |
2259 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2260 | "not vectorized: unhandled data-ref in basic " | |
2261 | "block.\n"); | |
b8698a0f | 2262 | |
a70d6342 IR |
2263 | destroy_bb_vec_info (bb_vinfo); |
2264 | return NULL; | |
2265 | } | |
2266 | ||
fcac74a1 | 2267 | if (BB_VINFO_DATAREFS (bb_vinfo).length () < 2) |
a70d6342 | 2268 | { |
73fbfcad | 2269 | if (dump_enabled_p ()) |
78c60e3d SS |
2270 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2271 | "not vectorized: not enough data-refs in " | |
2272 | "basic block.\n"); | |
a70d6342 IR |
2273 | |
2274 | destroy_bb_vec_info (bb_vinfo); | |
2275 | return NULL; | |
2276 | } | |
2277 | ||
5abe1e05 RB |
2278 | if (!vect_analyze_data_ref_accesses (NULL, bb_vinfo)) |
2279 | { | |
2280 | if (dump_enabled_p ()) | |
2281 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2282 | "not vectorized: unhandled data access in " | |
2283 | "basic block.\n"); | |
2284 | ||
2285 | destroy_bb_vec_info (bb_vinfo); | |
2286 | return NULL; | |
2287 | } | |
2288 | ||
f5709183 IR |
2289 | vect_pattern_recog (NULL, bb_vinfo); |
2290 | ||
a70d6342 IR |
2291 | if (!vect_analyze_data_refs_alignment (NULL, bb_vinfo)) |
2292 | { | |
73fbfcad | 2293 | if (dump_enabled_p ()) |
78c60e3d SS |
2294 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2295 | "not vectorized: bad data alignment in basic " | |
2296 | "block.\n"); | |
b8698a0f | 2297 | |
a70d6342 IR |
2298 | destroy_bb_vec_info (bb_vinfo); |
2299 | return NULL; | |
2300 | } | |
b8698a0f | 2301 | |
a70d6342 IR |
2302 | /* Check the SLP opportunities in the basic block, analyze and build SLP |
2303 | trees. */ | |
1428105c | 2304 | if (!vect_analyze_slp (NULL, bb_vinfo, n_stmts)) |
a70d6342 | 2305 | { |
73fbfcad | 2306 | if (dump_enabled_p ()) |
78c60e3d SS |
2307 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2308 | "not vectorized: failed to find SLP opportunities " | |
2309 | "in basic block.\n"); | |
a70d6342 IR |
2310 | |
2311 | destroy_bb_vec_info (bb_vinfo); | |
2312 | return NULL; | |
2313 | } | |
b8698a0f | 2314 | |
a70d6342 IR |
2315 | slp_instances = BB_VINFO_SLP_INSTANCES (bb_vinfo); |
2316 | ||
2317 | /* Mark all the statements that we want to vectorize as pure SLP and | |
2318 | relevant. */ | |
9771b263 | 2319 | FOR_EACH_VEC_ELT (slp_instances, i, instance) |
a70d6342 IR |
2320 | { |
2321 | vect_mark_slp_stmts (SLP_INSTANCE_TREE (instance), pure_slp, -1); | |
2322 | vect_mark_slp_stmts_relevant (SLP_INSTANCE_TREE (instance)); | |
b8698a0f | 2323 | } |
a70d6342 | 2324 | |
5e6667b2 RB |
2325 | /* Mark all the statements that we do not want to vectorize. */ |
2326 | for (gimple_stmt_iterator gsi = gsi_start_bb (BB_VINFO_BB (bb_vinfo)); | |
2327 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
2328 | { | |
2329 | stmt_vec_info vinfo = vinfo_for_stmt (gsi_stmt (gsi)); | |
2330 | if (STMT_SLP_TYPE (vinfo) != pure_slp) | |
2331 | STMT_VINFO_VECTORIZABLE (vinfo) = false; | |
2332 | } | |
2333 | ||
2334 | /* Analyze dependences. At this point all stmts not participating in | |
2335 | vectorization have to be marked. Dependence analysis assumes | |
2336 | that we either vectorize all SLP instances or none at all. */ | |
2337 | if (!vect_slp_analyze_data_ref_dependences (bb_vinfo)) | |
2338 | { | |
2339 | if (dump_enabled_p ()) | |
2340 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2341 | "not vectorized: unhandled data dependence " | |
2342 | "in basic block.\n"); | |
2343 | ||
2344 | destroy_bb_vec_info (bb_vinfo); | |
2345 | return NULL; | |
2346 | } | |
2347 | ||
c3e7ee41 | 2348 | if (!vect_verify_datarefs_alignment (NULL, bb_vinfo)) |
38eec4c6 | 2349 | { |
73fbfcad | 2350 | if (dump_enabled_p ()) |
78c60e3d SS |
2351 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2352 | "not vectorized: unsupported alignment in basic " | |
2353 | "block.\n"); | |
38eec4c6 UW |
2354 | destroy_bb_vec_info (bb_vinfo); |
2355 | return NULL; | |
2356 | } | |
2357 | ||
a70d6342 IR |
2358 | if (!vect_slp_analyze_operations (bb_vinfo)) |
2359 | { | |
73fbfcad | 2360 | if (dump_enabled_p ()) |
e645e942 | 2361 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
78c60e3d | 2362 | "not vectorized: bad operation in basic block.\n"); |
a70d6342 IR |
2363 | |
2364 | destroy_bb_vec_info (bb_vinfo); | |
2365 | return NULL; | |
2366 | } | |
2367 | ||
69f11a13 | 2368 | /* Cost model: check if the vectorization is worthwhile. */ |
8b5e1202 | 2369 | if (!unlimited_cost_model (NULL) |
69f11a13 IR |
2370 | && !vect_bb_vectorization_profitable_p (bb_vinfo)) |
2371 | { | |
73fbfcad | 2372 | if (dump_enabled_p ()) |
78c60e3d SS |
2373 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2374 | "not vectorized: vectorization is not " | |
2375 | "profitable.\n"); | |
69f11a13 IR |
2376 | |
2377 | destroy_bb_vec_info (bb_vinfo); | |
2378 | return NULL; | |
2379 | } | |
2380 | ||
73fbfcad | 2381 | if (dump_enabled_p ()) |
78c60e3d SS |
2382 | dump_printf_loc (MSG_NOTE, vect_location, |
2383 | "Basic block will be vectorized using SLP\n"); | |
a70d6342 IR |
2384 | |
2385 | return bb_vinfo; | |
2386 | } | |
2387 | ||
2388 | ||
8e19f5a1 IR |
2389 | bb_vec_info |
2390 | vect_slp_analyze_bb (basic_block bb) | |
2391 | { | |
2392 | bb_vec_info bb_vinfo; | |
2393 | int insns = 0; | |
2394 | gimple_stmt_iterator gsi; | |
2395 | unsigned int vector_sizes; | |
2396 | ||
73fbfcad | 2397 | if (dump_enabled_p ()) |
78c60e3d | 2398 | dump_printf_loc (MSG_NOTE, vect_location, "===vect_slp_analyze_bb===\n"); |
8e19f5a1 IR |
2399 | |
2400 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2401 | { | |
2402 | gimple stmt = gsi_stmt (gsi); | |
2403 | if (!is_gimple_debug (stmt) | |
2404 | && !gimple_nop_p (stmt) | |
2405 | && gimple_code (stmt) != GIMPLE_LABEL) | |
2406 | insns++; | |
2407 | } | |
2408 | ||
2409 | if (insns > PARAM_VALUE (PARAM_SLP_MAX_INSNS_IN_BB)) | |
2410 | { | |
73fbfcad | 2411 | if (dump_enabled_p ()) |
78c60e3d SS |
2412 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2413 | "not vectorized: too many instructions in " | |
2414 | "basic block.\n"); | |
8e19f5a1 IR |
2415 | |
2416 | return NULL; | |
2417 | } | |
2418 | ||
2419 | /* Autodetect first vector size we try. */ | |
2420 | current_vector_size = 0; | |
2421 | vector_sizes = targetm.vectorize.autovectorize_vector_sizes (); | |
2422 | ||
2423 | while (1) | |
2424 | { | |
2425 | bb_vinfo = vect_slp_analyze_bb_1 (bb); | |
2426 | if (bb_vinfo) | |
2427 | return bb_vinfo; | |
2428 | ||
2429 | destroy_bb_vec_info (bb_vinfo); | |
2430 | ||
2431 | vector_sizes &= ~current_vector_size; | |
2432 | if (vector_sizes == 0 | |
2433 | || current_vector_size == 0) | |
2434 | return NULL; | |
2435 | ||
2436 | /* Try the next biggest vector size. */ | |
2437 | current_vector_size = 1 << floor_log2 (vector_sizes); | |
73fbfcad | 2438 | if (dump_enabled_p ()) |
78c60e3d SS |
2439 | dump_printf_loc (MSG_NOTE, vect_location, |
2440 | "***** Re-trying analysis with " | |
2441 | "vector size %d\n", current_vector_size); | |
8e19f5a1 IR |
2442 | } |
2443 | } | |
2444 | ||
2445 | ||
b8698a0f | 2446 | /* SLP costs are calculated according to SLP instance unrolling factor (i.e., |
ff802fa1 IR |
2447 | the number of created vector stmts depends on the unrolling factor). |
2448 | However, the actual number of vector stmts for every SLP node depends on | |
2449 | VF which is set later in vect_analyze_operations (). Hence, SLP costs | |
2450 | should be updated. In this function we assume that the inside costs | |
2451 | calculated in vect_model_xxx_cost are linear in ncopies. */ | |
ebfd146a IR |
2452 | |
2453 | void | |
2454 | vect_update_slp_costs_according_to_vf (loop_vec_info loop_vinfo) | |
2455 | { | |
c3e7ee41 | 2456 | unsigned int i, j, vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); |
9771b263 | 2457 | vec<slp_instance> slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo); |
ebfd146a | 2458 | slp_instance instance; |
92345349 | 2459 | stmt_vector_for_cost body_cost_vec; |
c3e7ee41 | 2460 | stmt_info_for_cost *si; |
92345349 | 2461 | void *data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); |
ebfd146a | 2462 | |
73fbfcad | 2463 | if (dump_enabled_p ()) |
78c60e3d | 2464 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 2465 | "=== vect_update_slp_costs_according_to_vf ===\n"); |
ebfd146a | 2466 | |
9771b263 | 2467 | FOR_EACH_VEC_ELT (slp_instances, i, instance) |
c3e7ee41 BS |
2468 | { |
2469 | /* We assume that costs are linear in ncopies. */ | |
2470 | int ncopies = vf / SLP_INSTANCE_UNROLLING_FACTOR (instance); | |
2471 | ||
2472 | /* Record the instance's instructions in the target cost model. | |
2473 | This was delayed until here because the count of instructions | |
2474 | isn't known beforehand. */ | |
92345349 | 2475 | body_cost_vec = SLP_INSTANCE_BODY_COST_VEC (instance); |
c3e7ee41 | 2476 | |
9771b263 | 2477 | FOR_EACH_VEC_ELT (body_cost_vec, j, si) |
92345349 BS |
2478 | (void) add_stmt_cost (data, si->count * ncopies, si->kind, |
2479 | vinfo_for_stmt (si->stmt), si->misalign, | |
2480 | vect_body); | |
c3e7ee41 | 2481 | } |
ebfd146a IR |
2482 | } |
2483 | ||
a70d6342 | 2484 | |
b8698a0f L |
2485 | /* For constant and loop invariant defs of SLP_NODE this function returns |
2486 | (vector) defs (VEC_OPRNDS) that will be used in the vectorized stmts. | |
d59dc888 IR |
2487 | OP_NUM determines if we gather defs for operand 0 or operand 1 of the RHS of |
2488 | scalar stmts. NUMBER_OF_VECTORS is the number of vector defs to create. | |
b5aeb3bb IR |
2489 | REDUC_INDEX is the index of the reduction operand in the statements, unless |
2490 | it is -1. */ | |
ebfd146a IR |
2491 | |
2492 | static void | |
9dc3f7de | 2493 | vect_get_constant_vectors (tree op, slp_tree slp_node, |
9771b263 | 2494 | vec<tree> *vec_oprnds, |
b5aeb3bb IR |
2495 | unsigned int op_num, unsigned int number_of_vectors, |
2496 | int reduc_index) | |
ebfd146a | 2497 | { |
9771b263 DN |
2498 | vec<gimple> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
2499 | gimple stmt = stmts[0]; | |
ebfd146a | 2500 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
d2a12ae7 | 2501 | unsigned nunits; |
ebfd146a | 2502 | tree vec_cst; |
d2a12ae7 RG |
2503 | tree *elts; |
2504 | unsigned j, number_of_places_left_in_vector; | |
ebfd146a | 2505 | tree vector_type; |
9dc3f7de | 2506 | tree vop; |
9771b263 | 2507 | int group_size = stmts.length (); |
ebfd146a | 2508 | unsigned int vec_num, i; |
d2a12ae7 | 2509 | unsigned number_of_copies = 1; |
9771b263 DN |
2510 | vec<tree> voprnds; |
2511 | voprnds.create (number_of_vectors); | |
ebfd146a | 2512 | bool constant_p, is_store; |
b5aeb3bb | 2513 | tree neutral_op = NULL; |
bac430c9 | 2514 | enum tree_code code = gimple_expr_code (stmt); |
0e93a64e IR |
2515 | gimple def_stmt; |
2516 | struct loop *loop; | |
13396b6e | 2517 | gimple_seq ctor_seq = NULL; |
b5aeb3bb | 2518 | |
29ed4920 IR |
2519 | if (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
2520 | && reduc_index != -1) | |
b5aeb3bb | 2521 | { |
b5aeb3bb | 2522 | op_num = reduc_index - 1; |
9dc3f7de | 2523 | op = gimple_op (stmt, reduc_index); |
b5aeb3bb | 2524 | /* For additional copies (see the explanation of NUMBER_OF_COPIES below) |
ff802fa1 | 2525 | we need either neutral operands or the original operands. See |
b5aeb3bb IR |
2526 | get_initial_def_for_reduction() for details. */ |
2527 | switch (code) | |
2528 | { | |
2529 | case WIDEN_SUM_EXPR: | |
2530 | case DOT_PROD_EXPR: | |
2531 | case PLUS_EXPR: | |
2532 | case MINUS_EXPR: | |
2533 | case BIT_IOR_EXPR: | |
2534 | case BIT_XOR_EXPR: | |
2535 | if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (op))) | |
2536 | neutral_op = build_real (TREE_TYPE (op), dconst0); | |
2537 | else | |
2538 | neutral_op = build_int_cst (TREE_TYPE (op), 0); | |
2539 | ||
2540 | break; | |
2541 | ||
2542 | case MULT_EXPR: | |
b5aeb3bb IR |
2543 | if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (op))) |
2544 | neutral_op = build_real (TREE_TYPE (op), dconst1); | |
2545 | else | |
2546 | neutral_op = build_int_cst (TREE_TYPE (op), 1); | |
2547 | ||
2548 | break; | |
2549 | ||
c1e822d5 IR |
2550 | case BIT_AND_EXPR: |
2551 | neutral_op = build_int_cst (TREE_TYPE (op), -1); | |
2552 | break; | |
2553 | ||
f1485e5b RB |
2554 | /* For MIN/MAX we don't have an easy neutral operand but |
2555 | the initial values can be used fine here. Only for | |
2556 | a reduction chain we have to force a neutral element. */ | |
2557 | case MAX_EXPR: | |
2558 | case MIN_EXPR: | |
2559 | if (!GROUP_FIRST_ELEMENT (stmt_vinfo)) | |
2560 | neutral_op = NULL; | |
2561 | else | |
2562 | { | |
2563 | def_stmt = SSA_NAME_DEF_STMT (op); | |
2564 | loop = (gimple_bb (stmt))->loop_father; | |
2565 | neutral_op = PHI_ARG_DEF_FROM_EDGE (def_stmt, | |
2566 | loop_preheader_edge (loop)); | |
2567 | } | |
2568 | break; | |
0e93a64e | 2569 | |
b5aeb3bb | 2570 | default: |
0e93a64e | 2571 | neutral_op = NULL; |
b5aeb3bb IR |
2572 | } |
2573 | } | |
ebfd146a IR |
2574 | |
2575 | if (STMT_VINFO_DATA_REF (stmt_vinfo)) | |
2576 | { | |
2577 | is_store = true; | |
2578 | op = gimple_assign_rhs1 (stmt); | |
2579 | } | |
2580 | else | |
9dc3f7de IR |
2581 | is_store = false; |
2582 | ||
2583 | gcc_assert (op); | |
ebfd146a IR |
2584 | |
2585 | if (CONSTANT_CLASS_P (op)) | |
d59dc888 | 2586 | constant_p = true; |
ebfd146a | 2587 | else |
d59dc888 IR |
2588 | constant_p = false; |
2589 | ||
9dc3f7de | 2590 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
cd481d83 | 2591 | gcc_assert (vector_type); |
ebfd146a IR |
2592 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); |
2593 | ||
2594 | /* NUMBER_OF_COPIES is the number of times we need to use the same values in | |
b8698a0f | 2595 | created vectors. It is greater than 1 if unrolling is performed. |
ebfd146a IR |
2596 | |
2597 | For example, we have two scalar operands, s1 and s2 (e.g., group of | |
2598 | strided accesses of size two), while NUNITS is four (i.e., four scalars | |
f7e531cf IR |
2599 | of this type can be packed in a vector). The output vector will contain |
2600 | two copies of each scalar operand: {s1, s2, s1, s2}. (NUMBER_OF_COPIES | |
ebfd146a IR |
2601 | will be 2). |
2602 | ||
b8698a0f | 2603 | If GROUP_SIZE > NUNITS, the scalars will be split into several vectors |
ebfd146a IR |
2604 | containing the operands. |
2605 | ||
2606 | For example, NUNITS is four as before, and the group size is 8 | |
f7e531cf | 2607 | (s1, s2, ..., s8). We will create two vectors {s1, s2, s3, s4} and |
ebfd146a | 2608 | {s5, s6, s7, s8}. */ |
b8698a0f | 2609 | |
ebfd146a IR |
2610 | number_of_copies = least_common_multiple (nunits, group_size) / group_size; |
2611 | ||
2612 | number_of_places_left_in_vector = nunits; | |
d2a12ae7 | 2613 | elts = XALLOCAVEC (tree, nunits); |
ebfd146a IR |
2614 | for (j = 0; j < number_of_copies; j++) |
2615 | { | |
9771b263 | 2616 | for (i = group_size - 1; stmts.iterate (i, &stmt); i--) |
ebfd146a IR |
2617 | { |
2618 | if (is_store) | |
2619 | op = gimple_assign_rhs1 (stmt); | |
bac430c9 | 2620 | else |
f7e531cf | 2621 | { |
bac430c9 | 2622 | switch (code) |
f7e531cf | 2623 | { |
bac430c9 IR |
2624 | case COND_EXPR: |
2625 | if (op_num == 0 || op_num == 1) | |
2626 | { | |
2627 | tree cond = gimple_assign_rhs1 (stmt); | |
2628 | op = TREE_OPERAND (cond, op_num); | |
2629 | } | |
2630 | else | |
2631 | { | |
2632 | if (op_num == 2) | |
2633 | op = gimple_assign_rhs2 (stmt); | |
2634 | else | |
2635 | op = gimple_assign_rhs3 (stmt); | |
2636 | } | |
2637 | break; | |
2638 | ||
2639 | case CALL_EXPR: | |
2640 | op = gimple_call_arg (stmt, op_num); | |
2641 | break; | |
2642 | ||
b84b294a JJ |
2643 | case LSHIFT_EXPR: |
2644 | case RSHIFT_EXPR: | |
2645 | case LROTATE_EXPR: | |
2646 | case RROTATE_EXPR: | |
2647 | op = gimple_op (stmt, op_num + 1); | |
2648 | /* Unlike the other binary operators, shifts/rotates have | |
2649 | the shift count being int, instead of the same type as | |
2650 | the lhs, so make sure the scalar is the right type if | |
2651 | we are dealing with vectors of | |
2652 | long long/long/short/char. */ | |
793d9a16 | 2653 | if (op_num == 1 && TREE_CODE (op) == INTEGER_CST) |
b84b294a JJ |
2654 | op = fold_convert (TREE_TYPE (vector_type), op); |
2655 | break; | |
2656 | ||
bac430c9 IR |
2657 | default: |
2658 | op = gimple_op (stmt, op_num + 1); | |
b84b294a | 2659 | break; |
f7e531cf IR |
2660 | } |
2661 | } | |
b8698a0f | 2662 | |
b5aeb3bb IR |
2663 | if (reduc_index != -1) |
2664 | { | |
0e93a64e IR |
2665 | loop = (gimple_bb (stmt))->loop_father; |
2666 | def_stmt = SSA_NAME_DEF_STMT (op); | |
b5aeb3bb IR |
2667 | |
2668 | gcc_assert (loop); | |
b010117a IR |
2669 | |
2670 | /* Get the def before the loop. In reduction chain we have only | |
2671 | one initial value. */ | |
2672 | if ((j != (number_of_copies - 1) | |
2673 | || (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) | |
2674 | && i != 0)) | |
2675 | && neutral_op) | |
b5aeb3bb | 2676 | op = neutral_op; |
b010117a IR |
2677 | else |
2678 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, | |
2679 | loop_preheader_edge (loop)); | |
b5aeb3bb IR |
2680 | } |
2681 | ||
ebfd146a | 2682 | /* Create 'vect_ = {op0,op1,...,opn}'. */ |
ebfd146a | 2683 | number_of_places_left_in_vector--; |
13396b6e | 2684 | if (!types_compatible_p (TREE_TYPE (vector_type), TREE_TYPE (op))) |
50eeef09 | 2685 | { |
793d9a16 | 2686 | if (CONSTANT_CLASS_P (op)) |
13396b6e JJ |
2687 | { |
2688 | op = fold_unary (VIEW_CONVERT_EXPR, | |
2689 | TREE_TYPE (vector_type), op); | |
2690 | gcc_assert (op && CONSTANT_CLASS_P (op)); | |
2691 | } | |
2692 | else | |
2693 | { | |
b731b390 | 2694 | tree new_temp = make_ssa_name (TREE_TYPE (vector_type)); |
13396b6e | 2695 | gimple init_stmt; |
0d0e4a03 | 2696 | op = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (vector_type), op); |
13396b6e | 2697 | init_stmt |
0d0e4a03 | 2698 | = gimple_build_assign (new_temp, VIEW_CONVERT_EXPR, op); |
13396b6e JJ |
2699 | gimple_seq_add_stmt (&ctor_seq, init_stmt); |
2700 | op = new_temp; | |
2701 | } | |
50eeef09 | 2702 | } |
d2a12ae7 | 2703 | elts[number_of_places_left_in_vector] = op; |
793d9a16 RB |
2704 | if (!CONSTANT_CLASS_P (op)) |
2705 | constant_p = false; | |
ebfd146a IR |
2706 | |
2707 | if (number_of_places_left_in_vector == 0) | |
2708 | { | |
2709 | number_of_places_left_in_vector = nunits; | |
2710 | ||
2711 | if (constant_p) | |
d2a12ae7 | 2712 | vec_cst = build_vector (vector_type, elts); |
ebfd146a | 2713 | else |
d2a12ae7 | 2714 | { |
9771b263 | 2715 | vec<constructor_elt, va_gc> *v; |
d2a12ae7 | 2716 | unsigned k; |
9771b263 | 2717 | vec_alloc (v, nunits); |
d2a12ae7 RG |
2718 | for (k = 0; k < nunits; ++k) |
2719 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, elts[k]); | |
2720 | vec_cst = build_constructor (vector_type, v); | |
2721 | } | |
9771b263 DN |
2722 | voprnds.quick_push (vect_init_vector (stmt, vec_cst, |
2723 | vector_type, NULL)); | |
13396b6e JJ |
2724 | if (ctor_seq != NULL) |
2725 | { | |
9771b263 | 2726 | gimple init_stmt = SSA_NAME_DEF_STMT (voprnds.last ()); |
13396b6e JJ |
2727 | gimple_stmt_iterator gsi = gsi_for_stmt (init_stmt); |
2728 | gsi_insert_seq_before_without_update (&gsi, ctor_seq, | |
2729 | GSI_SAME_STMT); | |
2730 | ctor_seq = NULL; | |
2731 | } | |
ebfd146a IR |
2732 | } |
2733 | } | |
2734 | } | |
2735 | ||
b8698a0f | 2736 | /* Since the vectors are created in the reverse order, we should invert |
ebfd146a | 2737 | them. */ |
9771b263 | 2738 | vec_num = voprnds.length (); |
d2a12ae7 | 2739 | for (j = vec_num; j != 0; j--) |
ebfd146a | 2740 | { |
9771b263 DN |
2741 | vop = voprnds[j - 1]; |
2742 | vec_oprnds->quick_push (vop); | |
ebfd146a IR |
2743 | } |
2744 | ||
9771b263 | 2745 | voprnds.release (); |
ebfd146a IR |
2746 | |
2747 | /* In case that VF is greater than the unrolling factor needed for the SLP | |
b8698a0f L |
2748 | group of stmts, NUMBER_OF_VECTORS to be created is greater than |
2749 | NUMBER_OF_SCALARS/NUNITS or NUNITS/NUMBER_OF_SCALARS, and hence we have | |
ebfd146a | 2750 | to replicate the vectors. */ |
9771b263 | 2751 | while (number_of_vectors > vec_oprnds->length ()) |
ebfd146a | 2752 | { |
b5aeb3bb IR |
2753 | tree neutral_vec = NULL; |
2754 | ||
2755 | if (neutral_op) | |
2756 | { | |
2757 | if (!neutral_vec) | |
b9acc9f1 | 2758 | neutral_vec = build_vector_from_val (vector_type, neutral_op); |
b5aeb3bb | 2759 | |
9771b263 | 2760 | vec_oprnds->quick_push (neutral_vec); |
b5aeb3bb IR |
2761 | } |
2762 | else | |
2763 | { | |
9771b263 DN |
2764 | for (i = 0; vec_oprnds->iterate (i, &vop) && i < vec_num; i++) |
2765 | vec_oprnds->quick_push (vop); | |
b5aeb3bb | 2766 | } |
ebfd146a IR |
2767 | } |
2768 | } | |
2769 | ||
2770 | ||
2771 | /* Get vectorized definitions from SLP_NODE that contains corresponding | |
2772 | vectorized def-stmts. */ | |
2773 | ||
2774 | static void | |
9771b263 | 2775 | vect_get_slp_vect_defs (slp_tree slp_node, vec<tree> *vec_oprnds) |
ebfd146a IR |
2776 | { |
2777 | tree vec_oprnd; | |
2778 | gimple vec_def_stmt; | |
2779 | unsigned int i; | |
2780 | ||
9771b263 | 2781 | gcc_assert (SLP_TREE_VEC_STMTS (slp_node).exists ()); |
ebfd146a | 2782 | |
9771b263 | 2783 | FOR_EACH_VEC_ELT (SLP_TREE_VEC_STMTS (slp_node), i, vec_def_stmt) |
ebfd146a IR |
2784 | { |
2785 | gcc_assert (vec_def_stmt); | |
2786 | vec_oprnd = gimple_get_lhs (vec_def_stmt); | |
9771b263 | 2787 | vec_oprnds->quick_push (vec_oprnd); |
ebfd146a IR |
2788 | } |
2789 | } | |
2790 | ||
2791 | ||
b8698a0f L |
2792 | /* Get vectorized definitions for SLP_NODE. |
2793 | If the scalar definitions are loop invariants or constants, collect them and | |
ebfd146a IR |
2794 | call vect_get_constant_vectors() to create vector stmts. |
2795 | Otherwise, the def-stmts must be already vectorized and the vectorized stmts | |
d092494c IR |
2796 | must be stored in the corresponding child of SLP_NODE, and we call |
2797 | vect_get_slp_vect_defs () to retrieve them. */ | |
b8698a0f | 2798 | |
ebfd146a | 2799 | void |
9771b263 | 2800 | vect_get_slp_defs (vec<tree> ops, slp_tree slp_node, |
37b5ec8f | 2801 | vec<vec<tree> > *vec_oprnds, int reduc_index) |
ebfd146a | 2802 | { |
e44978dc | 2803 | gimple first_stmt; |
d092494c IR |
2804 | int number_of_vects = 0, i; |
2805 | unsigned int child_index = 0; | |
b8698a0f | 2806 | HOST_WIDE_INT lhs_size_unit, rhs_size_unit; |
d092494c | 2807 | slp_tree child = NULL; |
37b5ec8f | 2808 | vec<tree> vec_defs; |
e44978dc | 2809 | tree oprnd; |
d092494c | 2810 | bool vectorized_defs; |
ebfd146a | 2811 | |
9771b263 DN |
2812 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
2813 | FOR_EACH_VEC_ELT (ops, i, oprnd) | |
ebfd146a | 2814 | { |
d092494c IR |
2815 | /* For each operand we check if it has vectorized definitions in a child |
2816 | node or we need to create them (for invariants and constants). We | |
2817 | check if the LHS of the first stmt of the next child matches OPRND. | |
2818 | If it does, we found the correct child. Otherwise, we call | |
2819 | vect_get_constant_vectors (), and not advance CHILD_INDEX in order | |
2820 | to check this child node for the next operand. */ | |
2821 | vectorized_defs = false; | |
9771b263 | 2822 | if (SLP_TREE_CHILDREN (slp_node).length () > child_index) |
ebfd146a | 2823 | { |
01d8bf07 | 2824 | child = SLP_TREE_CHILDREN (slp_node)[child_index]; |
d092494c | 2825 | |
e44978dc RB |
2826 | /* We have to check both pattern and original def, if available. */ |
2827 | gimple first_def = SLP_TREE_SCALAR_STMTS (child)[0]; | |
2828 | gimple related = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (first_def)); | |
ebfd146a | 2829 | |
e44978dc RB |
2830 | if (operand_equal_p (oprnd, gimple_get_lhs (first_def), 0) |
2831 | || (related | |
2832 | && operand_equal_p (oprnd, gimple_get_lhs (related), 0))) | |
2833 | { | |
2834 | /* The number of vector defs is determined by the number of | |
2835 | vector statements in the node from which we get those | |
d092494c | 2836 | statements. */ |
e44978dc RB |
2837 | number_of_vects = SLP_TREE_NUMBER_OF_VEC_STMTS (child); |
2838 | vectorized_defs = true; | |
d092494c | 2839 | child_index++; |
e44978dc | 2840 | } |
d092494c | 2841 | } |
ebfd146a | 2842 | |
d092494c IR |
2843 | if (!vectorized_defs) |
2844 | { | |
2845 | if (i == 0) | |
2846 | { | |
2847 | number_of_vects = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
2848 | /* Number of vector stmts was calculated according to LHS in | |
2849 | vect_schedule_slp_instance (), fix it by replacing LHS with | |
2850 | RHS, if necessary. See vect_get_smallest_scalar_type () for | |
2851 | details. */ | |
2852 | vect_get_smallest_scalar_type (first_stmt, &lhs_size_unit, | |
2853 | &rhs_size_unit); | |
2854 | if (rhs_size_unit != lhs_size_unit) | |
2855 | { | |
2856 | number_of_vects *= rhs_size_unit; | |
2857 | number_of_vects /= lhs_size_unit; | |
2858 | } | |
2859 | } | |
2860 | } | |
b5aeb3bb | 2861 | |
d092494c | 2862 | /* Allocate memory for vectorized defs. */ |
37b5ec8f JJ |
2863 | vec_defs = vNULL; |
2864 | vec_defs.create (number_of_vects); | |
ebfd146a | 2865 | |
d092494c IR |
2866 | /* For reduction defs we call vect_get_constant_vectors (), since we are |
2867 | looking for initial loop invariant values. */ | |
2868 | if (vectorized_defs && reduc_index == -1) | |
2869 | /* The defs are already vectorized. */ | |
37b5ec8f | 2870 | vect_get_slp_vect_defs (child, &vec_defs); |
d092494c IR |
2871 | else |
2872 | /* Build vectors from scalar defs. */ | |
37b5ec8f | 2873 | vect_get_constant_vectors (oprnd, slp_node, &vec_defs, i, |
d092494c | 2874 | number_of_vects, reduc_index); |
ebfd146a | 2875 | |
37b5ec8f | 2876 | vec_oprnds->quick_push (vec_defs); |
ebfd146a | 2877 | |
d092494c IR |
2878 | /* For reductions, we only need initial values. */ |
2879 | if (reduc_index != -1) | |
2880 | return; | |
2881 | } | |
ebfd146a IR |
2882 | } |
2883 | ||
a70d6342 | 2884 | |
b8698a0f | 2885 | /* Create NCOPIES permutation statements using the mask MASK_BYTES (by |
ebfd146a IR |
2886 | building a vector of type MASK_TYPE from it) and two input vectors placed in |
2887 | DR_CHAIN at FIRST_VEC_INDX and SECOND_VEC_INDX for the first copy and | |
2888 | shifting by STRIDE elements of DR_CHAIN for every copy. | |
2889 | (STRIDE is the number of vectorized stmts for NODE divided by the number of | |
b8698a0f | 2890 | copies). |
ebfd146a IR |
2891 | VECT_STMTS_COUNTER specifies the index in the vectorized stmts of NODE, where |
2892 | the created stmts must be inserted. */ | |
2893 | ||
2894 | static inline void | |
b8698a0f | 2895 | vect_create_mask_and_perm (gimple stmt, gimple next_scalar_stmt, |
faf63e39 | 2896 | tree mask, int first_vec_indx, int second_vec_indx, |
b8698a0f | 2897 | gimple_stmt_iterator *gsi, slp_tree node, |
9771b263 | 2898 | tree vectype, vec<tree> dr_chain, |
ebfd146a IR |
2899 | int ncopies, int vect_stmts_counter) |
2900 | { | |
faf63e39 | 2901 | tree perm_dest; |
ebfd146a IR |
2902 | gimple perm_stmt = NULL; |
2903 | stmt_vec_info next_stmt_info; | |
0f900dfa | 2904 | int i, stride; |
ebfd146a | 2905 | tree first_vec, second_vec, data_ref; |
ebfd146a | 2906 | |
ebfd146a | 2907 | stride = SLP_TREE_NUMBER_OF_VEC_STMTS (node) / ncopies; |
ebfd146a | 2908 | |
b8698a0f | 2909 | /* Initialize the vect stmts of NODE to properly insert the generated |
ebfd146a | 2910 | stmts later. */ |
9771b263 | 2911 | for (i = SLP_TREE_VEC_STMTS (node).length (); |
ebfd146a | 2912 | i < (int) SLP_TREE_NUMBER_OF_VEC_STMTS (node); i++) |
9771b263 | 2913 | SLP_TREE_VEC_STMTS (node).quick_push (NULL); |
ebfd146a IR |
2914 | |
2915 | perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype); | |
2916 | for (i = 0; i < ncopies; i++) | |
2917 | { | |
9771b263 DN |
2918 | first_vec = dr_chain[first_vec_indx]; |
2919 | second_vec = dr_chain[second_vec_indx]; | |
ebfd146a | 2920 | |
ebfd146a | 2921 | /* Generate the permute statement. */ |
0d0e4a03 JJ |
2922 | perm_stmt = gimple_build_assign (perm_dest, VEC_PERM_EXPR, |
2923 | first_vec, second_vec, mask); | |
ebfd146a | 2924 | data_ref = make_ssa_name (perm_dest, perm_stmt); |
2635892a | 2925 | gimple_set_lhs (perm_stmt, data_ref); |
ebfd146a | 2926 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
ebfd146a | 2927 | |
b8698a0f | 2928 | /* Store the vector statement in NODE. */ |
9771b263 | 2929 | SLP_TREE_VEC_STMTS (node)[stride * i + vect_stmts_counter] = perm_stmt; |
ebfd146a IR |
2930 | |
2931 | first_vec_indx += stride; | |
2932 | second_vec_indx += stride; | |
2933 | } | |
2934 | ||
2935 | /* Mark the scalar stmt as vectorized. */ | |
2936 | next_stmt_info = vinfo_for_stmt (next_scalar_stmt); | |
2937 | STMT_VINFO_VEC_STMT (next_stmt_info) = perm_stmt; | |
2938 | } | |
2939 | ||
2940 | ||
b8698a0f | 2941 | /* Given FIRST_MASK_ELEMENT - the mask element in element representation, |
ebfd146a | 2942 | return in CURRENT_MASK_ELEMENT its equivalent in target specific |
ff802fa1 | 2943 | representation. Check that the mask is valid and return FALSE if not. |
ebfd146a IR |
2944 | Return TRUE in NEED_NEXT_VECTOR if the permutation requires to move to |
2945 | the next vector, i.e., the current first vector is not needed. */ | |
b8698a0f | 2946 | |
ebfd146a | 2947 | static bool |
b8698a0f | 2948 | vect_get_mask_element (gimple stmt, int first_mask_element, int m, |
ebfd146a | 2949 | int mask_nunits, bool only_one_vec, int index, |
22e4dee7 | 2950 | unsigned char *mask, int *current_mask_element, |
694a4f61 IR |
2951 | bool *need_next_vector, int *number_of_mask_fixes, |
2952 | bool *mask_fixed, bool *needs_first_vector) | |
ebfd146a IR |
2953 | { |
2954 | int i; | |
ebfd146a IR |
2955 | |
2956 | /* Convert to target specific representation. */ | |
2957 | *current_mask_element = first_mask_element + m; | |
2958 | /* Adjust the value in case it's a mask for second and third vectors. */ | |
694a4f61 | 2959 | *current_mask_element -= mask_nunits * (*number_of_mask_fixes - 1); |
ebfd146a IR |
2960 | |
2961 | if (*current_mask_element < mask_nunits) | |
694a4f61 | 2962 | *needs_first_vector = true; |
ebfd146a IR |
2963 | |
2964 | /* We have only one input vector to permute but the mask accesses values in | |
2965 | the next vector as well. */ | |
2966 | if (only_one_vec && *current_mask_element >= mask_nunits) | |
2967 | { | |
73fbfcad | 2968 | if (dump_enabled_p ()) |
ebfd146a | 2969 | { |
e645e942 | 2970 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
78c60e3d SS |
2971 | "permutation requires at least two vectors "); |
2972 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 2973 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a IR |
2974 | } |
2975 | ||
2976 | return false; | |
2977 | } | |
2978 | ||
2979 | /* The mask requires the next vector. */ | |
496d3346 | 2980 | while (*current_mask_element >= mask_nunits * 2) |
ebfd146a | 2981 | { |
694a4f61 | 2982 | if (*needs_first_vector || *mask_fixed) |
ebfd146a IR |
2983 | { |
2984 | /* We either need the first vector too or have already moved to the | |
b8698a0f | 2985 | next vector. In both cases, this permutation needs three |
ebfd146a | 2986 | vectors. */ |
73fbfcad | 2987 | if (dump_enabled_p ()) |
ebfd146a | 2988 | { |
78c60e3d SS |
2989 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2990 | "permutation requires at " | |
2991 | "least three vectors "); | |
2992 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 2993 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a IR |
2994 | } |
2995 | ||
2996 | return false; | |
2997 | } | |
2998 | ||
2999 | /* We move to the next vector, dropping the first one and working with | |
3000 | the second and the third - we need to adjust the values of the mask | |
3001 | accordingly. */ | |
694a4f61 | 3002 | *current_mask_element -= mask_nunits * *number_of_mask_fixes; |
ebfd146a IR |
3003 | |
3004 | for (i = 0; i < index; i++) | |
694a4f61 | 3005 | mask[i] -= mask_nunits * *number_of_mask_fixes; |
ebfd146a | 3006 | |
694a4f61 IR |
3007 | (*number_of_mask_fixes)++; |
3008 | *mask_fixed = true; | |
ebfd146a IR |
3009 | } |
3010 | ||
694a4f61 | 3011 | *need_next_vector = *mask_fixed; |
ebfd146a IR |
3012 | |
3013 | /* This was the last element of this mask. Start a new one. */ | |
3014 | if (index == mask_nunits - 1) | |
3015 | { | |
694a4f61 IR |
3016 | *number_of_mask_fixes = 1; |
3017 | *mask_fixed = false; | |
3018 | *needs_first_vector = false; | |
ebfd146a IR |
3019 | } |
3020 | ||
3021 | return true; | |
3022 | } | |
3023 | ||
3024 | ||
3025 | /* Generate vector permute statements from a list of loads in DR_CHAIN. | |
3026 | If ANALYZE_ONLY is TRUE, only check that it is possible to create valid | |
01d8bf07 RB |
3027 | permute statements for the SLP node NODE of the SLP instance |
3028 | SLP_NODE_INSTANCE. */ | |
3029 | ||
ebfd146a | 3030 | bool |
01d8bf07 | 3031 | vect_transform_slp_perm_load (slp_tree node, vec<tree> dr_chain, |
ebfd146a IR |
3032 | gimple_stmt_iterator *gsi, int vf, |
3033 | slp_instance slp_node_instance, bool analyze_only) | |
3034 | { | |
01d8bf07 | 3035 | gimple stmt = SLP_TREE_SCALAR_STMTS (node)[0]; |
ebfd146a IR |
3036 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
3037 | tree mask_element_type = NULL_TREE, mask_type; | |
2635892a | 3038 | int i, j, k, nunits, vec_index = 0, scalar_index; |
2635892a | 3039 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a IR |
3040 | gimple next_scalar_stmt; |
3041 | int group_size = SLP_INSTANCE_GROUP_SIZE (slp_node_instance); | |
3042 | int first_mask_element; | |
22e4dee7 RH |
3043 | int index, unroll_factor, current_mask_element, ncopies; |
3044 | unsigned char *mask; | |
ebfd146a IR |
3045 | bool only_one_vec = false, need_next_vector = false; |
3046 | int first_vec_index, second_vec_index, orig_vec_stmts_num, vect_stmts_counter; | |
694a4f61 IR |
3047 | int number_of_mask_fixes = 1; |
3048 | bool mask_fixed = false; | |
3049 | bool needs_first_vector = false; | |
ef4bddc2 | 3050 | machine_mode mode; |
ebfd146a | 3051 | |
22e4dee7 RH |
3052 | mode = TYPE_MODE (vectype); |
3053 | ||
3054 | if (!can_vec_perm_p (mode, false, NULL)) | |
ebfd146a | 3055 | { |
73fbfcad | 3056 | if (dump_enabled_p ()) |
ebfd146a | 3057 | { |
78c60e3d SS |
3058 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3059 | "no vect permute for "); | |
3060 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 3061 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
ebfd146a | 3062 | } |
2635892a | 3063 | return false; |
ebfd146a IR |
3064 | } |
3065 | ||
2635892a RH |
3066 | /* The generic VEC_PERM_EXPR code always uses an integral type of the |
3067 | same size as the vector element being permuted. */ | |
96f9265a RG |
3068 | mask_element_type = lang_hooks.types.type_for_mode |
3069 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
ebfd146a | 3070 | mask_type = get_vectype_for_scalar_type (mask_element_type); |
ebfd146a | 3071 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 | 3072 | mask = XALLOCAVEC (unsigned char, nunits); |
ebfd146a IR |
3073 | unroll_factor = SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance); |
3074 | ||
3075 | /* The number of vector stmts to generate based only on SLP_NODE_INSTANCE | |
3076 | unrolling factor. */ | |
b8698a0f | 3077 | orig_vec_stmts_num = group_size * |
ebfd146a IR |
3078 | SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance) / nunits; |
3079 | if (orig_vec_stmts_num == 1) | |
3080 | only_one_vec = true; | |
3081 | ||
b8698a0f | 3082 | /* Number of copies is determined by the final vectorization factor |
ebfd146a | 3083 | relatively to SLP_NODE_INSTANCE unrolling factor. */ |
b8698a0f | 3084 | ncopies = vf / SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance); |
ebfd146a | 3085 | |
01d8bf07 RB |
3086 | if (!STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
3087 | return false; | |
3088 | ||
b8698a0f L |
3089 | /* Generate permutation masks for every NODE. Number of masks for each NODE |
3090 | is equal to GROUP_SIZE. | |
3091 | E.g., we have a group of three nodes with three loads from the same | |
3092 | location in each node, and the vector size is 4. I.e., we have a | |
3093 | a0b0c0a1b1c1... sequence and we need to create the following vectors: | |
ebfd146a IR |
3094 | for a's: a0a0a0a1 a1a1a2a2 a2a3a3a3 |
3095 | for b's: b0b0b0b1 b1b1b2b2 b2b3b3b3 | |
3096 | ... | |
3097 | ||
2635892a | 3098 | The masks for a's should be: {0,0,0,3} {3,3,6,6} {6,9,9,9}. |
b8698a0f | 3099 | The last mask is illegal since we assume two operands for permute |
ff802fa1 IR |
3100 | operation, and the mask element values can't be outside that range. |
3101 | Hence, the last mask must be converted into {2,5,5,5}. | |
b8698a0f | 3102 | For the first two permutations we need the first and the second input |
ebfd146a | 3103 | vectors: {a0,b0,c0,a1} and {b1,c1,a2,b2}, and for the last permutation |
b8698a0f | 3104 | we need the second and the third vectors: {b1,c1,a2,b2} and |
ebfd146a IR |
3105 | {c2,a3,b3,c3}. */ |
3106 | ||
ebfd146a IR |
3107 | { |
3108 | scalar_index = 0; | |
3109 | index = 0; | |
3110 | vect_stmts_counter = 0; | |
3111 | vec_index = 0; | |
3112 | first_vec_index = vec_index++; | |
3113 | if (only_one_vec) | |
3114 | second_vec_index = first_vec_index; | |
3115 | else | |
3116 | second_vec_index = vec_index++; | |
3117 | ||
3118 | for (j = 0; j < unroll_factor; j++) | |
3119 | { | |
3120 | for (k = 0; k < group_size; k++) | |
3121 | { | |
01d8bf07 | 3122 | i = SLP_TREE_LOAD_PERMUTATION (node)[k]; |
2635892a RH |
3123 | first_mask_element = i + j * group_size; |
3124 | if (!vect_get_mask_element (stmt, first_mask_element, 0, | |
3125 | nunits, only_one_vec, index, | |
3126 | mask, ¤t_mask_element, | |
3127 | &need_next_vector, | |
3128 | &number_of_mask_fixes, &mask_fixed, | |
3129 | &needs_first_vector)) | |
3130 | return false; | |
496d3346 | 3131 | gcc_assert (current_mask_element < 2 * nunits); |
2635892a | 3132 | mask[index++] = current_mask_element; |
ebfd146a | 3133 | |
2635892a | 3134 | if (index == nunits) |
ebfd146a | 3135 | { |
01d8bf07 | 3136 | index = 0; |
22e4dee7 RH |
3137 | if (!can_vec_perm_p (mode, false, mask)) |
3138 | { | |
73fbfcad | 3139 | if (dump_enabled_p ()) |
22e4dee7 | 3140 | { |
78c60e3d SS |
3141 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, |
3142 | vect_location, | |
3143 | "unsupported vect permute { "); | |
22e4dee7 | 3144 | for (i = 0; i < nunits; ++i) |
78c60e3d SS |
3145 | dump_printf (MSG_MISSED_OPTIMIZATION, "%d ", |
3146 | mask[i]); | |
3147 | dump_printf (MSG_MISSED_OPTIMIZATION, "}\n"); | |
22e4dee7 RH |
3148 | } |
3149 | return false; | |
3150 | } | |
3151 | ||
ebfd146a IR |
3152 | if (!analyze_only) |
3153 | { | |
01d8bf07 RB |
3154 | int l; |
3155 | tree mask_vec, *mask_elts; | |
3156 | mask_elts = XALLOCAVEC (tree, nunits); | |
3157 | for (l = 0; l < nunits; ++l) | |
3158 | mask_elts[l] = build_int_cst (mask_element_type, | |
3159 | mask[l]); | |
3160 | mask_vec = build_vector (mask_type, mask_elts); | |
3161 | ||
3162 | if (need_next_vector) | |
ebfd146a IR |
3163 | { |
3164 | first_vec_index = second_vec_index; | |
3165 | second_vec_index = vec_index; | |
3166 | } | |
3167 | ||
9771b263 DN |
3168 | next_scalar_stmt |
3169 | = SLP_TREE_SCALAR_STMTS (node)[scalar_index++]; | |
ebfd146a IR |
3170 | |
3171 | vect_create_mask_and_perm (stmt, next_scalar_stmt, | |
faf63e39 | 3172 | mask_vec, first_vec_index, second_vec_index, |
2635892a | 3173 | gsi, node, vectype, dr_chain, |
faf63e39 | 3174 | ncopies, vect_stmts_counter++); |
ebfd146a | 3175 | } |
b8698a0f L |
3176 | } |
3177 | } | |
3178 | } | |
3179 | } | |
ebfd146a | 3180 | |
ebfd146a IR |
3181 | return true; |
3182 | } | |
3183 | ||
3184 | ||
3185 | ||
3186 | /* Vectorize SLP instance tree in postorder. */ | |
3187 | ||
3188 | static bool | |
3189 | vect_schedule_slp_instance (slp_tree node, slp_instance instance, | |
a70d6342 | 3190 | unsigned int vectorization_factor) |
ebfd146a IR |
3191 | { |
3192 | gimple stmt; | |
0d0293ac | 3193 | bool grouped_store, is_store; |
ebfd146a IR |
3194 | gimple_stmt_iterator si; |
3195 | stmt_vec_info stmt_info; | |
3196 | unsigned int vec_stmts_size, nunits, group_size; | |
3197 | tree vectype; | |
3198 | int i; | |
d755c7ef | 3199 | slp_tree child; |
ebfd146a IR |
3200 | |
3201 | if (!node) | |
3202 | return false; | |
3203 | ||
9771b263 | 3204 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) |
d755c7ef | 3205 | vect_schedule_slp_instance (child, instance, vectorization_factor); |
b8698a0f | 3206 | |
9771b263 | 3207 | stmt = SLP_TREE_SCALAR_STMTS (node)[0]; |
ebfd146a IR |
3208 | stmt_info = vinfo_for_stmt (stmt); |
3209 | ||
3210 | /* VECTYPE is the type of the destination. */ | |
b690cc0f | 3211 | vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a IR |
3212 | nunits = (unsigned int) TYPE_VECTOR_SUBPARTS (vectype); |
3213 | group_size = SLP_INSTANCE_GROUP_SIZE (instance); | |
3214 | ||
3215 | /* For each SLP instance calculate number of vector stmts to be created | |
ff802fa1 | 3216 | for the scalar stmts in each node of the SLP tree. Number of vector |
ebfd146a IR |
3217 | elements in one vector iteration is the number of scalar elements in |
3218 | one scalar iteration (GROUP_SIZE) multiplied by VF divided by vector | |
3219 | size. */ | |
3220 | vec_stmts_size = (vectorization_factor * group_size) / nunits; | |
3221 | ||
9771b263 | 3222 | if (!SLP_TREE_VEC_STMTS (node).exists ()) |
ebfd146a | 3223 | { |
9771b263 | 3224 | SLP_TREE_VEC_STMTS (node).create (vec_stmts_size); |
ebfd146a IR |
3225 | SLP_TREE_NUMBER_OF_VEC_STMTS (node) = vec_stmts_size; |
3226 | } | |
3227 | ||
73fbfcad | 3228 | if (dump_enabled_p ()) |
ebfd146a | 3229 | { |
78c60e3d SS |
3230 | dump_printf_loc (MSG_NOTE,vect_location, |
3231 | "------>vectorizing SLP node starting from: "); | |
3232 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 3233 | dump_printf (MSG_NOTE, "\n"); |
b8698a0f | 3234 | } |
ebfd146a IR |
3235 | |
3236 | /* Loads should be inserted before the first load. */ | |
3237 | if (SLP_INSTANCE_FIRST_LOAD_STMT (instance) | |
0d0293ac | 3238 | && STMT_VINFO_GROUPED_ACCESS (stmt_info) |
6aa904c4 | 3239 | && !REFERENCE_CLASS_P (gimple_get_lhs (stmt)) |
01d8bf07 | 3240 | && SLP_TREE_LOAD_PERMUTATION (node).exists ()) |
ebfd146a | 3241 | si = gsi_for_stmt (SLP_INSTANCE_FIRST_LOAD_STMT (instance)); |
9d5e7640 | 3242 | else if (is_pattern_stmt_p (stmt_info)) |
6aa904c4 | 3243 | si = gsi_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
ebfd146a IR |
3244 | else |
3245 | si = gsi_for_stmt (stmt); | |
b8698a0f | 3246 | |
e4a707c4 | 3247 | /* Stores should be inserted just before the last store. */ |
0d0293ac | 3248 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) |
e4a707c4 IR |
3249 | && REFERENCE_CLASS_P (gimple_get_lhs (stmt))) |
3250 | { | |
3251 | gimple last_store = vect_find_last_store_in_slp_instance (instance); | |
a024e70e IR |
3252 | if (is_pattern_stmt_p (vinfo_for_stmt (last_store))) |
3253 | last_store = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (last_store)); | |
e4a707c4 IR |
3254 | si = gsi_for_stmt (last_store); |
3255 | } | |
3256 | ||
b010117a IR |
3257 | /* Mark the first element of the reduction chain as reduction to properly |
3258 | transform the node. In the analysis phase only the last element of the | |
3259 | chain is marked as reduction. */ | |
0d0293ac | 3260 | if (GROUP_FIRST_ELEMENT (stmt_info) && !STMT_VINFO_GROUPED_ACCESS (stmt_info) |
b010117a IR |
3261 | && GROUP_FIRST_ELEMENT (stmt_info) == stmt) |
3262 | { | |
3263 | STMT_VINFO_DEF_TYPE (stmt_info) = vect_reduction_def; | |
3264 | STMT_VINFO_TYPE (stmt_info) = reduc_vec_info_type; | |
3265 | } | |
3266 | ||
0d0293ac | 3267 | is_store = vect_transform_stmt (stmt, &si, &grouped_store, node, instance); |
b5aeb3bb | 3268 | return is_store; |
ebfd146a IR |
3269 | } |
3270 | ||
dd34c087 JJ |
3271 | /* Replace scalar calls from SLP node NODE with setting of their lhs to zero. |
3272 | For loop vectorization this is done in vectorizable_call, but for SLP | |
3273 | it needs to be deferred until end of vect_schedule_slp, because multiple | |
3274 | SLP instances may refer to the same scalar stmt. */ | |
3275 | ||
3276 | static void | |
3277 | vect_remove_slp_scalar_calls (slp_tree node) | |
3278 | { | |
3279 | gimple stmt, new_stmt; | |
3280 | gimple_stmt_iterator gsi; | |
3281 | int i; | |
d755c7ef | 3282 | slp_tree child; |
dd34c087 JJ |
3283 | tree lhs; |
3284 | stmt_vec_info stmt_info; | |
3285 | ||
3286 | if (!node) | |
3287 | return; | |
3288 | ||
9771b263 | 3289 | FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) |
d755c7ef | 3290 | vect_remove_slp_scalar_calls (child); |
dd34c087 | 3291 | |
9771b263 | 3292 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt) |
dd34c087 JJ |
3293 | { |
3294 | if (!is_gimple_call (stmt) || gimple_bb (stmt) == NULL) | |
3295 | continue; | |
3296 | stmt_info = vinfo_for_stmt (stmt); | |
3297 | if (stmt_info == NULL | |
3298 | || is_pattern_stmt_p (stmt_info) | |
3299 | || !PURE_SLP_STMT (stmt_info)) | |
3300 | continue; | |
3301 | lhs = gimple_call_lhs (stmt); | |
3302 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
3303 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
3304 | set_vinfo_for_stmt (stmt, NULL); | |
3305 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
3306 | gsi = gsi_for_stmt (stmt); | |
3307 | gsi_replace (&gsi, new_stmt, false); | |
3308 | SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt; | |
3309 | } | |
3310 | } | |
ebfd146a | 3311 | |
ff802fa1 IR |
3312 | /* Generate vector code for all SLP instances in the loop/basic block. */ |
3313 | ||
ebfd146a | 3314 | bool |
a70d6342 | 3315 | vect_schedule_slp (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo) |
ebfd146a | 3316 | { |
9771b263 | 3317 | vec<slp_instance> slp_instances; |
ebfd146a | 3318 | slp_instance instance; |
01d8bf07 | 3319 | unsigned int i, vf; |
ebfd146a IR |
3320 | bool is_store = false; |
3321 | ||
a70d6342 IR |
3322 | if (loop_vinfo) |
3323 | { | |
3324 | slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo); | |
3325 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
b8698a0f | 3326 | } |
a70d6342 IR |
3327 | else |
3328 | { | |
3329 | slp_instances = BB_VINFO_SLP_INSTANCES (bb_vinfo); | |
3330 | vf = 1; | |
b8698a0f | 3331 | } |
a70d6342 | 3332 | |
9771b263 | 3333 | FOR_EACH_VEC_ELT (slp_instances, i, instance) |
ebfd146a IR |
3334 | { |
3335 | /* Schedule the tree of INSTANCE. */ | |
3336 | is_store = vect_schedule_slp_instance (SLP_INSTANCE_TREE (instance), | |
a70d6342 | 3337 | instance, vf); |
73fbfcad | 3338 | if (dump_enabled_p ()) |
78c60e3d | 3339 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3340 | "vectorizing stmts using SLP.\n"); |
ebfd146a IR |
3341 | } |
3342 | ||
9771b263 | 3343 | FOR_EACH_VEC_ELT (slp_instances, i, instance) |
b5aeb3bb IR |
3344 | { |
3345 | slp_tree root = SLP_INSTANCE_TREE (instance); | |
3346 | gimple store; | |
3347 | unsigned int j; | |
3348 | gimple_stmt_iterator gsi; | |
3349 | ||
c40eced0 RB |
3350 | /* Remove scalar call stmts. Do not do this for basic-block |
3351 | vectorization as not all uses may be vectorized. | |
3352 | ??? Why should this be necessary? DCE should be able to | |
3353 | remove the stmts itself. | |
3354 | ??? For BB vectorization we can as well remove scalar | |
3355 | stmts starting from the SLP tree root if they have no | |
3356 | uses. */ | |
3357 | if (loop_vinfo) | |
3358 | vect_remove_slp_scalar_calls (root); | |
dd34c087 | 3359 | |
9771b263 | 3360 | for (j = 0; SLP_TREE_SCALAR_STMTS (root).iterate (j, &store) |
b5aeb3bb IR |
3361 | && j < SLP_INSTANCE_GROUP_SIZE (instance); j++) |
3362 | { | |
3363 | if (!STMT_VINFO_DATA_REF (vinfo_for_stmt (store))) | |
3364 | break; | |
3365 | ||
a024e70e IR |
3366 | if (is_pattern_stmt_p (vinfo_for_stmt (store))) |
3367 | store = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (store)); | |
b5aeb3bb IR |
3368 | /* Free the attached stmt_vec_info and remove the stmt. */ |
3369 | gsi = gsi_for_stmt (store); | |
3d3f2249 | 3370 | unlink_stmt_vdef (store); |
b5aeb3bb | 3371 | gsi_remove (&gsi, true); |
3d3f2249 | 3372 | release_defs (store); |
b5aeb3bb IR |
3373 | free_stmt_vec_info (store); |
3374 | } | |
3375 | } | |
3376 | ||
ebfd146a IR |
3377 | return is_store; |
3378 | } | |
a70d6342 IR |
3379 | |
3380 | ||
3381 | /* Vectorize the basic block. */ | |
3382 | ||
3383 | void | |
3384 | vect_slp_transform_bb (basic_block bb) | |
3385 | { | |
3386 | bb_vec_info bb_vinfo = vec_info_for_bb (bb); | |
3387 | gimple_stmt_iterator si; | |
3388 | ||
3389 | gcc_assert (bb_vinfo); | |
3390 | ||
73fbfcad | 3391 | if (dump_enabled_p ()) |
78c60e3d | 3392 | dump_printf_loc (MSG_NOTE, vect_location, "SLPing BB\n"); |
a70d6342 IR |
3393 | |
3394 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
3395 | { | |
3396 | gimple stmt = gsi_stmt (si); | |
3397 | stmt_vec_info stmt_info; | |
3398 | ||
73fbfcad | 3399 | if (dump_enabled_p ()) |
a70d6342 | 3400 | { |
78c60e3d SS |
3401 | dump_printf_loc (MSG_NOTE, vect_location, |
3402 | "------>SLPing statement: "); | |
3403 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 3404 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
3405 | } |
3406 | ||
3407 | stmt_info = vinfo_for_stmt (stmt); | |
3408 | gcc_assert (stmt_info); | |
3409 | ||
3410 | /* Schedule all the SLP instances when the first SLP stmt is reached. */ | |
3411 | if (STMT_SLP_TYPE (stmt_info)) | |
3412 | { | |
3413 | vect_schedule_slp (NULL, bb_vinfo); | |
3414 | break; | |
3415 | } | |
3416 | } | |
3417 | ||
73fbfcad | 3418 | if (dump_enabled_p ()) |
5d318fd4 | 3419 | dump_printf_loc (MSG_NOTE, vect_location, |
ccb3ad87 | 3420 | "BASIC BLOCK VECTORIZED\n"); |
a70d6342 | 3421 | |
12aaf609 IR |
3422 | destroy_bb_vec_info (bb_vinfo); |
3423 | } |